4 October, 2015 Tejun Heo <tj@kernel.org>
6 This is the authoritative documentation on the design, interface and
7 conventions of cgroup v2. It describes all userland-visible aspects
8 of cgroup including core and specific controller behaviors. All
9 future changes must be reflected in this document. Documentation for
10 v1 is available under Documentation/cgroup-v1/.
19 2-2. Organizing Processes
20 2-3. [Un]populated Notification
21 2-4. Controlling Controllers
22 2-4-1. Enabling and Disabling
23 2-4-2. Top-down Constraint
24 2-4-3. No Internal Process Constraint
26 2-5-1. Model of Delegation
27 2-5-2. Delegation Containment
29 2-6-1. Organize Once and Control
30 2-6-2. Avoid Name Collisions
31 3. Resource Distribution Models
39 4-3. Core Interface Files
42 5-1-1. CPU Interface Files
44 5-2-1. Memory Interface Files
45 5-2-2. Usage Guidelines
46 5-2-3. Memory Ownership
48 5-3-1. IO Interface Files
50 P. Information on Kernel Programming
51 P-1. Filesystem Support for Writeback
52 D. Deprecated v1 Core Features
53 R. Issues with v1 and Rationales for v2
54 R-1. Multiple Hierarchies
55 R-2. Thread Granularity
56 R-3. Competition Between Inner Nodes and Threads
57 R-4. Other Interface Issues
58 R-5. Controller Issues and Remedies
66 "cgroup" stands for "control group" and is never capitalized. The
67 singular form is used to designate the whole feature and also as a
68 qualifier as in "cgroup controllers". When explicitly referring to
69 multiple individual control groups, the plural form "cgroups" is used.
74 cgroup is a mechanism to organize processes hierarchically and
75 distribute system resources along the hierarchy in a controlled and
78 cgroup is largely composed of two parts - the core and controllers.
79 cgroup core is primarily responsible for hierarchically organizing
80 processes. A cgroup controller is usually responsible for
81 distributing a specific type of system resource along the hierarchy
82 although there are utility controllers which serve purposes other than
83 resource distribution.
85 cgroups form a tree structure and every process in the system belongs
86 to one and only one cgroup. All threads of a process belong to the
87 same cgroup. On creation, all processes are put in the cgroup that
88 the parent process belongs to at the time. A process can be migrated
89 to another cgroup. Migration of a process doesn't affect already
90 existing descendant processes.
92 Following certain structural constraints, controllers may be enabled or
93 disabled selectively on a cgroup. All controller behaviors are
94 hierarchical - if a controller is enabled on a cgroup, it affects all
95 processes which belong to the cgroups consisting the inclusive
96 sub-hierarchy of the cgroup. When a controller is enabled on a nested
97 cgroup, it always restricts the resource distribution further. The
98 restrictions set closer to the root in the hierarchy can not be
99 overridden from further away.
106 Unlike v1, cgroup v2 has only single hierarchy. The cgroup v2
107 hierarchy can be mounted with the following mount command.
109 # mount -t cgroup2 none $MOUNT_POINT
111 cgroup2 filesystem has the magic number 0x63677270 ("cgrp"). All
112 controllers which support v2 and are not bound to a v1 hierarchy are
113 automatically bound to the v2 hierarchy and show up at the root.
114 Controllers which are not in active use in the v2 hierarchy can be
115 bound to other hierarchies. This allows mixing v2 hierarchy with the
116 legacy v1 multiple hierarchies in a fully backward compatible way.
118 A controller can be moved across hierarchies only after the controller
119 is no longer referenced in its current hierarchy. Because per-cgroup
120 controller states are destroyed asynchronously and controllers may
121 have lingering references, a controller may not show up immediately on
122 the v2 hierarchy after the final umount of the previous hierarchy.
123 Similarly, a controller should be fully disabled to be moved out of
124 the unified hierarchy and it may take some time for the disabled
125 controller to become available for other hierarchies; furthermore, due
126 to inter-controller dependencies, other controllers may need to be
129 While useful for development and manual configurations, moving
130 controllers dynamically between the v2 and other hierarchies is
131 strongly discouraged for production use. It is recommended to decide
132 the hierarchies and controller associations before starting using the
133 controllers after system boot.
136 2-2. Organizing Processes
138 Initially, only the root cgroup exists to which all processes belong.
139 A child cgroup can be created by creating a sub-directory.
143 A given cgroup may have multiple child cgroups forming a tree
144 structure. Each cgroup has a read-writable interface file
145 "cgroup.procs". When read, it lists the PIDs of all processes which
146 belong to the cgroup one-per-line. The PIDs are not ordered and the
147 same PID may show up more than once if the process got moved to
148 another cgroup and then back or the PID got recycled while reading.
150 A process can be migrated into a cgroup by writing its PID to the
151 target cgroup's "cgroup.procs" file. Only one process can be migrated
152 on a single write(2) call. If a process is composed of multiple
153 threads, writing the PID of any thread migrates all threads of the
156 When a process forks a child process, the new process is born into the
157 cgroup that the forking process belongs to at the time of the
158 operation. After exit, a process stays associated with the cgroup
159 that it belonged to at the time of exit until it's reaped; however, a
160 zombie process does not appear in "cgroup.procs" and thus can't be
161 moved to another cgroup.
163 A cgroup which doesn't have any children or live processes can be
164 destroyed by removing the directory. Note that a cgroup which doesn't
165 have any children and is associated only with zombie processes is
166 considered empty and can be removed.
