6 :Author: Tejun Heo <tj@kernel.org>
8 This is the authoritative documentation on the design, interface and
9 conventions of cgroup v2. It describes all userland-visible aspects
10 of cgroup including core and specific controller behaviors. All
11 future changes must be reflected in this document. Documentation for
12 v1 is available under :ref:`Documentation/admin-guide/cgroup-v1/index.rst <cgroup-v1>`.
21 2-2. Organizing Processes and Threads
24 2-3. [Un]populated Notification
25 2-4. Controlling Controllers
26 2-4-1. Enabling and Disabling
27 2-4-2. Top-down Constraint
28 2-4-3. No Internal Process Constraint
30 2-5-1. Model of Delegation
31 2-5-2. Delegation Containment
33 2-6-1. Organize Once and Control
34 2-6-2. Avoid Name Collisions
35 3. Resource Distribution Models
43 4-3. Core Interface Files
46 5-1-1. CPU Interface Files
48 5-2-1. Memory Interface Files
49 5-2-2. Usage Guidelines
50 5-2-3. Memory Ownership
52 5-3-1. IO Interface Files
55 5-3-3-1. How IO Latency Throttling Works
56 5-3-3-2. IO Latency Interface Files
58 5-4-1. PID Interface Files
60 5.5-1. Cpuset Interface Files
63 5-7-1. RDMA Interface Files
65 5.8-1. HugeTLB Interface Files
68 5-N. Non-normative information
69 5-N-1. CPU controller root cgroup process behaviour
70 5-N-2. IO controller root cgroup process behaviour
73 6-2. The Root and Views
74 6-3. Migration and setns(2)
75 6-4. Interaction with Other Namespaces
76 P. Information on Kernel Programming
77 P-1. Filesystem Support for Writeback
78 D. Deprecated v1 Core Features
79 R. Issues with v1 and Rationales for v2
80 R-1. Multiple Hierarchies
81 R-2. Thread Granularity
82 R-3. Competition Between Inner Nodes and Threads
83 R-4. Other Interface Issues
84 R-5. Controller Issues and Remedies
94 "cgroup" stands for "control group" and is never capitalized. The
95 singular form is used to designate the whole feature and also as a
96 qualifier as in "cgroup controllers". When explicitly referring to
97 multiple individual control groups, the plural form "cgroups" is used.
103 cgroup is a mechanism to organize processes hierarchically and
104 distribute system resources along the hierarchy in a controlled and
107 cgroup is largely composed of two parts - the core and controllers.
108 cgroup core is primarily responsible for hierarchically organizing
109 processes. A cgroup controller is usually responsible for
110 distributing a specific type of system resource along the hierarchy
111 although there are utility controllers which serve purposes other than
112 resource distribution.
114 cgroups form a tree structure and every process in the system belongs
115 to one and only one cgroup. All threads of a process belong to the
116 same cgroup. On creation, all processes are put in the cgroup that
117 the parent process belongs to at the time. A process can be migrated
118 to another cgroup. Migration of a process doesn't affect already
119 existing descendant processes.
121 Following certain structural constraints, controllers may be enabled or
122 disabled selectively on a cgroup. All controller behaviors are
123 hierarchical - if a controller is enabled on a cgroup, it affects all
124 processes which belong to the cgroups consisting the inclusive
125 sub-hierarchy of the cgroup. When a controller is enabled on a nested
126 cgroup, it always restricts the resource distribution further. The
127 restrictions set closer to the root in the hierarchy can not be
128 overridden from further away.
137 Unlike v1, cgroup v2 has only single hierarchy. The cgroup v2
138 hierarchy can be mounted with the following mount command::
140 # mount -t cgroup2 none $MOUNT_POINT
142 cgroup2 filesystem has the magic number 0x63677270 ("cgrp"). All
143 controllers which support v2 and are not bound to a v1 hierarchy are
144 automatically bound to the v2 hierarchy and show up at the root.
145 Controllers which are not in active use in the v2 hierarchy can be
146 bound to other hierarchies. This allows mixing v2 hierarchy with the
147 legacy v1 multiple hierarchies in a fully backward compatible way.
149 A controller can be moved across hierarchies only after the controller
150 is no longer referenced in its current hierarchy. Because per-cgroup
151 controller states are destroyed asynchronously and controllers may
152 have lingering references, a controller may not show up immediately on
153 the v2 hierarchy after the final umount of the previous hierarchy.
154 Similarly, a controller should be fully disabled to be moved out of
155 the unified hierarchy and it may take some time for the disabled
156 controller to become available for other hierarchies; furthermore, due
157 to inter-controller dependencies, other controllers may need to be
160 While useful for development and manual configurations, moving
161 controllers dynamically between the v2 and other hierarchies is
162 strongly discouraged for production use. It is recommended to decide
163 the hierarchies and controller associations before starting using the
164 controllers after system boot.
166 During transition to v2, system management software might still
167 automount the v1 cgroup filesystem and so hijack all controllers
168 during boot, before manual intervention is possible. To make testing
169 and experimenting easier, the kernel parameter cgroup_no_v1= allows
170 disabling controllers in v1 and make them always available in v2.
172 cgroup v2 currently supports the following mount options.
176 Consider cgroup namespaces as delegation boundaries. This
177 option is system wide and can only be set on mount or modified
178 through remount from the init namespace. The mount option is
179 ignored on non-init namespace mounts. Please refer to the
180 Delegation section for details.
184 Only populate memory.events with data for the current cgroup,
185 and not any subtrees. This is legacy behaviour, the default
186 behaviour without this option is to include subtree counts.
187 This option is system wide and can only be set on mount or
188 modified through remount from the init namespace. The mount
189 option is ignored on non-init namespace mounts.
193 Recursively apply memory.min and memory.low protection to
194 entire subtrees, without requiring explicit downward
195 propagation into leaf cgroups. This allows protecting entire
196 subtrees from one another, while retaining free competition
197 within those subtrees. This should have been the default
198 behavior but is a mount-option to avoid regressing setups
199 relying on the original semantics (e.g. specifying bogusly
200 high 'bypass' protection values at higher tree levels).
203 Organizing Processes and Threads
204 --------------------------------
209 Initially, only the root cgroup exists to which all processes belong.
210 A child cgroup can be created by creating a sub-directory::
214 A given cgroup may have multiple child cgroups forming a tree
215 structure. Each cgroup has a read-writable interface file
216 "cgroup.procs". When read, it lists the PIDs of all processes which
217 belong to the cgroup one-per-line. The PIDs are not ordered and the
218 same PID may show up more than once if the process got moved to
219 another cgroup and then back or the PID got recycled while reading.
221 A process can be migrated into a cgroup by writing its PID to the
222 target cgroup's "cgroup.procs" file. Only one process can be migrated
223 on a single write(2) call. If a process is composed of multiple
224 threads, writing the PID of any thread migrates all threads of the
227 When a process forks a child process, the new process is born into the
228 cgroup that the forking process belongs to at the time of the
229 operation. After exit, a process stays associated with the cgroup
230 that it belonged to at the time of exit until it's reaped; however, a
231 zombie process does not appear in "cgroup.procs" and thus can't be
232 moved to another cgroup.
234 A cgroup which doesn't have any children or live processes can be
235 destroyed by removing the directory. Note that a cgroup which doesn't
236 have any children and is associated only with zombie processes is
237 considered empty and can be removed::
241 "/proc/$PID/cgroup" lists a process's cgroup membership. If legacy
242 cgroup is in use in the system, this file may contain multiple lines,
243 one for each hierarchy. The entry for cgroup v2 is always in the
246 # cat /proc/842/cgroup
248 0::/test-cgroup/test-cgroup-nested
250 If the process becomes a zombie and the cgroup it was associated with
251 is removed subsequently, " (deleted)" is appended to the path::
253 # cat /proc/842/cgroup
255 0::/test-cgroup/test-cgroup-nested (deleted)
261 cgroup v2 supports thread granularity for a subset of controllers to
262 support use cases requiring hierarchical resource distribution across
263 the threads of a group of processes. By default, all threads of a
264 process belong to the same cgroup, which also serves as the resource
265 domain to host resource consumptions which are not specific to a
266 process or thread. The thread mode allows threads to be spread across
267 a subtree while still maintaining the common resource domain for them.
269 Controllers which support thread mode are called threaded controllers.
270 The ones which don't are called domain controllers.
272 Marking a cgroup threaded makes it join the resource domain of its
273 parent as a threaded cgroup. The parent may be another threaded
274 cgroup whose resource domain is further up in the hierarchy. The root
275 of a threaded subtree, that is, the nearest ancestor which is not
276 threaded, is called threaded domain or thread root interchangeably and
277 serves as the resource domain for the entire subtree.
279 Inside a threaded subtree, threads of a process can be put in
280 different cgroups and are not subject to the no internal process
281 constraint - threaded controllers can be enabled on non-leaf cgroups
282 whether they have threads in them or not.
284 As the threaded domain cgroup hosts all the domain resource
285 consumptions of the subtree, it is considered to have internal
286 resource consumptions whether there are processes in it or not and
287 can't have populated child cgroups which aren't threaded. Because the
288 root cgroup is not subject to no internal process constraint, it can
289 serve both as a threaded domain and a parent to domain cgroups.
291 The current operation mode or type of the cgroup is shown in the
292 "cgroup.type" file which indicates whether the cgroup is a normal
293 domain, a domain which is serving as the domain of a threaded subtree,
294 or a threaded cgroup.
296 On creation, a cgroup is always a domain cgroup and can be made
297 threaded by writing "threaded" to the "cgroup.type" file. The
298 operation is single direction::
300 # echo threaded > cgroup.type
302 Once threaded, the cgroup can't be made a domain again. To enable the
303 thread mode, the following conditions must be met.
305 - As the cgroup will join the parent's resource domain. The parent
306 must either be a valid (threaded) domain or a threaded cgroup.
308 - When the parent is an unthreaded domain, it must not have any domain
309 controllers enabled or populated domain children. The root is
310 exempt from this requirement.
312 Topology-wise, a cgroup can be in an invalid state. Please consider
313 the following topology::
315 A (threaded domain) - B (threaded) - C (domain, just created)
317 C is created as a domain but isn't connected to a parent which can
318 host child domains. C can't be used until it is turned into a
319 threaded cgroup. "cgroup.type" file will report "domain (invalid)" in
320 these cases. Operations which fail due to invalid topology use
321 EOPNOTSUPP as the errno.
323 A domain cgroup is turned into a threaded domain when one of its child
324 cgroup becomes threaded or threaded controllers are enabled in the
325 "cgroup.subtree_control" file while there are processes in the cgroup.
