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
66 5-N. Non-normative information
67 5-N-1. CPU controller root cgroup process behaviour
68 5-N-2. IO controller root cgroup process behaviour
71 6-2. The Root and Views
72 6-3. Migration and setns(2)
73 6-4. Interaction with Other Namespaces
74 P. Information on Kernel Programming
75 P-1. Filesystem Support for Writeback
76 D. Deprecated v1 Core Features
77 R. Issues with v1 and Rationales for v2
78 R-1. Multiple Hierarchies
79 R-2. Thread Granularity
80 R-3. Competition Between Inner Nodes and Threads
81 R-4. Other Interface Issues
82 R-5. Controller Issues and Remedies
92 "cgroup" stands for "control group" and is never capitalized. The
93 singular form is used to designate the whole feature and also as a
94 qualifier as in "cgroup controllers". When explicitly referring to
95 multiple individual control groups, the plural form "cgroups" is used.
101 cgroup is a mechanism to organize processes hierarchically and
102 distribute system resources along the hierarchy in a controlled and
105 cgroup is largely composed of two parts - the core and controllers.
106 cgroup core is primarily responsible for hierarchically organizing
107 processes. A cgroup controller is usually responsible for
108 distributing a specific type of system resource along the hierarchy
109 although there are utility controllers which serve purposes other than
110 resource distribution.
112 cgroups form a tree structure and every process in the system belongs
113 to one and only one cgroup. All threads of a process belong to the
114 same cgroup. On creation, all processes are put in the cgroup that
115 the parent process belongs to at the time. A process can be migrated
116 to another cgroup. Migration of a process doesn't affect already
117 existing descendant processes.
119 Following certain structural constraints, controllers may be enabled or
120 disabled selectively on a cgroup. All controller behaviors are
121 hierarchical - if a controller is enabled on a cgroup, it affects all
122 processes which belong to the cgroups consisting the inclusive
123 sub-hierarchy of the cgroup. When a controller is enabled on a nested
124 cgroup, it always restricts the resource distribution further. The
125 restrictions set closer to the root in the hierarchy can not be
126 overridden from further away.
135 Unlike v1, cgroup v2 has only single hierarchy. The cgroup v2
136 hierarchy can be mounted with the following mount command::
138 # mount -t cgroup2 none $MOUNT_POINT
140 cgroup2 filesystem has the magic number 0x63677270 ("cgrp"). All
141 controllers which support v2 and are not bound to a v1 hierarchy are
142 automatically bound to the v2 hierarchy and show up at the root.
143 Controllers which are not in active use in the v2 hierarchy can be
144 bound to other hierarchies. This allows mixing v2 hierarchy with the
145 legacy v1 multiple hierarchies in a fully backward compatible way.
147 A controller can be moved across hierarchies only after the controller
148 is no longer referenced in its current hierarchy. Because per-cgroup
149 controller states are destroyed asynchronously and controllers may
150 have lingering references, a controller may not show up immediately on
151 the v2 hierarchy after the final umount of the previous hierarchy.
152 Similarly, a controller should be fully disabled to be moved out of
153 the unified hierarchy and it may take some time for the disabled
154 controller to become available for other hierarchies; furthermore, due
155 to inter-controller dependencies, other controllers may need to be
158 While useful for development and manual configurations, moving
159 controllers dynamically between the v2 and other hierarchies is
160 strongly discouraged for production use. It is recommended to decide
161 the hierarchies and controller associations before starting using the
162 controllers after system boot.
164 During transition to v2, system management software might still
165 automount the v1 cgroup filesystem and so hijack all controllers
166 during boot, before manual intervention is possible. To make testing
167 and experimenting easier, the kernel parameter cgroup_no_v1= allows
168 disabling controllers in v1 and make them always available in v2.
170 cgroup v2 currently supports the following mount options.
174 Consider cgroup namespaces as delegation boundaries. This
175 option is system wide and can only be set on mount or modified
176 through remount from the init namespace. The mount option is
177 ignored on non-init namespace mounts. Please refer to the
178 Delegation section for details.
182 Only populate memory.events with data for the current cgroup,
183 and not any subtrees. This is legacy behaviour, the default
184 behaviour without this option is to include subtree counts.
185 This option is system wide and can only be set on mount or
186 modified through remount from the init namespace. The mount
187 option is ignored on non-init namespace mounts.
190 Organizing Processes and Threads
191 --------------------------------
196 Initially, only the root cgroup exists to which all processes belong.
197 A child cgroup can be created by creating a sub-directory::
201 A given cgroup may have multiple child cgroups forming a tree
202 structure. Each cgroup has a read-writable interface file
203 "cgroup.procs". When read, it lists the PIDs of all processes which
204 belong to the cgroup one-per-line. The PIDs are not ordered and the
205 same PID may show up more than once if the process got moved to
206 another cgroup and then back or the PID got recycled while reading.
208 A process can be migrated into a cgroup by writing its PID to the
209 target cgroup's "cgroup.procs" file. Only one process can be migrated
210 on a single write(2) call. If a process is composed of multiple
211 threads, writing the PID of any thread migrates all threads of the
214 When a process forks a child process, the new process is born into the
215 cgroup that the forking process belongs to at the time of the
216 operation. After exit, a process stays associated with the cgroup
217 that it belonged to at the time of exit until it's reaped; however, a
218 zombie process does not appear in "cgroup.procs" and thus can't be
219 moved to another cgroup.
221 A cgroup which doesn't have any children or live processes can be
222 destroyed by removing the directory. Note that a cgroup which doesn't
223 have any children and is associated only with zombie processes is
224 considered empty and can be removed::
228 "/proc/$PID/cgroup" lists a process's cgroup membership. If legacy
229 cgroup is in use in the system, this file may contain multiple lines,
230 one for each hierarchy. The entry for cgroup v2 is always in the
233 # cat /proc/842/cgroup
235 0::/test-cgroup/test-cgroup-nested
237 If the process becomes a zombie and the cgroup it was associated with
238 is removed subsequently, " (deleted)" is appended to the path::
240 # cat /proc/842/cgroup
242 0::/test-cgroup/test-cgroup-nested (deleted)
248 cgroup v2 supports thread granularity for a subset of controllers to
249 support use cases requiring hierarchical resource distribution across
250 the threads of a group of processes. By default, all threads of a
251 process belong to the same cgroup, which also serves as the resource
252 domain to host resource consumptions which are not specific to a
253 process or thread. The thread mode allows threads to be spread across
254 a subtree while still maintaining the common resource domain for them.
256 Controllers which support thread mode are called threaded controllers.
257 The ones which don't are called domain controllers.
259 Marking a cgroup threaded makes it join the resource domain of its
260 parent as a threaded cgroup. The parent may be another threaded
261 cgroup whose resource domain is further up in the hierarchy. The root
262 of a threaded subtree, that is, the nearest ancestor which is not
263 threaded, is called threaded domain or thread root interchangeably and
264 serves as the resource domain for the entire subtree.
266 Inside a threaded subtree, threads of a process can be put in
267 different cgroups and are not subject to the no internal process
268 constraint - threaded controllers can be enabled on non-leaf cgroups
269 whether they have threads in them or not.
271 As the threaded domain cgroup hosts all the domain resource
272 consumptions of the subtree, it is considered to have internal
273 resource consumptions whether there are processes in it or not and
274 can't have populated child cgroups which aren't threaded. Because the
275 root cgroup is not subject to no internal process constraint, it can
276 serve both as a threaded domain and a parent to domain cgroups.
278 The current operation mode or type of the cgroup is shown in the
279 "cgroup.type" file which indicates whether the cgroup is a normal
280 domain, a domain which is serving as the domain of a threaded subtree,
281 or a threaded cgroup.
283 On creation, a cgroup is always a domain cgroup and can be made
284 threaded by writing "threaded" to the "cgroup.type" file. The
285 operation is single direction::
287 # echo threaded > cgroup.type
289 Once threaded, the cgroup can't be made a domain again. To enable the
290 thread mode, the following conditions must be met.
292 - As the cgroup will join the parent's resource domain. The parent
293 must either be a valid (threaded) domain or a threaded cgroup.
295 - When the parent is an unthreaded domain, it must not have any domain
296 controllers enabled or populated domain children. The root is
297 exempt from this requirement.
299 Topology-wise, a cgroup can be in an invalid state. Please consider
300 the following topology::
302 A (threaded domain) - B (threaded) - C (domain, just created)
304 C is created as a domain but isn't connected to a parent which can
305 host child domains. C can't be used until it is turned into a
306 threaded cgroup. "cgroup.type" file will report "domain (invalid)" in
307 these cases. Operations which fail due to invalid topology use
308 EOPNOTSUPP as the errno.
310 A domain cgroup is turned into a threaded domain when one of its child
311 cgroup becomes threaded or threaded controllers are enabled in the
312 "cgroup.subtree_control" file while there are processes in the cgroup.
313 A threaded domain reverts to a normal domain when the conditions
316 When read, "cgroup.threads" contains the list of the thread IDs of all
317 threads in the cgroup. Except that the operations are per-thread
318 instead of per-process, "cgroup.threads" has the same format and
319 behaves the same way as "cgroup.procs". While "cgroup.threads" can be
320 written to in any cgroup, as it can only move threads inside the same
321 threaded domain, its operations are confined inside each threaded
324 The threaded domain cgroup serves as the resource domain for the whole
325 subtree, and, while the threads can be scattered across the subtree,
326 all the processes are considered to be in the threaded domain cgroup.
327 "cgroup.procs" in a threaded domain cgroup contains the PIDs of all
328 processes in the subtree and is not readable in the subtree proper.
329 However, "cgroup.procs" can be written to from anywhere in the subtree
330 to migrate all threads of the matching process to the cgroup.
332 Only threaded controllers can be enabled in a threaded subtree. When
333 a threaded controller is enabled inside a threaded subtree, it only
334 accounts for and controls resource consumptions associated with the
335 threads in the cgroup and its descendants. All consumptions which
336 aren't tied to a specific thread belong to the threaded domain cgroup.
338 Because a threaded subtree is exempt from no internal process
339 constraint, a threaded controller must be able to handle competition
340 between threads in a non-leaf cgroup and its child cgroups. Each
341 threaded controller defines how such competitions are handled.
