6 :Author: Tejun Heo <tj@kernel.org>
8 This is the authoritative documentation on the design, interface and
9 conventions of cgroup v2. It describes all userland-visible aspects
10 of cgroup including core and specific controller behaviors. All
11 future changes must be reflected in this document. Documentation for
12 v1 is available under Documentation/cgroup-v1/.
21 2-2. Organizing Processes and Threads
24 2-3. [Un]populated Notification
25 2-4. Controlling Controllers
26 2-4-1. Enabling and Disabling
27 2-4-2. Top-down Constraint
28 2-4-3. No Internal Process Constraint
30 2-5-1. Model of Delegation
31 2-5-2. Delegation Containment
33 2-6-1. Organize Once and Control
34 2-6-2. Avoid Name Collisions
35 3. Resource Distribution Models
43 4-3. Core Interface Files
46 5-1-1. CPU Interface Files
48 5-2-1. Memory Interface Files
49 5-2-2. Usage Guidelines
50 5-2-3. Memory Ownership
52 5-3-1. IO Interface Files
55 5-4-1. PID Interface Files
58 5-6-1. RDMA Interface Files
61 5-N. Non-normative information
62 5-N-1. CPU controller root cgroup process behaviour
63 5-N-2. IO controller root cgroup process behaviour
66 6-2. The Root and Views
67 6-3. Migration and setns(2)
68 6-4. Interaction with Other Namespaces
69 P. Information on Kernel Programming
70 P-1. Filesystem Support for Writeback
71 D. Deprecated v1 Core Features
72 R. Issues with v1 and Rationales for v2
73 R-1. Multiple Hierarchies
74 R-2. Thread Granularity
75 R-3. Competition Between Inner Nodes and Threads
76 R-4. Other Interface Issues
77 R-5. Controller Issues and Remedies
87 "cgroup" stands for "control group" and is never capitalized. The
88 singular form is used to designate the whole feature and also as a
89 qualifier as in "cgroup controllers". When explicitly referring to
90 multiple individual control groups, the plural form "cgroups" is used.
96 cgroup is a mechanism to organize processes hierarchically and
97 distribute system resources along the hierarchy in a controlled and
100 cgroup is largely composed of two parts - the core and controllers.
101 cgroup core is primarily responsible for hierarchically organizing
102 processes. A cgroup controller is usually responsible for
103 distributing a specific type of system resource along the hierarchy
104 although there are utility controllers which serve purposes other than
105 resource distribution.
107 cgroups form a tree structure and every process in the system belongs
108 to one and only one cgroup. All threads of a process belong to the
109 same cgroup. On creation, all processes are put in the cgroup that
110 the parent process belongs to at the time. A process can be migrated
111 to another cgroup. Migration of a process doesn't affect already
112 existing descendant processes.
114 Following certain structural constraints, controllers may be enabled or
115 disabled selectively on a cgroup. All controller behaviors are
116 hierarchical - if a controller is enabled on a cgroup, it affects all
117 processes which belong to the cgroups consisting the inclusive
118 sub-hierarchy of the cgroup. When a controller is enabled on a nested
119 cgroup, it always restricts the resource distribution further. The
120 restrictions set closer to the root in the hierarchy can not be
121 overridden from further away.
130 Unlike v1, cgroup v2 has only single hierarchy. The cgroup v2
131 hierarchy can be mounted with the following mount command::
133 # mount -t cgroup2 none $MOUNT_POINT
135 cgroup2 filesystem has the magic number 0x63677270 ("cgrp"). All
136 controllers which support v2 and are not bound to a v1 hierarchy are
137 automatically bound to the v2 hierarchy and show up at the root.
138 Controllers which are not in active use in the v2 hierarchy can be
139 bound to other hierarchies. This allows mixing v2 hierarchy with the
140 legacy v1 multiple hierarchies in a fully backward compatible way.
142 A controller can be moved across hierarchies only after the controller
143 is no longer referenced in its current hierarchy. Because per-cgroup
144 controller states are destroyed asynchronously and controllers may
145 have lingering references, a controller may not show up immediately on
146 the v2 hierarchy after the final umount of the previous hierarchy.
147 Similarly, a controller should be fully disabled to be moved out of
148 the unified hierarchy and it may take some time for the disabled
149 controller to become available for other hierarchies; furthermore, due
150 to inter-controller dependencies, other controllers may need to be
153 While useful for development and manual configurations, moving
154 controllers dynamically between the v2 and other hierarchies is
155 strongly discouraged for production use. It is recommended to decide
156 the hierarchies and controller associations before starting using the
157 controllers after system boot.
159 During transition to v2, system management software might still
160 automount the v1 cgroup filesystem and so hijack all controllers
161 during boot, before manual intervention is possible. To make testing
162 and experimenting easier, the kernel parameter cgroup_no_v1= allows
163 disabling controllers in v1 and make them always available in v2.
165 cgroup v2 currently supports the following mount options.
169 Consider cgroup namespaces as delegation boundaries. This
170 option is system wide and can only be set on mount or modified
171 through remount from the init namespace. The mount option is
172 ignored on non-init namespace mounts. Please refer to the
173 Delegation section for details.
176 Organizing Processes and Threads
177 --------------------------------
182 Initially, only the root cgroup exists to which all processes belong.
183 A child cgroup can be created by creating a sub-directory::
187 A given cgroup may have multiple child cgroups forming a tree
188 structure. Each cgroup has a read-writable interface file
189 "cgroup.procs". When read, it lists the PIDs of all processes which
190 belong to the cgroup one-per-line. The PIDs are not ordered and the
191 same PID may show up more than once if the process got moved to
192 another cgroup and then back or the PID got recycled while reading.
194 A process can be migrated into a cgroup by writing its PID to the
195 target cgroup's "cgroup.procs" file. Only one process can be migrated
196 on a single write(2) call. If a process is composed of multiple
197 threads, writing the PID of any thread migrates all threads of the
200 When a process forks a child process, the new process is born into the
201 cgroup that the forking process belongs to at the time of the
202 operation. After exit, a process stays associated with the cgroup
203 that it belonged to at the time of exit until it's reaped; however, a
204 zombie process does not appear in "cgroup.procs" and thus can't be
205 moved to another cgroup.
207 A cgroup which doesn't have any children or live processes can be
208 destroyed by removing the directory. Note that a cgroup which doesn't
209 have any children and is associated only with zombie processes is
210 considered empty and can be removed::
214 "/proc/$PID/cgroup" lists a process's cgroup membership. If legacy
215 cgroup is in use in the system, this file may contain multiple lines,
216 one for each hierarchy. The entry for cgroup v2 is always in the
219 # cat /proc/842/cgroup
221 0::/test-cgroup/test-cgroup-nested
223 If the process becomes a zombie and the cgroup it was associated with
224 is removed subsequently, " (deleted)" is appended to the path::
226 # cat /proc/842/cgroup
228 0::/test-cgroup/test-cgroup-nested (deleted)
234 cgroup v2 supports thread granularity for a subset of controllers to
235 support use cases requiring hierarchical resource distribution across
236 the threads of a group of processes. By default, all threads of a
237 process belong to the same cgroup, which also serves as the resource
238 domain to host resource consumptions which are not specific to a
239 process or thread. The thread mode allows threads to be spread across
240 a subtree while still maintaining the common resource domain for them.
242 Controllers which support thread mode are called threaded controllers.
243 The ones which don't are called domain controllers.
245 Marking a cgroup threaded makes it join the resource domain of its
246 parent as a threaded cgroup. The parent may be another threaded
247 cgroup whose resource domain is further up in the hierarchy. The root
248 of a threaded subtree, that is, the nearest ancestor which is not
249 threaded, is called threaded domain or thread root interchangeably and
250 serves as the resource domain for the entire subtree.
252 Inside a threaded subtree, threads of a process can be put in
253 different cgroups and are not subject to the no internal process
254 constraint - threaded controllers can be enabled on non-leaf cgroups
255 whether they have threads in them or not.
257 As the threaded domain cgroup hosts all the domain resource
258 consumptions of the subtree, it is considered to have internal
259 resource consumptions whether there are processes in it or not and
260 can't have populated child cgroups which aren't threaded. Because the
261 root cgroup is not subject to no internal process constraint, it can
262 serve both as a threaded domain and a parent to domain cgroups.