326 A threaded domain reverts to a normal domain when the conditions
329 When read, "cgroup.threads" contains the list of the thread IDs of all
330 threads in the cgroup. Except that the operations are per-thread
331 instead of per-process, "cgroup.threads" has the same format and
332 behaves the same way as "cgroup.procs". While "cgroup.threads" can be
333 written to in any cgroup, as it can only move threads inside the same
334 threaded domain, its operations are confined inside each threaded
337 The threaded domain cgroup serves as the resource domain for the whole
338 subtree, and, while the threads can be scattered across the subtree,
339 all the processes are considered to be in the threaded domain cgroup.
340 "cgroup.procs" in a threaded domain cgroup contains the PIDs of all
341 processes in the subtree and is not readable in the subtree proper.
342 However, "cgroup.procs" can be written to from anywhere in the subtree
343 to migrate all threads of the matching process to the cgroup.
345 Only threaded controllers can be enabled in a threaded subtree. When
346 a threaded controller is enabled inside a threaded subtree, it only
347 accounts for and controls resource consumptions associated with the
348 threads in the cgroup and its descendants. All consumptions which
349 aren't tied to a specific thread belong to the threaded domain cgroup.
351 Because a threaded subtree is exempt from no internal process
352 constraint, a threaded controller must be able to handle competition
353 between threads in a non-leaf cgroup and its child cgroups. Each
354 threaded controller defines how such competitions are handled.
357 [Un]populated Notification
358 --------------------------
360 Each non-root cgroup has a "cgroup.events" file which contains
361 "populated" field indicating whether the cgroup's sub-hierarchy has
362 live processes in it. Its value is 0 if there is no live process in
363 the cgroup and its descendants; otherwise, 1. poll and [id]notify
364 events are triggered when the value changes. This can be used, for
365 example, to start a clean-up operation after all processes of a given
366 sub-hierarchy have exited. The populated state updates and
367 notifications are recursive. Consider the following sub-hierarchy
368 where the numbers in the parentheses represent the numbers of processes
374 A, B and C's "populated" fields would be 1 while D's 0. After the one
375 process in C exits, B and C's "populated" fields would flip to "0" and
376 file modified events will be generated on the "cgroup.events" files of
380 Controlling Controllers
381 -----------------------
383 Enabling and Disabling
384 ~~~~~~~~~~~~~~~~~~~~~~
386 Each cgroup has a "cgroup.controllers" file which lists all
387 controllers available for the cgroup to enable::
389 # cat cgroup.controllers
392 No controller is enabled by default. Controllers can be enabled and
393 disabled by writing to the "cgroup.subtree_control" file::
395 # echo "+cpu +memory -io" > cgroup.subtree_control
397 Only controllers which are listed in "cgroup.controllers" can be
398 enabled. When multiple operations are specified as above, either they
399 all succeed or fail. If multiple operations on the same controller
400 are specified, the last one is effective.
402 Enabling a controller in a cgroup indicates that the distribution of
403 the target resource across its immediate children will be controlled.
404 Consider the following sub-hierarchy. The enabled controllers are
405 listed in parentheses::
407 A(cpu,memory) - B(memory) - C()
410 As A has "cpu" and "memory" enabled, A will control the distribution
411 of CPU cycles and memory to its children, in this case, B. As B has
412 "memory" enabled but not "CPU", C and D will compete freely on CPU
413 cycles but their division of memory available to B will be controlled.
415 As a controller regulates the distribution of the target resource to
416 the cgroup's children, enabling it creates the controller's interface
417 files in the child cgroups. In the above example, enabling "cpu" on B
418 would create the "cpu." prefixed controller interface files in C and
419 D. Likewise, disabling "memory" from B would remove the "memory."
420 prefixed controller interface files from C and D. This means that the
421 controller interface files - anything which doesn't start with
422 "cgroup." are owned by the parent rather than the cgroup itself.
428 Resources are distributed top-down and a cgroup can further distribute
429 a resource only if the resource has been distributed to it from the
430 parent. This means that all non-root "cgroup.subtree_control" files
431 can only contain controllers which are enabled in the parent's
432 "cgroup.subtree_control" file. A controller can be enabled only if
433 the parent has the controller enabled and a controller can't be
434 disabled if one or more children have it enabled.
437 No Internal Process Constraint
438 ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
440 Non-root cgroups can distribute domain resources to their children
441 only when they don't have any processes of their own. In other words,
442 only domain cgroups which don't contain any processes can have domain
443 controllers enabled in their "cgroup.subtree_control" files.
445 This guarantees that, when a domain controller is looking at the part
446 of the hierarchy which has it enabled, processes are always only on
447 the leaves. This rules out situations where child cgroups compete
448 against internal processes of the parent.
450 The root cgroup is exempt from this restriction. Root contains
451 processes and anonymous resource consumption which can't be associated
452 with any other cgroups and requires special treatment from most
453 controllers. How resource consumption in the root cgroup is governed
454 is up to each controller (for more information on this topic please
455 refer to the Non-normative information section in the Controllers
458 Note that the restriction doesn't get in the way if there is no
459 enabled controller in the cgroup's "cgroup.subtree_control". This is
460 important as otherwise it wouldn't be possible to create children of a
461 populated cgroup. To control resource distribution of a cgroup, the
462 cgroup must create children and transfer all its processes to the
463 children before enabling controllers in its "cgroup.subtree_control"
473 A cgroup can be delegated in two ways. First, to a less privileged
474 user by granting write access of the directory and its "cgroup.procs",
475 "cgroup.threads" and "cgroup.subtree_control" files to the user.
476 Second, if the "nsdelegate" mount option is set, automatically to a
477 cgroup namespace on namespace creation.
479 Because the resource control interface files in a given directory
480 control the distribution of the parent's resources, the delegatee
481 shouldn't be allowed to write to them. For the first method, this is
482 achieved by not granting access to these files. For the second, the
483 kernel rejects writes to all files other than "cgroup.procs" and
484 "cgroup.subtree_control" on a namespace root from inside the
487 The end results are equivalent for both delegation types. Once
488 delegated, the user can build sub-hierarchy under the directory,
489 organize processes inside it as it sees fit and further distribute the
490 resources it received from the parent. The limits and other settings
491 of all resource controllers are hierarchical and regardless of what
492 happens in the delegated sub-hierarchy, nothing can escape the
493 resource restrictions imposed by the parent.
495 Currently, cgroup doesn't impose any restrictions on the number of
496 cgroups in or nesting depth of a delegated sub-hierarchy; however,
497 this may be limited explicitly in the future.
500 Delegation Containment
501 ~~~~~~~~~~~~~~~~~~~~~~
503 A delegated sub-hierarchy is contained in the sense that processes
504 can't be moved into or out of the sub-hierarchy by the delegatee.
506 For delegations to a less privileged user, this is achieved by
507 requiring the following conditions for a process with a non-root euid
508 to migrate a target process into a cgroup by writing its PID to the
511 - The writer must have write access to the "cgroup.procs" file.
513 - The writer must have write access to the "cgroup.procs" file of the
514 common ancestor of the source and destination cgroups.
516 The above two constraints ensure that while a delegatee may migrate
517 processes around freely in the delegated sub-hierarchy it can't pull
518 in from or push out to outside the sub-hierarchy.
520 For an example, let's assume cgroups C0 and C1 have been delegated to
521 user U0 who created C00, C01 under C0 and C10 under C1 as follows and
522 all processes under C0 and C1 belong to U0::
524 ~~~~~~~~~~~~~ - C0 - C00
527 ~~~~~~~~~~~~~ - C1 - C10
529 Let's also say U0 wants to write the PID of a process which is
530 currently in C10 into "C00/cgroup.procs". U0 has write access to the
531 file; however, the common ancestor of the source cgroup C10 and the
532 destination cgroup C00 is above the points of delegation and U0 would
533 not have write access to its "cgroup.procs" files and thus the write
534 will be denied with -EACCES.
536 For delegations to namespaces, containment is achieved by requiring
537 that both the source and destination cgroups are reachable from the
538 namespace of the process which is attempting the migration. If either
539 is not reachable, the migration is rejected with -ENOENT.
545 Organize Once and Control
546 ~~~~~~~~~~~~~~~~~~~~~~~~~
548 Migrating a process across cgroups is a relatively expensive operation
549 and stateful resources such as memory are not moved together with the
550 process. This is an explicit design decision as there often exist
551 inherent trade-offs between migration and various hot paths in terms
552 of synchronization cost.
554 As such, migrating processes across cgroups frequently as a means to
555 apply different resource restrictions is discouraged. A workload
556 should be assigned to a cgroup according to the system's logical and
557 resource structure once on start-up. Dynamic adjustments to resource
558 distribution can be made by changing controller configuration through
562 Avoid Name Collisions
563 ~~~~~~~~~~~~~~~~~~~~~
565 Interface files for a cgroup and its children cgroups occupy the same
566 directory and it is possible to create children cgroups which collide
567 with interface files.
569 All cgroup core interface files are prefixed with "cgroup." and each
570 controller's interface files are prefixed with the controller name and
571 a dot. A controller's name is composed of lower case alphabets and
572 '_'s but never begins with an '_' so it can be used as the prefix
573 character for collision avoidance. Also, interface file names won't
574 start or end with terms which are often used in categorizing workloads
575 such as job, service, slice, unit or workload.
577 cgroup doesn't do anything to prevent name collisions and it's the
578 user's responsibility to avoid them.
581 Resource Distribution Models
582 ============================
584 cgroup controllers implement several resource distribution schemes
585 depending on the resource type and expected use cases. This section
586 describes major schemes in use along with their expected behaviors.
592 A parent's resource is distributed by adding up the weights of all
593 active children and giving each the fraction matching the ratio of its
594 weight against the sum. As only children which can make use of the
595 resource at the moment participate in the distribution, this is
596 work-conserving. Due to the dynamic nature, this model is usually
597 used for stateless resources.
599 All weights are in the range [1, 10000] with the default at 100. This
600 allows symmetric multiplicative biases in both directions at fine
601 enough granularity while staying in the intuitive range.
603 As long as the weight is in range, all configuration combinations are
604 valid and there is no reason to reject configuration changes or
607 "cpu.weight" proportionally distributes CPU cycles to active children
608 and is an example of this type.
614 A child can only consume upto the configured amount of the resource.
615 Limits can be over-committed - the sum of the limits of children can
616 exceed the amount of resource available to the parent.
618 Limits are in the range [0, max] and defaults to "max", which is noop.
620 As limits can be over-committed, all configuration combinations are
621 valid and there is no reason to reject configuration changes or
624 "io.max" limits the maximum BPS and/or IOPS that a cgroup can consume
625 on an IO device and is an example of this type.
631 A cgroup is protected upto the configured amount of the resource
632 as long as the usages of all its ancestors are under their
633 protected levels. Protections can be hard guarantees or best effort
634 soft boundaries. Protections can also be over-committed in which case
635 only upto the amount available to the parent is protected among
638 Protections are in the range [0, max] and defaults to 0, which is
641 As protections can be over-committed, all configuration combinations
642 are valid and there is no reason to reject configuration changes or
645 "memory.low" implements best-effort memory protection and is an
646 example of this type.
