brintos / linux-shallow public Read only

1.. _cgroup-v2:2 3================4Control Group v25================6 7:Date: October, 20158:Author: Tejun Heo <tj@kernel.org>9 10This is the authoritative documentation on the design, interface and11conventions of cgroup v2.  It describes all userland-visible aspects12of cgroup including core and specific controller behaviors.  All13future changes must be reflected in this document.  Documentation for14v1 is available under :ref:`Documentation/admin-guide/cgroup-v1/index.rst <cgroup-v1>`.15 16.. CONTENTS17 18   1. Introduction19     1-1. Terminology20     1-2. What is cgroup?21   2. Basic Operations22     2-1. Mounting23     2-2. Organizing Processes and Threads24       2-2-1. Processes25       2-2-2. Threads26     2-3. [Un]populated Notification27     2-4. Controlling Controllers28       2-4-1. Enabling and Disabling29       2-4-2. Top-down Constraint30       2-4-3. No Internal Process Constraint31     2-5. Delegation32       2-5-1. Model of Delegation33       2-5-2. Delegation Containment34     2-6. Guidelines35       2-6-1. Organize Once and Control36       2-6-2. Avoid Name Collisions37   3. Resource Distribution Models38     3-1. Weights39     3-2. Limits40     3-3. Protections41     3-4. Allocations42   4. Interface Files43     4-1. Format44     4-2. Conventions45     4-3. Core Interface Files46   5. Controllers47     5-1. CPU48       5-1-1. CPU Interface Files49     5-2. Memory50       5-2-1. Memory Interface Files51       5-2-2. Usage Guidelines52       5-2-3. Memory Ownership53     5-3. IO54       5-3-1. IO Interface Files55       5-3-2. Writeback56       5-3-3. IO Latency57         5-3-3-1. How IO Latency Throttling Works58         5-3-3-2. IO Latency Interface Files59       5-3-4. IO Priority60     5-4. PID61       5-4-1. PID Interface Files62     5-5. Cpuset63       5.5-1. Cpuset Interface Files64     5-6. Device65     5-7. RDMA66       5-7-1. RDMA Interface Files67     5-8. HugeTLB68       5.8-1. HugeTLB Interface Files69     5-9. Misc70       5.9-1 Miscellaneous cgroup Interface Files71       5.9-2 Migration and Ownership72     5-10. Others73       5-10-1. perf_event74     5-N. Non-normative information75       5-N-1. CPU controller root cgroup process behaviour76       5-N-2. IO controller root cgroup process behaviour77   6. Namespace78     6-1. Basics79     6-2. The Root and Views80     6-3. Migration and setns(2)81     6-4. Interaction with Other Namespaces82   P. Information on Kernel Programming83     P-1. Filesystem Support for Writeback84   D. Deprecated v1 Core Features85   R. Issues with v1 and Rationales for v286     R-1. Multiple Hierarchies87     R-2. Thread Granularity88     R-3. Competition Between Inner Nodes and Threads89     R-4. Other Interface Issues90     R-5. Controller Issues and Remedies91       R-5-1. Memory92 93 94Introduction95============96 97Terminology98-----------99 100"cgroup" stands for "control group" and is never capitalized.  The101singular form is used to designate the whole feature and also as a102qualifier as in "cgroup controllers".  When explicitly referring to103multiple individual control groups, the plural form "cgroups" is used.104 105 106What is cgroup?107---------------108 109cgroup is a mechanism to organize processes hierarchically and110distribute system resources along the hierarchy in a controlled and111configurable manner.112 113cgroup is largely composed of two parts - the core and controllers.114cgroup core is primarily responsible for hierarchically organizing115processes.  A cgroup controller is usually responsible for116distributing a specific type of system resource along the hierarchy117although there are utility controllers which serve purposes other than118resource distribution.119 120cgroups form a tree structure and every process in the system belongs121to one and only one cgroup.  All threads of a process belong to the122same cgroup.  On creation, all processes are put in the cgroup that123the parent process belongs to at the time.  A process can be migrated124to another cgroup.  Migration of a process doesn't affect already125existing descendant processes.126 127Following certain structural constraints, controllers may be enabled or128disabled selectively on a cgroup.  All controller behaviors are129hierarchical - if a controller is enabled on a cgroup, it affects all130processes which belong to the cgroups consisting the inclusive131sub-hierarchy of the cgroup.  When a controller is enabled on a nested132cgroup, it always restricts the resource distribution further.  The133restrictions set closer to the root in the hierarchy can not be134overridden from further away.135 136 137Basic Operations138================139 140Mounting141--------142 143Unlike v1, cgroup v2 has only single hierarchy.  The cgroup v2144hierarchy can be mounted with the following mount command::145 146  # mount -t cgroup2 none $MOUNT_POINT147 148cgroup2 filesystem has the magic number 0x63677270 ("cgrp").  All149controllers which support v2 and are not bound to a v1 hierarchy are150automatically bound to the v2 hierarchy and show up at the root.151Controllers which are not in active use in the v2 hierarchy can be152bound to other hierarchies.  This allows mixing v2 hierarchy with the153legacy v1 multiple hierarchies in a fully backward compatible way.154 155A controller can be moved across hierarchies only after the controller156is no longer referenced in its current hierarchy.  Because per-cgroup157controller states are destroyed asynchronously and controllers may158have lingering references, a controller may not show up immediately on159the v2 hierarchy after the final umount of the previous hierarchy.160Similarly, a controller should be fully disabled to be moved out of161the unified hierarchy and it may take some time for the disabled162controller to become available for other hierarchies; furthermore, due163to inter-controller dependencies, other controllers may need to be164disabled too.165 166While useful for development and manual configurations, moving167controllers dynamically between the v2 and other hierarchies is168strongly discouraged for production use.  It is recommended to decide169the hierarchies and controller associations before starting using the170controllers after system boot.171 172During transition to v2, system management software might still173automount the v1 cgroup filesystem and so hijack all controllers174during boot, before manual intervention is possible. To make testing175and experimenting easier, the kernel parameter cgroup_no_v1= allows176disabling controllers in v1 and make them always available in v2.177 178cgroup v2 currently supports the following mount options.179 180  nsdelegate181	Consider cgroup namespaces as delegation boundaries.  This182	option is system wide and can only be set on mount or modified183	through remount from the init namespace.  The mount option is184	ignored on non-init namespace mounts.  Please refer to the185	Delegation section for details.186 187  favordynmods188        Reduce the latencies of dynamic cgroup modifications such as189        task migrations and controller on/offs at the cost of making190        hot path operations such as forks and exits more expensive.191        The static usage pattern of creating a cgroup, enabling192        controllers, and then seeding it with CLONE_INTO_CGROUP is193        not affected by this option.194 195  memory_localevents196        Only populate memory.events with data for the current cgroup,197        and not any subtrees. This is legacy behaviour, the default198        behaviour without this option is to include subtree counts.199        This option is system wide and can only be set on mount or200        modified through remount from the init namespace. The mount201        option is ignored on non-init namespace mounts.202 203  memory_recursiveprot204        Recursively apply memory.min and memory.low protection to205        entire subtrees, without requiring explicit downward206        propagation into leaf cgroups.  This allows protecting entire207        subtrees from one another, while retaining free competition208        within those subtrees.  This should have been the default209        behavior but is a mount-option to avoid regressing setups210        relying on the original semantics (e.g. specifying bogusly211        high 'bypass' protection values at higher tree levels).212 213  memory_hugetlb_accounting214        Count HugeTLB memory usage towards the cgroup's overall215        memory usage for the memory controller (for the purpose of216        statistics reporting and memory protetion). This is a new217        behavior that could regress existing setups, so it must be218        explicitly opted in with this mount option.219 220        A few caveats to keep in mind:221 222        * There is no HugeTLB pool management involved in the memory223          controller. The pre-allocated pool does not belong to anyone.224          Specifically, when a new HugeTLB folio is allocated to225          the pool, it is not accounted for from the perspective of the226          memory controller. It is only charged to a cgroup when it is227          actually used (for e.g at page fault time). Host memory228          overcommit management has to consider this when configuring229          hard limits. In general, HugeTLB pool management should be230          done via other mechanisms (such as the HugeTLB controller).231        * Failure to charge a HugeTLB folio to the memory controller232          results in SIGBUS. This could happen even if the HugeTLB pool233          still has pages available (but the cgroup limit is hit and234          reclaim attempt fails).235        * Charging HugeTLB memory towards the memory controller affects236          memory protection and reclaim dynamics. Any userspace tuning237          (of low, min limits for e.g) needs to take this into account.238        * HugeTLB pages utilized while this option is not selected239          will not be tracked by the memory controller (even if cgroup240          v2 is remounted later on).241 242  pids_localevents243        The option restores v1-like behavior of pids.events:max, that is only244        local (inside cgroup proper) fork failures are counted. Without this245        option pids.events.max represents any pids.max enforcemnt across246        cgroup's subtree.247 248 249 250Organizing Processes and Threads251--------------------------------252 253Processes254~~~~~~~~~255 256Initially, only the root cgroup exists to which all processes belong.257A child cgroup can be created by creating a sub-directory::258 259  # mkdir $CGROUP_NAME260 261A given cgroup may have multiple child cgroups forming a tree262structure.  Each cgroup has a read-writable interface file263"cgroup.procs".  When read, it lists the PIDs of all processes which264belong to the cgroup one-per-line.  The PIDs are not ordered and the265same PID may show up more than once if the process got moved to266another cgroup and then back or the PID got recycled while reading.267 268A process can be migrated into a cgroup by writing its PID to the269target cgroup's "cgroup.procs" file.  Only one process can be migrated270on a single write(2) call.  If a process is composed of multiple271threads, writing the PID of any thread migrates all threads of the272process.273 274When a process forks a child process, the new process is born into the275cgroup that the forking process belongs to at the time of the276operation.  After exit, a process stays associated with the cgroup277that it belonged to at the time of exit until it's reaped; however, a278zombie process does not appear in "cgroup.procs" and thus can't be279moved to another cgroup.280 281A cgroup which doesn't have any children or live processes can be282destroyed by removing the directory.  Note that a cgroup which doesn't283have any children and is associated only with zombie processes is284considered empty and can be removed::285 286  # rmdir $CGROUP_NAME287 288"/proc/$PID/cgroup" lists a process's cgroup membership.  If legacy289cgroup is in use in the system, this file may contain multiple lines,290one for each hierarchy.  The entry for cgroup v2 is always in the291format "0::$PATH"::292 293  # cat /proc/842/cgroup294  ...295  0::/test-cgroup/test-cgroup-nested296 297If the process becomes a zombie and the cgroup it was associated with298is removed subsequently, " (deleted)" is appended to the path::299 300  # cat /proc/842/cgroup301  ...302  0::/test-cgroup/test-cgroup-nested (deleted)303 304 305Threads306~~~~~~~307 308cgroup v2 supports thread granularity for a subset of controllers to309support use cases requiring hierarchical resource distribution across310the threads of a group of processes.  By default, all threads of a311process belong to the same cgroup, which also serves as the resource312domain to host resource consumptions which are not specific to a313process or thread.  The thread mode allows threads to be spread across314a subtree while still maintaining the common resource domain for them.315 316Controllers which support thread mode are called threaded controllers.317The ones which don't are called domain controllers.318 319Marking a cgroup threaded makes it join the resource domain of its320parent as a threaded cgroup.  The parent may be another threaded321cgroup whose resource domain is further up in the hierarchy.  The root322of a threaded subtree, that is, the nearest ancestor which is not323threaded, is called threaded domain or thread root interchangeably and324serves as the resource domain for the entire subtree.325 326Inside a threaded subtree, threads of a process can be put in327different cgroups and are not subject to the no internal process328constraint - threaded controllers can be enabled on non-leaf cgroups329whether they have threads in them or not.330 331As the threaded domain cgroup hosts all the domain resource332consumptions of the subtree, it is considered to have internal333resource consumptions whether there are processes in it or not and334can't have populated child cgroups which aren't threaded.  Because the335root cgroup is not subject to no internal process constraint, it can336serve both as a threaded domain and a parent to domain cgroups.337 338The current operation mode or type of the cgroup is shown in the339"cgroup.type" file which indicates whether the cgroup is a normal340domain, a domain which is serving as the domain of a threaded subtree,341or a threaded cgroup.342 343On creation, a cgroup is always a domain cgroup and can be made344threaded by writing "threaded" to the "cgroup.type" file.  The345operation is single direction::346 347  # echo threaded > cgroup.type348 349Once threaded, the cgroup can't be made a domain again.  To enable the350thread mode, the following conditions must be met.351 352- As the cgroup will join the parent's resource domain.  The parent353  must either be a valid (threaded) domain or a threaded cgroup.354 355- When the parent is an unthreaded domain, it must not have any domain356  controllers enabled or populated domain children.  The root is357  exempt from this requirement.358 359Topology-wise, a cgroup can be in an invalid state.  Please consider360the following topology::361 362  A (threaded domain) - B (threaded) - C (domain, just created)363 364C is created as a domain but isn't connected to a parent which can365host child domains.  C can't be used until it is turned into a366threaded cgroup.  "cgroup.type" file will report "domain (invalid)" in367these cases.  Operations which fail due to invalid topology use368EOPNOTSUPP as the errno.369 370A domain cgroup is turned into a threaded domain when one of its child371cgroup becomes threaded or threaded controllers are enabled in the372"cgroup.subtree_control" file while there are processes in the cgroup.373A threaded domain reverts to a normal domain when the conditions374clear.375 376When read, "cgroup.threads" contains the list of the thread IDs of all377threads in the cgroup.  Except that the operations are per-thread378instead of per-process, "cgroup.threads" has the same format and379behaves the same way as "cgroup.procs".  