264 The current operation mode or type of the cgroup is shown in the
265 "cgroup.type" file which indicates whether the cgroup is a normal
266 domain, a domain which is serving as the domain of a threaded subtree,
267 or a threaded cgroup.
269 On creation, a cgroup is always a domain cgroup and can be made
270 threaded by writing "threaded" to the "cgroup.type" file. The
271 operation is single direction::
273 # echo threaded > cgroup.type
275 Once threaded, the cgroup can't be made a domain again. To enable the
276 thread mode, the following conditions must be met.
278 - As the cgroup will join the parent's resource domain. The parent
279 must either be a valid (threaded) domain or a threaded cgroup.
281 - When the parent is an unthreaded domain, it must not have any domain
282 controllers enabled or populated domain children. The root is
283 exempt from this requirement.
285 Topology-wise, a cgroup can be in an invalid state. Please consider
286 the following topology::
288 A (threaded domain) - B (threaded) - C (domain, just created)
290 C is created as a domain but isn't connected to a parent which can
291 host child domains. C can't be used until it is turned into a
292 threaded cgroup. "cgroup.type" file will report "domain (invalid)" in
293 these cases. Operations which fail due to invalid topology use
294 EOPNOTSUPP as the errno.
296 A domain cgroup is turned into a threaded domain when one of its child
297 cgroup becomes threaded or threaded controllers are enabled in the
298 "cgroup.subtree_control" file while there are processes in the cgroup.
299 A threaded domain reverts to a normal domain when the conditions
302 When read, "cgroup.threads" contains the list of the thread IDs of all
303 threads in the cgroup. Except that the operations are per-thread
304 instead of per-process, "cgroup.threads" has the same format and
305 behaves the same way as "cgroup.procs". While "cgroup.threads" can be
306 written to in any cgroup, as it can only move threads inside the same
307 threaded domain, its operations are confined inside each threaded
310 The threaded domain cgroup serves as the resource domain for the whole
311 subtree, and, while the threads can be scattered across the subtree,
312 all the processes are considered to be in the threaded domain cgroup.
313 "cgroup.procs" in a threaded domain cgroup contains the PIDs of all
314 processes in the subtree and is not readable in the subtree proper.
315 However, "cgroup.procs" can be written to from anywhere in the subtree
316 to migrate all threads of the matching process to the cgroup.
318 Only threaded controllers can be enabled in a threaded subtree. When
319 a threaded controller is enabled inside a threaded subtree, it only
320 accounts for and controls resource consumptions associated with the
321 threads in the cgroup and its descendants. All consumptions which
322 aren't tied to a specific thread belong to the threaded domain cgroup.
324 Because a threaded subtree is exempt from no internal process
325 constraint, a threaded controller must be able to handle competition
326 between threads in a non-leaf cgroup and its child cgroups. Each
327 threaded controller defines how such competitions are handled.
330 [Un]populated Notification
331 --------------------------
333 Each non-root cgroup has a "cgroup.events" file which contains
334 "populated" field indicating whether the cgroup's sub-hierarchy has
335 live processes in it. Its value is 0 if there is no live process in
336 the cgroup and its descendants; otherwise, 1. poll and [id]notify
337 events are triggered when the value changes. This can be used, for
338 example, to start a clean-up operation after all processes of a given
339 sub-hierarchy have exited. The populated state updates and
340 notifications are recursive. Consider the following sub-hierarchy
341 where the numbers in the parentheses represent the numbers of processes
347 A, B and C's "populated" fields would be 1 while D's 0. After the one
348 process in C exits, B and C's "populated" fields would flip to "0" and
349 file modified events will be generated on the "cgroup.events" files of
353 Controlling Controllers
354 -----------------------
356 Enabling and Disabling
357 ~~~~~~~~~~~~~~~~~~~~~~
359 Each cgroup has a "cgroup.controllers" file which lists all
360 controllers available for the cgroup to enable::
362 # cat cgroup.controllers
365 No controller is enabled by default. Controllers can be enabled and
366 disabled by writing to the "cgroup.subtree_control" file::
368 # echo "+cpu +memory -io" > cgroup.subtree_control
370 Only controllers which are listed in "cgroup.controllers" can be
371 enabled. When multiple operations are specified as above, either they
372 all succeed or fail. If multiple operations on the same controller
373 are specified, the last one is effective.
375 Enabling a controller in a cgroup indicates that the distribution of
376 the target resource across its immediate children will be controlled.
377 Consider the following sub-hierarchy. The enabled controllers are
378 listed in parentheses::
380 A(cpu,memory) - B(memory) - C()
383 As A has "cpu" and "memory" enabled, A will control the distribution
384 of CPU cycles and memory to its children, in this case, B. As B has
385 "memory" enabled but not "CPU", C and D will compete freely on CPU
386 cycles but their division of memory available to B will be controlled.
388 As a controller regulates the distribution of the target resource to
389 the cgroup's children, enabling it creates the controller's interface
390 files in the child cgroups. In the above example, enabling "cpu" on B
391 would create the "cpu." prefixed controller interface files in C and
392 D. Likewise, disabling "memory" from B would remove the "memory."
393 prefixed controller interface files from C and D. This means that the
394 controller interface files - anything which doesn't start with
395 "cgroup." are owned by the parent rather than the cgroup itself.
401 Resources are distributed top-down and a cgroup can further distribute
402 a resource only if the resource has been distributed to it from the
403 parent. This means that all non-root "cgroup.subtree_control" files
404 can only contain controllers which are enabled in the parent's
405 "cgroup.subtree_control" file. A controller can be enabled only if
406 the parent has the controller enabled and a controller can't be
407 disabled if one or more children have it enabled.
410 No Internal Process Constraint
411 ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
413 Non-root cgroups can distribute domain resources to their children
414 only when they don't have any processes of their own. In other words,
415 only domain cgroups which don't contain any processes can have domain
416 controllers enabled in their "cgroup.subtree_control" files.
418 This guarantees that, when a domain controller is looking at the part
419 of the hierarchy which has it enabled, processes are always only on
420 the leaves. This rules out situations where child cgroups compete
421 against internal processes of the parent.
423 The root cgroup is exempt from this restriction. Root contains
424 processes and anonymous resource consumption which can't be associated
425 with any other cgroups and requires special treatment from most
426 controllers. How resource consumption in the root cgroup is governed
427 is up to each controller (for more information on this topic please
428 refer to the Non-normative information section in the Controllers
431 Note that the restriction doesn't get in the way if there is no
432 enabled controller in the cgroup's "cgroup.subtree_control". This is
433 important as otherwise it wouldn't be possible to create children of a
434 populated cgroup. To control resource distribution of a cgroup, the
435 cgroup must create children and transfer all its processes to the
436 children before enabling controllers in its "cgroup.subtree_control"
446 A cgroup can be delegated in two ways. First, to a less privileged
447 user by granting write access of the directory and its "cgroup.procs",
448 "cgroup.threads" and "cgroup.subtree_control" files to the user.
449 Second, if the "nsdelegate" mount option is set, automatically to a
450 cgroup namespace on namespace creation.
452 Because the resource control interface files in a given directory
453 control the distribution of the parent's resources, the delegatee
454 shouldn't be allowed to write to them. For the first method, this is
455 achieved by not granting access to these files. For the second, the
456 kernel rejects writes to all files other than "cgroup.procs" and
457 "cgroup.subtree_control" on a namespace root from inside the
460 The end results are equivalent for both delegation types. Once
461 delegated, the user can build sub-hierarchy under the directory,
462 organize processes inside it as it sees fit and further distribute the
463 resources it received from the parent. The limits and other settings
464 of all resource controllers are hierarchical and regardless of what
465 happens in the delegated sub-hierarchy, nothing can escape the
466 resource restrictions imposed by the parent.
468 Currently, cgroup doesn't impose any restrictions on the number of
469 cgroups in or nesting depth of a delegated sub-hierarchy; however,
470 this may be limited explicitly in the future.
473 Delegation Containment
474 ~~~~~~~~~~~~~~~~~~~~~~
476 A delegated sub-hierarchy is contained in the sense that processes
477 can't be moved into or out of the sub-hierarchy by the delegatee.