652 A cgroup is exclusively allocated a certain amount of a finite
653 resource. Allocations can't be over-committed - the sum of the
654 allocations of children can not exceed the amount of resource
655 available to the parent.
657 Allocations are in the range [0, max] and defaults to 0, which is no
660 As allocations can't be over-committed, some configuration
661 combinations are invalid and should be rejected. Also, if the
662 resource is mandatory for execution of processes, process migrations
665 "cpu.rt.max" hard-allocates realtime slices and is an example of this
675 All interface files should be in one of the following formats whenever
678 New-line separated values
679 (when only one value can be written at once)
685 Space separated values
686 (when read-only or multiple values can be written at once)
698 KEY0 SUB_KEY0=VAL00 SUB_KEY1=VAL01...
699 KEY1 SUB_KEY0=VAL10 SUB_KEY1=VAL11...
702 For a writable file, the format for writing should generally match
703 reading; however, controllers may allow omitting later fields or
704 implement restricted shortcuts for most common use cases.
706 For both flat and nested keyed files, only the values for a single key
707 can be written at a time. For nested keyed files, the sub key pairs
708 may be specified in any order and not all pairs have to be specified.
714 - Settings for a single feature should be contained in a single file.
716 - The root cgroup should be exempt from resource control and thus
717 shouldn't have resource control interface files.
719 - The default time unit is microseconds. If a different unit is ever
720 used, an explicit unit suffix must be present.
722 - A parts-per quantity should use a percentage decimal with at least
723 two digit fractional part - e.g. 13.40.
725 - If a controller implements weight based resource distribution, its
726 interface file should be named "weight" and have the range [1,
727 10000] with 100 as the default. The values are chosen to allow
728 enough and symmetric bias in both directions while keeping it
729 intuitive (the default is 100%).
731 - If a controller implements an absolute resource guarantee and/or
732 limit, the interface files should be named "min" and "max"
733 respectively. If a controller implements best effort resource
734 guarantee and/or limit, the interface files should be named "low"
735 and "high" respectively.
737 In the above four control files, the special token "max" should be
738 used to represent upward infinity for both reading and writing.
740 - If a setting has a configurable default value and keyed specific
741 overrides, the default entry should be keyed with "default" and
742 appear as the first entry in the file.
744 The default value can be updated by writing either "default $VAL" or
747 When writing to update a specific override, "default" can be used as
748 the value to indicate removal of the override. Override entries
749 with "default" as the value must not appear when read.
751 For example, a setting which is keyed by major:minor device numbers
752 with integer values may look like the following::
754 # cat cgroup-example-interface-file
758 The default value can be updated by::
760 # echo 125 > cgroup-example-interface-file
764 # echo "default 125" > cgroup-example-interface-file
766 An override can be set by::
768 # echo "8:16 170" > cgroup-example-interface-file
772 # echo "8:0 default" > cgroup-example-interface-file
773 # cat cgroup-example-interface-file
777 - For events which are not very high frequency, an interface file
778 "events" should be created which lists event key value pairs.
779 Whenever a notifiable event happens, file modified event should be
780 generated on the file.
786 All cgroup core files are prefixed with "cgroup."
790 A read-write single value file which exists on non-root
793 When read, it indicates the current type of the cgroup, which
794 can be one of the following values.
796 - "domain" : A normal valid domain cgroup.
798 - "domain threaded" : A threaded domain cgroup which is
799 serving as the root of a threaded subtree.
801 - "domain invalid" : A cgroup which is in an invalid state.
802 It can't be populated or have controllers enabled. It may
803 be allowed to become a threaded cgroup.
805 - "threaded" : A threaded cgroup which is a member of a
808 A cgroup can be turned into a threaded cgroup by writing
809 "threaded" to this file.
812 A read-write new-line separated values file which exists on
815 When read, it lists the PIDs of all processes which belong to
816 the cgroup one-per-line. The PIDs are not ordered and the
817 same PID may show up more than once if the process got moved
818 to another cgroup and then back or the PID got recycled while
821 A PID can be written to migrate the process associated with
822 the PID to the cgroup. The writer should match all of the
823 following conditions.
825 - It must have write access to the "cgroup.procs" file.
827 - It must have write access to the "cgroup.procs" file of the
828 common ancestor of the source and destination cgroups.
830 When delegating a sub-hierarchy, write access to this file
831 should be granted along with the containing directory.
833 In a threaded cgroup, reading this file fails with EOPNOTSUPP
834 as all the processes belong to the thread root. Writing is
835 supported and moves every thread of the process to the cgroup.
838 A read-write new-line separated values file which exists on
841 When read, it lists the TIDs of all threads which belong to
842 the cgroup one-per-line. The TIDs are not ordered and the
843 same TID may show up more than once if the thread got moved to
844 another cgroup and then back or the TID got recycled while
847 A TID can be written to migrate the thread associated with the
848 TID to the cgroup. The writer should match all of the
849 following conditions.
851 - It must have write access to the "cgroup.threads" file.
853 - The cgroup that the thread is currently in must be in the
854 same resource domain as the destination cgroup.
856 - It must have write access to the "cgroup.procs" file of the
857 common ancestor of the source and destination cgroups.
859 When delegating a sub-hierarchy, write access to this file
860 should be granted along with the containing directory.
863 A read-only space separated values file which exists on all
866 It shows space separated list of all controllers available to
867 the cgroup. The controllers are not ordered.
869 cgroup.subtree_control
870 A read-write space separated values file which exists on all
871 cgroups. Starts out empty.
873 When read, it shows space separated list of the controllers
874 which are enabled to control resource distribution from the
875 cgroup to its children.
877 Space separated list of controllers prefixed with '+' or '-'
878 can be written to enable or disable controllers. A controller
879 name prefixed with '+' enables the controller and '-'
880 disables. If a controller appears more than once on the list,
881 the last one is effective. When multiple enable and disable
882 operations are specified, either all succeed or all fail.
885 A read-only flat-keyed file which exists on non-root cgroups.
886 The following entries are defined. Unless specified
887 otherwise, a value change in this file generates a file
891 1 if the cgroup or its descendants contains any live
892 processes; otherwise, 0.
894 1 if the cgroup is frozen; otherwise, 0.
896 cgroup.max.descendants
897 A read-write single value files. The default is "max".
899 Maximum allowed number of descent cgroups.
900 If the actual number of descendants is equal or larger,
901 an attempt to create a new cgroup in the hierarchy will fail.
904 A read-write single value files. The default is "max".
906 Maximum allowed descent depth below the current cgroup.
907 If the actual descent depth is equal or larger,
908 an attempt to create a new child cgroup will fail.
911 A read-only flat-keyed file with the following entries:
914 Total number of visible descendant cgroups.
917 Total number of dying descendant cgroups. A cgroup becomes
918 dying after being deleted by a user. The cgroup will remain
919 in dying state for some time undefined time (which can depend
920 on system load) before being completely destroyed.
922 A process can't enter a dying cgroup under any circumstances,
923 a dying cgroup can't revive.
925 A dying cgroup can consume system resources not exceeding
926 limits, which were active at the moment of cgroup deletion.
929 A read-write single value file which exists on non-root cgroups.
930 Allowed values are "0" and "1". The default is "0".
932 Writing "1" to the file causes freezing of the cgroup and all
933 descendant cgroups. This means that all belonging processes will
934 be stopped and will not run until the cgroup will be explicitly
935 unfrozen. Freezing of the cgroup may take some time; when this action
936 is completed, the "frozen" value in the cgroup.events control file
937 will be updated to "1" and the corresponding notification will be
940 A cgroup can be frozen either by its own settings, or by settings
941 of any ancestor cgroups. If any of ancestor cgroups is frozen, the
942 cgroup will remain frozen.
944 Processes in the frozen cgroup can be killed by a fatal signal.
945 They also can enter and leave a frozen cgroup: either by an explicit
946 move by a user, or if freezing of the cgroup races with fork().
947 If a process is moved to a frozen cgroup, it stops. If a process is
948 moved out of a frozen cgroup, it becomes running.
950 Frozen status of a cgroup doesn't affect any cgroup tree operations:
951 it's possible to delete a frozen (and empty) cgroup, as well as
952 create new sub-cgroups.
960 The "cpu" controllers regulates distribution of CPU cycles. This
961 controller implements weight and absolute bandwidth limit models for
962 normal scheduling policy and absolute bandwidth allocation model for
963 realtime scheduling policy.
965 In all the above models, cycles distribution is defined only on a temporal
966 base and it does not account for the frequency at which tasks are executed.
967 The (optional) utilization clamping support allows to hint the schedutil
968 cpufreq governor about the minimum desired frequency which should always be
969 provided by a CPU, as well as the maximum desired frequency, which should not
970 be exceeded by a CPU.
972 WARNING: cgroup2 doesn't yet support control of realtime processes and
973 the cpu controller can only be enabled when all RT processes are in
974 the root cgroup. Be aware that system management software may already
975 have placed RT processes into nonroot cgroups during the system boot
976 process, and these processes may need to be moved to the root cgroup
977 before the cpu controller can be enabled.
983 All time durations are in microseconds.
986 A read-only flat-keyed file.
987 This file exists whether the controller is enabled or not.
989 It always reports the following three stats:
995 and the following three when the controller is enabled:
1002 A read-write single value file which exists on non-root
1003 cgroups. The default is "100".
1005 The weight in the range [1, 10000].
1008 A read-write single value file which exists on non-root
1009 cgroups. The default is "0".
1011 The nice value is in the range [-20, 19].
1013 This interface file is an alternative interface for
1014 "cpu.weight" and allows reading and setting weight using the
1015 same values used by nice(2). Because the range is smaller and
1016 granularity is coarser for the nice values, the read value is
1017 the closest approximation of the current weight.
1020 A read-write two value file which exists on non-root cgroups.
1021 The default is "max 100000".
1023 The maximum bandwidth limit. It's in the following format::
1027 which indicates that the group may consume upto $MAX in each
1028 $PERIOD duration. "max" for $MAX indicates no limit. If only
1029 one number is written, $MAX is updated.
1032 A read-only nested-key file which exists on non-root cgroups.
1034 Shows pressure stall information for CPU. See
1035 :ref:`Documentation/accounting/psi.rst <psi>` for details.
1038 A read-write single value file which exists on non-root cgroups.
1039 The default is "0", i.e. no utilization boosting.
1041 The requested minimum utilization (protection) as a percentage
1042 rational number, e.g. 12.34 for 12.34%.
1044 This interface allows reading and setting minimum utilization clamp
1045 values similar to the sched_setattr(2). This minimum utilization
1046 value is used to clamp the task specific minimum utilization clamp.
1048 The requested minimum utilization (protection) is always capped by
1049 the current value for the maximum utilization (limit), i.e.