170 "/proc/$PID/cgroup" lists a process's cgroup membership. If legacy
171 cgroup is in use in the system, this file may contain multiple lines,
172 one for each hierarchy. The entry for cgroup v2 is always in the
175 # cat /proc/842/cgroup
177 0::/test-cgroup/test-cgroup-nested
179 If the process becomes a zombie and the cgroup it was associated with
180 is removed subsequently, " (deleted)" is appended to the path.
182 # cat /proc/842/cgroup
184 0::/test-cgroup/test-cgroup-nested (deleted)
187 2-3. [Un]populated Notification
189 Each non-root cgroup has a "cgroup.events" file which contains
190 "populated" field indicating whether the cgroup's sub-hierarchy has
191 live processes in it. Its value is 0 if there is no live process in
192 the cgroup and its descendants; otherwise, 1. poll and [id]notify
193 events are triggered when the value changes. This can be used, for
194 example, to start a clean-up operation after all processes of a given
195 sub-hierarchy have exited. The populated state updates and
196 notifications are recursive. Consider the following sub-hierarchy
197 where the numbers in the parentheses represent the numbers of processes
203 A, B and C's "populated" fields would be 1 while D's 0. After the one
204 process in C exits, B and C's "populated" fields would flip to "0" and
205 file modified events will be generated on the "cgroup.events" files of
209 2-4. Controlling Controllers
211 2-4-1. Enabling and Disabling
213 Each cgroup has a "cgroup.controllers" file which lists all
214 controllers available for the cgroup to enable.
216 # cat cgroup.controllers
219 No controller is enabled by default. Controllers can be enabled and
220 disabled by writing to the "cgroup.subtree_control" file.
222 # echo "+cpu +memory -io" > cgroup.subtree_control
224 Only controllers which are listed in "cgroup.controllers" can be
225 enabled. When multiple operations are specified as above, either they
226 all succeed or fail. If multiple operations on the same controller
227 are specified, the last one is effective.
229 Enabling a controller in a cgroup indicates that the distribution of
230 the target resource across its immediate children will be controlled.
231 Consider the following sub-hierarchy. The enabled controllers are
232 listed in parentheses.
234 A(cpu,memory) - B(memory) - C()
237 As A has "cpu" and "memory" enabled, A will control the distribution
238 of CPU cycles and memory to its children, in this case, B. As B has
239 "memory" enabled but not "CPU", C and D will compete freely on CPU
240 cycles but their division of memory available to B will be controlled.
242 As a controller regulates the distribution of the target resource to
243 the cgroup's children, enabling it creates the controller's interface
244 files in the child cgroups. In the above example, enabling "cpu" on B
245 would create the "cpu." prefixed controller interface files in C and
246 D. Likewise, disabling "memory" from B would remove the "memory."
247 prefixed controller interface files from C and D. This means that the
248 controller interface files - anything which doesn't start with
249 "cgroup." are owned by the parent rather than the cgroup itself.
252 2-4-2. Top-down Constraint
254 Resources are distributed top-down and a cgroup can further distribute
255 a resource only if the resource has been distributed to it from the
256 parent. This means that all non-root "cgroup.subtree_control" files
257 can only contain controllers which are enabled in the parent's
258 "cgroup.subtree_control" file. A controller can be enabled only if
259 the parent has the controller enabled and a controller can't be
260 disabled if one or more children have it enabled.
263 2-4-3. No Internal Process Constraint
265 Non-root cgroups can only distribute resources to their children when
266 they don't have any processes of their own. In other words, only
267 cgroups which don't contain any processes can have controllers enabled
268 in their "cgroup.subtree_control" files.
270 This guarantees that, when a controller is looking at the part of the
271 hierarchy which has it enabled, processes are always only on the
272 leaves. This rules out situations where child cgroups compete against
273 internal processes of the parent.
275 The root cgroup is exempt from this restriction. Root contains
276 processes and anonymous resource consumption which can't be associated
277 with any other cgroups and requires special treatment from most
278 controllers. How resource consumption in the root cgroup is governed
279 is up to each controller.
281 Note that the restriction doesn't get in the way if there is no
282 enabled controller in the cgroup's "cgroup.subtree_control". This is
283 important as otherwise it wouldn't be possible to create children of a
284 populated cgroup. To control resource distribution of a cgroup, the
285 cgroup must create children and transfer all its processes to the
286 children before enabling controllers in its "cgroup.subtree_control"
292 2-5-1. Model of Delegation
294 A cgroup can be delegated to a less privileged user by granting write
295 access of the directory and its "cgroup.procs" file to the user. Note
296 that resource control interface files in a given directory control the
297 distribution of the parent's resources and thus must not be delegated
298 along with the directory.
300 Once delegated, the user can build sub-hierarchy under the directory,
301 organize processes as it sees fit and further distribute the resources
302 it received from the parent. The limits and other settings of all
303 resource controllers are hierarchical and regardless of what happens
304 in the delegated sub-hierarchy, nothing can escape the resource
305 restrictions imposed by the parent.
307 Currently, cgroup doesn't impose any restrictions on the number of
308 cgroups in or nesting depth of a delegated sub-hierarchy; however,
309 this may be limited explicitly in the future.
312 2-5-2. Delegation Containment
314 A delegated sub-hierarchy is contained in the sense that processes
315 can't be moved into or out of the sub-hierarchy by the delegatee. For
316 a process with a non-root euid to migrate a target process into a
317 cgroup by writing its PID to the "cgroup.procs" file, the following
318 conditions must be met.
320 - The writer's euid must match either uid or suid of the target process.
322 - The writer must have write access to the "cgroup.procs" file.
324 - The writer must have write access to the "cgroup.procs" file of the
325 common ancestor of the source and destination cgroups.
