8 :Author: Tejun Heo <tj@kernel.org>
10 This is the authoritative documentation on the design, interface and
11 conventions of cgroup v2. It describes all userland-visible aspects
12 of cgroup including core and specific controller behaviors. All
13 future changes must be reflected in this document. Documentation for
14 v1 is available under :ref:`Documentation/admin-guide/cgroup-v1/index.rst <cgroup-v1>`.
23 2-2. Organizing Processes and Threads
26 2-3. [Un]populated Notification
27 2-4. Controlling Controllers
28 2-4-1. Enabling and Disabling
29 2-4-2. Top-down Constraint
30 2-4-3. No Internal Process Constraint
32 2-5-1. Model of Delegation
33 2-5-2. Delegation Containment
35 2-6-1. Organize Once and Control
36 2-6-2. Avoid Name Collisions
37 3. Resource Distribution Models
45 4-3. Core Interface Files
48 5-1-1. CPU Interface Files
50 5-2-1. Memory Interface Files
51 5-2-2. Usage Guidelines
52 5-2-3. Memory Ownership
54 5-3-1. IO Interface Files
57 5-3-3-1. How IO Latency Throttling Works
58 5-3-3-2. IO Latency Interface Files
60 5-4-1. PID Interface Files
62 5.5-1. Cpuset Interface Files
65 5-7-1. RDMA Interface Files
67 5.8-1. HugeTLB Interface Files
69 5.9-1 Miscellaneous cgroup Interface Files
70 5.9-2 Migration and Ownership
73 5-N. Non-normative information
74 5-N-1. CPU controller root cgroup process behaviour
75 5-N-2. IO controller root cgroup process behaviour
78 6-2. The Root and Views
79 6-3. Migration and setns(2)
80 6-4. Interaction with Other Namespaces
81 P. Information on Kernel Programming
82 P-1. Filesystem Support for Writeback
83 D. Deprecated v1 Core Features
84 R. Issues with v1 and Rationales for v2
85 R-1. Multiple Hierarchies
86 R-2. Thread Granularity
87 R-3. Competition Between Inner Nodes and Threads
88 R-4. Other Interface Issues
89 R-5. Controller Issues and Remedies
99 "cgroup" stands for "control group" and is never capitalized. The
100 singular form is used to designate the whole feature and also as a
101 qualifier as in "cgroup controllers". When explicitly referring to
102 multiple individual control groups, the plural form "cgroups" is used.
108 cgroup is a mechanism to organize processes hierarchically and
109 distribute system resources along the hierarchy in a controlled and
112 cgroup is largely composed of two parts - the core and controllers.
113 cgroup core is primarily responsible for hierarchically organizing
114 processes. A cgroup controller is usually responsible for
115 distributing a specific type of system resource along the hierarchy
116 although there are utility controllers which serve purposes other than
117 resource distribution.
119 cgroups form a tree structure and every process in the system belongs
120 to one and only one cgroup. All threads of a process belong to the
121 same cgroup. On creation, all processes are put in the cgroup that
122 the parent process belongs to at the time. A process can be migrated
123 to another cgroup. Migration of a process doesn't affect already
124 existing descendant processes.
126 Following certain structural constraints, controllers may be enabled or
127 disabled selectively on a cgroup. All controller behaviors are
128 hierarchical - if a controller is enabled on a cgroup, it affects all
129 processes which belong to the cgroups consisting the inclusive
130 sub-hierarchy of the cgroup. When a controller is enabled on a nested
131 cgroup, it always restricts the resource distribution further. The
132 restrictions set closer to the root in the hierarchy can not be
133 overridden from further away.
142 Unlike v1, cgroup v2 has only single hierarchy. The cgroup v2
143 hierarchy can be mounted with the following mount command::
145 # mount -t cgroup2 none $MOUNT_POINT
147 cgroup2 filesystem has the magic number 0x63677270 ("cgrp"). All
148 controllers which support v2 and are not bound to a v1 hierarchy are
149 automatically bound to the v2 hierarchy and show up at the root.
150 Controllers which are not in active use in the v2 hierarchy can be
151 bound to other hierarchies. This allows mixing v2 hierarchy with the
152 legacy v1 multiple hierarchies in a fully backward compatible way.
154 A controller can be moved across hierarchies only after the controller
155 is no longer referenced in its current hierarchy. Because per-cgroup
156 controller states are destroyed asynchronously and controllers may
157 have lingering references, a controller may not show up immediately on
158 the v2 hierarchy after the final umount of the previous hierarchy.
159 Similarly, a controller should be fully disabled to be moved out of
160 the unified hierarchy and it may take some time for the disabled
161 controller to become available for other hierarchies; furthermore, due
162 to inter-controller dependencies, other controllers may need to be
165 While useful for development and manual configurations, moving
166 controllers dynamically between the v2 and other hierarchies is
167 strongly discouraged for production use. It is recommended to decide
168 the hierarchies and controller associations before starting using the
169 controllers after system boot.
171 During transition to v2, system management software might still
172 automount the v1 cgroup filesystem and so hijack all controllers
173 during boot, before manual intervention is possible. To make testing
174 and experimenting easier, the kernel parameter cgroup_no_v1= allows
175 disabling controllers in v1 and make them always available in v2.
177 cgroup v2 currently supports the following mount options.
180 Consider cgroup namespaces as delegation boundaries. This
181 option is system wide and can only be set on mount or modified
182 through remount from the init namespace. The mount option is
183 ignored on non-init namespace mounts. Please refer to the
184 Delegation section for details.
187 Only populate memory.events with data for the current cgroup,
188 and not any subtrees. This is legacy behaviour, the default
189 behaviour without this option is to include subtree counts.
190 This option is system wide and can only be set on mount or
191 modified through remount from the init namespace. The mount
192 option is ignored on non-init namespace mounts.
195 Recursively apply memory.min and memory.low protection to
196 entire subtrees, without requiring explicit downward
197 propagation into leaf cgroups. This allows protecting entire
198 subtrees from one another, while retaining free competition
199 within those subtrees. This should have been the default
200 behavior but is a mount-option to avoid regressing setups
201 relying on the original semantics (e.g. specifying bogusly
202 high 'bypass' protection values at higher tree levels).
205 Organizing Processes and Threads
206 --------------------------------
211 Initially, only the root cgroup exists to which all processes belong.
212 A child cgroup can be created by creating a sub-directory::
216 A given cgroup may have multiple child cgroups forming a tree
217 structure. Each cgroup has a read-writable interface file
218 "cgroup.procs". When read, it lists the PIDs of all processes which
219 belong to the cgroup one-per-line. The PIDs are not ordered and the
220 same PID may show up more than once if the process got moved to
221 another cgroup and then back or the PID got recycled while reading.
223 A process can be migrated into a cgroup by writing its PID to the
224 target cgroup's "cgroup.procs" file. Only one process can be migrated
225 on a single write(2) call. If a process is composed of multiple
226 threads, writing the PID of any thread migrates all threads of the
229 When a process forks a child process, the new process is born into the
230 cgroup that the forking process belongs to at the time of the
231 operation. After exit, a process stays associated with the cgroup
232 that it belonged to at the time of exit until it's reaped; however, a
233 zombie process does not appear in "cgroup.procs" and thus can't be
234 moved to another cgroup.
236 A cgroup which doesn't have any children or live processes can be
237 destroyed by removing the directory. Note that a cgroup which doesn't
238 have any children and is associated only with zombie processes is
239 considered empty and can be removed::
243 "/proc/$PID/cgroup" lists a process's cgroup membership. If legacy
244 cgroup is in use in the system, this file may contain multiple lines,
245 one for each hierarchy. The entry for cgroup v2 is always in the
248 # cat /proc/842/cgroup
250 0::/test-cgroup/test-cgroup-nested
252 If the process becomes a zombie and the cgroup it was associated with
253 is removed subsequently, " (deleted)" is appended to the path::
255 # cat /proc/842/cgroup
257 0::/test-cgroup/test-cgroup-nested (deleted)
263 cgroup v2 supports thread granularity for a subset of controllers to
264 support use cases requiring hierarchical resource distribution across
265 the threads of a group of processes. By default, all threads of a
266 process belong to the same cgroup, which also serves as the resource
267 domain to host resource consumptions which are not specific to a
268 process or thread. The thread mode allows threads to be spread across
269 a subtree while still maintaining the common resource domain for them.
271 Controllers which support thread mode are called threaded controllers.
272 The ones which don't are called domain controllers.
274 Marking a cgroup threaded makes it join the resource domain of its
275 parent as a threaded cgroup. The parent may be another threaded
276 cgroup whose resource domain is further up in the hierarchy. The root
277 of a threaded subtree, that is, the nearest ancestor which is not
278 threaded, is called threaded domain or thread root interchangeably and
279 serves as the resource domain for the entire subtree.
281 Inside a threaded subtree, threads of a process can be put in
282 different cgroups and are not subject to the no internal process
283 constraint - threaded controllers can be enabled on non-leaf cgroups
284 whether they have threads in them or not.
286 As the threaded domain cgroup hosts all the domain resource
287 consumptions of the subtree, it is considered to have internal
288 resource consumptions whether there are processes in it or not and
289 can't have populated child cgroups which aren't threaded. Because the
290 root cgroup is not subject to no internal process constraint, it can
291 serve both as a threaded domain and a parent to domain cgroups.
293 The current operation mode or type of the cgroup is shown in the
294 "cgroup.type" file which indicates whether the cgroup is a normal
295 domain, a domain which is serving as the domain of a threaded subtree,
296 or a threaded cgroup.
298 On creation, a cgroup is always a domain cgroup and can be made
299 threaded by writing "threaded" to the "cgroup.type" file. The
300 operation is single direction::
302 # echo threaded > cgroup.type
304 Once threaded, the cgroup can't be made a domain again. To enable the
305 thread mode, the following conditions must be met.
307 - As the cgroup will join the parent's resource domain. The parent
308 must either be a valid (threaded) domain or a threaded cgroup.
310 - When the parent is an unthreaded domain, it must not have any domain
311 controllers enabled or populated domain children. The root is
312 exempt from this requirement.
314 Topology-wise, a cgroup can be in an invalid state. Please consider
315 the following topology::
317 A (threaded domain) - B (threaded) - C (domain, just created)
319 C is created as a domain but isn't connected to a parent which can
320 host child domains. C can't be used until it is turned into a
321 threaded cgroup. "cgroup.type" file will report "domain (invalid)" in
322 these cases. Operations which fail due to invalid topology use
323 EOPNOTSUPP as the errno.
325 A domain cgroup is turned into a threaded domain when one of its child
326 cgroup becomes threaded or threaded controllers are enabled in the
327 "cgroup.subtree_control" file while there are processes in the cgroup.
328 A threaded domain reverts to a normal domain when the conditions
331 When read, "cgroup.threads" contains the list of the thread IDs of all
332 threads in the cgroup. Except that the operations are per-thread
333 instead of per-process, "cgroup.threads" has the same format and
334 behaves the same way as "cgroup.procs". While "cgroup.threads" can be
335 written to in any cgroup, as it can only move threads inside the same
336 threaded domain, its operations are confined inside each threaded
339 The threaded domain cgroup serves as the resource domain for the whole
340 subtree, and, while the threads can be scattered across the subtree,
341 all the processes are considered to be in the threaded domain cgroup.