While "cgroup.threads" can be380written to in any cgroup, as it can only move threads inside the same381threaded domain, its operations are confined inside each threaded382subtree.383 384The threaded domain cgroup serves as the resource domain for the whole385subtree, and, while the threads can be scattered across the subtree,386all the processes are considered to be in the threaded domain cgroup.387"cgroup.procs" in a threaded domain cgroup contains the PIDs of all388processes in the subtree and is not readable in the subtree proper.389However, "cgroup.procs" can be written to from anywhere in the subtree390to migrate all threads of the matching process to the cgroup.391 392Only threaded controllers can be enabled in a threaded subtree.  When393a threaded controller is enabled inside a threaded subtree, it only394accounts for and controls resource consumptions associated with the395threads in the cgroup and its descendants.  All consumptions which396aren't tied to a specific thread belong to the threaded domain cgroup.397 398Because a threaded subtree is exempt from no internal process399constraint, a threaded controller must be able to handle competition400between threads in a non-leaf cgroup and its child cgroups.  Each401threaded controller defines how such competitions are handled.402 403Currently, the following controllers are threaded and can be enabled404in a threaded cgroup::405 406- cpu407- cpuset408- perf_event409- pids410 411[Un]populated Notification412--------------------------413 414Each non-root cgroup has a "cgroup.events" file which contains415"populated" field indicating whether the cgroup's sub-hierarchy has416live processes in it.  Its value is 0 if there is no live process in417the cgroup and its descendants; otherwise, 1.  poll and [id]notify418events are triggered when the value changes.  This can be used, for419example, to start a clean-up operation after all processes of a given420sub-hierarchy have exited.  The populated state updates and421notifications are recursive.  Consider the following sub-hierarchy422where the numbers in the parentheses represent the numbers of processes423in each cgroup::424 425  A(4) - B(0) - C(1)426              \ D(0)427 428A, B and C's "populated" fields would be 1 while D's 0.  After the one429process in C exits, B and C's "populated" fields would flip to "0" and430file modified events will be generated on the "cgroup.events" files of431both cgroups.432 433 434Controlling Controllers435-----------------------436 437Enabling and Disabling438~~~~~~~~~~~~~~~~~~~~~~439 440Each cgroup has a "cgroup.controllers" file which lists all441controllers available for the cgroup to enable::442 443  # cat cgroup.controllers444  cpu io memory445 446No controller is enabled by default.  Controllers can be enabled and447disabled by writing to the "cgroup.subtree_control" file::448 449  # echo "+cpu +memory -io" > cgroup.subtree_control450 451Only controllers which are listed in "cgroup.controllers" can be452enabled.  When multiple operations are specified as above, either they453all succeed or fail.  If multiple operations on the same controller454are specified, the last one is effective.455 456Enabling a controller in a cgroup indicates that the distribution of457the target resource across its immediate children will be controlled.458Consider the following sub-hierarchy.  The enabled controllers are459listed in parentheses::460 461  A(cpu,memory) - B(memory) - C()462                            \ D()463 464As A has "cpu" and "memory" enabled, A will control the distribution465of CPU cycles and memory to its children, in this case, B.  As B has466"memory" enabled but not "CPU", C and D will compete freely on CPU467cycles but their division of memory available to B will be controlled.468 469As a controller regulates the distribution of the target resource to470the cgroup's children, enabling it creates the controller's interface471files in the child cgroups.  In the above example, enabling "cpu" on B472would create the "cpu." prefixed controller interface files in C and473D.  Likewise, disabling "memory" from B would remove the "memory."474prefixed controller interface files from C and D.  This means that the475controller interface files - anything which doesn't start with476"cgroup." are owned by the parent rather than the cgroup itself.477 478 479Top-down Constraint480~~~~~~~~~~~~~~~~~~~481 482Resources are distributed top-down and a cgroup can further distribute483a resource only if the resource has been distributed to it from the484parent.  This means that all non-root "cgroup.subtree_control" files485can only contain controllers which are enabled in the parent's486"cgroup.subtree_control" file.  A controller can be enabled only if487the parent has the controller enabled and a controller can't be488disabled if one or more children have it enabled.489 490 491No Internal Process Constraint492~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~493 494Non-root cgroups can distribute domain resources to their children495only when they don't have any processes of their own.  In other words,496only domain cgroups which don't contain any processes can have domain497controllers enabled in their "cgroup.subtree_control" files.498 499This guarantees that, when a domain controller is looking at the part500of the hierarchy which has it enabled, processes are always only on501the leaves.  This rules out situations where child cgroups compete502against internal processes of the parent.503 504The root cgroup is exempt from this restriction.  Root contains505processes and anonymous resource consumption which can't be associated506with any other cgroups and requires special treatment from most507controllers.  How resource consumption in the root cgroup is governed508is up to each controller (for more information on this topic please509refer to the Non-normative information section in the Controllers510chapter).511 512Note that the restriction doesn't get in the way if there is no513enabled controller in the cgroup's "cgroup.subtree_control".  This is514important as otherwise it wouldn't be possible to create children of a515populated cgroup.  To control resource distribution of a cgroup, the516cgroup must create children and transfer all its processes to the517children before enabling controllers in its "cgroup.subtree_control"518file.519 520 521Delegation522----------523 524Model of Delegation525~~~~~~~~~~~~~~~~~~~526 527A cgroup can be delegated in two ways.  First, to a less privileged528user by granting write access of the directory and its "cgroup.procs",529"cgroup.threads" and "cgroup.subtree_control" files to the user.530Second, if the "nsdelegate" mount option is set, automatically to a531cgroup namespace on namespace creation.532 533Because the resource control interface files in a given directory534control the distribution of the parent's resources, the delegatee535shouldn't be allowed to write to them.  For the first method, this is536achieved by not granting access to these files.  For the second, files537outside the namespace should be hidden from the delegatee by the means538of at least mount namespacing, and the kernel rejects writes to all539files on a namespace root from inside the cgroup namespace, except for540those files listed in "/sys/kernel/cgroup/delegate" (including541"cgroup.procs", "cgroup.threads", "cgroup.subtree_control", etc.).542 543The end results are equivalent for both delegation types.  Once544delegated, the user can build sub-hierarchy under the directory,545organize processes inside it as it sees fit and further distribute the546resources it received from the parent.  The limits and other settings547of all resource controllers are hierarchical and regardless of what548happens in the delegated sub-hierarchy, nothing can escape the549resource restrictions imposed by the parent.550 551Currently, cgroup doesn't impose any restrictions on the number of552cgroups in or nesting depth of a delegated sub-hierarchy; however,553this may be limited explicitly in the future.554 555 556Delegation Containment557~~~~~~~~~~~~~~~~~~~~~~558 559A delegated sub-hierarchy is contained in the sense that processes560can't be moved into or out of the sub-hierarchy by the delegatee.561 562For delegations to a less privileged user, this is achieved by563requiring the following conditions for a process with a non-root euid564to migrate a target process into a cgroup by writing its PID to the565"cgroup.procs" file.566 567- The writer must have write access to the "cgroup.procs" file.568 569- The writer must have write access to the "cgroup.procs" file of the570  common ancestor of the source and destination cgroups.571 572The above two constraints ensure that while a delegatee may migrate573processes around freely in the delegated sub-hierarchy it can't pull574in from or push out to outside the sub-hierarchy.575 576For an example, let's assume cgroups C0 and C1 have been delegated to577user U0 who created C00, C01 under C0 and C10 under C1 as follows and578all processes under C0 and C1 belong to U0::579 580  ~~~~~~~~~~~~~ - C0 - C00581  ~ cgroup    ~      \ C01582  ~ hierarchy ~583  ~~~~~~~~~~~~~ - C1 - C10584 585Let's also say U0 wants to write the PID of a process which is586currently in C10 into "C00/cgroup.procs".  U0 has write access to the587file; however, the common ancestor of the source cgroup C10 and the588destination cgroup C00 is above the points of delegation and U0 would589not have write access to its "cgroup.procs" files and thus the write590will be denied with -EACCES.591 592For delegations to namespaces, containment is achieved by requiring593that both the source and destination cgroups are reachable from the594namespace of the process which is attempting the migration.  If either595is not reachable, the migration is rejected with -ENOENT.596 597 598Guidelines599----------600 601Organize Once and Control602~~~~~~~~~~~~~~~~~~~~~~~~~603 604Migrating a process across cgroups is a relatively expensive operation605and stateful resources such as memory are not moved together with the606process.  This is an explicit design decision as there often exist607inherent trade-offs between migration and various hot paths in terms608of synchronization cost.609 610As such, migrating processes across cgroups frequently as a means to611apply different resource restrictions is discouraged.  A workload612should be assigned to a cgroup according to the system's logical and613resource structure once on start-up.  Dynamic adjustments to resource614distribution can be made by changing controller configuration through615the interface files.616 617 618Avoid Name Collisions619~~~~~~~~~~~~~~~~~~~~~620 621Interface files for a cgroup and its children cgroups occupy the same622directory and it is possible to create children cgroups which collide623with interface files.624 625All cgroup core interface files are prefixed with "cgroup." and each626controller's interface files are prefixed with the controller name and627a dot.  A controller's name is composed of lower case alphabets and628'_'s but never begins with an '_' so it can be used as the prefix629character for collision avoidance.  Also, interface file names won't630start or end with terms which are often used in categorizing workloads631such as job, service, slice, unit or workload.632 633cgroup doesn't do anything to prevent name collisions and it's the634user's responsibility to avoid them.635 636 637Resource Distribution Models638============================639 640cgroup controllers implement several resource distribution schemes641depending on the resource type and expected use cases.  This section642describes major schemes in use along with their expected behaviors.643 644 645Weights646-------647 648A parent's resource is distributed by adding up the weights of all649active children and giving each the fraction matching the ratio of its650weight against the sum.  As only children which can make use of the651resource at the moment participate in the distribution, this is652work-conserving.  Due to the dynamic nature, this model is usually653used for stateless resources.654 655All weights are in the range [1, 10000] with the default at 100.  This656allows symmetric multiplicative biases in both directions at fine657enough granularity while staying in the intuitive range.658 659As long as the weight is in range, all configuration combinations are660valid and there is no reason to reject configuration changes or661process migrations.662 663"cpu.weight" proportionally distributes CPU cycles to active children664and is an example of this type.665 666 667.. _cgroupv2-limits-distributor:668 669Limits670------671 672A child can only consume up to the configured amount of the resource.673Limits can be over-committed - the sum of the limits of children can674exceed the amount of resource available to the parent.675 676Limits are in the range [0, max] and defaults to "max", which is noop.677 678As limits can be over-committed, all configuration combinations are679valid and there is no reason to reject configuration changes or680process migrations.681 682"io.max" limits the maximum BPS and/or IOPS that a cgroup can consume683on an IO device and is an example of this type.684 685.. _cgroupv2-protections-distributor:686 687Protections688-----------689 690A cgroup is protected up to the configured amount of the resource691as long as the usages of all its ancestors are under their692protected levels.  Protections can be hard guarantees or best effort693soft boundaries.  Protections can also be over-committed in which case694only up to the amount available to the parent is protected among695children.696 697Protections are in the range [0, max] and defaults to 0, which is698noop.699 700As protections can be over-committed, all configuration combinations701are valid and there is no reason to reject configuration changes or702process migrations.703 704"memory.low" implements best-effort memory protection and is an705example of this type.706 707 708Allocations709-----------710 711A cgroup is exclusively allocated a certain amount of a finite712resource.  Allocations can't be over-committed - the sum of the713allocations of children can not exceed the amount of resource714available to the parent.715 716Allocations are in the range [0, max] and defaults to 0, which is no717resource.718 719As allocations can't be over-committed, some configuration720combinations are invalid and should be rejected.  Also, if the721resource is mandatory for execution of processes, process migrations722may be rejected.723 724"cpu.rt.max" hard-allocates realtime slices and is an example of this725type.726 727 728Interface Files729===============730 731Format732------733 734All interface files should be in one of the following formats whenever735possible::736 737  New-line separated values738  (when only one value can be written at once)739 740	VAL0\n741	VAL1\n742	...743 744  Space separated values745  (when read-only or multiple values can be written at once)746 747	VAL0 VAL1 ...\n748 749  Flat keyed750 751	KEY0 VAL0\n752	KEY1 VAL1\n753	...754 755  Nested keyed756 757	KEY0 SUB_KEY0=VAL00 SUB_KEY1=VAL01...758	KEY1 SUB_KEY0=VAL10 SUB_KEY1=VAL11...759	...760 761For a writable file, the format for writing should generally match762reading; however, controllers may allow omitting later fields or763implement restricted shortcuts for most common use cases.764 765For both flat and nested keyed files, only the values for a single key766can be written at a time.  For nested keyed files, the sub key pairs767may be specified in any order and not all pairs have to be specified.768 769 770Conventions771-----------772 773- Settings for a single feature should be contained in a single file.774 775- The root cgroup should be exempt from resource control and thus776  shouldn't have resource control interface files.777 778- The default time unit is microseconds.  If a different unit is ever779  used, an explicit unit suffix must be present.780 781- A parts-per quantity should use a percentage decimal with at least782  two digit fractional part - e.g. 13.40.783 784- If a controller implements weight based resource distribution, its785  interface file should be named "weight" and have the range [1,786  10000] with 100 as the default.  The values are chosen to allow787  enough and symmetric bias in both directions while keeping it788  intuitive (the default is 100%).789 790- If a controller implements an absolute resource guarantee and/or791  limit, the interface files should be named "min" and "max"792  respectively.  