6 :Author: Tejun Heo <tj@kernel.org>
8 This is the authoritative documentation on the design, interface and
9 conventions of cgroup v2. It describes all userland-visible aspects
10 of cgroup including core and specific controller behaviors. All
11 future changes must be reflected in this document. Documentation for
12 v1 is available under Documentation/admin-guide/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.
192 Organizing Processes and Threads
193 --------------------------------
198 Initially, only the root cgroup exists to which all processes belong.
199 A child cgroup can be created by creating a sub-directory::
203 A given cgroup may have multiple child cgroups forming a tree
204 structure. Each cgroup has a read-writable interface file
205 "cgroup.procs". When read, it lists the PIDs of all processes which
206 belong to the cgroup one-per-line. The PIDs are not ordered and the
207 same PID may show up more than once if the process got moved to
208 another cgroup and then back or the PID got recycled while reading.
210 A process can be migrated into a cgroup by writing its PID to the
211 target cgroup's "cgroup.procs" file. Only one process can be migrated
212 on a single write(2) call. If a process is composed of multiple
213 threads, writing the PID of any thread migrates all threads of the
216 When a process forks a child process, the new process is born into the
217 cgroup that the forking process belongs to at the time of the
218 operation. After exit, a process stays associated with the cgroup
219 that it belonged to at the time of exit until it's reaped; however, a
220 zombie process does not appear in "cgroup.procs" and thus can't be
221 moved to another cgroup.
223 A cgroup which doesn't have any children or live processes can be
224 destroyed by removing the directory. Note that a cgroup which doesn't
225 have any children and is associated only with zombie processes is
226 considered empty and can be removed::
230 "/proc/$PID/cgroup" lists a process's cgroup membership. If legacy
231 cgroup is in use in the system, this file may contain multiple lines,
232 one for each hierarchy. The entry for cgroup v2 is always in the
235 # cat /proc/842/cgroup
237 0::/test-cgroup/test-cgroup-nested
239 If the process becomes a zombie and the cgroup it was associated with
240 is removed subsequently, " (deleted)" is appended to the path::
242 # cat /proc/842/cgroup
244 0::/test-cgroup/test-cgroup-nested (deleted)
250 cgroup v2 supports thread granularity for a subset of controllers to
251 support use cases requiring hierarchical resource distribution across
252 the threads of a group of processes. By default, all threads of a
253 process belong to the same cgroup, which also serves as the resource
254 domain to host resource consumptions which are not specific to a
255 process or thread. The thread mode allows threads to be spread across
256 a subtree while still maintaining the common resource domain for them.
258 Controllers which support thread mode are called threaded controllers.
259 The ones which don't are called domain controllers.
261 Marking a cgroup threaded makes it join the resource domain of its
262 parent as a threaded cgroup. The parent may be another threaded
263 cgroup whose resource domain is further up in the hierarchy. The root
264 of a threaded subtree, that is, the nearest ancestor which is not
265 threaded, is called threaded domain or thread root interchangeably and
266 serves as the resource domain for the entire subtree.
268 Inside a threaded subtree, threads of a process can be put in
269 different cgroups and are not subject to the no internal process
270 constraint - threaded controllers can be enabled on non-leaf cgroups
271 whether they have threads in them or not.
273 As the threaded domain cgroup hosts all the domain resource
274 consumptions of the subtree, it is considered to have internal
275 resource consumptions whether there are processes in it or not and
276 can't have populated child cgroups which aren't threaded. Because the
277 root cgroup is not subject to no internal process constraint, it can
278 serve both as a threaded domain and a parent to domain cgroups.
280 The current operation mode or type of the cgroup is shown in the
281 "cgroup.type" file which indicates whether the cgroup is a normal
282 domain, a domain which is serving as the domain of a threaded subtree,
283 or a threaded cgroup.
285 On creation, a cgroup is always a domain cgroup and can be made
286 threaded by writing "threaded" to the "cgroup.type" file. The
287 operation is single direction::
289 # echo threaded > cgroup.type
291 Once threaded, the cgroup can't be made a domain again. To enable the
292 thread mode, the following conditions must be met.
294 - As the cgroup will join the parent's resource domain. The parent
295 must either be a valid (threaded) domain or a threaded cgroup.
297 - When the parent is an unthreaded domain, it must not have any domain
298 controllers enabled or populated domain children. The root is
299 exempt from this requirement.
301 Topology-wise, a cgroup can be in an invalid state. Please consider
302 the following topology::
304 A (threaded domain) - B (threaded) - C (domain, just created)
306 C is created as a domain but isn't connected to a parent which can
307 host child domains. C can't be used until it is turned into a
308 threaded cgroup. "cgroup.type" file will report "domain (invalid)" in
309 these cases. Operations which fail due to invalid topology use
310 EOPNOTSUPP as the errno.
312 A domain cgroup is turned into a threaded domain when one of its child
313 cgroup becomes threaded or threaded controllers are enabled in the
314 "cgroup.subtree_control" file while there are processes in the cgroup.
315 A threaded domain reverts to a normal domain when the conditions
318 When read, "cgroup.threads" contains the list of the thread IDs of all
319 threads in the cgroup. Except that the operations are per-thread
320 instead of per-process, "cgroup.threads" has the same format and
321 behaves the same way as "cgroup.procs". While "cgroup.threads" can be
322 written to in any cgroup, as it can only move threads inside the same
323 threaded domain, its operations are confined inside each threaded
326 The threaded domain cgroup serves as the resource domain for the whole
327 subtree, and, while the threads can be scattered across the subtree,
328 all the processes are considered to be in the threaded domain cgroup.
329 "cgroup.procs" in a threaded domain cgroup contains the PIDs of all
330 processes in the subtree and is not readable in the subtree proper.
331 However, "cgroup.procs" can be written to from anywhere in the subtree
332 to migrate all threads of the matching process to the cgroup.
334 Only threaded controllers can be enabled in a threaded subtree. When
335 a threaded controller is enabled inside a threaded subtree, it only
336 accounts for and controls resource consumptions associated with the
337 threads in the cgroup and its descendants. All consumptions which
338 aren't tied to a specific thread belong to the threaded domain cgroup.
340 Because a threaded subtree is exempt from no internal process
341 constraint, a threaded controller must be able to handle competition
342 between threads in a non-leaf cgroup and its child cgroups. Each
343 threaded controller defines how such competitions are handled.
346 [Un]populated Notification
347 --------------------------
349 Each non-root cgroup has a "cgroup.events" file which contains
350 "populated" field indicating whether the cgroup's sub-hierarchy has
351 live processes in it. Its value is 0 if there is no live process in
352 the cgroup and its descendants; otherwise, 1. poll and [id]notify
353 events are triggered when the value changes. This can be used, for
354 example, to start a clean-up operation after all processes of a given
355 sub-hierarchy have exited. The populated state updates and
356 notifications are recursive. Consider the following sub-hierarchy
357 where the numbers in the parentheses represent the numbers of processes
363 A, B and C's "populated" fields would be 1 while D's 0. After the one
364 process in C exits, B and C's "populated" fields would flip to "0" and
365 file modified events will be generated on the "cgroup.events" files of
369 Controlling Controllers
370 -----------------------
372 Enabling and Disabling
373 ~~~~~~~~~~~~~~~~~~~~~~
375 Each cgroup has a "cgroup.controllers" file which lists all
376 controllers available for the cgroup to enable::
378 # cat cgroup.controllers
381 No controller is enabled by default. Controllers can be enabled and
382 disabled by writing to the "cgroup.subtree_control" file::
384 # echo "+cpu +memory -io" > cgroup.subtree_control
386 Only controllers which are listed in "cgroup.controllers" can be
387 enabled. When multiple operations are specified as above, either they
388 all succeed or fail. If multiple operations on the same controller
389 are specified, the last one is effective.
391 Enabling a controller in a cgroup indicates that the distribution of
392 the target resource across its immediate children will be controlled.
393 Consider the following sub-hierarchy. The enabled controllers are
394 listed in parentheses::
396 A(cpu,memory) - B(memory) - C()
399 As A has "cpu" and "memory" enabled, A will control the distribution
400 of CPU cycles and memory to its children, in this case, B. As B has
401 "memory" enabled but not "CPU", C and D will compete freely on CPU
402 cycles but their division of memory available to B will be controlled.
404 As a controller regulates the distribution of the target resource to
405 the cgroup's children, enabling it creates the controller's interface
406 files in the child cgroups. In the above example, enabling "cpu" on B
407 would create the "cpu." prefixed controller interface files in C and
408 D. Likewise, disabling "memory" from B would remove the "memory."
409 prefixed controller interface files from C and D. This means that the
410 controller interface files - anything which doesn't start with
411 "cgroup." are owned by the parent rather than the cgroup itself.
417 Resources are distributed top-down and a cgroup can further distribute
418 a resource only if the resource has been distributed to it from the
419 parent. This means that all non-root "cgroup.subtree_control" files
420 can only contain controllers which are enabled in the parent's
421 "cgroup.subtree_control" file. A controller can be enabled only if
422 the parent has the controller enabled and a controller can't be
423 disabled if one or more children have it enabled.
426 No Internal Process Constraint
427 ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
429 Non-root cgroups can distribute domain resources to their children
430 only when they don't have any processes of their own. In other words,
431 only domain cgroups which don't contain any processes can have domain
432 controllers enabled in their "cgroup.subtree_control" files.
434 This guarantees that, when a domain controller is looking at the part
435 of the hierarchy which has it enabled, processes are always only on
436 the leaves. This rules out situations where child cgroups compete
437 against internal processes of the parent.
439 The root cgroup is exempt from this restriction. Root contains
440 processes and anonymous resource consumption which can't be associated
441 with any other cgroups and requires special treatment from most
442 controllers. How resource consumption in the root cgroup is governed
443 is up to each controller (for more information on this topic please
444 refer to the Non-normative information section in the Controllers
447 Note that the restriction doesn't get in the way if there is no
448 enabled controller in the cgroup's "cgroup.subtree_control". This is
449 important as otherwise it wouldn't be possible to create children of a
450 populated cgroup. To control resource distribution of a cgroup, the
451 cgroup must create children and transfer all its processes to the
452 children before enabling controllers in its "cgroup.subtree_control"
462 A cgroup can be delegated in two ways. First, to a less privileged
463 user by granting write access of the directory and its "cgroup.procs",
464 "cgroup.threads" and "cgroup.subtree_control" files to the user.