342 "cgroup.procs" in a threaded domain cgroup contains the PIDs of all
343 processes in the subtree and is not readable in the subtree proper.
344 However, "cgroup.procs" can be written to from anywhere in the subtree
345 to migrate all threads of the matching process to the cgroup.
347 Only threaded controllers can be enabled in a threaded subtree. When
348 a threaded controller is enabled inside a threaded subtree, it only
349 accounts for and controls resource consumptions associated with the
350 threads in the cgroup and its descendants. All consumptions which
351 aren't tied to a specific thread belong to the threaded domain cgroup.
353 Because a threaded subtree is exempt from no internal process
354 constraint, a threaded controller must be able to handle competition
355 between threads in a non-leaf cgroup and its child cgroups. Each
356 threaded controller defines how such competitions are handled.
359 [Un]populated Notification
360 --------------------------
362 Each non-root cgroup has a "cgroup.events" file which contains
363 "populated" field indicating whether the cgroup's sub-hierarchy has
364 live processes in it. Its value is 0 if there is no live process in
365 the cgroup and its descendants; otherwise, 1. poll and [id]notify
366 events are triggered when the value changes. This can be used, for
367 example, to start a clean-up operation after all processes of a given
368 sub-hierarchy have exited. The populated state updates and
369 notifications are recursive. Consider the following sub-hierarchy
370 where the numbers in the parentheses represent the numbers of processes
376 A, B and C's "populated" fields would be 1 while D's 0. After the one
377 process in C exits, B and C's "populated" fields would flip to "0" and
378 file modified events will be generated on the "cgroup.events" files of
382 Controlling Controllers
383 -----------------------
385 Enabling and Disabling
386 ~~~~~~~~~~~~~~~~~~~~~~
388 Each cgroup has a "cgroup.controllers" file which lists all
389 controllers available for the cgroup to enable::
391 # cat cgroup.controllers
394 No controller is enabled by default. Controllers can be enabled and
395 disabled by writing to the "cgroup.subtree_control" file::
397 # echo "+cpu +memory -io" > cgroup.subtree_control
399 Only controllers which are listed in "cgroup.controllers" can be
400 enabled. When multiple operations are specified as above, either they
401 all succeed or fail. If multiple operations on the same controller
402 are specified, the last one is effective.
404 Enabling a controller in a cgroup indicates that the distribution of
405 the target resource across its immediate children will be controlled.
406 Consider the following sub-hierarchy. The enabled controllers are
407 listed in parentheses::
409 A(cpu,memory) - B(memory) - C()
412 As A has "cpu" and "memory" enabled, A will control the distribution
413 of CPU cycles and memory to its children, in this case, B. As B has
414 "memory" enabled but not "CPU", C and D will compete freely on CPU
415 cycles but their division of memory available to B will be controlled.
417 As a controller regulates the distribution of the target resource to
418 the cgroup's children, enabling it creates the controller's interface
419 files in the child cgroups. In the above example, enabling "cpu" on B
420 would create the "cpu." prefixed controller interface files in C and
421 D. Likewise, disabling "memory" from B would remove the "memory."
422 prefixed controller interface files from C and D. This means that the
423 controller interface files - anything which doesn't start with
424 "cgroup." are owned by the parent rather than the cgroup itself.
430 Resources are distributed top-down and a cgroup can further distribute
431 a resource only if the resource has been distributed to it from the
432 parent. This means that all non-root "cgroup.subtree_control" files
433 can only contain controllers which are enabled in the parent's
434 "cgroup.subtree_control" file. A controller can be enabled only if
435 the parent has the controller enabled and a controller can't be
436 disabled if one or more children have it enabled.
439 No Internal Process Constraint
440 ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
442 Non-root cgroups can distribute domain resources to their children
443 only when they don't have any processes of their own. In other words,
444 only domain cgroups which don't contain any processes can have domain
445 controllers enabled in their "cgroup.subtree_control" files.
447 This guarantees that, when a domain controller is looking at the part
448 of the hierarchy which has it enabled, processes are always only on
449 the leaves. This rules out situations where child cgroups compete
450 against internal processes of the parent.
452 The root cgroup is exempt from this restriction. Root contains
453 processes and anonymous resource consumption which can't be associated
454 with any other cgroups and requires special treatment from most
455 controllers. How resource consumption in the root cgroup is governed
456 is up to each controller (for more information on this topic please
457 refer to the Non-normative information section in the Controllers
460 Note that the restriction doesn't get in the way if there is no
461 enabled controller in the cgroup's "cgroup.subtree_control". This is
462 important as otherwise it wouldn't be possible to create children of a
463 populated cgroup. To control resource distribution of a cgroup, the
464 cgroup must create children and transfer all its processes to the
465 children before enabling controllers in its "cgroup.subtree_control"
475 A cgroup can be delegated in two ways. First, to a less privileged
476 user by granting write access of the directory and its "cgroup.procs",
477 "cgroup.threads" and "cgroup.subtree_control" files to the user.
478 Second, if the "nsdelegate" mount option is set, automatically to a
479 cgroup namespace on namespace creation.
481 Because the resource control interface files in a given directory
482 control the distribution of the parent's resources, the delegatee
483 shouldn't be allowed to write to them. For the first method, this is
484 achieved by not granting access to these files. For the second, the
485 kernel rejects writes to all files other than "cgroup.procs" and
486 "cgroup.subtree_control" on a namespace root from inside the
489 The end results are equivalent for both delegation types. Once
490 delegated, the user can build sub-hierarchy under the directory,
491 organize processes inside it as it sees fit and further distribute the
492 resources it received from the parent. The limits and other settings
493 of all resource controllers are hierarchical and regardless of what
494 happens in the delegated sub-hierarchy, nothing can escape the
495 resource restrictions imposed by the parent.
497 Currently, cgroup doesn't impose any restrictions on the number of
498 cgroups in or nesting depth of a delegated sub-hierarchy; however,
499 this may be limited explicitly in the future.
502 Delegation Containment
503 ~~~~~~~~~~~~~~~~~~~~~~
505 A delegated sub-hierarchy is contained in the sense that processes
506 can't be moved into or out of the sub-hierarchy by the delegatee.
508 For delegations to a less privileged user, this is achieved by
509 requiring the following conditions for a process with a non-root euid
510 to migrate a target process into a cgroup by writing its PID to the
513 - The writer must have write access to the "cgroup.procs" file.
515 - The writer must have write access to the "cgroup.procs" file of the
516 common ancestor of the source and destination cgroups.
518 The above two constraints ensure that while a delegatee may migrate
519 processes around freely in the delegated sub-hierarchy it can't pull
520 in from or push out to outside the sub-hierarchy.
522 For an example, let's assume cgroups C0 and C1 have been delegated to
523 user U0 who created C00, C01 under C0 and C10 under C1 as follows and
524 all processes under C0 and C1 belong to U0::
526 ~~~~~~~~~~~~~ - C0 - C00
529 ~~~~~~~~~~~~~ - C1 - C10
531 Let's also say U0 wants to write the PID of a process which is
532 currently in C10 into "C00/cgroup.procs". U0 has write access to the
533 file; however, the common ancestor of the source cgroup C10 and the
534 destination cgroup C00 is above the points of delegation and U0 would
535 not have write access to its "cgroup.procs" files and thus the write
536 will be denied with -EACCES.
538 For delegations to namespaces, containment is achieved by requiring
539 that both the source and destination cgroups are reachable from the
540 namespace of the process which is attempting the migration. If either
541 is not reachable, the migration is rejected with -ENOENT.
547 Organize Once and Control
548 ~~~~~~~~~~~~~~~~~~~~~~~~~
550 Migrating a process across cgroups is a relatively expensive operation
551 and stateful resources such as memory are not moved together with the
552 process. This is an explicit design decision as there often exist
553 inherent trade-offs between migration and various hot paths in terms
554 of synchronization cost.
556 As such, migrating processes across cgroups frequently as a means to
557 apply different resource restrictions is discouraged. A workload
558 should be assigned to a cgroup according to the system's logical and
559 resource structure once on start-up. Dynamic adjustments to resource
560 distribution can be made by changing controller configuration through
564 Avoid Name Collisions
565 ~~~~~~~~~~~~~~~~~~~~~
567 Interface files for a cgroup and its children cgroups occupy the same
568 directory and it is possible to create children cgroups which collide
569 with interface files.
571 All cgroup core interface files are prefixed with "cgroup." and each
572 controller's interface files are prefixed with the controller name and
573 a dot. A controller's name is composed of lower case alphabets and
574 '_'s but never begins with an '_' so it can be used as the prefix
575 character for collision avoidance. Also, interface file names won't
576 start or end with terms which are often used in categorizing workloads
577 such as job, service, slice, unit or workload.
579 cgroup doesn't do anything to prevent name collisions and it's the
580 user's responsibility to avoid them.
583 Resource Distribution Models
584 ============================
586 cgroup controllers implement several resource distribution schemes
587 depending on the resource type and expected use cases. This section
588 describes major schemes in use along with their expected behaviors.
594 A parent's resource is distributed by adding up the weights of all
595 active children and giving each the fraction matching the ratio of its
596 weight against the sum. As only children which can make use of the
597 resource at the moment participate in the distribution, this is
598 work-conserving. Due to the dynamic nature, this model is usually
599 used for stateless resources.
601 All weights are in the range [1, 10000] with the default at 100. This
602 allows symmetric multiplicative biases in both directions at fine
603 enough granularity while staying in the intuitive range.
605 As long as the weight is in range, all configuration combinations are
606 valid and there is no reason to reject configuration changes or
609 "cpu.weight" proportionally distributes CPU cycles to active children
610 and is an example of this type.
616 A child can only consume upto the configured amount of the resource.
617 Limits can be over-committed - the sum of the limits of children can
618 exceed the amount of resource available to the parent.
620 Limits are in the range [0, max] and defaults to "max", which is noop.
622 As limits can be over-committed, all configuration combinations are
623 valid and there is no reason to reject configuration changes or
626 "io.max" limits the maximum BPS and/or IOPS that a cgroup can consume
627 on an IO device and is an example of this type.
633 A cgroup is protected upto the configured amount of the resource
634 as long as the usages of all its ancestors are under their
635 protected levels. Protections can be hard guarantees or best effort
636 soft boundaries. Protections can also be over-committed in which case
637 only upto the amount available to the parent is protected among
640 Protections are in the range [0, max] and defaults to 0, which is
643 As protections can be over-committed, all configuration combinations
644 are valid and there is no reason to reject configuration changes or
647 "memory.low" implements best-effort memory protection and is an
648 example of this type.
654 A cgroup is exclusively allocated a certain amount of a finite
655 resource. Allocations can't be over-committed - the sum of the
656 allocations of children can not exceed the amount of resource
657 available to the parent.
659 Allocations are in the range [0, max] and defaults to 0, which is no
662 As allocations can't be over-committed, some configuration
663 combinations are invalid and should be rejected. Also, if the
664 resource is mandatory for execution of processes, process migrations
667 "cpu.rt.max" hard-allocates realtime slices and is an example of this
677 All interface files should be in one of the following formats whenever
680 New-line separated values
681 (when only one value can be written at once)
687 Space separated values
688 (when read-only or multiple values can be written at once)
700 KEY0 SUB_KEY0=VAL00 SUB_KEY1=VAL01...
701 KEY1 SUB_KEY0=VAL10 SUB_KEY1=VAL11...
704 For a writable file, the format for writing should generally match
705 reading; however, controllers may allow omitting later fields or
706 implement restricted shortcuts for most common use cases.
