xref: /linux/Documentation/admin-guide/cgroup-v2.rst (revision f898c16a0624e7f2dcb0b1cda6916c9be6489197)
1.. _cgroup-v2:
2
3================
4Control Group v2
5================
6
7:Date: October, 2015
8:Author: Tejun Heo <tj@kernel.org>
9
10This is the authoritative documentation on the design, interface and
11conventions of cgroup v2.  It describes all userland-visible aspects
12of cgroup including core and specific controller behaviors.  All
13future changes must be reflected in this document.  Documentation for
14v1 is available under :ref:`Documentation/admin-guide/cgroup-v1/index.rst <cgroup-v1>`.
15
16.. CONTENTS
17
18   1. Introduction
19     1-1. Terminology
20     1-2. What is cgroup?
21   2. Basic Operations
22     2-1. Mounting
23     2-2. Organizing Processes and Threads
24       2-2-1. Processes
25       2-2-2. Threads
26     2-3. [Un]populated Notification
27     2-4. Controlling Controllers
28       2-4-1. Enabling and Disabling
29       2-4-2. Top-down Constraint
30       2-4-3. No Internal Process Constraint
31     2-5. Delegation
32       2-5-1. Model of Delegation
33       2-5-2. Delegation Containment
34     2-6. Guidelines
35       2-6-1. Organize Once and Control
36       2-6-2. Avoid Name Collisions
37   3. Resource Distribution Models
38     3-1. Weights
39     3-2. Limits
40     3-3. Protections
41     3-4. Allocations
42   4. Interface Files
43     4-1. Format
44     4-2. Conventions
45     4-3. Core Interface Files
46   5. Controllers
47     5-1. CPU
48       5-1-1. CPU Interface Files
49     5-2. Memory
50       5-2-1. Memory Interface Files
51       5-2-2. Usage Guidelines
52       5-2-3. Memory Ownership
53     5-3. IO
54       5-3-1. IO Interface Files
55       5-3-2. Writeback
56       5-3-3. IO Latency
57         5-3-3-1. How IO Latency Throttling Works
58         5-3-3-2. IO Latency Interface Files
59       5-3-4. IO Priority
60     5-4. PID
61       5-4-1. PID Interface Files
62     5-5. Cpuset
63       5.5-1. Cpuset Interface Files
64     5-6. Device
65     5-7. RDMA
66       5-7-1. RDMA Interface Files
67     5-8. HugeTLB
68       5.8-1. HugeTLB Interface Files
69     5-9. Misc
70       5.9-1 Miscellaneous cgroup Interface Files
71       5.9-2 Migration and Ownership
72     5-10. Others
73       5-10-1. perf_event
74     5-N. Non-normative information
75       5-N-1. CPU controller root cgroup process behaviour
76       5-N-2. IO controller root cgroup process behaviour
77   6. Namespace
78     6-1. Basics
79     6-2. The Root and Views
80     6-3. Migration and setns(2)
81     6-4. Interaction with Other Namespaces
82   P. Information on Kernel Programming
83     P-1. Filesystem Support for Writeback
84   D. Deprecated v1 Core Features
85   R. Issues with v1 and Rationales for v2
86     R-1. Multiple Hierarchies
87     R-2. Thread Granularity
88     R-3. Competition Between Inner Nodes and Threads
89     R-4. Other Interface Issues
90     R-5. Controller Issues and Remedies
91       R-5-1. Memory
92
93
94Introduction
95============
96
97Terminology
98-----------
99
100"cgroup" stands for "control group" and is never capitalized.  The
101singular form is used to designate the whole feature and also as a
102qualifier as in "cgroup controllers".  When explicitly referring to
103multiple individual control groups, the plural form "cgroups" is used.
104
105
106What is cgroup?
107---------------
108
109cgroup is a mechanism to organize processes hierarchically and
110distribute system resources along the hierarchy in a controlled and
111configurable manner.
112
113cgroup is largely composed of two parts - the core and controllers.
114cgroup core is primarily responsible for hierarchically organizing
115processes.  A cgroup controller is usually responsible for
116distributing a specific type of system resource along the hierarchy
117although there are utility controllers which serve purposes other than
118resource distribution.
119
120cgroups form a tree structure and every process in the system belongs
121to one and only one cgroup.  All threads of a process belong to the
122same cgroup.  On creation, all processes are put in the cgroup that
123the parent process belongs to at the time.  A process can be migrated
124to another cgroup.  Migration of a process doesn't affect already
125existing descendant processes.
126
127Following certain structural constraints, controllers may be enabled or
128disabled selectively on a cgroup.  All controller behaviors are
129hierarchical - if a controller is enabled on a cgroup, it affects all
130processes which belong to the cgroups consisting the inclusive
131sub-hierarchy of the cgroup.  When a controller is enabled on a nested
132cgroup, it always restricts the resource distribution further.  The
133restrictions set closer to the root in the hierarchy can not be
134overridden from further away.
135
136
137Basic Operations
138================
139
140Mounting
141--------
142
143Unlike v1, cgroup v2 has only single hierarchy.  The cgroup v2
144hierarchy can be mounted with the following mount command::
145
146  # mount -t cgroup2 none $MOUNT_POINT
147
148cgroup2 filesystem has the magic number 0x63677270 ("cgrp").  All
149controllers which support v2 and are not bound to a v1 hierarchy are
150automatically bound to the v2 hierarchy and show up at the root.
151Controllers which are not in active use in the v2 hierarchy can be
152bound to other hierarchies.  This allows mixing v2 hierarchy with the
153legacy v1 multiple hierarchies in a fully backward compatible way.
154
155A controller can be moved across hierarchies only after the controller
156is no longer referenced in its current hierarchy.  Because per-cgroup
157controller states are destroyed asynchronously and controllers may
158have lingering references, a controller may not show up immediately on
159the v2 hierarchy after the final umount of the previous hierarchy.
160Similarly, a controller should be fully disabled to be moved out of
161the unified hierarchy and it may take some time for the disabled
162controller to become available for other hierarchies; furthermore, due
163to inter-controller dependencies, other controllers may need to be
164disabled too.
165
166While useful for development and manual configurations, moving
167controllers dynamically between the v2 and other hierarchies is
168strongly discouraged for production use.  It is recommended to decide
169the hierarchies and controller associations before starting using the
170controllers after system boot.
171
172During transition to v2, system management software might still
173automount the v1 cgroup filesystem and so hijack all controllers
174during boot, before manual intervention is possible. To make testing
175and experimenting easier, the kernel parameter cgroup_no_v1= allows
176disabling controllers in v1 and make them always available in v2.
177
178cgroup v2 currently supports the following mount options.
179
180  nsdelegate
181	Consider cgroup namespaces as delegation boundaries.  This
182	option is system wide and can only be set on mount or modified
183	through remount from the init namespace.  The mount option is
184	ignored on non-init namespace mounts.  Please refer to the
185	Delegation section for details.
186
187  favordynmods
188        Reduce the latencies of dynamic cgroup modifications such as
189        task migrations and controller on/offs at the cost of making
190        hot path operations such as forks and exits more expensive.
191        The static usage pattern of creating a cgroup, enabling
192        controllers, and then seeding it with CLONE_INTO_CGROUP is
193        not affected by this option.
194
195  memory_localevents
196        Only populate memory.events with data for the current cgroup,
197        and not any subtrees. This is legacy behaviour, the default
198        behaviour without this option is to include subtree counts.
199        This option is system wide and can only be set on mount or
200        modified through remount from the init namespace. The mount
201        option is ignored on non-init namespace mounts.
202
203  memory_recursiveprot
204        Recursively apply memory.min and memory.low protection to
205        entire subtrees, without requiring explicit downward
206        propagation into leaf cgroups.  This allows protecting entire
207        subtrees from one another, while retaining free competition
208        within those subtrees.  This should have been the default
209        behavior but is a mount-option to avoid regressing setups
210        relying on the original semantics (e.g. specifying bogusly
211        high 'bypass' protection values at higher tree levels).
212
213  memory_hugetlb_accounting
214        Count HugeTLB memory usage towards the cgroup's overall
215        memory usage for the memory controller (for the purpose of
216        statistics reporting and memory protetion). This is a new
217        behavior that could regress existing setups, so it must be
218        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 memory
223          controller. The pre-allocated pool does not belong to anyone.
224          Specifically, when a new HugeTLB folio is allocated to
225          the pool, it is not accounted for from the perspective of the
226          memory controller. It is only charged to a cgroup when it is
227          actually used (for e.g at page fault time). Host memory
228          overcommit management has to consider this when configuring
229          hard limits. In general, HugeTLB pool management should be
230          done via other mechanisms (such as the HugeTLB controller).
231        * Failure to charge a HugeTLB folio to the memory controller
232          results in SIGBUS. This could happen even if the HugeTLB pool
233          still has pages available (but the cgroup limit is hit and
234          reclaim attempt fails).
235        * Charging HugeTLB memory towards the memory controller affects
236          memory protection and reclaim dynamics. Any userspace tuning
237          (of low, min limits for e.g) needs to take this into account.
238        * HugeTLB pages utilized while this option is not selected
239          will not be tracked by the memory controller (even if cgroup
240          v2 is remounted later on).
241
242
243Organizing Processes and Threads
244--------------------------------
245
246Processes
247~~~~~~~~~
248
249Initially, only the root cgroup exists to which all processes belong.
250A child cgroup can be created by creating a sub-directory::
251
252  # mkdir $CGROUP_NAME
253
254A given cgroup may have multiple child cgroups forming a tree
255structure.  Each cgroup has a read-writable interface file
256"cgroup.procs".  When read, it lists the PIDs of all processes which
257belong to the cgroup one-per-line.  The PIDs are not ordered and the
258same PID may show up more than once if the process got moved to
259another cgroup and then back or the PID got recycled while reading.
260
261A process can be migrated into a cgroup by writing its PID to the
262target cgroup's "cgroup.procs" file.  Only one process can be migrated
263on a single write(2) call.  If a process is composed of multiple
264threads, writing the PID of any thread migrates all threads of the
265process.
266
267When a process forks a child process, the new process is born into the
268cgroup that the forking process belongs to at the time of the
269operation.  After exit, a process stays associated with the cgroup
270that it belonged to at the time of exit until it's reaped; however, a
271zombie process does not appear in "cgroup.procs" and thus can't be
272moved to another cgroup.
273
274A cgroup which doesn't have any children or live processes can be
275destroyed by removing the directory.  Note that a cgroup which doesn't
276have any children and is associated only with zombie processes is
277considered empty and can be removed::
278
279  # rmdir $CGROUP_NAME
280
281"/proc/$PID/cgroup" lists a process's cgroup membership.  If legacy
282cgroup is in use in the system, this file may contain multiple lines,
283one for each hierarchy.  The entry for cgroup v2 is always in the
284format "0::$PATH"::
285
286  # cat /proc/842/cgroup
287  ...
288  0::/test-cgroup/test-cgroup-nested
289
290If the process becomes a zombie and the cgroup it was associated with
291is removed subsequently, " (deleted)" is appended to the path::
292
293  # cat /proc/842/cgroup
294  ...
295  0::/test-cgroup/test-cgroup-nested (deleted)
296
297
298Threads
299~~~~~~~
300
301cgroup v2 supports thread granularity for a subset of controllers to
302support use cases requiring hierarchical resource distribution across
303the threads of a group of processes.  By default, all threads of a
304process belong to the same cgroup, which also serves as the resource
305domain to host resource consumptions which are not specific to a
306process or thread.  The thread mode allows threads to be spread across
307a subtree while still maintaining the common resource domain for them.
308
309Controllers which support thread mode are called threaded controllers.
310The ones which don't are called domain controllers.
311
312Marking a cgroup threaded makes it join the resource domain of its
313parent as a threaded cgroup.  The parent may be another threaded
314cgroup whose resource domain is further up in the hierarchy.  The root
315of a threaded subtree, that is, the nearest ancestor which is not
316threaded, is called threaded domain or thread root interchangeably and
317serves as the resource domain for the entire subtree.
318
319Inside a threaded subtree, threads of a process can be put in
320different cgroups and are not subject to the no internal process
321constraint - threaded controllers can be enabled on non-leaf cgroups
322whether they have threads in them or not.
323
324As the threaded domain cgroup hosts all the domain resource
325consumptions of the subtree, it is considered to have internal
326resource consumptions whether there are processes in it or not and
327can't have populated child cgroups which aren't threaded.  Because the
328root cgroup is not subject to no internal process constraint, it can
329serve both as a threaded domain and a parent to domain cgroups.
330
331The current operation mode or type of the cgroup is shown in the
332"cgroup.type" file which indicates whether the cgroup is a normal
333domain, a domain which is serving as the domain of a threaded subtree,
334or a threaded cgroup.
335
336On creation, a cgroup is always a domain cgroup and can be made
337threaded by writing "threaded" to the "cgroup.type" file.  The
338operation is single direction::
339
340  # echo threaded > cgroup.type
341
342Once threaded, the cgroup can't be made a domain again.  To enable the
343thread mode, the following conditions must be met.
344
345- As the cgroup will join the parent's resource domain.  The parent
346  must either be a valid (threaded) domain or a threaded cgroup.
347
348- When the parent is an unthreaded domain, it must not have any domain
349  controllers enabled or populated domain children.  The root is
350  exempt from this requirement.
351
352Topology-wise, a cgroup can be in an invalid state.  Please consider
353the following topology::
354
355  A (threaded domain) - B (threaded) - C (domain, just created)
356
357C is created as a domain but isn't connected to a parent which can
358host child domains.  C can't be used until it is turned into a
359threaded cgroup.  "cgroup.type" file will report "domain (invalid)" in
360these cases.  Operations which fail due to invalid topology use
361EOPNOTSUPP as the errno.
362
363A domain cgroup is turned into a threaded domain when one of its child
364cgroup becomes threaded or threaded controllers are enabled in the
365"cgroup.subtree_control" file while there are processes in the cgroup.
366A threaded domain reverts to a normal domain when the conditions
367clear.
368
369When read, "cgroup.threads" contains the list of the thread IDs of all
370threads in the cgroup.  Except that the operations are per-thread
371instead of per-process, "cgroup.threads" has the same format and
372behaves the same way as "cgroup.procs".  While "cgroup.threads" can be
373written to in any cgroup, as it can only move threads inside the same
374threaded domain, its operations are confined inside each threaded
375subtree.
376
377The threaded domain cgroup serves as the resource domain for the whole
378subtree, and, while the threads can be scattered across the subtree,
379all the processes are considered to be in the threaded domain cgroup.
380"cgroup.procs" in a threaded domain cgroup contains the PIDs of all
381processes in the subtree and is not readable in the subtree proper.
382However, "cgroup.procs" can be written to from anywhere in the subtree
383to migrate all threads of the matching process to the cgroup.