465 Second, if the "nsdelegate" mount option is set, automatically to a
466 cgroup namespace on namespace creation.
468 Because the resource control interface files in a given directory
469 control the distribution of the parent's resources, the delegatee
470 shouldn't be allowed to write to them. For the first method, this is
471 achieved by not granting access to these files. For the second, the
472 kernel rejects writes to all files other than "cgroup.procs" and
473 "cgroup.subtree_control" on a namespace root from inside the
476 The end results are equivalent for both delegation types. Once
477 delegated, the user can build sub-hierarchy under the directory,
478 organize processes inside it as it sees fit and further distribute the
479 resources it received from the parent. The limits and other settings
480 of all resource controllers are hierarchical and regardless of what
481 happens in the delegated sub-hierarchy, nothing can escape the
482 resource restrictions imposed by the parent.
484 Currently, cgroup doesn't impose any restrictions on the number of
485 cgroups in or nesting depth of a delegated sub-hierarchy; however,
486 this may be limited explicitly in the future.
489 Delegation Containment
490 ~~~~~~~~~~~~~~~~~~~~~~
492 A delegated sub-hierarchy is contained in the sense that processes
493 can't be moved into or out of the sub-hierarchy by the delegatee.
495 For delegations to a less privileged user, this is achieved by
496 requiring the following conditions for a process with a non-root euid
497 to migrate a target process into a cgroup by writing its PID to the
500 - The writer must have write access to the "cgroup.procs" file.
502 - The writer must have write access to the "cgroup.procs" file of the
503 common ancestor of the source and destination cgroups.
505 The above two constraints ensure that while a delegatee may migrate
506 processes around freely in the delegated sub-hierarchy it can't pull
507 in from or push out to outside the sub-hierarchy.
509 For an example, let's assume cgroups C0 and C1 have been delegated to
510 user U0 who created C00, C01 under C0 and C10 under C1 as follows and
511 all processes under C0 and C1 belong to U0::
513 ~~~~~~~~~~~~~ - C0 - C00
516 ~~~~~~~~~~~~~ - C1 - C10
518 Let's also say U0 wants to write the PID of a process which is
519 currently in C10 into "C00/cgroup.procs". U0 has write access to the
520 file; however, the common ancestor of the source cgroup C10 and the
521 destination cgroup C00 is above the points of delegation and U0 would
522 not have write access to its "cgroup.procs" files and thus the write
523 will be denied with -EACCES.
525 For delegations to namespaces, containment is achieved by requiring
526 that both the source and destination cgroups are reachable from the
527 namespace of the process which is attempting the migration. If either
528 is not reachable, the migration is rejected with -ENOENT.
534 Organize Once and Control
535 ~~~~~~~~~~~~~~~~~~~~~~~~~
537 Migrating a process across cgroups is a relatively expensive operation
538 and stateful resources such as memory are not moved together with the
539 process. This is an explicit design decision as there often exist
540 inherent trade-offs between migration and various hot paths in terms
541 of synchronization cost.
543 As such, migrating processes across cgroups frequently as a means to
544 apply different resource restrictions is discouraged. A workload
545 should be assigned to a cgroup according to the system's logical and
546 resource structure once on start-up. Dynamic adjustments to resource
547 distribution can be made by changing controller configuration through
551 Avoid Name Collisions
552 ~~~~~~~~~~~~~~~~~~~~~
554 Interface files for a cgroup and its children cgroups occupy the same
555 directory and it is possible to create children cgroups which collide
556 with interface files.
558 All cgroup core interface files are prefixed with "cgroup." and each
559 controller's interface files are prefixed with the controller name and
560 a dot. A controller's name is composed of lower case alphabets and
561 '_'s but never begins with an '_' so it can be used as the prefix
562 character for collision avoidance. Also, interface file names won't
563 start or end with terms which are often used in categorizing workloads
564 such as job, service, slice, unit or workload.
566 cgroup doesn't do anything to prevent name collisions and it's the
567 user's responsibility to avoid them.
570 Resource Distribution Models
571 ============================
573 cgroup controllers implement several resource distribution schemes
574 depending on the resource type and expected use cases. This section
575 describes major schemes in use along with their expected behaviors.
581 A parent's resource is distributed by adding up the weights of all
582 active children and giving each the fraction matching the ratio of its
583 weight against the sum. As only children which can make use of the
584 resource at the moment participate in the distribution, this is
585 work-conserving. Due to the dynamic nature, this model is usually
586 used for stateless resources.
588 All weights are in the range [1, 10000] with the default at 100. This
589 allows symmetric multiplicative biases in both directions at fine
590 enough granularity while staying in the intuitive range.
592 As long as the weight is in range, all configuration combinations are
593 valid and there is no reason to reject configuration changes or
596 "cpu.weight" proportionally distributes CPU cycles to active children
597 and is an example of this type.
603 A child can only consume upto the configured amount of the resource.
604 Limits can be over-committed - the sum of the limits of children can
605 exceed the amount of resource available to the parent.
607 Limits are in the range [0, max] and defaults to "max", which is noop.
609 As limits can be over-committed, all configuration combinations are
610 valid and there is no reason to reject configuration changes or
613 "io.max" limits the maximum BPS and/or IOPS that a cgroup can consume
614 on an IO device and is an example of this type.
620 A cgroup is protected upto the configured amount of the resource
621 as long as the usages of all its ancestors are under their
622 protected levels. Protections can be hard guarantees or best effort
623 soft boundaries. Protections can also be over-committed in which case
624 only upto the amount available to the parent is protected among
627 Protections are in the range [0, max] and defaults to 0, which is
630 As protections can be over-committed, all configuration combinations
631 are valid and there is no reason to reject configuration changes or
634 "memory.low" implements best-effort memory protection and is an
635 example of this type.
641 A cgroup is exclusively allocated a certain amount of a finite
642 resource. Allocations can't be over-committed - the sum of the
643 allocations of children can not exceed the amount of resource
644 available to the parent.
646 Allocations are in the range [0, max] and defaults to 0, which is no
649 As allocations can't be over-committed, some configuration
650 combinations are invalid and should be rejected. Also, if the
651 resource is mandatory for execution of processes, process migrations
654 "cpu.rt.max" hard-allocates realtime slices and is an example of this
664 All interface files should be in one of the following formats whenever
667 New-line separated values
668 (when only one value can be written at once)
674 Space separated values
675 (when read-only or multiple values can be written at once)
687 KEY0 SUB_KEY0=VAL00 SUB_KEY1=VAL01...
688 KEY1 SUB_KEY0=VAL10 SUB_KEY1=VAL11...
691 For a writable file, the format for writing should generally match
692 reading; however, controllers may allow omitting later fields or
693 implement restricted shortcuts for most common use cases.
695 For both flat and nested keyed files, only the values for a single key
696 can be written at a time. For nested keyed files, the sub key pairs
697 may be specified in any order and not all pairs have to be specified.
703 - Settings for a single feature should be contained in a single file.
705 - The root cgroup should be exempt from resource control and thus
706 shouldn't have resource control interface files. Also,
707 informational files on the root cgroup which end up showing global
708 information available elsewhere shouldn't exist.
710 - The default time unit is microseconds. If a different unit is ever
711 used, an explicit unit suffix must be present.
713 - A parts-per quantity should use a percentage decimal with at least
714 two digit fractional part - e.g. 13.40.
716 - If a controller implements weight based resource distribution, its
717 interface file should be named "weight" and have the range [1,
718 10000] with 100 as the default. The values are chosen to allow
719 enough and symmetric bias in both directions while keeping it
720 intuitive (the default is 100%).
722 - If a controller implements an absolute resource guarantee and/or
723 limit, the interface files should be named "min" and "max"
724 respectively. If a controller implements best effort resource
725 guarantee and/or limit, the interface files should be named "low"
726 and "high" respectively.
728 In the above four control files, the special token "max" should be
729 used to represent upward infinity for both reading and writing.
731 - If a setting has a configurable default value and keyed specific
732 overrides, the default entry should be keyed with "default" and
733 appear as the first entry in the file.
735 The default value can be updated by writing either "default $VAL" or
738 When writing to update a specific override, "default" can be used as
739 the value to indicate removal of the override. Override entries
740 with "default" as the value must not appear when read.
742 For example, a setting which is keyed by major:minor device numbers
743 with integer values may look like the following::
745 # cat cgroup-example-interface-file
749 The default value can be updated by::
751 # echo 125 > cgroup-example-interface-file
755 # echo "default 125" > cgroup-example-interface-file
757 An override can be set by::
759 # echo "8:16 170" > cgroup-example-interface-file
763 # echo "8:0 default" > cgroup-example-interface-file
764 # cat cgroup-example-interface-file
768 - For events which are not very high frequency, an interface file
769 "events" should be created which lists event key value pairs.
770 Whenever a notifiable event happens, file modified event should be
771 generated on the file.
777 All cgroup core files are prefixed with "cgroup."
781 A read-write single value file which exists on non-root
784 When read, it indicates the current type of the cgroup, which
785 can be one of the following values.
787 - "domain" : A normal valid domain cgroup.
789 - "domain threaded" : A threaded domain cgroup which is
790 serving as the root of a threaded subtree.
792 - "domain invalid" : A cgroup which is in an invalid state.
793 It can't be populated or have controllers enabled. It may
794 be allowed to become a threaded cgroup.
796 - "threaded" : A threaded cgroup which is a member of a
799 A cgroup can be turned into a threaded cgroup by writing
800 "threaded" to this file.