344 [Un]populated Notification
345 --------------------------
347 Each non-root cgroup has a "cgroup.events" file which contains
348 "populated" field indicating whether the cgroup's sub-hierarchy has
349 live processes in it. Its value is 0 if there is no live process in
350 the cgroup and its descendants; otherwise, 1. poll and [id]notify
351 events are triggered when the value changes. This can be used, for
352 example, to start a clean-up operation after all processes of a given
353 sub-hierarchy have exited. The populated state updates and
354 notifications are recursive. Consider the following sub-hierarchy
355 where the numbers in the parentheses represent the numbers of processes
361 A, B and C's "populated" fields would be 1 while D's 0. After the one
362 process in C exits, B and C's "populated" fields would flip to "0" and
363 file modified events will be generated on the "cgroup.events" files of
367 Controlling Controllers
368 -----------------------
370 Enabling and Disabling
371 ~~~~~~~~~~~~~~~~~~~~~~
373 Each cgroup has a "cgroup.controllers" file which lists all
374 controllers available for the cgroup to enable::
376 # cat cgroup.controllers
379 No controller is enabled by default. Controllers can be enabled and
380 disabled by writing to the "cgroup.subtree_control" file::
382 # echo "+cpu +memory -io" > cgroup.subtree_control
384 Only controllers which are listed in "cgroup.controllers" can be
385 enabled. When multiple operations are specified as above, either they
386 all succeed or fail. If multiple operations on the same controller
387 are specified, the last one is effective.
389 Enabling a controller in a cgroup indicates that the distribution of
390 the target resource across its immediate children will be controlled.
391 Consider the following sub-hierarchy. The enabled controllers are
392 listed in parentheses::
394 A(cpu,memory) - B(memory) - C()
397 As A has "cpu" and "memory" enabled, A will control the distribution
398 of CPU cycles and memory to its children, in this case, B. As B has
399 "memory" enabled but not "CPU", C and D will compete freely on CPU
400 cycles but their division of memory available to B will be controlled.
402 As a controller regulates the distribution of the target resource to
403 the cgroup's children, enabling it creates the controller's interface
404 files in the child cgroups. In the above example, enabling "cpu" on B
405 would create the "cpu." prefixed controller interface files in C and
406 D. Likewise, disabling "memory" from B would remove the "memory."
407 prefixed controller interface files from C and D. This means that the
408 controller interface files - anything which doesn't start with
409 "cgroup." are owned by the parent rather than the cgroup itself.
415 Resources are distributed top-down and a cgroup can further distribute
416 a resource only if the resource has been distributed to it from the
417 parent. This means that all non-root "cgroup.subtree_control" files
418 can only contain controllers which are enabled in the parent's
419 "cgroup.subtree_control" file. A controller can be enabled only if
420 the parent has the controller enabled and a controller can't be
421 disabled if one or more children have it enabled.
424 No Internal Process Constraint
425 ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
427 Non-root cgroups can distribute domain resources to their children
428 only when they don't have any processes of their own. In other words,
429 only domain cgroups which don't contain any processes can have domain
430 controllers enabled in their "cgroup.subtree_control" files.
432 This guarantees that, when a domain controller is looking at the part
433 of the hierarchy which has it enabled, processes are always only on
434 the leaves. This rules out situations where child cgroups compete
435 against internal processes of the parent.
437 The root cgroup is exempt from this restriction. Root contains
438 processes and anonymous resource consumption which can't be associated
439 with any other cgroups and requires special treatment from most
440 controllers. How resource consumption in the root cgroup is governed
441 is up to each controller (for more information on this topic please
442 refer to the Non-normative information section in the Controllers
445 Note that the restriction doesn't get in the way if there is no
446 enabled controller in the cgroup's "cgroup.subtree_control". This is
447 important as otherwise it wouldn't be possible to create children of a
448 populated cgroup. To control resource distribution of a cgroup, the
449 cgroup must create children and transfer all its processes to the
450 children before enabling controllers in its "cgroup.subtree_control"
460 A cgroup can be delegated in two ways. First, to a less privileged
461 user by granting write access of the directory and its "cgroup.procs",
462 "cgroup.threads" and "cgroup.subtree_control" files to the user.
463 Second, if the "nsdelegate" mount option is set, automatically to a
464 cgroup namespace on namespace creation.
466 Because the resource control interface files in a given directory
467 control the distribution of the parent's resources, the delegatee
468 shouldn't be allowed to write to them. For the first method, this is
469 achieved by not granting access to these files. For the second, the
470 kernel rejects writes to all files other than "cgroup.procs" and
471 "cgroup.subtree_control" on a namespace root from inside the
474 The end results are equivalent for both delegation types. Once
475 delegated, the user can build sub-hierarchy under the directory,
476 organize processes inside it as it sees fit and further distribute the
477 resources it received from the parent. The limits and other settings
478 of all resource controllers are hierarchical and regardless of what
479 happens in the delegated sub-hierarchy, nothing can escape the
480 resource restrictions imposed by the parent.
482 Currently, cgroup doesn't impose any restrictions on the number of
483 cgroups in or nesting depth of a delegated sub-hierarchy; however,
484 this may be limited explicitly in the future.
487 Delegation Containment
488 ~~~~~~~~~~~~~~~~~~~~~~
490 A delegated sub-hierarchy is contained in the sense that processes
491 can't be moved into or out of the sub-hierarchy by the delegatee.
493 For delegations to a less privileged user, this is achieved by
494 requiring the following conditions for a process with a non-root euid
495 to migrate a target process into a cgroup by writing its PID to the
498 - The writer must have write access to the "cgroup.procs" file.
500 - The writer must have write access to the "cgroup.procs" file of the
501 common ancestor of the source and destination cgroups.
503 The above two constraints ensure that while a delegatee may migrate
504 processes around freely in the delegated sub-hierarchy it can't pull
505 in from or push out to outside the sub-hierarchy.
507 For an example, let's assume cgroups C0 and C1 have been delegated to
508 user U0 who created C00, C01 under C0 and C10 under C1 as follows and
509 all processes under C0 and C1 belong to U0::
511 ~~~~~~~~~~~~~ - C0 - C00
514 ~~~~~~~~~~~~~ - C1 - C10
516 Let's also say U0 wants to write the PID of a process which is
517 currently in C10 into "C00/cgroup.procs". U0 has write access to the
518 file; however, the common ancestor of the source cgroup C10 and the
519 destination cgroup C00 is above the points of delegation and U0 would
520 not have write access to its "cgroup.procs" files and thus the write
521 will be denied with -EACCES.
523 For delegations to namespaces, containment is achieved by requiring
524 that both the source and destination cgroups are reachable from the
525 namespace of the process which is attempting the migration. If either
526 is not reachable, the migration is rejected with -ENOENT.
532 Organize Once and Control
533 ~~~~~~~~~~~~~~~~~~~~~~~~~
535 Migrating a process across cgroups is a relatively expensive operation
536 and stateful resources such as memory are not moved together with the
537 process. This is an explicit design decision as there often exist
538 inherent trade-offs between migration and various hot paths in terms
539 of synchronization cost.
541 As such, migrating processes across cgroups frequently as a means to
542 apply different resource restrictions is discouraged. A workload
543 should be assigned to a cgroup according to the system's logical and
544 resource structure once on start-up. Dynamic adjustments to resource
545 distribution can be made by changing controller configuration through
549 Avoid Name Collisions
550 ~~~~~~~~~~~~~~~~~~~~~
552 Interface files for a cgroup and its children cgroups occupy the same
553 directory and it is possible to create children cgroups which collide
554 with interface files.
556 All cgroup core interface files are prefixed with "cgroup." and each
557 controller's interface files are prefixed with the controller name and
558 a dot. A controller's name is composed of lower case alphabets and
559 '_'s but never begins with an '_' so it can be used as the prefix
560 character for collision avoidance. Also, interface file names won't
561 start or end with terms which are often used in categorizing workloads
562 such as job, service, slice, unit or workload.
564 cgroup doesn't do anything to prevent name collisions and it's the
565 user's responsibility to avoid them.
568 Resource Distribution Models
569 ============================
571 cgroup controllers implement several resource distribution schemes
572 depending on the resource type and expected use cases. This section
573 describes major schemes in use along with their expected behaviors.
579 A parent's resource is distributed by adding up the weights of all
580 active children and giving each the fraction matching the ratio of its
581 weight against the sum. As only children which can make use of the
582 resource at the moment participate in the distribution, this is
583 work-conserving. Due to the dynamic nature, this model is usually
584 used for stateless resources.
586 All weights are in the range [1, 10000] with the default at 100. This
587 allows symmetric multiplicative biases in both directions at fine
588 enough granularity while staying in the intuitive range.
590 As long as the weight is in range, all configuration combinations are
591 valid and there is no reason to reject configuration changes or
594 "cpu.weight" proportionally distributes CPU cycles to active children
595 and is an example of this type.
601 A child can only consume upto the configured amount of the resource.
602 Limits can be over-committed - the sum of the limits of children can
603 exceed the amount of resource available to the parent.
605 Limits are in the range [0, max] and defaults to "max", which is noop.
607 As limits can be over-committed, all configuration combinations are
608 valid and there is no reason to reject configuration changes or
611 "io.max" limits the maximum BPS and/or IOPS that a cgroup can consume
612 on an IO device and is an example of this type.
618 A cgroup is protected upto the configured amount of the resource
619 as long as the usages of all its ancestors are under their
620 protected levels. Protections can be hard guarantees or best effort
621 soft boundaries. Protections can also be over-committed in which case
622 only upto the amount available to the parent is protected among
625 Protections are in the range [0, max] and defaults to 0, which is
628 As protections can be over-committed, all configuration combinations
629 are valid and there is no reason to reject configuration changes or
632 "memory.low" implements best-effort memory protection and is an
633 example of this type.
639 A cgroup is exclusively allocated a certain amount of a finite
640 resource. Allocations can't be over-committed - the sum of the
641 allocations of children can not exceed the amount of resource
642 available to the parent.
644 Allocations are in the range [0, max] and defaults to 0, which is no
647 As allocations can't be over-committed, some configuration
648 combinations are invalid and should be rejected. Also, if the
649 resource is mandatory for execution of processes, process migrations
652 "cpu.rt.max" hard-allocates realtime slices and is an example of this
662 All interface files should be in one of the following formats whenever
665 New-line separated values
666 (when only one value can be written at once)
672 Space separated values
673 (when read-only or multiple values can be written at once)
685 KEY0 SUB_KEY0=VAL00 SUB_KEY1=VAL01...
686 KEY1 SUB_KEY0=VAL10 SUB_KEY1=VAL11...
689 For a writable file, the format for writing should generally match
690 reading; however, controllers may allow omitting later fields or
691 implement restricted shortcuts for most common use cases.