If a controller implements best effort resource793  guarantee and/or limit, the interface files should be named "low"794  and "high" respectively.795 796  In the above four control files, the special token "max" should be797  used to represent upward infinity for both reading and writing.798 799- If a setting has a configurable default value and keyed specific800  overrides, the default entry should be keyed with "default" and801  appear as the first entry in the file.802 803  The default value can be updated by writing either "default $VAL" or804  "$VAL".805 806  When writing to update a specific override, "default" can be used as807  the value to indicate removal of the override.  Override entries808  with "default" as the value must not appear when read.809 810  For example, a setting which is keyed by major:minor device numbers811  with integer values may look like the following::812 813    # cat cgroup-example-interface-file814    default 150815    8:0 300816 817  The default value can be updated by::818 819    # echo 125 > cgroup-example-interface-file820 821  or::822 823    # echo "default 125" > cgroup-example-interface-file824 825  An override can be set by::826 827    # echo "8:16 170" > cgroup-example-interface-file828 829  and cleared by::830 831    # echo "8:0 default" > cgroup-example-interface-file832    # cat cgroup-example-interface-file833    default 125834    8:16 170835 836- For events which are not very high frequency, an interface file837  "events" should be created which lists event key value pairs.838  Whenever a notifiable event happens, file modified event should be839  generated on the file.840 841 842Core Interface Files843--------------------844 845All cgroup core files are prefixed with "cgroup."846 847  cgroup.type848	A read-write single value file which exists on non-root849	cgroups.850 851	When read, it indicates the current type of the cgroup, which852	can be one of the following values.853 854	- "domain" : A normal valid domain cgroup.855 856	- "domain threaded" : A threaded domain cgroup which is857          serving as the root of a threaded subtree.858 859	- "domain invalid" : A cgroup which is in an invalid state.860	  It can't be populated or have controllers enabled.  It may861	  be allowed to become a threaded cgroup.862 863	- "threaded" : A threaded cgroup which is a member of a864          threaded subtree.865 866	A cgroup can be turned into a threaded cgroup by writing867	"threaded" to this file.868 869  cgroup.procs870	A read-write new-line separated values file which exists on871	all cgroups.872 873	When read, it lists the PIDs of all processes which belong to874	the cgroup one-per-line.  The PIDs are not ordered and the875	same PID may show up more than once if the process got moved876	to another cgroup and then back or the PID got recycled while877	reading.878 879	A PID can be written to migrate the process associated with880	the PID to the cgroup.  The writer should match all of the881	following conditions.882 883	- It must have write access to the "cgroup.procs" file.884 885	- It must have write access to the "cgroup.procs" file of the886	  common ancestor of the source and destination cgroups.887 888	When delegating a sub-hierarchy, write access to this file889	should be granted along with the containing directory.890 891	In a threaded cgroup, reading this file fails with EOPNOTSUPP892	as all the processes belong to the thread root.  Writing is893	supported and moves every thread of the process to the cgroup.894 895  cgroup.threads896	A read-write new-line separated values file which exists on897	all cgroups.898 899	When read, it lists the TIDs of all threads which belong to900	the cgroup one-per-line.  The TIDs are not ordered and the901	same TID may show up more than once if the thread got moved to902	another cgroup and then back or the TID got recycled while903	reading.904 905	A TID can be written to migrate the thread associated with the906	TID to the cgroup.  The writer should match all of the907	following conditions.908 909	- It must have write access to the "cgroup.threads" file.910 911	- The cgroup that the thread is currently in must be in the912          same resource domain as the destination cgroup.913 914	- It must have write access to the "cgroup.procs" file of the915	  common ancestor of the source and destination cgroups.916 917	When delegating a sub-hierarchy, write access to this file918	should be granted along with the containing directory.919 920  cgroup.controllers921	A read-only space separated values file which exists on all922	cgroups.923 924	It shows space separated list of all controllers available to925	the cgroup.  The controllers are not ordered.926 927  cgroup.subtree_control928	A read-write space separated values file which exists on all929	cgroups.  Starts out empty.930 931	When read, it shows space separated list of the controllers932	which are enabled to control resource distribution from the933	cgroup to its children.934 935	Space separated list of controllers prefixed with '+' or '-'936	can be written to enable or disable controllers.  A controller937	name prefixed with '+' enables the controller and '-'938	disables.  If a controller appears more than once on the list,939	the last one is effective.  When multiple enable and disable940	operations are specified, either all succeed or all fail.941 942  cgroup.events943	A read-only flat-keyed file which exists on non-root cgroups.944	The following entries are defined.  Unless specified945	otherwise, a value change in this file generates a file946	modified event.947 948	  populated949		1 if the cgroup or its descendants contains any live950		processes; otherwise, 0.951	  frozen952		1 if the cgroup is frozen; otherwise, 0.953 954  cgroup.max.descendants955	A read-write single value files.  The default is "max".956 957	Maximum allowed number of descent cgroups.958	If the actual number of descendants is equal or larger,959	an attempt to create a new cgroup in the hierarchy will fail.960 961  cgroup.max.depth962	A read-write single value files.  The default is "max".963 964	Maximum allowed descent depth below the current cgroup.965	If the actual descent depth is equal or larger,966	an attempt to create a new child cgroup will fail.967 968  cgroup.stat969	A read-only flat-keyed file with the following entries:970 971	  nr_descendants972		Total number of visible descendant cgroups.973 974	  nr_dying_descendants975		Total number of dying descendant cgroups. A cgroup becomes976		dying after being deleted by a user. The cgroup will remain977		in dying state for some time undefined time (which can depend978		on system load) before being completely destroyed.979 980		A process can't enter a dying cgroup under any circumstances,981		a dying cgroup can't revive.982 983		A dying cgroup can consume system resources not exceeding984		limits, which were active at the moment of cgroup deletion.985 986	  nr_subsys_<cgroup_subsys>987		Total number of live cgroup subsystems (e.g memory988		cgroup) at and beneath the current cgroup.989 990	  nr_dying_subsys_<cgroup_subsys>991		Total number of dying cgroup subsystems (e.g. memory992		cgroup) at and beneath the current cgroup.993 994  cgroup.freeze995	A read-write single value file which exists on non-root cgroups.996	Allowed values are "0" and "1". The default is "0".997 998	Writing "1" to the file causes freezing of the cgroup and all999	descendant cgroups. This means that all belonging processes will1000	be stopped and will not run until the cgroup will be explicitly1001	unfrozen. Freezing of the cgroup may take some time; when this action1002	is completed, the "frozen" value in the cgroup.events control file1003	will be updated to "1" and the corresponding notification will be1004	issued.1005 1006	A cgroup can be frozen either by its own settings, or by settings1007	of any ancestor cgroups. If any of ancestor cgroups is frozen, the1008	cgroup will remain frozen.1009 1010	Processes in the frozen cgroup can be killed by a fatal signal.1011	They also can enter and leave a frozen cgroup: either by an explicit1012	move by a user, or if freezing of the cgroup races with fork().1013	If a process is moved to a frozen cgroup, it stops. If a process is1014	moved out of a frozen cgroup, it becomes running.1015 1016	Frozen status of a cgroup doesn't affect any cgroup tree operations:1017	it's possible to delete a frozen (and empty) cgroup, as well as1018	create new sub-cgroups.1019 1020  cgroup.kill1021	A write-only single value file which exists in non-root cgroups.1022	The only allowed value is "1".1023 1024	Writing "1" to the file causes the cgroup and all descendant cgroups to1025	be killed. This means that all processes located in the affected cgroup1026	tree will be killed via SIGKILL.1027 1028	Killing a cgroup tree will deal with concurrent forks appropriately and1029	is protected against migrations.1030 1031	In a threaded cgroup, writing this file fails with EOPNOTSUPP as1032	killing cgroups is a process directed operation, i.e. it affects1033	the whole thread-group.1034 1035  cgroup.pressure1036	A read-write single value file that allowed values are "0" and "1".1037	The default is "1".1038 1039	Writing "0" to the file will disable the cgroup PSI accounting.1040	Writing "1" to the file will re-enable the cgroup PSI accounting.1041 1042	This control attribute is not hierarchical, so disable or enable PSI1043	accounting in a cgroup does not affect PSI accounting in descendants1044	and doesn't need pass enablement via ancestors from root.1045 1046	The reason this control attribute exists is that PSI accounts stalls for1047	each cgroup separately and aggregates it at each level of the hierarchy.1048	This may cause non-negligible overhead for some workloads when under1049	deep level of the hierarchy, in which case this control attribute can1050	be used to disable PSI accounting in the non-leaf cgroups.1051 1052  irq.pressure1053	A read-write nested-keyed file.1054 1055	Shows pressure stall information for IRQ/SOFTIRQ. See1056	:ref:`Documentation/accounting/psi.rst <psi>` for details.1057 1058Controllers1059===========1060 1061.. _cgroup-v2-cpu:1062 1063CPU1064---1065 1066The "cpu" controllers regulates distribution of CPU cycles.  This1067controller implements weight and absolute bandwidth limit models for1068normal scheduling policy and absolute bandwidth allocation model for1069realtime scheduling policy.1070 1071In all the above models, cycles distribution is defined only on a temporal1072base and it does not account for the frequency at which tasks are executed.1073The (optional) utilization clamping support allows to hint the schedutil1074cpufreq governor about the minimum desired frequency which should always be1075provided by a CPU, as well as the maximum desired frequency, which should not1076be exceeded by a CPU.1077 1078WARNING: cgroup2 doesn't yet support control of realtime processes. For1079a kernel built with the CONFIG_RT_GROUP_SCHED option enabled for group1080scheduling of realtime processes, the cpu controller can only be enabled1081when all RT processes are in the root cgroup.  This limitation does1082not apply if CONFIG_RT_GROUP_SCHED is disabled.  Be aware that system1083management software may already have placed RT processes into nonroot1084cgroups during the system boot process, and these processes may need1085to be moved to the root cgroup before the cpu controller can be enabled1086with a CONFIG_RT_GROUP_SCHED enabled kernel.1087 1088 1089CPU Interface Files1090~~~~~~~~~~~~~~~~~~~1091 1092All time durations are in microseconds.1093 1094  cpu.stat1095	A read-only flat-keyed file.1096	This file exists whether the controller is enabled or not.1097 1098	It always reports the following three stats:1099 1100	- usage_usec1101	- user_usec1102	- system_usec1103 1104	and the following five when the controller is enabled:1105 1106	- nr_periods1107	- nr_throttled1108	- throttled_usec1109	- nr_bursts1110	- burst_usec1111 1112  cpu.weight1113	A read-write single value file which exists on non-root1114	cgroups.  The default is "100".1115 1116	For non idle groups (cpu.idle = 0), the weight is in the1117	range [1, 10000].1118 1119	If the cgroup has been configured to be SCHED_IDLE (cpu.idle = 1),1120	then the weight will show as a 0.1121 1122  cpu.weight.nice1123	A read-write single value file which exists on non-root1124	cgroups.  The default is "0".1125 1126	The nice value is in the range [-20, 19].1127 1128	This interface file is an alternative interface for1129	"cpu.weight" and allows reading and setting weight using the1130	same values used by nice(2).  Because the range is smaller and1131	granularity is coarser for the nice values, the read value is1132	the closest approximation of the current weight.1133 1134  cpu.max1135	A read-write two value file which exists on non-root cgroups.1136	The default is "max 100000".1137 1138	The maximum bandwidth limit.  It's in the following format::1139 1140	  $MAX $PERIOD1141 1142	which indicates that the group may consume up to $MAX in each1143	$PERIOD duration.  "max" for $MAX indicates no limit.  If only1144	one number is written, $MAX is updated.1145 1146  cpu.max.burst1147	A read-write single value file which exists on non-root1148	cgroups.  The default is "0".1149 1150	The burst in the range [0, $MAX].1151 1152  cpu.pressure1153	A read-write nested-keyed file.1154 1155	Shows pressure stall information for CPU. See1156	:ref:`Documentation/accounting/psi.rst <psi>` for details.1157 1158  cpu.uclamp.min1159        A read-write single value file which exists on non-root cgroups.1160        The default is "0", i.e. no utilization boosting.1161 1162        The requested minimum utilization (protection) as a percentage1163        rational number, e.g. 12.34 for 12.34%.1164 1165        This interface allows reading and setting minimum utilization clamp1166        values similar to the sched_setattr(2). This minimum utilization1167        value is used to clamp the task specific minimum utilization clamp.1168 1169        The requested minimum utilization (protection) is always capped by1170        the current value for the maximum utilization (limit), i.e.1171        `cpu.uclamp.max`.1172 1173  cpu.uclamp.max1174        A read-write single value file which exists on non-root cgroups.1175        The default is "max". i.e. no utilization capping1176 1177        The requested maximum utilization (limit) as a percentage rational1178        number, e.g. 98.76 for 98.76%.1179 1180        This interface allows reading and setting maximum utilization clamp1181        values similar to the sched_setattr(2). This maximum utilization1182        value is used to clamp the task specific maximum utilization clamp.1183 1184  cpu.idle1185	A read-write single value file which exists on non-root cgroups.1186	The default is 0.1187 1188	This is the cgroup analog of the per-task SCHED_IDLE sched policy.1189	Setting this value to a 1 will make the scheduling policy of the1190	cgroup SCHED_IDLE. The threads inside the cgroup will retain their1191	own relative priorities, but the cgroup itself will be treated as1192	very low priority relative to its peers.1193 1194 1195 1196Memory1197------1198 1199The "memory" controller regulates distribution of memory.  Memory is1200stateful and implements both limit and protection models.  Due to the1201intertwining between memory usage and reclaim pressure and the1202stateful nature of memory, the distribution model is relatively1203complex.1204 1205While not completely water-tight, all major memory usages by a given1206cgroup are tracked so that the total memory consumption can be1207accounted and controlled to a reasonable extent.  Currently, the1208following types of memory usages are tracked.1209 1210- Userland memory - page cache and anonymous memory.1211 1212- Kernel data structures such as dentries and inodes.1213 1214- TCP socket buffers.1215 1216The above list may expand in the future for better coverage.1217 1218 1219Memory Interface Files1220~~~~~~~~~~~~~~~~~~~~~~1221 1222All memory amounts are in bytes.  If a value which is not aligned to1223PAGE_SIZE is written, the value may be rounded up to the closest1224PAGE_SIZE multiple when read back.1225 1226  memory.current1227	A read-only single value file which exists on non-root1228	cgroups.1229 1230	The total amount of memory currently being used by the cgroup1231	and its descendants.1232 1233  memory.min1234	A read-write single value file which exists on non-root1235	cgroups.  