384
385Only threaded controllers can be enabled in a threaded subtree.  When
386a threaded controller is enabled inside a threaded subtree, it only
387accounts for and controls resource consumptions associated with the
388threads in the cgroup and its descendants.  All consumptions which
389aren't tied to a specific thread belong to the threaded domain cgroup.
390
391Because a threaded subtree is exempt from no internal process
392constraint, a threaded controller must be able to handle competition
393between threads in a non-leaf cgroup and its child cgroups.  Each
394threaded controller defines how such competitions are handled.
395
396Currently, the following controllers are threaded and can be enabled
397in a threaded cgroup::
398
399- cpu
400- cpuset
401- perf_event
402- pids
403
404[Un]populated Notification
405--------------------------
406
407Each non-root cgroup has a "cgroup.events" file which contains
408"populated" field indicating whether the cgroup's sub-hierarchy has
409live processes in it.  Its value is 0 if there is no live process in
410the cgroup and its descendants; otherwise, 1.  poll and [id]notify
411events are triggered when the value changes.  This can be used, for
412example, to start a clean-up operation after all processes of a given
413sub-hierarchy have exited.  The populated state updates and
414notifications are recursive.  Consider the following sub-hierarchy
415where the numbers in the parentheses represent the numbers of processes
416in each cgroup::
417
418  A(4) - B(0) - C(1)
419              \ D(0)
420
421A, B and C's "populated" fields would be 1 while D's 0.  After the one
422process in C exits, B and C's "populated" fields would flip to "0" and
423file modified events will be generated on the "cgroup.events" files of
424both cgroups.
425
426
427Controlling Controllers
428-----------------------
429
430Enabling and Disabling
431~~~~~~~~~~~~~~~~~~~~~~
432
433Each cgroup has a "cgroup.controllers" file which lists all
434controllers available for the cgroup to enable::
435
436  # cat cgroup.controllers
437  cpu io memory
438
439No controller is enabled by default.  Controllers can be enabled and
440disabled by writing to the "cgroup.subtree_control" file::
441
442  # echo "+cpu +memory -io" > cgroup.subtree_control
443
444Only controllers which are listed in "cgroup.controllers" can be
445enabled.  When multiple operations are specified as above, either they
446all succeed or fail.  If multiple operations on the same controller
447are specified, the last one is effective.
448
449Enabling a controller in a cgroup indicates that the distribution of
450the target resource across its immediate children will be controlled.
451Consider the following sub-hierarchy.  The enabled controllers are
452listed in parentheses::
453
454  A(cpu,memory) - B(memory) - C()
455                            \ D()
456
457As A has "cpu" and "memory" enabled, A will control the distribution
458of CPU cycles and memory to its children, in this case, B.  As B has
459"memory" enabled but not "CPU", C and D will compete freely on CPU
460cycles but their division of memory available to B will be controlled.
461
462As a controller regulates the distribution of the target resource to
463the cgroup's children, enabling it creates the controller's interface
464files in the child cgroups.  In the above example, enabling "cpu" on B
465would create the "cpu." prefixed controller interface files in C and
466D.  Likewise, disabling "memory" from B would remove the "memory."
467prefixed controller interface files from C and D.  This means that the
468controller interface files - anything which doesn't start with
469"cgroup." are owned by the parent rather than the cgroup itself.
470
471
472Top-down Constraint
473~~~~~~~~~~~~~~~~~~~
474
475Resources are distributed top-down and a cgroup can further distribute
476a resource only if the resource has been distributed to it from the
477parent.  This means that all non-root "cgroup.subtree_control" files
478can only contain controllers which are enabled in the parent's
479"cgroup.subtree_control" file.  A controller can be enabled only if
480the parent has the controller enabled and a controller can't be
481disabled if one or more children have it enabled.
482
483
484No Internal Process Constraint
485~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
486
487Non-root cgroups can distribute domain resources to their children
488only when they don't have any processes of their own.  In other words,
489only domain cgroups which don't contain any processes can have domain
490controllers enabled in their "cgroup.subtree_control" files.
491
492This guarantees that, when a domain controller is looking at the part
493of the hierarchy which has it enabled, processes are always only on
494the leaves.  This rules out situations where child cgroups compete
495against internal processes of the parent.
496
497The root cgroup is exempt from this restriction.  Root contains
498processes and anonymous resource consumption which can't be associated
499with any other cgroups and requires special treatment from most
500controllers.  How resource consumption in the root cgroup is governed
501is up to each controller (for more information on this topic please
502refer to the Non-normative information section in the Controllers
503chapter).
504
505Note that the restriction doesn't get in the way if there is no
506enabled controller in the cgroup's "cgroup.subtree_control".  This is
507important as otherwise it wouldn't be possible to create children of a
508populated cgroup.  To control resource distribution of a cgroup, the
509cgroup must create children and transfer all its processes to the
510children before enabling controllers in its "cgroup.subtree_control"
511file.
512
513
514Delegation
515----------
516
517Model of Delegation
518~~~~~~~~~~~~~~~~~~~
519
520A cgroup can be delegated in two ways.  First, to a less privileged
521user by granting write access of the directory and its "cgroup.procs",
522"cgroup.threads" and "cgroup.subtree_control" files to the user.
523Second, if the "nsdelegate" mount option is set, automatically to a
524cgroup namespace on namespace creation.
525
526Because the resource control interface files in a given directory
527control the distribution of the parent's resources, the delegatee
528shouldn't be allowed to write to them.  For the first method, this is
529achieved by not granting access to these files.  For the second, the
530kernel rejects writes to all files other than "cgroup.procs" and
531"cgroup.subtree_control" on a namespace root from inside the
532namespace.
533
534The end results are equivalent for both delegation types.  Once
535delegated, the user can build sub-hierarchy under the directory,
536organize processes inside it as it sees fit and further distribute the
537resources it received from the parent.  The limits and other settings
538of all resource controllers are hierarchical and regardless of what
539happens in the delegated sub-hierarchy, nothing can escape the
540resource restrictions imposed by the parent.
541
542Currently, cgroup doesn't impose any restrictions on the number of
543cgroups in or nesting depth of a delegated sub-hierarchy; however,
544this may be limited explicitly in the future.
545
546
547Delegation Containment
548~~~~~~~~~~~~~~~~~~~~~~
549
550A delegated sub-hierarchy is contained in the sense that processes
551can't be moved into or out of the sub-hierarchy by the delegatee.
552
553For delegations to a less privileged user, this is achieved by
554requiring the following conditions for a process with a non-root euid
555to migrate a target process into a cgroup by writing its PID to the
556"cgroup.procs" file.
557
558- The writer must have write access to the "cgroup.procs" file.
559
560- The writer must have write access to the "cgroup.procs" file of the
561  common ancestor of the source and destination cgroups.
562
563The above two constraints ensure that while a delegatee may migrate
564processes around freely in the delegated sub-hierarchy it can't pull
565in from or push out to outside the sub-hierarchy.
566
567For an example, let's assume cgroups C0 and C1 have been delegated to
568user U0 who created C00, C01 under C0 and C10 under C1 as follows and
569all processes under C0 and C1 belong to U0::
570
571  ~~~~~~~~~~~~~ - C0 - C00
572  ~ cgroup    ~      \ C01
573  ~ hierarchy ~
574  ~~~~~~~~~~~~~ - C1 - C10
575
576Let's also say U0 wants to write the PID of a process which is
577currently in C10 into "C00/cgroup.procs".  U0 has write access to the
578file; however, the common ancestor of the source cgroup C10 and the
579destination cgroup C00 is above the points of delegation and U0 would
580not have write access to its "cgroup.procs" files and thus the write
581will be denied with -EACCES.
582
583For delegations to namespaces, containment is achieved by requiring
584that both the source and destination cgroups are reachable from the
585namespace of the process which is attempting the migration.  If either
586is not reachable, the migration is rejected with -ENOENT.
587
588
589Guidelines
590----------
591
592Organize Once and Control
593~~~~~~~~~~~~~~~~~~~~~~~~~
594
595Migrating a process across cgroups is a relatively expensive operation
596and stateful resources such as memory are not moved together with the
597process.  This is an explicit design decision as there often exist
598inherent trade-offs between migration and various hot paths in terms
599of synchronization cost.
600
601As such, migrating processes across cgroups frequently as a means to
602apply different resource restrictions is discouraged.  A workload
603should be assigned to a cgroup according to the system's logical and
604resource structure once on start-up.  Dynamic adjustments to resource
605distribution can be made by changing controller configuration through
606the interface files.
607
608
609Avoid Name Collisions
610~~~~~~~~~~~~~~~~~~~~~
611
612Interface files for a cgroup and its children cgroups occupy the same
613directory and it is possible to create children cgroups which collide
614with interface files.
615
616All cgroup core interface files are prefixed with "cgroup." and each
617controller's interface files are prefixed with the controller name and
618a dot.  A controller's name is composed of lower case alphabets and
619'_'s but never begins with an '_' so it can be used as the prefix
620character for collision avoidance.  Also, interface file names won't
621start or end with terms which are often used in categorizing workloads
622such as job, service, slice, unit or workload.
623
624cgroup doesn't do anything to prevent name collisions and it's the
625user's responsibility to avoid them.
626
627
628Resource Distribution Models
629============================
630
631cgroup controllers implement several resource distribution schemes
632depending on the resource type and expected use cases.  This section
633describes major schemes in use along with their expected behaviors.
634
635
636Weights
637-------
638
639A parent's resource is distributed by adding up the weights of all
640active children and giving each the fraction matching the ratio of its
641weight against the sum.  As only children which can make use of the
642resource at the moment participate in the distribution, this is
643work-conserving.  Due to the dynamic nature, this model is usually
644used for stateless resources.
645
646All weights are in the range [1, 10000] with the default at 100.  This
647allows symmetric multiplicative biases in both directions at fine
648enough granularity while staying in the intuitive range.
649
650As long as the weight is in range, all configuration combinations are
651valid and there is no reason to reject configuration changes or
652process migrations.
653
654"cpu.weight" proportionally distributes CPU cycles to active children
655and is an example of this type.
656
657
658.. _cgroupv2-limits-distributor:
659
660Limits
661------
662
663A child can only consume up to the configured amount of the resource.
664Limits can be over-committed - the sum of the limits of children can
665exceed the amount of resource available to the parent.
666
667Limits are in the range [0, max] and defaults to "max", which is noop.
668
669As limits can be over-committed, all configuration combinations are
670valid and there is no reason to reject configuration changes or
671process migrations.
672
673"io.max" limits the maximum BPS and/or IOPS that a cgroup can consume
674on an IO device and is an example of this type.
675
676.. _cgroupv2-protections-distributor:
677
678Protections
679-----------
680
681A cgroup is protected up to the configured amount of the resource
682as long as the usages of all its ancestors are under their
683protected levels.  Protections can be hard guarantees or best effort
684soft boundaries.  Protections can also be over-committed in which case
685only up to the amount available to the parent is protected among
686children.
687
688Protections are in the range [0, max] and defaults to 0, which is
689noop.
690
691As protections can be over-committed, all configuration combinations
692are valid and there is no reason to reject configuration changes or
693process migrations.
694
695"memory.low" implements best-effort memory protection and is an
696example of this type.
697
698
699Allocations
700-----------
701
702A cgroup is exclusively allocated a certain amount of a finite
703resource.  Allocations can't be over-committed - the sum of the
704allocations of children can not exceed the amount of resource
705available to the parent.
706
707Allocations are in the range [0, max] and defaults to 0, which is no
708resource.
709
710As allocations can't be over-committed, some configuration
711combinations are invalid and should be rejected.  Also, if the
712resource is mandatory for execution of processes, process migrations
713may be rejected.
714
715"cpu.rt.max" hard-allocates realtime slices and is an example of this
716type.
717
718
719Interface Files
720===============
721
722Format
723------
724
725All interface files should be in one of the following formats whenever
726possible::
727
728  New-line separated values
729  (when only one value can be written at once)
730
731	VAL0\n
732	VAL1\n
733	...
734
735  Space separated values
736  (when read-only or multiple values can be written at once)
737
738	VAL0 VAL1 ...\n
739
740  Flat keyed
741
742	KEY0 VAL0\n
743	KEY1 VAL1\n
744	...
745
746  Nested keyed
747
748	KEY0 SUB_KEY0=VAL00 SUB_KEY1=VAL01...
749	KEY1 SUB_KEY0=VAL10 SUB_KEY1=VAL11...
750	...
751
752For a writable file, the format for writing should generally match
753reading; however, controllers may allow omitting later fields or
754implement restricted shortcuts for most common use cases.
755
756For both flat and nested keyed files, only the values for a single key
757can be written at a time.  For nested keyed files, the sub key pairs
758may be specified in any order and not all pairs have to be specified.
759
760
761Conventions
762-----------
763
764- Settings for a single feature should be contained in a single file.
765
766- The root cgroup should be exempt from resource control and thus
767  shouldn't have resource control interface files.
768
769- The default time unit is microseconds.  If a different unit is ever
770  used, an explicit unit suffix must be present.
771
772- A parts-per quantity should use a percentage decimal with at least
773  two digit fractional part - e.g. 13.40.
774
775- If a controller implements weight based resource distribution, its
776  interface file should be named "weight" and have the range [1,
777  10000] with 100 as the default.  The values are chosen to allow
778  enough and symmetric bias in both directions while keeping it
779  intuitive (the default is 100%).
780
781- If a controller implements an absolute resource guarantee and/or
782  limit, the interface files should be named "min" and "max"
783  respectively.  If a controller implements best effort resource
784  guarantee and/or limit, the interface files should be named "low"
785  and "high" respectively.
786
787  In the above four control files, the special token "max" should be
788  used to represent upward infinity for both reading and writing.
789
790- If a setting has a configurable default value and keyed specific
791  overrides, the default entry should be keyed with "default" and
792  appear as the first entry in the file.
793
794  The default value can be updated by writing either "default $VAL" or
795  "$VAL".
796
797  When writing to update a specific override, "default" can be used as
798  the value to indicate removal of the override.  Override entries
799  with "default" as the value must not appear when read.
800
801  For example, a setting which is keyed by major:minor device numbers
802  with integer values may look like the following::
803
804    # cat cgroup-example-interface-file
805    default 150
806    8:0 300
807
808  The default value can be updated by::
809
810    # echo 125 > cgroup-example-interface-file
811
812  or::
813
814    # echo "default 125" > cgroup-example-interface-file
815
816  An override can be set by::
817
818    # echo "8:16 170" > cgroup-example-interface-file
819
820  and cleared by::
821
822    # echo "8:0 default" > cgroup-example-interface-file
823    # cat cgroup-example-interface-file
824    default 125
825    8:16 170
826
827- For events which are not very high frequency, an interface file
828  "events" should be created which lists event key value pairs.
829  Whenever a notifiable event happens, file modified event should be
830  generated on the file.