327 The above three constraints ensure that while a delegatee may migrate
328 processes around freely in the delegated sub-hierarchy it can't pull
329 in from or push out to outside the sub-hierarchy.
331 For an example, let's assume cgroups C0 and C1 have been delegated to
332 user U0 who created C00, C01 under C0 and C10 under C1 as follows and
333 all processes under C0 and C1 belong to U0.
335 ~~~~~~~~~~~~~ - C0 - C00
338 ~~~~~~~~~~~~~ - C1 - C10
340 Let's also say U0 wants to write the PID of a process which is
341 currently in C10 into "C00/cgroup.procs". U0 has write access to the
342 file and uid match on the process; however, the common ancestor of the
343 source cgroup C10 and the destination cgroup C00 is above the points
344 of delegation and U0 would not have write access to its "cgroup.procs"
345 files and thus the write will be denied with -EACCES.
350 2-6-1. Organize Once and Control
352 Migrating a process across cgroups is a relatively expensive operation
353 and stateful resources such as memory are not moved together with the
354 process. This is an explicit design decision as there often exist
355 inherent trade-offs between migration and various hot paths in terms
356 of synchronization cost.
358 As such, migrating processes across cgroups frequently as a means to
359 apply different resource restrictions is discouraged. A workload
360 should be assigned to a cgroup according to the system's logical and
361 resource structure once on start-up. Dynamic adjustments to resource
362 distribution can be made by changing controller configuration through
366 2-6-2. Avoid Name Collisions
368 Interface files for a cgroup and its children cgroups occupy the same
369 directory and it is possible to create children cgroups which collide
370 with interface files.
372 All cgroup core interface files are prefixed with "cgroup." and each
373 controller's interface files are prefixed with the controller name and
374 a dot. A controller's name is composed of lower case alphabets and
375 '_'s but never begins with an '_' so it can be used as the prefix
376 character for collision avoidance. Also, interface file names won't
377 start or end with terms which are often used in categorizing workloads
378 such as job, service, slice, unit or workload.
380 cgroup doesn't do anything to prevent name collisions and it's the
381 user's responsibility to avoid them.
384 3. Resource Distribution Models
386 cgroup controllers implement several resource distribution schemes
387 depending on the resource type and expected use cases. This section
388 describes major schemes in use along with their expected behaviors.
393 A parent's resource is distributed by adding up the weights of all
394 active children and giving each the fraction matching the ratio of its
395 weight against the sum. As only children which can make use of the
396 resource at the moment participate in the distribution, this is
397 work-conserving. Due to the dynamic nature, this model is usually
398 used for stateless resources.
400 All weights are in the range [1, 10000] with the default at 100. This
401 allows symmetric multiplicative biases in both directions at fine
402 enough granularity while staying in the intuitive range.
404 As long as the weight is in range, all configuration combinations are
405 valid and there is no reason to reject configuration changes or
408 "cpu.weight" proportionally distributes CPU cycles to active children
409 and is an example of this type.
414 A child can only consume upto the configured amount of the resource.
415 Limits can be over-committed - the sum of the limits of children can
416 exceed the amount of resource available to the parent.
418 Limits are in the range [0, max] and defaults to "max", which is noop.
420 As limits can be over-committed, all configuration combinations are
421 valid and there is no reason to reject configuration changes or
424 "io.max" limits the maximum BPS and/or IOPS that a cgroup can consume
425 on an IO device and is an example of this type.
430 A cgroup is protected to be allocated upto the configured amount of
431 the resource if the usages of all its ancestors are under their
432 protected levels. Protections can be hard guarantees or best effort
433 soft boundaries. Protections can also be over-committed in which case
434 only upto the amount available to the parent is protected among
437 Protections are in the range [0, max] and defaults to 0, which is
440 As protections can be over-committed, all configuration combinations
441 are valid and there is no reason to reject configuration changes or
444 "memory.low" implements best-effort memory protection and is an
445 example of this type.
450 A cgroup is exclusively allocated a certain amount of a finite
451 resource. Allocations can't be over-committed - the sum of the
452 allocations of children can not exceed the amount of resource
453 available to the parent.
455 Allocations are in the range [0, max] and defaults to 0, which is no
458 As allocations can't be over-committed, some configuration
459 combinations are invalid and should be rejected. Also, if the
460 resource is mandatory for execution of processes, process migrations
463 "cpu.rt.max" hard-allocates realtime slices and is an example of this
471 All interface files should be in one of the following formats whenever
474 New-line separated values
475 (when only one value can be written at once)
481 Space separated values
482 (when read-only or multiple values can be written at once)
494 KEY0 SUB_KEY0=VAL00 SUB_KEY1=VAL01...
495 KEY1 SUB_KEY0=VAL10 SUB_KEY1=VAL11...
498 For a writable file, the format for writing should generally match
499 reading; however, controllers may allow omitting later fields or
500 implement restricted shortcuts for most common use cases.
502 For both flat and nested keyed files, only the values for a single key
503 can be written at a time. For nested keyed files, the sub key pairs
504 may be specified in any order and not all pairs have to be specified.
509 - Settings for a single feature should be contained in a single file.
511 - The root cgroup should be exempt from resource control and thus
512 shouldn't have resource control interface files. Also,
513 informational files on the root cgroup which end up showing global
514 information available elsewhere shouldn't exist.
516 - If a controller implements weight based resource distribution, its
517 interface file should be named "weight" and have the range [1,
518 10000] with 100 as the default. The values are chosen to allow
519 enough and symmetric bias in both directions while keeping it
520 intuitive (the default is 100%).