708 For both flat and nested keyed files, only the values for a single key
709 can be written at a time. For nested keyed files, the sub key pairs
710 may be specified in any order and not all pairs have to be specified.
716 - Settings for a single feature should be contained in a single file.
718 - The root cgroup should be exempt from resource control and thus
719 shouldn't have resource control interface files.
721 - The default time unit is microseconds. If a different unit is ever
722 used, an explicit unit suffix must be present.
724 - A parts-per quantity should use a percentage decimal with at least
725 two digit fractional part - e.g. 13.40.
727 - If a controller implements weight based resource distribution, its
728 interface file should be named "weight" and have the range [1,
729 10000] with 100 as the default. The values are chosen to allow
730 enough and symmetric bias in both directions while keeping it
731 intuitive (the default is 100%).
733 - If a controller implements an absolute resource guarantee and/or
734 limit, the interface files should be named "min" and "max"
735 respectively. If a controller implements best effort resource
736 guarantee and/or limit, the interface files should be named "low"
737 and "high" respectively.
739 In the above four control files, the special token "max" should be
740 used to represent upward infinity for both reading and writing.
742 - If a setting has a configurable default value and keyed specific
743 overrides, the default entry should be keyed with "default" and
744 appear as the first entry in the file.
746 The default value can be updated by writing either "default $VAL" or
749 When writing to update a specific override, "default" can be used as
750 the value to indicate removal of the override. Override entries
751 with "default" as the value must not appear when read.
753 For example, a setting which is keyed by major:minor device numbers
754 with integer values may look like the following::
756 # cat cgroup-example-interface-file
760 The default value can be updated by::
762 # echo 125 > cgroup-example-interface-file
766 # echo "default 125" > cgroup-example-interface-file
768 An override can be set by::
770 # echo "8:16 170" > cgroup-example-interface-file
774 # echo "8:0 default" > cgroup-example-interface-file
775 # cat cgroup-example-interface-file
779 - For events which are not very high frequency, an interface file
780 "events" should be created which lists event key value pairs.
781 Whenever a notifiable event happens, file modified event should be
782 generated on the file.
788 All cgroup core files are prefixed with "cgroup."
791 A read-write single value file which exists on non-root
794 When read, it indicates the current type of the cgroup, which
795 can be one of the following values.
797 - "domain" : A normal valid domain cgroup.
799 - "domain threaded" : A threaded domain cgroup which is
800 serving as the root of a threaded subtree.
802 - "domain invalid" : A cgroup which is in an invalid state.
803 It can't be populated or have controllers enabled. It may
804 be allowed to become a threaded cgroup.
806 - "threaded" : A threaded cgroup which is a member of a
809 A cgroup can be turned into a threaded cgroup by writing
810 "threaded" to this file.
813 A read-write new-line separated values file which exists on
816 When read, it lists the PIDs of all processes which belong to
817 the cgroup one-per-line. The PIDs are not ordered and the
818 same PID may show up more than once if the process got moved
819 to another cgroup and then back or the PID got recycled while
822 A PID can be written to migrate the process associated with
823 the PID to the cgroup. The writer should match all of the
824 following conditions.
826 - It must have write access to the "cgroup.procs" file.
828 - It must have write access to the "cgroup.procs" file of the
829 common ancestor of the source and destination cgroups.
831 When delegating a sub-hierarchy, write access to this file
832 should be granted along with the containing directory.
834 In a threaded cgroup, reading this file fails with EOPNOTSUPP
835 as all the processes belong to the thread root. Writing is
836 supported and moves every thread of the process to the cgroup.
839 A read-write new-line separated values file which exists on
842 When read, it lists the TIDs of all threads which belong to
843 the cgroup one-per-line. The TIDs are not ordered and the
844 same TID may show up more than once if the thread got moved to
845 another cgroup and then back or the TID got recycled while
848 A TID can be written to migrate the thread associated with the
849 TID to the cgroup. The writer should match all of the
850 following conditions.
852 - It must have write access to the "cgroup.threads" file.
854 - The cgroup that the thread is currently in must be in the
855 same resource domain as the destination cgroup.
857 - It must have write access to the "cgroup.procs" file of the
858 common ancestor of the source and destination cgroups.
860 When delegating a sub-hierarchy, write access to this file
861 should be granted along with the containing directory.
864 A read-only space separated values file which exists on all
867 It shows space separated list of all controllers available to
868 the cgroup. The controllers are not ordered.
870 cgroup.subtree_control
871 A read-write space separated values file which exists on all
872 cgroups. Starts out empty.
874 When read, it shows space separated list of the controllers
875 which are enabled to control resource distribution from the
876 cgroup to its children.
878 Space separated list of controllers prefixed with '+' or '-'
879 can be written to enable or disable controllers. A controller
880 name prefixed with '+' enables the controller and '-'
881 disables. If a controller appears more than once on the list,
882 the last one is effective. When multiple enable and disable
883 operations are specified, either all succeed or all fail.
886 A read-only flat-keyed file which exists on non-root cgroups.
887 The following entries are defined. Unless specified
888 otherwise, a value change in this file generates a file
892 1 if the cgroup or its descendants contains any live
893 processes; otherwise, 0.
895 1 if the cgroup is frozen; otherwise, 0.
897 cgroup.max.descendants
898 A read-write single value files. The default is "max".
900 Maximum allowed number of descent cgroups.
901 If the actual number of descendants is equal or larger,
902 an attempt to create a new cgroup in the hierarchy will fail.
905 A read-write single value files. The default is "max".
907 Maximum allowed descent depth below the current cgroup.
908 If the actual descent depth is equal or larger,
909 an attempt to create a new child cgroup will fail.
912 A read-only flat-keyed file with the following entries:
915 Total number of visible descendant cgroups.
918 Total number of dying descendant cgroups. A cgroup becomes
919 dying after being deleted by a user. The cgroup will remain
920 in dying state for some time undefined time (which can depend
921 on system load) before being completely destroyed.
923 A process can't enter a dying cgroup under any circumstances,
924 a dying cgroup can't revive.
926 A dying cgroup can consume system resources not exceeding
927 limits, which were active at the moment of cgroup deletion.
930 A read-write single value file which exists on non-root cgroups.
931 Allowed values are "0" and "1". The default is "0".
933 Writing "1" to the file causes freezing of the cgroup and all
934 descendant cgroups. This means that all belonging processes will
935 be stopped and will not run until the cgroup will be explicitly
936 unfrozen. Freezing of the cgroup may take some time; when this action
937 is completed, the "frozen" value in the cgroup.events control file
938 will be updated to "1" and the corresponding notification will be
941 A cgroup can be frozen either by its own settings, or by settings
942 of any ancestor cgroups. If any of ancestor cgroups is frozen, the
943 cgroup will remain frozen.
945 Processes in the frozen cgroup can be killed by a fatal signal.
946 They also can enter and leave a frozen cgroup: either by an explicit
947 move by a user, or if freezing of the cgroup races with fork().
948 If a process is moved to a frozen cgroup, it stops. If a process is
949 moved out of a frozen cgroup, it becomes running.
951 Frozen status of a cgroup doesn't affect any cgroup tree operations:
952 it's possible to delete a frozen (and empty) cgroup, as well as
953 create new sub-cgroups.
963 The "cpu" controllers regulates distribution of CPU cycles. This
964 controller implements weight and absolute bandwidth limit models for
965 normal scheduling policy and absolute bandwidth allocation model for
966 realtime scheduling policy.
968 In all the above models, cycles distribution is defined only on a temporal
969 base and it does not account for the frequency at which tasks are executed.
970 The (optional) utilization clamping support allows to hint the schedutil
971 cpufreq governor about the minimum desired frequency which should always be
972 provided by a CPU, as well as the maximum desired frequency, which should not
973 be exceeded by a CPU.
975 WARNING: cgroup2 doesn't yet support control of realtime processes and
976 the cpu controller can only be enabled when all RT processes are in
977 the root cgroup. Be aware that system management software may already
978 have placed RT processes into nonroot cgroups during the system boot
979 process, and these processes may need to be moved to the root cgroup
980 before the cpu controller can be enabled.
986 All time durations are in microseconds.
989 A read-only flat-keyed file.
990 This file exists whether the controller is enabled or not.
992 It always reports the following three stats:
998 and the following three when the controller is enabled:
1005 A read-write single value file which exists on non-root
1006 cgroups. The default is "100".
1008 The weight in the range [1, 10000].
1011 A read-write single value file which exists on non-root
1012 cgroups. The default is "0".
1014 The nice value is in the range [-20, 19].
1016 This interface file is an alternative interface for
1017 "cpu.weight" and allows reading and setting weight using the
1018 same values used by nice(2). Because the range is smaller and
1019 granularity is coarser for the nice values, the read value is
1020 the closest approximation of the current weight.
1023 A read-write two value file which exists on non-root cgroups.
1024 The default is "max 100000".
1026 The maximum bandwidth limit. It's in the following format::
1030 which indicates that the group may consume upto $MAX in each
1031 $PERIOD duration. "max" for $MAX indicates no limit. If only
1032 one number is written, $MAX is updated.
1035 A read-write nested-keyed file.
1037 Shows pressure stall information for CPU. See
1038 :ref:`Documentation/accounting/psi.rst <psi>` for details.
1041 A read-write single value file which exists on non-root cgroups.
1042 The default is "0", i.e. no utilization boosting.
1044 The requested minimum utilization (protection) as a percentage
1045 rational number, e.g. 12.34 for 12.34%.
1047 This interface allows reading and setting minimum utilization clamp
1048 values similar to the sched_setattr(2). This minimum utilization
1049 value is used to clamp the task specific minimum utilization clamp.
1051 The requested minimum utilization (protection) is always capped by
1052 the current value for the maximum utilization (limit), i.e.
1056 A read-write single value file which exists on non-root cgroups.
1057 The default is "max". i.e. no utilization capping
1059 The requested maximum utilization (limit) as a percentage rational
1060 number, e.g. 98.76 for 98.76%.
1062 This interface allows reading and setting maximum utilization clamp
1063 values similar to the sched_setattr(2). This maximum utilization
1064 value is used to clamp the task specific maximum utilization clamp.
1071 The "memory" controller regulates distribution of memory. Memory is
1072 stateful and implements both limit and protection models. Due to the
1073 intertwining between memory usage and reclaim pressure and the
1074 stateful nature of memory, the distribution model is relatively
1077 While not completely water-tight, all major memory usages by a given
1078 cgroup are tracked so that the total memory consumption can be
1079 accounted and controlled to a reasonable extent. Currently, the
1080 following types of memory usages are tracked.
1082 - Userland memory - page cache and anonymous memory.
1084 - Kernel data structures such as dentries and inodes.
1086 - TCP socket buffers.
1088 The above list may expand in the future for better coverage.
1091 Memory Interface Files
1092 ~~~~~~~~~~~~~~~~~~~~~~
1094 All memory amounts are in bytes. If a value which is not aligned to
1095 PAGE_SIZE is written, the value may be rounded up to the closest
1096 PAGE_SIZE multiple when read back.
1099 A read-only single value file which exists on non-root
1102 The total amount of memory currently being used by the cgroup
1103 and its descendants.