803 A read-write new-line separated values file which exists on
806 When read, it lists the PIDs of all processes which belong to
807 the cgroup one-per-line. The PIDs are not ordered and the
808 same PID may show up more than once if the process got moved
809 to another cgroup and then back or the PID got recycled while
812 A PID can be written to migrate the process associated with
813 the PID to the cgroup. The writer should match all of the
814 following conditions.
816 - It must have write access to the "cgroup.procs" file.
818 - It must have write access to the "cgroup.procs" file of the
819 common ancestor of the source and destination cgroups.
821 When delegating a sub-hierarchy, write access to this file
822 should be granted along with the containing directory.
824 In a threaded cgroup, reading this file fails with EOPNOTSUPP
825 as all the processes belong to the thread root. Writing is
826 supported and moves every thread of the process to the cgroup.
829 A read-write new-line separated values file which exists on
832 When read, it lists the TIDs of all threads which belong to
833 the cgroup one-per-line. The TIDs are not ordered and the
834 same TID may show up more than once if the thread got moved to
835 another cgroup and then back or the TID got recycled while
838 A TID can be written to migrate the thread associated with the
839 TID to the cgroup. The writer should match all of the
840 following conditions.
842 - It must have write access to the "cgroup.threads" file.
844 - The cgroup that the thread is currently in must be in the
845 same resource domain as the destination cgroup.
847 - It must have write access to the "cgroup.procs" file of the
848 common ancestor of the source and destination cgroups.
850 When delegating a sub-hierarchy, write access to this file
851 should be granted along with the containing directory.
854 A read-only space separated values file which exists on all
857 It shows space separated list of all controllers available to
858 the cgroup. The controllers are not ordered.
860 cgroup.subtree_control
861 A read-write space separated values file which exists on all
862 cgroups. Starts out empty.
864 When read, it shows space separated list of the controllers
865 which are enabled to control resource distribution from the
866 cgroup to its children.
868 Space separated list of controllers prefixed with '+' or '-'
869 can be written to enable or disable controllers. A controller
870 name prefixed with '+' enables the controller and '-'
871 disables. If a controller appears more than once on the list,
872 the last one is effective. When multiple enable and disable
873 operations are specified, either all succeed or all fail.
876 A read-only flat-keyed file which exists on non-root cgroups.
877 The following entries are defined. Unless specified
878 otherwise, a value change in this file generates a file
882 1 if the cgroup or its descendants contains any live
883 processes; otherwise, 0.
885 1 if the cgroup is frozen; otherwise, 0.
887 cgroup.max.descendants
888 A read-write single value files. The default is "max".
890 Maximum allowed number of descent cgroups.
891 If the actual number of descendants is equal or larger,
892 an attempt to create a new cgroup in the hierarchy will fail.
895 A read-write single value files. The default is "max".
897 Maximum allowed descent depth below the current cgroup.
898 If the actual descent depth is equal or larger,
899 an attempt to create a new child cgroup will fail.
902 A read-only flat-keyed file with the following entries:
905 Total number of visible descendant cgroups.
908 Total number of dying descendant cgroups. A cgroup becomes
909 dying after being deleted by a user. The cgroup will remain
910 in dying state for some time undefined time (which can depend
911 on system load) before being completely destroyed.
913 A process can't enter a dying cgroup under any circumstances,
914 a dying cgroup can't revive.
916 A dying cgroup can consume system resources not exceeding
917 limits, which were active at the moment of cgroup deletion.
920 A read-write single value file which exists on non-root cgroups.
921 Allowed values are "0" and "1". The default is "0".
923 Writing "1" to the file causes freezing of the cgroup and all
924 descendant cgroups. This means that all belonging processes will
925 be stopped and will not run until the cgroup will be explicitly
926 unfrozen. Freezing of the cgroup may take some time; when this action
927 is completed, the "frozen" value in the cgroup.events control file
928 will be updated to "1" and the corresponding notification will be
931 A cgroup can be frozen either by its own settings, or by settings
932 of any ancestor cgroups. If any of ancestor cgroups is frozen, the
933 cgroup will remain frozen.
935 Processes in the frozen cgroup can be killed by a fatal signal.
936 They also can enter and leave a frozen cgroup: either by an explicit
937 move by a user, or if freezing of the cgroup races with fork().
938 If a process is moved to a frozen cgroup, it stops. If a process is
939 moved out of a frozen cgroup, it becomes running.
941 Frozen status of a cgroup doesn't affect any cgroup tree operations:
942 it's possible to delete a frozen (and empty) cgroup, as well as
943 create new sub-cgroups.
951 The "cpu" controllers regulates distribution of CPU cycles. This
952 controller implements weight and absolute bandwidth limit models for
953 normal scheduling policy and absolute bandwidth allocation model for
954 realtime scheduling policy.
956 In all the above models, cycles distribution is defined only on a temporal
957 base and it does not account for the frequency at which tasks are executed.
958 The (optional) utilization clamping support allows to hint the schedutil
959 cpufreq governor about the minimum desired frequency which should always be
960 provided by a CPU, as well as the maximum desired frequency, which should not
961 be exceeded by a CPU.
963 WARNING: cgroup2 doesn't yet support control of realtime processes and
964 the cpu controller can only be enabled when all RT processes are in
965 the root cgroup. Be aware that system management software may already
966 have placed RT processes into nonroot cgroups during the system boot
967 process, and these processes may need to be moved to the root cgroup
968 before the cpu controller can be enabled.
974 All time durations are in microseconds.
977 A read-only flat-keyed file which exists on non-root cgroups.
978 This file exists whether the controller is enabled or not.
980 It always reports the following three stats:
986 and the following three when the controller is enabled:
993 A read-write single value file which exists on non-root
994 cgroups. The default is "100".
996 The weight in the range [1, 10000].
999 A read-write single value file which exists on non-root
1000 cgroups. The default is "0".
1002 The nice value is in the range [-20, 19].
1004 This interface file is an alternative interface for
1005 "cpu.weight" and allows reading and setting weight using the
1006 same values used by nice(2). Because the range is smaller and
1007 granularity is coarser for the nice values, the read value is
1008 the closest approximation of the current weight.
1011 A read-write two value file which exists on non-root cgroups.
1012 The default is "max 100000".
1014 The maximum bandwidth limit. It's in the following format::
1018 which indicates that the group may consume upto $MAX in each
1019 $PERIOD duration. "max" for $MAX indicates no limit. If only
1020 one number is written, $MAX is updated.
1023 A read-only nested-key file which exists on non-root cgroups.
1025 Shows pressure stall information for CPU. See
1026 Documentation/accounting/psi.rst for details.
1029 A read-write single value file which exists on non-root cgroups.
1030 The default is "0", i.e. no utilization boosting.
1032 The requested minimum utilization (protection) as a percentage
1033 rational number, e.g. 12.34 for 12.34%.
1035 This interface allows reading and setting minimum utilization clamp
1036 values similar to the sched_setattr(2). This minimum utilization
1037 value is used to clamp the task specific minimum utilization clamp.
1039 The requested minimum utilization (protection) is always capped by
1040 the current value for the maximum utilization (limit), i.e.
1044 A read-write single value file which exists on non-root cgroups.
1045 The default is "max". i.e. no utilization capping
1047 The requested maximum utilization (limit) as a percentage rational
1048 number, e.g. 98.76 for 98.76%.
1050 This interface allows reading and setting maximum utilization clamp
1051 values similar to the sched_setattr(2). This maximum utilization
1052 value is used to clamp the task specific maximum utilization clamp.
1059 The "memory" controller regulates distribution of memory. Memory is
1060 stateful and implements both limit and protection models. Due to the
1061 intertwining between memory usage and reclaim pressure and the
1062 stateful nature of memory, the distribution model is relatively
1065 While not completely water-tight, all major memory usages by a given
1066 cgroup are tracked so that the total memory consumption can be
1067 accounted and controlled to a reasonable extent. Currently, the
1068 following types of memory usages are tracked.
1070 - Userland memory - page cache and anonymous memory.
1072 - Kernel data structures such as dentries and inodes.
1074 - TCP socket buffers.
1076 The above list may expand in the future for better coverage.
1079 Memory Interface Files
1080 ~~~~~~~~~~~~~~~~~~~~~~
1082 All memory amounts are in bytes. If a value which is not aligned to
1083 PAGE_SIZE is written, the value may be rounded up to the closest
1084 PAGE_SIZE multiple when read back.
1087 A read-only single value file which exists on non-root
1090 The total amount of memory currently being used by the cgroup
1091 and its descendants.
1094 A read-write single value file which exists on non-root
1095 cgroups. The default is "0".
1097 Hard memory protection. If the memory usage of a cgroup
1098 is within its effective min boundary, the cgroup's memory
1099 won't be reclaimed under any conditions. If there is no
1100 unprotected reclaimable memory available, OOM killer
1101 is invoked. Above the effective min boundary (or
1102 effective low boundary if it is higher), pages are reclaimed
1103 proportionally to the overage, reducing reclaim pressure for
1106 Effective min boundary is limited by memory.min values of
1107 all ancestor cgroups. If there is memory.min overcommitment
1108 (child cgroup or cgroups are requiring more protected memory
1109 than parent will allow), then each child cgroup will get
1110 the part of parent's protection proportional to its
1111 actual memory usage below memory.min.
1113 Putting more memory than generally available under this
1114 protection is discouraged and may lead to constant OOMs.
1116 If a memory cgroup is not populated with processes,
1117 its memory.min is ignored.
1120 A read-write single value file which exists on non-root
1121 cgroups. The default is "0".
1123 Best-effort memory protection. If the memory usage of a
1124 cgroup is within its effective low boundary, the cgroup's
1125 memory won't be reclaimed unless there is no reclaimable
1126 memory available in unprotected cgroups.
1127 Above the effective low boundary (or
1128 effective min boundary if it is higher), pages are reclaimed
1129 proportionally to the overage, reducing reclaim pressure for
1132 Effective low boundary is limited by memory.low values of
1133 all ancestor cgroups. If there is memory.low overcommitment
1134 (child cgroup or cgroups are requiring more protected memory
1135 than parent will allow), then each child cgroup will get
1136 the part of parent's protection proportional to its
1137 actual memory usage below memory.low.