693 For both flat and nested keyed files, only the values for a single key
694 can be written at a time. For nested keyed files, the sub key pairs
695 may be specified in any order and not all pairs have to be specified.
701 - Settings for a single feature should be contained in a single file.
703 - The root cgroup should be exempt from resource control and thus
704 shouldn't have resource control interface files. Also,
705 informational files on the root cgroup which end up showing global
706 information available elsewhere shouldn't exist.
708 - The default time unit is microseconds. If a different unit is ever
709 used, an explicit unit suffix must be present.
711 - A parts-per quantity should use a percentage decimal with at least
712 two digit fractional part - e.g. 13.40.
714 - If a controller implements weight based resource distribution, its
715 interface file should be named "weight" and have the range [1,
716 10000] with 100 as the default. The values are chosen to allow
717 enough and symmetric bias in both directions while keeping it
718 intuitive (the default is 100%).
720 - If a controller implements an absolute resource guarantee and/or
721 limit, the interface files should be named "min" and "max"
722 respectively. If a controller implements best effort resource
723 guarantee and/or limit, the interface files should be named "low"
724 and "high" respectively.
726 In the above four control files, the special token "max" should be
727 used to represent upward infinity for both reading and writing.
729 - If a setting has a configurable default value and keyed specific
730 overrides, the default entry should be keyed with "default" and
731 appear as the first entry in the file.
733 The default value can be updated by writing either "default $VAL" or
736 When writing to update a specific override, "default" can be used as
737 the value to indicate removal of the override. Override entries
738 with "default" as the value must not appear when read.
740 For example, a setting which is keyed by major:minor device numbers
741 with integer values may look like the following::
743 # cat cgroup-example-interface-file
747 The default value can be updated by::
749 # echo 125 > cgroup-example-interface-file
753 # echo "default 125" > cgroup-example-interface-file
755 An override can be set by::
757 # echo "8:16 170" > cgroup-example-interface-file
761 # echo "8:0 default" > cgroup-example-interface-file
762 # cat cgroup-example-interface-file
766 - For events which are not very high frequency, an interface file
767 "events" should be created which lists event key value pairs.
768 Whenever a notifiable event happens, file modified event should be
769 generated on the file.
775 All cgroup core files are prefixed with "cgroup."
779 A read-write single value file which exists on non-root
782 When read, it indicates the current type of the cgroup, which
783 can be one of the following values.
785 - "domain" : A normal valid domain cgroup.
787 - "domain threaded" : A threaded domain cgroup which is
788 serving as the root of a threaded subtree.
790 - "domain invalid" : A cgroup which is in an invalid state.
791 It can't be populated or have controllers enabled. It may
792 be allowed to become a threaded cgroup.
794 - "threaded" : A threaded cgroup which is a member of a
797 A cgroup can be turned into a threaded cgroup by writing
798 "threaded" to this file.
801 A read-write new-line separated values file which exists on
804 When read, it lists the PIDs of all processes which belong to
805 the cgroup one-per-line. The PIDs are not ordered and the
806 same PID may show up more than once if the process got moved
807 to another cgroup and then back or the PID got recycled while
810 A PID can be written to migrate the process associated with
811 the PID to the cgroup. The writer should match all of the
812 following conditions.
814 - It must have write access to the "cgroup.procs" file.
816 - It must have write access to the "cgroup.procs" file of the
817 common ancestor of the source and destination cgroups.
819 When delegating a sub-hierarchy, write access to this file
820 should be granted along with the containing directory.
822 In a threaded cgroup, reading this file fails with EOPNOTSUPP
823 as all the processes belong to the thread root. Writing is
824 supported and moves every thread of the process to the cgroup.
827 A read-write new-line separated values file which exists on
830 When read, it lists the TIDs of all threads which belong to
831 the cgroup one-per-line. The TIDs are not ordered and the
832 same TID may show up more than once if the thread got moved to
833 another cgroup and then back or the TID got recycled while
836 A TID can be written to migrate the thread associated with the
837 TID to the cgroup. The writer should match all of the
838 following conditions.
840 - It must have write access to the "cgroup.threads" file.
842 - The cgroup that the thread is currently in must be in the
843 same resource domain as the destination cgroup.
845 - It must have write access to the "cgroup.procs" file of the
846 common ancestor of the source and destination cgroups.
848 When delegating a sub-hierarchy, write access to this file
849 should be granted along with the containing directory.
852 A read-only space separated values file which exists on all
855 It shows space separated list of all controllers available to
856 the cgroup. The controllers are not ordered.
858 cgroup.subtree_control
859 A read-write space separated values file which exists on all
860 cgroups. Starts out empty.
862 When read, it shows space separated list of the controllers
863 which are enabled to control resource distribution from the
864 cgroup to its children.
866 Space separated list of controllers prefixed with '+' or '-'
867 can be written to enable or disable controllers. A controller
868 name prefixed with '+' enables the controller and '-'
869 disables. If a controller appears more than once on the list,
870 the last one is effective. When multiple enable and disable
871 operations are specified, either all succeed or all fail.
874 A read-only flat-keyed file which exists on non-root cgroups.
875 The following entries are defined. Unless specified
876 otherwise, a value change in this file generates a file
880 1 if the cgroup or its descendants contains any live
881 processes; otherwise, 0.
883 1 if the cgroup is frozen; otherwise, 0.
885 cgroup.max.descendants
886 A read-write single value files. The default is "max".
888 Maximum allowed number of descent cgroups.
889 If the actual number of descendants is equal or larger,
890 an attempt to create a new cgroup in the hierarchy will fail.
893 A read-write single value files. The default is "max".
895 Maximum allowed descent depth below the current cgroup.
896 If the actual descent depth is equal or larger,
897 an attempt to create a new child cgroup will fail.
900 A read-only flat-keyed file with the following entries:
903 Total number of visible descendant cgroups.
906 Total number of dying descendant cgroups. A cgroup becomes
907 dying after being deleted by a user. The cgroup will remain
908 in dying state for some time undefined time (which can depend
909 on system load) before being completely destroyed.
911 A process can't enter a dying cgroup under any circumstances,
912 a dying cgroup can't revive.
914 A dying cgroup can consume system resources not exceeding
915 limits, which were active at the moment of cgroup deletion.
918 A read-write single value file which exists on non-root cgroups.
919 Allowed values are "0" and "1". The default is "0".
921 Writing "1" to the file causes freezing of the cgroup and all
922 descendant cgroups. This means that all belonging processes will
923 be stopped and will not run until the cgroup will be explicitly
924 unfrozen. Freezing of the cgroup may take some time; when this action
925 is completed, the "frozen" value in the cgroup.events control file
926 will be updated to "1" and the corresponding notification will be
929 A cgroup can be frozen either by its own settings, or by settings
930 of any ancestor cgroups. If any of ancestor cgroups is frozen, the
931 cgroup will remain frozen.
933 Processes in the frozen cgroup can be killed by a fatal signal.
934 They also can enter and leave a frozen cgroup: either by an explicit
935 move by a user, or if freezing of the cgroup races with fork().
936 If a process is moved to a frozen cgroup, it stops. If a process is
937 moved out of a frozen cgroup, it becomes running.
939 Frozen status of a cgroup doesn't affect any cgroup tree operations:
940 it's possible to delete a frozen (and empty) cgroup, as well as
941 create new sub-cgroups.
949 The "cpu" controllers regulates distribution of CPU cycles. This
950 controller implements weight and absolute bandwidth limit models for
951 normal scheduling policy and absolute bandwidth allocation model for
952 realtime scheduling policy.
954 In all the above models, cycles distribution is defined only on a temporal
955 base and it does not account for the frequency at which tasks are executed.
956 The (optional) utilization clamping support allows to hint the schedutil
957 cpufreq governor about the minimum desired frequency which should always be
958 provided by a CPU, as well as the maximum desired frequency, which should not
959 be exceeded by a CPU.
961 WARNING: cgroup2 doesn't yet support control of realtime processes and
962 the cpu controller can only be enabled when all RT processes are in
963 the root cgroup. Be aware that system management software may already
964 have placed RT processes into nonroot cgroups during the system boot
965 process, and these processes may need to be moved to the root cgroup
966 before the cpu controller can be enabled.
972 All time durations are in microseconds.
975 A read-only flat-keyed file which exists on non-root cgroups.
976 This file exists whether the controller is enabled or not.
978 It always reports the following three stats:
984 and the following three when the controller is enabled:
991 A read-write single value file which exists on non-root
992 cgroups. The default is "100".
994 The weight in the range [1, 10000].
997 A read-write single value file which exists on non-root
998 cgroups. The default is "0".
1000 The nice value is in the range [-20, 19].
1002 This interface file is an alternative interface for
1003 "cpu.weight" and allows reading and setting weight using the
1004 same values used by nice(2). Because the range is smaller and
1005 granularity is coarser for the nice values, the read value is
1006 the closest approximation of the current weight.
1009 A read-write two value file which exists on non-root cgroups.
1010 The default is "max 100000".
1012 The maximum bandwidth limit. It's in the following format::
1016 which indicates that the group may consume upto $MAX in each
1017 $PERIOD duration. "max" for $MAX indicates no limit. If only
1018 one number is written, $MAX is updated.
1021 A read-only nested-key file which exists on non-root cgroups.
1023 Shows pressure stall information for CPU. See
1024 Documentation/accounting/psi.rst for details.
1027 A read-write single value file which exists on non-root cgroups.
1028 The default is "0", i.e. no utilization boosting.
1030 The requested minimum utilization (protection) as a percentage
1031 rational number, e.g. 12.34 for 12.34%.
1033 This interface allows reading and setting minimum utilization clamp
1034 values similar to the sched_setattr(2). This minimum utilization
1035 value is used to clamp the task specific minimum utilization clamp.
1037 The requested minimum utilization (protection) is always capped by
1038 the current value for the maximum utilization (limit), i.e.
1042 A read-write single value file which exists on non-root cgroups.
1043 The default is "max". i.e. no utilization capping
1045 The requested maximum utilization (limit) as a percentage rational
1046 number, e.g. 98.76 for 98.76%.
1048 This interface allows reading and setting maximum utilization clamp
1049 values similar to the sched_setattr(2). This maximum utilization
1050 value is used to clamp the task specific maximum utilization clamp.