The default is "0".1236 1237	Hard memory protection.  If the memory usage of a cgroup1238	is within its effective min boundary, the cgroup's memory1239	won't be reclaimed under any conditions. If there is no1240	unprotected reclaimable memory available, OOM killer1241	is invoked. Above the effective min boundary (or1242	effective low boundary if it is higher), pages are reclaimed1243	proportionally to the overage, reducing reclaim pressure for1244	smaller overages.1245 1246	Effective min boundary is limited by memory.min values of1247	all ancestor cgroups. If there is memory.min overcommitment1248	(child cgroup or cgroups are requiring more protected memory1249	than parent will allow), then each child cgroup will get1250	the part of parent's protection proportional to its1251	actual memory usage below memory.min.1252 1253	Putting more memory than generally available under this1254	protection is discouraged and may lead to constant OOMs.1255 1256	If a memory cgroup is not populated with processes,1257	its memory.min is ignored.1258 1259  memory.low1260	A read-write single value file which exists on non-root1261	cgroups.  The default is "0".1262 1263	Best-effort memory protection.  If the memory usage of a1264	cgroup is within its effective low boundary, the cgroup's1265	memory won't be reclaimed unless there is no reclaimable1266	memory available in unprotected cgroups.1267	Above the effective low	boundary (or 1268	effective min boundary if it is higher), pages are reclaimed1269	proportionally to the overage, reducing reclaim pressure for1270	smaller overages.1271 1272	Effective low boundary is limited by memory.low values of1273	all ancestor cgroups. If there is memory.low overcommitment1274	(child cgroup or cgroups are requiring more protected memory1275	than parent will allow), then each child cgroup will get1276	the part of parent's protection proportional to its1277	actual memory usage below memory.low.1278 1279	Putting more memory than generally available under this1280	protection is discouraged.1281 1282  memory.high1283	A read-write single value file which exists on non-root1284	cgroups.  The default is "max".1285 1286	Memory usage throttle limit.  If a cgroup's usage goes1287	over the high boundary, the processes of the cgroup are1288	throttled and put under heavy reclaim pressure.1289 1290	Going over the high limit never invokes the OOM killer and1291	under extreme conditions the limit may be breached. The high1292	limit should be used in scenarios where an external process1293	monitors the limited cgroup to alleviate heavy reclaim1294	pressure.1295 1296  memory.max1297	A read-write single value file which exists on non-root1298	cgroups.  The default is "max".1299 1300	Memory usage hard limit.  This is the main mechanism to limit1301	memory usage of a cgroup.  If a cgroup's memory usage reaches1302	this limit and can't be reduced, the OOM killer is invoked in1303	the cgroup. Under certain circumstances, the usage may go1304	over the limit temporarily.1305 1306	In default configuration regular 0-order allocations always1307	succeed unless OOM killer chooses current task as a victim.1308 1309	Some kinds of allocations don't invoke the OOM killer.1310	Caller could retry them differently, return into userspace1311	as -ENOMEM or silently ignore in cases like disk readahead.1312 1313  memory.reclaim1314	A write-only nested-keyed file which exists for all cgroups.1315 1316	This is a simple interface to trigger memory reclaim in the1317	target cgroup.1318 1319	Example::1320 1321	  echo "1G" > memory.reclaim1322 1323	Please note that the kernel can over or under reclaim from1324	the target cgroup. If less bytes are reclaimed than the1325	specified amount, -EAGAIN is returned.1326 1327	Please note that the proactive reclaim (triggered by this1328	interface) is not meant to indicate memory pressure on the1329	memory cgroup. Therefore socket memory balancing triggered by1330	the memory reclaim normally is not exercised in this case.1331	This means that the networking layer will not adapt based on1332	reclaim induced by memory.reclaim.1333 1334The following nested keys are defined.1335 1336	  ==========            ================================1337	  swappiness            Swappiness value to reclaim with1338	  ==========            ================================1339 1340	Specifying a swappiness value instructs the kernel to perform1341	the reclaim with that swappiness value. Note that this has the1342	same semantics as vm.swappiness applied to memcg reclaim with1343	all the existing limitations and potential future extensions.1344 1345  memory.peak1346	A read-write single value file which exists on non-root cgroups.1347 1348	The max memory usage recorded for the cgroup and its descendants since1349	either the creation of the cgroup or the most recent reset for that FD.1350 1351	A write of any non-empty string to this file resets it to the1352	current memory usage for subsequent reads through the same1353	file descriptor.1354 1355  memory.oom.group1356	A read-write single value file which exists on non-root1357	cgroups.  The default value is "0".1358 1359	Determines whether the cgroup should be treated as1360	an indivisible workload by the OOM killer. If set,1361	all tasks belonging to the cgroup or to its descendants1362	(if the memory cgroup is not a leaf cgroup) are killed1363	together or not at all. This can be used to avoid1364	partial kills to guarantee workload integrity.1365 1366	Tasks with the OOM protection (oom_score_adj set to -1000)1367	are treated as an exception and are never killed.1368 1369	If the OOM killer is invoked in a cgroup, it's not going1370	to kill any tasks outside of this cgroup, regardless1371	memory.oom.group values of ancestor cgroups.1372 1373  memory.events1374	A read-only flat-keyed file which exists on non-root cgroups.1375	The following entries are defined.  Unless specified1376	otherwise, a value change in this file generates a file1377	modified event.1378 1379	Note that all fields in this file are hierarchical and the1380	file modified event can be generated due to an event down the1381	hierarchy. For the local events at the cgroup level see1382	memory.events.local.1383 1384	  low1385		The number of times the cgroup is reclaimed due to1386		high memory pressure even though its usage is under1387		the low boundary.  This usually indicates that the low1388		boundary is over-committed.1389 1390	  high1391		The number of times processes of the cgroup are1392		throttled and routed to perform direct memory reclaim1393		because the high memory boundary was exceeded.  For a1394		cgroup whose memory usage is capped by the high limit1395		rather than global memory pressure, this event's1396		occurrences are expected.1397 1398	  max1399		The number of times the cgroup's memory usage was1400		about to go over the max boundary.  If direct reclaim1401		fails to bring it down, the cgroup goes to OOM state.1402 1403	  oom1404		The number of time the cgroup's memory usage was1405		reached the limit and allocation was about to fail.1406 1407		This event is not raised if the OOM killer is not1408		considered as an option, e.g. for failed high-order1409		allocations or if caller asked to not retry attempts.1410 1411	  oom_kill1412		The number of processes belonging to this cgroup1413		killed by any kind of OOM killer.1414 1415          oom_group_kill1416                The number of times a group OOM has occurred.1417 1418  memory.events.local1419	Similar to memory.events but the fields in the file are local1420	to the cgroup i.e. not hierarchical. The file modified event1421	generated on this file reflects only the local events.1422 1423  memory.stat1424	A read-only flat-keyed file which exists on non-root cgroups.1425 1426	This breaks down the cgroup's memory footprint into different1427	types of memory, type-specific details, and other information1428	on the state and past events of the memory management system.1429 1430	All memory amounts are in bytes.1431 1432	The entries are ordered to be human readable, and new entries1433	can show up in the middle. Don't rely on items remaining in a1434	fixed position; use the keys to look up specific values!1435 1436	If the entry has no per-node counter (or not show in the1437	memory.numa_stat). We use 'npn' (non-per-node) as the tag1438	to indicate that it will not show in the memory.numa_stat.1439 1440	  anon1441		Amount of memory used in anonymous mappings such as1442		brk(), sbrk(), and mmap(MAP_ANONYMOUS)1443 1444	  file1445		Amount of memory used to cache filesystem data,1446		including tmpfs and shared memory.1447 1448	  kernel (npn)1449		Amount of total kernel memory, including1450		(kernel_stack, pagetables, percpu, vmalloc, slab) in1451		addition to other kernel memory use cases.1452 1453	  kernel_stack1454		Amount of memory allocated to kernel stacks.1455 1456	  pagetables1457                Amount of memory allocated for page tables.1458 1459	  sec_pagetables1460		Amount of memory allocated for secondary page tables,1461		this currently includes KVM mmu allocations on x861462		and arm64 and IOMMU page tables.1463 1464	  percpu (npn)1465		Amount of memory used for storing per-cpu kernel1466		data structures.1467 1468	  sock (npn)1469		Amount of memory used in network transmission buffers1470 1471	  vmalloc (npn)1472		Amount of memory used for vmap backed memory.1473 1474	  shmem1475		Amount of cached filesystem data that is swap-backed,1476		such as tmpfs, shm segments, shared anonymous mmap()s1477 1478	  zswap1479		Amount of memory consumed by the zswap compression backend.1480 1481	  zswapped1482		Amount of application memory swapped out to zswap.1483 1484	  file_mapped1485		Amount of cached filesystem data mapped with mmap()1486 1487	  file_dirty1488		Amount of cached filesystem data that was modified but1489		not yet written back to disk1490 1491	  file_writeback1492		Amount of cached filesystem data that was modified and1493		is currently being written back to disk1494 1495	  swapcached1496		Amount of swap cached in memory. The swapcache is accounted1497		against both memory and swap usage.1498 1499	  anon_thp1500		Amount of memory used in anonymous mappings backed by1501		transparent hugepages1502 1503	  file_thp1504		Amount of cached filesystem data backed by transparent1505		hugepages1506 1507	  shmem_thp1508		Amount of shm, tmpfs, shared anonymous mmap()s backed by1509		transparent hugepages1510 1511	  inactive_anon, active_anon, inactive_file, active_file, unevictable1512		Amount of memory, swap-backed and filesystem-backed,1513		on the internal memory management lists used by the1514		page reclaim algorithm.1515 1516		As these represent internal list state (eg. shmem pages are on anon1517		memory management lists), inactive_foo + active_foo may not be equal to1518		the value for the foo counter, since the foo counter is type-based, not1519		list-based.1520 1521	  slab_reclaimable1522		Part of "slab" that might be reclaimed, such as1523		dentries and inodes.1524 1525	  slab_unreclaimable1526		Part of "slab" that cannot be reclaimed on memory1527		pressure.1528 1529	  slab (npn)1530		Amount of memory used for storing in-kernel data1531		structures.1532 1533	  workingset_refault_anon1534		Number of refaults of previously evicted anonymous pages.1535 1536	  workingset_refault_file1537		Number of refaults of previously evicted file pages.1538 1539	  workingset_activate_anon1540		Number of refaulted anonymous pages that were immediately1541		activated.1542 1543	  workingset_activate_file1544		Number of refaulted file pages that were immediately activated.1545 1546	  workingset_restore_anon1547		Number of restored anonymous pages which have been detected as1548		an active workingset before they got reclaimed.1549 1550	  workingset_restore_file1551		Number of restored file pages which have been detected as an1552		active workingset before they got reclaimed.1553 1554	  workingset_nodereclaim1555		Number of times a shadow node has been reclaimed1556 1557	  pgscan (npn)1558		Amount of scanned pages (in an inactive LRU list)1559 1560	  pgsteal (npn)1561		Amount of reclaimed pages1562 1563	  pgscan_kswapd (npn)1564		Amount of scanned pages by kswapd (in an inactive LRU list)1565 1566	  pgscan_direct (npn)1567		Amount of scanned pages directly  (in an inactive LRU list)1568 1569	  pgscan_khugepaged (npn)1570		Amount of scanned pages by khugepaged  (in an inactive LRU list)1571 1572	  pgsteal_kswapd (npn)1573		Amount of reclaimed pages by kswapd1574 1575	  pgsteal_direct (npn)1576		Amount of reclaimed pages directly1577 1578	  pgsteal_khugepaged (npn)1579		Amount of reclaimed pages by khugepaged1580 1581	  pgfault (npn)1582		Total number of page faults incurred1583 1584	  pgmajfault (npn)1585		Number of major page faults incurred1586 1587	  pgrefill (npn)1588		Amount of scanned pages (in an active LRU list)1589 1590	  pgactivate (npn)1591		Amount of pages moved to the active LRU list1592 1593	  pgdeactivate (npn)1594		Amount of pages moved to the inactive LRU list1595 1596	  pglazyfree (npn)1597		Amount of pages postponed to be freed under memory pressure1598 1599	  pglazyfreed (npn)1600		Amount of reclaimed lazyfree pages1601 1602	  swpin_zero1603		Number of pages swapped into memory and filled with zero, where I/O1604		was optimized out because the page content was detected to be zero1605		during swapout.1606 1607	  swpout_zero1608		Number of zero-filled pages swapped out with I/O skipped due to the1609		content being detected as zero.1610 1611	  zswpin1612		Number of pages moved in to memory from zswap.1613 1614	  zswpout1615		Number of pages moved out of memory to zswap.1616 1617	  zswpwb1618		Number of pages written from zswap to swap.1619 1620	  thp_fault_alloc (npn)1621		Number of transparent hugepages which were allocated to satisfy1622		a page fault. This counter is not present when CONFIG_TRANSPARENT_HUGEPAGE1623                is not set.1624 1625	  thp_collapse_alloc (npn)1626		Number of transparent hugepages which were allocated to allow1627		collapsing an existing range of pages. This counter is not1628		present when CONFIG_TRANSPARENT_HUGEPAGE is not set.1629 1630	  thp_swpout (npn)1631		Number of transparent hugepages which are swapout in one piece1632		without splitting.1633 1634	  thp_swpout_fallback (npn)1635		Number of transparent hugepages which were split before swapout.1636		Usually because failed to allocate some continuous swap space1637		for the huge page.1638 1639	  numa_pages_migrated (npn)1640		Number of pages migrated by NUMA balancing.1641 1642	  numa_pte_updates (npn)1643		Number of pages whose page table entries are modified by1644		NUMA balancing to produce NUMA hinting faults on access.1645 1646	  numa_hint_faults (npn)1647		Number of NUMA hinting faults.1648 1649	  pgdemote_kswapd1650		Number of pages demoted by kswapd.1651 1652	  pgdemote_direct1653		Number of pages demoted directly.1654 1655	  pgdemote_khugepaged1656		Number of pages demoted by khugepaged.1657 1658  memory.numa_stat1659	A read-only nested-keyed file which exists on non-root cgroups.1660 1661	This breaks down the cgroup's memory footprint into different1662	types of memory, type-specific details, and other information1663	per node on the state of the memory management system.1664 1665	This is useful for providing visibility into the NUMA locality1666	information within an memcg since the pages are allowed to be1667	allocated from any physical node. One of the use case is evaluating1668	application performance by combining this information with the1669	application's CPU allocation.1670 1671	All memory amounts are in bytes.1672 1673	The output format of memory.numa_stat is::1674 1675	  type N0=<bytes in node 0> N1=<bytes in node 1> ...1676 1677	The entries are ordered to be human readable, and new entries1678	can show up in the middle. Don't rely on items remaining in a1679	fixed position; use the keys to look up specific values!1680 1681	The entries can refer to the memory.stat.1682 1683  memory.swap.current1684	A read-only single value file which exists on non-root1685	cgroups.1686 1687	The total amount of swap currently being used by the cgroup1688	and its descendants.1689 1690  memory.swap.high1691	A read-write single value file which exists on non-root1692	cgroups.  The default is "max".1693 1694	Swap usage throttle limit.  