1================
2Control Group v2
3================
4
5:Date: October, 2015
6:Author: Tejun Heo <tj@kernel.org>
7
8This is the authoritative documentation on the design, interface and
9conventions of cgroup v2.  It describes all userland-visible aspects
10of cgroup including core and specific controller behaviors.  All
11future changes must be reflected in this document.  Documentation for
12v1 is available under Documentation/admin-guide/cgroup-v1/.
13
14.. CONTENTS
15
16   1. Introduction
17     1-1. Terminology
18     1-2. What is cgroup?
19   2. Basic Operations
20     2-1. Mounting
21     2-2. Organizing Processes and Threads
22       2-2-1. Processes
23       2-2-2. Threads
24     2-3. [Un]populated Notification
25     2-4. Controlling Controllers
26       2-4-1. Enabling and Disabling
27       2-4-2. Top-down Constraint
28       2-4-3. No Internal Process Constraint
29     2-5. Delegation
30       2-5-1. Model of Delegation
31       2-5-2. Delegation Containment
32     2-6. Guidelines
33       2-6-1. Organize Once and Control
34       2-6-2. Avoid Name Collisions
35   3. Resource Distribution Models
36     3-1. Weights
37     3-2. Limits
38     3-3. Protections
39     3-4. Allocations
40   4. Interface Files
41     4-1. Format
42     4-2. Conventions
43     4-3. Core Interface Files
44   5. Controllers
45     5-1. CPU
46       5-1-1. CPU Interface Files
47     5-2. Memory
48       5-2-1. Memory Interface Files
49       5-2-2. Usage Guidelines
50       5-2-3. Memory Ownership
51     5-3. IO
52       5-3-1. IO Interface Files
53       5-3-2. Writeback
54       5-3-3. IO Latency
55         5-3-3-1. How IO Latency Throttling Works
56         5-3-3-2. IO Latency Interface Files
57     5-4. PID
58       5-4-1. PID Interface Files
59     5-5. Cpuset
60       5.5-1. Cpuset Interface Files
61     5-6. Device
62     5-7. RDMA
63       5-7-1. RDMA Interface Files
64     5-8. Misc
65       5-8-1. perf_event
66     5-N. Non-normative information
67       5-N-1. CPU controller root cgroup process behaviour
68       5-N-2. IO controller root cgroup process behaviour
69   6. Namespace
70     6-1. Basics
71     6-2. The Root and Views
72     6-3. Migration and setns(2)
73     6-4. Interaction with Other Namespaces
74   P. Information on Kernel Programming
75     P-1. Filesystem Support for Writeback
76   D. Deprecated v1 Core Features
77   R. Issues with v1 and Rationales for v2
78     R-1. Multiple Hierarchies
79     R-2. Thread Granularity
80     R-3. Competition Between Inner Nodes and Threads
81     R-4. Other Interface Issues
82     R-5. Controller Issues and Remedies
83       R-5-1. Memory
84
85
86Introduction
87============
88
89Terminology
90-----------
91
92"cgroup" stands for "control group" and is never capitalized.  The
93singular form is used to designate the whole feature and also as a
94qualifier as in "cgroup controllers".  When explicitly referring to
95multiple individual control groups, the plural form "cgroups" is used.
96
97
98What is cgroup?
99---------------
100
101cgroup is a mechanism to organize processes hierarchically and
102distribute system resources along the hierarchy in a controlled and
103configurable manner.
104
105cgroup is largely composed of two parts - the core and controllers.
106cgroup core is primarily responsible for hierarchically organizing
107processes.  A cgroup controller is usually responsible for
108distributing a specific type of system resource along the hierarchy
109although there are utility controllers which serve purposes other than
110resource distribution.
111
112cgroups form a tree structure and every process in the system belongs
113to one and only one cgroup.  All threads of a process belong to the
114same cgroup.  On creation, all processes are put in the cgroup that
115the parent process belongs to at the time.  A process can be migrated
116to another cgroup.  Migration of a process doesn't affect already
117existing descendant processes.
118
119Following certain structural constraints, controllers may be enabled or
120disabled selectively on a cgroup.  All controller behaviors are
121hierarchical - if a controller is enabled on a cgroup, it affects all
122processes which belong to the cgroups consisting the inclusive
123sub-hierarchy of the cgroup.  When a controller is enabled on a nested
124cgroup, it always restricts the resource distribution further.  The
125restrictions set closer to the root in the hierarchy can not be
126overridden from further away.
127
128
129Basic Operations
130================
131
132Mounting
133--------
134
135Unlike v1, cgroup v2 has only single hierarchy.  The cgroup v2
136hierarchy can be mounted with the following mount command::
137
138  # mount -t cgroup2 none $MOUNT_POINT
139
140cgroup2 filesystem has the magic number 0x63677270 ("cgrp").  All
141controllers which support v2 and are not bound to a v1 hierarchy are
142automatically bound to the v2 hierarchy and show up at the root.