831
832
833Core Interface Files
834--------------------
835
836All cgroup core files are prefixed with "cgroup."
837
838  cgroup.type
839	A read-write single value file which exists on non-root
840	cgroups.
841
842	When read, it indicates the current type of the cgroup, which
843	can be one of the following values.
844
845	- "domain" : A normal valid domain cgroup.
846
847	- "domain threaded" : A threaded domain cgroup which is
848          serving as the root of a threaded subtree.
849
850	- "domain invalid" : A cgroup which is in an invalid state.
851	  It can't be populated or have controllers enabled.  It may
852	  be allowed to become a threaded cgroup.
853
854	- "threaded" : A threaded cgroup which is a member of a
855          threaded subtree.
856
857	A cgroup can be turned into a threaded cgroup by writing
858	"threaded" to this file.
859
860  cgroup.procs
861	A read-write new-line separated values file which exists on
862	all cgroups.
863
864	When read, it lists the PIDs of all processes which belong to
865	the cgroup one-per-line.  The PIDs are not ordered and the
866	same PID may show up more than once if the process got moved
867	to another cgroup and then back or the PID got recycled while
868	reading.
869
870	A PID can be written to migrate the process associated with
871	the PID to the cgroup.  The writer should match all of the
872	following conditions.
873
874	- It must have write access to the "cgroup.procs" file.
875
876	- It must have write access to the "cgroup.procs" file of the
877	  common ancestor of the source and destination cgroups.
878
879	When delegating a sub-hierarchy, write access to this file
880	should be granted along with the containing directory.
881
882	In a threaded cgroup, reading this file fails with EOPNOTSUPP
883	as all the processes belong to the thread root.  Writing is
884	supported and moves every thread of the process to the cgroup.
885
886  cgroup.threads
887	A read-write new-line separated values file which exists on
888	all cgroups.
889
890	When read, it lists the TIDs of all threads which belong to
891	the cgroup one-per-line.  The TIDs are not ordered and the
892	same TID may show up more than once if the thread got moved to
893	another cgroup and then back or the TID got recycled while
894	reading.
895
896	A TID can be written to migrate the thread associated with the
897	TID to the cgroup.  The writer should match all of the
898	following conditions.
899
900	- It must have write access to the "cgroup.threads" file.
901
902	- The cgroup that the thread is currently in must be in the
903          same resource domain as the destination cgroup.
904
905	- It must have write access to the "cgroup.procs" file of the
906	  common ancestor of the source and destination cgroups.
907
908	When delegating a sub-hierarchy, write access to this file
909	should be granted along with the containing directory.
910
911  cgroup.controllers
912	A read-only space separated values file which exists on all
913	cgroups.
914
915	It shows space separated list of all controllers available to
916	the cgroup.  The controllers are not ordered.
917
918  cgroup.subtree_control
919	A read-write space separated values file which exists on all
920	cgroups.  Starts out empty.
921
922	When read, it shows space separated list of the controllers
923	which are enabled to control resource distribution from the
924	cgroup to its children.
925
926	Space separated list of controllers prefixed with '+' or '-'
927	can be written to enable or disable controllers.  A controller
928	name prefixed with '+' enables the controller and '-'
929	disables.  If a controller appears more than once on the list,
930	the last one is effective.  When multiple enable and disable
931	operations are specified, either all succeed or all fail.
932
933  cgroup.events
934	A read-only flat-keyed file which exists on non-root cgroups.
935	The following entries are defined.  Unless specified
936	otherwise, a value change in this file generates a file
937	modified event.
938
939	  populated
940		1 if the cgroup or its descendants contains any live
941		processes; otherwise, 0.
942	  frozen
943		1 if the cgroup is frozen; otherwise, 0.
944
945  cgroup.max.descendants
946	A read-write single value files.  The default is "max".
947
948	Maximum allowed number of descent cgroups.
949	If the actual number of descendants is equal or larger,
950	an attempt to create a new cgroup in the hierarchy will fail.
951
952  cgroup.max.depth
953	A read-write single value files.  The default is "max".
954
955	Maximum allowed descent depth below the current cgroup.
956	If the actual descent depth is equal or larger,
957	an attempt to create a new child cgroup will fail.
958
959  cgroup.stat
960	A read-only flat-keyed file with the following entries:
961
962	  nr_descendants
963		Total number of visible descendant cgroups.
964
965	  nr_dying_descendants
966		Total number of dying descendant cgroups. A cgroup becomes
967		dying after being deleted by a user. The cgroup will remain
968		in dying state for some time undefined time (which can depend
969		on system load) before being completely destroyed.
970
971		A process can't enter a dying cgroup under any circumstances,
972		a dying cgroup can't revive.
973
974		A dying cgroup can consume system resources not exceeding
975		limits, which were active at the moment of cgroup deletion.
976
977  cgroup.freeze
978	A read-write single value file which exists on non-root cgroups.
979	Allowed values are "0" and "1". The default is "0".
980
981	Writing "1" to the file causes freezing of the cgroup and all
982	descendant cgroups. This means that all belonging processes will
983	be stopped and will not run until the cgroup will be explicitly
984	unfrozen. Freezing of the cgroup may take some time; when this action
985	is completed, the "frozen" value in the cgroup.events control file
986	will be updated to "1" and the corresponding notification will be
987	issued.
988
989	A cgroup can be frozen either by its own settings, or by settings
990	of any ancestor cgroups. If any of ancestor cgroups is frozen, the
991	cgroup will remain frozen.
992
993	Processes in the frozen cgroup can be killed by a fatal signal.
994	They also can enter and leave a frozen cgroup: either by an explicit
995	move by a user, or if freezing of the cgroup races with fork().
996	If a process is moved to a frozen cgroup, it stops. If a process is
997	moved out of a frozen cgroup, it becomes running.
998
999	Frozen status of a cgroup doesn't affect any cgroup tree operations:
1000	it's possible to delete a frozen (and empty) cgroup, as well as
1001	create new sub-cgroups.
1002
1003  cgroup.kill
1004	A write-only single value file which exists in non-root cgroups.
1005	The only allowed value is "1".
1006
1007	Writing "1" to the file causes the cgroup and all descendant cgroups to
1008	be killed. This means that all processes located in the affected cgroup
1009	tree will be killed via SIGKILL.
1010
1011	Killing a cgroup tree will deal with concurrent forks appropriately and
1012	is protected against migrations.
1013
1014	In a threaded cgroup, writing this file fails with EOPNOTSUPP as
1015	killing cgroups is a process directed operation, i.e. it affects
1016	the whole thread-group.
1017
1018  cgroup.pressure
1019	A read-write single value file that allowed values are "0" and "1".
1020	The default is "1".
1021
1022	Writing "0" to the file will disable the cgroup PSI accounting.
1023	Writing "1" to the file will re-enable the cgroup PSI accounting.
1024
1025	This control attribute is not hierarchical, so disable or enable PSI
1026	accounting in a cgroup does not affect PSI accounting in descendants
1027	and doesn't need pass enablement via ancestors from root.
1028
1029	The reason this control attribute exists is that PSI accounts stalls for
1030	each cgroup separately and aggregates it at each level of the hierarchy.
1031	This may cause non-negligible overhead for some workloads when under
1032	deep level of the hierarchy, in which case this control attribute can
1033	be used to disable PSI accounting in the non-leaf cgroups.
1034
1035  irq.pressure
1036	A read-write nested-keyed file.
1037
1038	Shows pressure stall information for IRQ/SOFTIRQ. See
1039	:ref:`Documentation/accounting/psi.rst <psi>` for details.
1040
1041Controllers
1042===========
1043
1044.. _cgroup-v2-cpu:
1045
1046CPU
1047---
1048
1049The "cpu" controllers regulates distribution of CPU cycles.  This
1050controller implements weight and absolute bandwidth limit models for
1051normal scheduling policy and absolute bandwidth allocation model for
1052realtime scheduling policy.
1053
1054In all the above models, cycles distribution is defined only on a temporal
1055base and it does not account for the frequency at which tasks are executed.
1056The (optional) utilization clamping support allows to hint the schedutil
1057cpufreq governor about the minimum desired frequency which should always be
1058provided by a CPU, as well as the maximum desired frequency, which should not
1059be exceeded by a CPU.
1060
1061WARNING: cgroup2 doesn't yet support control of realtime processes. For
1062a kernel built with the CONFIG_RT_GROUP_SCHED option enabled for group
1063scheduling of realtime processes, the cpu controller can only be enabled
1064when all RT processes are in the root cgroup.  This limitation does
1065not apply if CONFIG_RT_GROUP_SCHED is disabled.  Be aware that system
1066management software may already have placed RT processes into nonroot
1067cgroups during the system boot process, and these processes may need
1068to be moved to the root cgroup before the cpu controller can be enabled
1069with a CONFIG_RT_GROUP_SCHED enabled kernel.
1070
1071
1072CPU Interface Files
1073~~~~~~~~~~~~~~~~~~~
1074
1075All time durations are in microseconds.
1076
1077  cpu.stat
1078	A read-only flat-keyed file.
1079	This file exists whether the controller is enabled or not.
1080
1081	It always reports the following three stats:
1082
1083	- usage_usec
1084	- user_usec
1085	- system_usec
1086
1087	and the following five when the controller is enabled:
1088
1089	- nr_periods
1090	- nr_throttled
1091	- throttled_usec
1092	- nr_bursts
1093	- burst_usec
1094
1095  cpu.weight
1096	A read-write single value file which exists on non-root
1097	cgroups.  The default is "100".
1098
1099	For non idle groups (cpu.idle = 0), the weight is in the
1100	range [1, 10000].
1101
1102	If the cgroup has been configured to be SCHED_IDLE (cpu.idle = 1),
1103	then the weight will show as a 0.
1104
1105  cpu.weight.nice
1106	A read-write single value file which exists on non-root
1107	cgroups.  The default is "0".
1108
1109	The nice value is in the range [-20, 19].
1110
1111	This interface file is an alternative interface for
1112	"cpu.weight" and allows reading and setting weight using the
1113	same values used by nice(2).  Because the range is smaller and
1114	granularity is coarser for the nice values, the read value is
1115	the closest approximation of the current weight.
1116
1117  cpu.max
1118	A read-write two value file which exists on non-root cgroups.
1119	The default is "max 100000".
1120
1121	The maximum bandwidth limit.  It's in the following format::
1122
1123	  $MAX $PERIOD
1124
1125	which indicates that the group may consume up to $MAX in each
1126	$PERIOD duration.  "max" for $MAX indicates no limit.  If only
1127	one number is written, $MAX is updated.
1128
1129  cpu.max.burst
1130	A read-write single value file which exists on non-root
1131	cgroups.  The default is "0".
1132
1133	The burst in the range [0, $MAX].
1134
1135  cpu.pressure
1136	A read-write nested-keyed file.
1137
1138	Shows pressure stall information for CPU. See
1139	:ref:`Documentation/accounting/psi.rst <psi>` for details.
1140
1141  cpu.uclamp.min
1142        A read-write single value file which exists on non-root cgroups.
1143        The default is "0", i.e. no utilization boosting.
1144
1145        The requested minimum utilization (protection) as a percentage
1146        rational number, e.g. 12.34 for 12.34%.
1147
1148        This interface allows reading and setting minimum utilization clamp
1149        values similar to the sched_setattr(2). This minimum utilization
1150        value is used to clamp the task specific minimum utilization clamp.
1151
1152        The requested minimum utilization (protection) is always capped by
1153        the current value for the maximum utilization (limit), i.e.
1154        `cpu.uclamp.max`.
1155
1156  cpu.uclamp.max
1157        A read-write single value file which exists on non-root cgroups.
1158        The default is "max". i.e. no utilization capping
1159
1160        The requested maximum utilization (limit) as a percentage rational
1161        number, e.g. 98.76 for 98.76%.
1162
1163        This interface allows reading and setting maximum utilization clamp
1164        values similar to the sched_setattr(2). This maximum utilization
1165        value is used to clamp the task specific maximum utilization clamp.
1166
1167  cpu.idle
1168	A read-write single value file which exists on non-root cgroups.
1169	The default is 0.
1170
1171	This is the cgroup analog of the per-task SCHED_IDLE sched policy.
1172	Setting this value to a 1 will make the scheduling policy of the
1173	cgroup SCHED_IDLE. The threads inside the cgroup will retain their
1174	own relative priorities, but the cgroup itself will be treated as
1175	very low priority relative to its peers.
1176
1177
1178
1179Memory
1180------
1181
1182The "memory" controller regulates distribution of memory.  Memory is
1183stateful and implements both limit and protection models.  Due to the
1184intertwining between memory usage and reclaim pressure and the
1185stateful nature of memory, the distribution model is relatively
1186complex.
1187
1188While not completely water-tight, all major memory usages by a given
1189cgroup are tracked so that the total memory consumption can be
1190accounted and controlled to a reasonable extent.  Currently, the
1191following types of memory usages are tracked.
1192
1193- Userland memory - page cache and anonymous memory.
1194
1195- Kernel data structures such as dentries and inodes.
1196
1197- TCP socket buffers.
1198
1199The above list may expand in the future for better coverage.
1200
1201
1202Memory Interface Files
1203~~~~~~~~~~~~~~~~~~~~~~
1204
1205All memory amounts are in bytes.  If a value which is not aligned to
1206PAGE_SIZE is written, the value may be rounded up to the closest
1207PAGE_SIZE multiple when read back.
1208
1209  memory.current
1210	A read-only single value file which exists on non-root
1211	cgroups.
1212
1213	The total amount of memory currently being used by the cgroup
1214	and its descendants.
1215
1216  memory.min
1217	A read-write single value file which exists on non-root
1218	cgroups.  The default is "0".
1219
1220	Hard memory protection.  If the memory usage of a cgroup
1221	is within its effective min boundary, the cgroup's memory
1222	won't be reclaimed under any conditions. If there is no
1223	unprotected reclaimable memory available, OOM killer
1224	is invoked. Above the effective min boundary (or
1225	effective low boundary if it is higher), pages are reclaimed
1226	proportionally to the overage, reducing reclaim pressure for
1227	smaller overages.
1228
1229	Effective min boundary is limited by memory.min values of
1230	all ancestor cgroups. If there is memory.min overcommitment
1231	(child cgroup or cgroups are requiring more protected memory
1232	than parent will allow), then each child cgroup will get
1233	the part of parent's protection proportional to its
1234	actual memory usage below memory.min.
1235
1236	Putting more memory than generally available under this
1237	protection is discouraged and may lead to constant OOMs.
1238
1239	If a memory cgroup is not populated with processes,
1240	its memory.min is ignored.
1241
1242  memory.low
1243	A read-write single value file which exists on non-root
1244	cgroups.  The default is "0".