479 For delegations to a less privileged user, this is achieved by
480 requiring the following conditions for a process with a non-root euid
481 to migrate a target process into a cgroup by writing its PID to the
484 - The writer must have write access to the "cgroup.procs" file.
486 - The writer must have write access to the "cgroup.procs" file of the
487 common ancestor of the source and destination cgroups.
489 The above two constraints ensure that while a delegatee may migrate
490 processes around freely in the delegated sub-hierarchy it can't pull
491 in from or push out to outside the sub-hierarchy.
493 For an example, let's assume cgroups C0 and C1 have been delegated to
494 user U0 who created C00, C01 under C0 and C10 under C1 as follows and
495 all processes under C0 and C1 belong to U0::
497 ~~~~~~~~~~~~~ - C0 - C00
500 ~~~~~~~~~~~~~ - C1 - C10
502 Let's also say U0 wants to write the PID of a process which is
503 currently in C10 into "C00/cgroup.procs". U0 has write access to the
504 file; however, the common ancestor of the source cgroup C10 and the
505 destination cgroup C00 is above the points of delegation and U0 would
506 not have write access to its "cgroup.procs" files and thus the write
507 will be denied with -EACCES.
509 For delegations to namespaces, containment is achieved by requiring
510 that both the source and destination cgroups are reachable from the
511 namespace of the process which is attempting the migration. If either
512 is not reachable, the migration is rejected with -ENOENT.
518 Organize Once and Control
519 ~~~~~~~~~~~~~~~~~~~~~~~~~
521 Migrating a process across cgroups is a relatively expensive operation
522 and stateful resources such as memory are not moved together with the
523 process. This is an explicit design decision as there often exist
524 inherent trade-offs between migration and various hot paths in terms
525 of synchronization cost.
527 As such, migrating processes across cgroups frequently as a means to
528 apply different resource restrictions is discouraged. A workload
529 should be assigned to a cgroup according to the system's logical and
530 resource structure once on start-up. Dynamic adjustments to resource
531 distribution can be made by changing controller configuration through
535 Avoid Name Collisions
536 ~~~~~~~~~~~~~~~~~~~~~
538 Interface files for a cgroup and its children cgroups occupy the same
539 directory and it is possible to create children cgroups which collide
540 with interface files.
542 All cgroup core interface files are prefixed with "cgroup." and each
543 controller's interface files are prefixed with the controller name and
544 a dot. A controller's name is composed of lower case alphabets and
545 '_'s but never begins with an '_' so it can be used as the prefix
546 character for collision avoidance. Also, interface file names won't
547 start or end with terms which are often used in categorizing workloads
548 such as job, service, slice, unit or workload.
550 cgroup doesn't do anything to prevent name collisions and it's the
551 user's responsibility to avoid them.
554 Resource Distribution Models
555 ============================
557 cgroup controllers implement several resource distribution schemes
558 depending on the resource type and expected use cases. This section
559 describes major schemes in use along with their expected behaviors.
565 A parent's resource is distributed by adding up the weights of all
566 active children and giving each the fraction matching the ratio of its
567 weight against the sum. As only children which can make use of the
568 resource at the moment participate in the distribution, this is
569 work-conserving. Due to the dynamic nature, this model is usually
570 used for stateless resources.
572 All weights are in the range [1, 10000] with the default at 100. This
573 allows symmetric multiplicative biases in both directions at fine
574 enough granularity while staying in the intuitive range.
576 As long as the weight is in range, all configuration combinations are
577 valid and there is no reason to reject configuration changes or
580 "cpu.weight" proportionally distributes CPU cycles to active children
581 and is an example of this type.
587 A child can only consume upto the configured amount of the resource.
588 Limits can be over-committed - the sum of the limits of children can
589 exceed the amount of resource available to the parent.
591 Limits are in the range [0, max] and defaults to "max", which is noop.
593 As limits can be over-committed, all configuration combinations are
594 valid and there is no reason to reject configuration changes or
597 "io.max" limits the maximum BPS and/or IOPS that a cgroup can consume
598 on an IO device and is an example of this type.
604 A cgroup is protected to be allocated upto the configured amount of
605 the resource if the usages of all its ancestors are under their
606 protected levels. Protections can be hard guarantees or best effort
607 soft boundaries. Protections can also be over-committed in which case
608 only upto the amount available to the parent is protected among
611 Protections are in the range [0, max] and defaults to 0, which is
614 As protections can be over-committed, all configuration combinations
615 are valid and there is no reason to reject configuration changes or
618 "memory.low" implements best-effort memory protection and is an
619 example of this type.
625 A cgroup is exclusively allocated a certain amount of a finite
626 resource. Allocations can't be over-committed - the sum of the
627 allocations of children can not exceed the amount of resource
628 available to the parent.
630 Allocations are in the range [0, max] and defaults to 0, which is no
633 As allocations can't be over-committed, some configuration
634 combinations are invalid and should be rejected. Also, if the
635 resource is mandatory for execution of processes, process migrations
638 "cpu.rt.max" hard-allocates realtime slices and is an example of this
648 All interface files should be in one of the following formats whenever
651 New-line separated values
652 (when only one value can be written at once)
658 Space separated values
659 (when read-only or multiple values can be written at once)
671 KEY0 SUB_KEY0=VAL00 SUB_KEY1=VAL01...
672 KEY1 SUB_KEY0=VAL10 SUB_KEY1=VAL11...
675 For a writable file, the format for writing should generally match
676 reading; however, controllers may allow omitting later fields or
677 implement restricted shortcuts for most common use cases.
679 For both flat and nested keyed files, only the values for a single key
680 can be written at a time. For nested keyed files, the sub key pairs
681 may be specified in any order and not all pairs have to be specified.
687 - Settings for a single feature should be contained in a single file.
689 - The root cgroup should be exempt from resource control and thus
690 shouldn't have resource control interface files. Also,
691 informational files on the root cgroup which end up showing global
692 information available elsewhere shouldn't exist.
694 - If a controller implements weight based resource distribution, its
695 interface file should be named "weight" and have the range [1,
696 10000] with 100 as the default. The values are chosen to allow
697 enough and symmetric bias in both directions while keeping it
698 intuitive (the default is 100%).
700 - If a controller implements an absolute resource guarantee and/or
701 limit, the interface files should be named "min" and "max"
702 respectively. If a controller implements best effort resource
703 guarantee and/or limit, the interface files should be named "low"
704 and "high" respectively.
706 In the above four control files, the special token "max" should be
707 used to represent upward infinity for both reading and writing.
709 - If a setting has a configurable default value and keyed specific
710 overrides, the default entry should be keyed with "default" and
711 appear as the first entry in the file.
713 The default value can be updated by writing either "default $VAL" or
716 When writing to update a specific override, "default" can be used as
717 the value to indicate removal of the override. Override entries
718 with "default" as the value must not appear when read.
720 For example, a setting which is keyed by major:minor device numbers
721 with integer values may look like the following::
723 # cat cgroup-example-interface-file
727 The default value can be updated by::
729 # echo 125 > cgroup-example-interface-file
733 # echo "default 125" > cgroup-example-interface-file
735 An override can be set by::
737 # echo "8:16 170" > cgroup-example-interface-file
741 # echo "8:0 default" > cgroup-example-interface-file
742 # cat cgroup-example-interface-file
746 - For events which are not very high frequency, an interface file
747 "events" should be created which lists event key value pairs.
748 Whenever a notifiable event happens, file modified event should be
749 generated on the file.
755 All cgroup core files are prefixed with "cgroup."
759 A read-write single value file which exists on non-root
762 When read, it indicates the current type of the cgroup, which
763 can be one of the following values.
765 - "domain" : A normal valid domain cgroup.
767 - "domain threaded" : A threaded domain cgroup which is
768 serving as the root of a threaded subtree.
770 - "domain invalid" : A cgroup which is in an invalid state.
771 It can't be populated or have controllers enabled. It may
772 be allowed to become a threaded cgroup.
774 - "threaded" : A threaded cgroup which is a member of a
777 A cgroup can be turned into a threaded cgroup by writing
778 "threaded" to this file.