1139 Putting more memory than generally available under this
1140 protection is discouraged.
1143 A read-write single value file which exists on non-root
1144 cgroups. The default is "max".
1146 Memory usage throttle limit. This is the main mechanism to
1147 control memory usage of a cgroup. If a cgroup's usage goes
1148 over the high boundary, the processes of the cgroup are
1149 throttled and put under heavy reclaim pressure.
1151 Going over the high limit never invokes the OOM killer and
1152 under extreme conditions the limit may be breached.
1155 A read-write single value file which exists on non-root
1156 cgroups. The default is "max".
1158 Memory usage hard limit. This is the final protection
1159 mechanism. If a cgroup's memory usage reaches this limit and
1160 can't be reduced, the OOM killer is invoked in the cgroup.
1161 Under certain circumstances, the usage may go over the limit
1164 This is the ultimate protection mechanism. As long as the
1165 high limit is used and monitored properly, this limit's
1166 utility is limited to providing the final safety net.
1169 A read-write single value file which exists on non-root
1170 cgroups. The default value is "0".
1172 Determines whether the cgroup should be treated as
1173 an indivisible workload by the OOM killer. If set,
1174 all tasks belonging to the cgroup or to its descendants
1175 (if the memory cgroup is not a leaf cgroup) are killed
1176 together or not at all. This can be used to avoid
1177 partial kills to guarantee workload integrity.
1179 Tasks with the OOM protection (oom_score_adj set to -1000)
1180 are treated as an exception and are never killed.
1182 If the OOM killer is invoked in a cgroup, it's not going
1183 to kill any tasks outside of this cgroup, regardless
1184 memory.oom.group values of ancestor cgroups.
1187 A read-only flat-keyed file which exists on non-root cgroups.
1188 The following entries are defined. Unless specified
1189 otherwise, a value change in this file generates a file
1192 Note that all fields in this file are hierarchical and the
1193 file modified event can be generated due to an event down the
1194 hierarchy. For for the local events at the cgroup level see
1195 memory.events.local.
1198 The number of times the cgroup is reclaimed due to
1199 high memory pressure even though its usage is under
1200 the low boundary. This usually indicates that the low
1201 boundary is over-committed.
1204 The number of times processes of the cgroup are
1205 throttled and routed to perform direct memory reclaim
1206 because the high memory boundary was exceeded. For a
1207 cgroup whose memory usage is capped by the high limit
1208 rather than global memory pressure, this event's
1209 occurrences are expected.
1212 The number of times the cgroup's memory usage was
1213 about to go over the max boundary. If direct reclaim
1214 fails to bring it down, the cgroup goes to OOM state.
1217 The number of time the cgroup's memory usage was
1218 reached the limit and allocation was about to fail.
1220 Depending on context result could be invocation of OOM
1221 killer and retrying allocation or failing allocation.
1223 Failed allocation in its turn could be returned into
1224 userspace as -ENOMEM or silently ignored in cases like
1225 disk readahead. For now OOM in memory cgroup kills
1226 tasks iff shortage has happened inside page fault.
1228 This event is not raised if the OOM killer is not
1229 considered as an option, e.g. for failed high-order
1233 The number of processes belonging to this cgroup
1234 killed by any kind of OOM killer.
1237 Similar to memory.events but the fields in the file are local
1238 to the cgroup i.e. not hierarchical. The file modified event
1239 generated on this file reflects only the local events.
1242 A read-only flat-keyed file which exists on non-root cgroups.
1244 This breaks down the cgroup's memory footprint into different
1245 types of memory, type-specific details, and other information
1246 on the state and past events of the memory management system.
1248 All memory amounts are in bytes.
1250 The entries are ordered to be human readable, and new entries
1251 can show up in the middle. Don't rely on items remaining in a
1252 fixed position; use the keys to look up specific values!
1255 Amount of memory used in anonymous mappings such as
1256 brk(), sbrk(), and mmap(MAP_ANONYMOUS)
1259 Amount of memory used to cache filesystem data,
1260 including tmpfs and shared memory.
1263 Amount of memory allocated to kernel stacks.
1266 Amount of memory used for storing in-kernel data
1270 Amount of memory used in network transmission buffers
1273 Amount of cached filesystem data that is swap-backed,
1274 such as tmpfs, shm segments, shared anonymous mmap()s
1277 Amount of cached filesystem data mapped with mmap()
1280 Amount of cached filesystem data that was modified but
1281 not yet written back to disk
1284 Amount of cached filesystem data that was modified and
1285 is currently being written back to disk
1288 Amount of memory used in anonymous mappings backed by
1289 transparent hugepages
1291 inactive_anon, active_anon, inactive_file, active_file, unevictable
1292 Amount of memory, swap-backed and filesystem-backed,
1293 on the internal memory management lists used by the
1294 page reclaim algorithm.
1296 As these represent internal list state (eg. shmem pages are on anon
1297 memory management lists), inactive_foo + active_foo may not be equal to
1298 the value for the foo counter, since the foo counter is type-based, not
1302 Part of "slab" that might be reclaimed, such as
1303 dentries and inodes.
1306 Part of "slab" that cannot be reclaimed on memory
1310 Total number of page faults incurred
1313 Number of major page faults incurred
1317 Number of refaults of previously evicted pages
1321 Number of refaulted pages that were immediately activated
1323 workingset_nodereclaim
1325 Number of times a shadow node has been reclaimed
1329 Amount of scanned pages (in an active LRU list)
1333 Amount of scanned pages (in an inactive LRU list)
1337 Amount of reclaimed pages
1341 Amount of pages moved to the active LRU list
1345 Amount of pages moved to the inactive LRU list
1349 Amount of pages postponed to be freed under memory pressure
1353 Amount of reclaimed lazyfree pages
1357 Number of transparent hugepages which were allocated to satisfy
1358 a page fault, including COW faults. This counter is not present
1359 when CONFIG_TRANSPARENT_HUGEPAGE is not set.
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 hard limit. If a cgroup's swap usage reaches this
1379 limit, anonymous memory of the cgroup will not be swapped out.
1382 A read-only flat-keyed file which exists on non-root cgroups.
1383 The following entries are defined. Unless specified
1384 otherwise, a value change in this file generates a file
1388 The number of times the cgroup's swap usage was about
1389 to go over the max boundary and swap allocation
1393 The number of times swap allocation failed either
1394 because of running out of swap system-wide or max
1397 When reduced under the current usage, the existing swap
1398 entries are reclaimed gradually and the swap usage may stay
1399 higher than the limit for an extended period of time. This
1400 reduces the impact on the workload and memory management.
1403 A read-only nested-key file which exists on non-root cgroups.
1405 Shows pressure stall information for memory. See
1406 Documentation/accounting/psi.rst for details.
1412 "memory.high" is the main mechanism to control memory usage.
1413 Over-committing on high limit (sum of high limits > available memory)
1414 and letting global memory pressure to distribute memory according to
1415 usage is a viable strategy.
1417 Because breach of the high limit doesn't trigger the OOM killer but
1418 throttles the offending cgroup, a management agent has ample
1419 opportunities to monitor and take appropriate actions such as granting
1420 more memory or terminating the workload.
1422 Determining whether a cgroup has enough memory is not trivial as
1423 memory usage doesn't indicate whether the workload can benefit from
1424 more memory. For example, a workload which writes data received from
1425 network to a file can use all available memory but can also operate as
1426 performant with a small amount of memory. A measure of memory
1427 pressure - how much the workload is being impacted due to lack of
1428 memory - is necessary to determine whether a workload needs more
1429 memory; unfortunately, memory pressure monitoring mechanism isn't
1436 A memory area is charged to the cgroup which instantiated it and stays
1437 charged to the cgroup until the area is released. Migrating a process
1438 to a different cgroup doesn't move the memory usages that it
1439 instantiated while in the previous cgroup to the new cgroup.
1441 A memory area may be used by processes belonging to different cgroups.
1442 To which cgroup the area will be charged is in-deterministic; however,
1443 over time, the memory area is likely to end up in a cgroup which has
1444 enough memory allowance to avoid high reclaim pressure.
1446 If a cgroup sweeps a considerable amount of memory which is expected
1447 to be accessed repeatedly by other cgroups, it may make sense to use
1448 POSIX_FADV_DONTNEED to relinquish the ownership of memory areas
1449 belonging to the affected files to ensure correct memory ownership.
1455 The "io" controller regulates the distribution of IO resources. This
1456 controller implements both weight based and absolute bandwidth or IOPS
1457 limit distribution; however, weight based distribution is available
1458 only if cfq-iosched is in use and neither scheme is available for
1466 A read-only nested-keyed file which exists on non-root
1469 Lines are keyed by $MAJ:$MIN device numbers and not ordered.
1470 The following nested keys are defined.
1472 ====== =====================
1474 wbytes Bytes written
1475 rios Number of read IOs
1476 wios Number of write IOs
1477 dbytes Bytes discarded
1478 dios Number of discard IOs
1479 ====== =====================
1481 An example read output follows:
1483 8:16 rbytes=1459200 wbytes=314773504 rios=192 wios=353 dbytes=0 dios=0
1484 8:0 rbytes=90430464 wbytes=299008000 rios=8950 wios=1252 dbytes=50331648 dios=3021
1487 A read-write nested-keyed file with exists only on the root
1490 This file configures the Quality of Service of the IO cost
1491 model based controller (CONFIG_BLK_CGROUP_IOCOST) which
1492 currently implements "io.weight" proportional control. Lines
1493 are keyed by $MAJ:$MIN device numbers and not ordered. The
1494 line for a given device is populated on the first write for
1495 the device on "io.cost.qos" or "io.cost.model". The following
1496 nested keys are defined.