1053 A read-write single value file which exists on non-root cgroups.
1054 The default is "max". i.e. no utilization capping
1056 The requested maximum utilization (limit) as a percentage rational
1057 number, e.g. 98.76 for 98.76%.
1059 This interface allows reading and setting maximum utilization clamp
1060 values similar to the sched_setattr(2). This maximum utilization
1061 value is used to clamp the task specific maximum utilization clamp.
1068 The "memory" controller regulates distribution of memory. Memory is
1069 stateful and implements both limit and protection models. Due to the
1070 intertwining between memory usage and reclaim pressure and the
1071 stateful nature of memory, the distribution model is relatively
1074 While not completely water-tight, all major memory usages by a given
1075 cgroup are tracked so that the total memory consumption can be
1076 accounted and controlled to a reasonable extent. Currently, the
1077 following types of memory usages are tracked.
1079 - Userland memory - page cache and anonymous memory.
1081 - Kernel data structures such as dentries and inodes.
1083 - TCP socket buffers.
1085 The above list may expand in the future for better coverage.
1088 Memory Interface Files
1089 ~~~~~~~~~~~~~~~~~~~~~~
1091 All memory amounts are in bytes. If a value which is not aligned to
1092 PAGE_SIZE is written, the value may be rounded up to the closest
1093 PAGE_SIZE multiple when read back.
1096 A read-only single value file which exists on non-root
1099 The total amount of memory currently being used by the cgroup
1100 and its descendants.
1103 A read-write single value file which exists on non-root
1104 cgroups. The default is "0".
1106 Hard memory protection. If the memory usage of a cgroup
1107 is within its effective min boundary, the cgroup's memory
1108 won't be reclaimed under any conditions. If there is no
1109 unprotected reclaimable memory available, OOM killer
1110 is invoked. Above the effective min boundary (or
1111 effective low boundary if it is higher), pages are reclaimed
1112 proportionally to the overage, reducing reclaim pressure for
1115 Effective min boundary is limited by memory.min values of
1116 all ancestor cgroups. If there is memory.min overcommitment
1117 (child cgroup or cgroups are requiring more protected memory
1118 than parent will allow), then each child cgroup will get
1119 the part of parent's protection proportional to its
1120 actual memory usage below memory.min.
1122 Putting more memory than generally available under this
1123 protection is discouraged and may lead to constant OOMs.
1125 If a memory cgroup is not populated with processes,
1126 its memory.min is ignored.
1129 A read-write single value file which exists on non-root
1130 cgroups. The default is "0".
1132 Best-effort memory protection. If the memory usage of a
1133 cgroup is within its effective low boundary, the cgroup's
1134 memory won't be reclaimed unless there is no reclaimable
1135 memory available in unprotected cgroups.
1136 Above the effective low boundary (or
1137 effective min boundary if it is higher), pages are reclaimed
1138 proportionally to the overage, reducing reclaim pressure for
1141 Effective low boundary is limited by memory.low values of
1142 all ancestor cgroups. If there is memory.low overcommitment
1143 (child cgroup or cgroups are requiring more protected memory
1144 than parent will allow), then each child cgroup will get
1145 the part of parent's protection proportional to its
1146 actual memory usage below memory.low.
1148 Putting more memory than generally available under this
1149 protection is discouraged.
1152 A read-write single value file which exists on non-root
1153 cgroups. The default is "max".
1155 Memory usage throttle limit. This is the main mechanism to
1156 control memory usage of a cgroup. If a cgroup's usage goes
1157 over the high boundary, the processes of the cgroup are
1158 throttled and put under heavy reclaim pressure.
1160 Going over the high limit never invokes the OOM killer and
1161 under extreme conditions the limit may be breached.
1164 A read-write single value file which exists on non-root
1165 cgroups. The default is "max".
1167 Memory usage hard limit. This is the final protection
1168 mechanism. If a cgroup's memory usage reaches this limit and
1169 can't be reduced, the OOM killer is invoked in the cgroup.
1170 Under certain circumstances, the usage may go over the limit
1173 In default configuration regular 0-order allocations always
1174 succeed unless OOM killer chooses current task as a victim.
1176 Some kinds of allocations don't invoke the OOM killer.
1177 Caller could retry them differently, return into userspace
1178 as -ENOMEM or silently ignore in cases like disk readahead.
1180 This is the ultimate protection mechanism. As long as the
1181 high limit is used and monitored properly, this limit's
1182 utility is limited to providing the final safety net.
1185 A read-write single value file which exists on non-root
1186 cgroups. The default value is "0".
1188 Determines whether the cgroup should be treated as
1189 an indivisible workload by the OOM killer. If set,
1190 all tasks belonging to the cgroup or to its descendants
1191 (if the memory cgroup is not a leaf cgroup) are killed
1192 together or not at all. This can be used to avoid
1193 partial kills to guarantee workload integrity.
1195 Tasks with the OOM protection (oom_score_adj set to -1000)
1196 are treated as an exception and are never killed.
1198 If the OOM killer is invoked in a cgroup, it's not going
1199 to kill any tasks outside of this cgroup, regardless
1200 memory.oom.group values of ancestor cgroups.
1203 A read-only flat-keyed file which exists on non-root cgroups.
1204 The following entries are defined. Unless specified
1205 otherwise, a value change in this file generates a file
1208 Note that all fields in this file are hierarchical and the
1209 file modified event can be generated due to an event down the
1210 hierarchy. For for the local events at the cgroup level see
1211 memory.events.local.
1214 The number of times the cgroup is reclaimed due to
1215 high memory pressure even though its usage is under
1216 the low boundary. This usually indicates that the low
1217 boundary is over-committed.
1220 The number of times processes of the cgroup are
1221 throttled and routed to perform direct memory reclaim
1222 because the high memory boundary was exceeded. For a
1223 cgroup whose memory usage is capped by the high limit
1224 rather than global memory pressure, this event's
1225 occurrences are expected.
1228 The number of times the cgroup's memory usage was
1229 about to go over the max boundary. If direct reclaim
1230 fails to bring it down, the cgroup goes to OOM state.
1233 The number of time the cgroup's memory usage was
1234 reached the limit and allocation was about to fail.
1236 This event is not raised if the OOM killer is not
1237 considered as an option, e.g. for failed high-order
1238 allocations or if caller asked to not retry attempts.
1241 The number of processes belonging to this cgroup
1242 killed by any kind of OOM killer.
1245 Similar to memory.events but the fields in the file are local
1246 to the cgroup i.e. not hierarchical. The file modified event
1247 generated on this file reflects only the local events.
1250 A read-only flat-keyed file which exists on non-root cgroups.
1252 This breaks down the cgroup's memory footprint into different
1253 types of memory, type-specific details, and other information
1254 on the state and past events of the memory management system.
1256 All memory amounts are in bytes.
1258 The entries are ordered to be human readable, and new entries
1259 can show up in the middle. Don't rely on items remaining in a
1260 fixed position; use the keys to look up specific values!
1263 Amount of memory used in anonymous mappings such as
1264 brk(), sbrk(), and mmap(MAP_ANONYMOUS)
1267 Amount of memory used to cache filesystem data,
1268 including tmpfs and shared memory.
1271 Amount of memory allocated to kernel stacks.
1274 Amount of memory used for storing in-kernel data
1278 Amount of memory used in network transmission buffers
1281 Amount of cached filesystem data that is swap-backed,
1282 such as tmpfs, shm segments, shared anonymous mmap()s
1285 Amount of cached filesystem data mapped with mmap()
1288 Amount of cached filesystem data that was modified but
1289 not yet written back to disk
1292 Amount of cached filesystem data that was modified and
1293 is currently being written back to disk
1296 Amount of memory used in anonymous mappings backed by
1297 transparent hugepages
1299 inactive_anon, active_anon, inactive_file, active_file, unevictable
1300 Amount of memory, swap-backed and filesystem-backed,
1301 on the internal memory management lists used by the
1302 page reclaim algorithm.
1304 As these represent internal list state (eg. shmem pages are on anon
1305 memory management lists), inactive_foo + active_foo may not be equal to
1306 the value for the foo counter, since the foo counter is type-based, not
1310 Part of "slab" that might be reclaimed, such as
1311 dentries and inodes.
1314 Part of "slab" that cannot be reclaimed on memory
1318 Total number of page faults incurred
1321 Number of major page faults incurred
1324 Number of refaults of previously evicted pages
1327 Number of refaulted pages that were immediately activated
1330 Number of restored pages which have been detected as an active
1331 workingset before they got reclaimed.
1333 workingset_nodereclaim
1334 Number of times a shadow node has been reclaimed
1337 Amount of scanned pages (in an active LRU list)
1340 Amount of scanned pages (in an inactive LRU list)
1343 Amount of reclaimed pages
1346 Amount of pages moved to the active LRU list
1349 Amount of pages moved to the inactive LRU list
1352 Amount of pages postponed to be freed under memory pressure
1355 Amount of reclaimed lazyfree pages
1358 Number of transparent hugepages which were allocated to satisfy
1359 a page fault. This counter is not present when CONFIG_TRANSPARENT_HUGEPAGE
1363 Number of transparent hugepages which were allocated to allow
1364 collapsing an existing range of pages. This counter is not
1365 present when CONFIG_TRANSPARENT_HUGEPAGE is not set.
1368 A read-only single value file which exists on non-root
1371 The total amount of swap currently being used by the cgroup
1372 and its descendants.
1375 A read-write single value file which exists on non-root
1376 cgroups. The default is "max".
1378 Swap usage throttle limit. If a cgroup's swap usage exceeds
1379 this limit, all its further allocations will be throttled to
1380 allow userspace to implement custom out-of-memory procedures.
1382 This limit marks a point of no return for the cgroup. It is NOT
1383 designed to manage the amount of swapping a workload does
1384 during regular operation. Compare to memory.swap.max, which
1385 prohibits swapping past a set amount, but lets the cgroup
1386 continue unimpeded as long as other memory can be reclaimed.
1388 Healthy workloads are not expected to reach this limit.
1391 A read-write single value file which exists on non-root
1392 cgroups. The default is "max".
1394 Swap usage hard limit. If a cgroup's swap usage reaches this
1395 limit, anonymous memory of the cgroup will not be swapped out.
1398 A read-only flat-keyed file which exists on non-root cgroups.
1399 The following entries are defined. Unless specified
1400 otherwise, a value change in this file generates a file
1404 The number of times the cgroup's swap usage was over
1408 The number of times the cgroup's swap usage was about
1409 to go over the max boundary and swap allocation
1413 The number of times swap allocation failed either
1414 because of running out of swap system-wide or max
1417 When reduced under the current usage, the existing swap
1418 entries are reclaimed gradually and the swap usage may stay
1419 higher than the limit for an extended period of time. This
1420 reduces the impact on the workload and memory management.
1423 A read-only nested-key file which exists on non-root cgroups.
1425 Shows pressure stall information for memory. See
1426 :ref:`Documentation/accounting/psi.rst <psi>` for details.