1245
1246	Best-effort memory protection.  If the memory usage of a
1247	cgroup is within its effective low boundary, the cgroup's
1248	memory won't be reclaimed unless there is no reclaimable
1249	memory available in unprotected cgroups.
1250	Above the effective low	boundary (or
1251	effective min boundary if it is higher), pages are reclaimed
1252	proportionally to the overage, reducing reclaim pressure for
1253	smaller overages.
1254
1255	Effective low boundary is limited by memory.low values of
1256	all ancestor cgroups. If there is memory.low overcommitment
1257	(child cgroup or cgroups are requiring more protected memory
1258	than parent will allow), then each child cgroup will get
1259	the part of parent's protection proportional to its
1260	actual memory usage below memory.low.
1261
1262	Putting more memory than generally available under this
1263	protection is discouraged.
1264
1265  memory.high
1266	A read-write single value file which exists on non-root
1267	cgroups.  The default is "max".
1268
1269	Memory usage throttle limit.  If a cgroup's usage goes
1270	over the high boundary, the processes of the cgroup are
1271	throttled and put under heavy reclaim pressure.
1272
1273	Going over the high limit never invokes the OOM killer and
1274	under extreme conditions the limit may be breached. The high
1275	limit should be used in scenarios where an external process
1276	monitors the limited cgroup to alleviate heavy reclaim
1277	pressure.
1278
1279  memory.max
1280	A read-write single value file which exists on non-root
1281	cgroups.  The default is "max".
1282
1283	Memory usage hard limit.  This is the main mechanism to limit
1284	memory usage of a cgroup.  If a cgroup's memory usage reaches
1285	this limit and can't be reduced, the OOM killer is invoked in
1286	the cgroup. Under certain circumstances, the usage may go
1287	over the limit temporarily.
1288
1289	In default configuration regular 0-order allocations always
1290	succeed unless OOM killer chooses current task as a victim.
1291
1292	Some kinds of allocations don't invoke the OOM killer.
1293	Caller could retry them differently, return into userspace
1294	as -ENOMEM or silently ignore in cases like disk readahead.
1295
1296  memory.reclaim
1297	A write-only nested-keyed file which exists for all cgroups.
1298
1299	This is a simple interface to trigger memory reclaim in the
1300	target cgroup.
1301
1302	This file accepts a single key, the number of bytes to reclaim.
1303	No nested keys are currently supported.
1304
1305	Example::
1306
1307	  echo "1G" > memory.reclaim
1308
1309	The interface can be later extended with nested keys to
1310	configure the reclaim behavior. For example, specify the
1311	type of memory to reclaim from (anon, file, ..).
1312
1313	Please note that the kernel can over or under reclaim from
1314	the target cgroup. If less bytes are reclaimed than the
1315	specified amount, -EAGAIN is returned.
1316
1317	Please note that the proactive reclaim (triggered by this
1318	interface) is not meant to indicate memory pressure on the
1319	memory cgroup. Therefore socket memory balancing triggered by
1320	the memory reclaim normally is not exercised in this case.
1321	This means that the networking layer will not adapt based on
1322	reclaim induced by memory.reclaim.
1323
1324  memory.peak
1325	A read-only single value file which exists on non-root
1326	cgroups.
1327
1328	The max memory usage recorded for the cgroup and its
1329	descendants since the creation of the cgroup.
1330
1331  memory.oom.group
1332	A read-write single value file which exists on non-root
1333	cgroups.  The default value is "0".
1334
1335	Determines whether the cgroup should be treated as
1336	an indivisible workload by the OOM killer. If set,
1337	all tasks belonging to the cgroup or to its descendants
1338	(if the memory cgroup is not a leaf cgroup) are killed
1339	together or not at all. This can be used to avoid
1340	partial kills to guarantee workload integrity.
1341
1342	Tasks with the OOM protection (oom_score_adj set to -1000)
1343	are treated as an exception and are never killed.
1344
1345	If the OOM killer is invoked in a cgroup, it's not going
1346	to kill any tasks outside of this cgroup, regardless
1347	memory.oom.group values of ancestor cgroups.
1348
1349  memory.events
1350	A read-only flat-keyed file which exists on non-root cgroups.
1351	The following entries are defined.  Unless specified
1352	otherwise, a value change in this file generates a file
1353	modified event.
1354
1355	Note that all fields in this file are hierarchical and the
1356	file modified event can be generated due to an event down the
1357	hierarchy. For the local events at the cgroup level see
1358	memory.events.local.
1359
1360	  low
1361		The number of times the cgroup is reclaimed due to
1362		high memory pressure even though its usage is under
1363		the low boundary.  This usually indicates that the low
1364		boundary is over-committed.
1365
1366	  high
1367		The number of times processes of the cgroup are
1368		throttled and routed to perform direct memory reclaim
1369		because the high memory boundary was exceeded.  For a
1370		cgroup whose memory usage is capped by the high limit
1371		rather than global memory pressure, this event's
1372		occurrences are expected.
1373
1374	  max
1375		The number of times the cgroup's memory usage was
1376		about to go over the max boundary.  If direct reclaim
1377		fails to bring it down, the cgroup goes to OOM state.
1378
1379	  oom
1380		The number of time the cgroup's memory usage was
1381		reached the limit and allocation was about to fail.
1382
1383		This event is not raised if the OOM killer is not
1384		considered as an option, e.g. for failed high-order
1385		allocations or if caller asked to not retry attempts.
1386
1387	  oom_kill
1388		The number of processes belonging to this cgroup
1389		killed by any kind of OOM killer.
1390
1391          oom_group_kill
1392                The number of times a group OOM has occurred.
1393
1394  memory.events.local
1395	Similar to memory.events but the fields in the file are local
1396	to the cgroup i.e. not hierarchical. The file modified event
1397	generated on this file reflects only the local events.
1398
1399  memory.stat
1400	A read-only flat-keyed file which exists on non-root cgroups.
1401
1402	This breaks down the cgroup's memory footprint into different
1403	types of memory, type-specific details, and other information
1404	on the state and past events of the memory management system.
1405
1406	All memory amounts are in bytes.
1407
1408	The entries are ordered to be human readable, and new entries
1409	can show up in the middle. Don't rely on items remaining in a
1410	fixed position; use the keys to look up specific values!
1411
1412	If the entry has no per-node counter (or not show in the
1413	memory.numa_stat). We use 'npn' (non-per-node) as the tag
1414	to indicate that it will not show in the memory.numa_stat.
1415
1416	  anon
1417		Amount of memory used in anonymous mappings such as
1418		brk(), sbrk(), and mmap(MAP_ANONYMOUS)
1419
1420	  file
1421		Amount of memory used to cache filesystem data,
1422		including tmpfs and shared memory.
1423
1424	  kernel (npn)
1425		Amount of total kernel memory, including
1426		(kernel_stack, pagetables, percpu, vmalloc, slab) in
1427		addition to other kernel memory use cases.
1428
1429	  kernel_stack
1430		Amount of memory allocated to kernel stacks.
1431
1432	  pagetables
1433                Amount of memory allocated for page tables.
1434
1435	  sec_pagetables
1436		Amount of memory allocated for secondary page tables,
1437		this currently includes KVM mmu allocations on x86
1438		and arm64 and IOMMU page tables.
1439
1440	  percpu (npn)
1441		Amount of memory used for storing per-cpu kernel
1442		data structures.
1443
1444	  sock (npn)
1445		Amount of memory used in network transmission buffers
1446
1447	  vmalloc (npn)
1448		Amount of memory used for vmap backed memory.
1449
1450	  shmem
1451		Amount of cached filesystem data that is swap-backed,
1452		such as tmpfs, shm segments, shared anonymous mmap()s
1453
1454	  zswap
1455		Amount of memory consumed by the zswap compression backend.
1456
1457	  zswapped
1458		Amount of application memory swapped out to zswap.
1459
1460	  file_mapped
1461		Amount of cached filesystem data mapped with mmap()
1462
1463	  file_dirty
1464		Amount of cached filesystem data that was modified but
1465		not yet written back to disk
1466
1467	  file_writeback
1468		Amount of cached filesystem data that was modified and
1469		is currently being written back to disk
1470
1471	  swapcached
1472		Amount of swap cached in memory. The swapcache is accounted
1473		against both memory and swap usage.
1474
1475	  anon_thp
1476		Amount of memory used in anonymous mappings backed by
1477		transparent hugepages
1478
1479	  file_thp
1480		Amount of cached filesystem data backed by transparent
1481		hugepages
1482
1483	  shmem_thp
1484		Amount of shm, tmpfs, shared anonymous mmap()s backed by
1485		transparent hugepages
1486
1487	  inactive_anon, active_anon, inactive_file, active_file, unevictable
1488		Amount of memory, swap-backed and filesystem-backed,
1489		on the internal memory management lists used by the
1490		page reclaim algorithm.
1491
1492		As these represent internal list state (eg. shmem pages are on anon
1493		memory management lists), inactive_foo + active_foo may not be equal to
1494		the value for the foo counter, since the foo counter is type-based, not
1495		list-based.
1496
1497	  slab_reclaimable
1498		Part of "slab" that might be reclaimed, such as
1499		dentries and inodes.
1500
1501	  slab_unreclaimable
1502		Part of "slab" that cannot be reclaimed on memory
1503		pressure.
1504
1505	  slab (npn)
1506		Amount of memory used for storing in-kernel data
1507		structures.
1508
1509	  workingset_refault_anon
1510		Number of refaults of previously evicted anonymous pages.
1511
1512	  workingset_refault_file
1513		Number of refaults of previously evicted file pages.
1514
1515	  workingset_activate_anon
1516		Number of refaulted anonymous pages that were immediately
1517		activated.
1518
1519	  workingset_activate_file
1520		Number of refaulted file pages that were immediately activated.
1521
1522	  workingset_restore_anon
1523		Number of restored anonymous pages which have been detected as
1524		an active workingset before they got reclaimed.
1525
1526	  workingset_restore_file
1527		Number of restored file pages which have been detected as an
1528		active workingset before they got reclaimed.
1529
1530	  workingset_nodereclaim
1531		Number of times a shadow node has been reclaimed
1532
1533	  pgscan (npn)
1534		Amount of scanned pages (in an inactive LRU list)
1535
1536	  pgsteal (npn)
1537		Amount of reclaimed pages
1538
1539	  pgscan_kswapd (npn)
1540		Amount of scanned pages by kswapd (in an inactive LRU list)
1541
1542	  pgscan_direct (npn)
1543		Amount of scanned pages directly  (in an inactive LRU list)
1544
1545	  pgscan_khugepaged (npn)
1546		Amount of scanned pages by khugepaged  (in an inactive LRU list)
1547
1548	  pgsteal_kswapd (npn)
1549		Amount of reclaimed pages by kswapd
1550
1551	  pgsteal_direct (npn)
1552		Amount of reclaimed pages directly
1553
1554	  pgsteal_khugepaged (npn)
1555		Amount of reclaimed pages by khugepaged
1556
1557	  pgfault (npn)
1558		Total number of page faults incurred
1559
1560	  pgmajfault (npn)
1561		Number of major page faults incurred
1562
1563	  pgrefill (npn)
1564		Amount of scanned pages (in an active LRU list)
1565
1566	  pgactivate (npn)
1567		Amount of pages moved to the active LRU list
1568
1569	  pgdeactivate (npn)
1570		Amount of pages moved to the inactive LRU list
1571
1572	  pglazyfree (npn)
1573		Amount of pages postponed to be freed under memory pressure
1574
1575	  pglazyfreed (npn)
1576		Amount of reclaimed lazyfree pages
1577
1578	  zswpin
1579		Number of pages moved in to memory from zswap.
1580
1581	  zswpout
1582		Number of pages moved out of memory to zswap.
1583
1584	  zswpwb
1585		Number of pages written from zswap to swap.
1586
1587	  thp_fault_alloc (npn)
1588		Number of transparent hugepages which were allocated to satisfy
1589		a page fault. This counter is not present when CONFIG_TRANSPARENT_HUGEPAGE
1590                is not set.
1591
1592	  thp_collapse_alloc (npn)
1593		Number of transparent hugepages which were allocated to allow
1594		collapsing an existing range of pages. This counter is not
1595		present when CONFIG_TRANSPARENT_HUGEPAGE is not set.
1596
1597	  thp_swpout (npn)
1598		Number of transparent hugepages which are swapout in one piece
1599		without splitting.
1600
1601	  thp_swpout_fallback (npn)
1602		Number of transparent hugepages which were split before swapout.
1603		Usually because failed to allocate some continuous swap space
1604		for the huge page.
1605
1606  memory.numa_stat
1607	A read-only nested-keyed file which exists on non-root cgroups.
1608
1609	This breaks down the cgroup's memory footprint into different
1610	types of memory, type-specific details, and other information
1611	per node on the state of the memory management system.
1612
1613	This is useful for providing visibility into the NUMA locality
1614	information within an memcg since the pages are allowed to be
1615	allocated from any physical node. One of the use case is evaluating
1616	application performance by combining this information with the
1617	application's CPU allocation.
1618
1619	All memory amounts are in bytes.
1620
1621	The output format of memory.numa_stat is::
1622
1623	  type N0=<bytes in node 0> N1=<bytes in node 1> ...
1624
1625	The entries are ordered to be human readable, and new entries
1626	can show up in the middle. Don't rely on items remaining in a
1627	fixed position; use the keys to look up specific values!
1628
1629	The entries can refer to the memory.stat.
1630
1631  memory.swap.current
1632	A read-only single value file which exists on non-root
1633	cgroups.
1634
1635	The total amount of swap currently being used by the cgroup
1636	and its descendants.
1637
1638  memory.swap.high
1639	A read-write single value file which exists on non-root
1640	cgroups.  The default is "max".
1641
1642	Swap usage throttle limit.  If a cgroup's swap usage exceeds
1643	this limit, all its further allocations will be throttled to
1644	allow userspace to implement custom out-of-memory procedures.
1645
1646	This limit marks a point of no return for the cgroup. It is NOT
1647	designed to manage the amount of swapping a workload does
1648	during regular operation. Compare to memory.swap.max, which
1649	prohibits swapping past a set amount, but lets the cgroup
1650	continue unimpeded as long as other memory can be reclaimed.
1651
1652	Healthy workloads are not expected to reach this limit.
1653
1654  memory.swap.peak
1655	A read-only single value file which exists on non-root
1656	cgroups.
1657
1658	The max swap usage recorded for the cgroup and its
1659	descendants since the creation of the cgroup.
1660
1661  memory.swap.max
1662	A read-write single value file which exists on non-root
1663	cgroups.  The default is "max".
1664
1665	Swap usage hard limit.  If a cgroup's swap usage reaches this
1666	limit, anonymous memory of the cgroup will not be swapped out.
1667
1668  memory.swap.events
1669	A read-only flat-keyed file which exists on non-root cgroups.
1670	The following entries are defined.  Unless specified
1671	otherwise, a value change in this file generates a file
1672	modified event.