1057 The "memory" controller regulates distribution of memory. Memory is
1058 stateful and implements both limit and protection models. Due to the
1059 intertwining between memory usage and reclaim pressure and the
1060 stateful nature of memory, the distribution model is relatively
1063 While not completely water-tight, all major memory usages by a given
1064 cgroup are tracked so that the total memory consumption can be
1065 accounted and controlled to a reasonable extent. Currently, the
1066 following types of memory usages are tracked.
1068 - Userland memory - page cache and anonymous memory.
1070 - Kernel data structures such as dentries and inodes.
1072 - TCP socket buffers.
1074 The above list may expand in the future for better coverage.
1077 Memory Interface Files
1078 ~~~~~~~~~~~~~~~~~~~~~~
1080 All memory amounts are in bytes. If a value which is not aligned to
1081 PAGE_SIZE is written, the value may be rounded up to the closest
1082 PAGE_SIZE multiple when read back.
1085 A read-only single value file which exists on non-root
1088 The total amount of memory currently being used by the cgroup
1089 and its descendants.
1092 A read-write single value file which exists on non-root
1093 cgroups. The default is "0".
1095 Hard memory protection. If the memory usage of a cgroup
1096 is within its effective min boundary, the cgroup's memory
1097 won't be reclaimed under any conditions. If there is no
1098 unprotected reclaimable memory available, OOM killer
1099 is invoked. Above the effective min boundary (or
1100 effective low boundary if it is higher), pages are reclaimed
1101 proportionally to the overage, reducing reclaim pressure for
1104 Effective min boundary is limited by memory.min values of
1105 all ancestor cgroups. If there is memory.min overcommitment
1106 (child cgroup or cgroups are requiring more protected memory
1107 than parent will allow), then each child cgroup will get
1108 the part of parent's protection proportional to its
1109 actual memory usage below memory.min.
1111 Putting more memory than generally available under this
1112 protection is discouraged and may lead to constant OOMs.
1114 If a memory cgroup is not populated with processes,
1115 its memory.min is ignored.
1118 A read-write single value file which exists on non-root
1119 cgroups. The default is "0".
1121 Best-effort memory protection. If the memory usage of a
1122 cgroup is within its effective low boundary, the cgroup's
1123 memory won't be reclaimed unless there is no reclaimable
1124 memory available in unprotected cgroups.
1125 Above the effective low boundary (or
1126 effective min boundary if it is higher), pages are reclaimed
1127 proportionally to the overage, reducing reclaim pressure for
1130 Effective low boundary is limited by memory.low values of
1131 all ancestor cgroups. If there is memory.low overcommitment
1132 (child cgroup or cgroups are requiring more protected memory
1133 than parent will allow), then each child cgroup will get
1134 the part of parent's protection proportional to its
1135 actual memory usage below memory.low.
1137 Putting more memory than generally available under this
1138 protection is discouraged.
1141 A read-write single value file which exists on non-root
1142 cgroups. The default is "max".
1144 Memory usage throttle limit. This is the main mechanism to
1145 control memory usage of a cgroup. If a cgroup's usage goes
1146 over the high boundary, the processes of the cgroup are
1147 throttled and put under heavy reclaim pressure.
1149 Going over the high limit never invokes the OOM killer and
1150 under extreme conditions the limit may be breached.
1153 A read-write single value file which exists on non-root
1154 cgroups. The default is "max".
1156 Memory usage hard limit. This is the final protection
1157 mechanism. If a cgroup's memory usage reaches this limit and
1158 can't be reduced, the OOM killer is invoked in the cgroup.
1159 Under certain circumstances, the usage may go over the limit
1162 This is the ultimate protection mechanism. As long as the
1163 high limit is used and monitored properly, this limit's
1164 utility is limited to providing the final safety net.
1167 A read-write single value file which exists on non-root
1168 cgroups. The default value is "0".
1170 Determines whether the cgroup should be treated as
1171 an indivisible workload by the OOM killer. If set,
1172 all tasks belonging to the cgroup or to its descendants
1173 (if the memory cgroup is not a leaf cgroup) are killed
1174 together or not at all. This can be used to avoid
1175 partial kills to guarantee workload integrity.
1177 Tasks with the OOM protection (oom_score_adj set to -1000)
1178 are treated as an exception and are never killed.
1180 If the OOM killer is invoked in a cgroup, it's not going
1181 to kill any tasks outside of this cgroup, regardless
1182 memory.oom.group values of ancestor cgroups.
1185 A read-only flat-keyed file which exists on non-root cgroups.
1186 The following entries are defined. Unless specified
1187 otherwise, a value change in this file generates a file
1190 Note that all fields in this file are hierarchical and the
1191 file modified event can be generated due to an event down the
1192 hierarchy. For for the local events at the cgroup level see
1193 memory.events.local.
1196 The number of times the cgroup is reclaimed due to
1197 high memory pressure even though its usage is under
1198 the low boundary. This usually indicates that the low
1199 boundary is over-committed.
1202 The number of times processes of the cgroup are
1203 throttled and routed to perform direct memory reclaim
1204 because the high memory boundary was exceeded. For a
1205 cgroup whose memory usage is capped by the high limit
1206 rather than global memory pressure, this event's
1207 occurrences are expected.
1210 The number of times the cgroup's memory usage was
1211 about to go over the max boundary. If direct reclaim
1212 fails to bring it down, the cgroup goes to OOM state.
1215 The number of time the cgroup's memory usage was
1216 reached the limit and allocation was about to fail.
1218 Depending on context result could be invocation of OOM
1219 killer and retrying allocation or failing allocation.
1221 Failed allocation in its turn could be returned into
1222 userspace as -ENOMEM or silently ignored in cases like
1223 disk readahead. For now OOM in memory cgroup kills
1224 tasks iff shortage has happened inside page fault.
1226 This event is not raised if the OOM killer is not
1227 considered as an option, e.g. for failed high-order
1231 The number of processes belonging to this cgroup
1232 killed by any kind of OOM killer.
1235 Similar to memory.events but the fields in the file are local
1236 to the cgroup i.e. not hierarchical. The file modified event
1237 generated on this file reflects only the local events.
1240 A read-only flat-keyed file which exists on non-root cgroups.
1242 This breaks down the cgroup's memory footprint into different
1243 types of memory, type-specific details, and other information
1244 on the state and past events of the memory management system.
1246 All memory amounts are in bytes.
1248 The entries are ordered to be human readable, and new entries
1249 can show up in the middle. Don't rely on items remaining in a
1250 fixed position; use the keys to look up specific values!
1253 Amount of memory used in anonymous mappings such as
1254 brk(), sbrk(), and mmap(MAP_ANONYMOUS)
1257 Amount of memory used to cache filesystem data,
1258 including tmpfs and shared memory.
1261 Amount of memory allocated to kernel stacks.
1264 Amount of memory used for storing in-kernel data
1268 Amount of memory used in network transmission buffers
1271 Amount of cached filesystem data that is swap-backed,
1272 such as tmpfs, shm segments, shared anonymous mmap()s
1275 Amount of cached filesystem data mapped with mmap()
1278 Amount of cached filesystem data that was modified but
1279 not yet written back to disk
1282 Amount of cached filesystem data that was modified and
1283 is currently being written back to disk
1286 Amount of memory used in anonymous mappings backed by
1287 transparent hugepages
1289 inactive_anon, active_anon, inactive_file, active_file, unevictable
1290 Amount of memory, swap-backed and filesystem-backed,
1291 on the internal memory management lists used by the
1292 page reclaim algorithm
1295 Part of "slab" that might be reclaimed, such as
1296 dentries and inodes.
1299 Part of "slab" that cannot be reclaimed on memory
1303 Total number of page faults incurred
1306 Number of major page faults incurred
1310 Number of refaults of previously evicted pages
1314 Number of refaulted pages that were immediately activated
1316 workingset_nodereclaim
1318 Number of times a shadow node has been reclaimed
1322 Amount of scanned pages (in an active LRU list)
1326 Amount of scanned pages (in an inactive LRU list)
1330 Amount of reclaimed pages
1334 Amount of pages moved to the active LRU list
1338 Amount of pages moved to the inactive LRU lis
1342 Amount of pages postponed to be freed under memory pressure
1346 Amount of reclaimed lazyfree pages
1350 Number of transparent hugepages which were allocated to satisfy
1351 a page fault, including COW faults. This counter is not present
1352 when CONFIG_TRANSPARENT_HUGEPAGE is not set.
1356 Number of transparent hugepages which were allocated to allow
1357 collapsing an existing range of pages. This counter is not
1358 present when CONFIG_TRANSPARENT_HUGEPAGE is not set.
1361 A read-only single value file which exists on non-root
1364 The total amount of swap currently being used by the cgroup
1365 and its descendants.
1368 A read-write single value file which exists on non-root
1369 cgroups. The default is "max".
1371 Swap usage hard limit. If a cgroup's swap usage reaches this
1372 limit, anonymous memory of the cgroup will not be swapped out.
1375 A read-only flat-keyed file which exists on non-root cgroups.
1376 The following entries are defined. Unless specified
1377 otherwise, a value change in this file generates a file
1381 The number of times the cgroup's swap usage was about
1382 to go over the max boundary and swap allocation
1386 The number of times swap allocation failed either
1387 because of running out of swap system-wide or max
1390 When reduced under the current usage, the existing swap
1391 entries are reclaimed gradually and the swap usage may stay
1392 higher than the limit for an extended period of time. This
1393 reduces the impact on the workload and memory management.
1396 A read-only nested-key file which exists on non-root cgroups.
1398 Shows pressure stall information for memory. See
1399 Documentation/accounting/psi.rst for details.
1405 "memory.high" is the main mechanism to control memory usage.
1406 Over-committing on high limit (sum of high limits > available memory)
1407 and letting global memory pressure to distribute memory according to
1408 usage is a viable strategy.
1410 Because breach of the high limit doesn't trigger the OOM killer but
1411 throttles the offending cgroup, a management agent has ample
1412 opportunities to monitor and take appropriate actions such as granting
1413 more memory or terminating the workload.