1432 "memory.high" is the main mechanism to control memory usage.
1433 Over-committing on high limit (sum of high limits > available memory)
1434 and letting global memory pressure to distribute memory according to
1435 usage is a viable strategy.
1437 Because breach of the high limit doesn't trigger the OOM killer but
1438 throttles the offending cgroup, a management agent has ample
1439 opportunities to monitor and take appropriate actions such as granting
1440 more memory or terminating the workload.
1442 Determining whether a cgroup has enough memory is not trivial as
1443 memory usage doesn't indicate whether the workload can benefit from
1444 more memory. For example, a workload which writes data received from
1445 network to a file can use all available memory but can also operate as
1446 performant with a small amount of memory. A measure of memory
1447 pressure - how much the workload is being impacted due to lack of
1448 memory - is necessary to determine whether a workload needs more
1449 memory; unfortunately, memory pressure monitoring mechanism isn't
1456 A memory area is charged to the cgroup which instantiated it and stays
1457 charged to the cgroup until the area is released. Migrating a process
1458 to a different cgroup doesn't move the memory usages that it
1459 instantiated while in the previous cgroup to the new cgroup.
1461 A memory area may be used by processes belonging to different cgroups.
1462 To which cgroup the area will be charged is in-deterministic; however,
1463 over time, the memory area is likely to end up in a cgroup which has
1464 enough memory allowance to avoid high reclaim pressure.
1466 If a cgroup sweeps a considerable amount of memory which is expected
1467 to be accessed repeatedly by other cgroups, it may make sense to use
1468 POSIX_FADV_DONTNEED to relinquish the ownership of memory areas
1469 belonging to the affected files to ensure correct memory ownership.
1475 The "io" controller regulates the distribution of IO resources. This
1476 controller implements both weight based and absolute bandwidth or IOPS
1477 limit distribution; however, weight based distribution is available
1478 only if cfq-iosched is in use and neither scheme is available for
1486 A read-only nested-keyed file.
1488 Lines are keyed by $MAJ:$MIN device numbers and not ordered.
1489 The following nested keys are defined.
1491 ====== =====================
1493 wbytes Bytes written
1494 rios Number of read IOs
1495 wios Number of write IOs
1496 dbytes Bytes discarded
1497 dios Number of discard IOs
1498 ====== =====================
1500 An example read output follows::
1502 8:16 rbytes=1459200 wbytes=314773504 rios=192 wios=353 dbytes=0 dios=0
1503 8:0 rbytes=90430464 wbytes=299008000 rios=8950 wios=1252 dbytes=50331648 dios=3021
1506 A read-write nested-keyed file with exists only on the root
1509 This file configures the Quality of Service of the IO cost
1510 model based controller (CONFIG_BLK_CGROUP_IOCOST) which
1511 currently implements "io.weight" proportional control. Lines
1512 are keyed by $MAJ:$MIN device numbers and not ordered. The
1513 line for a given device is populated on the first write for
1514 the device on "io.cost.qos" or "io.cost.model". The following
1515 nested keys are defined.
1517 ====== =====================================
1518 enable Weight-based control enable
1519 ctrl "auto" or "user"
1520 rpct Read latency percentile [0, 100]
1521 rlat Read latency threshold
1522 wpct Write latency percentile [0, 100]
1523 wlat Write latency threshold
1524 min Minimum scaling percentage [1, 10000]
1525 max Maximum scaling percentage [1, 10000]
1526 ====== =====================================
1528 The controller is disabled by default and can be enabled by
1529 setting "enable" to 1. "rpct" and "wpct" parameters default
1530 to zero and the controller uses internal device saturation
1531 state to adjust the overall IO rate between "min" and "max".
1533 When a better control quality is needed, latency QoS
1534 parameters can be configured. For example::
1536 8:16 enable=1 ctrl=auto rpct=95.00 rlat=75000 wpct=95.00 wlat=150000 min=50.00 max=150.0
1538 shows that on sdb, the controller is enabled, will consider
1539 the device saturated if the 95th percentile of read completion
1540 latencies is above 75ms or write 150ms, and adjust the overall
1541 IO issue rate between 50% and 150% accordingly.
1543 The lower the saturation point, the better the latency QoS at
1544 the cost of aggregate bandwidth. The narrower the allowed
1545 adjustment range between "min" and "max", the more conformant
1546 to the cost model the IO behavior. Note that the IO issue
1547 base rate may be far off from 100% and setting "min" and "max"
1548 blindly can lead to a significant loss of device capacity or
1549 control quality. "min" and "max" are useful for regulating
1550 devices which show wide temporary behavior changes - e.g. a
1551 ssd which accepts writes at the line speed for a while and
1552 then completely stalls for multiple seconds.
1554 When "ctrl" is "auto", the parameters are controlled by the
1555 kernel and may change automatically. Setting "ctrl" to "user"
1556 or setting any of the percentile and latency parameters puts
1557 it into "user" mode and disables the automatic changes. The
1558 automatic mode can be restored by setting "ctrl" to "auto".
1561 A read-write nested-keyed file with exists only on the root
1564 This file configures the cost model of the IO cost model based
1565 controller (CONFIG_BLK_CGROUP_IOCOST) which currently
1566 implements "io.weight" proportional control. Lines are keyed
1567 by $MAJ:$MIN device numbers and not ordered. The line for a
1568 given device is populated on the first write for the device on
1569 "io.cost.qos" or "io.cost.model". The following nested keys
1572 ===== ================================
1573 ctrl "auto" or "user"
1574 model The cost model in use - "linear"
1575 ===== ================================
1577 When "ctrl" is "auto", the kernel may change all parameters
1578 dynamically. When "ctrl" is set to "user" or any other
1579 parameters are written to, "ctrl" become "user" and the
1580 automatic changes are disabled.
1582 When "model" is "linear", the following model parameters are
1585 ============= ========================================
1586 [r|w]bps The maximum sequential IO throughput
1587 [r|w]seqiops The maximum 4k sequential IOs per second
1588 [r|w]randiops The maximum 4k random IOs per second
1589 ============= ========================================
1591 From the above, the builtin linear model determines the base
1592 costs of a sequential and random IO and the cost coefficient
1593 for the IO size. While simple, this model can cover most
1594 common device classes acceptably.
1596 The IO cost model isn't expected to be accurate in absolute
1597 sense and is scaled to the device behavior dynamically.
1599 If needed, tools/cgroup/iocost_coef_gen.py can be used to
1600 generate device-specific coefficients.
1603 A read-write flat-keyed file which exists on non-root cgroups.
1604 The default is "default 100".
1606 The first line is the default weight applied to devices
1607 without specific override. The rest are overrides keyed by
1608 $MAJ:$MIN device numbers and not ordered. The weights are in
1609 the range [1, 10000] and specifies the relative amount IO time
1610 the cgroup can use in relation to its siblings.
1612 The default weight can be updated by writing either "default
1613 $WEIGHT" or simply "$WEIGHT". Overrides can be set by writing
1614 "$MAJ:$MIN $WEIGHT" and unset by writing "$MAJ:$MIN default".
1616 An example read output follows::
1623 A read-write nested-keyed file which exists on non-root
1626 BPS and IOPS based IO limit. Lines are keyed by $MAJ:$MIN
1627 device numbers and not ordered. The following nested keys are
1630 ===== ==================================
1631 rbps Max read bytes per second
1632 wbps Max write bytes per second
1633 riops Max read IO operations per second
1634 wiops Max write IO operations per second
1635 ===== ==================================
1637 When writing, any number of nested key-value pairs can be
1638 specified in any order. "max" can be specified as the value
1639 to remove a specific limit. If the same key is specified
1640 multiple times, the outcome is undefined.
1642 BPS and IOPS are measured in each IO direction and IOs are
1643 delayed if limit is reached. Temporary bursts are allowed.
1645 Setting read limit at 2M BPS and write at 120 IOPS for 8:16::
1647 echo "8:16 rbps=2097152 wiops=120" > io.max
1649 Reading returns the following::
1651 8:16 rbps=2097152 wbps=max riops=max wiops=120
1653 Write IOPS limit can be removed by writing the following::
1655 echo "8:16 wiops=max" > io.max
1657 Reading now returns the following::
1659 8:16 rbps=2097152 wbps=max riops=max wiops=max
1662 A read-only nested-key file which exists on non-root cgroups.
1664 Shows pressure stall information for IO. See
1665 :ref:`Documentation/accounting/psi.rst <psi>` for details.
1671 Page cache is dirtied through buffered writes and shared mmaps and
1672 written asynchronously to the backing filesystem by the writeback
1673 mechanism. Writeback sits between the memory and IO domains and
1674 regulates the proportion of dirty memory by balancing dirtying and
1677 The io controller, in conjunction with the memory controller,
1678 implements control of page cache writeback IOs. The memory controller
1679 defines the memory domain that dirty memory ratio is calculated and
1680 maintained for and the io controller defines the io domain which
1681 writes out dirty pages for the memory domain. Both system-wide and
1682 per-cgroup dirty memory states are examined and the more restrictive
1683 of the two is enforced.
1685 cgroup writeback requires explicit support from the underlying
1686 filesystem. Currently, cgroup writeback is implemented on ext2, ext4,
1687 btrfs, f2fs, and xfs. On other filesystems, all writeback IOs are
1688 attributed to the root cgroup.
1690 There are inherent differences in memory and writeback management
1691 which affects how cgroup ownership is tracked. Memory is tracked per
1692 page while writeback per inode. For the purpose of writeback, an
1693 inode is assigned to a cgroup and all IO requests to write dirty pages
1694 from the inode are attributed to that cgroup.
1696 As cgroup ownership for memory is tracked per page, there can be pages
1697 which are associated with different cgroups than the one the inode is
1698 associated with. These are called foreign pages. The writeback
1699 constantly keeps track of foreign pages and, if a particular foreign
1700 cgroup becomes the majority over a certain period of time, switches
1701 the ownership of the inode to that cgroup.
1703 While this model is enough for most use cases where a given inode is
1704 mostly dirtied by a single cgroup even when the main writing cgroup
1705 changes over time, use cases where multiple cgroups write to a single
1706 inode simultaneously are not supported well. In such circumstances, a
1707 significant portion of IOs are likely to be attributed incorrectly.
1708 As memory controller assigns page ownership on the first use and
1709 doesn't update it until the page is released, even if writeback
1710 strictly follows page ownership, multiple cgroups dirtying overlapping
1711 areas wouldn't work as expected. It's recommended to avoid such usage
1714 The sysctl knobs which affect writeback behavior are applied to cgroup
1715 writeback as follows.
1717 vm.dirty_background_ratio, vm.dirty_ratio
1718 These ratios apply the same to cgroup writeback with the
1719 amount of available memory capped by limits imposed by the
1720 memory controller and system-wide clean memory.