1498 ====== =====================================
1499 enable Weight-based control enable
1500 ctrl "auto" or "user"
1501 rpct Read latency percentile [0, 100]
1502 rlat Read latency threshold
1503 wpct Write latency percentile [0, 100]
1504 wlat Write latency threshold
1505 min Minimum scaling percentage [1, 10000]
1506 max Maximum scaling percentage [1, 10000]
1507 ====== =====================================
1509 The controller is disabled by default and can be enabled by
1510 setting "enable" to 1. "rpct" and "wpct" parameters default
1511 to zero and the controller uses internal device saturation
1512 state to adjust the overall IO rate between "min" and "max".
1514 When a better control quality is needed, latency QoS
1515 parameters can be configured. For example::
1517 8:16 enable=1 ctrl=auto rpct=95.00 rlat=75000 wpct=95.00 wlat=150000 min=50.00 max=150.0
1519 shows that on sdb, the controller is enabled, will consider
1520 the device saturated if the 95th percentile of read completion
1521 latencies is above 75ms or write 150ms, and adjust the overall
1522 IO issue rate between 50% and 150% accordingly.
1524 The lower the saturation point, the better the latency QoS at
1525 the cost of aggregate bandwidth. The narrower the allowed
1526 adjustment range between "min" and "max", the more conformant
1527 to the cost model the IO behavior. Note that the IO issue
1528 base rate may be far off from 100% and setting "min" and "max"
1529 blindly can lead to a significant loss of device capacity or
1530 control quality. "min" and "max" are useful for regulating
1531 devices which show wide temporary behavior changes - e.g. a
1532 ssd which accepts writes at the line speed for a while and
1533 then completely stalls for multiple seconds.
1535 When "ctrl" is "auto", the parameters are controlled by the
1536 kernel and may change automatically. Setting "ctrl" to "user"
1537 or setting any of the percentile and latency parameters puts
1538 it into "user" mode and disables the automatic changes. The
1539 automatic mode can be restored by setting "ctrl" to "auto".
1542 A read-write nested-keyed file with exists only on the root
1545 This file configures the cost model of the IO cost model based
1546 controller (CONFIG_BLK_CGROUP_IOCOST) which currently
1547 implements "io.weight" proportional control. Lines are keyed
1548 by $MAJ:$MIN device numbers and not ordered. The line for a
1549 given device is populated on the first write for the device on
1550 "io.cost.qos" or "io.cost.model". The following nested keys
1553 ===== ================================
1554 ctrl "auto" or "user"
1555 model The cost model in use - "linear"
1556 ===== ================================
1558 When "ctrl" is "auto", the kernel may change all parameters
1559 dynamically. When "ctrl" is set to "user" or any other
1560 parameters are written to, "ctrl" become "user" and the
1561 automatic changes are disabled.
1563 When "model" is "linear", the following model parameters are
1566 ============= ========================================
1567 [r|w]bps The maximum sequential IO throughput
1568 [r|w]seqiops The maximum 4k sequential IOs per second
1569 [r|w]randiops The maximum 4k random IOs per second
1570 ============= ========================================
1572 From the above, the builtin linear model determines the base
1573 costs of a sequential and random IO and the cost coefficient
1574 for the IO size. While simple, this model can cover most
1575 common device classes acceptably.
1577 The IO cost model isn't expected to be accurate in absolute
1578 sense and is scaled to the device behavior dynamically.
1580 If needed, tools/cgroup/iocost_coef_gen.py can be used to
1581 generate device-specific coefficients.
1584 A read-write flat-keyed file which exists on non-root cgroups.
1585 The default is "default 100".
1587 The first line is the default weight applied to devices
1588 without specific override. The rest are overrides keyed by
1589 $MAJ:$MIN device numbers and not ordered. The weights are in
1590 the range [1, 10000] and specifies the relative amount IO time
1591 the cgroup can use in relation to its siblings.
1593 The default weight can be updated by writing either "default
1594 $WEIGHT" or simply "$WEIGHT". Overrides can be set by writing
1595 "$MAJ:$MIN $WEIGHT" and unset by writing "$MAJ:$MIN default".
1597 An example read output follows::
1604 A read-write nested-keyed file which exists on non-root
1607 BPS and IOPS based IO limit. Lines are keyed by $MAJ:$MIN
1608 device numbers and not ordered. The following nested keys are
1611 ===== ==================================
1612 rbps Max read bytes per second
1613 wbps Max write bytes per second
1614 riops Max read IO operations per second
1615 wiops Max write IO operations per second
1616 ===== ==================================
1618 When writing, any number of nested key-value pairs can be
1619 specified in any order. "max" can be specified as the value
1620 to remove a specific limit. If the same key is specified
1621 multiple times, the outcome is undefined.
1623 BPS and IOPS are measured in each IO direction and IOs are
1624 delayed if limit is reached. Temporary bursts are allowed.
1626 Setting read limit at 2M BPS and write at 120 IOPS for 8:16::
1628 echo "8:16 rbps=2097152 wiops=120" > io.max
1630 Reading returns the following::
1632 8:16 rbps=2097152 wbps=max riops=max wiops=120
1634 Write IOPS limit can be removed by writing the following::
1636 echo "8:16 wiops=max" > io.max
1638 Reading now returns the following::
1640 8:16 rbps=2097152 wbps=max riops=max wiops=max
1643 A read-only nested-key file which exists on non-root cgroups.
1645 Shows pressure stall information for IO. See
1646 Documentation/accounting/psi.rst for details.
1652 Page cache is dirtied through buffered writes and shared mmaps and
1653 written asynchronously to the backing filesystem by the writeback
1654 mechanism. Writeback sits between the memory and IO domains and
1655 regulates the proportion of dirty memory by balancing dirtying and
1658 The io controller, in conjunction with the memory controller,
1659 implements control of page cache writeback IOs. The memory controller
1660 defines the memory domain that dirty memory ratio is calculated and
1661 maintained for and the io controller defines the io domain which
1662 writes out dirty pages for the memory domain. Both system-wide and
1663 per-cgroup dirty memory states are examined and the more restrictive
1664 of the two is enforced.
1666 cgroup writeback requires explicit support from the underlying
1667 filesystem. Currently, cgroup writeback is implemented on ext2, ext4
1668 and btrfs. On other filesystems, all writeback IOs are attributed to
1671 There are inherent differences in memory and writeback management
1672 which affects how cgroup ownership is tracked. Memory is tracked per
1673 page while writeback per inode. For the purpose of writeback, an
1674 inode is assigned to a cgroup and all IO requests to write dirty pages
1675 from the inode are attributed to that cgroup.
1677 As cgroup ownership for memory is tracked per page, there can be pages
1678 which are associated with different cgroups than the one the inode is
1679 associated with. These are called foreign pages. The writeback
1680 constantly keeps track of foreign pages and, if a particular foreign
1681 cgroup becomes the majority over a certain period of time, switches
1682 the ownership of the inode to that cgroup.
1684 While this model is enough for most use cases where a given inode is
1685 mostly dirtied by a single cgroup even when the main writing cgroup
1686 changes over time, use cases where multiple cgroups write to a single
1687 inode simultaneously are not supported well. In such circumstances, a
1688 significant portion of IOs are likely to be attributed incorrectly.
1689 As memory controller assigns page ownership on the first use and
1690 doesn't update it until the page is released, even if writeback
1691 strictly follows page ownership, multiple cgroups dirtying overlapping
1692 areas wouldn't work as expected. It's recommended to avoid such usage
1695 The sysctl knobs which affect writeback behavior are applied to cgroup
1696 writeback as follows.
1698 vm.dirty_background_ratio, vm.dirty_ratio
1699 These ratios apply the same to cgroup writeback with the
1700 amount of available memory capped by limits imposed by the
1701 memory controller and system-wide clean memory.
1703 vm.dirty_background_bytes, vm.dirty_bytes
1704 For cgroup writeback, this is calculated into ratio against
1705 total available memory and applied the same way as
1706 vm.dirty[_background]_ratio.
1712 This is a cgroup v2 controller for IO workload protection. You provide a group
1713 with a latency target, and if the average latency exceeds that target the
1714 controller will throttle any peers that have a lower latency target than the
1717 The limits are only applied at the peer level in the hierarchy. This means that
1718 in the diagram below, only groups A, B, and C will influence each other, and
1719 groups D and F will influence each other. Group G will influence nobody::
1728 So the ideal way to configure this is to set io.latency in groups A, B, and C.
1729 Generally you do not want to set a value lower than the latency your device
1730 supports. Experiment to find the value that works best for your workload.
1731 Start at higher than the expected latency for your device and watch the
1732 avg_lat value in io.stat for your workload group to get an idea of the
1733 latency you see during normal operation. Use the avg_lat value as a basis for
1734 your real setting, setting at 10-15% higher than the value in io.stat.
1736 How IO Latency Throttling Works
1737 ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
1739 io.latency is work conserving; so as long as everybody is meeting their latency
1740 target the controller doesn't do anything. Once a group starts missing its
1741 target it begins throttling any peer group that has a higher target than itself.
1742 This throttling takes 2 forms:
1744 - Queue depth throttling. This is the number of outstanding IO's a group is
1745 allowed to have. We will clamp down relatively quickly, starting at no limit
1746 and going all the way down to 1 IO at a time.
1748 - Artificial delay induction. There are certain types of IO that cannot be
1749 throttled without possibly adversely affecting higher priority groups. This
1750 includes swapping and metadata IO. These types of IO are allowed to occur
1751 normally, however they are "charged" to the originating group. If the
1752 originating group is being throttled you will see the use_delay and delay
1753 fields in io.stat increase. The delay value is how many microseconds that are
1754 being added to any process that runs in this group. Because this number can
1755 grow quite large if there is a lot of swapping or metadata IO occurring we
1756 limit the individual delay events to 1 second at a time.
1758 Once the victimized group starts meeting its latency target again it will start
1759 unthrottling any peer groups that were throttled previously. If the victimized
1760 group simply stops doing IO the global counter will unthrottle appropriately.
1762 IO Latency Interface Files
1763 ~~~~~~~~~~~~~~~~~~~~~~~~~~
1766 This takes a similar format as the other controllers.
1768 "MAJOR:MINOR target=<target time in microseconds"
1771 If the controller is enabled you will see extra stats in io.stat in
1772 addition to the normal ones.