781 A read-write new-line separated values file which exists on
784 When read, it lists the PIDs of all processes which belong to
785 the cgroup one-per-line. The PIDs are not ordered and the
786 same PID may show up more than once if the process got moved
787 to another cgroup and then back or the PID got recycled while
790 A PID can be written to migrate the process associated with
791 the PID to the cgroup. The writer should match all of the
792 following conditions.
794 - It must have write access to the "cgroup.procs" file.
796 - It must have write access to the "cgroup.procs" file of the
797 common ancestor of the source and destination cgroups.
799 When delegating a sub-hierarchy, write access to this file
800 should be granted along with the containing directory.
802 In a threaded cgroup, reading this file fails with EOPNOTSUPP
803 as all the processes belong to the thread root. Writing is
804 supported and moves every thread of the process to the cgroup.
807 A read-write new-line separated values file which exists on
810 When read, it lists the TIDs of all threads which belong to
811 the cgroup one-per-line. The TIDs are not ordered and the
812 same TID may show up more than once if the thread got moved to
813 another cgroup and then back or the TID got recycled while
816 A TID can be written to migrate the thread associated with the
817 TID to the cgroup. The writer should match all of the
818 following conditions.
820 - It must have write access to the "cgroup.threads" file.
822 - The cgroup that the thread is currently in must be in the
823 same resource domain as the destination cgroup.
825 - It must have write access to the "cgroup.procs" file of the
826 common ancestor of the source and destination cgroups.
828 When delegating a sub-hierarchy, write access to this file
829 should be granted along with the containing directory.
832 A read-only space separated values file which exists on all
835 It shows space separated list of all controllers available to
836 the cgroup. The controllers are not ordered.
838 cgroup.subtree_control
839 A read-write space separated values file which exists on all
840 cgroups. Starts out empty.
842 When read, it shows space separated list of the controllers
843 which are enabled to control resource distribution from the
844 cgroup to its children.
846 Space separated list of controllers prefixed with '+' or '-'
847 can be written to enable or disable controllers. A controller
848 name prefixed with '+' enables the controller and '-'
849 disables. If a controller appears more than once on the list,
850 the last one is effective. When multiple enable and disable
851 operations are specified, either all succeed or all fail.
854 A read-only flat-keyed file which exists on non-root cgroups.
855 The following entries are defined. Unless specified
856 otherwise, a value change in this file generates a file
860 1 if the cgroup or its descendants contains any live
861 processes; otherwise, 0.
863 cgroup.max.descendants
864 A read-write single value files. The default is "max".
866 Maximum allowed number of descent cgroups.
867 If the actual number of descendants is equal or larger,
868 an attempt to create a new cgroup in the hierarchy will fail.
871 A read-write single value files. The default is "max".
873 Maximum allowed descent depth below the current cgroup.
874 If the actual descent depth is equal or larger,
875 an attempt to create a new child cgroup will fail.
878 A read-only flat-keyed file with the following entries:
881 Total number of visible descendant cgroups.
884 Total number of dying descendant cgroups. A cgroup becomes
885 dying after being deleted by a user. The cgroup will remain
886 in dying state for some time undefined time (which can depend
887 on system load) before being completely destroyed.
889 A process can't enter a dying cgroup under any circumstances,
890 a dying cgroup can't revive.
892 A dying cgroup can consume system resources not exceeding
893 limits, which were active at the moment of cgroup deletion.
902 The "cpu" controllers regulates distribution of CPU cycles. This
903 controller implements weight and absolute bandwidth limit models for
904 normal scheduling policy and absolute bandwidth allocation model for
905 realtime scheduling policy.
907 WARNING: cgroup2 doesn't yet support control of realtime processes and
908 the cpu controller can only be enabled when all RT processes are in
909 the root cgroup. Be aware that system management software may already
910 have placed RT processes into nonroot cgroups during the system boot
911 process, and these processes may need to be moved to the root cgroup
912 before the cpu controller can be enabled.
918 All time durations are in microseconds.
921 A read-only flat-keyed file which exists on non-root cgroups.
922 This file exists whether the controller is enabled or not.
924 It always reports the following three stats:
930 and the following three when the controller is enabled:
937 A read-write single value file which exists on non-root
938 cgroups. The default is "100".
940 The weight in the range [1, 10000].
943 A read-write single value file which exists on non-root
944 cgroups. The default is "0".
946 The nice value is in the range [-20, 19].
948 This interface file is an alternative interface for
949 "cpu.weight" and allows reading and setting weight using the
950 same values used by nice(2). Because the range is smaller and
951 granularity is coarser for the nice values, the read value is
952 the closest approximation of the current weight.
955 A read-write two value file which exists on non-root cgroups.
956 The default is "max 100000".
958 The maximum bandwidth limit. It's in the following format::
962 which indicates that the group may consume upto $MAX in each
963 $PERIOD duration. "max" for $MAX indicates no limit. If only
964 one number is written, $MAX is updated.
970 The "memory" controller regulates distribution of memory. Memory is
971 stateful and implements both limit and protection models. Due to the
972 intertwining between memory usage and reclaim pressure and the
973 stateful nature of memory, the distribution model is relatively
976 While not completely water-tight, all major memory usages by a given
977 cgroup are tracked so that the total memory consumption can be
978 accounted and controlled to a reasonable extent. Currently, the
979 following types of memory usages are tracked.
981 - Userland memory - page cache and anonymous memory.
983 - Kernel data structures such as dentries and inodes.
985 - TCP socket buffers.
987 The above list may expand in the future for better coverage.
990 Memory Interface Files
991 ~~~~~~~~~~~~~~~~~~~~~~
993 All memory amounts are in bytes. If a value which is not aligned to
994 PAGE_SIZE is written, the value may be rounded up to the closest
995 PAGE_SIZE multiple when read back.
998 A read-only single value file which exists on non-root
1001 The total amount of memory currently being used by the cgroup
1002 and its descendants.
1005 A read-write single value file which exists on non-root
1006 cgroups. The default is "0".
1008 Best-effort memory protection. If the memory usages of a
1009 cgroup and all its ancestors are below their low boundaries,
1010 the cgroup's memory won't be reclaimed unless memory can be
1011 reclaimed from unprotected cgroups.
1013 Putting more memory than generally available under this
1014 protection is discouraged.
1017 A read-write single value file which exists on non-root
1018 cgroups. The default is "max".
1020 Memory usage throttle limit. This is the main mechanism to
1021 control memory usage of a cgroup. If a cgroup's usage goes
1022 over the high boundary, the processes of the cgroup are
1023 throttled and put under heavy reclaim pressure.
1025 Going over the high limit never invokes the OOM killer and
1026 under extreme conditions the limit may be breached.
1029 A read-write single value file which exists on non-root
1030 cgroups. The default is "max".
1032 Memory usage hard limit. This is the final protection
1033 mechanism. If a cgroup's memory usage reaches this limit and
1034 can't be reduced, the OOM killer is invoked in the cgroup.
1035 Under certain circumstances, the usage may go over the limit
1038 This is the ultimate protection mechanism. As long as the
1039 high limit is used and monitored properly, this limit's
1040 utility is limited to providing the final safety net.
1043 A read-only flat-keyed file which exists on non-root cgroups.
1044 The following entries are defined. Unless specified
1045 otherwise, a value change in this file generates a file
1049 The number of times the cgroup is reclaimed due to
1050 high memory pressure even though its usage is under
1051 the low boundary. This usually indicates that the low
1052 boundary is over-committed.
1055 The number of times processes of the cgroup are
1056 throttled and routed to perform direct memory reclaim
1057 because the high memory boundary was exceeded. For a
1058 cgroup whose memory usage is capped by the high limit
1059 rather than global memory pressure, this event's
1060 occurrences are expected.
1063 The number of times the cgroup's memory usage was
1064 about to go over the max boundary. If direct reclaim
1065 fails to bring it down, the cgroup goes to OOM state.
1068 The number of time the cgroup's memory usage was
1069 reached the limit and allocation was about to fail.
1071 Depending on context result could be invocation of OOM
1072 killer and retrying allocation or failing allocation.