1722 vm.dirty_background_bytes, vm.dirty_bytes
1723 For cgroup writeback, this is calculated into ratio against
1724 total available memory and applied the same way as
1725 vm.dirty[_background]_ratio.
1731 This is a cgroup v2 controller for IO workload protection. You provide a group
1732 with a latency target, and if the average latency exceeds that target the
1733 controller will throttle any peers that have a lower latency target than the
1736 The limits are only applied at the peer level in the hierarchy. This means that
1737 in the diagram below, only groups A, B, and C will influence each other, and
1738 groups D and F will influence each other. Group G will influence nobody::
1747 So the ideal way to configure this is to set io.latency in groups A, B, and C.
1748 Generally you do not want to set a value lower than the latency your device
1749 supports. Experiment to find the value that works best for your workload.
1750 Start at higher than the expected latency for your device and watch the
1751 avg_lat value in io.stat for your workload group to get an idea of the
1752 latency you see during normal operation. Use the avg_lat value as a basis for
1753 your real setting, setting at 10-15% higher than the value in io.stat.
1755 How IO Latency Throttling Works
1756 ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
1758 io.latency is work conserving; so as long as everybody is meeting their latency
1759 target the controller doesn't do anything. Once a group starts missing its
1760 target it begins throttling any peer group that has a higher target than itself.
1761 This throttling takes 2 forms:
1763 - Queue depth throttling. This is the number of outstanding IO's a group is
1764 allowed to have. We will clamp down relatively quickly, starting at no limit
1765 and going all the way down to 1 IO at a time.
1767 - Artificial delay induction. There are certain types of IO that cannot be
1768 throttled without possibly adversely affecting higher priority groups. This
1769 includes swapping and metadata IO. These types of IO are allowed to occur
1770 normally, however they are "charged" to the originating group. If the
1771 originating group is being throttled you will see the use_delay and delay
1772 fields in io.stat increase. The delay value is how many microseconds that are
1773 being added to any process that runs in this group. Because this number can
1774 grow quite large if there is a lot of swapping or metadata IO occurring we
1775 limit the individual delay events to 1 second at a time.
1777 Once the victimized group starts meeting its latency target again it will start
1778 unthrottling any peer groups that were throttled previously. If the victimized
1779 group simply stops doing IO the global counter will unthrottle appropriately.
1781 IO Latency Interface Files
1782 ~~~~~~~~~~~~~~~~~~~~~~~~~~
1785 This takes a similar format as the other controllers.
1787 "MAJOR:MINOR target=<target time in microseconds"
1790 If the controller is enabled you will see extra stats in io.stat in
1791 addition to the normal ones.
1794 This is the current queue depth for the group.
1797 This is an exponential moving average with a decay rate of 1/exp
1798 bound by the sampling interval. The decay rate interval can be
1799 calculated by multiplying the win value in io.stat by the
1800 corresponding number of samples based on the win value.
1803 The sampling window size in milliseconds. This is the minimum
1804 duration of time between evaluation events. Windows only elapse
1805 with IO activity. Idle periods extend the most recent window.
1810 The process number controller is used to allow a cgroup to stop any
1811 new tasks from being fork()'d or clone()'d after a specified limit is
1814 The number of tasks in a cgroup can be exhausted in ways which other
1815 controllers cannot prevent, thus warranting its own controller. For
1816 example, a fork bomb is likely to exhaust the number of tasks before
1817 hitting memory restrictions.
1819 Note that PIDs used in this controller refer to TIDs, process IDs as
1827 A read-write single value file which exists on non-root
1828 cgroups. The default is "max".
1830 Hard limit of number of processes.
1833 A read-only single value file which exists on all cgroups.
1835 The number of processes currently in the cgroup and its
1838 Organisational operations are not blocked by cgroup policies, so it is
1839 possible to have pids.current > pids.max. This can be done by either
1840 setting the limit to be smaller than pids.current, or attaching enough
1841 processes to the cgroup such that pids.current is larger than
1842 pids.max. However, it is not possible to violate a cgroup PID policy
1843 through fork() or clone(). These will return -EAGAIN if the creation
1844 of a new process would cause a cgroup policy to be violated.
1850 The "cpuset" controller provides a mechanism for constraining
1851 the CPU and memory node placement of tasks to only the resources
1852 specified in the cpuset interface files in a task's current cgroup.
1853 This is especially valuable on large NUMA systems where placing jobs
1854 on properly sized subsets of the systems with careful processor and
1855 memory placement to reduce cross-node memory access and contention
1856 can improve overall system performance.
1858 The "cpuset" controller is hierarchical. That means the controller
1859 cannot use CPUs or memory nodes not allowed in its parent.
1862 Cpuset Interface Files
1863 ~~~~~~~~~~~~~~~~~~~~~~
1866 A read-write multiple values file which exists on non-root
1867 cpuset-enabled cgroups.
1869 It lists the requested CPUs to be used by tasks within this
1870 cgroup. The actual list of CPUs to be granted, however, is
1871 subjected to constraints imposed by its parent and can differ
1872 from the requested CPUs.
1874 The CPU numbers are comma-separated numbers or ranges.
1880 An empty value indicates that the cgroup is using the same
1881 setting as the nearest cgroup ancestor with a non-empty
1882 "cpuset.cpus" or all the available CPUs if none is found.
1884 The value of "cpuset.cpus" stays constant until the next update
1885 and won't be affected by any CPU hotplug events.
1887 cpuset.cpus.effective
1888 A read-only multiple values file which exists on all
1889 cpuset-enabled cgroups.
1891 It lists the onlined CPUs that are actually granted to this
1892 cgroup by its parent. These CPUs are allowed to be used by
1893 tasks within the current cgroup.
1895 If "cpuset.cpus" is empty, the "cpuset.cpus.effective" file shows
1896 all the CPUs from the parent cgroup that can be available to
1897 be used by this cgroup. Otherwise, it should be a subset of
1898 "cpuset.cpus" unless none of the CPUs listed in "cpuset.cpus"
1899 can be granted. In this case, it will be treated just like an
1900 empty "cpuset.cpus".
1902 Its value will be affected by CPU hotplug events.
1905 A read-write multiple values file which exists on non-root
1906 cpuset-enabled cgroups.
1908 It lists the requested memory nodes to be used by tasks within
1909 this cgroup. The actual list of memory nodes granted, however,
1910 is subjected to constraints imposed by its parent and can differ
1911 from the requested memory nodes.
1913 The memory node numbers are comma-separated numbers or ranges.
1919 An empty value indicates that the cgroup is using the same
1920 setting as the nearest cgroup ancestor with a non-empty
1921 "cpuset.mems" or all the available memory nodes if none
1924 The value of "cpuset.mems" stays constant until the next update
1925 and won't be affected by any memory nodes hotplug events.
1927 cpuset.mems.effective
1928 A read-only multiple values file which exists on all
1929 cpuset-enabled cgroups.
1931 It lists the onlined memory nodes that are actually granted to
1932 this cgroup by its parent. These memory nodes are allowed to
1933 be used by tasks within the current cgroup.
1935 If "cpuset.mems" is empty, it shows all the memory nodes from the
1936 parent cgroup that will be available to be used by this cgroup.
1937 Otherwise, it should be a subset of "cpuset.mems" unless none of
1938 the memory nodes listed in "cpuset.mems" can be granted. In this
1939 case, it will be treated just like an empty "cpuset.mems".
1941 Its value will be affected by memory nodes hotplug events.
1943 cpuset.cpus.partition
1944 A read-write single value file which exists on non-root
1945 cpuset-enabled cgroups. This flag is owned by the parent cgroup
1946 and is not delegatable.
1948 It accepts only the following input values when written to.
1950 "root" - a partition root
1951 "member" - a non-root member of a partition
1953 When set to be a partition root, the current cgroup is the
1954 root of a new partition or scheduling domain that comprises
1955 itself and all its descendants except those that are separate
1956 partition roots themselves and their descendants. The root
1957 cgroup is always a partition root.
1959 There are constraints on where a partition root can be set.
1960 It can only be set in a cgroup if all the following conditions
1963 1) The "cpuset.cpus" is not empty and the list of CPUs are
1964 exclusive, i.e. they are not shared by any of its siblings.
1965 2) The parent cgroup is a partition root.
1966 3) The "cpuset.cpus" is also a proper subset of the parent's
1967 "cpuset.cpus.effective".
1968 4) There is no child cgroups with cpuset enabled. This is for
1969 eliminating corner cases that have to be handled if such a
1970 condition is allowed.
1972 Setting it to partition root will take the CPUs away from the
1973 effective CPUs of the parent cgroup. Once it is set, this
1974 file cannot be reverted back to "member" if there are any child
1975 cgroups with cpuset enabled.
1977 A parent partition cannot distribute all its CPUs to its
1978 child partitions. There must be at least one cpu left in the
1981 Once becoming a partition root, changes to "cpuset.cpus" is
1982 generally allowed as long as the first condition above is true,
1983 the change will not take away all the CPUs from the parent
1984 partition and the new "cpuset.cpus" value is a superset of its
1985 children's "cpuset.cpus" values.
1987 Sometimes, external factors like changes to ancestors'
1988 "cpuset.cpus" or cpu hotplug can cause the state of the partition
1989 root to change. On read, the "cpuset.sched.partition" file
1990 can show the following values.
1992 "member" Non-root member of a partition
1993 "root" Partition root
1994 "root invalid" Invalid partition root
1996 It is a partition root if the first 2 partition root conditions
1997 above are true and at least one CPU from "cpuset.cpus" is
1998 granted by the parent cgroup.
2000 A partition root can become invalid if none of CPUs requested
2001 in "cpuset.cpus" can be granted by the parent cgroup or the
2002 parent cgroup is no longer a partition root itself. In this
2003 case, it is not a real partition even though the restriction
2004 of the first partition root condition above will still apply.
2005 The cpu affinity of all the tasks in the cgroup will then be
2006 associated with CPUs in the nearest ancestor partition.
2008 An invalid partition root can be transitioned back to a
2009 real partition root if at least one of the requested CPUs
2010 can now be granted by its parent. In this case, the cpu
2011 affinity of all the tasks in the formerly invalid partition
2012 will be associated to the CPUs of the newly formed partition.
2013 Changing the partition state of an invalid partition root to
2014 "member" is always allowed even if child cpusets are present.
2020 Device controller manages access to device files. It includes both
2021 creation of new device files (using mknod), and access to the
2022 existing device files.
2024 Cgroup v2 device controller has no interface files and is implemented
2025 on top of cgroup BPF. To control access to device files, a user may
2026 create bpf programs of the BPF_CGROUP_DEVICE type and attach them
2027 to cgroups. On an attempt to access a device file, corresponding
2028 BPF programs will be executed, and depending on the return value
2029 the attempt will succeed or fail with -EPERM.