1775 This is the current queue depth for the group.
1778 This is an exponential moving average with a decay rate of 1/exp
1779 bound by the sampling interval. The decay rate interval can be
1780 calculated by multiplying the win value in io.stat by the
1781 corresponding number of samples based on the win value.
1784 The sampling window size in milliseconds. This is the minimum
1785 duration of time between evaluation events. Windows only elapse
1786 with IO activity. Idle periods extend the most recent window.
1791 The process number controller is used to allow a cgroup to stop any
1792 new tasks from being fork()'d or clone()'d after a specified limit is
1795 The number of tasks in a cgroup can be exhausted in ways which other
1796 controllers cannot prevent, thus warranting its own controller. For
1797 example, a fork bomb is likely to exhaust the number of tasks before
1798 hitting memory restrictions.
1800 Note that PIDs used in this controller refer to TIDs, process IDs as
1808 A read-write single value file which exists on non-root
1809 cgroups. The default is "max".
1811 Hard limit of number of processes.
1814 A read-only single value file which exists on all cgroups.
1816 The number of processes currently in the cgroup and its
1819 Organisational operations are not blocked by cgroup policies, so it is
1820 possible to have pids.current > pids.max. This can be done by either
1821 setting the limit to be smaller than pids.current, or attaching enough
1822 processes to the cgroup such that pids.current is larger than
1823 pids.max. However, it is not possible to violate a cgroup PID policy
1824 through fork() or clone(). These will return -EAGAIN if the creation
1825 of a new process would cause a cgroup policy to be violated.
1831 The "cpuset" controller provides a mechanism for constraining
1832 the CPU and memory node placement of tasks to only the resources
1833 specified in the cpuset interface files in a task's current cgroup.
1834 This is especially valuable on large NUMA systems where placing jobs
1835 on properly sized subsets of the systems with careful processor and
1836 memory placement to reduce cross-node memory access and contention
1837 can improve overall system performance.
1839 The "cpuset" controller is hierarchical. That means the controller
1840 cannot use CPUs or memory nodes not allowed in its parent.
1843 Cpuset Interface Files
1844 ~~~~~~~~~~~~~~~~~~~~~~
1847 A read-write multiple values file which exists on non-root
1848 cpuset-enabled cgroups.
1850 It lists the requested CPUs to be used by tasks within this
1851 cgroup. The actual list of CPUs to be granted, however, is
1852 subjected to constraints imposed by its parent and can differ
1853 from the requested CPUs.
1855 The CPU numbers are comma-separated numbers or ranges.
1861 An empty value indicates that the cgroup is using the same
1862 setting as the nearest cgroup ancestor with a non-empty
1863 "cpuset.cpus" or all the available CPUs if none is found.
1865 The value of "cpuset.cpus" stays constant until the next update
1866 and won't be affected by any CPU hotplug events.
1868 cpuset.cpus.effective
1869 A read-only multiple values file which exists on all
1870 cpuset-enabled cgroups.
1872 It lists the onlined CPUs that are actually granted to this
1873 cgroup by its parent. These CPUs are allowed to be used by
1874 tasks within the current cgroup.
1876 If "cpuset.cpus" is empty, the "cpuset.cpus.effective" file shows
1877 all the CPUs from the parent cgroup that can be available to
1878 be used by this cgroup. Otherwise, it should be a subset of
1879 "cpuset.cpus" unless none of the CPUs listed in "cpuset.cpus"
1880 can be granted. In this case, it will be treated just like an
1881 empty "cpuset.cpus".
1883 Its value will be affected by CPU hotplug events.
1886 A read-write multiple values file which exists on non-root
1887 cpuset-enabled cgroups.
1889 It lists the requested memory nodes to be used by tasks within
1890 this cgroup. The actual list of memory nodes granted, however,
1891 is subjected to constraints imposed by its parent and can differ
1892 from the requested memory nodes.
1894 The memory node numbers are comma-separated numbers or ranges.
1900 An empty value indicates that the cgroup is using the same
1901 setting as the nearest cgroup ancestor with a non-empty
1902 "cpuset.mems" or all the available memory nodes if none
1905 The value of "cpuset.mems" stays constant until the next update
1906 and won't be affected by any memory nodes hotplug events.
1908 cpuset.mems.effective
1909 A read-only multiple values file which exists on all
1910 cpuset-enabled cgroups.
1912 It lists the onlined memory nodes that are actually granted to
1913 this cgroup by its parent. These memory nodes are allowed to
1914 be used by tasks within the current cgroup.
1916 If "cpuset.mems" is empty, it shows all the memory nodes from the
1917 parent cgroup that will be available to be used by this cgroup.
1918 Otherwise, it should be a subset of "cpuset.mems" unless none of
1919 the memory nodes listed in "cpuset.mems" can be granted. In this
1920 case, it will be treated just like an empty "cpuset.mems".
1922 Its value will be affected by memory nodes hotplug events.
1924 cpuset.cpus.partition
1925 A read-write single value file which exists on non-root
1926 cpuset-enabled cgroups. This flag is owned by the parent cgroup
1927 and is not delegatable.
1929 It accepts only the following input values when written to.
1931 "root" - a partition root
1932 "member" - a non-root member of a partition
1934 When set to be a partition root, the current cgroup is the
1935 root of a new partition or scheduling domain that comprises
1936 itself and all its descendants except those that are separate
1937 partition roots themselves and their descendants. The root
1938 cgroup is always a partition root.
1940 There are constraints on where a partition root can be set.
1941 It can only be set in a cgroup if all the following conditions
1944 1) The "cpuset.cpus" is not empty and the list of CPUs are
1945 exclusive, i.e. they are not shared by any of its siblings.
1946 2) The parent cgroup is a partition root.
1947 3) The "cpuset.cpus" is also a proper subset of the parent's
1948 "cpuset.cpus.effective".
1949 4) There is no child cgroups with cpuset enabled. This is for
1950 eliminating corner cases that have to be handled if such a
1951 condition is allowed.
1953 Setting it to partition root will take the CPUs away from the
1954 effective CPUs of the parent cgroup. Once it is set, this
1955 file cannot be reverted back to "member" if there are any child
1956 cgroups with cpuset enabled.
1958 A parent partition cannot distribute all its CPUs to its
1959 child partitions. There must be at least one cpu left in the
1962 Once becoming a partition root, changes to "cpuset.cpus" is
1963 generally allowed as long as the first condition above is true,
1964 the change will not take away all the CPUs from the parent
1965 partition and the new "cpuset.cpus" value is a superset of its
1966 children's "cpuset.cpus" values.
1968 Sometimes, external factors like changes to ancestors'
1969 "cpuset.cpus" or cpu hotplug can cause the state of the partition
1970 root to change. On read, the "cpuset.sched.partition" file
1971 can show the following values.
1973 "member" Non-root member of a partition
1974 "root" Partition root
1975 "root invalid" Invalid partition root
1977 It is a partition root if the first 2 partition root conditions
1978 above are true and at least one CPU from "cpuset.cpus" is
1979 granted by the parent cgroup.
1981 A partition root can become invalid if none of CPUs requested
1982 in "cpuset.cpus" can be granted by the parent cgroup or the
1983 parent cgroup is no longer a partition root itself. In this
1984 case, it is not a real partition even though the restriction
1985 of the first partition root condition above will still apply.
1986 The cpu affinity of all the tasks in the cgroup will then be
1987 associated with CPUs in the nearest ancestor partition.
1989 An invalid partition root can be transitioned back to a
1990 real partition root if at least one of the requested CPUs
1991 can now be granted by its parent. In this case, the cpu
1992 affinity of all the tasks in the formerly invalid partition
1993 will be associated to the CPUs of the newly formed partition.
1994 Changing the partition state of an invalid partition root to
1995 "member" is always allowed even if child cpusets are present.
2001 Device controller manages access to device files. It includes both
2002 creation of new device files (using mknod), and access to the
2003 existing device files.
2005 Cgroup v2 device controller has no interface files and is implemented
2006 on top of cgroup BPF. To control access to device files, a user may
2007 create bpf programs of the BPF_CGROUP_DEVICE type and attach them
2008 to cgroups. On an attempt to access a device file, corresponding
2009 BPF programs will be executed, and depending on the return value
2010 the attempt will succeed or fail with -EPERM.
2012 A BPF_CGROUP_DEVICE program takes a pointer to the bpf_cgroup_dev_ctx
2013 structure, which describes the device access attempt: access type
2014 (mknod/read/write) and device (type, major and minor numbers).
2015 If the program returns 0, the attempt fails with -EPERM, otherwise
2018 An example of BPF_CGROUP_DEVICE program may be found in the kernel
2019 source tree in the tools/testing/selftests/bpf/dev_cgroup.c file.
2025 The "rdma" controller regulates the distribution and accounting of
2028 RDMA Interface Files
2029 ~~~~~~~~~~~~~~~~~~~~
2032 A readwrite nested-keyed file that exists for all the cgroups
2033 except root that describes current configured resource limit
2034 for a RDMA/IB device.
2036 Lines are keyed by device name and are not ordered.
2037 Each line contains space separated resource name and its configured
2038 limit that can be distributed.
2040 The following nested keys are defined.
2042 ========== =============================
2043 hca_handle Maximum number of HCA Handles
2044 hca_object Maximum number of HCA Objects
2045 ========== =============================
2047 An example for mlx4 and ocrdma device follows::
2049 mlx4_0 hca_handle=2 hca_object=2000
2050 ocrdma1 hca_handle=3 hca_object=max
2053 A read-only file that describes current resource usage.
2054 It exists for all the cgroup except root.
2056 An example for mlx4 and ocrdma device follows::
2058 mlx4_0 hca_handle=1 hca_object=20
2059 ocrdma1 hca_handle=1 hca_object=23
2064 The HugeTLB controller allows to limit the HugeTLB usage per control group and
2065 enforces the controller limit during page fault.
2067 HugeTLB Interface Files
2068 ~~~~~~~~~~~~~~~~~~~~~~~
2070 hugetlb.<hugepagesize>.current
2071 Show current usage for "hugepagesize" hugetlb. It exists for all
2072 the cgroup except root.