1106 A read-write single value file which exists on non-root
1107 cgroups. The default is "0".
1109 Hard memory protection. If the memory usage of a cgroup
1110 is within its effective min boundary, the cgroup's memory
1111 won't be reclaimed under any conditions. If there is no
1112 unprotected reclaimable memory available, OOM killer
1113 is invoked. Above the effective min boundary (or
1114 effective low boundary if it is higher), pages are reclaimed
1115 proportionally to the overage, reducing reclaim pressure for
1118 Effective min boundary is limited by memory.min values of
1119 all ancestor cgroups. If there is memory.min overcommitment
1120 (child cgroup or cgroups are requiring more protected memory
1121 than parent will allow), then each child cgroup will get
1122 the part of parent's protection proportional to its
1123 actual memory usage below memory.min.
1125 Putting more memory than generally available under this
1126 protection is discouraged and may lead to constant OOMs.
1128 If a memory cgroup is not populated with processes,
1129 its memory.min is ignored.
1132 A read-write single value file which exists on non-root
1133 cgroups. The default is "0".
1135 Best-effort memory protection. If the memory usage of a
1136 cgroup is within its effective low boundary, the cgroup's
1137 memory won't be reclaimed unless there is no reclaimable
1138 memory available in unprotected cgroups.
1139 Above the effective low boundary (or
1140 effective min boundary if it is higher), pages are reclaimed
1141 proportionally to the overage, reducing reclaim pressure for
1144 Effective low boundary is limited by memory.low values of
1145 all ancestor cgroups. If there is memory.low overcommitment
1146 (child cgroup or cgroups are requiring more protected memory
1147 than parent will allow), then each child cgroup will get
1148 the part of parent's protection proportional to its
1149 actual memory usage below memory.low.
1151 Putting more memory than generally available under this
1152 protection is discouraged.
1155 A read-write single value file which exists on non-root
1156 cgroups. The default is "max".
1158 Memory usage throttle limit. This is the main mechanism to
1159 control memory usage of a cgroup. If a cgroup's usage goes
1160 over the high boundary, the processes of the cgroup are
1161 throttled and put under heavy reclaim pressure.
1163 Going over the high limit never invokes the OOM killer and
1164 under extreme conditions the limit may be breached.
1167 A read-write single value file which exists on non-root
1168 cgroups. The default is "max".
1170 Memory usage hard limit. This is the final protection
1171 mechanism. If a cgroup's memory usage reaches this limit and
1172 can't be reduced, the OOM killer is invoked in the cgroup.
1173 Under certain circumstances, the usage may go over the limit
1176 In default configuration regular 0-order allocations always
1177 succeed unless OOM killer chooses current task as a victim.
1179 Some kinds of allocations don't invoke the OOM killer.
1180 Caller could retry them differently, return into userspace
1181 as -ENOMEM or silently ignore in cases like disk readahead.
1183 This is the ultimate protection mechanism. As long as the
1184 high limit is used and monitored properly, this limit's
1185 utility is limited to providing the final safety net.
1188 A read-write single value file which exists on non-root
1189 cgroups. The default value is "0".
1191 Determines whether the cgroup should be treated as
1192 an indivisible workload by the OOM killer. If set,
1193 all tasks belonging to the cgroup or to its descendants
1194 (if the memory cgroup is not a leaf cgroup) are killed
1195 together or not at all. This can be used to avoid
1196 partial kills to guarantee workload integrity.
1198 Tasks with the OOM protection (oom_score_adj set to -1000)
1199 are treated as an exception and are never killed.
1201 If the OOM killer is invoked in a cgroup, it's not going
1202 to kill any tasks outside of this cgroup, regardless
1203 memory.oom.group values of ancestor cgroups.
1206 A read-only flat-keyed file which exists on non-root cgroups.
1207 The following entries are defined. Unless specified
1208 otherwise, a value change in this file generates a file
1211 Note that all fields in this file are hierarchical and the
1212 file modified event can be generated due to an event down the
1213 hierarchy. For for the local events at the cgroup level see
1214 memory.events.local.
1217 The number of times the cgroup is reclaimed due to
1218 high memory pressure even though its usage is under
1219 the low boundary. This usually indicates that the low
1220 boundary is over-committed.
1223 The number of times processes of the cgroup are
1224 throttled and routed to perform direct memory reclaim
1225 because the high memory boundary was exceeded. For a
1226 cgroup whose memory usage is capped by the high limit
1227 rather than global memory pressure, this event's
1228 occurrences are expected.
1231 The number of times the cgroup's memory usage was
1232 about to go over the max boundary. If direct reclaim
1233 fails to bring it down, the cgroup goes to OOM state.
1236 The number of time the cgroup's memory usage was
1237 reached the limit and allocation was about to fail.
1239 This event is not raised if the OOM killer is not
1240 considered as an option, e.g. for failed high-order
1241 allocations or if caller asked to not retry attempts.
1244 The number of processes belonging to this cgroup
1245 killed by any kind of OOM killer.
1248 Similar to memory.events but the fields in the file are local
1249 to the cgroup i.e. not hierarchical. The file modified event
1250 generated on this file reflects only the local events.
1253 A read-only flat-keyed file which exists on non-root cgroups.
1255 This breaks down the cgroup's memory footprint into different
1256 types of memory, type-specific details, and other information
1257 on the state and past events of the memory management system.
1259 All memory amounts are in bytes.
1261 The entries are ordered to be human readable, and new entries
1262 can show up in the middle. Don't rely on items remaining in a
1263 fixed position; use the keys to look up specific values!
1265 If the entry has no per-node counter (or not show in the
1266 memory.numa_stat). We use 'npn' (non-per-node) as the tag
1267 to indicate that it will not show in the memory.numa_stat.
1270 Amount of memory used in anonymous mappings such as
1271 brk(), sbrk(), and mmap(MAP_ANONYMOUS)
1274 Amount of memory used to cache filesystem data,
1275 including tmpfs and shared memory.
1278 Amount of memory allocated to kernel stacks.
1281 Amount of memory allocated for page tables.
1284 Amount of memory used for storing per-cpu kernel
1288 Amount of memory used in network transmission buffers
1291 Amount of cached filesystem data that is swap-backed,
1292 such as tmpfs, shm segments, shared anonymous mmap()s
1295 Amount of cached filesystem data mapped with mmap()
1298 Amount of cached filesystem data that was modified but
1299 not yet written back to disk
1302 Amount of cached filesystem data that was modified and
1303 is currently being written back to disk
1306 Amount of swap cached in memory. The swapcache is accounted
1307 against both memory and swap usage.
1310 Amount of memory used in anonymous mappings backed by
1311 transparent hugepages
1314 Amount of cached filesystem data backed by transparent
1318 Amount of shm, tmpfs, shared anonymous mmap()s backed by
1319 transparent hugepages
1321 inactive_anon, active_anon, inactive_file, active_file, unevictable
1322 Amount of memory, swap-backed and filesystem-backed,
1323 on the internal memory management lists used by the
1324 page reclaim algorithm.
1326 As these represent internal list state (eg. shmem pages are on anon
1327 memory management lists), inactive_foo + active_foo may not be equal to
1328 the value for the foo counter, since the foo counter is type-based, not
1332 Part of "slab" that might be reclaimed, such as
1333 dentries and inodes.
1336 Part of "slab" that cannot be reclaimed on memory
1340 Amount of memory used for storing in-kernel data
1343 workingset_refault_anon
1344 Number of refaults of previously evicted anonymous pages.
1346 workingset_refault_file
1347 Number of refaults of previously evicted file pages.
1349 workingset_activate_anon
1350 Number of refaulted anonymous pages that were immediately
1353 workingset_activate_file
1354 Number of refaulted file pages that were immediately activated.
1356 workingset_restore_anon
1357 Number of restored anonymous pages which have been detected as
1358 an active workingset before they got reclaimed.
1360 workingset_restore_file
1361 Number of restored file pages which have been detected as an
1362 active workingset before they got reclaimed.
1364 workingset_nodereclaim
1365 Number of times a shadow node has been reclaimed
1368 Total number of page faults incurred
1371 Number of major page faults incurred
1374 Amount of scanned pages (in an active LRU list)
1377 Amount of scanned pages (in an inactive LRU list)
1380 Amount of reclaimed pages
1383 Amount of pages moved to the active LRU list
1386 Amount of pages moved to the inactive LRU list
1389 Amount of pages postponed to be freed under memory pressure
1392 Amount of reclaimed lazyfree pages
1394 thp_fault_alloc (npn)
1395 Number of transparent hugepages which were allocated to satisfy
1396 a page fault. This counter is not present when CONFIG_TRANSPARENT_HUGEPAGE
1399 thp_collapse_alloc (npn)
1400 Number of transparent hugepages which were allocated to allow
1401 collapsing an existing range of pages. This counter is not
1402 present when CONFIG_TRANSPARENT_HUGEPAGE is not set.
1405 A read-only nested-keyed file which exists on non-root cgroups.
1407 This breaks down the cgroup's memory footprint into different
1408 types of memory, type-specific details, and other information
1409 per node on the state of the memory management system.
1411 This is useful for providing visibility into the NUMA locality
1412 information within an memcg since the pages are allowed to be
1413 allocated from any physical node. One of the use case is evaluating
1414 application performance by combining this information with the
1415 application's CPU allocation.
1417 All memory amounts are in bytes.
1419 The output format of memory.numa_stat is::
1421 type N0=<bytes in node 0> N1=<bytes in node 1> ...
1423 The entries are ordered to be human readable, and new entries
1424 can show up in the middle. Don't rely on items remaining in a
1425 fixed position; use the keys to look up specific values!
1427 The entries can refer to the memory.stat.
1430 A read-only single value file which exists on non-root
1433 The total amount of swap currently being used by the cgroup
1434 and its descendants.
1437 A read-write single value file which exists on non-root
1438 cgroups. The default is "max".
1440 Swap usage throttle limit. If a cgroup's swap usage exceeds
1441 this limit, all its further allocations will be throttled to
1442 allow userspace to implement custom out-of-memory procedures.
1444 This limit marks a point of no return for the cgroup. It is NOT
1445 designed to manage the amount of swapping a workload does
1446 during regular operation. Compare to memory.swap.max, which
1447 prohibits swapping past a set amount, but lets the cgroup
1448 continue unimpeded as long as other memory can be reclaimed.
1450 Healthy workloads are not expected to reach this limit.
1453 A read-write single value file which exists on non-root
1454 cgroups. The default is "max".
1456 Swap usage hard limit. If a cgroup's swap usage reaches this
1457 limit, anonymous memory of the cgroup will not be swapped out.
1460 A read-only flat-keyed file which exists on non-root cgroups.
1461 The following entries are defined. Unless specified
1462 otherwise, a value change in this file generates a file
1466 The number of times the cgroup's swap usage was over
1470 The number of times the cgroup's swap usage was about
1471 to go over the max boundary and swap allocation
1475 The number of times swap allocation failed either
1476 because of running out of swap system-wide or max
1479 When reduced under the current usage, the existing swap
1480 entries are reclaimed gradually and the swap usage may stay
1481 higher than the limit for an extended period of time. This
1482 reduces the impact on the workload and memory management.
1485 A read-only nested-keyed file.
1487 Shows pressure stall information for memory. See
1488 :ref:`Documentation/accounting/psi.rst <psi>` for details.