2031 A BPF_CGROUP_DEVICE program takes a pointer to the bpf_cgroup_dev_ctx
2032 structure, which describes the device access attempt: access type
2033 (mknod/read/write) and device (type, major and minor numbers).
2034 If the program returns 0, the attempt fails with -EPERM, otherwise
2037 An example of BPF_CGROUP_DEVICE program may be found in the kernel
2038 source tree in the tools/testing/selftests/bpf/dev_cgroup.c file.
2044 The "rdma" controller regulates the distribution and accounting of
2047 RDMA Interface Files
2048 ~~~~~~~~~~~~~~~~~~~~
2051 A readwrite nested-keyed file that exists for all the cgroups
2052 except root that describes current configured resource limit
2053 for a RDMA/IB device.
2055 Lines are keyed by device name and are not ordered.
2056 Each line contains space separated resource name and its configured
2057 limit that can be distributed.
2059 The following nested keys are defined.
2061 ========== =============================
2062 hca_handle Maximum number of HCA Handles
2063 hca_object Maximum number of HCA Objects
2064 ========== =============================
2066 An example for mlx4 and ocrdma device follows::
2068 mlx4_0 hca_handle=2 hca_object=2000
2069 ocrdma1 hca_handle=3 hca_object=max
2072 A read-only file that describes current resource usage.
2073 It exists for all the cgroup except root.
2075 An example for mlx4 and ocrdma device follows::
2077 mlx4_0 hca_handle=1 hca_object=20
2078 ocrdma1 hca_handle=1 hca_object=23
2083 The HugeTLB controller allows to limit the HugeTLB usage per control group and
2084 enforces the controller limit during page fault.
2086 HugeTLB Interface Files
2087 ~~~~~~~~~~~~~~~~~~~~~~~
2089 hugetlb.<hugepagesize>.current
2090 Show current usage for "hugepagesize" hugetlb. It exists for all
2091 the cgroup except root.
2093 hugetlb.<hugepagesize>.max
2094 Set/show the hard limit of "hugepagesize" hugetlb usage.
2095 The default value is "max". It exists for all the cgroup except root.
2097 hugetlb.<hugepagesize>.events
2098 A read-only flat-keyed file which exists on non-root cgroups.
2101 The number of allocation failure due to HugeTLB limit
2103 hugetlb.<hugepagesize>.events.local
2104 Similar to hugetlb.<hugepagesize>.events but the fields in the file
2105 are local to the cgroup i.e. not hierarchical. The file modified event
2106 generated on this file reflects only the local events.
2114 perf_event controller, if not mounted on a legacy hierarchy, is
2115 automatically enabled on the v2 hierarchy so that perf events can
2116 always be filtered by cgroup v2 path. The controller can still be
2117 moved to a legacy hierarchy after v2 hierarchy is populated.
2120 Non-normative information
2121 -------------------------
2123 This section contains information that isn't considered to be a part of
2124 the stable kernel API and so is subject to change.
2127 CPU controller root cgroup process behaviour
2128 ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
2130 When distributing CPU cycles in the root cgroup each thread in this
2131 cgroup is treated as if it was hosted in a separate child cgroup of the
2132 root cgroup. This child cgroup weight is dependent on its thread nice
2135 For details of this mapping see sched_prio_to_weight array in
2136 kernel/sched/core.c file (values from this array should be scaled
2137 appropriately so the neutral - nice 0 - value is 100 instead of 1024).
2140 IO controller root cgroup process behaviour
2141 ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
2143 Root cgroup processes are hosted in an implicit leaf child node.
2144 When distributing IO resources this implicit child node is taken into
2145 account as if it was a normal child cgroup of the root cgroup with a
2146 weight value of 200.
2155 cgroup namespace provides a mechanism to virtualize the view of the
2156 "/proc/$PID/cgroup" file and cgroup mounts. The CLONE_NEWCGROUP clone
2157 flag can be used with clone(2) and unshare(2) to create a new cgroup
2158 namespace. The process running inside the cgroup namespace will have
2159 its "/proc/$PID/cgroup" output restricted to cgroupns root. The
2160 cgroupns root is the cgroup of the process at the time of creation of
2161 the cgroup namespace.
2163 Without cgroup namespace, the "/proc/$PID/cgroup" file shows the
2164 complete path of the cgroup of a process. In a container setup where
2165 a set of cgroups and namespaces are intended to isolate processes the
2166 "/proc/$PID/cgroup" file may leak potential system level information
2167 to the isolated processes. For Example::
2169 # cat /proc/self/cgroup
2170 0::/batchjobs/container_id1
2172 The path '/batchjobs/container_id1' can be considered as system-data
2173 and undesirable to expose to the isolated processes. cgroup namespace
2174 can be used to restrict visibility of this path. For example, before
2175 creating a cgroup namespace, one would see::
2177 # ls -l /proc/self/ns/cgroup
2178 lrwxrwxrwx 1 root root 0 2014-07-15 10:37 /proc/self/ns/cgroup -> cgroup:[4026531835]
2179 # cat /proc/self/cgroup
2180 0::/batchjobs/container_id1
2182 After unsharing a new namespace, the view changes::
2184 # ls -l /proc/self/ns/cgroup
2185 lrwxrwxrwx 1 root root 0 2014-07-15 10:35 /proc/self/ns/cgroup -> cgroup:[4026532183]
2186 # cat /proc/self/cgroup
2189 When some thread from a multi-threaded process unshares its cgroup
2190 namespace, the new cgroupns gets applied to the entire process (all
2191 the threads). This is natural for the v2 hierarchy; however, for the
2192 legacy hierarchies, this may be unexpected.
2194 A cgroup namespace is alive as long as there are processes inside or
2195 mounts pinning it. When the last usage goes away, the cgroup
2196 namespace is destroyed. The cgroupns root and the actual cgroups
2203 The 'cgroupns root' for a cgroup namespace is the cgroup in which the
2204 process calling unshare(2) is running. For example, if a process in
2205 /batchjobs/container_id1 cgroup calls unshare, cgroup
2206 /batchjobs/container_id1 becomes the cgroupns root. For the
2207 init_cgroup_ns, this is the real root ('/') cgroup.
2209 The cgroupns root cgroup does not change even if the namespace creator
2210 process later moves to a different cgroup::
2212 # ~/unshare -c # unshare cgroupns in some cgroup
2213 # cat /proc/self/cgroup
2216 # echo 0 > sub_cgrp_1/cgroup.procs
2217 # cat /proc/self/cgroup
2220 Each process gets its namespace-specific view of "/proc/$PID/cgroup"
2222 Processes running inside the cgroup namespace will be able to see
2223 cgroup paths (in /proc/self/cgroup) only inside their root cgroup.
2224 From within an unshared cgroupns::
2228 # echo 7353 > sub_cgrp_1/cgroup.procs
2229 # cat /proc/7353/cgroup
2232 From the initial cgroup namespace, the real cgroup path will be
2235 $ cat /proc/7353/cgroup
2236 0::/batchjobs/container_id1/sub_cgrp_1
2238 From a sibling cgroup namespace (that is, a namespace rooted at a
2239 different cgroup), the cgroup path relative to its own cgroup
2240 namespace root will be shown. For instance, if PID 7353's cgroup
2241 namespace root is at '/batchjobs/container_id2', then it will see::
2243 # cat /proc/7353/cgroup
2244 0::/../container_id2/sub_cgrp_1
2246 Note that the relative path always starts with '/' to indicate that
2247 its relative to the cgroup namespace root of the caller.
2250 Migration and setns(2)
2251 ----------------------
2253 Processes inside a cgroup namespace can move into and out of the
2254 namespace root if they have proper access to external cgroups. For
2255 example, from inside a namespace with cgroupns root at
2256 /batchjobs/container_id1, and assuming that the global hierarchy is
2257 still accessible inside cgroupns::
2259 # cat /proc/7353/cgroup
2261 # echo 7353 > batchjobs/container_id2/cgroup.procs
2262 # cat /proc/7353/cgroup
2263 0::/../container_id2
2265 Note that this kind of setup is not encouraged. A task inside cgroup
2266 namespace should only be exposed to its own cgroupns hierarchy.
2268 setns(2) to another cgroup namespace is allowed when:
2270 (a) the process has CAP_SYS_ADMIN against its current user namespace
2271 (b) the process has CAP_SYS_ADMIN against the target cgroup
2274 No implicit cgroup changes happen with attaching to another cgroup
2275 namespace. It is expected that the someone moves the attaching
2276 process under the target cgroup namespace root.
2279 Interaction with Other Namespaces
2280 ---------------------------------
2282 Namespace specific cgroup hierarchy can be mounted by a process
2283 running inside a non-init cgroup namespace::
2285 # mount -t cgroup2 none $MOUNT_POINT
2287 This will mount the unified cgroup hierarchy with cgroupns root as the
2288 filesystem root. The process needs CAP_SYS_ADMIN against its user and
2291 The virtualization of /proc/self/cgroup file combined with restricting
2292 the view of cgroup hierarchy by namespace-private cgroupfs mount
2293 provides a properly isolated cgroup view inside the container.
2296 Information on Kernel Programming
2297 =================================
2299 This section contains kernel programming information in the areas
2300 where interacting with cgroup is necessary. cgroup core and
2301 controllers are not covered.
2304 Filesystem Support for Writeback
2305 --------------------------------
2307 A filesystem can support cgroup writeback by updating
2308 address_space_operations->writepage[s]() to annotate bio's using the
2309 following two functions.
2311 wbc_init_bio(@wbc, @bio)
2312 Should be called for each bio carrying writeback data and
2313 associates the bio with the inode's owner cgroup and the
2314 corresponding request queue. This must be called after
2315 a queue (device) has been associated with the bio and
2318 wbc_account_cgroup_owner(@wbc, @page, @bytes)
2319 Should be called for each data segment being written out.
2320 While this function doesn't care exactly when it's called
2321 during the writeback session, it's the easiest and most
2322 natural to call it as data segments are added to a bio.
2324 With writeback bio's annotated, cgroup support can be enabled per
2325 super_block by setting SB_I_CGROUPWB in ->s_iflags. This allows for
2326 selective disabling of cgroup writeback support which is helpful when
2327 certain filesystem features, e.g. journaled data mode, are
2330 wbc_init_bio() binds the specified bio to its cgroup. Depending on
2331 the configuration, the bio may be executed at a lower priority and if
2332 the writeback session is holding shared resources, e.g. a journal
2333 entry, may lead to priority inversion. There is no one easy solution
2334 for the problem. Filesystems can try to work around specific problem
2335 cases by skipping wbc_init_bio() and using bio_associate_blkg()
2339 Deprecated v1 Core Features
2340 ===========================
2342 - Multiple hierarchies including named ones are not supported.
2344 - All v1 mount options are not supported.
2346 - The "tasks" file is removed and "cgroup.procs" is not sorted.
2348 - "cgroup.clone_children" is removed.
2350 - /proc/cgroups is meaningless for v2. Use "cgroup.controllers" file
2351 at the root instead.