1074 Failed allocation in its turn could be returned into
1075 userspace as -ENOMEM or silently ignored in cases like
1076 disk readahead. For now OOM in memory cgroup kills
1077 tasks iff shortage has happened inside page fault.
1080 The number of processes belonging to this cgroup
1081 killed by any kind of OOM killer.
1084 A read-only flat-keyed file which exists on non-root cgroups.
1086 This breaks down the cgroup's memory footprint into different
1087 types of memory, type-specific details, and other information
1088 on the state and past events of the memory management system.
1090 All memory amounts are in bytes.
1092 The entries are ordered to be human readable, and new entries
1093 can show up in the middle. Don't rely on items remaining in a
1094 fixed position; use the keys to look up specific values!
1097 Amount of memory used in anonymous mappings such as
1098 brk(), sbrk(), and mmap(MAP_ANONYMOUS)
1101 Amount of memory used to cache filesystem data,
1102 including tmpfs and shared memory.
1105 Amount of memory allocated to kernel stacks.
1108 Amount of memory used for storing in-kernel data
1112 Amount of memory used in network transmission buffers
1115 Amount of cached filesystem data that is swap-backed,
1116 such as tmpfs, shm segments, shared anonymous mmap()s
1119 Amount of cached filesystem data mapped with mmap()
1122 Amount of cached filesystem data that was modified but
1123 not yet written back to disk
1126 Amount of cached filesystem data that was modified and
1127 is currently being written back to disk
1129 inactive_anon, active_anon, inactive_file, active_file, unevictable
1130 Amount of memory, swap-backed and filesystem-backed,
1131 on the internal memory management lists used by the
1132 page reclaim algorithm
1135 Part of "slab" that might be reclaimed, such as
1136 dentries and inodes.
1139 Part of "slab" that cannot be reclaimed on memory
1143 Total number of page faults incurred
1146 Number of major page faults incurred
1150 Number of refaults of previously evicted pages
1154 Number of refaulted pages that were immediately activated
1156 workingset_nodereclaim
1158 Number of times a shadow node has been reclaimed
1162 Amount of scanned pages (in an active LRU list)
1166 Amount of scanned pages (in an inactive LRU list)
1170 Amount of reclaimed pages
1174 Amount of pages moved to the active LRU list
1178 Amount of pages moved to the inactive LRU lis
1182 Amount of pages postponed to be freed under memory pressure
1186 Amount of reclaimed lazyfree pages
1189 A read-only single value file which exists on non-root
1192 The total amount of swap currently being used by the cgroup
1193 and its descendants.
1196 A read-write single value file which exists on non-root
1197 cgroups. The default is "max".
1199 Swap usage hard limit. If a cgroup's swap usage reaches this
1200 limit, anonymous memory of the cgroup will not be swapped out.
1203 A read-only flat-keyed file which exists on non-root cgroups.
1204 The following entries are defined. Unless specified
1205 otherwise, a value change in this file generates a file
1209 The number of times the cgroup's swap usage was about
1210 to go over the max boundary and swap allocation
1214 The number of times swap allocation failed either
1215 because of running out of swap system-wide or max
1222 "memory.high" is the main mechanism to control memory usage.
1223 Over-committing on high limit (sum of high limits > available memory)
1224 and letting global memory pressure to distribute memory according to
1225 usage is a viable strategy.
1227 Because breach of the high limit doesn't trigger the OOM killer but
1228 throttles the offending cgroup, a management agent has ample
1229 opportunities to monitor and take appropriate actions such as granting
1230 more memory or terminating the workload.
1232 Determining whether a cgroup has enough memory is not trivial as
1233 memory usage doesn't indicate whether the workload can benefit from
1234 more memory. For example, a workload which writes data received from
1235 network to a file can use all available memory but can also operate as
1236 performant with a small amount of memory. A measure of memory
1237 pressure - how much the workload is being impacted due to lack of
1238 memory - is necessary to determine whether a workload needs more
1239 memory; unfortunately, memory pressure monitoring mechanism isn't
1246 A memory area is charged to the cgroup which instantiated it and stays
1247 charged to the cgroup until the area is released. Migrating a process
1248 to a different cgroup doesn't move the memory usages that it
1249 instantiated while in the previous cgroup to the new cgroup.
1251 A memory area may be used by processes belonging to different cgroups.
1252 To which cgroup the area will be charged is in-deterministic; however,
1253 over time, the memory area is likely to end up in a cgroup which has
1254 enough memory allowance to avoid high reclaim pressure.
1256 If a cgroup sweeps a considerable amount of memory which is expected
1257 to be accessed repeatedly by other cgroups, it may make sense to use
1258 POSIX_FADV_DONTNEED to relinquish the ownership of memory areas
1259 belonging to the affected files to ensure correct memory ownership.
1265 The "io" controller regulates the distribution of IO resources. This
1266 controller implements both weight based and absolute bandwidth or IOPS
1267 limit distribution; however, weight based distribution is available
1268 only if cfq-iosched is in use and neither scheme is available for
1276 A read-only nested-keyed file which exists on non-root
1279 Lines are keyed by $MAJ:$MIN device numbers and not ordered.
1280 The following nested keys are defined.
1282 ====== ===================
1284 wbytes Bytes written
1285 rios Number of read IOs
1286 wios Number of write IOs
1287 ====== ===================
1289 An example read output follows:
1291 8:16 rbytes=1459200 wbytes=314773504 rios=192 wios=353
1292 8:0 rbytes=90430464 wbytes=299008000 rios=8950 wios=1252
1295 A read-write flat-keyed file which exists on non-root cgroups.
1296 The default is "default 100".
1298 The first line is the default weight applied to devices
1299 without specific override. The rest are overrides keyed by
1300 $MAJ:$MIN device numbers and not ordered. The weights are in
1301 the range [1, 10000] and specifies the relative amount IO time
1302 the cgroup can use in relation to its siblings.
1304 The default weight can be updated by writing either "default
1305 $WEIGHT" or simply "$WEIGHT". Overrides can be set by writing
1306 "$MAJ:$MIN $WEIGHT" and unset by writing "$MAJ:$MIN default".
1308 An example read output follows::
1315 A read-write nested-keyed file which exists on non-root
1318 BPS and IOPS based IO limit. Lines are keyed by $MAJ:$MIN
1319 device numbers and not ordered. The following nested keys are
1322 ===== ==================================
1323 rbps Max read bytes per second
1324 wbps Max write bytes per second
1325 riops Max read IO operations per second
1326 wiops Max write IO operations per second
1327 ===== ==================================
1329 When writing, any number of nested key-value pairs can be
1330 specified in any order. "max" can be specified as the value
1331 to remove a specific limit. If the same key is specified
1332 multiple times, the outcome is undefined.
1334 BPS and IOPS are measured in each IO direction and IOs are
1335 delayed if limit is reached. Temporary bursts are allowed.
1337 Setting read limit at 2M BPS and write at 120 IOPS for 8:16::
1339 echo "8:16 rbps=2097152 wiops=120" > io.max
1341 Reading returns the following::
1343 8:16 rbps=2097152 wbps=max riops=max wiops=120
1345 Write IOPS limit can be removed by writing the following::
1347 echo "8:16 wiops=max" > io.max
1349 Reading now returns the following::
1351 8:16 rbps=2097152 wbps=max riops=max wiops=max
1357 Page cache is dirtied through buffered writes and shared mmaps and
1358 written asynchronously to the backing filesystem by the writeback
1359 mechanism. Writeback sits between the memory and IO domains and
1360 regulates the proportion of dirty memory by balancing dirtying and
1363 The io controller, in conjunction with the memory controller,
1364 implements control of page cache writeback IOs. The memory controller
1365 defines the memory domain that dirty memory ratio is calculated and
1366 maintained for and the io controller defines the io domain which
1367 writes out dirty pages for the memory domain. Both system-wide and
1368 per-cgroup dirty memory states are examined and the more restrictive
1369 of the two is enforced.
1371 cgroup writeback requires explicit support from the underlying
1372 filesystem. Currently, cgroup writeback is implemented on ext2, ext4
1373 and btrfs. On other filesystems, all writeback IOs are attributed to
1376 There are inherent differences in memory and writeback management
1377 which affects how cgroup ownership is tracked. Memory is tracked per
1378 page while writeback per inode. For the purpose of writeback, an
1379 inode is assigned to a cgroup and all IO requests to write dirty pages
1380 from the inode are attributed to that cgroup.