1415 Determining whether a cgroup has enough memory is not trivial as
1416 memory usage doesn't indicate whether the workload can benefit from
1417 more memory. For example, a workload which writes data received from
1418 network to a file can use all available memory but can also operate as
1419 performant with a small amount of memory. A measure of memory
1420 pressure - how much the workload is being impacted due to lack of
1421 memory - is necessary to determine whether a workload needs more
1422 memory; unfortunately, memory pressure monitoring mechanism isn't
1429 A memory area is charged to the cgroup which instantiated it and stays
1430 charged to the cgroup until the area is released. Migrating a process
1431 to a different cgroup doesn't move the memory usages that it
1432 instantiated while in the previous cgroup to the new cgroup.
1434 A memory area may be used by processes belonging to different cgroups.
1435 To which cgroup the area will be charged is in-deterministic; however,
1436 over time, the memory area is likely to end up in a cgroup which has
1437 enough memory allowance to avoid high reclaim pressure.
1439 If a cgroup sweeps a considerable amount of memory which is expected
1440 to be accessed repeatedly by other cgroups, it may make sense to use
1441 POSIX_FADV_DONTNEED to relinquish the ownership of memory areas
1442 belonging to the affected files to ensure correct memory ownership.
1448 The "io" controller regulates the distribution of IO resources. This
1449 controller implements both weight based and absolute bandwidth or IOPS
1450 limit distribution; however, weight based distribution is available
1451 only if cfq-iosched is in use and neither scheme is available for
1459 A read-only nested-keyed file which exists on non-root
1462 Lines are keyed by $MAJ:$MIN device numbers and not ordered.
1463 The following nested keys are defined.
1465 ====== =====================
1467 wbytes Bytes written
1468 rios Number of read IOs
1469 wios Number of write IOs
1470 dbytes Bytes discarded
1471 dios Number of discard IOs
1472 ====== =====================
1474 An example read output follows:
1476 8:16 rbytes=1459200 wbytes=314773504 rios=192 wios=353 dbytes=0 dios=0
1477 8:0 rbytes=90430464 wbytes=299008000 rios=8950 wios=1252 dbytes=50331648 dios=3021
1480 A read-write nested-keyed file with exists only on the root
1483 This file configures the Quality of Service of the IO cost
1484 model based controller (CONFIG_BLK_CGROUP_IOCOST) which
1485 currently implements "io.weight" proportional control. Lines
1486 are keyed by $MAJ:$MIN device numbers and not ordered. The
1487 line for a given device is populated on the first write for
1488 the device on "io.cost.qos" or "io.cost.model". The following
1489 nested keys are defined.
1491 ====== =====================================
1492 enable Weight-based control enable
1493 ctrl "auto" or "user"
1494 rpct Read latency percentile [0, 100]
1495 rlat Read latency threshold
1496 wpct Write latency percentile [0, 100]
1497 wlat Write latency threshold
1498 min Minimum scaling percentage [1, 10000]
1499 max Maximum scaling percentage [1, 10000]
1500 ====== =====================================
1502 The controller is disabled by default and can be enabled by
1503 setting "enable" to 1. "rpct" and "wpct" parameters default
1504 to zero and the controller uses internal device saturation
1505 state to adjust the overall IO rate between "min" and "max".
1507 When a better control quality is needed, latency QoS
1508 parameters can be configured. For example::
1510 8:16 enable=1 ctrl=auto rpct=95.00 rlat=75000 wpct=95.00 wlat=150000 min=50.00 max=150.0
1512 shows that on sdb, the controller is enabled, will consider
1513 the device saturated if the 95th percentile of read completion
1514 latencies is above 75ms or write 150ms, and adjust the overall
1515 IO issue rate between 50% and 150% accordingly.
1517 The lower the saturation point, the better the latency QoS at
1518 the cost of aggregate bandwidth. The narrower the allowed
1519 adjustment range between "min" and "max", the more conformant
1520 to the cost model the IO behavior. Note that the IO issue
1521 base rate may be far off from 100% and setting "min" and "max"
1522 blindly can lead to a significant loss of device capacity or
1523 control quality. "min" and "max" are useful for regulating
1524 devices which show wide temporary behavior changes - e.g. a
1525 ssd which accepts writes at the line speed for a while and
1526 then completely stalls for multiple seconds.
1528 When "ctrl" is "auto", the parameters are controlled by the
1529 kernel and may change automatically. Setting "ctrl" to "user"
1530 or setting any of the percentile and latency parameters puts
1531 it into "user" mode and disables the automatic changes. The
1532 automatic mode can be restored by setting "ctrl" to "auto".
1535 A read-write nested-keyed file with exists only on the root
1538 This file configures the cost model of the IO cost model based
1539 controller (CONFIG_BLK_CGROUP_IOCOST) which currently
1540 implements "io.weight" proportional control. Lines are keyed
1541 by $MAJ:$MIN device numbers and not ordered. The line for a
1542 given device is populated on the first write for the device on
1543 "io.cost.qos" or "io.cost.model". The following nested keys
1546 ===== ================================
1547 ctrl "auto" or "user"
1548 model The cost model in use - "linear"
1549 ===== ================================
1551 When "ctrl" is "auto", the kernel may change all parameters
1552 dynamically. When "ctrl" is set to "user" or any other
1553 parameters are written to, "ctrl" become "user" and the
1554 automatic changes are disabled.
1556 When "model" is "linear", the following model parameters are
1559 ============= ========================================
1560 [r|w]bps The maximum sequential IO throughput
1561 [r|w]seqiops The maximum 4k sequential IOs per second
1562 [r|w]randiops The maximum 4k random IOs per second
1563 ============= ========================================
1565 From the above, the builtin linear model determines the base
1566 costs of a sequential and random IO and the cost coefficient
1567 for the IO size. While simple, this model can cover most
1568 common device classes acceptably.
1570 The IO cost model isn't expected to be accurate in absolute
1571 sense and is scaled to the device behavior dynamically.
1573 If needed, tools/cgroup/iocost_coef_gen.py can be used to
1574 generate device-specific coefficients.
1577 A read-write flat-keyed file which exists on non-root cgroups.
1578 The default is "default 100".
1580 The first line is the default weight applied to devices
1581 without specific override. The rest are overrides keyed by
1582 $MAJ:$MIN device numbers and not ordered. The weights are in
1583 the range [1, 10000] and specifies the relative amount IO time
1584 the cgroup can use in relation to its siblings.
1586 The default weight can be updated by writing either "default
1587 $WEIGHT" or simply "$WEIGHT". Overrides can be set by writing
1588 "$MAJ:$MIN $WEIGHT" and unset by writing "$MAJ:$MIN default".
1590 An example read output follows::
1597 A read-write nested-keyed file which exists on non-root
1600 BPS and IOPS based IO limit. Lines are keyed by $MAJ:$MIN
1601 device numbers and not ordered. The following nested keys are
1604 ===== ==================================
1605 rbps Max read bytes per second
1606 wbps Max write bytes per second
1607 riops Max read IO operations per second
1608 wiops Max write IO operations per second
1609 ===== ==================================
1611 When writing, any number of nested key-value pairs can be
1612 specified in any order. "max" can be specified as the value
1613 to remove a specific limit. If the same key is specified
1614 multiple times, the outcome is undefined.
1616 BPS and IOPS are measured in each IO direction and IOs are
1617 delayed if limit is reached. Temporary bursts are allowed.
1619 Setting read limit at 2M BPS and write at 120 IOPS for 8:16::
1621 echo "8:16 rbps=2097152 wiops=120" > io.max
1623 Reading returns the following::
1625 8:16 rbps=2097152 wbps=max riops=max wiops=120
1627 Write IOPS limit can be removed by writing the following::
1629 echo "8:16 wiops=max" > io.max
1631 Reading now returns the following::
1633 8:16 rbps=2097152 wbps=max riops=max wiops=max
1636 A read-only nested-key file which exists on non-root cgroups.
1638 Shows pressure stall information for IO. See
1639 Documentation/accounting/psi.rst for details.
1645 Page cache is dirtied through buffered writes and shared mmaps and
1646 written asynchronously to the backing filesystem by the writeback
1647 mechanism. Writeback sits between the memory and IO domains and
1648 regulates the proportion of dirty memory by balancing dirtying and
1651 The io controller, in conjunction with the memory controller,
1652 implements control of page cache writeback IOs. The memory controller
1653 defines the memory domain that dirty memory ratio is calculated and
1654 maintained for and the io controller defines the io domain which
1655 writes out dirty pages for the memory domain. Both system-wide and
1656 per-cgroup dirty memory states are examined and the more restrictive
1657 of the two is enforced.
1659 cgroup writeback requires explicit support from the underlying
1660 filesystem. Currently, cgroup writeback is implemented on ext2, ext4
1661 and btrfs. On other filesystems, all writeback IOs are attributed to
1664 There are inherent differences in memory and writeback management
1665 which affects how cgroup ownership is tracked. Memory is tracked per
1666 page while writeback per inode. For the purpose of writeback, an
1667 inode is assigned to a cgroup and all IO requests to write dirty pages
1668 from the inode are attributed to that cgroup.
1670 As cgroup ownership for memory is tracked per page, there can be pages
1671 which are associated with different cgroups than the one the inode is
1672 associated with. These are called foreign pages. The writeback
1673 constantly keeps track of foreign pages and, if a particular foreign
1674 cgroup becomes the majority over a certain period of time, switches
1675 the ownership of the inode to that cgroup.
1677 While this model is enough for most use cases where a given inode is
1678 mostly dirtied by a single cgroup even when the main writing cgroup
1679 changes over time, use cases where multiple cgroups write to a single
1680 inode simultaneously are not supported well. In such circumstances, a
1681 significant portion of IOs are likely to be attributed incorrectly.
1682 As memory controller assigns page ownership on the first use and
1683 doesn't update it until the page is released, even if writeback
1684 strictly follows page ownership, multiple cgroups dirtying overlapping
1685 areas wouldn't work as expected. It's recommended to avoid such usage
1688 The sysctl knobs which affect writeback behavior are applied to cgroup
1689 writeback as follows.
1691 vm.dirty_background_ratio, vm.dirty_ratio
1692 These ratios apply the same to cgroup writeback with the
1693 amount of available memory capped by limits imposed by the
1694 memory controller and system-wide clean memory.
1696 vm.dirty_background_bytes, vm.dirty_bytes
1697 For cgroup writeback, this is calculated into ratio against
1698 total available memory and applied the same way as
1699 vm.dirty[_background]_ratio.