522 - If a controller implements an absolute resource guarantee and/or
523 limit, the interface files should be named "min" and "max"
524 respectively. If a controller implements best effort resource
525 guarantee and/or limit, the interface files should be named "low"
526 and "high" respectively.
528 In the above four control files, the special token "max" should be
529 used to represent upward infinity for both reading and writing.
531 - If a setting has a configurable default value and keyed specific
532 overrides, the default entry should be keyed with "default" and
533 appear as the first entry in the file.
535 The default value can be updated by writing either "default $VAL" or
538 When writing to update a specific override, "default" can be used as
539 the value to indicate removal of the override. Override entries
540 with "default" as the value must not appear when read.
542 For example, a setting which is keyed by major:minor device numbers
543 with integer values may look like the following.
545 # cat cgroup-example-interface-file
549 The default value can be updated by
551 # echo 125 > cgroup-example-interface-file
555 # echo "default 125" > cgroup-example-interface-file
557 An override can be set by
559 # echo "8:16 170" > cgroup-example-interface-file
563 # echo "8:0 default" > cgroup-example-interface-file
564 # cat cgroup-example-interface-file
568 - For events which are not very high frequency, an interface file
569 "events" should be created which lists event key value pairs.
570 Whenever a notifiable event happens, file modified event should be
571 generated on the file.
574 4-3. Core Interface Files
576 All cgroup core files are prefixed with "cgroup."
580 A read-write new-line separated values file which exists on
583 When read, it lists the PIDs of all processes which belong to
584 the cgroup one-per-line. The PIDs are not ordered and the
585 same PID may show up more than once if the process got moved
586 to another cgroup and then back or the PID got recycled while
589 A PID can be written to migrate the process associated with
590 the PID to the cgroup. The writer should match all of the
591 following conditions.
593 - Its euid is either root or must match either uid or suid of
596 - It must have write access to the "cgroup.procs" file.
598 - It must have write access to the "cgroup.procs" file of the
599 common ancestor of the source and destination cgroups.
601 When delegating a sub-hierarchy, write access to this file
602 should be granted along with the containing directory.
606 A read-only space separated values file which exists on all
609 It shows space separated list of all controllers available to
610 the cgroup. The controllers are not ordered.
612 cgroup.subtree_control
614 A read-write space separated values file which exists on all
615 cgroups. Starts out empty.
617 When read, it shows space separated list of the controllers
618 which are enabled to control resource distribution from the
619 cgroup to its children.
621 Space separated list of controllers prefixed with '+' or '-'
622 can be written to enable or disable controllers. A controller
623 name prefixed with '+' enables the controller and '-'
624 disables. If a controller appears more than once on the list,
625 the last one is effective. When multiple enable and disable
626 operations are specified, either all succeed or all fail.
630 A read-only flat-keyed file which exists on non-root cgroups.
631 The following entries are defined. Unless specified
632 otherwise, a value change in this file generates a file
637 1 if the cgroup or its descendants contains any live
638 processes; otherwise, 0.
645 [NOTE: The interface for the cpu controller hasn't been merged yet]
647 The "cpu" controllers regulates distribution of CPU cycles. This
648 controller implements weight and absolute bandwidth limit models for
649 normal scheduling policy and absolute bandwidth allocation model for
650 realtime scheduling policy.
653 5-1-1. CPU Interface Files
655 All time durations are in microseconds.
659 A read-only flat-keyed file which exists on non-root cgroups.
661 It reports the following six stats.
672 A read-write single value file which exists on non-root
673 cgroups. The default is "100".
675 The weight in the range [1, 10000].
679 A read-write two value file which exists on non-root cgroups.
680 The default is "max 100000".
682 The maximum bandwidth limit. It's in the following format.
686 which indicates that the group may consume upto $MAX in each
687 $PERIOD duration. "max" for $MAX indicates no limit. If only
688 one number is written, $MAX is updated.
692 [NOTE: The semantics of this file is still under discussion and the
693 interface hasn't been merged yet]
695 A read-write two value file which exists on all cgroups.
696 The default is "0 100000".
698 The maximum realtime runtime allocation. Over-committing
699 configurations are disallowed and process migrations are
700 rejected if not enough bandwidth is available. It's in the
705 which indicates that the group may consume upto $MAX in each
706 $PERIOD duration. If only one number is written, $MAX is
712 The "memory" controller regulates distribution of memory. Memory is
713 stateful and implements both limit and protection models. Due to the
714 intertwining between memory usage and reclaim pressure and the
715 stateful nature of memory, the distribution model is relatively
718 While not completely water-tight, all major memory usages by a given
719 cgroup are tracked so that the total memory consumption can be
720 accounted and controlled to a reasonable extent. Currently, the
721 following types of memory usages are tracked.
723 - Userland memory - page cache and anonymous memory.
725 - Kernel data structures such as dentries and inodes.
727 - TCP socket buffers.
729 The above list may expand in the future for better coverage.
732 5-2-1. Memory Interface Files
734 All memory amounts are in bytes. If a value which is not aligned to
735 PAGE_SIZE is written, the value may be rounded up to the closest
736 PAGE_SIZE multiple when read back.
740 A read-only single value file which exists on non-root
743 The total amount of memory currently being used by the cgroup
748 A read-write single value file which exists on non-root
749 cgroups. The default is "0".
751 Best-effort memory protection. If the memory usages of a
752 cgroup and all its ancestors are below their low boundaries,
753 the cgroup's memory won't be reclaimed unless memory can be
754 reclaimed from unprotected cgroups.
756 Putting more memory than generally available under this
757 protection is discouraged.
761 A read-write single value file which exists on non-root
762 cgroups. The default is "max".