1382 As cgroup ownership for memory is tracked per page, there can be pages
1383 which are associated with different cgroups than the one the inode is
1384 associated with. These are called foreign pages. The writeback
1385 constantly keeps track of foreign pages and, if a particular foreign
1386 cgroup becomes the majority over a certain period of time, switches
1387 the ownership of the inode to that cgroup.
1389 While this model is enough for most use cases where a given inode is
1390 mostly dirtied by a single cgroup even when the main writing cgroup
1391 changes over time, use cases where multiple cgroups write to a single
1392 inode simultaneously are not supported well. In such circumstances, a
1393 significant portion of IOs are likely to be attributed incorrectly.
1394 As memory controller assigns page ownership on the first use and
1395 doesn't update it until the page is released, even if writeback
1396 strictly follows page ownership, multiple cgroups dirtying overlapping
1397 areas wouldn't work as expected. It's recommended to avoid such usage
1400 The sysctl knobs which affect writeback behavior are applied to cgroup
1401 writeback as follows.
1403 vm.dirty_background_ratio, vm.dirty_ratio
1404 These ratios apply the same to cgroup writeback with the
1405 amount of available memory capped by limits imposed by the
1406 memory controller and system-wide clean memory.
1408 vm.dirty_background_bytes, vm.dirty_bytes
1409 For cgroup writeback, this is calculated into ratio against
1410 total available memory and applied the same way as
1411 vm.dirty[_background]_ratio.
1417 The process number controller is used to allow a cgroup to stop any
1418 new tasks from being fork()'d or clone()'d after a specified limit is
1421 The number of tasks in a cgroup can be exhausted in ways which other
1422 controllers cannot prevent, thus warranting its own controller. For
1423 example, a fork bomb is likely to exhaust the number of tasks before
1424 hitting memory restrictions.
1426 Note that PIDs used in this controller refer to TIDs, process IDs as
1434 A read-write single value file which exists on non-root
1435 cgroups. The default is "max".
1437 Hard limit of number of processes.
1440 A read-only single value file which exists on all cgroups.
1442 The number of processes currently in the cgroup and its
1445 Organisational operations are not blocked by cgroup policies, so it is
1446 possible to have pids.current > pids.max. This can be done by either
1447 setting the limit to be smaller than pids.current, or attaching enough
1448 processes to the cgroup such that pids.current is larger than
1449 pids.max. However, it is not possible to violate a cgroup PID policy
1450 through fork() or clone(). These will return -EAGAIN if the creation
1451 of a new process would cause a cgroup policy to be violated.
1457 Device controller manages access to device files. It includes both
1458 creation of new device files (using mknod), and access to the
1459 existing device files.
1461 Cgroup v2 device controller has no interface files and is implemented
1462 on top of cgroup BPF. To control access to device files, a user may
1463 create bpf programs of the BPF_CGROUP_DEVICE type and attach them
1464 to cgroups. On an attempt to access a device file, corresponding
1465 BPF programs will be executed, and depending on the return value
1466 the attempt will succeed or fail with -EPERM.
1468 A BPF_CGROUP_DEVICE program takes a pointer to the bpf_cgroup_dev_ctx
1469 structure, which describes the device access attempt: access type
1470 (mknod/read/write) and device (type, major and minor numbers).
1471 If the program returns 0, the attempt fails with -EPERM, otherwise
1474 An example of BPF_CGROUP_DEVICE program may be found in the kernel
1475 source tree in the tools/testing/selftests/bpf/dev_cgroup.c file.
1481 The "rdma" controller regulates the distribution and accounting of
1484 RDMA Interface Files
1485 ~~~~~~~~~~~~~~~~~~~~
1488 A readwrite nested-keyed file that exists for all the cgroups
1489 except root that describes current configured resource limit
1490 for a RDMA/IB device.
1492 Lines are keyed by device name and are not ordered.
1493 Each line contains space separated resource name and its configured
1494 limit that can be distributed.
1496 The following nested keys are defined.
1498 ========== =============================
1499 hca_handle Maximum number of HCA Handles
1500 hca_object Maximum number of HCA Objects
1501 ========== =============================
1503 An example for mlx4 and ocrdma device follows::
1505 mlx4_0 hca_handle=2 hca_object=2000
1506 ocrdma1 hca_handle=3 hca_object=max
1509 A read-only file that describes current resource usage.
1510 It exists for all the cgroup except root.
1512 An example for mlx4 and ocrdma device follows::
1514 mlx4_0 hca_handle=1 hca_object=20
1515 ocrdma1 hca_handle=1 hca_object=23
1524 perf_event controller, if not mounted on a legacy hierarchy, is
1525 automatically enabled on the v2 hierarchy so that perf events can
1526 always be filtered by cgroup v2 path. The controller can still be
1527 moved to a legacy hierarchy after v2 hierarchy is populated.
1530 Non-normative information
1531 -------------------------
1533 This section contains information that isn't considered to be a part of
1534 the stable kernel API and so is subject to change.
1537 CPU controller root cgroup process behaviour
1538 ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
1540 When distributing CPU cycles in the root cgroup each thread in this
1541 cgroup is treated as if it was hosted in a separate child cgroup of the
1542 root cgroup. This child cgroup weight is dependent on its thread nice
1545 For details of this mapping see sched_prio_to_weight array in
1546 kernel/sched/core.c file (values from this array should be scaled
1547 appropriately so the neutral - nice 0 - value is 100 instead of 1024).
1550 IO controller root cgroup process behaviour
1551 ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
1553 Root cgroup processes are hosted in an implicit leaf child node.
1554 When distributing IO resources this implicit child node is taken into
1555 account as if it was a normal child cgroup of the root cgroup with a
1556 weight value of 200.
1565 cgroup namespace provides a mechanism to virtualize the view of the
1566 "/proc/$PID/cgroup" file and cgroup mounts. The CLONE_NEWCGROUP clone
1567 flag can be used with clone(2) and unshare(2) to create a new cgroup
1568 namespace. The process running inside the cgroup namespace will have
1569 its "/proc/$PID/cgroup" output restricted to cgroupns root. The
1570 cgroupns root is the cgroup of the process at the time of creation of
1571 the cgroup namespace.
1573 Without cgroup namespace, the "/proc/$PID/cgroup" file shows the
1574 complete path of the cgroup of a process. In a container setup where
1575 a set of cgroups and namespaces are intended to isolate processes the
1576 "/proc/$PID/cgroup" file may leak potential system level information
1577 to the isolated processes. For Example::
1579 # cat /proc/self/cgroup
1580 0::/batchjobs/container_id1
1582 The path '/batchjobs/container_id1' can be considered as system-data
1583 and undesirable to expose to the isolated processes. cgroup namespace
1584 can be used to restrict visibility of this path. For example, before
1585 creating a cgroup namespace, one would see::
1587 # ls -l /proc/self/ns/cgroup
1588 lrwxrwxrwx 1 root root 0 2014-07-15 10:37 /proc/self/ns/cgroup -> cgroup:[4026531835]
1589 # cat /proc/self/cgroup
1590 0::/batchjobs/container_id1
1592 After unsharing a new namespace, the view changes::
1594 # ls -l /proc/self/ns/cgroup
1595 lrwxrwxrwx 1 root root 0 2014-07-15 10:35 /proc/self/ns/cgroup -> cgroup:[4026532183]
1596 # cat /proc/self/cgroup
1599 When some thread from a multi-threaded process unshares its cgroup
1600 namespace, the new cgroupns gets applied to the entire process (all
1601 the threads). This is natural for the v2 hierarchy; however, for the
1602 legacy hierarchies, this may be unexpected.
1604 A cgroup namespace is alive as long as there are processes inside or
1605 mounts pinning it. When the last usage goes away, the cgroup
1606 namespace is destroyed. The cgroupns root and the actual cgroups
1613 The 'cgroupns root' for a cgroup namespace is the cgroup in which the
1614 process calling unshare(2) is running. For example, if a process in
1615 /batchjobs/container_id1 cgroup calls unshare, cgroup
1616 /batchjobs/container_id1 becomes the cgroupns root. For the
1617 init_cgroup_ns, this is the real root ('/') cgroup.