2074 hugetlb.<hugepagesize>.max
2075 Set/show the hard limit of "hugepagesize" hugetlb usage.
2076 The default value is "max". It exists for all the cgroup except root.
2078 hugetlb.<hugepagesize>.events
2079 A read-only flat-keyed file which exists on non-root cgroups.
2082 The number of allocation failure due to HugeTLB limit
2084 hugetlb.<hugepagesize>.events.local
2085 Similar to hugetlb.<hugepagesize>.events but the fields in the file
2086 are local to the cgroup i.e. not hierarchical. The file modified event
2087 generated on this file reflects only the local events.
2095 perf_event controller, if not mounted on a legacy hierarchy, is
2096 automatically enabled on the v2 hierarchy so that perf events can
2097 always be filtered by cgroup v2 path. The controller can still be
2098 moved to a legacy hierarchy after v2 hierarchy is populated.
2101 Non-normative information
2102 -------------------------
2104 This section contains information that isn't considered to be a part of
2105 the stable kernel API and so is subject to change.
2108 CPU controller root cgroup process behaviour
2109 ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
2111 When distributing CPU cycles in the root cgroup each thread in this
2112 cgroup is treated as if it was hosted in a separate child cgroup of the
2113 root cgroup. This child cgroup weight is dependent on its thread nice
2116 For details of this mapping see sched_prio_to_weight array in
2117 kernel/sched/core.c file (values from this array should be scaled
2118 appropriately so the neutral - nice 0 - value is 100 instead of 1024).
2121 IO controller root cgroup process behaviour
2122 ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
2124 Root cgroup processes are hosted in an implicit leaf child node.
2125 When distributing IO resources this implicit child node is taken into
2126 account as if it was a normal child cgroup of the root cgroup with a
2127 weight value of 200.
2136 cgroup namespace provides a mechanism to virtualize the view of the
2137 "/proc/$PID/cgroup" file and cgroup mounts. The CLONE_NEWCGROUP clone
2138 flag can be used with clone(2) and unshare(2) to create a new cgroup
2139 namespace. The process running inside the cgroup namespace will have
2140 its "/proc/$PID/cgroup" output restricted to cgroupns root. The
2141 cgroupns root is the cgroup of the process at the time of creation of
2142 the cgroup namespace.
2144 Without cgroup namespace, the "/proc/$PID/cgroup" file shows the
2145 complete path of the cgroup of a process. In a container setup where
2146 a set of cgroups and namespaces are intended to isolate processes the
2147 "/proc/$PID/cgroup" file may leak potential system level information
2148 to the isolated processes. For Example::
2150 # cat /proc/self/cgroup
2151 0::/batchjobs/container_id1
2153 The path '/batchjobs/container_id1' can be considered as system-data
2154 and undesirable to expose to the isolated processes. cgroup namespace
2155 can be used to restrict visibility of this path. For example, before
2156 creating a cgroup namespace, one would see::
2158 # ls -l /proc/self/ns/cgroup
2159 lrwxrwxrwx 1 root root 0 2014-07-15 10:37 /proc/self/ns/cgroup -> cgroup:[4026531835]
2160 # cat /proc/self/cgroup
2161 0::/batchjobs/container_id1
2163 After unsharing a new namespace, the view changes::
2165 # ls -l /proc/self/ns/cgroup
2166 lrwxrwxrwx 1 root root 0 2014-07-15 10:35 /proc/self/ns/cgroup -> cgroup:[4026532183]
2167 # cat /proc/self/cgroup
2170 When some thread from a multi-threaded process unshares its cgroup
2171 namespace, the new cgroupns gets applied to the entire process (all
2172 the threads). This is natural for the v2 hierarchy; however, for the
2173 legacy hierarchies, this may be unexpected.
2175 A cgroup namespace is alive as long as there are processes inside or
2176 mounts pinning it. When the last usage goes away, the cgroup
2177 namespace is destroyed. The cgroupns root and the actual cgroups
2184 The 'cgroupns root' for a cgroup namespace is the cgroup in which the
2185 process calling unshare(2) is running. For example, if a process in
2186 /batchjobs/container_id1 cgroup calls unshare, cgroup
2187 /batchjobs/container_id1 becomes the cgroupns root. For the
2188 init_cgroup_ns, this is the real root ('/') cgroup.
2190 The cgroupns root cgroup does not change even if the namespace creator
2191 process later moves to a different cgroup::
2193 # ~/unshare -c # unshare cgroupns in some cgroup
2194 # cat /proc/self/cgroup
2197 # echo 0 > sub_cgrp_1/cgroup.procs
2198 # cat /proc/self/cgroup
2201 Each process gets its namespace-specific view of "/proc/$PID/cgroup"
2203 Processes running inside the cgroup namespace will be able to see
2204 cgroup paths (in /proc/self/cgroup) only inside their root cgroup.
2205 From within an unshared cgroupns::
2209 # echo 7353 > sub_cgrp_1/cgroup.procs
2210 # cat /proc/7353/cgroup
2213 From the initial cgroup namespace, the real cgroup path will be
2216 $ cat /proc/7353/cgroup
2217 0::/batchjobs/container_id1/sub_cgrp_1
2219 From a sibling cgroup namespace (that is, a namespace rooted at a
2220 different cgroup), the cgroup path relative to its own cgroup
2221 namespace root will be shown. For instance, if PID 7353's cgroup
2222 namespace root is at '/batchjobs/container_id2', then it will see::
2224 # cat /proc/7353/cgroup
2225 0::/../container_id2/sub_cgrp_1
2227 Note that the relative path always starts with '/' to indicate that
2228 its relative to the cgroup namespace root of the caller.
2231 Migration and setns(2)
2232 ----------------------
2234 Processes inside a cgroup namespace can move into and out of the
2235 namespace root if they have proper access to external cgroups. For
2236 example, from inside a namespace with cgroupns root at
2237 /batchjobs/container_id1, and assuming that the global hierarchy is
2238 still accessible inside cgroupns::
2240 # cat /proc/7353/cgroup
2242 # echo 7353 > batchjobs/container_id2/cgroup.procs
2243 # cat /proc/7353/cgroup
2244 0::/../container_id2
2246 Note that this kind of setup is not encouraged. A task inside cgroup
2247 namespace should only be exposed to its own cgroupns hierarchy.
2249 setns(2) to another cgroup namespace is allowed when:
2251 (a) the process has CAP_SYS_ADMIN against its current user namespace
2252 (b) the process has CAP_SYS_ADMIN against the target cgroup
2255 No implicit cgroup changes happen with attaching to another cgroup
2256 namespace. It is expected that the someone moves the attaching
2257 process under the target cgroup namespace root.
2260 Interaction with Other Namespaces
2261 ---------------------------------
2263 Namespace specific cgroup hierarchy can be mounted by a process
2264 running inside a non-init cgroup namespace::
2266 # mount -t cgroup2 none $MOUNT_POINT
2268 This will mount the unified cgroup hierarchy with cgroupns root as the
2269 filesystem root. The process needs CAP_SYS_ADMIN against its user and
2272 The virtualization of /proc/self/cgroup file combined with restricting
2273 the view of cgroup hierarchy by namespace-private cgroupfs mount
2274 provides a properly isolated cgroup view inside the container.
2277 Information on Kernel Programming
2278 =================================
2280 This section contains kernel programming information in the areas
2281 where interacting with cgroup is necessary. cgroup core and
2282 controllers are not covered.
2285 Filesystem Support for Writeback
2286 --------------------------------
2288 A filesystem can support cgroup writeback by updating
2289 address_space_operations->writepage[s]() to annotate bio's using the
2290 following two functions.
2292 wbc_init_bio(@wbc, @bio)
2293 Should be called for each bio carrying writeback data and
2294 associates the bio with the inode's owner cgroup and the
2295 corresponding request queue. This must be called after
2296 a queue (device) has been associated with the bio and
2299 wbc_account_cgroup_owner(@wbc, @page, @bytes)
2300 Should be called for each data segment being written out.
2301 While this function doesn't care exactly when it's called
2302 during the writeback session, it's the easiest and most
2303 natural to call it as data segments are added to a bio.
2305 With writeback bio's annotated, cgroup support can be enabled per
2306 super_block by setting SB_I_CGROUPWB in ->s_iflags. This allows for
2307 selective disabling of cgroup writeback support which is helpful when
2308 certain filesystem features, e.g. journaled data mode, are
2311 wbc_init_bio() binds the specified bio to its cgroup. Depending on
2312 the configuration, the bio may be executed at a lower priority and if
2313 the writeback session is holding shared resources, e.g. a journal
2314 entry, may lead to priority inversion. There is no one easy solution
2315 for the problem. Filesystems can try to work around specific problem
2316 cases by skipping wbc_init_bio() and using bio_associate_blkg()
2320 Deprecated v1 Core Features
2321 ===========================
2323 - Multiple hierarchies including named ones are not supported.
2325 - All v1 mount options are not supported.
2327 - The "tasks" file is removed and "cgroup.procs" is not sorted.
2329 - "cgroup.clone_children" is removed.
2331 - /proc/cgroups is meaningless for v2. Use "cgroup.controllers" file
2332 at the root instead.
2335 Issues with v1 and Rationales for v2
2336 ====================================
2338 Multiple Hierarchies
2339 --------------------
2341 cgroup v1 allowed an arbitrary number of hierarchies and each
2342 hierarchy could host any number of controllers. While this seemed to
2343 provide a high level of flexibility, it wasn't useful in practice.
2345 For example, as there is only one instance of each controller, utility
2346 type controllers such as freezer which can be useful in all
2347 hierarchies could only be used in one. The issue is exacerbated by
2348 the fact that controllers couldn't be moved to another hierarchy once
2349 hierarchies were populated. Another issue was that all controllers
2350 bound to a hierarchy were forced to have exactly the same view of the
2351 hierarchy. It wasn't possible to vary the granularity depending on
2352 the specific controller.