1705 This is a cgroup v2 controller for IO workload protection. You provide a group
1706 with a latency target, and if the average latency exceeds that target the
1707 controller will throttle any peers that have a lower latency target than the
1710 The limits are only applied at the peer level in the hierarchy. This means that
1711 in the diagram below, only groups A, B, and C will influence each other, and
1712 groups D and F will influence each other. Group G will influence nobody::
1721 So the ideal way to configure this is to set io.latency in groups A, B, and C.
1722 Generally you do not want to set a value lower than the latency your device
1723 supports. Experiment to find the value that works best for your workload.
1724 Start at higher than the expected latency for your device and watch the
1725 avg_lat value in io.stat for your workload group to get an idea of the
1726 latency you see during normal operation. Use the avg_lat value as a basis for
1727 your real setting, setting at 10-15% higher than the value in io.stat.
1729 How IO Latency Throttling Works
1730 ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
1732 io.latency is work conserving; so as long as everybody is meeting their latency
1733 target the controller doesn't do anything. Once a group starts missing its
1734 target it begins throttling any peer group that has a higher target than itself.
1735 This throttling takes 2 forms:
1737 - Queue depth throttling. This is the number of outstanding IO's a group is
1738 allowed to have. We will clamp down relatively quickly, starting at no limit
1739 and going all the way down to 1 IO at a time.
1741 - Artificial delay induction. There are certain types of IO that cannot be
1742 throttled without possibly adversely affecting higher priority groups. This
1743 includes swapping and metadata IO. These types of IO are allowed to occur
1744 normally, however they are "charged" to the originating group. If the
1745 originating group is being throttled you will see the use_delay and delay
1746 fields in io.stat increase. The delay value is how many microseconds that are
1747 being added to any process that runs in this group. Because this number can
1748 grow quite large if there is a lot of swapping or metadata IO occurring we
1749 limit the individual delay events to 1 second at a time.
1751 Once the victimized group starts meeting its latency target again it will start
1752 unthrottling any peer groups that were throttled previously. If the victimized
1753 group simply stops doing IO the global counter will unthrottle appropriately.
1755 IO Latency Interface Files
1756 ~~~~~~~~~~~~~~~~~~~~~~~~~~
1759 This takes a similar format as the other controllers.
1761 "MAJOR:MINOR target=<target time in microseconds"
1764 If the controller is enabled you will see extra stats in io.stat in
1765 addition to the normal ones.
1768 This is the current queue depth for the group.
1771 This is an exponential moving average with a decay rate of 1/exp
1772 bound by the sampling interval. The decay rate interval can be
1773 calculated by multiplying the win value in io.stat by the
1774 corresponding number of samples based on the win value.
1777 The sampling window size in milliseconds. This is the minimum
1778 duration of time between evaluation events. Windows only elapse
1779 with IO activity. Idle periods extend the most recent window.
1784 The process number controller is used to allow a cgroup to stop any
1785 new tasks from being fork()'d or clone()'d after a specified limit is
1788 The number of tasks in a cgroup can be exhausted in ways which other
1789 controllers cannot prevent, thus warranting its own controller. For
1790 example, a fork bomb is likely to exhaust the number of tasks before
1791 hitting memory restrictions.
1793 Note that PIDs used in this controller refer to TIDs, process IDs as
1801 A read-write single value file which exists on non-root
1802 cgroups. The default is "max".
1804 Hard limit of number of processes.
1807 A read-only single value file which exists on all cgroups.
1809 The number of processes currently in the cgroup and its
1812 Organisational operations are not blocked by cgroup policies, so it is
1813 possible to have pids.current > pids.max. This can be done by either
1814 setting the limit to be smaller than pids.current, or attaching enough
1815 processes to the cgroup such that pids.current is larger than
1816 pids.max. However, it is not possible to violate a cgroup PID policy
1817 through fork() or clone(). These will return -EAGAIN if the creation
1818 of a new process would cause a cgroup policy to be violated.
1824 The "cpuset" controller provides a mechanism for constraining
1825 the CPU and memory node placement of tasks to only the resources
1826 specified in the cpuset interface files in a task's current cgroup.
1827 This is especially valuable on large NUMA systems where placing jobs
1828 on properly sized subsets of the systems with careful processor and
1829 memory placement to reduce cross-node memory access and contention
1830 can improve overall system performance.
1832 The "cpuset" controller is hierarchical. That means the controller
1833 cannot use CPUs or memory nodes not allowed in its parent.
1836 Cpuset Interface Files
1837 ~~~~~~~~~~~~~~~~~~~~~~
1840 A read-write multiple values file which exists on non-root
1841 cpuset-enabled cgroups.
1843 It lists the requested CPUs to be used by tasks within this
1844 cgroup. The actual list of CPUs to be granted, however, is
1845 subjected to constraints imposed by its parent and can differ
1846 from the requested CPUs.
1848 The CPU numbers are comma-separated numbers or ranges.
1854 An empty value indicates that the cgroup is using the same
1855 setting as the nearest cgroup ancestor with a non-empty
1856 "cpuset.cpus" or all the available CPUs if none is found.
1858 The value of "cpuset.cpus" stays constant until the next update
1859 and won't be affected by any CPU hotplug events.
1861 cpuset.cpus.effective
1862 A read-only multiple values file which exists on all
1863 cpuset-enabled cgroups.
1865 It lists the onlined CPUs that are actually granted to this
1866 cgroup by its parent. These CPUs are allowed to be used by
1867 tasks within the current cgroup.
1869 If "cpuset.cpus" is empty, the "cpuset.cpus.effective" file shows
1870 all the CPUs from the parent cgroup that can be available to
1871 be used by this cgroup. Otherwise, it should be a subset of
1872 "cpuset.cpus" unless none of the CPUs listed in "cpuset.cpus"
1873 can be granted. In this case, it will be treated just like an
1874 empty "cpuset.cpus".
1876 Its value will be affected by CPU hotplug events.
1879 A read-write multiple values file which exists on non-root
1880 cpuset-enabled cgroups.
1882 It lists the requested memory nodes to be used by tasks within
1883 this cgroup. The actual list of memory nodes granted, however,
1884 is subjected to constraints imposed by its parent and can differ
1885 from the requested memory nodes.
1887 The memory node numbers are comma-separated numbers or ranges.
1893 An empty value indicates that the cgroup is using the same
1894 setting as the nearest cgroup ancestor with a non-empty
1895 "cpuset.mems" or all the available memory nodes if none
1898 The value of "cpuset.mems" stays constant until the next update
1899 and won't be affected by any memory nodes hotplug events.
1901 cpuset.mems.effective
1902 A read-only multiple values file which exists on all
1903 cpuset-enabled cgroups.
1905 It lists the onlined memory nodes that are actually granted to
1906 this cgroup by its parent. These memory nodes are allowed to
1907 be used by tasks within the current cgroup.
1909 If "cpuset.mems" is empty, it shows all the memory nodes from the
1910 parent cgroup that will be available to be used by this cgroup.
1911 Otherwise, it should be a subset of "cpuset.mems" unless none of
1912 the memory nodes listed in "cpuset.mems" can be granted. In this
1913 case, it will be treated just like an empty "cpuset.mems".
1915 Its value will be affected by memory nodes hotplug events.
1917 cpuset.cpus.partition
1918 A read-write single value file which exists on non-root
1919 cpuset-enabled cgroups. This flag is owned by the parent cgroup
1920 and is not delegatable.
1922 It accepts only the following input values when written to.
1924 "root" - a partition root
1925 "member" - a non-root member of a partition
1927 When set to be a partition root, the current cgroup is the
1928 root of a new partition or scheduling domain that comprises
1929 itself and all its descendants except those that are separate
1930 partition roots themselves and their descendants. The root
1931 cgroup is always a partition root.
1933 There are constraints on where a partition root can be set.
1934 It can only be set in a cgroup if all the following conditions
1937 1) The "cpuset.cpus" is not empty and the list of CPUs are
1938 exclusive, i.e. they are not shared by any of its siblings.
1939 2) The parent cgroup is a partition root.
1940 3) The "cpuset.cpus" is also a proper subset of the parent's
1941 "cpuset.cpus.effective".
1942 4) There is no child cgroups with cpuset enabled. This is for
1943 eliminating corner cases that have to be handled if such a
1944 condition is allowed.
1946 Setting it to partition root will take the CPUs away from the
1947 effective CPUs of the parent cgroup. Once it is set, this
1948 file cannot be reverted back to "member" if there are any child
1949 cgroups with cpuset enabled.
1951 A parent partition cannot distribute all its CPUs to its
1952 child partitions. There must be at least one cpu left in the
1955 Once becoming a partition root, changes to "cpuset.cpus" is
1956 generally allowed as long as the first condition above is true,
1957 the change will not take away all the CPUs from the parent
1958 partition and the new "cpuset.cpus" value is a superset of its
1959 children's "cpuset.cpus" values.
1961 Sometimes, external factors like changes to ancestors'
1962 "cpuset.cpus" or cpu hotplug can cause the state of the partition
1963 root to change. On read, the "cpuset.sched.partition" file
1964 can show the following values.
1966 "member" Non-root member of a partition
1967 "root" Partition root
1968 "root invalid" Invalid partition root
1970 It is a partition root if the first 2 partition root conditions
1971 above are true and at least one CPU from "cpuset.cpus" is
1972 granted by the parent cgroup.
1974 A partition root can become invalid if none of CPUs requested
1975 in "cpuset.cpus" can be granted by the parent cgroup or the
1976 parent cgroup is no longer a partition root itself. In this
1977 case, it is not a real partition even though the restriction
1978 of the first partition root condition above will still apply.
1979 The cpu affinity of all the tasks in the cgroup will then be
1980 associated with CPUs in the nearest ancestor partition.
1982 An invalid partition root can be transitioned back to a
1983 real partition root if at least one of the requested CPUs
1984 can now be granted by its parent. In this case, the cpu
1985 affinity of all the tasks in the formerly invalid partition
1986 will be associated to the CPUs of the newly formed partition.
1987 Changing the partition state of an invalid partition root to
1988 "member" is always allowed even if child cpusets are present.
1994 Device controller manages access to device files. It includes both
1995 creation of new device files (using mknod), and access to the
1996 existing device files.
1998 Cgroup v2 device controller has no interface files and is implemented
1999 on top of cgroup BPF. To control access to device files, a user may
2000 create bpf programs of the BPF_CGROUP_DEVICE type and attach them
2001 to cgroups. On an attempt to access a device file, corresponding
2002 BPF programs will be executed, and depending on the return value
2003 the attempt will succeed or fail with -EPERM.