If a cgroup's swap usage exceeds1695	this limit, all its further allocations will be throttled to1696	allow userspace to implement custom out-of-memory procedures.1697 1698	This limit marks a point of no return for the cgroup. It is NOT1699	designed to manage the amount of swapping a workload does1700	during regular operation. Compare to memory.swap.max, which1701	prohibits swapping past a set amount, but lets the cgroup1702	continue unimpeded as long as other memory can be reclaimed.1703 1704	Healthy workloads are not expected to reach this limit.1705 1706  memory.swap.peak1707	A read-write single value file which exists on non-root cgroups.1708 1709	The max swap usage recorded for the cgroup and its descendants since1710	the creation of the cgroup or the most recent reset for that FD.1711 1712	A write of any non-empty string to this file resets it to the1713	current memory usage for subsequent reads through the same1714	file descriptor.1715 1716  memory.swap.max1717	A read-write single value file which exists on non-root1718	cgroups.  The default is "max".1719 1720	Swap usage hard limit.  If a cgroup's swap usage reaches this1721	limit, anonymous memory of the cgroup will not be swapped out.1722 1723  memory.swap.events1724	A read-only flat-keyed file which exists on non-root cgroups.1725	The following entries are defined.  Unless specified1726	otherwise, a value change in this file generates a file1727	modified event.1728 1729	  high1730		The number of times the cgroup's swap usage was over1731		the high threshold.1732 1733	  max1734		The number of times the cgroup's swap usage was about1735		to go over the max boundary and swap allocation1736		failed.1737 1738	  fail1739		The number of times swap allocation failed either1740		because of running out of swap system-wide or max1741		limit.1742 1743	When reduced under the current usage, the existing swap1744	entries are reclaimed gradually and the swap usage may stay1745	higher than the limit for an extended period of time.  This1746	reduces the impact on the workload and memory management.1747 1748  memory.zswap.current1749	A read-only single value file which exists on non-root1750	cgroups.1751 1752	The total amount of memory consumed by the zswap compression1753	backend.1754 1755  memory.zswap.max1756	A read-write single value file which exists on non-root1757	cgroups.  The default is "max".1758 1759	Zswap usage hard limit. If a cgroup's zswap pool reaches this1760	limit, it will refuse to take any more stores before existing1761	entries fault back in or are written out to disk.1762 1763  memory.zswap.writeback1764	A read-write single value file. The default value is "1".1765	Note that this setting is hierarchical, i.e. the writeback would be1766	implicitly disabled for child cgroups if the upper hierarchy1767	does so.1768 1769	When this is set to 0, all swapping attempts to swapping devices1770	are disabled. This included both zswap writebacks, and swapping due1771	to zswap store failures. If the zswap store failures are recurring1772	(for e.g if the pages are incompressible), users can observe1773	reclaim inefficiency after disabling writeback (because the same1774	pages might be rejected again and again).1775 1776	Note that this is subtly different from setting memory.swap.max to1777	0, as it still allows for pages to be written to the zswap pool.1778	This setting has no effect if zswap is disabled, and swapping1779	is allowed unless memory.swap.max is set to 0.1780 1781  memory.pressure1782	A read-only nested-keyed file.1783 1784	Shows pressure stall information for memory. See1785	:ref:`Documentation/accounting/psi.rst <psi>` for details.1786 1787 1788Usage Guidelines1789~~~~~~~~~~~~~~~~1790 1791"memory.high" is the main mechanism to control memory usage.1792Over-committing on high limit (sum of high limits > available memory)1793and letting global memory pressure to distribute memory according to1794usage is a viable strategy.1795 1796Because breach of the high limit doesn't trigger the OOM killer but1797throttles the offending cgroup, a management agent has ample1798opportunities to monitor and take appropriate actions such as granting1799more memory or terminating the workload.1800 1801Determining whether a cgroup has enough memory is not trivial as1802memory usage doesn't indicate whether the workload can benefit from1803more memory.  For example, a workload which writes data received from1804network to a file can use all available memory but can also operate as1805performant with a small amount of memory.  A measure of memory1806pressure - how much the workload is being impacted due to lack of1807memory - is necessary to determine whether a workload needs more1808memory; unfortunately, memory pressure monitoring mechanism isn't1809implemented yet.1810 1811 1812Memory Ownership1813~~~~~~~~~~~~~~~~1814 1815A memory area is charged to the cgroup which instantiated it and stays1816charged to the cgroup until the area is released.  Migrating a process1817to a different cgroup doesn't move the memory usages that it1818instantiated while in the previous cgroup to the new cgroup.1819 1820A memory area may be used by processes belonging to different cgroups.1821To which cgroup the area will be charged is in-deterministic; however,1822over time, the memory area is likely to end up in a cgroup which has1823enough memory allowance to avoid high reclaim pressure.1824 1825If a cgroup sweeps a considerable amount of memory which is expected1826to be accessed repeatedly by other cgroups, it may make sense to use1827POSIX_FADV_DONTNEED to relinquish the ownership of memory areas1828belonging to the affected files to ensure correct memory ownership.1829 1830 1831IO1832--1833 1834The "io" controller regulates the distribution of IO resources.  This1835controller implements both weight based and absolute bandwidth or IOPS1836limit distribution; however, weight based distribution is available1837only if cfq-iosched is in use and neither scheme is available for1838blk-mq devices.1839 1840 1841IO Interface Files1842~~~~~~~~~~~~~~~~~~1843 1844  io.stat1845	A read-only nested-keyed file.1846 1847	Lines are keyed by $MAJ:$MIN device numbers and not ordered.1848	The following nested keys are defined.1849 1850	  ======	=====================1851	  rbytes	Bytes read1852	  wbytes	Bytes written1853	  rios		Number of read IOs1854	  wios		Number of write IOs1855	  dbytes	Bytes discarded1856	  dios		Number of discard IOs1857	  ======	=====================1858 1859	An example read output follows::1860 1861	  8:16 rbytes=1459200 wbytes=314773504 rios=192 wios=353 dbytes=0 dios=01862	  8:0 rbytes=90430464 wbytes=299008000 rios=8950 wios=1252 dbytes=50331648 dios=30211863 1864  io.cost.qos1865	A read-write nested-keyed file which exists only on the root1866	cgroup.1867 1868	This file configures the Quality of Service of the IO cost1869	model based controller (CONFIG_BLK_CGROUP_IOCOST) which1870	currently implements "io.weight" proportional control.  Lines1871	are keyed by $MAJ:$MIN device numbers and not ordered.  The1872	line for a given device is populated on the first write for1873	the device on "io.cost.qos" or "io.cost.model".  The following1874	nested keys are defined.1875 1876	  ======	=====================================1877	  enable	Weight-based control enable1878	  ctrl		"auto" or "user"1879	  rpct		Read latency percentile    [0, 100]1880	  rlat		Read latency threshold1881	  wpct		Write latency percentile   [0, 100]1882	  wlat		Write latency threshold1883	  min		Minimum scaling percentage [1, 10000]1884	  max		Maximum scaling percentage [1, 10000]1885	  ======	=====================================1886 1887	The controller is disabled by default and can be enabled by1888	setting "enable" to 1.  "rpct" and "wpct" parameters default1889	to zero and the controller uses internal device saturation1890	state to adjust the overall IO rate between "min" and "max".1891 1892	When a better control quality is needed, latency QoS1893	parameters can be configured.  For example::1894 1895	  8:16 enable=1 ctrl=auto rpct=95.00 rlat=75000 wpct=95.00 wlat=150000 min=50.00 max=150.01896 1897	shows that on sdb, the controller is enabled, will consider1898	the device saturated if the 95th percentile of read completion1899	latencies is above 75ms or write 150ms, and adjust the overall1900	IO issue rate between 50% and 150% accordingly.1901 1902	The lower the saturation point, the better the latency QoS at1903	the cost of aggregate bandwidth.  The narrower the allowed1904	adjustment range between "min" and "max", the more conformant1905	to the cost model the IO behavior.  Note that the IO issue1906	base rate may be far off from 100% and setting "min" and "max"1907	blindly can lead to a significant loss of device capacity or1908	control quality.  "min" and "max" are useful for regulating1909	devices which show wide temporary behavior changes - e.g. a1910	ssd which accepts writes at the line speed for a while and1911	then completely stalls for multiple seconds.1912 1913	When "ctrl" is "auto", the parameters are controlled by the1914	kernel and may change automatically.  Setting "ctrl" to "user"1915	or setting any of the percentile and latency parameters puts1916	it into "user" mode and disables the automatic changes.  The1917	automatic mode can be restored by setting "ctrl" to "auto".1918 1919  io.cost.model1920	A read-write nested-keyed file which exists only on the root1921	cgroup.1922 1923	This file configures the cost model of the IO cost model based1924	controller (CONFIG_BLK_CGROUP_IOCOST) which currently1925	implements "io.weight" proportional control.  Lines are keyed1926	by $MAJ:$MIN device numbers and not ordered.  The line for a1927	given device is populated on the first write for the device on1928	"io.cost.qos" or "io.cost.model".  The following nested keys1929	are defined.1930 1931	  =====		================================1932	  ctrl		"auto" or "user"1933	  model		The cost model in use - "linear"1934	  =====		================================1935 1936	When "ctrl" is "auto", the kernel may change all parameters1937	dynamically.  When "ctrl" is set to "user" or any other1938	parameters are written to, "ctrl" become "user" and the1939	automatic changes are disabled.1940 1941	When "model" is "linear", the following model parameters are1942	defined.1943 1944	  =============	========================================1945	  [r|w]bps	The maximum sequential IO throughput1946	  [r|w]seqiops	The maximum 4k sequential IOs per second1947	  [r|w]randiops	The maximum 4k random IOs per second1948	  =============	========================================1949 1950	From the above, the builtin linear model determines the base1951	costs of a sequential and random IO and the cost coefficient1952	for the IO size.  While simple, this model can cover most1953	common device classes acceptably.1954 1955	The IO cost model isn't expected to be accurate in absolute1956	sense and is scaled to the device behavior dynamically.1957 1958	If needed, tools/cgroup/iocost_coef_gen.py can be used to1959	generate device-specific coefficients.1960 1961  io.weight1962	A read-write flat-keyed file which exists on non-root cgroups.1963	The default is "default 100".1964 1965	The first line is the default weight applied to devices1966	without specific override.  The rest are overrides keyed by1967	$MAJ:$MIN device numbers and not ordered.  The weights are in1968	the range [1, 10000] and specifies the relative amount IO time1969	the cgroup can use in relation to its siblings.1970 1971	The default weight can be updated by writing either "default1972	$WEIGHT" or simply "$WEIGHT".  Overrides can be set by writing1973	"$MAJ:$MIN $WEIGHT" and unset by writing "$MAJ:$MIN default".1974 1975	An example read output follows::1976 1977	  default 1001978	  8:16 2001979	  8:0 501980 1981  io.max1982	A read-write nested-keyed file which exists on non-root1983	cgroups.1984 1985	BPS and IOPS based IO limit.  Lines are keyed by $MAJ:$MIN1986	device numbers and not ordered.  The following nested keys are1987	defined.1988 1989	  =====		==================================1990	  rbps		Max read bytes per second1991	  wbps		Max write bytes per second1992	  riops		Max read IO operations per second1993	  wiops		Max write IO operations per second1994	  =====		==================================1995 1996	When writing, any number of nested key-value pairs can be1997	specified in any order.  "max" can be specified as the value1998	to remove a specific limit.  If the same key is specified1999	multiple times, the outcome is undefined.2000 2001	BPS and IOPS are measured in each IO direction and IOs are2002	delayed if limit is reached.  Temporary bursts are allowed.2003 2004	Setting read limit at 2M BPS and write at 120 IOPS for 8:16::2005 2006	  echo "8:16 rbps=2097152 wiops=120" > io.max2007 2008	Reading returns the following::2009 2010	  8:16 rbps=2097152 wbps=max riops=max wiops=1202011 2012	Write IOPS limit can be removed by writing the following::2013 2014	  echo "8:16 wiops=max" > io.max2015 2016	Reading now returns the following::2017 2018	  8:16 rbps=2097152 wbps=max riops=max wiops=max2019 2020  io.pressure2021	A read-only nested-keyed file.2022 2023	Shows pressure stall information for IO. See2024	:ref:`Documentation/accounting/psi.rst <psi>` for details.2025 2026 2027Writeback2028~~~~~~~~~2029 2030Page cache is dirtied through buffered writes and shared mmaps and2031written asynchronously to the backing filesystem by the writeback2032mechanism.  Writeback sits between the memory and IO domains and2033regulates the proportion of dirty memory by balancing dirtying and2034write IOs.2035 2036The io controller, in conjunction with the memory controller,2037implements control of page cache writeback IOs.  The memory controller2038defines the memory domain that dirty memory ratio is calculated and2039maintained for and the io controller defines the io domain which2040writes out dirty pages for the memory domain.  Both system-wide and2041per-cgroup dirty memory states are examined and the more restrictive2042of the two is enforced.2043 2044cgroup writeback requires explicit support from the underlying2045filesystem.  Currently, cgroup writeback is implemented on ext2, ext4,2046btrfs, f2fs, and xfs.  On other filesystems, all writeback IOs are 2047attributed to the root cgroup.2048 2049There are inherent differences in memory and writeback management2050which affects how cgroup ownership is tracked.  Memory is tracked per2051page while writeback per inode.  For the purpose of writeback, an2052inode is assigned to a cgroup and all IO requests to write dirty pages2053from the inode are attributed to that cgroup.2054 2055As cgroup ownership for memory is tracked per page, there can be pages2056which are associated with different cgroups than the one the inode is2057associated with.  These are called foreign pages.  The writeback2058constantly keeps track of foreign pages and, if a particular foreign2059cgroup becomes the majority over a certain period of time, switches2060the ownership of the inode to that cgroup.2061 2062While this model is enough for most use cases where a given inode is2063mostly dirtied by a single cgroup even when the main writing cgroup2064changes over time, use cases where multiple cgroups write to a single2065inode simultaneously are not supported well.  In such circumstances, a2066significant portion of IOs are likely to be attributed incorrectly.2067As memory controller assigns page ownership on the first use and2068doesn't update it until the page is released, even if writeback2069strictly follows page ownership, multiple cgroups dirtying overlapping2070areas wouldn't work as expected.  It's recommended to avoid such usage2071patterns.2072 2073The sysctl knobs which affect writeback behavior are applied to cgroup2074writeback as follows.2075 2076  vm.dirty_background_ratio, vm.dirty_ratio2077	These ratios apply the same to cgroup writeback with the2078	amount of available memory capped by limits imposed by the2079	memory controller and system-wide clean memory.2080 2081  vm.dirty_background_bytes, vm.dirty_bytes2082	For cgroup writeback, this is calculated into ratio against2083	total available memory and applied the same way as2084	vm.dirty[_background]_ratio.2085 2086 2087IO Latency2088~~~~~~~~~~2089 2090This is a cgroup v2 controller for IO workload protection.  