764 Memory usage throttle limit. This is the main mechanism to
765 control memory usage of a cgroup. If a cgroup's usage goes
766 over the high boundary, the processes of the cgroup are
767 throttled and put under heavy reclaim pressure.
769 Going over the high limit never invokes the OOM killer and
770 under extreme conditions the limit may be breached.
774 A read-write single value file which exists on non-root
775 cgroups. The default is "max".
777 Memory usage hard limit. This is the final protection
778 mechanism. If a cgroup's memory usage reaches this limit and
779 can't be reduced, the OOM killer is invoked in the cgroup.
780 Under certain circumstances, the usage may go over the limit
783 This is the ultimate protection mechanism. As long as the
784 high limit is used and monitored properly, this limit's
785 utility is limited to providing the final safety net.
789 A read-only flat-keyed file which exists on non-root cgroups.
790 The following entries are defined. Unless specified
791 otherwise, a value change in this file generates a file
796 The number of times the cgroup is reclaimed due to
797 high memory pressure even though its usage is under
798 the low boundary. This usually indicates that the low
799 boundary is over-committed.
803 The number of times processes of the cgroup are
804 throttled and routed to perform direct memory reclaim
805 because the high memory boundary was exceeded. For a
806 cgroup whose memory usage is capped by the high limit
807 rather than global memory pressure, this event's
808 occurrences are expected.
812 The number of times the cgroup's memory usage was
813 about to go over the max boundary. If direct reclaim
814 fails to bring it down, the OOM killer is invoked.
818 The number of times the OOM killer has been invoked in
819 the cgroup. This may not exactly match the number of
820 processes killed but should generally be close.
824 A read-only flat-keyed file which exists on non-root cgroups.
826 This breaks down the cgroup's memory footprint into different
827 types of memory, type-specific details, and other information
828 on the state and past events of the memory management system.
830 All memory amounts are in bytes.
832 The entries are ordered to be human readable, and new entries
833 can show up in the middle. Don't rely on items remaining in a
834 fixed position; use the keys to look up specific values!
838 Amount of memory used in anonymous mappings such as
839 brk(), sbrk(), and mmap(MAP_ANONYMOUS)
843 Amount of memory used to cache filesystem data,
844 including tmpfs and shared memory.
848 Amount of memory used in network transmission buffers
852 Amount of cached filesystem data mapped with mmap()
856 Amount of cached filesystem data that was modified but
857 not yet written back to disk
861 Amount of cached filesystem data that was modified and
862 is currently being written back to disk
870 Amount of memory, swap-backed and filesystem-backed,
871 on the internal memory management lists used by the
872 page reclaim algorithm
876 Total number of page faults incurred
880 Number of major page faults incurred
884 A read-only single value file which exists on non-root
887 The total amount of swap currently being used by the cgroup
892 A read-write single value file which exists on non-root
893 cgroups. The default is "max".
895 Swap usage hard limit. If a cgroup's swap usage reaches this
896 limit, anonymous meomry of the cgroup will not be swapped out.
901 "memory.high" is the main mechanism to control memory usage.
902 Over-committing on high limit (sum of high limits > available memory)
903 and letting global memory pressure to distribute memory according to
904 usage is a viable strategy.
906 Because breach of the high limit doesn't trigger the OOM killer but
907 throttles the offending cgroup, a management agent has ample
908 opportunities to monitor and take appropriate actions such as granting
909 more memory or terminating the workload.
911 Determining whether a cgroup has enough memory is not trivial as
912 memory usage doesn't indicate whether the workload can benefit from
913 more memory. For example, a workload which writes data received from
914 network to a file can use all available memory but can also operate as
915 performant with a small amount of memory. A measure of memory
916 pressure - how much the workload is being impacted due to lack of
917 memory - is necessary to determine whether a workload needs more
918 memory; unfortunately, memory pressure monitoring mechanism isn't
922 5-2-3. Memory Ownership
924 A memory area is charged to the cgroup which instantiated it and stays
925 charged to the cgroup until the area is released. Migrating a process
926 to a different cgroup doesn't move the memory usages that it
927 instantiated while in the previous cgroup to the new cgroup.
929 A memory area may be used by processes belonging to different cgroups.
930 To which cgroup the area will be charged is in-deterministic; however,
931 over time, the memory area is likely to end up in a cgroup which has
932 enough memory allowance to avoid high reclaim pressure.
934 If a cgroup sweeps a considerable amount of memory which is expected
935 to be accessed repeatedly by other cgroups, it may make sense to use
936 POSIX_FADV_DONTNEED to relinquish the ownership of memory areas
937 belonging to the affected files to ensure correct memory ownership.
942 The "io" controller regulates the distribution of IO resources. This
943 controller implements both weight based and absolute bandwidth or IOPS
944 limit distribution; however, weight based distribution is available
945 only if cfq-iosched is in use and neither scheme is available for
949 5-3-1. IO Interface Files
953 A read-only nested-keyed file which exists on non-root
956 Lines are keyed by $MAJ:$MIN device numbers and not ordered.
957 The following nested keys are defined.
961 rios Number of read IOs
962 wios Number of write IOs
964 An example read output follows.
966 8:16 rbytes=1459200 wbytes=314773504 rios=192 wios=353
967 8:0 rbytes=90430464 wbytes=299008000 rios=8950 wios=1252
971 A read-write flat-keyed file which exists on non-root cgroups.
972 The default is "default 100".
974 The first line is the default weight applied to devices
975 without specific override. The rest are overrides keyed by
976 $MAJ:$MIN device numbers and not ordered. The weights are in
977 the range [1, 10000] and specifies the relative amount IO time
978 the cgroup can use in relation to its siblings.