1619 The cgroupns root cgroup does not change even if the namespace creator
1620 process later moves to a different cgroup::
1622 # ~/unshare -c # unshare cgroupns in some cgroup
1623 # cat /proc/self/cgroup
1626 # echo 0 > sub_cgrp_1/cgroup.procs
1627 # cat /proc/self/cgroup
1630 Each process gets its namespace-specific view of "/proc/$PID/cgroup"
1632 Processes running inside the cgroup namespace will be able to see
1633 cgroup paths (in /proc/self/cgroup) only inside their root cgroup.
1634 From within an unshared cgroupns::
1638 # echo 7353 > sub_cgrp_1/cgroup.procs
1639 # cat /proc/7353/cgroup
1642 From the initial cgroup namespace, the real cgroup path will be
1645 $ cat /proc/7353/cgroup
1646 0::/batchjobs/container_id1/sub_cgrp_1
1648 From a sibling cgroup namespace (that is, a namespace rooted at a
1649 different cgroup), the cgroup path relative to its own cgroup
1650 namespace root will be shown. For instance, if PID 7353's cgroup
1651 namespace root is at '/batchjobs/container_id2', then it will see::
1653 # cat /proc/7353/cgroup
1654 0::/../container_id2/sub_cgrp_1
1656 Note that the relative path always starts with '/' to indicate that
1657 its relative to the cgroup namespace root of the caller.
1660 Migration and setns(2)
1661 ----------------------
1663 Processes inside a cgroup namespace can move into and out of the
1664 namespace root if they have proper access to external cgroups. For
1665 example, from inside a namespace with cgroupns root at
1666 /batchjobs/container_id1, and assuming that the global hierarchy is
1667 still accessible inside cgroupns::
1669 # cat /proc/7353/cgroup
1671 # echo 7353 > batchjobs/container_id2/cgroup.procs
1672 # cat /proc/7353/cgroup
1673 0::/../container_id2
1675 Note that this kind of setup is not encouraged. A task inside cgroup
1676 namespace should only be exposed to its own cgroupns hierarchy.
1678 setns(2) to another cgroup namespace is allowed when:
1680 (a) the process has CAP_SYS_ADMIN against its current user namespace
1681 (b) the process has CAP_SYS_ADMIN against the target cgroup
1684 No implicit cgroup changes happen with attaching to another cgroup
1685 namespace. It is expected that the someone moves the attaching
1686 process under the target cgroup namespace root.
1689 Interaction with Other Namespaces
1690 ---------------------------------
1692 Namespace specific cgroup hierarchy can be mounted by a process
1693 running inside a non-init cgroup namespace::
1695 # mount -t cgroup2 none $MOUNT_POINT
1697 This will mount the unified cgroup hierarchy with cgroupns root as the
1698 filesystem root. The process needs CAP_SYS_ADMIN against its user and
1701 The virtualization of /proc/self/cgroup file combined with restricting
1702 the view of cgroup hierarchy by namespace-private cgroupfs mount
1703 provides a properly isolated cgroup view inside the container.
1706 Information on Kernel Programming
1707 =================================
1709 This section contains kernel programming information in the areas
1710 where interacting with cgroup is necessary. cgroup core and
1711 controllers are not covered.
1714 Filesystem Support for Writeback
1715 --------------------------------
1717 A filesystem can support cgroup writeback by updating
1718 address_space_operations->writepage[s]() to annotate bio's using the
1719 following two functions.
1721 wbc_init_bio(@wbc, @bio)
1722 Should be called for each bio carrying writeback data and
1723 associates the bio with the inode's owner cgroup. Can be
1724 called anytime between bio allocation and submission.
1726 wbc_account_io(@wbc, @page, @bytes)
1727 Should be called for each data segment being written out.
1728 While this function doesn't care exactly when it's called
1729 during the writeback session, it's the easiest and most
1730 natural to call it as data segments are added to a bio.
1732 With writeback bio's annotated, cgroup support can be enabled per
1733 super_block by setting SB_I_CGROUPWB in ->s_iflags. This allows for
1734 selective disabling of cgroup writeback support which is helpful when
1735 certain filesystem features, e.g. journaled data mode, are
1738 wbc_init_bio() binds the specified bio to its cgroup. Depending on
1739 the configuration, the bio may be executed at a lower priority and if
1740 the writeback session is holding shared resources, e.g. a journal
1741 entry, may lead to priority inversion. There is no one easy solution
1742 for the problem. Filesystems can try to work around specific problem
1743 cases by skipping wbc_init_bio() or using bio_associate_blkcg()
1747 Deprecated v1 Core Features
1748 ===========================
1750 - Multiple hierarchies including named ones are not supported.
1752 - All v1 mount options are not supported.
1754 - The "tasks" file is removed and "cgroup.procs" is not sorted.
1756 - "cgroup.clone_children" is removed.
1758 - /proc/cgroups is meaningless for v2. Use "cgroup.controllers" file
1759 at the root instead.
1762 Issues with v1 and Rationales for v2
1763 ====================================
1765 Multiple Hierarchies
1766 --------------------
1768 cgroup v1 allowed an arbitrary number of hierarchies and each
1769 hierarchy could host any number of controllers. While this seemed to
1770 provide a high level of flexibility, it wasn't useful in practice.
1772 For example, as there is only one instance of each controller, utility
1773 type controllers such as freezer which can be useful in all
1774 hierarchies could only be used in one. The issue is exacerbated by
1775 the fact that controllers couldn't be moved to another hierarchy once
1776 hierarchies were populated. Another issue was that all controllers
1777 bound to a hierarchy were forced to have exactly the same view of the
1778 hierarchy. It wasn't possible to vary the granularity depending on
1779 the specific controller.
1781 In practice, these issues heavily limited which controllers could be
1782 put on the same hierarchy and most configurations resorted to putting
1783 each controller on its own hierarchy. Only closely related ones, such
1784 as the cpu and cpuacct controllers, made sense to be put on the same
1785 hierarchy. This often meant that userland ended up managing multiple
1786 similar hierarchies repeating the same steps on each hierarchy
1787 whenever a hierarchy management operation was necessary.
1789 Furthermore, support for multiple hierarchies came at a steep cost.
1790 It greatly complicated cgroup core implementation but more importantly
1791 the support for multiple hierarchies restricted how cgroup could be
1792 used in general and what controllers was able to do.
1794 There was no limit on how many hierarchies there might be, which meant
1795 that a thread's cgroup membership couldn't be described in finite
1796 length. The key might contain any number of entries and was unlimited
1797 in length, which made it highly awkward to manipulate and led to
1798 addition of controllers which existed only to identify membership,
1799 which in turn exacerbated the original problem of proliferating number
1802 Also, as a controller couldn't have any expectation regarding the
1803 topologies of hierarchies other controllers might be on, each
1804 controller had to assume that all other controllers were attached to
1805 completely orthogonal hierarchies. This made it impossible, or at
1806 least very cumbersome, for controllers to cooperate with each other.
1808 In most use cases, putting controllers on hierarchies which are
1809 completely orthogonal to each other isn't necessary. What usually is
1810 called for is the ability to have differing levels of granularity
1811 depending on the specific controller. In other words, hierarchy may
1812 be collapsed from leaf towards root when viewed from specific
1813 controllers. For example, a given configuration might not care about
1814 how memory is distributed beyond a certain level while still wanting
1815 to control how CPU cycles are distributed.
1821 cgroup v1 allowed threads of a process to belong to different cgroups.
1822 This didn't make sense for some controllers and those controllers
1823 ended up implementing different ways to ignore such situations but
1824 much more importantly it blurred the line between API exposed to
1825 individual applications and system management interface.
1827 Generally, in-process knowledge is available only to the process
1828 itself; thus, unlike service-level organization of processes,
1829 categorizing threads of a process requires active participation from
1830 the application which owns the target process.