You provide a group2091with a latency target, and if the average latency exceeds that target the2092controller will throttle any peers that have a lower latency target than the2093protected workload.2094 2095The limits are only applied at the peer level in the hierarchy.  This means that2096in the diagram below, only groups A, B, and C will influence each other, and2097groups D and F will influence each other.  Group G will influence nobody::2098 2099			[root]2100		/	   |		\2101		A	   B		C2102	       /  \        |2103	      D    F	   G2104 2105 2106So the ideal way to configure this is to set io.latency in groups A, B, and C.2107Generally you do not want to set a value lower than the latency your device2108supports.  Experiment to find the value that works best for your workload.2109Start at higher than the expected latency for your device and watch the2110avg_lat value in io.stat for your workload group to get an idea of the2111latency you see during normal operation.  Use the avg_lat value as a basis for2112your real setting, setting at 10-15% higher than the value in io.stat.2113 2114How IO Latency Throttling Works2115~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~2116 2117io.latency is work conserving; so as long as everybody is meeting their latency2118target the controller doesn't do anything.  Once a group starts missing its2119target it begins throttling any peer group that has a higher target than itself.2120This throttling takes 2 forms:2121 2122- Queue depth throttling.  This is the number of outstanding IO's a group is2123  allowed to have.  We will clamp down relatively quickly, starting at no limit2124  and going all the way down to 1 IO at a time.2125 2126- Artificial delay induction.  There are certain types of IO that cannot be2127  throttled without possibly adversely affecting higher priority groups.  This2128  includes swapping and metadata IO.  These types of IO are allowed to occur2129  normally, however they are "charged" to the originating group.  If the2130  originating group is being throttled you will see the use_delay and delay2131  fields in io.stat increase.  The delay value is how many microseconds that are2132  being added to any process that runs in this group.  Because this number can2133  grow quite large if there is a lot of swapping or metadata IO occurring we2134  limit the individual delay events to 1 second at a time.2135 2136Once the victimized group starts meeting its latency target again it will start2137unthrottling any peer groups that were throttled previously.  If the victimized2138group simply stops doing IO the global counter will unthrottle appropriately.2139 2140IO Latency Interface Files2141~~~~~~~~~~~~~~~~~~~~~~~~~~2142 2143  io.latency2144	This takes a similar format as the other controllers.2145 2146		"MAJOR:MINOR target=<target time in microseconds>"2147 2148  io.stat2149	If the controller is enabled you will see extra stats in io.stat in2150	addition to the normal ones.2151 2152	  depth2153		This is the current queue depth for the group.2154 2155	  avg_lat2156		This is an exponential moving average with a decay rate of 1/exp2157		bound by the sampling interval.  The decay rate interval can be2158		calculated by multiplying the win value in io.stat by the2159		corresponding number of samples based on the win value.2160 2161	  win2162		The sampling window size in milliseconds.  This is the minimum2163		duration of time between evaluation events.  Windows only elapse2164		with IO activity.  Idle periods extend the most recent window.2165 2166IO Priority2167~~~~~~~~~~~2168 2169A single attribute controls the behavior of the I/O priority cgroup policy,2170namely the io.prio.class attribute. The following values are accepted for2171that attribute:2172 2173  no-change2174	Do not modify the I/O priority class.2175 2176  promote-to-rt2177	For requests that have a non-RT I/O priority class, change it into RT.2178	Also change the priority level of these requests to 4. Do not modify2179	the I/O priority of requests that have priority class RT.2180 2181  restrict-to-be2182	For requests that do not have an I/O priority class or that have I/O2183	priority class RT, change it into BE. Also change the priority level2184	of these requests to 0. Do not modify the I/O priority class of2185	requests that have priority class IDLE.2186 2187  idle2188	Change the I/O priority class of all requests into IDLE, the lowest2189	I/O priority class.2190 2191  none-to-rt2192	Deprecated. Just an alias for promote-to-rt.2193 2194The following numerical values are associated with the I/O priority policies:2195 2196+----------------+---+2197| no-change      | 0 |2198+----------------+---+2199| promote-to-rt  | 1 |2200+----------------+---+2201| restrict-to-be | 2 |2202+----------------+---+2203| idle           | 3 |2204+----------------+---+2205 2206The numerical value that corresponds to each I/O priority class is as follows:2207 2208+-------------------------------+---+2209| IOPRIO_CLASS_NONE             | 0 |2210+-------------------------------+---+2211| IOPRIO_CLASS_RT (real-time)   | 1 |2212+-------------------------------+---+2213| IOPRIO_CLASS_BE (best effort) | 2 |2214+-------------------------------+---+2215| IOPRIO_CLASS_IDLE             | 3 |2216+-------------------------------+---+2217 2218The algorithm to set the I/O priority class for a request is as follows:2219 2220- If I/O priority class policy is promote-to-rt, change the request I/O2221  priority class to IOPRIO_CLASS_RT and change the request I/O priority2222  level to 4.2223- If I/O priority class policy is not promote-to-rt, translate the I/O priority2224  class policy into a number, then change the request I/O priority class2225  into the maximum of the I/O priority class policy number and the numerical2226  I/O priority class.2227 2228PID2229---2230 2231The process number controller is used to allow a cgroup to stop any2232new tasks from being fork()'d or clone()'d after a specified limit is2233reached.2234 2235The number of tasks in a cgroup can be exhausted in ways which other2236controllers cannot prevent, thus warranting its own controller.  For2237example, a fork bomb is likely to exhaust the number of tasks before2238hitting memory restrictions.2239 2240Note that PIDs used in this controller refer to TIDs, process IDs as2241used by the kernel.2242 2243 2244PID Interface Files2245~~~~~~~~~~~~~~~~~~~2246 2247  pids.max2248	A read-write single value file which exists on non-root2249	cgroups.  The default is "max".2250 2251	Hard limit of number of processes.2252 2253  pids.current2254	A read-only single value file which exists on non-root cgroups.2255 2256	The number of processes currently in the cgroup and its2257	descendants.2258 2259  pids.peak2260	A read-only single value file which exists on non-root cgroups.2261 2262	The maximum value that the number of processes in the cgroup and its2263	descendants has ever reached.2264 2265  pids.events2266	A read-only flat-keyed file which exists on non-root cgroups. Unless2267	specified otherwise, a value change in this file generates a file2268	modified event. The following entries are defined.2269 2270	  max2271		The number of times the cgroup's total number of processes hit the pids.max2272		limit (see also pids_localevents).2273 2274  pids.events.local2275	Similar to pids.events but the fields in the file are local2276	to the cgroup i.e. not hierarchical. The file modified event2277	generated on this file reflects only the local events.2278 2279Organisational operations are not blocked by cgroup policies, so it is2280possible to have pids.current > pids.max.  This can be done by either2281setting the limit to be smaller than pids.current, or attaching enough2282processes to the cgroup such that pids.current is larger than2283pids.max.  However, it is not possible to violate a cgroup PID policy2284through fork() or clone(). These will return -EAGAIN if the creation2285of a new process would cause a cgroup policy to be violated.2286 2287 2288Cpuset2289------2290 2291The "cpuset" controller provides a mechanism for constraining2292the CPU and memory node placement of tasks to only the resources2293specified in the cpuset interface files in a task's current cgroup.2294This is especially valuable on large NUMA systems where placing jobs2295on properly sized subsets of the systems with careful processor and2296memory placement to reduce cross-node memory access and contention2297can improve overall system performance.2298 2299The "cpuset" controller is hierarchical.  That means the controller2300cannot use CPUs or memory nodes not allowed in its parent.2301 2302 2303Cpuset Interface Files2304~~~~~~~~~~~~~~~~~~~~~~2305 2306  cpuset.cpus2307	A read-write multiple values file which exists on non-root2308	cpuset-enabled cgroups.2309 2310	It lists the requested CPUs to be used by tasks within this2311	cgroup.  The actual list of CPUs to be granted, however, is2312	subjected to constraints imposed by its parent and can differ2313	from the requested CPUs.2314 2315	The CPU numbers are comma-separated numbers or ranges.2316	For example::2317 2318	  # cat cpuset.cpus2319	  0-4,6,8-102320 2321	An empty value indicates that the cgroup is using the same2322	setting as the nearest cgroup ancestor with a non-empty2323	"cpuset.cpus" or all the available CPUs if none is found.2324 2325	The value of "cpuset.cpus" stays constant until the next update2326	and won't be affected by any CPU hotplug events.2327 2328  cpuset.cpus.effective2329	A read-only multiple values file which exists on all2330	cpuset-enabled cgroups.2331 2332	It lists the onlined CPUs that are actually granted to this2333	cgroup by its parent.  These CPUs are allowed to be used by2334	tasks within the current cgroup.2335 2336	If "cpuset.cpus" is empty, the "cpuset.cpus.effective" file shows2337	all the CPUs from the parent cgroup that can be available to2338	be used by this cgroup.  Otherwise, it should be a subset of2339	"cpuset.cpus" unless none of the CPUs listed in "cpuset.cpus"2340	can be granted.  In this case, it will be treated just like an2341	empty "cpuset.cpus".2342 2343	Its value will be affected by CPU hotplug events.2344 2345  cpuset.mems2346	A read-write multiple values file which exists on non-root2347	cpuset-enabled cgroups.2348 2349	It lists the requested memory nodes to be used by tasks within2350	this cgroup.  The actual list of memory nodes granted, however,2351	is subjected to constraints imposed by its parent and can differ2352	from the requested memory nodes.2353 2354	The memory node numbers are comma-separated numbers or ranges.2355	For example::2356 2357	  # cat cpuset.mems2358	  0-1,32359 2360	An empty value indicates that the cgroup is using the same2361	setting as the nearest cgroup ancestor with a non-empty2362	"cpuset.mems" or all the available memory nodes if none2363	is found.2364 2365	The value of "cpuset.mems" stays constant until the next update2366	and won't be affected by any memory nodes hotplug events.2367 2368	Setting a non-empty value to "cpuset.mems" causes memory of2369	tasks within the cgroup to be migrated to the designated nodes if2370	they are currently using memory outside of the designated nodes.2371 2372	There is a cost for this memory migration.  The migration2373	may not be complete and some memory pages may be left behind.2374	So it is recommended that "cpuset.mems" should be set properly2375	before spawning new tasks into the cpuset.  Even if there is2376	a need to change "cpuset.mems" with active tasks, it shouldn't2377	be done frequently.2378 2379  cpuset.mems.effective2380	A read-only multiple values file which exists on all2381	cpuset-enabled cgroups.2382 2383	It lists the onlined memory nodes that are actually granted to2384	this cgroup by its parent. These memory nodes are allowed to2385	be used by tasks within the current cgroup.2386 2387	If "cpuset.mems" is empty, it shows all the memory nodes from the2388	parent cgroup that will be available to be used by this cgroup.2389	Otherwise, it should be a subset of "cpuset.mems" unless none of2390	the memory nodes listed in "cpuset.mems" can be granted.  In this2391	case, it will be treated just like an empty "cpuset.mems".2392 2393	Its value will be affected by memory nodes hotplug events.2394 2395  cpuset.cpus.exclusive2396	A read-write multiple values file which exists on non-root2397	cpuset-enabled cgroups.2398 2399	It lists all the exclusive CPUs that are allowed to be used2400	to create a new cpuset partition.  Its value is not used2401	unless the cgroup becomes a valid partition root.  See the2402	"cpuset.cpus.partition" section below for a description of what2403	a cpuset partition is.2404 2405	When the cgroup becomes a partition root, the actual exclusive2406	CPUs that are allocated to that partition are listed in2407	"cpuset.cpus.exclusive.effective" which may be different2408	from "cpuset.cpus.exclusive".  If "cpuset.cpus.exclusive"2409	has previously been set, "cpuset.cpus.exclusive.effective"2410	is always a subset of it.2411 2412	Users can manually set it to a value that is different from2413	"cpuset.cpus".	One constraint in setting it is that the list of2414	CPUs must be exclusive with respect to "cpuset.cpus.exclusive"2415	of its sibling.  If "cpuset.cpus.exclusive" of a sibling cgroup2416	isn't set, its "cpuset.cpus" value, if set, cannot be a subset2417	of it to leave at least one CPU available when the exclusive2418	CPUs are taken away.2419 2420	For a parent cgroup, any one of its exclusive CPUs can only2421	be distributed to at most one of its child cgroups.  Having an2422	exclusive CPU appearing in two or more of its child cgroups is2423	not allowed (the exclusivity rule).  A value that violates the2424	exclusivity rule will be rejected with a write error.2425 2426	The root cgroup is a partition root and all its available CPUs2427	are in its exclusive CPU set.2428 2429  cpuset.cpus.exclusive.effective2430	A read-only multiple values file which exists on all non-root2431	cpuset-enabled cgroups.2432 2433	This file shows the effective set of exclusive CPUs that2434	can be used to create a partition root.  The content2435	of this file will always be a subset of its parent's2436	"cpuset.cpus.exclusive.effective" if its parent is not the root2437	cgroup.  It will also be a subset of "cpuset.cpus.exclusive"2438	if it is set.  If "cpuset.cpus.exclusive" is not set, it is2439	treated to have an implicit value of "cpuset.cpus" in the2440	formation of local partition.2441 2442  cpuset.cpus.isolated2443	A read-only and root cgroup only multiple values file.2444 2445	This file shows the set of all isolated CPUs used in existing2446	isolated partitions. It will be empty if no isolated partition2447	is created.2448 2449  cpuset.cpus.partition2450	A read-write single value file which exists on non-root2451	cpuset-enabled cgroups.  This flag is owned by the parent cgroup2452	and is not delegatable.2453 2454	It accepts only the following input values when written to.2455 2456	  ==========	=====================================2457	  "member"	Non-root member of a partition2458	  "root"	Partition root2459	  "isolated"	Partition root without load balancing2460	  ==========	=====================================2461 2462	A cpuset partition is a collection of cpuset-enabled cgroups with2463	a partition root at the top of the hierarchy and its descendants2464	except those that are separate partition roots themselves and2465	their descendants.  A partition has exclusive access to the2466	set of exclusive CPUs allocated to it.	Other cgroups outside2467	of that partition cannot use any CPUs in that set.2468 2469	There are two types of partitions - local and remote.  A local2470	partition is one whose parent cgroup is also a valid partition2471	root.  A remote partition is one whose parent cgroup is not a2472	valid partition root itself.  Writing to "cpuset.cpus.exclusive"2473	is optional for the creation of a local partition as its2474	"cpuset.cpus.exclusive" file will assume an implicit value that2475	is the same as "cpuset.cpus" if it is not set.	Writing the2476	proper "cpuset.cpus.exclusive" values down the cgroup hierarchy2477	before the target partition root is mandatory for the creation2478	of a remote partition.2479 2480	Currently, a remote partition cannot be created under a local2481	partition.  