1494 "memory.high" is the main mechanism to control memory usage.
1495 Over-committing on high limit (sum of high limits > available memory)
1496 and letting global memory pressure to distribute memory according to
1497 usage is a viable strategy.
1499 Because breach of the high limit doesn't trigger the OOM killer but
1500 throttles the offending cgroup, a management agent has ample
1501 opportunities to monitor and take appropriate actions such as granting
1502 more memory or terminating the workload.
1504 Determining whether a cgroup has enough memory is not trivial as
1505 memory usage doesn't indicate whether the workload can benefit from
1506 more memory. For example, a workload which writes data received from
1507 network to a file can use all available memory but can also operate as
1508 performant with a small amount of memory. A measure of memory
1509 pressure - how much the workload is being impacted due to lack of
1510 memory - is necessary to determine whether a workload needs more
1511 memory; unfortunately, memory pressure monitoring mechanism isn't
1518 A memory area is charged to the cgroup which instantiated it and stays
1519 charged to the cgroup until the area is released. Migrating a process
1520 to a different cgroup doesn't move the memory usages that it
1521 instantiated while in the previous cgroup to the new cgroup.
1523 A memory area may be used by processes belonging to different cgroups.
1524 To which cgroup the area will be charged is in-deterministic; however,
1525 over time, the memory area is likely to end up in a cgroup which has
1526 enough memory allowance to avoid high reclaim pressure.
1528 If a cgroup sweeps a considerable amount of memory which is expected
1529 to be accessed repeatedly by other cgroups, it may make sense to use
1530 POSIX_FADV_DONTNEED to relinquish the ownership of memory areas
1531 belonging to the affected files to ensure correct memory ownership.
1537 The "io" controller regulates the distribution of IO resources. This
1538 controller implements both weight based and absolute bandwidth or IOPS
1539 limit distribution; however, weight based distribution is available
1540 only if cfq-iosched is in use and neither scheme is available for
1548 A read-only nested-keyed file.
1550 Lines are keyed by $MAJ:$MIN device numbers and not ordered.
1551 The following nested keys are defined.
1553 ====== =====================
1555 wbytes Bytes written
1556 rios Number of read IOs
1557 wios Number of write IOs
1558 dbytes Bytes discarded
1559 dios Number of discard IOs
1560 ====== =====================
1562 An example read output follows::
1564 8:16 rbytes=1459200 wbytes=314773504 rios=192 wios=353 dbytes=0 dios=0
1565 8:0 rbytes=90430464 wbytes=299008000 rios=8950 wios=1252 dbytes=50331648 dios=3021
1568 A read-write nested-keyed file which exists only on the root
1571 This file configures the Quality of Service of the IO cost
1572 model based controller (CONFIG_BLK_CGROUP_IOCOST) which
1573 currently implements "io.weight" proportional control. Lines
1574 are keyed by $MAJ:$MIN device numbers and not ordered. The
1575 line for a given device is populated on the first write for
1576 the device on "io.cost.qos" or "io.cost.model". The following
1577 nested keys are defined.
1579 ====== =====================================
1580 enable Weight-based control enable
1581 ctrl "auto" or "user"
1582 rpct Read latency percentile [0, 100]
1583 rlat Read latency threshold
1584 wpct Write latency percentile [0, 100]
1585 wlat Write latency threshold
1586 min Minimum scaling percentage [1, 10000]
1587 max Maximum scaling percentage [1, 10000]
1588 ====== =====================================
1590 The controller is disabled by default and can be enabled by
1591 setting "enable" to 1. "rpct" and "wpct" parameters default
1592 to zero and the controller uses internal device saturation
1593 state to adjust the overall IO rate between "min" and "max".
1595 When a better control quality is needed, latency QoS
1596 parameters can be configured. For example::
1598 8:16 enable=1 ctrl=auto rpct=95.00 rlat=75000 wpct=95.00 wlat=150000 min=50.00 max=150.0
1600 shows that on sdb, the controller is enabled, will consider
1601 the device saturated if the 95th percentile of read completion
1602 latencies is above 75ms or write 150ms, and adjust the overall
1603 IO issue rate between 50% and 150% accordingly.
1605 The lower the saturation point, the better the latency QoS at
1606 the cost of aggregate bandwidth. The narrower the allowed
1607 adjustment range between "min" and "max", the more conformant
1608 to the cost model the IO behavior. Note that the IO issue
1609 base rate may be far off from 100% and setting "min" and "max"
1610 blindly can lead to a significant loss of device capacity or
1611 control quality. "min" and "max" are useful for regulating
1612 devices which show wide temporary behavior changes - e.g. a
1613 ssd which accepts writes at the line speed for a while and
1614 then completely stalls for multiple seconds.
1616 When "ctrl" is "auto", the parameters are controlled by the
1617 kernel and may change automatically. Setting "ctrl" to "user"
1618 or setting any of the percentile and latency parameters puts
1619 it into "user" mode and disables the automatic changes. The
1620 automatic mode can be restored by setting "ctrl" to "auto".
1623 A read-write nested-keyed file which exists only on the root
1626 This file configures the cost model of the IO cost model based
1627 controller (CONFIG_BLK_CGROUP_IOCOST) which currently
1628 implements "io.weight" proportional control. Lines are keyed
1629 by $MAJ:$MIN device numbers and not ordered. The line for a
1630 given device is populated on the first write for the device on
1631 "io.cost.qos" or "io.cost.model". The following nested keys
1634 ===== ================================
1635 ctrl "auto" or "user"
1636 model The cost model in use - "linear"
1637 ===== ================================
1639 When "ctrl" is "auto", the kernel may change all parameters
1640 dynamically. When "ctrl" is set to "user" or any other
1641 parameters are written to, "ctrl" become "user" and the
1642 automatic changes are disabled.
1644 When "model" is "linear", the following model parameters are
1647 ============= ========================================
1648 [r|w]bps The maximum sequential IO throughput
1649 [r|w]seqiops The maximum 4k sequential IOs per second
1650 [r|w]randiops The maximum 4k random IOs per second
1651 ============= ========================================
1653 From the above, the builtin linear model determines the base
1654 costs of a sequential and random IO and the cost coefficient
1655 for the IO size. While simple, this model can cover most
1656 common device classes acceptably.
1658 The IO cost model isn't expected to be accurate in absolute
1659 sense and is scaled to the device behavior dynamically.
1661 If needed, tools/cgroup/iocost_coef_gen.py can be used to
1662 generate device-specific coefficients.
1665 A read-write flat-keyed file which exists on non-root cgroups.
1666 The default is "default 100".
1668 The first line is the default weight applied to devices
1669 without specific override. The rest are overrides keyed by
1670 $MAJ:$MIN device numbers and not ordered. The weights are in
1671 the range [1, 10000] and specifies the relative amount IO time
1672 the cgroup can use in relation to its siblings.
1674 The default weight can be updated by writing either "default
1675 $WEIGHT" or simply "$WEIGHT". Overrides can be set by writing
1676 "$MAJ:$MIN $WEIGHT" and unset by writing "$MAJ:$MIN default".
1678 An example read output follows::
1685 A read-write nested-keyed file which exists on non-root
1688 BPS and IOPS based IO limit. Lines are keyed by $MAJ:$MIN
1689 device numbers and not ordered. The following nested keys are
1692 ===== ==================================
1693 rbps Max read bytes per second
1694 wbps Max write bytes per second
1695 riops Max read IO operations per second
1696 wiops Max write IO operations per second
1697 ===== ==================================
1699 When writing, any number of nested key-value pairs can be
1700 specified in any order. "max" can be specified as the value
1701 to remove a specific limit. If the same key is specified
1702 multiple times, the outcome is undefined.
1704 BPS and IOPS are measured in each IO direction and IOs are
1705 delayed if limit is reached. Temporary bursts are allowed.
1707 Setting read limit at 2M BPS and write at 120 IOPS for 8:16::
1709 echo "8:16 rbps=2097152 wiops=120" > io.max
1711 Reading returns the following::
1713 8:16 rbps=2097152 wbps=max riops=max wiops=120
1715 Write IOPS limit can be removed by writing the following::
1717 echo "8:16 wiops=max" > io.max
1719 Reading now returns the following::
1721 8:16 rbps=2097152 wbps=max riops=max wiops=max
1724 A read-only nested-keyed file.
1726 Shows pressure stall information for IO. See
1727 :ref:`Documentation/accounting/psi.rst <psi>` for details.
1733 Page cache is dirtied through buffered writes and shared mmaps and
1734 written asynchronously to the backing filesystem by the writeback
1735 mechanism. Writeback sits between the memory and IO domains and
1736 regulates the proportion of dirty memory by balancing dirtying and
1739 The io controller, in conjunction with the memory controller,
1740 implements control of page cache writeback IOs. The memory controller
1741 defines the memory domain that dirty memory ratio is calculated and
1742 maintained for and the io controller defines the io domain which
1743 writes out dirty pages for the memory domain. Both system-wide and
1744 per-cgroup dirty memory states are examined and the more restrictive
1745 of the two is enforced.
1747 cgroup writeback requires explicit support from the underlying
1748 filesystem. Currently, cgroup writeback is implemented on ext2, ext4,
1749 btrfs, f2fs, and xfs. On other filesystems, all writeback IOs are
1750 attributed to the root cgroup.
1752 There are inherent differences in memory and writeback management
1753 which affects how cgroup ownership is tracked. Memory is tracked per
1754 page while writeback per inode. For the purpose of writeback, an
1755 inode is assigned to a cgroup and all IO requests to write dirty pages
1756 from the inode are attributed to that cgroup.
1758 As cgroup ownership for memory is tracked per page, there can be pages
1759 which are associated with different cgroups than the one the inode is
1760 associated with. These are called foreign pages. The writeback
1761 constantly keeps track of foreign pages and, if a particular foreign
1762 cgroup becomes the majority over a certain period of time, switches
1763 the ownership of the inode to that cgroup.
1765 While this model is enough for most use cases where a given inode is
1766 mostly dirtied by a single cgroup even when the main writing cgroup
1767 changes over time, use cases where multiple cgroups write to a single
1768 inode simultaneously are not supported well. In such circumstances, a
1769 significant portion of IOs are likely to be attributed incorrectly.
1770 As memory controller assigns page ownership on the first use and
1771 doesn't update it until the page is released, even if writeback
1772 strictly follows page ownership, multiple cgroups dirtying overlapping
1773 areas wouldn't work as expected. It's recommended to avoid such usage
1776 The sysctl knobs which affect writeback behavior are applied to cgroup
1777 writeback as follows.
1779 vm.dirty_background_ratio, vm.dirty_ratio
1780 These ratios apply the same to cgroup writeback with the
1781 amount of available memory capped by limits imposed by the
1782 memory controller and system-wide clean memory.
1784 vm.dirty_background_bytes, vm.dirty_bytes
1785 For cgroup writeback, this is calculated into ratio against
1786 total available memory and applied the same way as
1787 vm.dirty[_background]_ratio.
1793 This is a cgroup v2 controller for IO workload protection. You provide a group
1794 with a latency target, and if the average latency exceeds that target the
1795 controller will throttle any peers that have a lower latency target than the
1798 The limits are only applied at the peer level in the hierarchy. This means that
1799 in the diagram below, only groups A, B, and C will influence each other, and
1800 groups D and F will influence each other. Group G will influence nobody::
1809 So the ideal way to configure this is to set io.latency in groups A, B, and C.
1810 Generally you do not want to set a value lower than the latency your device
1811 supports. Experiment to find the value that works best for your workload.