1673
1674	  high
1675		The number of times the cgroup's swap usage was over
1676		the high threshold.
1677
1678	  max
1679		The number of times the cgroup's swap usage was about
1680		to go over the max boundary and swap allocation
1681		failed.
1682
1683	  fail
1684		The number of times swap allocation failed either
1685		because of running out of swap system-wide or max
1686		limit.
1687
1688	When reduced under the current usage, the existing swap
1689	entries are reclaimed gradually and the swap usage may stay
1690	higher than the limit for an extended period of time.  This
1691	reduces the impact on the workload and memory management.
1692
1693  memory.zswap.current
1694	A read-only single value file which exists on non-root
1695	cgroups.
1696
1697	The total amount of memory consumed by the zswap compression
1698	backend.
1699
1700  memory.zswap.max
1701	A read-write single value file which exists on non-root
1702	cgroups.  The default is "max".
1703
1704	Zswap usage hard limit. If a cgroup's zswap pool reaches this
1705	limit, it will refuse to take any more stores before existing
1706	entries fault back in or are written out to disk.
1707
1708  memory.zswap.writeback
1709	A read-write single value file. The default value is "1". The
1710	initial value of the root cgroup is 1, and when a new cgroup is
1711	created, it inherits the current value of its parent.
1712
1713	When this is set to 0, all swapping attempts to swapping devices
1714	are disabled. This included both zswap writebacks, and swapping due
1715	to zswap store failures. If the zswap store failures are recurring
1716	(for e.g if the pages are incompressible), users can observe
1717	reclaim inefficiency after disabling writeback (because the same
1718	pages might be rejected again and again).
1719
1720	Note that this is subtly different from setting memory.swap.max to
1721	0, as it still allows for pages to be written to the zswap pool.
1722
1723  memory.pressure
1724	A read-only nested-keyed file.
1725
1726	Shows pressure stall information for memory. See
1727	:ref:`Documentation/accounting/psi.rst <psi>` for details.
1728
1729
1730Usage Guidelines
1731~~~~~~~~~~~~~~~~
1732
1733"memory.high" is the main mechanism to control memory usage.
1734Over-committing on high limit (sum of high limits > available memory)
1735and letting global memory pressure to distribute memory according to
1736usage is a viable strategy.
1737
1738Because breach of the high limit doesn't trigger the OOM killer but
1739throttles the offending cgroup, a management agent has ample
1740opportunities to monitor and take appropriate actions such as granting
1741more memory or terminating the workload.
1742
1743Determining whether a cgroup has enough memory is not trivial as
1744memory usage doesn't indicate whether the workload can benefit from
1745more memory.  For example, a workload which writes data received from
1746network to a file can use all available memory but can also operate as
1747performant with a small amount of memory.  A measure of memory
1748pressure - how much the workload is being impacted due to lack of
1749memory - is necessary to determine whether a workload needs more
1750memory; unfortunately, memory pressure monitoring mechanism isn't
1751implemented yet.
1752
1753
1754Memory Ownership
1755~~~~~~~~~~~~~~~~
1756
1757A memory area is charged to the cgroup which instantiated it and stays
1758charged to the cgroup until the area is released.  Migrating a process
1759to a different cgroup doesn't move the memory usages that it
1760instantiated while in the previous cgroup to the new cgroup.
1761
1762A memory area may be used by processes belonging to different cgroups.
1763To which cgroup the area will be charged is in-deterministic; however,
1764over time, the memory area is likely to end up in a cgroup which has
1765enough memory allowance to avoid high reclaim pressure.
1766
1767If a cgroup sweeps a considerable amount of memory which is expected
1768to be accessed repeatedly by other cgroups, it may make sense to use
1769POSIX_FADV_DONTNEED to relinquish the ownership of memory areas
1770belonging to the affected files to ensure correct memory ownership.
1771
1772
1773IO
1774--
1775
1776The "io" controller regulates the distribution of IO resources.  This
1777controller implements both weight based and absolute bandwidth or IOPS
1778limit distribution; however, weight based distribution is available
1779only if cfq-iosched is in use and neither scheme is available for
1780blk-mq devices.
1781
1782
1783IO Interface Files
1784~~~~~~~~~~~~~~~~~~
1785
1786  io.stat
1787	A read-only nested-keyed file.
1788
1789	Lines are keyed by $MAJ:$MIN device numbers and not ordered.
1790	The following nested keys are defined.
1791
1792	  ======	=====================
1793	  rbytes	Bytes read
1794	  wbytes	Bytes written
1795	  rios		Number of read IOs
1796	  wios		Number of write IOs
1797	  dbytes	Bytes discarded
1798	  dios		Number of discard IOs
1799	  ======	=====================
1800
1801	An example read output follows::
1802
1803	  8:16 rbytes=1459200 wbytes=314773504 rios=192 wios=353 dbytes=0 dios=0
1804	  8:0 rbytes=90430464 wbytes=299008000 rios=8950 wios=1252 dbytes=50331648 dios=3021
1805
1806  io.cost.qos
1807	A read-write nested-keyed file which exists only on the root
1808	cgroup.
1809
1810	This file configures the Quality of Service of the IO cost
1811	model based controller (CONFIG_BLK_CGROUP_IOCOST) which
1812	currently implements "io.weight" proportional control.  Lines
1813	are keyed by $MAJ:$MIN device numbers and not ordered.  The
1814	line for a given device is populated on the first write for
1815	the device on "io.cost.qos" or "io.cost.model".  The following
1816	nested keys are defined.
1817
1818	  ======	=====================================
1819	  enable	Weight-based control enable
1820	  ctrl		"auto" or "user"
1821	  rpct		Read latency percentile    [0, 100]
1822	  rlat		Read latency threshold
1823	  wpct		Write latency percentile   [0, 100]
1824	  wlat		Write latency threshold
1825	  min		Minimum scaling percentage [1, 10000]
1826	  max		Maximum scaling percentage [1, 10000]
1827	  ======	=====================================
1828
1829	The controller is disabled by default and can be enabled by
1830	setting "enable" to 1.  "rpct" and "wpct" parameters default
1831	to zero and the controller uses internal device saturation
1832	state to adjust the overall IO rate between "min" and "max".
1833
1834	When a better control quality is needed, latency QoS
1835	parameters can be configured.  For example::
1836
1837	  8:16 enable=1 ctrl=auto rpct=95.00 rlat=75000 wpct=95.00 wlat=150000 min=50.00 max=150.0
1838
1839	shows that on sdb, the controller is enabled, will consider
1840	the device saturated if the 95th percentile of read completion
1841	latencies is above 75ms or write 150ms, and adjust the overall
1842	IO issue rate between 50% and 150% accordingly.
1843
1844	The lower the saturation point, the better the latency QoS at
1845	the cost of aggregate bandwidth.  The narrower the allowed
1846	adjustment range between "min" and "max", the more conformant
1847	to the cost model the IO behavior.  Note that the IO issue
1848	base rate may be far off from 100% and setting "min" and "max"
1849	blindly can lead to a significant loss of device capacity or
1850	control quality.  "min" and "max" are useful for regulating
1851	devices which show wide temporary behavior changes - e.g. a
1852	ssd which accepts writes at the line speed for a while and
1853	then completely stalls for multiple seconds.
1854
1855	When "ctrl" is "auto", the parameters are controlled by the
1856	kernel and may change automatically.  Setting "ctrl" to "user"
1857	or setting any of the percentile and latency parameters puts
1858	it into "user" mode and disables the automatic changes.  The
1859	automatic mode can be restored by setting "ctrl" to "auto".
1860
1861  io.cost.model
1862	A read-write nested-keyed file which exists only on the root
1863	cgroup.
1864
1865	This file configures the cost model of the IO cost model based
1866	controller (CONFIG_BLK_CGROUP_IOCOST) which currently
1867	implements "io.weight" proportional control.  Lines are keyed
1868	by $MAJ:$MIN device numbers and not ordered.  The line for a
1869	given device is populated on the first write for the device on
1870	"io.cost.qos" or "io.cost.model".  The following nested keys
1871	are defined.
1872
1873	  =====		================================
1874	  ctrl		"auto" or "user"
1875	  model		The cost model in use - "linear"
1876	  =====		================================
1877
1878	When "ctrl" is "auto", the kernel may change all parameters
1879	dynamically.  When "ctrl" is set to "user" or any other
1880	parameters are written to, "ctrl" become "user" and the
1881	automatic changes are disabled.
1882
1883	When "model" is "linear", the following model parameters are
1884	defined.
1885
1886	  =============	========================================
1887	  [r|w]bps	The maximum sequential IO throughput
1888	  [r|w]seqiops	The maximum 4k sequential IOs per second
1889	  [r|w]randiops	The maximum 4k random IOs per second
1890	  =============	========================================
1891
1892	From the above, the builtin linear model determines the base
1893	costs of a sequential and random IO and the cost coefficient
1894	for the IO size.  While simple, this model can cover most
1895	common device classes acceptably.
1896
1897	The IO cost model isn't expected to be accurate in absolute
1898	sense and is scaled to the device behavior dynamically.
1899
1900	If needed, tools/cgroup/iocost_coef_gen.py can be used to
1901	generate device-specific coefficients.
1902
1903  io.weight
1904	A read-write flat-keyed file which exists on non-root cgroups.
1905	The default is "default 100".
1906
1907	The first line is the default weight applied to devices
1908	without specific override.  The rest are overrides keyed by
1909	$MAJ:$MIN device numbers and not ordered.  The weights are in
1910	the range [1, 10000] and specifies the relative amount IO time
1911	the cgroup can use in relation to its siblings.
1912
1913	The default weight can be updated by writing either "default
1914	$WEIGHT" or simply "$WEIGHT".  Overrides can be set by writing
1915	"$MAJ:$MIN $WEIGHT" and unset by writing "$MAJ:$MIN default".
1916
1917	An example read output follows::
1918
1919	  default 100
1920	  8:16 200
1921	  8:0 50
1922
1923  io.max
1924	A read-write nested-keyed file which exists on non-root
1925	cgroups.
1926
1927	BPS and IOPS based IO limit.  Lines are keyed by $MAJ:$MIN
1928	device numbers and not ordered.  The following nested keys are
1929	defined.
1930
1931	  =====		==================================
1932	  rbps		Max read bytes per second
1933	  wbps		Max write bytes per second
1934	  riops		Max read IO operations per second
1935	  wiops		Max write IO operations per second
1936	  =====		==================================
1937
1938	When writing, any number of nested key-value pairs can be
1939	specified in any order.  "max" can be specified as the value
1940	to remove a specific limit.  If the same key is specified
1941	multiple times, the outcome is undefined.
1942
1943	BPS and IOPS are measured in each IO direction and IOs are
1944	delayed if limit is reached.  Temporary bursts are allowed.
1945
1946	Setting read limit at 2M BPS and write at 120 IOPS for 8:16::
1947
1948	  echo "8:16 rbps=2097152 wiops=120" > io.max
1949
1950	Reading returns the following::
1951
1952	  8:16 rbps=2097152 wbps=max riops=max wiops=120
1953
1954	Write IOPS limit can be removed by writing the following::
1955
1956	  echo "8:16 wiops=max" > io.max
1957
1958	Reading now returns the following::
1959
1960	  8:16 rbps=2097152 wbps=max riops=max wiops=max
1961
1962  io.pressure
1963	A read-only nested-keyed file.
1964
1965	Shows pressure stall information for IO. See
1966	:ref:`Documentation/accounting/psi.rst <psi>` for details.
1967
1968
1969Writeback
1970~~~~~~~~~
1971
1972Page cache is dirtied through buffered writes and shared mmaps and
1973written asynchronously to the backing filesystem by the writeback
1974mechanism.  Writeback sits between the memory and IO domains and
1975regulates the proportion of dirty memory by balancing dirtying and
1976write IOs.
1977
1978The io controller, in conjunction with the memory controller,
1979implements control of page cache writeback IOs.  The memory controller
1980defines the memory domain that dirty memory ratio is calculated and
1981maintained for and the io controller defines the io domain which
1982writes out dirty pages for the memory domain.  Both system-wide and
1983per-cgroup dirty memory states are examined and the more restrictive
1984of the two is enforced.
1985
1986cgroup writeback requires explicit support from the underlying
1987filesystem.  Currently, cgroup writeback is implemented on ext2, ext4,
1988btrfs, f2fs, and xfs.  On other filesystems, all writeback IOs are
1989attributed to the root cgroup.
1990
1991There are inherent differences in memory and writeback management
1992which affects how cgroup ownership is tracked.  Memory is tracked per
1993page while writeback per inode.  For the purpose of writeback, an
1994inode is assigned to a cgroup and all IO requests to write dirty pages
1995from the inode are attributed to that cgroup.
1996
1997As cgroup ownership for memory is tracked per page, there can be pages
1998which are associated with different cgroups than the one the inode is
1999associated with.  These are called foreign pages.  The writeback
2000constantly keeps track of foreign pages and, if a particular foreign
2001cgroup becomes the majority over a certain period of time, switches
2002the ownership of the inode to that cgroup.
2003
2004While this model is enough for most use cases where a given inode is
2005mostly dirtied by a single cgroup even when the main writing cgroup
2006changes over time, use cases where multiple cgroups write to a single
2007inode simultaneously are not supported well.  In such circumstances, a
2008significant portion of IOs are likely to be attributed incorrectly.
2009As memory controller assigns page ownership on the first use and
2010doesn't update it until the page is released, even if writeback
2011strictly follows page ownership, multiple cgroups dirtying overlapping
2012areas wouldn't work as expected.  It's recommended to avoid such usage
2013patterns.
2014
2015The sysctl knobs which affect writeback behavior are applied to cgroup
2016writeback as follows.
2017
2018  vm.dirty_background_ratio, vm.dirty_ratio
2019	These ratios apply the same to cgroup writeback with the
2020	amount of available memory capped by limits imposed by the
2021	memory controller and system-wide clean memory.
2022
2023  vm.dirty_background_bytes, vm.dirty_bytes
2024	For cgroup writeback, this is calculated into ratio against
2025	total available memory and applied the same way as
2026	vm.dirty[_background]_ratio.
2027
2028
2029IO Latency
2030~~~~~~~~~~
2031
2032This is a cgroup v2 controller for IO workload protection.  You provide a group
2033with a latency target, and if the average latency exceeds that target the
2034controller will throttle any peers that have a lower latency target than the
2035protected workload.
2036
2037The limits are only applied at the peer level in the hierarchy.  This means that
2038in the diagram below, only groups A, B, and C will influence each other, and
2039groups D and F will influence each other.  Group G will influence nobody::
2040
2041			[root]
2042		/	   |		\
2043		A	   B		C
2044	       /  \        |
2045	      D    F	   G
2046
2047
2048So the ideal way to configure this is to set io.latency in groups A, B, and C.