All the ancestors of a remote partition root except2482	the root cgroup cannot be a partition root.2483 2484	The root cgroup is always a partition root and its state cannot2485	be changed.  All other non-root cgroups start out as "member".2486 2487	When set to "root", the current cgroup is the root of a new2488	partition or scheduling domain.  The set of exclusive CPUs is2489	determined by the value of its "cpuset.cpus.exclusive.effective".2490 2491	When set to "isolated", the CPUs in that partition will be in2492	an isolated state without any load balancing from the scheduler2493	and excluded from the unbound workqueues.  Tasks placed in such2494	a partition with multiple CPUs should be carefully distributed2495	and bound to each of the individual CPUs for optimal performance.2496 2497	A partition root ("root" or "isolated") can be in one of the2498	two possible states - valid or invalid.  An invalid partition2499	root is in a degraded state where some state information may2500	be retained, but behaves more like a "member".2501 2502	All possible state transitions among "member", "root" and2503	"isolated" are allowed.2504 2505	On read, the "cpuset.cpus.partition" file can show the following2506	values.2507 2508	  =============================	=====================================2509	  "member"			Non-root member of a partition2510	  "root"			Partition root2511	  "isolated"			Partition root without load balancing2512	  "root invalid (<reason>)"	Invalid partition root2513	  "isolated invalid (<reason>)"	Invalid isolated partition root2514	  =============================	=====================================2515 2516	In the case of an invalid partition root, a descriptive string on2517	why the partition is invalid is included within parentheses.2518 2519	For a local partition root to be valid, the following conditions2520	must be met.2521 2522	1) The parent cgroup is a valid partition root.2523	2) The "cpuset.cpus.exclusive.effective" file cannot be empty,2524	   though it may contain offline CPUs.2525	3) The "cpuset.cpus.effective" cannot be empty unless there is2526	   no task associated with this partition.2527 2528	For a remote partition root to be valid, all the above conditions2529	except the first one must be met.2530 2531	External events like hotplug or changes to "cpuset.cpus" or2532	"cpuset.cpus.exclusive" can cause a valid partition root to2533	become invalid and vice versa.	Note that a task cannot be2534	moved to a cgroup with empty "cpuset.cpus.effective".2535 2536	A valid non-root parent partition may distribute out all its CPUs2537	to its child local partitions when there is no task associated2538	with it.2539 2540	Care must be taken to change a valid partition root to "member"2541	as all its child local partitions, if present, will become2542	invalid causing disruption to tasks running in those child2543	partitions. These inactivated partitions could be recovered if2544	their parent is switched back to a partition root with a proper2545	value in "cpuset.cpus" or "cpuset.cpus.exclusive".2546 2547	Poll and inotify events are triggered whenever the state of2548	"cpuset.cpus.partition" changes.  That includes changes caused2549	by write to "cpuset.cpus.partition", cpu hotplug or other2550	changes that modify the validity status of the partition.2551	This will allow user space agents to monitor unexpected changes2552	to "cpuset.cpus.partition" without the need to do continuous2553	polling.2554 2555	A user can pre-configure certain CPUs to an isolated state2556	with load balancing disabled at boot time with the "isolcpus"2557	kernel boot command line option.  If those CPUs are to be put2558	into a partition, they have to be used in an isolated partition.2559 2560 2561Device controller2562-----------------2563 2564Device controller manages access to device files. It includes both2565creation of new device files (using mknod), and access to the2566existing device files.2567 2568Cgroup v2 device controller has no interface files and is implemented2569on top of cgroup BPF. To control access to device files, a user may2570create bpf programs of type BPF_PROG_TYPE_CGROUP_DEVICE and attach2571them to cgroups with BPF_CGROUP_DEVICE flag. On an attempt to access a2572device file, corresponding BPF programs will be executed, and depending2573on the return value the attempt will succeed or fail with -EPERM.2574 2575A BPF_PROG_TYPE_CGROUP_DEVICE program takes a pointer to the2576bpf_cgroup_dev_ctx structure, which describes the device access attempt:2577access type (mknod/read/write) and device (type, major and minor numbers).2578If the program returns 0, the attempt fails with -EPERM, otherwise it2579succeeds.2580 2581An example of BPF_PROG_TYPE_CGROUP_DEVICE program may be found in2582tools/testing/selftests/bpf/progs/dev_cgroup.c in the kernel source tree.2583 2584 2585RDMA2586----2587 2588The "rdma" controller regulates the distribution and accounting of2589RDMA resources.2590 2591RDMA Interface Files2592~~~~~~~~~~~~~~~~~~~~2593 2594  rdma.max2595	A readwrite nested-keyed file that exists for all the cgroups2596	except root that describes current configured resource limit2597	for a RDMA/IB device.2598 2599	Lines are keyed by device name and are not ordered.2600	Each line contains space separated resource name and its configured2601	limit that can be distributed.2602 2603	The following nested keys are defined.2604 2605	  ==========	=============================2606	  hca_handle	Maximum number of HCA Handles2607	  hca_object 	Maximum number of HCA Objects2608	  ==========	=============================2609 2610	An example for mlx4 and ocrdma device follows::2611 2612	  mlx4_0 hca_handle=2 hca_object=20002613	  ocrdma1 hca_handle=3 hca_object=max2614 2615  rdma.current2616	A read-only file that describes current resource usage.2617	It exists for all the cgroup except root.2618 2619	An example for mlx4 and ocrdma device follows::2620 2621	  mlx4_0 hca_handle=1 hca_object=202622	  ocrdma1 hca_handle=1 hca_object=232623 2624HugeTLB2625-------2626 2627The HugeTLB controller allows to limit the HugeTLB usage per control group and2628enforces the controller limit during page fault.2629 2630HugeTLB Interface Files2631~~~~~~~~~~~~~~~~~~~~~~~2632 2633  hugetlb.<hugepagesize>.current2634	Show current usage for "hugepagesize" hugetlb.  It exists for all2635	the cgroup except root.2636 2637  hugetlb.<hugepagesize>.max2638	Set/show the hard limit of "hugepagesize" hugetlb usage.2639	The default value is "max".  It exists for all the cgroup except root.2640 2641  hugetlb.<hugepagesize>.events2642	A read-only flat-keyed file which exists on non-root cgroups.2643 2644	  max2645		The number of allocation failure due to HugeTLB limit2646 2647  hugetlb.<hugepagesize>.events.local2648	Similar to hugetlb.<hugepagesize>.events but the fields in the file2649	are local to the cgroup i.e. not hierarchical. The file modified event2650	generated on this file reflects only the local events.2651 2652  hugetlb.<hugepagesize>.numa_stat2653	Similar to memory.numa_stat, it shows the numa information of the2654        hugetlb pages of <hugepagesize> in this cgroup.  Only active in2655        use hugetlb pages are included.  The per-node values are in bytes.2656 2657Misc2658----2659 2660The Miscellaneous cgroup provides the resource limiting and tracking2661mechanism for the scalar resources which cannot be abstracted like the other2662cgroup resources. Controller is enabled by the CONFIG_CGROUP_MISC config2663option.2664 2665A resource can be added to the controller via enum misc_res_type{} in the2666include/linux/misc_cgroup.h file and the corresponding name via misc_res_name[]2667in the kernel/cgroup/misc.c file. Provider of the resource must set its2668capacity prior to using the resource by calling misc_cg_set_capacity().2669 2670Once a capacity is set then the resource usage can be updated using charge and2671uncharge APIs. All of the APIs to interact with misc controller are in2672include/linux/misc_cgroup.h.2673 2674Misc Interface Files2675~~~~~~~~~~~~~~~~~~~~2676 2677Miscellaneous controller provides 3 interface files. If two misc resources (res_a and res_b) are registered then:2678 2679  misc.capacity2680        A read-only flat-keyed file shown only in the root cgroup.  It shows2681        miscellaneous scalar resources available on the platform along with2682        their quantities::2683 2684	  $ cat misc.capacity2685	  res_a 502686	  res_b 102687 2688  misc.current2689        A read-only flat-keyed file shown in the all cgroups.  It shows2690        the current usage of the resources in the cgroup and its children.::2691 2692	  $ cat misc.current2693	  res_a 32694	  res_b 02695 2696  misc.peak2697        A read-only flat-keyed file shown in all cgroups.  It shows the2698        historical maximum usage of the resources in the cgroup and its2699        children.::2700 2701	  $ cat misc.peak2702	  res_a 102703	  res_b 82704 2705  misc.max2706        A read-write flat-keyed file shown in the non root cgroups. Allowed2707        maximum usage of the resources in the cgroup and its children.::2708 2709	  $ cat misc.max2710	  res_a max2711	  res_b 42712 2713	Limit can be set by::2714 2715	  # echo res_a 1 > misc.max2716 2717	Limit can be set to max by::2718 2719	  # echo res_a max > misc.max2720 2721        Limits can be set higher than the capacity value in the misc.capacity2722        file.2723 2724  misc.events2725	A read-only flat-keyed file which exists on non-root cgroups. The2726	following entries are defined. Unless specified otherwise, a value2727	change in this file generates a file modified event. All fields in2728	this file are hierarchical.2729 2730	  max2731		The number of times the cgroup's resource usage was2732		about to go over the max boundary.2733 2734  misc.events.local2735        Similar to misc.events but the fields in the file are local to the2736        cgroup i.e. not hierarchical. The file modified event generated on2737        this file reflects only the local events.2738 2739Migration and Ownership2740~~~~~~~~~~~~~~~~~~~~~~~2741 2742A miscellaneous scalar resource is charged to the cgroup in which it is used2743first, and stays charged to that cgroup until that resource is freed. Migrating2744a process to a different cgroup does not move the charge to the destination2745cgroup where the process has moved.2746 2747Others2748------2749 2750perf_event2751~~~~~~~~~~2752 2753perf_event controller, if not mounted on a legacy hierarchy, is2754automatically enabled on the v2 hierarchy so that perf events can2755always be filtered by cgroup v2 path.  The controller can still be2756moved to a legacy hierarchy after v2 hierarchy is populated.2757 2758 2759Non-normative information2760-------------------------2761 2762This section contains information that isn't considered to be a part of2763the stable kernel API and so is subject to change.2764 2765 2766CPU controller root cgroup process behaviour2767~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~2768 2769When distributing CPU cycles in the root cgroup each thread in this2770cgroup is treated as if it was hosted in a separate child cgroup of the2771root cgroup. This child cgroup weight is dependent on its thread nice2772level.2773 2774For details of this mapping see sched_prio_to_weight array in2775kernel/sched/core.c file (values from this array should be scaled2776appropriately so the neutral - nice 0 - value is 100 instead of 1024).2777 2778 2779IO controller root cgroup process behaviour2780~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~2781 2782Root cgroup processes are hosted in an implicit leaf child node.2783When distributing IO resources this implicit child node is taken into2784account as if it was a normal child cgroup of the root cgroup with a2785weight value of 200.2786 2787 2788Namespace2789=========2790 2791Basics2792------2793 2794cgroup namespace provides a mechanism to virtualize the view of the2795"/proc/$PID/cgroup" file and cgroup mounts.  The CLONE_NEWCGROUP clone2796flag can be used with clone(2) and unshare(2) to create a new cgroup2797namespace.  The process running inside the cgroup namespace will have2798its "/proc/$PID/cgroup" output restricted to cgroupns root.  The2799cgroupns root is the cgroup of the process at the time of creation of2800the cgroup namespace.2801 2802Without cgroup namespace, the "/proc/$PID/cgroup" file shows the2803complete path of the cgroup of a process.  In a container setup where2804a set of cgroups and namespaces are intended to isolate processes the2805"/proc/$PID/cgroup" file may leak potential system level information2806to the isolated processes.  For example::2807 2808  # cat /proc/self/cgroup2809  0::/batchjobs/container_id12810 2811The path '/batchjobs/container_id1' can be considered as system-data2812and undesirable to expose to the isolated processes.  cgroup namespace2813can be used to restrict visibility of this path.  For example, before2814creating a cgroup namespace, one would see::2815 2816  # ls -l /proc/self/ns/cgroup2817  lrwxrwxrwx 1 root root 0 2014-07-15 10:37 /proc/self/ns/cgroup -> cgroup:[4026531835]2818  # cat /proc/self/cgroup2819  0::/batchjobs/container_id12820 2821After unsharing a new namespace, the view changes::2822 2823  # ls -l /proc/self/ns/cgroup2824  lrwxrwxrwx 1 root root 0 2014-07-15 10:35 /proc/self/ns/cgroup -> cgroup:[4026532183]2825  # cat /proc/self/cgroup2826  0::/2827 2828When some thread from a multi-threaded process unshares its cgroup2829namespace, the new cgroupns gets applied to the entire process (all2830the threads).  This is natural for the v2 hierarchy; however, for the2831legacy hierarchies, this may be unexpected.2832 2833A cgroup namespace is alive as long as there are processes inside or2834mounts pinning it.  When the last usage goes away, the cgroup2835namespace is destroyed.  The cgroupns root and the actual cgroups2836remain.2837 2838 2839The Root and Views2840------------------2841 2842The 'cgroupns root' for a cgroup namespace is the cgroup in which the2843process calling unshare(2) is running.  For example, if a process in2844/batchjobs/container_id1 cgroup calls unshare, cgroup2845/batchjobs/container_id1 becomes the cgroupns root.  For the2846init_cgroup_ns, this is the real root ('/') cgroup.2847 2848The cgroupns root cgroup does not change even if the namespace creator2849process later moves to a different cgroup::2850 2851  # ~/unshare -c # unshare cgroupns in some cgroup2852  # cat /proc/self/cgroup2853  0::/2854  # mkdir sub_cgrp_12855  # echo 0 > sub_cgrp_1/cgroup.procs2856  # cat /proc/self/cgroup2857  0::/sub_cgrp_12858 2859Each process gets its namespace-specific view of "/proc/$PID/cgroup"2860 2861Processes running inside the cgroup namespace will be able to see2862cgroup paths (in /proc/self/cgroup) only inside their root cgroup.2863From within an unshared cgroupns::2864 2865  # sleep 100000 &2866  [1] 73532867  # echo 7353 > sub_cgrp_1/cgroup.procs2868  # cat /proc/7353/cgroup2869  0::/sub_cgrp_12870 2871From the initial cgroup namespace, the real cgroup path will be2872visible::2873 2874  $ cat /proc/7353/cgroup2875  0::/batchjobs/container_id1/sub_cgrp_12876 2877From a sibling cgroup namespace (that is, a namespace rooted at a2878different cgroup), the cgroup path relative to its own cgroup2879namespace root will be shown.  For instance, if PID 7353's cgroup2880namespace root is at '/batchjobs/container_id2', then it will see::2881 2882  # cat /proc/7353/cgroup2883  0::/../container_id2/sub_cgrp_12884 2885Note that the relative path always starts with '/' to indicate that2886its relative to the cgroup namespace root of the caller.2887 2888 2889Migration and setns(2)2890----------------------2891 2892Processes inside a cgroup namespace can move into and out of the2893namespace root if they have proper access to external cgroups.  For2894example, from inside a namespace with cgroupns root at2895/batchjobs/container_id1, and assuming that the global hierarchy is2896still accessible inside cgroupns::2897 2898  # cat /proc/7353/cgroup2899  0::/sub_cgrp_12900  # echo 7353 > batchjobs/container_id2/cgroup.procs2901  # cat /proc/7353/cgroup2902  0::/../container_id22903 2904Note that this kind of setup is not encouraged.  A task inside cgroup2905namespace should only be exposed to its own cgroupns hierarchy.2906 2907setns(2) to another cgroup namespace is allowed when:2908 2909(a) the process has CAP_SYS_ADMIN against its current user namespace2910(b) the process has CAP_SYS_ADMIN against the target cgroup2911    namespace's userns2912 2913No implicit cgroup changes happen with attaching to another cgroup2914namespace.  