1812 Start at higher than the expected latency for your device and watch the
1813 avg_lat value in io.stat for your workload group to get an idea of the
1814 latency you see during normal operation. Use the avg_lat value as a basis for
1815 your real setting, setting at 10-15% higher than the value in io.stat.
1817 How IO Latency Throttling Works
1818 ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
1820 io.latency is work conserving; so as long as everybody is meeting their latency
1821 target the controller doesn't do anything. Once a group starts missing its
1822 target it begins throttling any peer group that has a higher target than itself.
1823 This throttling takes 2 forms:
1825 - Queue depth throttling. This is the number of outstanding IO's a group is
1826 allowed to have. We will clamp down relatively quickly, starting at no limit
1827 and going all the way down to 1 IO at a time.
1829 - Artificial delay induction. There are certain types of IO that cannot be
1830 throttled without possibly adversely affecting higher priority groups. This
1831 includes swapping and metadata IO. These types of IO are allowed to occur
1832 normally, however they are "charged" to the originating group. If the
1833 originating group is being throttled you will see the use_delay and delay
1834 fields in io.stat increase. The delay value is how many microseconds that are
1835 being added to any process that runs in this group. Because this number can
1836 grow quite large if there is a lot of swapping or metadata IO occurring we
1837 limit the individual delay events to 1 second at a time.
1839 Once the victimized group starts meeting its latency target again it will start
1840 unthrottling any peer groups that were throttled previously. If the victimized
1841 group simply stops doing IO the global counter will unthrottle appropriately.
1843 IO Latency Interface Files
1844 ~~~~~~~~~~~~~~~~~~~~~~~~~~
1847 This takes a similar format as the other controllers.
1849 "MAJOR:MINOR target=<target time in microseconds"
1852 If the controller is enabled you will see extra stats in io.stat in
1853 addition to the normal ones.
1856 This is the current queue depth for the group.
1859 This is an exponential moving average with a decay rate of 1/exp
1860 bound by the sampling interval. The decay rate interval can be
1861 calculated by multiplying the win value in io.stat by the
1862 corresponding number of samples based on the win value.
1865 The sampling window size in milliseconds. This is the minimum
1866 duration of time between evaluation events. Windows only elapse
1867 with IO activity. Idle periods extend the most recent window.
1872 The process number controller is used to allow a cgroup to stop any
1873 new tasks from being fork()'d or clone()'d after a specified limit is
1876 The number of tasks in a cgroup can be exhausted in ways which other
1877 controllers cannot prevent, thus warranting its own controller. For
1878 example, a fork bomb is likely to exhaust the number of tasks before
1879 hitting memory restrictions.
1881 Note that PIDs used in this controller refer to TIDs, process IDs as
1889 A read-write single value file which exists on non-root
1890 cgroups. The default is "max".
1892 Hard limit of number of processes.
1895 A read-only single value file which exists on all cgroups.
1897 The number of processes currently in the cgroup and its
1900 Organisational operations are not blocked by cgroup policies, so it is
1901 possible to have pids.current > pids.max. This can be done by either
1902 setting the limit to be smaller than pids.current, or attaching enough
1903 processes to the cgroup such that pids.current is larger than
1904 pids.max. However, it is not possible to violate a cgroup PID policy
1905 through fork() or clone(). These will return -EAGAIN if the creation
1906 of a new process would cause a cgroup policy to be violated.
1912 The "cpuset" controller provides a mechanism for constraining
1913 the CPU and memory node placement of tasks to only the resources
1914 specified in the cpuset interface files in a task's current cgroup.
1915 This is especially valuable on large NUMA systems where placing jobs
1916 on properly sized subsets of the systems with careful processor and
1917 memory placement to reduce cross-node memory access and contention
1918 can improve overall system performance.
1920 The "cpuset" controller is hierarchical. That means the controller
1921 cannot use CPUs or memory nodes not allowed in its parent.
1924 Cpuset Interface Files
1925 ~~~~~~~~~~~~~~~~~~~~~~
1928 A read-write multiple values file which exists on non-root
1929 cpuset-enabled cgroups.
1931 It lists the requested CPUs to be used by tasks within this
1932 cgroup. The actual list of CPUs to be granted, however, is
1933 subjected to constraints imposed by its parent and can differ
1934 from the requested CPUs.
1936 The CPU numbers are comma-separated numbers or ranges.
1942 An empty value indicates that the cgroup is using the same
1943 setting as the nearest cgroup ancestor with a non-empty
1944 "cpuset.cpus" or all the available CPUs if none is found.
1946 The value of "cpuset.cpus" stays constant until the next update
1947 and won't be affected by any CPU hotplug events.
1949 cpuset.cpus.effective
1950 A read-only multiple values file which exists on all
1951 cpuset-enabled cgroups.
1953 It lists the onlined CPUs that are actually granted to this
1954 cgroup by its parent. These CPUs are allowed to be used by
1955 tasks within the current cgroup.
1957 If "cpuset.cpus" is empty, the "cpuset.cpus.effective" file shows
1958 all the CPUs from the parent cgroup that can be available to
1959 be used by this cgroup. Otherwise, it should be a subset of
1960 "cpuset.cpus" unless none of the CPUs listed in "cpuset.cpus"
1961 can be granted. In this case, it will be treated just like an
1962 empty "cpuset.cpus".
1964 Its value will be affected by CPU hotplug events.
1967 A read-write multiple values file which exists on non-root
1968 cpuset-enabled cgroups.
1970 It lists the requested memory nodes to be used by tasks within
1971 this cgroup. The actual list of memory nodes granted, however,
1972 is subjected to constraints imposed by its parent and can differ
1973 from the requested memory nodes.
1975 The memory node numbers are comma-separated numbers or ranges.
1981 An empty value indicates that the cgroup is using the same
1982 setting as the nearest cgroup ancestor with a non-empty
1983 "cpuset.mems" or all the available memory nodes if none
1986 The value of "cpuset.mems" stays constant until the next update
1987 and won't be affected by any memory nodes hotplug events.
1989 cpuset.mems.effective
1990 A read-only multiple values file which exists on all
1991 cpuset-enabled cgroups.
1993 It lists the onlined memory nodes that are actually granted to
1994 this cgroup by its parent. These memory nodes are allowed to
1995 be used by tasks within the current cgroup.
1997 If "cpuset.mems" is empty, it shows all the memory nodes from the
1998 parent cgroup that will be available to be used by this cgroup.
1999 Otherwise, it should be a subset of "cpuset.mems" unless none of
2000 the memory nodes listed in "cpuset.mems" can be granted. In this
2001 case, it will be treated just like an empty "cpuset.mems".
2003 Its value will be affected by memory nodes hotplug events.
2005 cpuset.cpus.partition
2006 A read-write single value file which exists on non-root
2007 cpuset-enabled cgroups. This flag is owned by the parent cgroup
2008 and is not delegatable.
2010 It accepts only the following input values when written to.
2012 ======== ================================
2013 "root" a partition root
2014 "member" a non-root member of a partition
2015 ======== ================================
2017 When set to be a partition root, the current cgroup is the
2018 root of a new partition or scheduling domain that comprises
2019 itself and all its descendants except those that are separate
2020 partition roots themselves and their descendants. The root
2021 cgroup is always a partition root.
2023 There are constraints on where a partition root can be set.
2024 It can only be set in a cgroup if all the following conditions
2027 1) The "cpuset.cpus" is not empty and the list of CPUs are
2028 exclusive, i.e. they are not shared by any of its siblings.
2029 2) The parent cgroup is a partition root.
2030 3) The "cpuset.cpus" is also a proper subset of the parent's
2031 "cpuset.cpus.effective".
2032 4) There is no child cgroups with cpuset enabled. This is for
2033 eliminating corner cases that have to be handled if such a
2034 condition is allowed.
2036 Setting it to partition root will take the CPUs away from the
2037 effective CPUs of the parent cgroup. Once it is set, this
2038 file cannot be reverted back to "member" if there are any child
2039 cgroups with cpuset enabled.
2041 A parent partition cannot distribute all its CPUs to its
2042 child partitions. There must be at least one cpu left in the
2045 Once becoming a partition root, changes to "cpuset.cpus" is
2046 generally allowed as long as the first condition above is true,
2047 the change will not take away all the CPUs from the parent
2048 partition and the new "cpuset.cpus" value is a superset of its
2049 children's "cpuset.cpus" values.
2051 Sometimes, external factors like changes to ancestors'
2052 "cpuset.cpus" or cpu hotplug can cause the state of the partition
2053 root to change. On read, the "cpuset.sched.partition" file
2054 can show the following values.
2056 ============== ==============================
2057 "member" Non-root member of a partition
2058 "root" Partition root
2059 "root invalid" Invalid partition root
2060 ============== ==============================
2062 It is a partition root if the first 2 partition root conditions
2063 above are true and at least one CPU from "cpuset.cpus" is
2064 granted by the parent cgroup.
2066 A partition root can become invalid if none of CPUs requested
2067 in "cpuset.cpus" can be granted by the parent cgroup or the
2068 parent cgroup is no longer a partition root itself. In this
2069 case, it is not a real partition even though the restriction
2070 of the first partition root condition above will still apply.
2071 The cpu affinity of all the tasks in the cgroup will then be
2072 associated with CPUs in the nearest ancestor partition.
2074 An invalid partition root can be transitioned back to a
2075 real partition root if at least one of the requested CPUs
2076 can now be granted by its parent. In this case, the cpu
2077 affinity of all the tasks in the formerly invalid partition
2078 will be associated to the CPUs of the newly formed partition.
2079 Changing the partition state of an invalid partition root to
2080 "member" is always allowed even if child cpusets are present.
2086 Device controller manages access to device files. It includes both
2087 creation of new device files (using mknod), and access to the
2088 existing device files.
2090 Cgroup v2 device controller has no interface files and is implemented
2091 on top of cgroup BPF. To control access to device files, a user may
2092 create bpf programs of the BPF_CGROUP_DEVICE type and attach them
2093 to cgroups. On an attempt to access a device file, corresponding
2094 BPF programs will be executed, and depending on the return value
2095 the attempt will succeed or fail with -EPERM.
2097 A BPF_CGROUP_DEVICE program takes a pointer to the bpf_cgroup_dev_ctx
2098 structure, which describes the device access attempt: access type
2099 (mknod/read/write) and device (type, major and minor numbers).
2100 If the program returns 0, the attempt fails with -EPERM, otherwise
2103 An example of BPF_CGROUP_DEVICE program may be found in the kernel
2104 source tree in the tools/testing/selftests/bpf/progs/dev_cgroup.c file.
2110 The "rdma" controller regulates the distribution and accounting of
2113 RDMA Interface Files
2114 ~~~~~~~~~~~~~~~~~~~~
2117 A readwrite nested-keyed file that exists for all the cgroups
2118 except root that describes current configured resource limit
2119 for a RDMA/IB device.
2121 Lines are keyed by device name and are not ordered.
2122 Each line contains space separated resource name and its configured
2123 limit that can be distributed.