2049Generally you do not want to set a value lower than the latency your device
2050supports.  Experiment to find the value that works best for your workload.
2051Start at higher than the expected latency for your device and watch the
2052avg_lat value in io.stat for your workload group to get an idea of the
2053latency you see during normal operation.  Use the avg_lat value as a basis for
2054your real setting, setting at 10-15% higher than the value in io.stat.
2055
2056How IO Latency Throttling Works
2057~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
2058
2059io.latency is work conserving; so as long as everybody is meeting their latency
2060target the controller doesn't do anything.  Once a group starts missing its
2061target it begins throttling any peer group that has a higher target than itself.
2062This throttling takes 2 forms:
2063
2064- Queue depth throttling.  This is the number of outstanding IO's a group is
2065  allowed to have.  We will clamp down relatively quickly, starting at no limit
2066  and going all the way down to 1 IO at a time.
2067
2068- Artificial delay induction.  There are certain types of IO that cannot be
2069  throttled without possibly adversely affecting higher priority groups.  This
2070  includes swapping and metadata IO.  These types of IO are allowed to occur
2071  normally, however they are "charged" to the originating group.  If the
2072  originating group is being throttled you will see the use_delay and delay
2073  fields in io.stat increase.  The delay value is how many microseconds that are
2074  being added to any process that runs in this group.  Because this number can
2075  grow quite large if there is a lot of swapping or metadata IO occurring we
2076  limit the individual delay events to 1 second at a time.
2077
2078Once the victimized group starts meeting its latency target again it will start
2079unthrottling any peer groups that were throttled previously.  If the victimized
2080group simply stops doing IO the global counter will unthrottle appropriately.
2081
2082IO Latency Interface Files
2083~~~~~~~~~~~~~~~~~~~~~~~~~~
2084
2085  io.latency
2086	This takes a similar format as the other controllers.
2087
2088		"MAJOR:MINOR target=<target time in microseconds>"
2089
2090  io.stat
2091	If the controller is enabled you will see extra stats in io.stat in
2092	addition to the normal ones.
2093
2094	  depth
2095		This is the current queue depth for the group.
2096
2097	  avg_lat
2098		This is an exponential moving average with a decay rate of 1/exp
2099		bound by the sampling interval.  The decay rate interval can be
2100		calculated by multiplying the win value in io.stat by the
2101		corresponding number of samples based on the win value.
2102
2103	  win
2104		The sampling window size in milliseconds.  This is the minimum
2105		duration of time between evaluation events.  Windows only elapse
2106		with IO activity.  Idle periods extend the most recent window.
2107
2108IO Priority
2109~~~~~~~~~~~
2110
2111A single attribute controls the behavior of the I/O priority cgroup policy,
2112namely the io.prio.class attribute. The following values are accepted for
2113that attribute:
2114
2115  no-change
2116	Do not modify the I/O priority class.
2117
2118  promote-to-rt
2119	For requests that have a non-RT I/O priority class, change it into RT.
2120	Also change the priority level of these requests to 4. Do not modify
2121	the I/O priority of requests that have priority class RT.
2122
2123  restrict-to-be
2124	For requests that do not have an I/O priority class or that have I/O
2125	priority class RT, change it into BE. Also change the priority level
2126	of these requests to 0. Do not modify the I/O priority class of
2127	requests that have priority class IDLE.
2128
2129  idle
2130	Change the I/O priority class of all requests into IDLE, the lowest
2131	I/O priority class.
2132
2133  none-to-rt
2134	Deprecated. Just an alias for promote-to-rt.
2135
2136The following numerical values are associated with the I/O priority policies:
2137
2138+----------------+---+
2139| no-change      | 0 |
2140+----------------+---+
2141| promote-to-rt  | 1 |
2142+----------------+---+
2143| restrict-to-be | 2 |
2144+----------------+---+
2145| idle           | 3 |
2146+----------------+---+
2147
2148The numerical value that corresponds to each I/O priority class is as follows:
2149
2150+-------------------------------+---+
2151| IOPRIO_CLASS_NONE             | 0 |
2152+-------------------------------+---+
2153| IOPRIO_CLASS_RT (real-time)   | 1 |
2154+-------------------------------+---+
2155| IOPRIO_CLASS_BE (best effort) | 2 |
2156+-------------------------------+---+
2157| IOPRIO_CLASS_IDLE             | 3 |
2158+-------------------------------+---+
2159
2160The algorithm to set the I/O priority class for a request is as follows:
2161
2162- If I/O priority class policy is promote-to-rt, change the request I/O
2163  priority class to IOPRIO_CLASS_RT and change the request I/O priority
2164  level to 4.
2165- If I/O priority class policy is not promote-to-rt, translate the I/O priority
2166  class policy into a number, then change the request I/O priority class
2167  into the maximum of the I/O priority class policy number and the numerical
2168  I/O priority class.
2169
2170PID
2171---
2172
2173The process number controller is used to allow a cgroup to stop any
2174new tasks from being fork()'d or clone()'d after a specified limit is
2175reached.
2176
2177The number of tasks in a cgroup can be exhausted in ways which other
2178controllers cannot prevent, thus warranting its own controller.  For
2179example, a fork bomb is likely to exhaust the number of tasks before
2180hitting memory restrictions.
2181
2182Note that PIDs used in this controller refer to TIDs, process IDs as
2183used by the kernel.
2184
2185
2186PID Interface Files
2187~~~~~~~~~~~~~~~~~~~
2188
2189  pids.max
2190	A read-write single value file which exists on non-root
2191	cgroups.  The default is "max".
2192
2193	Hard limit of number of processes.
2194
2195  pids.current
2196	A read-only single value file which exists on non-root cgroups.
2197
2198	The number of processes currently in the cgroup and its
2199	descendants.
2200
2201  pids.peak
2202	A read-only single value file which exists on non-root cgroups.
2203
2204	The maximum value that the number of processes in the cgroup and its
2205	descendants has ever reached.
2206
2207  pids.events
2208	A read-only flat-keyed file which exists on non-root cgroups. The
2209	following entries are defined. Unless specified otherwise, a value
2210	change in this file generates a file modified event.
2211
2212	  max
2213		Number of times fork failed because limit was hit.
2214
2215Organisational operations are not blocked by cgroup policies, so it is
2216possible to have pids.current > pids.max.  This can be done by either
2217setting the limit to be smaller than pids.current, or attaching enough
2218processes to the cgroup such that pids.current is larger than
2219pids.max.  However, it is not possible to violate a cgroup PID policy
2220through fork() or clone(). These will return -EAGAIN if the creation
2221of a new process would cause a cgroup policy to be violated.
2222
2223
2224Cpuset
2225------
2226
2227The "cpuset" controller provides a mechanism for constraining
2228the CPU and memory node placement of tasks to only the resources
2229specified in the cpuset interface files in a task's current cgroup.
2230This is especially valuable on large NUMA systems where placing jobs
2231on properly sized subsets of the systems with careful processor and
2232memory placement to reduce cross-node memory access and contention
2233can improve overall system performance.
2234
2235The "cpuset" controller is hierarchical.  That means the controller
2236cannot use CPUs or memory nodes not allowed in its parent.
2237
2238
2239Cpuset Interface Files
2240~~~~~~~~~~~~~~~~~~~~~~
2241
2242  cpuset.cpus
2243	A read-write multiple values file which exists on non-root
2244	cpuset-enabled cgroups.
2245
2246	It lists the requested CPUs to be used by tasks within this
2247	cgroup.  The actual list of CPUs to be granted, however, is
2248	subjected to constraints imposed by its parent and can differ
2249	from the requested CPUs.
2250
2251	The CPU numbers are comma-separated numbers or ranges.
2252	For example::
2253
2254	  # cat cpuset.cpus
2255	  0-4,6,8-10
2256
2257	An empty value indicates that the cgroup is using the same
2258	setting as the nearest cgroup ancestor with a non-empty
2259	"cpuset.cpus" or all the available CPUs if none is found.
2260
2261	The value of "cpuset.cpus" stays constant until the next update
2262	and won't be affected by any CPU hotplug events.
2263
2264  cpuset.cpus.effective
2265	A read-only multiple values file which exists on all
2266	cpuset-enabled cgroups.
2267
2268	It lists the onlined CPUs that are actually granted to this
2269	cgroup by its parent.  These CPUs are allowed to be used by
2270	tasks within the current cgroup.
2271
2272	If "cpuset.cpus" is empty, the "cpuset.cpus.effective" file shows
2273	all the CPUs from the parent cgroup that can be available to
2274	be used by this cgroup.  Otherwise, it should be a subset of
2275	"cpuset.cpus" unless none of the CPUs listed in "cpuset.cpus"
2276	can be granted.  In this case, it will be treated just like an
2277	empty "cpuset.cpus".
2278
2279	Its value will be affected by CPU hotplug events.
2280
2281  cpuset.mems
2282	A read-write multiple values file which exists on non-root
2283	cpuset-enabled cgroups.
2284
2285	It lists the requested memory nodes to be used by tasks within
2286	this cgroup.  The actual list of memory nodes granted, however,
2287	is subjected to constraints imposed by its parent and can differ
2288	from the requested memory nodes.
2289
2290	The memory node numbers are comma-separated numbers or ranges.
2291	For example::
2292
2293	  # cat cpuset.mems
2294	  0-1,3
2295
2296	An empty value indicates that the cgroup is using the same
2297	setting as the nearest cgroup ancestor with a non-empty
2298	"cpuset.mems" or all the available memory nodes if none
2299	is found.
2300
2301	The value of "cpuset.mems" stays constant until the next update
2302	and won't be affected by any memory nodes hotplug events.
2303
2304	Setting a non-empty value to "cpuset.mems" causes memory of
2305	tasks within the cgroup to be migrated to the designated nodes if
2306	they are currently using memory outside of the designated nodes.
2307
2308	There is a cost for this memory migration.  The migration
2309	may not be complete and some memory pages may be left behind.
2310	So it is recommended that "cpuset.mems" should be set properly
2311	before spawning new tasks into the cpuset.  Even if there is
2312	a need to change "cpuset.mems" with active tasks, it shouldn't
2313	be done frequently.
2314
2315  cpuset.mems.effective
2316	A read-only multiple values file which exists on all
2317	cpuset-enabled cgroups.
2318
2319	It lists the onlined memory nodes that are actually granted to
2320	this cgroup by its parent. These memory nodes are allowed to
2321	be used by tasks within the current cgroup.
2322
2323	If "cpuset.mems" is empty, it shows all the memory nodes from the
2324	parent cgroup that will be available to be used by this cgroup.
2325	Otherwise, it should be a subset of "cpuset.mems" unless none of
2326	the memory nodes listed in "cpuset.mems" can be granted.  In this
2327	case, it will be treated just like an empty "cpuset.mems".
2328
2329	Its value will be affected by memory nodes hotplug events.
2330
2331  cpuset.cpus.exclusive
2332	A read-write multiple values file which exists on non-root
2333	cpuset-enabled cgroups.
2334
2335	It lists all the exclusive CPUs that are allowed to be used
2336	to create a new cpuset partition.  Its value is not used
2337	unless the cgroup becomes a valid partition root.  See the
2338	"cpuset.cpus.partition" section below for a description of what
2339	a cpuset partition is.
2340
2341	When the cgroup becomes a partition root, the actual exclusive
2342	CPUs that are allocated to that partition are listed in
2343	"cpuset.cpus.exclusive.effective" which may be different
2344	from "cpuset.cpus.exclusive".  If "cpuset.cpus.exclusive"
2345	has previously been set, "cpuset.cpus.exclusive.effective"
2346	is always a subset of it.
2347
2348	Users can manually set it to a value that is different from
2349	"cpuset.cpus".	The only constraint in setting it is that the
2350	list of CPUs must be exclusive with respect to its sibling.
2351
2352	For a parent cgroup, any one of its exclusive CPUs can only
2353	be distributed to at most one of its child cgroups.  Having an
2354	exclusive CPU appearing in two or more of its child cgroups is
2355	not allowed (the exclusivity rule).  A value that violates the
2356	exclusivity rule will be rejected with a write error.
2357
2358	The root cgroup is a partition root and all its available CPUs
2359	are in its exclusive CPU set.
2360
2361  cpuset.cpus.exclusive.effective
2362	A read-only multiple values file which exists on all non-root
2363	cpuset-enabled cgroups.
2364
2365	This file shows the effective set of exclusive CPUs that
2366	can be used to create a partition root.  The content of this
2367	file will always be a subset of "cpuset.cpus" and its parent's
2368	"cpuset.cpus.exclusive.effective" if its parent is not the root
2369	cgroup.  It will also be a subset of "cpuset.cpus.exclusive"
2370	if it is set.  If "cpuset.cpus.exclusive" is not set, it is
2371	treated to have an implicit value of "cpuset.cpus" in the
2372	formation of local partition.
2373
2374  cpuset.cpus.isolated
2375	A read-only and root cgroup only multiple values file.
2376
2377	This file shows the set of all isolated CPUs used in existing
2378	isolated partitions. It will be empty if no isolated partition
2379	is created.
2380
2381  cpuset.cpus.partition
2382	A read-write single value file which exists on non-root
2383	cpuset-enabled cgroups.  This flag is owned by the parent cgroup
2384	and is not delegatable.
2385
2386	It accepts only the following input values when written to.
2387
2388	  ==========	=====================================
2389	  "member"	Non-root member of a partition
2390	  "root"	Partition root
2391	  "isolated"	Partition root without load balancing
2392	  ==========	=====================================
2393
2394	A cpuset partition is a collection of cpuset-enabled cgroups with
2395	a partition root at the top of the hierarchy and its descendants
2396	except those that are separate partition roots themselves and
2397	their descendants.  A partition has exclusive access to the
2398	set of exclusive CPUs allocated to it.	Other cgroups outside
2399	of that partition cannot use any CPUs in that set.
2400
2401	There are two types of partitions - local and remote.  A local
2402	partition is one whose parent cgroup is also a valid partition
2403	root.  A remote partition is one whose parent cgroup is not a
2404	valid partition root itself.  Writing to "cpuset.cpus.exclusive"
2405	is optional for the creation of a local partition as its
2406	"cpuset.cpus.exclusive" file will assume an implicit value that
2407	is the same as "cpuset.cpus" if it is not set.	Writing the
2408	proper "cpuset.cpus.exclusive" values down the cgroup hierarchy
2409	before the target partition root is mandatory for the creation
2410	of a remote partition.
2411
2412	Currently, a remote partition cannot be created under a local
2413	partition.  All the ancestors of a remote partition root except
2414	the root cgroup cannot be a partition root.
2415
2416	The root cgroup is always a partition root and its state cannot
2417	be changed.  All other non-root cgroups start out as "member".