2005 A BPF_CGROUP_DEVICE program takes a pointer to the bpf_cgroup_dev_ctx
2006 structure, which describes the device access attempt: access type
2007 (mknod/read/write) and device (type, major and minor numbers).
2008 If the program returns 0, the attempt fails with -EPERM, otherwise
2011 An example of BPF_CGROUP_DEVICE program may be found in the kernel
2012 source tree in the tools/testing/selftests/bpf/dev_cgroup.c file.
2018 The "rdma" controller regulates the distribution and accounting of
2021 RDMA Interface Files
2022 ~~~~~~~~~~~~~~~~~~~~
2025 A readwrite nested-keyed file that exists for all the cgroups
2026 except root that describes current configured resource limit
2027 for a RDMA/IB device.
2029 Lines are keyed by device name and are not ordered.
2030 Each line contains space separated resource name and its configured
2031 limit that can be distributed.
2033 The following nested keys are defined.
2035 ========== =============================
2036 hca_handle Maximum number of HCA Handles
2037 hca_object Maximum number of HCA Objects
2038 ========== =============================
2040 An example for mlx4 and ocrdma device follows::
2042 mlx4_0 hca_handle=2 hca_object=2000
2043 ocrdma1 hca_handle=3 hca_object=max
2046 A read-only file that describes current resource usage.
2047 It exists for all the cgroup except root.
2049 An example for mlx4 and ocrdma device follows::
2051 mlx4_0 hca_handle=1 hca_object=20
2052 ocrdma1 hca_handle=1 hca_object=23
2061 perf_event controller, if not mounted on a legacy hierarchy, is
2062 automatically enabled on the v2 hierarchy so that perf events can
2063 always be filtered by cgroup v2 path. The controller can still be
2064 moved to a legacy hierarchy after v2 hierarchy is populated.
2067 Non-normative information
2068 -------------------------
2070 This section contains information that isn't considered to be a part of
2071 the stable kernel API and so is subject to change.
2074 CPU controller root cgroup process behaviour
2075 ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
2077 When distributing CPU cycles in the root cgroup each thread in this
2078 cgroup is treated as if it was hosted in a separate child cgroup of the
2079 root cgroup. This child cgroup weight is dependent on its thread nice
2082 For details of this mapping see sched_prio_to_weight array in
2083 kernel/sched/core.c file (values from this array should be scaled
2084 appropriately so the neutral - nice 0 - value is 100 instead of 1024).
2087 IO controller root cgroup process behaviour
2088 ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
2090 Root cgroup processes are hosted in an implicit leaf child node.
2091 When distributing IO resources this implicit child node is taken into
2092 account as if it was a normal child cgroup of the root cgroup with a
2093 weight value of 200.
2102 cgroup namespace provides a mechanism to virtualize the view of the
2103 "/proc/$PID/cgroup" file and cgroup mounts. The CLONE_NEWCGROUP clone
2104 flag can be used with clone(2) and unshare(2) to create a new cgroup
2105 namespace. The process running inside the cgroup namespace will have
2106 its "/proc/$PID/cgroup" output restricted to cgroupns root. The
2107 cgroupns root is the cgroup of the process at the time of creation of
2108 the cgroup namespace.
2110 Without cgroup namespace, the "/proc/$PID/cgroup" file shows the
2111 complete path of the cgroup of a process. In a container setup where
2112 a set of cgroups and namespaces are intended to isolate processes the
2113 "/proc/$PID/cgroup" file may leak potential system level information
2114 to the isolated processes. For Example::
2116 # cat /proc/self/cgroup
2117 0::/batchjobs/container_id1
2119 The path '/batchjobs/container_id1' can be considered as system-data
2120 and undesirable to expose to the isolated processes. cgroup namespace
2121 can be used to restrict visibility of this path. For example, before
2122 creating a cgroup namespace, one would see::
2124 # ls -l /proc/self/ns/cgroup
2125 lrwxrwxrwx 1 root root 0 2014-07-15 10:37 /proc/self/ns/cgroup -> cgroup:[4026531835]
2126 # cat /proc/self/cgroup
2127 0::/batchjobs/container_id1
2129 After unsharing a new namespace, the view changes::
2131 # ls -l /proc/self/ns/cgroup
2132 lrwxrwxrwx 1 root root 0 2014-07-15 10:35 /proc/self/ns/cgroup -> cgroup:[4026532183]
2133 # cat /proc/self/cgroup
2136 When some thread from a multi-threaded process unshares its cgroup
2137 namespace, the new cgroupns gets applied to the entire process (all
2138 the threads). This is natural for the v2 hierarchy; however, for the
2139 legacy hierarchies, this may be unexpected.
2141 A cgroup namespace is alive as long as there are processes inside or
2142 mounts pinning it. When the last usage goes away, the cgroup
2143 namespace is destroyed. The cgroupns root and the actual cgroups
2150 The 'cgroupns root' for a cgroup namespace is the cgroup in which the
2151 process calling unshare(2) is running. For example, if a process in
2152 /batchjobs/container_id1 cgroup calls unshare, cgroup
2153 /batchjobs/container_id1 becomes the cgroupns root. For the
2154 init_cgroup_ns, this is the real root ('/') cgroup.
2156 The cgroupns root cgroup does not change even if the namespace creator
2157 process later moves to a different cgroup::
2159 # ~/unshare -c # unshare cgroupns in some cgroup
2160 # cat /proc/self/cgroup
2163 # echo 0 > sub_cgrp_1/cgroup.procs
2164 # cat /proc/self/cgroup
2167 Each process gets its namespace-specific view of "/proc/$PID/cgroup"
2169 Processes running inside the cgroup namespace will be able to see
2170 cgroup paths (in /proc/self/cgroup) only inside their root cgroup.
2171 From within an unshared cgroupns::
2175 # echo 7353 > sub_cgrp_1/cgroup.procs
2176 # cat /proc/7353/cgroup
2179 From the initial cgroup namespace, the real cgroup path will be
2182 $ cat /proc/7353/cgroup
2183 0::/batchjobs/container_id1/sub_cgrp_1
2185 From a sibling cgroup namespace (that is, a namespace rooted at a
2186 different cgroup), the cgroup path relative to its own cgroup
2187 namespace root will be shown. For instance, if PID 7353's cgroup
2188 namespace root is at '/batchjobs/container_id2', then it will see::
2190 # cat /proc/7353/cgroup
2191 0::/../container_id2/sub_cgrp_1
2193 Note that the relative path always starts with '/' to indicate that
2194 its relative to the cgroup namespace root of the caller.
2197 Migration and setns(2)
2198 ----------------------
2200 Processes inside a cgroup namespace can move into and out of the
2201 namespace root if they have proper access to external cgroups. For
2202 example, from inside a namespace with cgroupns root at
2203 /batchjobs/container_id1, and assuming that the global hierarchy is
2204 still accessible inside cgroupns::
2206 # cat /proc/7353/cgroup
2208 # echo 7353 > batchjobs/container_id2/cgroup.procs
2209 # cat /proc/7353/cgroup
2210 0::/../container_id2
2212 Note that this kind of setup is not encouraged. A task inside cgroup
2213 namespace should only be exposed to its own cgroupns hierarchy.
2215 setns(2) to another cgroup namespace is allowed when:
2217 (a) the process has CAP_SYS_ADMIN against its current user namespace
2218 (b) the process has CAP_SYS_ADMIN against the target cgroup
2221 No implicit cgroup changes happen with attaching to another cgroup
2222 namespace. It is expected that the someone moves the attaching
2223 process under the target cgroup namespace root.
2226 Interaction with Other Namespaces
2227 ---------------------------------
2229 Namespace specific cgroup hierarchy can be mounted by a process
2230 running inside a non-init cgroup namespace::
2232 # mount -t cgroup2 none $MOUNT_POINT
2234 This will mount the unified cgroup hierarchy with cgroupns root as the
2235 filesystem root. The process needs CAP_SYS_ADMIN against its user and
2238 The virtualization of /proc/self/cgroup file combined with restricting
2239 the view of cgroup hierarchy by namespace-private cgroupfs mount
2240 provides a properly isolated cgroup view inside the container.
2243 Information on Kernel Programming
2244 =================================
2246 This section contains kernel programming information in the areas
2247 where interacting with cgroup is necessary. cgroup core and
2248 controllers are not covered.
2251 Filesystem Support for Writeback
2252 --------------------------------
2254 A filesystem can support cgroup writeback by updating
2255 address_space_operations->writepage[s]() to annotate bio's using the
2256 following two functions.
2258 wbc_init_bio(@wbc, @bio)
2259 Should be called for each bio carrying writeback data and
2260 associates the bio with the inode's owner cgroup and the
2261 corresponding request queue. This must be called after
2262 a queue (device) has been associated with the bio and
2265 wbc_account_cgroup_owner(@wbc, @page, @bytes)
2266 Should be called for each data segment being written out.
2267 While this function doesn't care exactly when it's called
2268 during the writeback session, it's the easiest and most
2269 natural to call it as data segments are added to a bio.
2271 With writeback bio's annotated, cgroup support can be enabled per
2272 super_block by setting SB_I_CGROUPWB in ->s_iflags. This allows for
2273 selective disabling of cgroup writeback support which is helpful when
2274 certain filesystem features, e.g. journaled data mode, are
2277 wbc_init_bio() binds the specified bio to its cgroup. Depending on
2278 the configuration, the bio may be executed at a lower priority and if
2279 the writeback session is holding shared resources, e.g. a journal
2280 entry, may lead to priority inversion. There is no one easy solution
2281 for the problem. Filesystems can try to work around specific problem
2282 cases by skipping wbc_init_bio() and using bio_associate_blkg()
2286 Deprecated v1 Core Features
2287 ===========================
2289 - Multiple hierarchies including named ones are not supported.
2291 - All v1 mount options are not supported.
2293 - The "tasks" file is removed and "cgroup.procs" is not sorted.
2295 - "cgroup.clone_children" is removed.
2297 - /proc/cgroups is meaningless for v2. Use "cgroup.controllers" file
2298 at the root instead.
2301 Issues with v1 and Rationales for v2
2302 ====================================
2304 Multiple Hierarchies
2305 --------------------
2307 cgroup v1 allowed an arbitrary number of hierarchies and each
2308 hierarchy could host any number of controllers. While this seemed to
2309 provide a high level of flexibility, it wasn't useful in practice.