2125 The following nested keys are defined.
2127 ========== =============================
2128 hca_handle Maximum number of HCA Handles
2129 hca_object Maximum number of HCA Objects
2130 ========== =============================
2132 An example for mlx4 and ocrdma device follows::
2134 mlx4_0 hca_handle=2 hca_object=2000
2135 ocrdma1 hca_handle=3 hca_object=max
2138 A read-only file that describes current resource usage.
2139 It exists for all the cgroup except root.
2141 An example for mlx4 and ocrdma device follows::
2143 mlx4_0 hca_handle=1 hca_object=20
2144 ocrdma1 hca_handle=1 hca_object=23
2149 The HugeTLB controller allows to limit the HugeTLB usage per control group and
2150 enforces the controller limit during page fault.
2152 HugeTLB Interface Files
2153 ~~~~~~~~~~~~~~~~~~~~~~~
2155 hugetlb.<hugepagesize>.current
2156 Show current usage for "hugepagesize" hugetlb. It exists for all
2157 the cgroup except root.
2159 hugetlb.<hugepagesize>.max
2160 Set/show the hard limit of "hugepagesize" hugetlb usage.
2161 The default value is "max". It exists for all the cgroup except root.
2163 hugetlb.<hugepagesize>.events
2164 A read-only flat-keyed file which exists on non-root cgroups.
2167 The number of allocation failure due to HugeTLB limit
2169 hugetlb.<hugepagesize>.events.local
2170 Similar to hugetlb.<hugepagesize>.events but the fields in the file
2171 are local to the cgroup i.e. not hierarchical. The file modified event
2172 generated on this file reflects only the local events.
2177 The Miscellaneous cgroup provides the resource limiting and tracking
2178 mechanism for the scalar resources which cannot be abstracted like the other
2179 cgroup resources. Controller is enabled by the CONFIG_CGROUP_MISC config
2182 A resource can be added to the controller via enum misc_res_type{} in the
2183 include/linux/misc_cgroup.h file and the corresponding name via misc_res_name[]
2184 in the kernel/cgroup/misc.c file. Provider of the resource must set its
2185 capacity prior to using the resource by calling misc_cg_set_capacity().
2187 Once a capacity is set then the resource usage can be updated using charge and
2188 uncharge APIs. All of the APIs to interact with misc controller are in
2189 include/linux/misc_cgroup.h.
2191 Misc Interface Files
2192 ~~~~~~~~~~~~~~~~~~~~
2194 Miscellaneous controller provides 3 interface files. If two misc resources (res_a and res_b) are registered then:
2197 A read-only flat-keyed file shown only in the root cgroup. It shows
2198 miscellaneous scalar resources available on the platform along with
2206 A read-only flat-keyed file shown in the non-root cgroups. It shows
2207 the current usage of the resources in the cgroup and its children.::
2214 A read-write flat-keyed file shown in the non root cgroups. Allowed
2215 maximum usage of the resources in the cgroup and its children.::
2221 Limit can be set by::
2223 # echo res_a 1 > misc.max
2225 Limit can be set to max by::
2227 # echo res_a max > misc.max
2229 Limits can be set higher than the capacity value in the misc.capacity
2232 Migration and Ownership
2233 ~~~~~~~~~~~~~~~~~~~~~~~
2235 A miscellaneous scalar resource is charged to the cgroup in which it is used
2236 first, and stays charged to that cgroup until that resource is freed. Migrating
2237 a process to a different cgroup does not move the charge to the destination
2238 cgroup where the process has moved.
2246 perf_event controller, if not mounted on a legacy hierarchy, is
2247 automatically enabled on the v2 hierarchy so that perf events can
2248 always be filtered by cgroup v2 path. The controller can still be
2249 moved to a legacy hierarchy after v2 hierarchy is populated.
2252 Non-normative information
2253 -------------------------
2255 This section contains information that isn't considered to be a part of
2256 the stable kernel API and so is subject to change.
2259 CPU controller root cgroup process behaviour
2260 ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
2262 When distributing CPU cycles in the root cgroup each thread in this
2263 cgroup is treated as if it was hosted in a separate child cgroup of the
2264 root cgroup. This child cgroup weight is dependent on its thread nice
2267 For details of this mapping see sched_prio_to_weight array in
2268 kernel/sched/core.c file (values from this array should be scaled
2269 appropriately so the neutral - nice 0 - value is 100 instead of 1024).
2272 IO controller root cgroup process behaviour
2273 ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
2275 Root cgroup processes are hosted in an implicit leaf child node.
2276 When distributing IO resources this implicit child node is taken into
2277 account as if it was a normal child cgroup of the root cgroup with a
2278 weight value of 200.
2287 cgroup namespace provides a mechanism to virtualize the view of the
2288 "/proc/$PID/cgroup" file and cgroup mounts. The CLONE_NEWCGROUP clone
2289 flag can be used with clone(2) and unshare(2) to create a new cgroup
2290 namespace. The process running inside the cgroup namespace will have
2291 its "/proc/$PID/cgroup" output restricted to cgroupns root. The
2292 cgroupns root is the cgroup of the process at the time of creation of
2293 the cgroup namespace.
2295 Without cgroup namespace, the "/proc/$PID/cgroup" file shows the
2296 complete path of the cgroup of a process. In a container setup where
2297 a set of cgroups and namespaces are intended to isolate processes the
2298 "/proc/$PID/cgroup" file may leak potential system level information
2299 to the isolated processes. For example::
2301 # cat /proc/self/cgroup
2302 0::/batchjobs/container_id1
2304 The path '/batchjobs/container_id1' can be considered as system-data
2305 and undesirable to expose to the isolated processes. cgroup namespace
2306 can be used to restrict visibility of this path. For example, before
2307 creating a cgroup namespace, one would see::
2309 # ls -l /proc/self/ns/cgroup
2310 lrwxrwxrwx 1 root root 0 2014-07-15 10:37 /proc/self/ns/cgroup -> cgroup:[4026531835]
2311 # cat /proc/self/cgroup
2312 0::/batchjobs/container_id1
2314 After unsharing a new namespace, the view changes::
2316 # ls -l /proc/self/ns/cgroup
2317 lrwxrwxrwx 1 root root 0 2014-07-15 10:35 /proc/self/ns/cgroup -> cgroup:[4026532183]
2318 # cat /proc/self/cgroup
2321 When some thread from a multi-threaded process unshares its cgroup
2322 namespace, the new cgroupns gets applied to the entire process (all
2323 the threads). This is natural for the v2 hierarchy; however, for the
2324 legacy hierarchies, this may be unexpected.
2326 A cgroup namespace is alive as long as there are processes inside or
2327 mounts pinning it. When the last usage goes away, the cgroup
2328 namespace is destroyed. The cgroupns root and the actual cgroups
2335 The 'cgroupns root' for a cgroup namespace is the cgroup in which the
2336 process calling unshare(2) is running. For example, if a process in
2337 /batchjobs/container_id1 cgroup calls unshare, cgroup
2338 /batchjobs/container_id1 becomes the cgroupns root. For the
2339 init_cgroup_ns, this is the real root ('/') cgroup.
2341 The cgroupns root cgroup does not change even if the namespace creator
2342 process later moves to a different cgroup::
2344 # ~/unshare -c # unshare cgroupns in some cgroup
2345 # cat /proc/self/cgroup
2348 # echo 0 > sub_cgrp_1/cgroup.procs
2349 # cat /proc/self/cgroup
2352 Each process gets its namespace-specific view of "/proc/$PID/cgroup"
2354 Processes running inside the cgroup namespace will be able to see
2355 cgroup paths (in /proc/self/cgroup) only inside their root cgroup.
2356 From within an unshared cgroupns::
2360 # echo 7353 > sub_cgrp_1/cgroup.procs
2361 # cat /proc/7353/cgroup
2364 From the initial cgroup namespace, the real cgroup path will be
2367 $ cat /proc/7353/cgroup
2368 0::/batchjobs/container_id1/sub_cgrp_1
2370 From a sibling cgroup namespace (that is, a namespace rooted at a
2371 different cgroup), the cgroup path relative to its own cgroup
2372 namespace root will be shown. For instance, if PID 7353's cgroup
2373 namespace root is at '/batchjobs/container_id2', then it will see::
2375 # cat /proc/7353/cgroup
2376 0::/../container_id2/sub_cgrp_1
2378 Note that the relative path always starts with '/' to indicate that
2379 its relative to the cgroup namespace root of the caller.
2382 Migration and setns(2)
2383 ----------------------
2385 Processes inside a cgroup namespace can move into and out of the
2386 namespace root if they have proper access to external cgroups. For
2387 example, from inside a namespace with cgroupns root at
2388 /batchjobs/container_id1, and assuming that the global hierarchy is
2389 still accessible inside cgroupns::
2391 # cat /proc/7353/cgroup
2393 # echo 7353 > batchjobs/container_id2/cgroup.procs
2394 # cat /proc/7353/cgroup
2395 0::/../container_id2
2397 Note that this kind of setup is not encouraged. A task inside cgroup
2398 namespace should only be exposed to its own cgroupns hierarchy.
2400 setns(2) to another cgroup namespace is allowed when:
2402 (a) the process has CAP_SYS_ADMIN against its current user namespace
2403 (b) the process has CAP_SYS_ADMIN against the target cgroup
2406 No implicit cgroup changes happen with attaching to another cgroup
2407 namespace. It is expected that the someone moves the attaching
2408 process under the target cgroup namespace root.
2411 Interaction with Other Namespaces
2412 ---------------------------------
2414 Namespace specific cgroup hierarchy can be mounted by a process
2415 running inside a non-init cgroup namespace::
2417 # mount -t cgroup2 none $MOUNT_POINT
2419 This will mount the unified cgroup hierarchy with cgroupns root as the
2420 filesystem root. The process needs CAP_SYS_ADMIN against its user and
2423 The virtualization of /proc/self/cgroup file combined with restricting
2424 the view of cgroup hierarchy by namespace-private cgroupfs mount
2425 provides a properly isolated cgroup view inside the container.
2428 Information on Kernel Programming
2429 =================================
2431 This section contains kernel programming information in the areas
2432 where interacting with cgroup is necessary. cgroup core and
2433 controllers are not covered.
2436 Filesystem Support for Writeback
2437 --------------------------------
2439 A filesystem can support cgroup writeback by updating
2440 address_space_operations->writepage[s]() to annotate bio's using the
2441 following two functions.
2443 wbc_init_bio(@wbc, @bio)
2444 Should be called for each bio carrying writeback data and
2445 associates the bio with the inode's owner cgroup and the
2446 corresponding request queue. This must be called after
2447 a queue (device) has been associated with the bio and
2450 wbc_account_cgroup_owner(@wbc, @page, @bytes)
2451 Should be called for each data segment being written out.
2452 While this function doesn't care exactly when it's called
2453 during the writeback session, it's the easiest and most
2454 natural to call it as data segments are added to a bio.
2456 With writeback bio's annotated, cgroup support can be enabled per
2457 super_block by setting SB_I_CGROUPWB in ->s_iflags. This allows for
2458 selective disabling of cgroup writeback support which is helpful when
2459 certain filesystem features, e.g. journaled data mode, are
2462 wbc_init_bio() binds the specified bio to its cgroup. Depending on
2463 the configuration, the bio may be executed at a lower priority and if
2464 the writeback session is holding shared resources, e.g. a journal
2465 entry, may lead to priority inversion. There is no one easy solution
2466 for the problem. Filesystems can try to work around specific problem
2467 cases by skipping wbc_init_bio() and using bio_associate_blkg()
2471 Deprecated v1 Core Features
2472 ===========================
2474 - Multiple hierarchies including named ones are not supported.