It is expected that the someone moves the attaching2915process under the target cgroup namespace root.2916 2917 2918Interaction with Other Namespaces2919---------------------------------2920 2921Namespace specific cgroup hierarchy can be mounted by a process2922running inside a non-init cgroup namespace::2923 2924  # mount -t cgroup2 none $MOUNT_POINT2925 2926This will mount the unified cgroup hierarchy with cgroupns root as the2927filesystem root.  The process needs CAP_SYS_ADMIN against its user and2928mount namespaces.2929 2930The virtualization of /proc/self/cgroup file combined with restricting2931the view of cgroup hierarchy by namespace-private cgroupfs mount2932provides a properly isolated cgroup view inside the container.2933 2934 2935Information on Kernel Programming2936=================================2937 2938This section contains kernel programming information in the areas2939where interacting with cgroup is necessary.  cgroup core and2940controllers are not covered.2941 2942 2943Filesystem Support for Writeback2944--------------------------------2945 2946A filesystem can support cgroup writeback by updating2947address_space_operations->writepage[s]() to annotate bio's using the2948following two functions.2949 2950  wbc_init_bio(@wbc, @bio)2951	Should be called for each bio carrying writeback data and2952	associates the bio with the inode's owner cgroup and the2953	corresponding request queue.  This must be called after2954	a queue (device) has been associated with the bio and2955	before submission.2956 2957  wbc_account_cgroup_owner(@wbc, @page, @bytes)2958	Should be called for each data segment being written out.2959	While this function doesn't care exactly when it's called2960	during the writeback session, it's the easiest and most2961	natural to call it as data segments are added to a bio.2962 2963With writeback bio's annotated, cgroup support can be enabled per2964super_block by setting SB_I_CGROUPWB in ->s_iflags.  This allows for2965selective disabling of cgroup writeback support which is helpful when2966certain filesystem features, e.g. journaled data mode, are2967incompatible.2968 2969wbc_init_bio() binds the specified bio to its cgroup.  Depending on2970the configuration, the bio may be executed at a lower priority and if2971the writeback session is holding shared resources, e.g. a journal2972entry, may lead to priority inversion.  There is no one easy solution2973for the problem.  Filesystems can try to work around specific problem2974cases by skipping wbc_init_bio() and using bio_associate_blkg()2975directly.2976 2977 2978Deprecated v1 Core Features2979===========================2980 2981- Multiple hierarchies including named ones are not supported.2982 2983- All v1 mount options are not supported.2984 2985- The "tasks" file is removed and "cgroup.procs" is not sorted.2986 2987- "cgroup.clone_children" is removed.2988 2989- /proc/cgroups is meaningless for v2.  Use "cgroup.controllers" or2990  "cgroup.stat" files at the root instead.2991 2992 2993Issues with v1 and Rationales for v22994====================================2995 2996Multiple Hierarchies2997--------------------2998 2999cgroup v1 allowed an arbitrary number of hierarchies and each3000hierarchy could host any number of controllers.  While this seemed to3001provide a high level of flexibility, it wasn't useful in practice.3002 3003For example, as there is only one instance of each controller, utility3004type controllers such as freezer which can be useful in all3005hierarchies could only be used in one.  The issue is exacerbated by3006the fact that controllers couldn't be moved to another hierarchy once3007hierarchies were populated.  Another issue was that all controllers3008bound to a hierarchy were forced to have exactly the same view of the3009hierarchy.  It wasn't possible to vary the granularity depending on3010the specific controller.3011 3012In practice, these issues heavily limited which controllers could be3013put on the same hierarchy and most configurations resorted to putting3014each controller on its own hierarchy.  Only closely related ones, such3015as the cpu and cpuacct controllers, made sense to be put on the same3016hierarchy.  This often meant that userland ended up managing multiple3017similar hierarchies repeating the same steps on each hierarchy3018whenever a hierarchy management operation was necessary.3019 3020Furthermore, support for multiple hierarchies came at a steep cost.3021It greatly complicated cgroup core implementation but more importantly3022the support for multiple hierarchies restricted how cgroup could be3023used in general and what controllers was able to do.3024 3025There was no limit on how many hierarchies there might be, which meant3026that a thread's cgroup membership couldn't be described in finite3027length.  The key might contain any number of entries and was unlimited3028in length, which made it highly awkward to manipulate and led to3029addition of controllers which existed only to identify membership,3030which in turn exacerbated the original problem of proliferating number3031of hierarchies.3032 3033Also, as a controller couldn't have any expectation regarding the3034topologies of hierarchies other controllers might be on, each3035controller had to assume that all other controllers were attached to3036completely orthogonal hierarchies.  This made it impossible, or at3037least very cumbersome, for controllers to cooperate with each other.3038 3039In most use cases, putting controllers on hierarchies which are3040completely orthogonal to each other isn't necessary.  What usually is3041called for is the ability to have differing levels of granularity3042depending on the specific controller.  In other words, hierarchy may3043be collapsed from leaf towards root when viewed from specific3044controllers.  For example, a given configuration might not care about3045how memory is distributed beyond a certain level while still wanting3046to control how CPU cycles are distributed.3047 3048 3049Thread Granularity3050------------------3051 3052cgroup v1 allowed threads of a process to belong to different cgroups.3053This didn't make sense for some controllers and those controllers3054ended up implementing different ways to ignore such situations but3055much more importantly it blurred the line between API exposed to3056individual applications and system management interface.3057 3058Generally, in-process knowledge is available only to the process3059itself; thus, unlike service-level organization of processes,3060categorizing threads of a process requires active participation from3061the application which owns the target process.3062 3063cgroup v1 had an ambiguously defined delegation model which got abused3064in combination with thread granularity.  cgroups were delegated to3065individual applications so that they can create and manage their own3066sub-hierarchies and control resource distributions along them.  This3067effectively raised cgroup to the status of a syscall-like API exposed3068to lay programs.3069 3070First of all, cgroup has a fundamentally inadequate interface to be3071exposed this way.  For a process to access its own knobs, it has to3072extract the path on the target hierarchy from /proc/self/cgroup,3073construct the path by appending the name of the knob to the path, open3074and then read and/or write to it.  This is not only extremely clunky3075and unusual but also inherently racy.  There is no conventional way to3076define transaction across the required steps and nothing can guarantee3077that the process would actually be operating on its own sub-hierarchy.3078 3079cgroup controllers implemented a number of knobs which would never be3080accepted as public APIs because they were just adding control knobs to3081system-management pseudo filesystem.  cgroup ended up with interface3082knobs which were not properly abstracted or refined and directly3083revealed kernel internal details.  These knobs got exposed to3084individual applications through the ill-defined delegation mechanism3085effectively abusing cgroup as a shortcut to implementing public APIs3086without going through the required scrutiny.3087 3088This was painful for both userland and kernel.  Userland ended up with3089misbehaving and poorly abstracted interfaces and kernel exposing and3090locked into constructs inadvertently.3091 3092 3093Competition Between Inner Nodes and Threads3094-------------------------------------------3095 3096cgroup v1 allowed threads to be in any cgroups which created an3097interesting problem where threads belonging to a parent cgroup and its3098children cgroups competed for resources.  This was nasty as two3099different types of entities competed and there was no obvious way to3100settle it.  Different controllers did different things.3101 3102The cpu controller considered threads and cgroups as equivalents and3103mapped nice levels to cgroup weights.  This worked for some cases but3104fell flat when children wanted to be allocated specific ratios of CPU3105cycles and the number of internal threads fluctuated - the ratios3106constantly changed as the number of competing entities fluctuated.3107There also were other issues.  The mapping from nice level to weight3108wasn't obvious or universal, and there were various other knobs which3109simply weren't available for threads.3110 3111The io controller implicitly created a hidden leaf node for each3112cgroup to host the threads.  The hidden leaf had its own copies of all3113the knobs with ``leaf_`` prefixed.  While this allowed equivalent3114control over internal threads, it was with serious drawbacks.  It3115always added an extra layer of nesting which wouldn't be necessary3116otherwise, made the interface messy and significantly complicated the3117implementation.3118 3119The memory controller didn't have a way to control what happened3120between internal tasks and child cgroups and the behavior was not3121clearly defined.  There were attempts to add ad-hoc behaviors and3122knobs to tailor the behavior to specific workloads which would have3123led to problems extremely difficult to resolve in the long term.3124 3125Multiple controllers struggled with internal tasks and came up with3126different ways to deal with it; unfortunately, all the approaches were3127severely flawed and, furthermore, the widely different behaviors3128made cgroup as a whole highly inconsistent.3129 3130This clearly is a problem which needs to be addressed from cgroup core3131in a uniform way.3132 3133 3134Other Interface Issues3135----------------------3136 3137cgroup v1 grew without oversight and developed a large number of3138idiosyncrasies and inconsistencies.  One issue on the cgroup core side3139was how an empty cgroup was notified - a userland helper binary was3140forked and executed for each event.  The event delivery wasn't3141recursive or delegatable.  The limitations of the mechanism also led3142to in-kernel event delivery filtering mechanism further complicating3143the interface.3144 3145Controller interfaces were problematic too.  An extreme example is3146controllers completely ignoring hierarchical organization and treating3147all cgroups as if they were all located directly under the root3148cgroup.  Some controllers exposed a large amount of inconsistent3149implementation details to userland.3150 3151There also was no consistency across controllers.  When a new cgroup3152was created, some controllers defaulted to not imposing extra3153restrictions while others disallowed any resource usage until3154explicitly configured.  Configuration knobs for the same type of3155control used widely differing naming schemes and formats.  Statistics3156and information knobs were named arbitrarily and used different3157formats and units even in the same controller.3158 3159cgroup v2 establishes common conventions where appropriate and updates3160controllers so that they expose minimal and consistent interfaces.3161 3162 3163Controller Issues and Remedies3164------------------------------3165 3166Memory3167~~~~~~3168 3169The original lower boundary, the soft limit, is defined as a limit3170that is per default unset.  As a result, the set of cgroups that3171global reclaim prefers is opt-in, rather than opt-out.  The costs for3172optimizing these mostly negative lookups are so high that the3173implementation, despite its enormous size, does not even provide the3174basic desirable behavior.  First off, the soft limit has no3175hierarchical meaning.  All configured groups are organized in a global3176rbtree and treated like equal peers, regardless where they are located3177in the hierarchy.  This makes subtree delegation impossible.  Second,3178the soft limit reclaim pass is so aggressive that it not just3179introduces high allocation latencies into the system, but also impacts3180system performance due to overreclaim, to the point where the feature3181becomes self-defeating.3182 3183The memory.low boundary on the other hand is a top-down allocated3184reserve.  A cgroup enjoys reclaim protection when it's within its3185effective low, which makes delegation of subtrees possible. It also3186enjoys having reclaim pressure proportional to its overage when3187above its effective low.3188 3189The original high boundary, the hard limit, is defined as a strict3190limit that can not budge, even if the OOM killer has to be called.3191But this generally goes against the goal of making the most out of the3192available memory.  The memory consumption of workloads varies during3193runtime, and that requires users to overcommit.  But doing that with a3194strict upper limit requires either a fairly accurate prediction of the3195working set size or adding slack to the limit.  Since working set size3196estimation is hard and error prone, and getting it wrong results in3197OOM kills, most users tend to err on the side of a looser limit and3198end up wasting precious resources.3199 3200The memory.high boundary on the other hand can be set much more3201conservatively.  When hit, it throttles allocations by forcing them3202into direct reclaim to work off the excess, but it never invokes the3203OOM killer.  As a result, a high boundary that is chosen too3204aggressively will not terminate the processes, but instead it will3205lead to gradual performance degradation.  The user can monitor this3206and make corrections until the minimal memory footprint that still3207gives acceptable performance is found.3208 3209In extreme cases, with many concurrent allocations and a complete3210breakdown of reclaim progress within the group, the high boundary can3211be exceeded.  But even then it's mostly better to satisfy the3212allocation from the slack available in other groups or the rest of the3213system than killing the group.  Otherwise, memory.max is there to3214limit this type of spillover and ultimately contain buggy or even3215malicious applications.3216 3217Setting the original memory.limit_in_bytes below the current usage was3218subject to a race condition, where concurrent charges could cause the3219limit setting to fail. memory.max on the other hand will first set the3220limit to prevent new charges, and then reclaim and OOM kill until the3221new limit is met - or the task writing to memory.max is killed.3222 3223The combined memory+swap accounting and limiting is replaced by real3224control over swap space.3225 3226The main argument for a combined memory+swap facility in the original3227cgroup design was that global or parental pressure would always be3228able to swap all anonymous memory of a child group, regardless of the3229child's own (possibly untrusted) configuration.  However, untrusted3230groups can sabotage swapping by other means - such as referencing its3231anonymous memory in a tight loop - and an admin can not assume full3232swappability when overcommitting untrusted jobs.3233 3234For trusted jobs, on the other hand, a combined counter is not an3235intuitive userspace interface, and it flies in the face of the idea3236that cgroup controllers should account and limit specific physical3237resources.  Swap space is a resource like all others in the system,3238and that's why unified hierarchy allows distributing it separately.3239