980 The default weight can be updated by writing either "default
981 $WEIGHT" or simply "$WEIGHT". Overrides can be set by writing
982 "$MAJ:$MIN $WEIGHT" and unset by writing "$MAJ:$MIN default".
984 An example read output follows.
992 A read-write nested-keyed file which exists on non-root
995 BPS and IOPS based IO limit. Lines are keyed by $MAJ:$MIN
996 device numbers and not ordered. The following nested keys are
999 rbps Max read bytes per second
1000 wbps Max write bytes per second
1001 riops Max read IO operations per second
1002 wiops Max write IO operations per second
1004 When writing, any number of nested key-value pairs can be
1005 specified in any order. "max" can be specified as the value
1006 to remove a specific limit. If the same key is specified
1007 multiple times, the outcome is undefined.
1009 BPS and IOPS are measured in each IO direction and IOs are
1010 delayed if limit is reached. Temporary bursts are allowed.
1012 Setting read limit at 2M BPS and write at 120 IOPS for 8:16.
1014 echo "8:16 rbps=2097152 wiops=120" > io.max
1016 Reading returns the following.
1018 8:16 rbps=2097152 wbps=max riops=max wiops=120
1020 Write IOPS limit can be removed by writing the following.
1022 echo "8:16 wiops=max" > io.max
1024 Reading now returns the following.
1026 8:16 rbps=2097152 wbps=max riops=max wiops=max
1031 Page cache is dirtied through buffered writes and shared mmaps and
1032 written asynchronously to the backing filesystem by the writeback
1033 mechanism. Writeback sits between the memory and IO domains and
1034 regulates the proportion of dirty memory by balancing dirtying and
1037 The io controller, in conjunction with the memory controller,
1038 implements control of page cache writeback IOs. The memory controller
1039 defines the memory domain that dirty memory ratio is calculated and
1040 maintained for and the io controller defines the io domain which
1041 writes out dirty pages for the memory domain. Both system-wide and
1042 per-cgroup dirty memory states are examined and the more restrictive
1043 of the two is enforced.
1045 cgroup writeback requires explicit support from the underlying
1046 filesystem. Currently, cgroup writeback is implemented on ext2, ext4
1047 and btrfs. On other filesystems, all writeback IOs are attributed to
1050 There are inherent differences in memory and writeback management
1051 which affects how cgroup ownership is tracked. Memory is tracked per
1052 page while writeback per inode. For the purpose of writeback, an
1053 inode is assigned to a cgroup and all IO requests to write dirty pages
1054 from the inode are attributed to that cgroup.
1056 As cgroup ownership for memory is tracked per page, there can be pages
1057 which are associated with different cgroups than the one the inode is
1058 associated with. These are called foreign pages. The writeback
1059 constantly keeps track of foreign pages and, if a particular foreign
1060 cgroup becomes the majority over a certain period of time, switches
1061 the ownership of the inode to that cgroup.
1063 While this model is enough for most use cases where a given inode is
1064 mostly dirtied by a single cgroup even when the main writing cgroup
1065 changes over time, use cases where multiple cgroups write to a single
1066 inode simultaneously are not supported well. In such circumstances, a
1067 significant portion of IOs are likely to be attributed incorrectly.
1068 As memory controller assigns page ownership on the first use and
1069 doesn't update it until the page is released, even if writeback
1070 strictly follows page ownership, multiple cgroups dirtying overlapping
1071 areas wouldn't work as expected. It's recommended to avoid such usage
1074 The sysctl knobs which affect writeback behavior are applied to cgroup
1075 writeback as follows.
1077 vm.dirty_background_ratio
1080 These ratios apply the same to cgroup writeback with the
1081 amount of available memory capped by limits imposed by the
1082 memory controller and system-wide clean memory.
1084 vm.dirty_background_bytes
1087 For cgroup writeback, this is calculated into ratio against
1088 total available memory and applied the same way as
1089 vm.dirty[_background]_ratio.
1092 P. Information on Kernel Programming
1094 This section contains kernel programming information in the areas
1095 where interacting with cgroup is necessary. cgroup core and
1096 controllers are not covered.
1099 P-1. Filesystem Support for Writeback
1101 A filesystem can support cgroup writeback by updating
1102 address_space_operations->writepage[s]() to annotate bio's using the
1103 following two functions.
1105 wbc_init_bio(@wbc, @bio)
1107 Should be called for each bio carrying writeback data and
1108 associates the bio with the inode's owner cgroup. Can be
1109 called anytime between bio allocation and submission.
1111 wbc_account_io(@wbc, @page, @bytes)
1113 Should be called for each data segment being written out.
1114 While this function doesn't care exactly when it's called
1115 during the writeback session, it's the easiest and most
1116 natural to call it as data segments are added to a bio.
1118 With writeback bio's annotated, cgroup support can be enabled per
1119 super_block by setting SB_I_CGROUPWB in ->s_iflags. This allows for
1120 selective disabling of cgroup writeback support which is helpful when
1121 certain filesystem features, e.g. journaled data mode, are
1124 wbc_init_bio() binds the specified bio to its cgroup. Depending on
1125 the configuration, the bio may be executed at a lower priority and if
1126 the writeback session is holding shared resources, e.g. a journal
1127 entry, may lead to priority inversion. There is no one easy solution
1128 for the problem. Filesystems can try to work around specific problem
1129 cases by skipping wbc_init_bio() or using bio_associate_blkcg()
1133 D. Deprecated v1 Core Features
1135 - Multiple hierarchies including named ones are not supported.
1137 - All mount options and remounting are not supported.
1139 - The "tasks" file is removed and "cgroup.procs" is not sorted.