2354 In practice, these issues heavily limited which controllers could be
2355 put on the same hierarchy and most configurations resorted to putting
2356 each controller on its own hierarchy. Only closely related ones, such
2357 as the cpu and cpuacct controllers, made sense to be put on the same
2358 hierarchy. This often meant that userland ended up managing multiple
2359 similar hierarchies repeating the same steps on each hierarchy
2360 whenever a hierarchy management operation was necessary.
2362 Furthermore, support for multiple hierarchies came at a steep cost.
2363 It greatly complicated cgroup core implementation but more importantly
2364 the support for multiple hierarchies restricted how cgroup could be
2365 used in general and what controllers was able to do.
2367 There was no limit on how many hierarchies there might be, which meant
2368 that a thread's cgroup membership couldn't be described in finite
2369 length. The key might contain any number of entries and was unlimited
2370 in length, which made it highly awkward to manipulate and led to
2371 addition of controllers which existed only to identify membership,
2372 which in turn exacerbated the original problem of proliferating number
2375 Also, as a controller couldn't have any expectation regarding the
2376 topologies of hierarchies other controllers might be on, each
2377 controller had to assume that all other controllers were attached to
2378 completely orthogonal hierarchies. This made it impossible, or at
2379 least very cumbersome, for controllers to cooperate with each other.
2381 In most use cases, putting controllers on hierarchies which are
2382 completely orthogonal to each other isn't necessary. What usually is
2383 called for is the ability to have differing levels of granularity
2384 depending on the specific controller. In other words, hierarchy may
2385 be collapsed from leaf towards root when viewed from specific
2386 controllers. For example, a given configuration might not care about
2387 how memory is distributed beyond a certain level while still wanting
2388 to control how CPU cycles are distributed.
2394 cgroup v1 allowed threads of a process to belong to different cgroups.
2395 This didn't make sense for some controllers and those controllers
2396 ended up implementing different ways to ignore such situations but
2397 much more importantly it blurred the line between API exposed to
2398 individual applications and system management interface.
2400 Generally, in-process knowledge is available only to the process
2401 itself; thus, unlike service-level organization of processes,
2402 categorizing threads of a process requires active participation from
2403 the application which owns the target process.
2405 cgroup v1 had an ambiguously defined delegation model which got abused
2406 in combination with thread granularity. cgroups were delegated to
2407 individual applications so that they can create and manage their own
2408 sub-hierarchies and control resource distributions along them. This
2409 effectively raised cgroup to the status of a syscall-like API exposed
2412 First of all, cgroup has a fundamentally inadequate interface to be
2413 exposed this way. For a process to access its own knobs, it has to
2414 extract the path on the target hierarchy from /proc/self/cgroup,
2415 construct the path by appending the name of the knob to the path, open
2416 and then read and/or write to it. This is not only extremely clunky
2417 and unusual but also inherently racy. There is no conventional way to
2418 define transaction across the required steps and nothing can guarantee
2419 that the process would actually be operating on its own sub-hierarchy.
2421 cgroup controllers implemented a number of knobs which would never be
2422 accepted as public APIs because they were just adding control knobs to
2423 system-management pseudo filesystem. cgroup ended up with interface
2424 knobs which were not properly abstracted or refined and directly
2425 revealed kernel internal details. These knobs got exposed to
2426 individual applications through the ill-defined delegation mechanism
2427 effectively abusing cgroup as a shortcut to implementing public APIs
2428 without going through the required scrutiny.
2430 This was painful for both userland and kernel. Userland ended up with
2431 misbehaving and poorly abstracted interfaces and kernel exposing and
2432 locked into constructs inadvertently.
2435 Competition Between Inner Nodes and Threads
2436 -------------------------------------------
2438 cgroup v1 allowed threads to be in any cgroups which created an
2439 interesting problem where threads belonging to a parent cgroup and its
2440 children cgroups competed for resources. This was nasty as two
2441 different types of entities competed and there was no obvious way to
2442 settle it. Different controllers did different things.
2444 The cpu controller considered threads and cgroups as equivalents and
2445 mapped nice levels to cgroup weights. This worked for some cases but
2446 fell flat when children wanted to be allocated specific ratios of CPU
2447 cycles and the number of internal threads fluctuated - the ratios
2448 constantly changed as the number of competing entities fluctuated.
2449 There also were other issues. The mapping from nice level to weight
2450 wasn't obvious or universal, and there were various other knobs which
2451 simply weren't available for threads.
2453 The io controller implicitly created a hidden leaf node for each
2454 cgroup to host the threads. The hidden leaf had its own copies of all
2455 the knobs with ``leaf_`` prefixed. While this allowed equivalent
2456 control over internal threads, it was with serious drawbacks. It
2457 always added an extra layer of nesting which wouldn't be necessary
2458 otherwise, made the interface messy and significantly complicated the
2461 The memory controller didn't have a way to control what happened
2462 between internal tasks and child cgroups and the behavior was not
2463 clearly defined. There were attempts to add ad-hoc behaviors and
2464 knobs to tailor the behavior to specific workloads which would have
2465 led to problems extremely difficult to resolve in the long term.
2467 Multiple controllers struggled with internal tasks and came up with
2468 different ways to deal with it; unfortunately, all the approaches were
2469 severely flawed and, furthermore, the widely different behaviors
2470 made cgroup as a whole highly inconsistent.
2472 This clearly is a problem which needs to be addressed from cgroup core
2476 Other Interface Issues
2477 ----------------------
2479 cgroup v1 grew without oversight and developed a large number of
2480 idiosyncrasies and inconsistencies. One issue on the cgroup core side
2481 was how an empty cgroup was notified - a userland helper binary was
2482 forked and executed for each event. The event delivery wasn't
2483 recursive or delegatable. The limitations of the mechanism also led
2484 to in-kernel event delivery filtering mechanism further complicating
2487 Controller interfaces were problematic too. An extreme example is
2488 controllers completely ignoring hierarchical organization and treating
2489 all cgroups as if they were all located directly under the root
2490 cgroup. Some controllers exposed a large amount of inconsistent
2491 implementation details to userland.
2493 There also was no consistency across controllers. When a new cgroup
2494 was created, some controllers defaulted to not imposing extra
2495 restrictions while others disallowed any resource usage until
2496 explicitly configured. Configuration knobs for the same type of
2497 control used widely differing naming schemes and formats. Statistics
2498 and information knobs were named arbitrarily and used different
2499 formats and units even in the same controller.
2501 cgroup v2 establishes common conventions where appropriate and updates
2502 controllers so that they expose minimal and consistent interfaces.
2505 Controller Issues and Remedies
2506 ------------------------------
2511 The original lower boundary, the soft limit, is defined as a limit
2512 that is per default unset. As a result, the set of cgroups that
2513 global reclaim prefers is opt-in, rather than opt-out. The costs for
2514 optimizing these mostly negative lookups are so high that the
2515 implementation, despite its enormous size, does not even provide the
2516 basic desirable behavior. First off, the soft limit has no
2517 hierarchical meaning. All configured groups are organized in a global
2518 rbtree and treated like equal peers, regardless where they are located
2519 in the hierarchy. This makes subtree delegation impossible. Second,
2520 the soft limit reclaim pass is so aggressive that it not just
2521 introduces high allocation latencies into the system, but also impacts
2522 system performance due to overreclaim, to the point where the feature
2523 becomes self-defeating.
2525 The memory.low boundary on the other hand is a top-down allocated
2526 reserve. A cgroup enjoys reclaim protection when it's within its
2527 effective low, which makes delegation of subtrees possible. It also
2528 enjoys having reclaim pressure proportional to its overage when
2529 above its effective low.
2531 The original high boundary, the hard limit, is defined as a strict
2532 limit that can not budge, even if the OOM killer has to be called.
2533 But this generally goes against the goal of making the most out of the
2534 available memory. The memory consumption of workloads varies during
2535 runtime, and that requires users to overcommit. But doing that with a
2536 strict upper limit requires either a fairly accurate prediction of the
2537 working set size or adding slack to the limit. Since working set size
2538 estimation is hard and error prone, and getting it wrong results in
2539 OOM kills, most users tend to err on the side of a looser limit and
2540 end up wasting precious resources.
2542 The memory.high boundary on the other hand can be set much more
2543 conservatively. When hit, it throttles allocations by forcing them
2544 into direct reclaim to work off the excess, but it never invokes the
2545 OOM killer. As a result, a high boundary that is chosen too
2546 aggressively will not terminate the processes, but instead it will
2547 lead to gradual performance degradation. The user can monitor this
2548 and make corrections until the minimal memory footprint that still
2549 gives acceptable performance is found.
2551 In extreme cases, with many concurrent allocations and a complete
2552 breakdown of reclaim progress within the group, the high boundary can
2553 be exceeded. But even then it's mostly better to satisfy the
2554 allocation from the slack available in other groups or the rest of the
2555 system than killing the group. Otherwise, memory.max is there to
2556 limit this type of spillover and ultimately contain buggy or even
2557 malicious applications.
2559 Setting the original memory.limit_in_bytes below the current usage was
2560 subject to a race condition, where concurrent charges could cause the
2561 limit setting to fail. memory.max on the other hand will first set the
2562 limit to prevent new charges, and then reclaim and OOM kill until the
2563 new limit is met - or the task writing to memory.max is killed.
2565 The combined memory+swap accounting and limiting is replaced by real
2566 control over swap space.
2568 The main argument for a combined memory+swap facility in the original
2569 cgroup design was that global or parental pressure would always be
2570 able to swap all anonymous memory of a child group, regardless of the
2571 child's own (possibly untrusted) configuration. However, untrusted
2572 groups can sabotage swapping by other means - such as referencing its
2573 anonymous memory in a tight loop - and an admin can not assume full
2574 swappability when overcommitting untrusted jobs.
2576 For trusted jobs, on the other hand, a combined counter is not an
2577 intuitive userspace interface, and it flies in the face of the idea
2578 that cgroup controllers should account and limit specific physical
2579 resources. Swap space is a resource like all others in the system,
2580 and that's why unified hierarchy allows distributing it separately.