2418
2419	When set to "root", the current cgroup is the root of a new
2420	partition or scheduling domain.  The set of exclusive CPUs is
2421	determined by the value of its "cpuset.cpus.exclusive.effective".
2422
2423	When set to "isolated", the CPUs in that partition will be in
2424	an isolated state without any load balancing from the scheduler
2425	and excluded from the unbound workqueues.  Tasks placed in such
2426	a partition with multiple CPUs should be carefully distributed
2427	and bound to each of the individual CPUs for optimal performance.
2428
2429	A partition root ("root" or "isolated") can be in one of the
2430	two possible states - valid or invalid.  An invalid partition
2431	root is in a degraded state where some state information may
2432	be retained, but behaves more like a "member".
2433
2434	All possible state transitions among "member", "root" and
2435	"isolated" are allowed.
2436
2437	On read, the "cpuset.cpus.partition" file can show the following
2438	values.
2439
2440	  =============================	=====================================
2441	  "member"			Non-root member of a partition
2442	  "root"			Partition root
2443	  "isolated"			Partition root without load balancing
2444	  "root invalid (<reason>)"	Invalid partition root
2445	  "isolated invalid (<reason>)"	Invalid isolated partition root
2446	  =============================	=====================================
2447
2448	In the case of an invalid partition root, a descriptive string on
2449	why the partition is invalid is included within parentheses.
2450
2451	For a local partition root to be valid, the following conditions
2452	must be met.
2453
2454	1) The parent cgroup is a valid partition root.
2455	2) The "cpuset.cpus.exclusive.effective" file cannot be empty,
2456	   though it may contain offline CPUs.
2457	3) The "cpuset.cpus.effective" cannot be empty unless there is
2458	   no task associated with this partition.
2459
2460	For a remote partition root to be valid, all the above conditions
2461	except the first one must be met.
2462
2463	External events like hotplug or changes to "cpuset.cpus" or
2464	"cpuset.cpus.exclusive" can cause a valid partition root to
2465	become invalid and vice versa.	Note that a task cannot be
2466	moved to a cgroup with empty "cpuset.cpus.effective".
2467
2468	A valid non-root parent partition may distribute out all its CPUs
2469	to its child local partitions when there is no task associated
2470	with it.
2471
2472	Care must be taken to change a valid partition root to "member"
2473	as all its child local partitions, if present, will become
2474	invalid causing disruption to tasks running in those child
2475	partitions. These inactivated partitions could be recovered if
2476	their parent is switched back to a partition root with a proper
2477	value in "cpuset.cpus" or "cpuset.cpus.exclusive".
2478
2479	Poll and inotify events are triggered whenever the state of
2480	"cpuset.cpus.partition" changes.  That includes changes caused
2481	by write to "cpuset.cpus.partition", cpu hotplug or other
2482	changes that modify the validity status of the partition.
2483	This will allow user space agents to monitor unexpected changes
2484	to "cpuset.cpus.partition" without the need to do continuous
2485	polling.
2486
2487	A user can pre-configure certain CPUs to an isolated state
2488	with load balancing disabled at boot time with the "isolcpus"
2489	kernel boot command line option.  If those CPUs are to be put
2490	into a partition, they have to be used in an isolated partition.
2491
2492
2493Device controller
2494-----------------
2495
2496Device controller manages access to device files. It includes both
2497creation of new device files (using mknod), and access to the
2498existing device files.
2499
2500Cgroup v2 device controller has no interface files and is implemented
2501on top of cgroup BPF. To control access to device files, a user may
2502create bpf programs of type BPF_PROG_TYPE_CGROUP_DEVICE and attach
2503them to cgroups with BPF_CGROUP_DEVICE flag. On an attempt to access a
2504device file, corresponding BPF programs will be executed, and depending
2505on the return value the attempt will succeed or fail with -EPERM.
2506
2507A BPF_PROG_TYPE_CGROUP_DEVICE program takes a pointer to the
2508bpf_cgroup_dev_ctx structure, which describes the device access attempt:
2509access type (mknod/read/write) and device (type, major and minor numbers).
2510If the program returns 0, the attempt fails with -EPERM, otherwise it
2511succeeds.
2512
2513An example of BPF_PROG_TYPE_CGROUP_DEVICE program may be found in
2514tools/testing/selftests/bpf/progs/dev_cgroup.c in the kernel source tree.
2515
2516
2517RDMA
2518----
2519
2520The "rdma" controller regulates the distribution and accounting of
2521RDMA resources.
2522
2523RDMA Interface Files
2524~~~~~~~~~~~~~~~~~~~~
2525
2526  rdma.max
2527	A readwrite nested-keyed file that exists for all the cgroups
2528	except root that describes current configured resource limit
2529	for a RDMA/IB device.
2530
2531	Lines are keyed by device name and are not ordered.
2532	Each line contains space separated resource name and its configured
2533	limit that can be distributed.
2534
2535	The following nested keys are defined.
2536
2537	  ==========	=============================
2538	  hca_handle	Maximum number of HCA Handles
2539	  hca_object 	Maximum number of HCA Objects
2540	  ==========	=============================
2541
2542	An example for mlx4 and ocrdma device follows::
2543
2544	  mlx4_0 hca_handle=2 hca_object=2000
2545	  ocrdma1 hca_handle=3 hca_object=max
2546
2547  rdma.current
2548	A read-only file that describes current resource usage.
2549	It exists for all the cgroup except root.
2550
2551	An example for mlx4 and ocrdma device follows::
2552
2553	  mlx4_0 hca_handle=1 hca_object=20
2554	  ocrdma1 hca_handle=1 hca_object=23
2555
2556HugeTLB
2557-------
2558
2559The HugeTLB controller allows to limit the HugeTLB usage per control group and
2560enforces the controller limit during page fault.
2561
2562HugeTLB Interface Files
2563~~~~~~~~~~~~~~~~~~~~~~~
2564
2565  hugetlb.<hugepagesize>.current
2566	Show current usage for "hugepagesize" hugetlb.  It exists for all
2567	the cgroup except root.
2568
2569  hugetlb.<hugepagesize>.max
2570	Set/show the hard limit of "hugepagesize" hugetlb usage.
2571	The default value is "max".  It exists for all the cgroup except root.
2572
2573  hugetlb.<hugepagesize>.events
2574	A read-only flat-keyed file which exists on non-root cgroups.
2575
2576	  max
2577		The number of allocation failure due to HugeTLB limit
2578
2579  hugetlb.<hugepagesize>.events.local
2580	Similar to hugetlb.<hugepagesize>.events but the fields in the file
2581	are local to the cgroup i.e. not hierarchical. The file modified event
2582	generated on this file reflects only the local events.
2583
2584  hugetlb.<hugepagesize>.numa_stat
2585	Similar to memory.numa_stat, it shows the numa information of the
2586        hugetlb pages of <hugepagesize> in this cgroup.  Only active in
2587        use hugetlb pages are included.  The per-node values are in bytes.
2588
2589Misc
2590----
2591
2592The Miscellaneous cgroup provides the resource limiting and tracking
2593mechanism for the scalar resources which cannot be abstracted like the other
2594cgroup resources. Controller is enabled by the CONFIG_CGROUP_MISC config
2595option.
2596
2597A resource can be added to the controller via enum misc_res_type{} in the
2598include/linux/misc_cgroup.h file and the corresponding name via misc_res_name[]
2599in the kernel/cgroup/misc.c file. Provider of the resource must set its
2600capacity prior to using the resource by calling misc_cg_set_capacity().
2601
2602Once a capacity is set then the resource usage can be updated using charge and
2603uncharge APIs. All of the APIs to interact with misc controller are in
2604include/linux/misc_cgroup.h.
2605
2606Misc Interface Files
2607~~~~~~~~~~~~~~~~~~~~
2608
2609Miscellaneous controller provides 3 interface files. If two misc resources (res_a and res_b) are registered then:
2610
2611  misc.capacity
2612        A read-only flat-keyed file shown only in the root cgroup.  It shows
2613        miscellaneous scalar resources available on the platform along with
2614        their quantities::
2615
2616	  $ cat misc.capacity
2617	  res_a 50
2618	  res_b 10
2619
2620  misc.current
2621        A read-only flat-keyed file shown in the all cgroups.  It shows
2622        the current usage of the resources in the cgroup and its children.::
2623
2624	  $ cat misc.current
2625	  res_a 3
2626	  res_b 0
2627
2628  misc.max
2629        A read-write flat-keyed file shown in the non root cgroups. Allowed
2630        maximum usage of the resources in the cgroup and its children.::
2631
2632	  $ cat misc.max
2633	  res_a max
2634	  res_b 4
2635
2636	Limit can be set by::
2637
2638	  # echo res_a 1 > misc.max
2639
2640	Limit can be set to max by::
2641
2642	  # echo res_a max > misc.max
2643
2644        Limits can be set higher than the capacity value in the misc.capacity
2645        file.
2646
2647  misc.events
2648	A read-only flat-keyed file which exists on non-root cgroups. The
2649	following entries are defined. Unless specified otherwise, a value
2650	change in this file generates a file modified event. All fields in
2651	this file are hierarchical.
2652
2653	  max
2654		The number of times the cgroup's resource usage was
2655		about to go over the max boundary.
2656
2657Migration and Ownership
2658~~~~~~~~~~~~~~~~~~~~~~~
2659
2660A miscellaneous scalar resource is charged to the cgroup in which it is used
2661first, and stays charged to that cgroup until that resource is freed. Migrating
2662a process to a different cgroup does not move the charge to the destination
2663cgroup where the process has moved.
2664
2665Others
2666------
2667
2668perf_event
2669~~~~~~~~~~
2670
2671perf_event controller, if not mounted on a legacy hierarchy, is
2672automatically enabled on the v2 hierarchy so that perf events can
2673always be filtered by cgroup v2 path.  The controller can still be
2674moved to a legacy hierarchy after v2 hierarchy is populated.
2675
2676
2677Non-normative information
2678-------------------------
2679
2680This section contains information that isn't considered to be a part of
2681the stable kernel API and so is subject to change.
2682
2683
2684CPU controller root cgroup process behaviour
2685~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
2686
2687When distributing CPU cycles in the root cgroup each thread in this
2688cgroup is treated as if it was hosted in a separate child cgroup of the
2689root cgroup. This child cgroup weight is dependent on its thread nice
2690level.
2691
2692For details of this mapping see sched_prio_to_weight array in
2693kernel/sched/core.c file (values from this array should be scaled
2694appropriately so the neutral - nice 0 - value is 100 instead of 1024).
2695
2696
2697IO controller root cgroup process behaviour
2698~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
2699
2700Root cgroup processes are hosted in an implicit leaf child node.
2701When distributing IO resources this implicit child node is taken into
2702account as if it was a normal child cgroup of the root cgroup with a
2703weight value of 200.
2704
2705
2706Namespace
2707=========
2708
2709Basics
2710------
2711
2712cgroup namespace provides a mechanism to virtualize the view of the
2713"/proc/$PID/cgroup" file and cgroup mounts.  The CLONE_NEWCGROUP clone
2714flag can be used with clone(2) and unshare(2) to create a new cgroup
2715namespace.  The process running inside the cgroup namespace will have
2716its "/proc/$PID/cgroup" output restricted to cgroupns root.  The
2717cgroupns root is the cgroup of the process at the time of creation of
2718the cgroup namespace.
2719
2720Without cgroup namespace, the "/proc/$PID/cgroup" file shows the
2721complete path of the cgroup of a process.  In a container setup where
2722a set of cgroups and namespaces are intended to isolate processes the
2723"/proc/$PID/cgroup" file may leak potential system level information
2724to the isolated processes.  For example::
2725
2726  # cat /proc/self/cgroup
2727  0::/batchjobs/container_id1
2728
2729The path '/batchjobs/container_id1' can be considered as system-data
2730and undesirable to expose to the isolated processes.  cgroup namespace
2731can be used to restrict visibility of this path.  For example, before
2732creating a cgroup namespace, one would see::
2733
2734  # ls -l /proc/self/ns/cgroup
2735  lrwxrwxrwx 1 root root 0 2014-07-15 10:37 /proc/self/ns/cgroup -> cgroup:[4026531835]
2736  # cat /proc/self/cgroup
2737  0::/batchjobs/container_id1
2738
2739After unsharing a new namespace, the view changes::
2740
2741  # ls -l /proc/self/ns/cgroup
2742  lrwxrwxrwx 1 root root 0 2014-07-15 10:35 /proc/self/ns/cgroup -> cgroup:[4026532183]
2743  # cat /proc/self/cgroup
2744  0::/
2745
2746When some thread from a multi-threaded process unshares its cgroup
2747namespace, the new cgroupns gets applied to the entire process (all
2748the threads).  This is natural for the v2 hierarchy; however, for the
2749legacy hierarchies, this may be unexpected.
2750
2751A cgroup namespace is alive as long as there are processes inside or
2752mounts pinning it.  When the last usage goes away, the cgroup
2753namespace is destroyed.  The cgroupns root and the actual cgroups
2754remain.
2755
2756
2757The Root and Views
2758------------------
2759
2760The 'cgroupns root' for a cgroup namespace is the cgroup in which the
2761process calling unshare(2) is running.  For example, if a process in
2762/batchjobs/container_id1 cgroup calls unshare, cgroup
2763/batchjobs/container_id1 becomes the cgroupns root.  For the
2764init_cgroup_ns, this is the real root ('/') cgroup.
2765
2766The cgroupns root cgroup does not change even if the namespace creator
2767process later moves to a different cgroup::
2768
2769  # ~/unshare -c # unshare cgroupns in some cgroup
2770  # cat /proc/self/cgroup
2771  0::/
2772  # mkdir sub_cgrp_1
2773  # echo 0 > sub_cgrp_1/cgroup.procs
2774  # cat /proc/self/cgroup
2775  0::/sub_cgrp_1
2776
2777Each process gets its namespace-specific view of "/proc/$PID/cgroup"
2778
2779Processes running inside the cgroup namespace will be able to see
2780cgroup paths (in /proc/self/cgroup) only inside their root cgroup.
2781From within an unshared cgroupns::
2782
2783  # sleep 100000 &
2784  [1] 7353
2785  # echo 7353 > sub_cgrp_1/cgroup.procs
2786  # cat /proc/7353/cgroup
2787  0::/sub_cgrp_1
2788
2789From the initial cgroup namespace, the real cgroup path will be
2790visible::
2791
2792  $ cat /proc/7353/cgroup
2793  0::/batchjobs/container_id1/sub_cgrp_1
2794
2795From a sibling cgroup namespace (that is, a namespace rooted at a
2796different cgroup), the cgroup path relative to its own cgroup
2797namespace root will be shown.  For instance, if PID 7353's cgroup
2798namespace root is at '/batchjobs/container_id2', then it will see::
2799
2800  # cat /proc/7353/cgroup
2801  0::/../container_id2/sub_cgrp_1
2802
2803Note that the relative path always starts with '/' to indicate that
2804its relative to the cgroup namespace root of the caller.