1832 cgroup v1 had an ambiguously defined delegation model which got abused
1833 in combination with thread granularity. cgroups were delegated to
1834 individual applications so that they can create and manage their own
1835 sub-hierarchies and control resource distributions along them. This
1836 effectively raised cgroup to the status of a syscall-like API exposed
1839 First of all, cgroup has a fundamentally inadequate interface to be
1840 exposed this way. For a process to access its own knobs, it has to
1841 extract the path on the target hierarchy from /proc/self/cgroup,
1842 construct the path by appending the name of the knob to the path, open
1843 and then read and/or write to it. This is not only extremely clunky
1844 and unusual but also inherently racy. There is no conventional way to
1845 define transaction across the required steps and nothing can guarantee
1846 that the process would actually be operating on its own sub-hierarchy.
1848 cgroup controllers implemented a number of knobs which would never be
1849 accepted as public APIs because they were just adding control knobs to
1850 system-management pseudo filesystem. cgroup ended up with interface
1851 knobs which were not properly abstracted or refined and directly
1852 revealed kernel internal details. These knobs got exposed to
1853 individual applications through the ill-defined delegation mechanism
1854 effectively abusing cgroup as a shortcut to implementing public APIs
1855 without going through the required scrutiny.
1857 This was painful for both userland and kernel. Userland ended up with
1858 misbehaving and poorly abstracted interfaces and kernel exposing and
1859 locked into constructs inadvertently.
1862 Competition Between Inner Nodes and Threads
1863 -------------------------------------------
1865 cgroup v1 allowed threads to be in any cgroups which created an
1866 interesting problem where threads belonging to a parent cgroup and its
1867 children cgroups competed for resources. This was nasty as two
1868 different types of entities competed and there was no obvious way to
1869 settle it. Different controllers did different things.
1871 The cpu controller considered threads and cgroups as equivalents and
1872 mapped nice levels to cgroup weights. This worked for some cases but
1873 fell flat when children wanted to be allocated specific ratios of CPU
1874 cycles and the number of internal threads fluctuated - the ratios
1875 constantly changed as the number of competing entities fluctuated.
1876 There also were other issues. The mapping from nice level to weight
1877 wasn't obvious or universal, and there were various other knobs which
1878 simply weren't available for threads.
1880 The io controller implicitly created a hidden leaf node for each
1881 cgroup to host the threads. The hidden leaf had its own copies of all
1882 the knobs with ``leaf_`` prefixed. While this allowed equivalent
1883 control over internal threads, it was with serious drawbacks. It
1884 always added an extra layer of nesting which wouldn't be necessary
1885 otherwise, made the interface messy and significantly complicated the
1888 The memory controller didn't have a way to control what happened
1889 between internal tasks and child cgroups and the behavior was not
1890 clearly defined. There were attempts to add ad-hoc behaviors and
1891 knobs to tailor the behavior to specific workloads which would have
1892 led to problems extremely difficult to resolve in the long term.
1894 Multiple controllers struggled with internal tasks and came up with
1895 different ways to deal with it; unfortunately, all the approaches were
1896 severely flawed and, furthermore, the widely different behaviors
1897 made cgroup as a whole highly inconsistent.
1899 This clearly is a problem which needs to be addressed from cgroup core
1903 Other Interface Issues
1904 ----------------------
1906 cgroup v1 grew without oversight and developed a large number of
1907 idiosyncrasies and inconsistencies. One issue on the cgroup core side
1908 was how an empty cgroup was notified - a userland helper binary was
1909 forked and executed for each event. The event delivery wasn't
1910 recursive or delegatable. The limitations of the mechanism also led
1911 to in-kernel event delivery filtering mechanism further complicating
1914 Controller interfaces were problematic too. An extreme example is
1915 controllers completely ignoring hierarchical organization and treating
1916 all cgroups as if they were all located directly under the root
1917 cgroup. Some controllers exposed a large amount of inconsistent
1918 implementation details to userland.
1920 There also was no consistency across controllers. When a new cgroup
1921 was created, some controllers defaulted to not imposing extra
1922 restrictions while others disallowed any resource usage until
1923 explicitly configured. Configuration knobs for the same type of
1924 control used widely differing naming schemes and formats. Statistics
1925 and information knobs were named arbitrarily and used different
1926 formats and units even in the same controller.
1928 cgroup v2 establishes common conventions where appropriate and updates
1929 controllers so that they expose minimal and consistent interfaces.
1932 Controller Issues and Remedies
1933 ------------------------------
1938 The original lower boundary, the soft limit, is defined as a limit
1939 that is per default unset. As a result, the set of cgroups that
1940 global reclaim prefers is opt-in, rather than opt-out. The costs for
1941 optimizing these mostly negative lookups are so high that the
1942 implementation, despite its enormous size, does not even provide the
1943 basic desirable behavior. First off, the soft limit has no
1944 hierarchical meaning. All configured groups are organized in a global
1945 rbtree and treated like equal peers, regardless where they are located
1946 in the hierarchy. This makes subtree delegation impossible. Second,
1947 the soft limit reclaim pass is so aggressive that it not just
1948 introduces high allocation latencies into the system, but also impacts
1949 system performance due to overreclaim, to the point where the feature
1950 becomes self-defeating.
1952 The memory.low boundary on the other hand is a top-down allocated
1953 reserve. A cgroup enjoys reclaim protection when it and all its
1954 ancestors are below their low boundaries, which makes delegation of
1955 subtrees possible. Secondly, new cgroups have no reserve per default
1956 and in the common case most cgroups are eligible for the preferred
1957 reclaim pass. This allows the new low boundary to be efficiently
1958 implemented with just a minor addition to the generic reclaim code,
1959 without the need for out-of-band data structures and reclaim passes.
1960 Because the generic reclaim code considers all cgroups except for the
1961 ones running low in the preferred first reclaim pass, overreclaim of
1962 individual groups is eliminated as well, resulting in much better
1963 overall workload performance.
1965 The original high boundary, the hard limit, is defined as a strict
1966 limit that can not budge, even if the OOM killer has to be called.
1967 But this generally goes against the goal of making the most out of the
1968 available memory. The memory consumption of workloads varies during
1969 runtime, and that requires users to overcommit. But doing that with a
1970 strict upper limit requires either a fairly accurate prediction of the
1971 working set size or adding slack to the limit. Since working set size
1972 estimation is hard and error prone, and getting it wrong results in
1973 OOM kills, most users tend to err on the side of a looser limit and
1974 end up wasting precious resources.
1976 The memory.high boundary on the other hand can be set much more
1977 conservatively. When hit, it throttles allocations by forcing them
1978 into direct reclaim to work off the excess, but it never invokes the
1979 OOM killer. As a result, a high boundary that is chosen too
1980 aggressively will not terminate the processes, but instead it will
1981 lead to gradual performance degradation. The user can monitor this
1982 and make corrections until the minimal memory footprint that still
1983 gives acceptable performance is found.
1985 In extreme cases, with many concurrent allocations and a complete
1986 breakdown of reclaim progress within the group, the high boundary can
1987 be exceeded. But even then it's mostly better to satisfy the
1988 allocation from the slack available in other groups or the rest of the
1989 system than killing the group. Otherwise, memory.max is there to
1990 limit this type of spillover and ultimately contain buggy or even
1991 malicious applications.
1993 Setting the original memory.limit_in_bytes below the current usage was
1994 subject to a race condition, where concurrent charges could cause the
1995 limit setting to fail. memory.max on the other hand will first set the
1996 limit to prevent new charges, and then reclaim and OOM kill until the
1997 new limit is met - or the task writing to memory.max is killed.
1999 The combined memory+swap accounting and limiting is replaced by real
2000 control over swap space.
2002 The main argument for a combined memory+swap facility in the original
2003 cgroup design was that global or parental pressure would always be
2004 able to swap all anonymous memory of a child group, regardless of the
2005 child's own (possibly untrusted) configuration. However, untrusted
2006 groups can sabotage swapping by other means - such as referencing its
2007 anonymous memory in a tight loop - and an admin can not assume full
2008 swappability when overcommitting untrusted jobs.
2010 For trusted jobs, on the other hand, a combined counter is not an
2011 intuitive userspace interface, and it flies in the face of the idea
2012 that cgroup controllers should account and limit specific physical
2013 resources. Swap space is a resource like all others in the system,
2014 and that's why unified hierarchy allows distributing it separately.