2311 For example, as there is only one instance of each controller, utility
2312 type controllers such as freezer which can be useful in all
2313 hierarchies could only be used in one. The issue is exacerbated by
2314 the fact that controllers couldn't be moved to another hierarchy once
2315 hierarchies were populated. Another issue was that all controllers
2316 bound to a hierarchy were forced to have exactly the same view of the
2317 hierarchy. It wasn't possible to vary the granularity depending on
2318 the specific controller.
2320 In practice, these issues heavily limited which controllers could be
2321 put on the same hierarchy and most configurations resorted to putting
2322 each controller on its own hierarchy. Only closely related ones, such
2323 as the cpu and cpuacct controllers, made sense to be put on the same
2324 hierarchy. This often meant that userland ended up managing multiple
2325 similar hierarchies repeating the same steps on each hierarchy
2326 whenever a hierarchy management operation was necessary.
2328 Furthermore, support for multiple hierarchies came at a steep cost.
2329 It greatly complicated cgroup core implementation but more importantly
2330 the support for multiple hierarchies restricted how cgroup could be
2331 used in general and what controllers was able to do.
2333 There was no limit on how many hierarchies there might be, which meant
2334 that a thread's cgroup membership couldn't be described in finite
2335 length. The key might contain any number of entries and was unlimited
2336 in length, which made it highly awkward to manipulate and led to
2337 addition of controllers which existed only to identify membership,
2338 which in turn exacerbated the original problem of proliferating number
2341 Also, as a controller couldn't have any expectation regarding the
2342 topologies of hierarchies other controllers might be on, each
2343 controller had to assume that all other controllers were attached to
2344 completely orthogonal hierarchies. This made it impossible, or at
2345 least very cumbersome, for controllers to cooperate with each other.
2347 In most use cases, putting controllers on hierarchies which are
2348 completely orthogonal to each other isn't necessary. What usually is
2349 called for is the ability to have differing levels of granularity
2350 depending on the specific controller. In other words, hierarchy may
2351 be collapsed from leaf towards root when viewed from specific
2352 controllers. For example, a given configuration might not care about
2353 how memory is distributed beyond a certain level while still wanting
2354 to control how CPU cycles are distributed.
2360 cgroup v1 allowed threads of a process to belong to different cgroups.
2361 This didn't make sense for some controllers and those controllers
2362 ended up implementing different ways to ignore such situations but
2363 much more importantly it blurred the line between API exposed to
2364 individual applications and system management interface.
2366 Generally, in-process knowledge is available only to the process
2367 itself; thus, unlike service-level organization of processes,
2368 categorizing threads of a process requires active participation from
2369 the application which owns the target process.
2371 cgroup v1 had an ambiguously defined delegation model which got abused
2372 in combination with thread granularity. cgroups were delegated to
2373 individual applications so that they can create and manage their own
2374 sub-hierarchies and control resource distributions along them. This
2375 effectively raised cgroup to the status of a syscall-like API exposed
2378 First of all, cgroup has a fundamentally inadequate interface to be
2379 exposed this way. For a process to access its own knobs, it has to
2380 extract the path on the target hierarchy from /proc/self/cgroup,
2381 construct the path by appending the name of the knob to the path, open
2382 and then read and/or write to it. This is not only extremely clunky
2383 and unusual but also inherently racy. There is no conventional way to
2384 define transaction across the required steps and nothing can guarantee
2385 that the process would actually be operating on its own sub-hierarchy.
2387 cgroup controllers implemented a number of knobs which would never be
2388 accepted as public APIs because they were just adding control knobs to
2389 system-management pseudo filesystem. cgroup ended up with interface
2390 knobs which were not properly abstracted or refined and directly
2391 revealed kernel internal details. These knobs got exposed to
2392 individual applications through the ill-defined delegation mechanism
2393 effectively abusing cgroup as a shortcut to implementing public APIs
2394 without going through the required scrutiny.
2396 This was painful for both userland and kernel. Userland ended up with
2397 misbehaving and poorly abstracted interfaces and kernel exposing and
2398 locked into constructs inadvertently.
2401 Competition Between Inner Nodes and Threads
2402 -------------------------------------------
2404 cgroup v1 allowed threads to be in any cgroups which created an
2405 interesting problem where threads belonging to a parent cgroup and its
2406 children cgroups competed for resources. This was nasty as two
2407 different types of entities competed and there was no obvious way to
2408 settle it. Different controllers did different things.
2410 The cpu controller considered threads and cgroups as equivalents and
2411 mapped nice levels to cgroup weights. This worked for some cases but
2412 fell flat when children wanted to be allocated specific ratios of CPU
2413 cycles and the number of internal threads fluctuated - the ratios
2414 constantly changed as the number of competing entities fluctuated.
2415 There also were other issues. The mapping from nice level to weight
2416 wasn't obvious or universal, and there were various other knobs which
2417 simply weren't available for threads.
2419 The io controller implicitly created a hidden leaf node for each
2420 cgroup to host the threads. The hidden leaf had its own copies of all
2421 the knobs with ``leaf_`` prefixed. While this allowed equivalent
2422 control over internal threads, it was with serious drawbacks. It
2423 always added an extra layer of nesting which wouldn't be necessary
2424 otherwise, made the interface messy and significantly complicated the
2427 The memory controller didn't have a way to control what happened
2428 between internal tasks and child cgroups and the behavior was not
2429 clearly defined. There were attempts to add ad-hoc behaviors and
2430 knobs to tailor the behavior to specific workloads which would have
2431 led to problems extremely difficult to resolve in the long term.
2433 Multiple controllers struggled with internal tasks and came up with
2434 different ways to deal with it; unfortunately, all the approaches were
2435 severely flawed and, furthermore, the widely different behaviors
2436 made cgroup as a whole highly inconsistent.
2438 This clearly is a problem which needs to be addressed from cgroup core
2442 Other Interface Issues
2443 ----------------------
2445 cgroup v1 grew without oversight and developed a large number of
2446 idiosyncrasies and inconsistencies. One issue on the cgroup core side
2447 was how an empty cgroup was notified - a userland helper binary was
2448 forked and executed for each event. The event delivery wasn't
2449 recursive or delegatable. The limitations of the mechanism also led
2450 to in-kernel event delivery filtering mechanism further complicating
2453 Controller interfaces were problematic too. An extreme example is
2454 controllers completely ignoring hierarchical organization and treating
2455 all cgroups as if they were all located directly under the root
2456 cgroup. Some controllers exposed a large amount of inconsistent
2457 implementation details to userland.
2459 There also was no consistency across controllers. When a new cgroup
2460 was created, some controllers defaulted to not imposing extra
2461 restrictions while others disallowed any resource usage until
2462 explicitly configured. Configuration knobs for the same type of
2463 control used widely differing naming schemes and formats. Statistics
2464 and information knobs were named arbitrarily and used different
2465 formats and units even in the same controller.
2467 cgroup v2 establishes common conventions where appropriate and updates
2468 controllers so that they expose minimal and consistent interfaces.
2471 Controller Issues and Remedies
2472 ------------------------------
2477 The original lower boundary, the soft limit, is defined as a limit
2478 that is per default unset. As a result, the set of cgroups that
2479 global reclaim prefers is opt-in, rather than opt-out. The costs for
2480 optimizing these mostly negative lookups are so high that the
2481 implementation, despite its enormous size, does not even provide the
2482 basic desirable behavior. First off, the soft limit has no
2483 hierarchical meaning. All configured groups are organized in a global
2484 rbtree and treated like equal peers, regardless where they are located
2485 in the hierarchy. This makes subtree delegation impossible. Second,
2486 the soft limit reclaim pass is so aggressive that it not just
2487 introduces high allocation latencies into the system, but also impacts
2488 system performance due to overreclaim, to the point where the feature
2489 becomes self-defeating.
2491 The memory.low boundary on the other hand is a top-down allocated
2492 reserve. A cgroup enjoys reclaim protection when it's within its
2493 effective low, which makes delegation of subtrees possible. It also
2494 enjoys having reclaim pressure proportional to its overage when
2495 above its effective low.
2497 The original high boundary, the hard limit, is defined as a strict
2498 limit that can not budge, even if the OOM killer has to be called.
2499 But this generally goes against the goal of making the most out of the
2500 available memory. The memory consumption of workloads varies during
2501 runtime, and that requires users to overcommit. But doing that with a
2502 strict upper limit requires either a fairly accurate prediction of the
2503 working set size or adding slack to the limit. Since working set size
2504 estimation is hard and error prone, and getting it wrong results in
2505 OOM kills, most users tend to err on the side of a looser limit and
2506 end up wasting precious resources.
2508 The memory.high boundary on the other hand can be set much more
2509 conservatively. When hit, it throttles allocations by forcing them
2510 into direct reclaim to work off the excess, but it never invokes the
2511 OOM killer. As a result, a high boundary that is chosen too
2512 aggressively will not terminate the processes, but instead it will
2513 lead to gradual performance degradation. The user can monitor this
2514 and make corrections until the minimal memory footprint that still
2515 gives acceptable performance is found.
2517 In extreme cases, with many concurrent allocations and a complete
2518 breakdown of reclaim progress within the group, the high boundary can
2519 be exceeded. But even then it's mostly better to satisfy the
2520 allocation from the slack available in other groups or the rest of the
2521 system than killing the group. Otherwise, memory.max is there to
2522 limit this type of spillover and ultimately contain buggy or even
2523 malicious applications.
2525 Setting the original memory.limit_in_bytes below the current usage was
2526 subject to a race condition, where concurrent charges could cause the
2527 limit setting to fail. memory.max on the other hand will first set the
2528 limit to prevent new charges, and then reclaim and OOM kill until the
2529 new limit is met - or the task writing to memory.max is killed.
2531 The combined memory+swap accounting and limiting is replaced by real
2532 control over swap space.
2534 The main argument for a combined memory+swap facility in the original
2535 cgroup design was that global or parental pressure would always be
2536 able to swap all anonymous memory of a child group, regardless of the
2537 child's own (possibly untrusted) configuration. However, untrusted
2538 groups can sabotage swapping by other means - such as referencing its
2539 anonymous memory in a tight loop - and an admin can not assume full
2540 swappability when overcommitting untrusted jobs.
2542 For trusted jobs, on the other hand, a combined counter is not an
2543 intuitive userspace interface, and it flies in the face of the idea
2544 that cgroup controllers should account and limit specific physical
2545 resources. Swap space is a resource like all others in the system,
2546 and that's why unified hierarchy allows distributing it separately.