2805
2806
2807Migration and setns(2)
2808----------------------
2809
2810Processes inside a cgroup namespace can move into and out of the
2811namespace root if they have proper access to external cgroups.  For
2812example, from inside a namespace with cgroupns root at
2813/batchjobs/container_id1, and assuming that the global hierarchy is
2814still accessible inside cgroupns::
2815
2816  # cat /proc/7353/cgroup
2817  0::/sub_cgrp_1
2818  # echo 7353 > batchjobs/container_id2/cgroup.procs
2819  # cat /proc/7353/cgroup
2820  0::/../container_id2
2821
2822Note that this kind of setup is not encouraged.  A task inside cgroup
2823namespace should only be exposed to its own cgroupns hierarchy.
2824
2825setns(2) to another cgroup namespace is allowed when:
2826
2827(a) the process has CAP_SYS_ADMIN against its current user namespace
2828(b) the process has CAP_SYS_ADMIN against the target cgroup
2829    namespace's userns
2830
2831No implicit cgroup changes happen with attaching to another cgroup
2832namespace.  It is expected that the someone moves the attaching
2833process under the target cgroup namespace root.
2834
2835
2836Interaction with Other Namespaces
2837---------------------------------
2838
2839Namespace specific cgroup hierarchy can be mounted by a process
2840running inside a non-init cgroup namespace::
2841
2842  # mount -t cgroup2 none $MOUNT_POINT
2843
2844This will mount the unified cgroup hierarchy with cgroupns root as the
2845filesystem root.  The process needs CAP_SYS_ADMIN against its user and
2846mount namespaces.
2847
2848The virtualization of /proc/self/cgroup file combined with restricting
2849the view of cgroup hierarchy by namespace-private cgroupfs mount
2850provides a properly isolated cgroup view inside the container.
2851
2852
2853Information on Kernel Programming
2854=================================
2855
2856This section contains kernel programming information in the areas
2857where interacting with cgroup is necessary.  cgroup core and
2858controllers are not covered.
2859
2860
2861Filesystem Support for Writeback
2862--------------------------------
2863
2864A filesystem can support cgroup writeback by updating
2865address_space_operations->writepage[s]() to annotate bio's using the
2866following two functions.
2867
2868  wbc_init_bio(@wbc, @bio)
2869	Should be called for each bio carrying writeback data and
2870	associates the bio with the inode's owner cgroup and the
2871	corresponding request queue.  This must be called after
2872	a queue (device) has been associated with the bio and
2873	before submission.
2874
2875  wbc_account_cgroup_owner(@wbc, @page, @bytes)
2876	Should be called for each data segment being written out.
2877	While this function doesn't care exactly when it's called
2878	during the writeback session, it's the easiest and most
2879	natural to call it as data segments are added to a bio.
2880
2881With writeback bio's annotated, cgroup support can be enabled per
2882super_block by setting SB_I_CGROUPWB in ->s_iflags.  This allows for
2883selective disabling of cgroup writeback support which is helpful when
2884certain filesystem features, e.g. journaled data mode, are
2885incompatible.
2886
2887wbc_init_bio() binds the specified bio to its cgroup.  Depending on
2888the configuration, the bio may be executed at a lower priority and if
2889the writeback session is holding shared resources, e.g. a journal
2890entry, may lead to priority inversion.  There is no one easy solution
2891for the problem.  Filesystems can try to work around specific problem
2892cases by skipping wbc_init_bio() and using bio_associate_blkg()
2893directly.
2894
2895
2896Deprecated v1 Core Features
2897===========================
2898
2899- Multiple hierarchies including named ones are not supported.
2900
2901- All v1 mount options are not supported.
2902
2903- The "tasks" file is removed and "cgroup.procs" is not sorted.
2904
2905- "cgroup.clone_children" is removed.
2906
2907- /proc/cgroups is meaningless for v2.  Use "cgroup.controllers" file
2908  at the root instead.
2909
2910
2911Issues with v1 and Rationales for v2
2912====================================
2913
2914Multiple Hierarchies
2915--------------------
2916
2917cgroup v1 allowed an arbitrary number of hierarchies and each
2918hierarchy could host any number of controllers.  While this seemed to
2919provide a high level of flexibility, it wasn't useful in practice.
2920
2921For example, as there is only one instance of each controller, utility
2922type controllers such as freezer which can be useful in all
2923hierarchies could only be used in one.  The issue is exacerbated by
2924the fact that controllers couldn't be moved to another hierarchy once
2925hierarchies were populated.  Another issue was that all controllers
2926bound to a hierarchy were forced to have exactly the same view of the
2927hierarchy.  It wasn't possible to vary the granularity depending on
2928the specific controller.
2929
2930In practice, these issues heavily limited which controllers could be
2931put on the same hierarchy and most configurations resorted to putting
2932each controller on its own hierarchy.  Only closely related ones, such
2933as the cpu and cpuacct controllers, made sense to be put on the same
2934hierarchy.  This often meant that userland ended up managing multiple
2935similar hierarchies repeating the same steps on each hierarchy
2936whenever a hierarchy management operation was necessary.
2937
2938Furthermore, support for multiple hierarchies came at a steep cost.
2939It greatly complicated cgroup core implementation but more importantly
2940the support for multiple hierarchies restricted how cgroup could be
2941used in general and what controllers was able to do.
2942
2943There was no limit on how many hierarchies there might be, which meant
2944that a thread's cgroup membership couldn't be described in finite
2945length.  The key might contain any number of entries and was unlimited
2946in length, which made it highly awkward to manipulate and led to
2947addition of controllers which existed only to identify membership,
2948which in turn exacerbated the original problem of proliferating number
2949of hierarchies.
2950
2951Also, as a controller couldn't have any expectation regarding the
2952topologies of hierarchies other controllers might be on, each
2953controller had to assume that all other controllers were attached to
2954completely orthogonal hierarchies.  This made it impossible, or at
2955least very cumbersome, for controllers to cooperate with each other.
2956
2957In most use cases, putting controllers on hierarchies which are
2958completely orthogonal to each other isn't necessary.  What usually is
2959called for is the ability to have differing levels of granularity
2960depending on the specific controller.  In other words, hierarchy may
2961be collapsed from leaf towards root when viewed from specific
2962controllers.  For example, a given configuration might not care about
2963how memory is distributed beyond a certain level while still wanting
2964to control how CPU cycles are distributed.
2965
2966
2967Thread Granularity
2968------------------
2969
2970cgroup v1 allowed threads of a process to belong to different cgroups.
2971This didn't make sense for some controllers and those controllers
2972ended up implementing different ways to ignore such situations but
2973much more importantly it blurred the line between API exposed to
2974individual applications and system management interface.
2975
2976Generally, in-process knowledge is available only to the process
2977itself; thus, unlike service-level organization of processes,
2978categorizing threads of a process requires active participation from
2979the application which owns the target process.
2980
2981cgroup v1 had an ambiguously defined delegation model which got abused
2982in combination with thread granularity.  cgroups were delegated to
2983individual applications so that they can create and manage their own
2984sub-hierarchies and control resource distributions along them.  This
2985effectively raised cgroup to the status of a syscall-like API exposed
2986to lay programs.
2987
2988First of all, cgroup has a fundamentally inadequate interface to be
2989exposed this way.  For a process to access its own knobs, it has to
2990extract the path on the target hierarchy from /proc/self/cgroup,
2991construct the path by appending the name of the knob to the path, open
2992and then read and/or write to it.  This is not only extremely clunky
2993and unusual but also inherently racy.  There is no conventional way to
2994define transaction across the required steps and nothing can guarantee
2995that the process would actually be operating on its own sub-hierarchy.
2996
2997cgroup controllers implemented a number of knobs which would never be
2998accepted as public APIs because they were just adding control knobs to
2999system-management pseudo filesystem.  cgroup ended up with interface
3000knobs which were not properly abstracted or refined and directly
3001revealed kernel internal details.  These knobs got exposed to
3002individual applications through the ill-defined delegation mechanism
3003effectively abusing cgroup as a shortcut to implementing public APIs
3004without going through the required scrutiny.
3005
3006This was painful for both userland and kernel.  Userland ended up with
3007misbehaving and poorly abstracted interfaces and kernel exposing and
3008locked into constructs inadvertently.
3009
3010
3011Competition Between Inner Nodes and Threads
3012-------------------------------------------
3013
3014cgroup v1 allowed threads to be in any cgroups which created an
3015interesting problem where threads belonging to a parent cgroup and its
3016children cgroups competed for resources.  This was nasty as two
3017different types of entities competed and there was no obvious way to
3018settle it.  Different controllers did different things.
3019
3020The cpu controller considered threads and cgroups as equivalents and
3021mapped nice levels to cgroup weights.  This worked for some cases but
3022fell flat when children wanted to be allocated specific ratios of CPU
3023cycles and the number of internal threads fluctuated - the ratios
3024constantly changed as the number of competing entities fluctuated.
3025There also were other issues.  The mapping from nice level to weight
3026wasn't obvious or universal, and there were various other knobs which
3027simply weren't available for threads.
3028
3029The io controller implicitly created a hidden leaf node for each
3030cgroup to host the threads.  The hidden leaf had its own copies of all
3031the knobs with ``leaf_`` prefixed.  While this allowed equivalent
3032control over internal threads, it was with serious drawbacks.  It
3033always added an extra layer of nesting which wouldn't be necessary
3034otherwise, made the interface messy and significantly complicated the
3035implementation.
3036
3037The memory controller didn't have a way to control what happened
3038between internal tasks and child cgroups and the behavior was not
3039clearly defined.  There were attempts to add ad-hoc behaviors and
3040knobs to tailor the behavior to specific workloads which would have
3041led to problems extremely difficult to resolve in the long term.
3042
3043Multiple controllers struggled with internal tasks and came up with
3044different ways to deal with it; unfortunately, all the approaches were
3045severely flawed and, furthermore, the widely different behaviors
3046made cgroup as a whole highly inconsistent.
3047
3048This clearly is a problem which needs to be addressed from cgroup core
3049in a uniform way.
3050
3051
3052Other Interface Issues
3053----------------------
3054
3055cgroup v1 grew without oversight and developed a large number of
3056idiosyncrasies and inconsistencies.  One issue on the cgroup core side
3057was how an empty cgroup was notified - a userland helper binary was
3058forked and executed for each event.  The event delivery wasn't
3059recursive or delegatable.  The limitations of the mechanism also led
3060to in-kernel event delivery filtering mechanism further complicating
3061the interface.
3062
3063Controller interfaces were problematic too.  An extreme example is
3064controllers completely ignoring hierarchical organization and treating
3065all cgroups as if they were all located directly under the root
3066cgroup.  Some controllers exposed a large amount of inconsistent
3067implementation details to userland.
3068
3069There also was no consistency across controllers.  When a new cgroup
3070was created, some controllers defaulted to not imposing extra
3071restrictions while others disallowed any resource usage until
3072explicitly configured.  Configuration knobs for the same type of
3073control used widely differing naming schemes and formats.  Statistics
3074and information knobs were named arbitrarily and used different
3075formats and units even in the same controller.
3076
3077cgroup v2 establishes common conventions where appropriate and updates
3078controllers so that they expose minimal and consistent interfaces.
3079
3080
3081Controller Issues and Remedies
3082------------------------------
3083
3084Memory
3085~~~~~~
3086
3087The original lower boundary, the soft limit, is defined as a limit
3088that is per default unset.  As a result, the set of cgroups that
3089global reclaim prefers is opt-in, rather than opt-out.  The costs for
3090optimizing these mostly negative lookups are so high that the
3091implementation, despite its enormous size, does not even provide the
3092basic desirable behavior.  First off, the soft limit has no
3093hierarchical meaning.  All configured groups are organized in a global
3094rbtree and treated like equal peers, regardless where they are located
3095in the hierarchy.  This makes subtree delegation impossible.  Second,
3096the soft limit reclaim pass is so aggressive that it not just
3097introduces high allocation latencies into the system, but also impacts
3098system performance due to overreclaim, to the point where the feature
3099becomes self-defeating.
3100
3101The memory.low boundary on the other hand is a top-down allocated
3102reserve.  A cgroup enjoys reclaim protection when it's within its
3103effective low, which makes delegation of subtrees possible. It also
3104enjoys having reclaim pressure proportional to its overage when
3105above its effective low.
3106
3107The original high boundary, the hard limit, is defined as a strict
3108limit that can not budge, even if the OOM killer has to be called.
3109But this generally goes against the goal of making the most out of the
3110available memory.  The memory consumption of workloads varies during
3111runtime, and that requires users to overcommit.  But doing that with a
3112strict upper limit requires either a fairly accurate prediction of the
3113working set size or adding slack to the limit.  Since working set size
3114estimation is hard and error prone, and getting it wrong results in
3115OOM kills, most users tend to err on the side of a looser limit and
3116end up wasting precious resources.
3117
3118The memory.high boundary on the other hand can be set much more
3119conservatively.  When hit, it throttles allocations by forcing them
3120into direct reclaim to work off the excess, but it never invokes the
3121OOM killer.  As a result, a high boundary that is chosen too
3122aggressively will not terminate the processes, but instead it will
3123lead to gradual performance degradation.  The user can monitor this
3124and make corrections until the minimal memory footprint that still
3125gives acceptable performance is found.
3126
3127In extreme cases, with many concurrent allocations and a complete
3128breakdown of reclaim progress within the group, the high boundary can
3129be exceeded.  But even then it's mostly better to satisfy the
3130allocation from the slack available in other groups or the rest of the
3131system than killing the group.  Otherwise, memory.max is there to
3132limit this type of spillover and ultimately contain buggy or even
3133malicious applications.
3134
3135Setting the original memory.limit_in_bytes below the current usage was
3136subject to a race condition, where concurrent charges could cause the
3137limit setting to fail. memory.max on the other hand will first set the
3138limit to prevent new charges, and then reclaim and OOM kill until the
3139new limit is met - or the task writing to memory.max is killed.
3140
3141The combined memory+swap accounting and limiting is replaced by real
3142control over swap space.
3143
3144The main argument for a combined memory+swap facility in the original
3145cgroup design was that global or parental pressure would always be
3146able to swap all anonymous memory of a child group, regardless of the
3147child's own (possibly untrusted) configuration.  However, untrusted
3148groups can sabotage swapping by other means - such as referencing its
3149anonymous memory in a tight loop - and an admin can not assume full
3150swappability when overcommitting untrusted jobs.
3151
3152For trusted jobs, on the other hand, a combined counter is not an
3153intuitive userspace interface, and it flies in the face of the idea
3154that cgroup controllers should account and limit specific physical
3155resources.  Swap space is a resource like all others in the system,
3156and that's why unified hierarchy allows distributing it separately.
3157