1141 - "cgroup.clone_children" is removed.
1143 - /proc/cgroups is meaningless for v2. Use "cgroup.controllers" file
1144 at the root instead.
1147 R. Issues with v1 and Rationales for v2
1149 R-1. Multiple Hierarchies
1151 cgroup v1 allowed an arbitrary number of hierarchies and each
1152 hierarchy could host any number of controllers. While this seemed to
1153 provide a high level of flexibility, it wasn't useful in practice.
1155 For example, as there is only one instance of each controller, utility
1156 type controllers such as freezer which can be useful in all
1157 hierarchies could only be used in one. The issue is exacerbated by
1158 the fact that controllers couldn't be moved to another hierarchy once
1159 hierarchies were populated. Another issue was that all controllers
1160 bound to a hierarchy were forced to have exactly the same view of the
1161 hierarchy. It wasn't possible to vary the granularity depending on
1162 the specific controller.
1164 In practice, these issues heavily limited which controllers could be
1165 put on the same hierarchy and most configurations resorted to putting
1166 each controller on its own hierarchy. Only closely related ones, such
1167 as the cpu and cpuacct controllers, made sense to be put on the same
1168 hierarchy. This often meant that userland ended up managing multiple
1169 similar hierarchies repeating the same steps on each hierarchy
1170 whenever a hierarchy management operation was necessary.
1172 Furthermore, support for multiple hierarchies came at a steep cost.
1173 It greatly complicated cgroup core implementation but more importantly
1174 the support for multiple hierarchies restricted how cgroup could be
1175 used in general and what controllers was able to do.
1177 There was no limit on how many hierarchies there might be, which meant
1178 that a thread's cgroup membership couldn't be described in finite
1179 length. The key might contain any number of entries and was unlimited
1180 in length, which made it highly awkward to manipulate and led to
1181 addition of controllers which existed only to identify membership,
1182 which in turn exacerbated the original problem of proliferating number
1185 Also, as a controller couldn't have any expectation regarding the
1186 topologies of hierarchies other controllers might be on, each
1187 controller had to assume that all other controllers were attached to
1188 completely orthogonal hierarchies. This made it impossible, or at
1189 least very cumbersome, for controllers to cooperate with each other.
1191 In most use cases, putting controllers on hierarchies which are
1192 completely orthogonal to each other isn't necessary. What usually is
1193 called for is the ability to have differing levels of granularity
1194 depending on the specific controller. In other words, hierarchy may
1195 be collapsed from leaf towards root when viewed from specific
1196 controllers. For example, a given configuration might not care about
1197 how memory is distributed beyond a certain level while still wanting
1198 to control how CPU cycles are distributed.
1201 R-2. Thread Granularity
1203 cgroup v1 allowed threads of a process to belong to different cgroups.
1204 This didn't make sense for some controllers and those controllers
1205 ended up implementing different ways to ignore such situations but
1206 much more importantly it blurred the line between API exposed to
1207 individual applications and system management interface.
1209 Generally, in-process knowledge is available only to the process
1210 itself; thus, unlike service-level organization of processes,
1211 categorizing threads of a process requires active participation from
1212 the application which owns the target process.
1214 cgroup v1 had an ambiguously defined delegation model which got abused
1215 in combination with thread granularity. cgroups were delegated to
1216 individual applications so that they can create and manage their own
1217 sub-hierarchies and control resource distributions along them. This
1218 effectively raised cgroup to the status of a syscall-like API exposed
1221 First of all, cgroup has a fundamentally inadequate interface to be
1222 exposed this way. For a process to access its own knobs, it has to
1223 extract the path on the target hierarchy from /proc/self/cgroup,
1224 construct the path by appending the name of the knob to the path, open
1225 and then read and/or write to it. This is not only extremely clunky
1226 and unusual but also inherently racy. There is no conventional way to
1227 define transaction across the required steps and nothing can guarantee
1228 that the process would actually be operating on its own sub-hierarchy.
1230 cgroup controllers implemented a number of knobs which would never be
1231 accepted as public APIs because they were just adding control knobs to
1232 system-management pseudo filesystem. cgroup ended up with interface
1233 knobs which were not properly abstracted or refined and directly
1234 revealed kernel internal details. These knobs got exposed to
1235 individual applications through the ill-defined delegation mechanism
1236 effectively abusing cgroup as a shortcut to implementing public APIs
1237 without going through the required scrutiny.
1239 This was painful for both userland and kernel. Userland ended up with
1240 misbehaving and poorly abstracted interfaces and kernel exposing and
1241 locked into constructs inadvertently.
1244 R-3. Competition Between Inner Nodes and Threads
1246 cgroup v1 allowed threads to be in any cgroups which created an
1247 interesting problem where threads belonging to a parent cgroup and its
1248 children cgroups competed for resources. This was nasty as two
1249 different types of entities competed and there was no obvious way to
1250 settle it. Different controllers did different things.
1252 The cpu controller considered threads and cgroups as equivalents and
1253 mapped nice levels to cgroup weights. This worked for some cases but
1254 fell flat when children wanted to be allocated specific ratios of CPU
1255 cycles and the number of internal threads fluctuated - the ratios
1256 constantly changed as the number of competing entities fluctuated.
1257 There also were other issues. The mapping from nice level to weight
1258 wasn't obvious or universal, and there were various other knobs which
1259 simply weren't available for threads.