2354 Issues with v1 and Rationales for v2
2355 ====================================
2357 Multiple Hierarchies
2358 --------------------
2360 cgroup v1 allowed an arbitrary number of hierarchies and each
2361 hierarchy could host any number of controllers. While this seemed to
2362 provide a high level of flexibility, it wasn't useful in practice.
2364 For example, as there is only one instance of each controller, utility
2365 type controllers such as freezer which can be useful in all
2366 hierarchies could only be used in one. The issue is exacerbated by
2367 the fact that controllers couldn't be moved to another hierarchy once
2368 hierarchies were populated. Another issue was that all controllers
2369 bound to a hierarchy were forced to have exactly the same view of the
2370 hierarchy. It wasn't possible to vary the granularity depending on
2371 the specific controller.
2373 In practice, these issues heavily limited which controllers could be
2374 put on the same hierarchy and most configurations resorted to putting
2375 each controller on its own hierarchy. Only closely related ones, such
2376 as the cpu and cpuacct controllers, made sense to be put on the same
2377 hierarchy. This often meant that userland ended up managing multiple
2378 similar hierarchies repeating the same steps on each hierarchy
2379 whenever a hierarchy management operation was necessary.
2381 Furthermore, support for multiple hierarchies came at a steep cost.
2382 It greatly complicated cgroup core implementation but more importantly
2383 the support for multiple hierarchies restricted how cgroup could be
2384 used in general and what controllers was able to do.
2386 There was no limit on how many hierarchies there might be, which meant
2387 that a thread's cgroup membership couldn't be described in finite
2388 length. The key might contain any number of entries and was unlimited
2389 in length, which made it highly awkward to manipulate and led to
2390 addition of controllers which existed only to identify membership,
2391 which in turn exacerbated the original problem of proliferating number
2394 Also, as a controller couldn't have any expectation regarding the
2395 topologies of hierarchies other controllers might be on, each
2396 controller had to assume that all other controllers were attached to
2397 completely orthogonal hierarchies. This made it impossible, or at
2398 least very cumbersome, for controllers to cooperate with each other.
2400 In most use cases, putting controllers on hierarchies which are
2401 completely orthogonal to each other isn't necessary. What usually is
2402 called for is the ability to have differing levels of granularity
2403 depending on the specific controller. In other words, hierarchy may
2404 be collapsed from leaf towards root when viewed from specific
2405 controllers. For example, a given configuration might not care about
2406 how memory is distributed beyond a certain level while still wanting
2407 to control how CPU cycles are distributed.
2413 cgroup v1 allowed threads of a process to belong to different cgroups.
2414 This didn't make sense for some controllers and those controllers
2415 ended up implementing different ways to ignore such situations but
2416 much more importantly it blurred the line between API exposed to
2417 individual applications and system management interface.
2419 Generally, in-process knowledge is available only to the process
2420 itself; thus, unlike service-level organization of processes,
2421 categorizing threads of a process requires active participation from
2422 the application which owns the target process.
2424 cgroup v1 had an ambiguously defined delegation model which got abused
2425 in combination with thread granularity. cgroups were delegated to
2426 individual applications so that they can create and manage their own
2427 sub-hierarchies and control resource distributions along them. This
2428 effectively raised cgroup to the status of a syscall-like API exposed
2431 First of all, cgroup has a fundamentally inadequate interface to be
2432 exposed this way. For a process to access its own knobs, it has to
2433 extract the path on the target hierarchy from /proc/self/cgroup,
2434 construct the path by appending the name of the knob to the path, open
2435 and then read and/or write to it. This is not only extremely clunky
2436 and unusual but also inherently racy. There is no conventional way to
2437 define transaction across the required steps and nothing can guarantee
2438 that the process would actually be operating on its own sub-hierarchy.
2440 cgroup controllers implemented a number of knobs which would never be
2441 accepted as public APIs because they were just adding control knobs to
2442 system-management pseudo filesystem. cgroup ended up with interface
2443 knobs which were not properly abstracted or refined and directly
2444 revealed kernel internal details. These knobs got exposed to
2445 individual applications through the ill-defined delegation mechanism
2446 effectively abusing cgroup as a shortcut to implementing public APIs
2447 without going through the required scrutiny.
2449 This was painful for both userland and kernel. Userland ended up with
2450 misbehaving and poorly abstracted interfaces and kernel exposing and
2451 locked into constructs inadvertently.
2454 Competition Between Inner Nodes and Threads
2455 -------------------------------------------
2457 cgroup v1 allowed threads to be in any cgroups which created an
2458 interesting problem where threads belonging to a parent cgroup and its
2459 children cgroups competed for resources. This was nasty as two
2460 different types of entities competed and there was no obvious way to
2461 settle it. Different controllers did different things.
2463 The cpu controller considered threads and cgroups as equivalents and
2464 mapped nice levels to cgroup weights. This worked for some cases but
2465 fell flat when children wanted to be allocated specific ratios of CPU
2466 cycles and the number of internal threads fluctuated - the ratios
2467 constantly changed as the number of competing entities fluctuated.
2468 There also were other issues. The mapping from nice level to weight
2469 wasn't obvious or universal, and there were various other knobs which
2470 simply weren't available for threads.
2472 The io controller implicitly created a hidden leaf node for each
2473 cgroup to host the threads. The hidden leaf had its own copies of all
2474 the knobs with ``leaf_`` prefixed. While this allowed equivalent
2475 control over internal threads, it was with serious drawbacks. It
2476 always added an extra layer of nesting which wouldn't be necessary
2477 otherwise, made the interface messy and significantly complicated the
2480 The memory controller didn't have a way to control what happened
2481 between internal tasks and child cgroups and the behavior was not
2482 clearly defined. There were attempts to add ad-hoc behaviors and
2483 knobs to tailor the behavior to specific workloads which would have
2484 led to problems extremely difficult to resolve in the long term.
2486 Multiple controllers struggled with internal tasks and came up with
2487 different ways to deal with it; unfortunately, all the approaches were
2488 severely flawed and, furthermore, the widely different behaviors
2489 made cgroup as a whole highly inconsistent.
2491 This clearly is a problem which needs to be addressed from cgroup core
2495 Other Interface Issues
2496 ----------------------
2498 cgroup v1 grew without oversight and developed a large number of
2499 idiosyncrasies and inconsistencies. One issue on the cgroup core side
2500 was how an empty cgroup was notified - a userland helper binary was
2501 forked and executed for each event. The event delivery wasn't
2502 recursive or delegatable. The limitations of the mechanism also led
2503 to in-kernel event delivery filtering mechanism further complicating
2506 Controller interfaces were problematic too. An extreme example is
2507 controllers completely ignoring hierarchical organization and treating
2508 all cgroups as if they were all located directly under the root
2509 cgroup. Some controllers exposed a large amount of inconsistent
2510 implementation details to userland.
2512 There also was no consistency across controllers. When a new cgroup
2513 was created, some controllers defaulted to not imposing extra
2514 restrictions while others disallowed any resource usage until
2515 explicitly configured. Configuration knobs for the same type of
2516 control used widely differing naming schemes and formats. Statistics
2517 and information knobs were named arbitrarily and used different
2518 formats and units even in the same controller.
2520 cgroup v2 establishes common conventions where appropriate and updates
2521 controllers so that they expose minimal and consistent interfaces.
2524 Controller Issues and Remedies
2525 ------------------------------
2530 The original lower boundary, the soft limit, is defined as a limit
2531 that is per default unset. As a result, the set of cgroups that
2532 global reclaim prefers is opt-in, rather than opt-out. The costs for
2533 optimizing these mostly negative lookups are so high that the
2534 implementation, despite its enormous size, does not even provide the
2535 basic desirable behavior. First off, the soft limit has no
2536 hierarchical meaning. All configured groups are organized in a global
2537 rbtree and treated like equal peers, regardless where they are located
2538 in the hierarchy. This makes subtree delegation impossible. Second,
2539 the soft limit reclaim pass is so aggressive that it not just
2540 introduces high allocation latencies into the system, but also impacts
2541 system performance due to overreclaim, to the point where the feature
2542 becomes self-defeating.
2544 The memory.low boundary on the other hand is a top-down allocated
2545 reserve. A cgroup enjoys reclaim protection when it's within its
2546 effective low, which makes delegation of subtrees possible. It also
2547 enjoys having reclaim pressure proportional to its overage when
2548 above its effective low.
2550 The original high boundary, the hard limit, is defined as a strict
2551 limit that can not budge, even if the OOM killer has to be called.
2552 But this generally goes against the goal of making the most out of the
2553 available memory. The memory consumption of workloads varies during
2554 runtime, and that requires users to overcommit. But doing that with a
2555 strict upper limit requires either a fairly accurate prediction of the
2556 working set size or adding slack to the limit. Since working set size
2557 estimation is hard and error prone, and getting it wrong results in
2558 OOM kills, most users tend to err on the side of a looser limit and
2559 end up wasting precious resources.
2561 The memory.high boundary on the other hand can be set much more
2562 conservatively. When hit, it throttles allocations by forcing them
2563 into direct reclaim to work off the excess, but it never invokes the
2564 OOM killer. As a result, a high boundary that is chosen too
2565 aggressively will not terminate the processes, but instead it will
2566 lead to gradual performance degradation. The user can monitor this
2567 and make corrections until the minimal memory footprint that still
2568 gives acceptable performance is found.
2570 In extreme cases, with many concurrent allocations and a complete
2571 breakdown of reclaim progress within the group, the high boundary can
2572 be exceeded. But even then it's mostly better to satisfy the
2573 allocation from the slack available in other groups or the rest of the
2574 system than killing the group. Otherwise, memory.max is there to
2575 limit this type of spillover and ultimately contain buggy or even
2576 malicious applications.
2578 Setting the original memory.limit_in_bytes below the current usage was
2579 subject to a race condition, where concurrent charges could cause the
2580 limit setting to fail. memory.max on the other hand will first set the
2581 limit to prevent new charges, and then reclaim and OOM kill until the
2582 new limit is met - or the task writing to memory.max is killed.
2584 The combined memory+swap accounting and limiting is replaced by real
2585 control over swap space.
2587 The main argument for a combined memory+swap facility in the original
2588 cgroup design was that global or parental pressure would always be
2589 able to swap all anonymous memory of a child group, regardless of the
2590 child's own (possibly untrusted) configuration. However, untrusted
2591 groups can sabotage swapping by other means - such as referencing its
2592 anonymous memory in a tight loop - and an admin can not assume full
2593 swappability when overcommitting untrusted jobs.
2595 For trusted jobs, on the other hand, a combined counter is not an
2596 intuitive userspace interface, and it flies in the face of the idea
2597 that cgroup controllers should account and limit specific physical
2598 resources. Swap space is a resource like all others in the system,
2599 and that's why unified hierarchy allows distributing it separately.