2476 - All v1 mount options are not supported.
2478 - The "tasks" file is removed and "cgroup.procs" is not sorted.
2480 - "cgroup.clone_children" is removed.
2482 - /proc/cgroups is meaningless for v2. Use "cgroup.controllers" file
2483 at the root instead.
2486 Issues with v1 and Rationales for v2
2487 ====================================
2489 Multiple Hierarchies
2490 --------------------
2492 cgroup v1 allowed an arbitrary number of hierarchies and each
2493 hierarchy could host any number of controllers. While this seemed to
2494 provide a high level of flexibility, it wasn't useful in practice.
2496 For example, as there is only one instance of each controller, utility
2497 type controllers such as freezer which can be useful in all
2498 hierarchies could only be used in one. The issue is exacerbated by
2499 the fact that controllers couldn't be moved to another hierarchy once
2500 hierarchies were populated. Another issue was that all controllers
2501 bound to a hierarchy were forced to have exactly the same view of the
2502 hierarchy. It wasn't possible to vary the granularity depending on
2503 the specific controller.
2505 In practice, these issues heavily limited which controllers could be
2506 put on the same hierarchy and most configurations resorted to putting
2507 each controller on its own hierarchy. Only closely related ones, such
2508 as the cpu and cpuacct controllers, made sense to be put on the same
2509 hierarchy. This often meant that userland ended up managing multiple
2510 similar hierarchies repeating the same steps on each hierarchy
2511 whenever a hierarchy management operation was necessary.
2513 Furthermore, support for multiple hierarchies came at a steep cost.
2514 It greatly complicated cgroup core implementation but more importantly
2515 the support for multiple hierarchies restricted how cgroup could be
2516 used in general and what controllers was able to do.
2518 There was no limit on how many hierarchies there might be, which meant
2519 that a thread's cgroup membership couldn't be described in finite
2520 length. The key might contain any number of entries and was unlimited
2521 in length, which made it highly awkward to manipulate and led to
2522 addition of controllers which existed only to identify membership,
2523 which in turn exacerbated the original problem of proliferating number
2526 Also, as a controller couldn't have any expectation regarding the
2527 topologies of hierarchies other controllers might be on, each
2528 controller had to assume that all other controllers were attached to
2529 completely orthogonal hierarchies. This made it impossible, or at
2530 least very cumbersome, for controllers to cooperate with each other.
2532 In most use cases, putting controllers on hierarchies which are
2533 completely orthogonal to each other isn't necessary. What usually is
2534 called for is the ability to have differing levels of granularity
2535 depending on the specific controller. In other words, hierarchy may
2536 be collapsed from leaf towards root when viewed from specific
2537 controllers. For example, a given configuration might not care about
2538 how memory is distributed beyond a certain level while still wanting
2539 to control how CPU cycles are distributed.
2545 cgroup v1 allowed threads of a process to belong to different cgroups.
2546 This didn't make sense for some controllers and those controllers
2547 ended up implementing different ways to ignore such situations but
2548 much more importantly it blurred the line between API exposed to
2549 individual applications and system management interface.
2551 Generally, in-process knowledge is available only to the process
2552 itself; thus, unlike service-level organization of processes,
2553 categorizing threads of a process requires active participation from
2554 the application which owns the target process.
2556 cgroup v1 had an ambiguously defined delegation model which got abused
2557 in combination with thread granularity. cgroups were delegated to
2558 individual applications so that they can create and manage their own
2559 sub-hierarchies and control resource distributions along them. This
2560 effectively raised cgroup to the status of a syscall-like API exposed
2563 First of all, cgroup has a fundamentally inadequate interface to be
2564 exposed this way. For a process to access its own knobs, it has to
2565 extract the path on the target hierarchy from /proc/self/cgroup,
2566 construct the path by appending the name of the knob to the path, open
2567 and then read and/or write to it. This is not only extremely clunky
2568 and unusual but also inherently racy. There is no conventional way to
2569 define transaction across the required steps and nothing can guarantee
2570 that the process would actually be operating on its own sub-hierarchy.
2572 cgroup controllers implemented a number of knobs which would never be
2573 accepted as public APIs because they were just adding control knobs to
2574 system-management pseudo filesystem. cgroup ended up with interface
2575 knobs which were not properly abstracted or refined and directly
2576 revealed kernel internal details. These knobs got exposed to
2577 individual applications through the ill-defined delegation mechanism
2578 effectively abusing cgroup as a shortcut to implementing public APIs
2579 without going through the required scrutiny.
2581 This was painful for both userland and kernel. Userland ended up with
2582 misbehaving and poorly abstracted interfaces and kernel exposing and
2583 locked into constructs inadvertently.
2586 Competition Between Inner Nodes and Threads
2587 -------------------------------------------
2589 cgroup v1 allowed threads to be in any cgroups which created an
2590 interesting problem where threads belonging to a parent cgroup and its
2591 children cgroups competed for resources. This was nasty as two
2592 different types of entities competed and there was no obvious way to
2593 settle it. Different controllers did different things.
2595 The cpu controller considered threads and cgroups as equivalents and
2596 mapped nice levels to cgroup weights. This worked for some cases but
2597 fell flat when children wanted to be allocated specific ratios of CPU
2598 cycles and the number of internal threads fluctuated - the ratios
2599 constantly changed as the number of competing entities fluctuated.
2600 There also were other issues. The mapping from nice level to weight
2601 wasn't obvious or universal, and there were various other knobs which
2602 simply weren't available for threads.
2604 The io controller implicitly created a hidden leaf node for each
2605 cgroup to host the threads. The hidden leaf had its own copies of all
2606 the knobs with ``leaf_`` prefixed. While this allowed equivalent
2607 control over internal threads, it was with serious drawbacks. It
2608 always added an extra layer of nesting which wouldn't be necessary
2609 otherwise, made the interface messy and significantly complicated the
2612 The memory controller didn't have a way to control what happened
2613 between internal tasks and child cgroups and the behavior was not
2614 clearly defined. There were attempts to add ad-hoc behaviors and
2615 knobs to tailor the behavior to specific workloads which would have
2616 led to problems extremely difficult to resolve in the long term.
2618 Multiple controllers struggled with internal tasks and came up with
2619 different ways to deal with it; unfortunately, all the approaches were
2620 severely flawed and, furthermore, the widely different behaviors
2621 made cgroup as a whole highly inconsistent.
2623 This clearly is a problem which needs to be addressed from cgroup core
2627 Other Interface Issues
2628 ----------------------
2630 cgroup v1 grew without oversight and developed a large number of
2631 idiosyncrasies and inconsistencies. One issue on the cgroup core side
2632 was how an empty cgroup was notified - a userland helper binary was
2633 forked and executed for each event. The event delivery wasn't
2634 recursive or delegatable. The limitations of the mechanism also led
2635 to in-kernel event delivery filtering mechanism further complicating
2638 Controller interfaces were problematic too. An extreme example is
2639 controllers completely ignoring hierarchical organization and treating
2640 all cgroups as if they were all located directly under the root
2641 cgroup. Some controllers exposed a large amount of inconsistent
2642 implementation details to userland.
2644 There also was no consistency across controllers. When a new cgroup
2645 was created, some controllers defaulted to not imposing extra
2646 restrictions while others disallowed any resource usage until
2647 explicitly configured. Configuration knobs for the same type of
2648 control used widely differing naming schemes and formats. Statistics
2649 and information knobs were named arbitrarily and used different
2650 formats and units even in the same controller.
2652 cgroup v2 establishes common conventions where appropriate and updates
2653 controllers so that they expose minimal and consistent interfaces.
2656 Controller Issues and Remedies
2657 ------------------------------
2662 The original lower boundary, the soft limit, is defined as a limit
2663 that is per default unset. As a result, the set of cgroups that
2664 global reclaim prefers is opt-in, rather than opt-out. The costs for
2665 optimizing these mostly negative lookups are so high that the
2666 implementation, despite its enormous size, does not even provide the
2667 basic desirable behavior. First off, the soft limit has no
2668 hierarchical meaning. All configured groups are organized in a global
2669 rbtree and treated like equal peers, regardless where they are located
2670 in the hierarchy. This makes subtree delegation impossible. Second,
2671 the soft limit reclaim pass is so aggressive that it not just
2672 introduces high allocation latencies into the system, but also impacts
2673 system performance due to overreclaim, to the point where the feature
2674 becomes self-defeating.
2676 The memory.low boundary on the other hand is a top-down allocated
2677 reserve. A cgroup enjoys reclaim protection when it's within its
2678 effective low, which makes delegation of subtrees possible. It also
2679 enjoys having reclaim pressure proportional to its overage when
2680 above its effective low.
2682 The original high boundary, the hard limit, is defined as a strict
2683 limit that can not budge, even if the OOM killer has to be called.
2684 But this generally goes against the goal of making the most out of the
2685 available memory. The memory consumption of workloads varies during
2686 runtime, and that requires users to overcommit. But doing that with a
2687 strict upper limit requires either a fairly accurate prediction of the
2688 working set size or adding slack to the limit. Since working set size
2689 estimation is hard and error prone, and getting it wrong results in
2690 OOM kills, most users tend to err on the side of a looser limit and
2691 end up wasting precious resources.
2693 The memory.high boundary on the other hand can be set much more
2694 conservatively. When hit, it throttles allocations by forcing them
2695 into direct reclaim to work off the excess, but it never invokes the
2696 OOM killer. As a result, a high boundary that is chosen too
2697 aggressively will not terminate the processes, but instead it will
2698 lead to gradual performance degradation. The user can monitor this
2699 and make corrections until the minimal memory footprint that still
2700 gives acceptable performance is found.
2702 In extreme cases, with many concurrent allocations and a complete
2703 breakdown of reclaim progress within the group, the high boundary can
2704 be exceeded. But even then it's mostly better to satisfy the
2705 allocation from the slack available in other groups or the rest of the
2706 system than killing the group. Otherwise, memory.max is there to
2707 limit this type of spillover and ultimately contain buggy or even
2708 malicious applications.
2710 Setting the original memory.limit_in_bytes below the current usage was
2711 subject to a race condition, where concurrent charges could cause the
2712 limit setting to fail. memory.max on the other hand will first set the
2713 limit to prevent new charges, and then reclaim and OOM kill until the
2714 new limit is met - or the task writing to memory.max is killed.
2716 The combined memory+swap accounting and limiting is replaced by real
2717 control over swap space.
2719 The main argument for a combined memory+swap facility in the original
2720 cgroup design was that global or parental pressure would always be
2721 able to swap all anonymous memory of a child group, regardless of the
2722 child's own (possibly untrusted) configuration. However, untrusted
2723 groups can sabotage swapping by other means - such as referencing its
2724 anonymous memory in a tight loop - and an admin can not assume full
2725 swappability when overcommitting untrusted jobs.
2727 For trusted jobs, on the other hand, a combined counter is not an
2728 intuitive userspace interface, and it flies in the face of the idea
2729 that cgroup controllers should account and limit specific physical
2730 resources. Swap space is a resource like all others in the system,
2731 and that's why unified hierarchy allows distributing it separately.