1261 The io controller implicitly created a hidden leaf node for each
1262 cgroup to host the threads. The hidden leaf had its own copies of all
1263 the knobs with "leaf_" prefixed. While this allowed equivalent
1264 control over internal threads, it was with serious drawbacks. It
1265 always added an extra layer of nesting which wouldn't be necessary
1266 otherwise, made the interface messy and significantly complicated the
1269 The memory controller didn't have a way to control what happened
1270 between internal tasks and child cgroups and the behavior was not
1271 clearly defined. There were attempts to add ad-hoc behaviors and
1272 knobs to tailor the behavior to specific workloads which would have
1273 led to problems extremely difficult to resolve in the long term.
1275 Multiple controllers struggled with internal tasks and came up with
1276 different ways to deal with it; unfortunately, all the approaches were
1277 severely flawed and, furthermore, the widely different behaviors
1278 made cgroup as a whole highly inconsistent.
1280 This clearly is a problem which needs to be addressed from cgroup core
1284 R-4. Other Interface Issues
1286 cgroup v1 grew without oversight and developed a large number of
1287 idiosyncrasies and inconsistencies. One issue on the cgroup core side
1288 was how an empty cgroup was notified - a userland helper binary was
1289 forked and executed for each event. The event delivery wasn't
1290 recursive or delegatable. The limitations of the mechanism also led
1291 to in-kernel event delivery filtering mechanism further complicating
1294 Controller interfaces were problematic too. An extreme example is
1295 controllers completely ignoring hierarchical organization and treating
1296 all cgroups as if they were all located directly under the root
1297 cgroup. Some controllers exposed a large amount of inconsistent
1298 implementation details to userland.
1300 There also was no consistency across controllers. When a new cgroup
1301 was created, some controllers defaulted to not imposing extra
1302 restrictions while others disallowed any resource usage until
1303 explicitly configured. Configuration knobs for the same type of
1304 control used widely differing naming schemes and formats. Statistics
1305 and information knobs were named arbitrarily and used different
1306 formats and units even in the same controller.
1308 cgroup v2 establishes common conventions where appropriate and updates
1309 controllers so that they expose minimal and consistent interfaces.
1312 R-5. Controller Issues and Remedies
1316 The original lower boundary, the soft limit, is defined as a limit
1317 that is per default unset. As a result, the set of cgroups that
1318 global reclaim prefers is opt-in, rather than opt-out. The costs for
1319 optimizing these mostly negative lookups are so high that the
1320 implementation, despite its enormous size, does not even provide the
1321 basic desirable behavior. First off, the soft limit has no
1322 hierarchical meaning. All configured groups are organized in a global
1323 rbtree and treated like equal peers, regardless where they are located
1324 in the hierarchy. This makes subtree delegation impossible. Second,
1325 the soft limit reclaim pass is so aggressive that it not just
1326 introduces high allocation latencies into the system, but also impacts
1327 system performance due to overreclaim, to the point where the feature
1328 becomes self-defeating.
1330 The memory.low boundary on the other hand is a top-down allocated
1331 reserve. A cgroup enjoys reclaim protection when it and all its
1332 ancestors are below their low boundaries, which makes delegation of
1333 subtrees possible. Secondly, new cgroups have no reserve per default
1334 and in the common case most cgroups are eligible for the preferred
1335 reclaim pass. This allows the new low boundary to be efficiently
1336 implemented with just a minor addition to the generic reclaim code,
1337 without the need for out-of-band data structures and reclaim passes.
1338 Because the generic reclaim code considers all cgroups except for the
1339 ones running low in the preferred first reclaim pass, overreclaim of
1340 individual groups is eliminated as well, resulting in much better
1341 overall workload performance.
1343 The original high boundary, the hard limit, is defined as a strict
1344 limit that can not budge, even if the OOM killer has to be called.
1345 But this generally goes against the goal of making the most out of the
1346 available memory. The memory consumption of workloads varies during
1347 runtime, and that requires users to overcommit. But doing that with a
1348 strict upper limit requires either a fairly accurate prediction of the
1349 working set size or adding slack to the limit. Since working set size
1350 estimation is hard and error prone, and getting it wrong results in
1351 OOM kills, most users tend to err on the side of a looser limit and
1352 end up wasting precious resources.
1354 The memory.high boundary on the other hand can be set much more
1355 conservatively. When hit, it throttles allocations by forcing them
1356 into direct reclaim to work off the excess, but it never invokes the
1357 OOM killer. As a result, a high boundary that is chosen too
1358 aggressively will not terminate the processes, but instead it will
1359 lead to gradual performance degradation. The user can monitor this
1360 and make corrections until the minimal memory footprint that still
1361 gives acceptable performance is found.
1363 In extreme cases, with many concurrent allocations and a complete
1364 breakdown of reclaim progress within the group, the high boundary can
1365 be exceeded. But even then it's mostly better to satisfy the
1366 allocation from the slack available in other groups or the rest of the
1367 system than killing the group. Otherwise, memory.max is there to
1368 limit this type of spillover and ultimately contain buggy or even
1369 malicious applications.
1371 The combined memory+swap accounting and limiting is replaced by real
1372 control over swap space.
1374 The main argument for a combined memory+swap facility in the original
1375 cgroup design was that global or parental pressure would always be
1376 able to swap all anonymous memory of a child group, regardless of the
1377 child's own (possibly untrusted) configuration. However, untrusted
1378 groups can sabotage swapping by other means - such as referencing its
1379 anonymous memory in a tight loop - and an admin can not assume full
1380 swappability when overcommitting untrusted jobs.
1382 For trusted jobs, on the other hand, a combined counter is not an
1383 intuitive userspace interface, and it flies in the face of the idea
1384 that cgroup controllers should account and limit specific physical
1385 resources. Swap space is a resource like all others in the system,
1386 and that's why unified hierarchy allows distributing it separately.