143Controllers which are not in active use in the v2 hierarchy can be
144bound to other hierarchies.  This allows mixing v2 hierarchy with the
145legacy v1 multiple hierarchies in a fully backward compatible way.
146
147A controller can be moved across hierarchies only after the controller
148is no longer referenced in its current hierarchy.  Because per-cgroup
149controller states are destroyed asynchronously and controllers may
150have lingering references, a controller may not show up immediately on
151the v2 hierarchy after the final umount of the previous hierarchy.
152Similarly, a controller should be fully disabled to be moved out of
153the unified hierarchy and it may take some time for the disabled
154controller to become available for other hierarchies; furthermore, due
155to inter-controller dependencies, other controllers may need to be
156disabled too.
157
158While useful for development and manual configurations, moving
159controllers dynamically between the v2 and other hierarchies is
160strongly discouraged for production use.  It is recommended to decide
161the hierarchies and controller associations before starting using the
162controllers after system boot.
163
164During transition to v2, system management software might still
165automount the v1 cgroup filesystem and so hijack all controllers
166during boot, before manual intervention is possible. To make testing
167and experimenting easier, the kernel parameter cgroup_no_v1= allows
168disabling controllers in v1 and make them always available in v2.
169
170cgroup v2 currently supports the following mount options.
171
172  nsdelegate
173
174	Consider cgroup namespaces as delegation boundaries.  This
175	option is system wide and can only be set on mount or modified
176	through remount from the init namespace.  The mount option is
177	ignored on non-init namespace mounts.  Please refer to the
178	Delegation section for details.
179
180  memory_localevents
181
182        Only populate memory.events with data for the current cgroup,
183        and not any subtrees. This is legacy behaviour, the default
184        behaviour without this option is to include subtree counts.
185        This option is system wide and can only be set on mount or
186        modified through remount from the init namespace. The mount
187        option is ignored on non-init namespace mounts.
188
189
190Organizing Processes and Threads
191--------------------------------
192
193Processes
194~~~~~~~~~
195
196Initially, only the root cgroup exists to which all processes belong.
197A child cgroup can be created by creating a sub-directory::
198
199  # mkdir $CGROUP_NAME
200
201A given cgroup may have multiple child cgroups forming a tree
202structure.  Each cgroup has a read-writable interface file
203"cgroup.procs".  When read, it lists the PIDs of all processes which
204belong to the cgroup one-per-line.  The PIDs are not ordered and the
205same PID may show up more than once if the process got moved to
206another cgroup and then back or the PID got recycled while reading.
207
208A process can be migrated into a cgroup by writing its PID to the
209target cgroup's "cgroup.procs" file.  Only one process can be migrated
210on a single write(2) call.  If a process is composed of multiple
211threads, writing the PID of any thread migrates all threads of the
212process.
213
214When a process forks a child process, the new process is born into the
215cgroup that the forking process belongs to at the time of the
216operation.  After exit, a process stays associated with the cgroup
217that it belonged to at the time of exit until it's reaped; however, a
218zombie process does not appear in "cgroup.procs" and thus can't be
219moved to another cgroup.
220
221A cgroup which doesn't have any children or live processes can be
222destroyed by removing the directory.  Note that a cgroup which doesn't
223have any children and is associated only with zombie processes is
224considered empty and can be removed::
225
226  # rmdir $CGROUP_NAME
227
228"/proc/$PID/cgroup" lists a process's cgroup membership.  If legacy
229cgroup is in use in the system, this file may contain multiple lines,
230one for each hierarchy.  The entry for cgroup v2 is always in the
231format "0::$PATH"::
232
233  # cat /proc/842/cgroup
234  ...
235  0::/test-cgroup/test-cgroup-nested
236
237If the process becomes a zombie and the cgroup it was associated with
238is removed subsequently, " (deleted)" is appended to the path::
239
240  # cat /proc/842/cgroup
241  ...
242  0::/test-cgroup/test-cgroup-nested (deleted)
243
244
245Threads
246~~~~~~~
247
248cgroup v2 supports thread granularity for a subset of controllers to
249support use cases requiring hierarchical resource distribution across
250the threads of a group of processes.  By default, all threads of a
251process belong to the same cgroup, which also serves as the resource
252domain to host resource consumptions which are not specific to a
253process or thread.  The thread mode allows threads to be spread across
254a subtree while still maintaining the common resource domain for them.
255
256Controllers which support thread mode are called threaded controllers.
257The ones which don't are called domain controllers.
258
259Marking a cgroup threaded makes it join the resource domain of its
260parent as a threaded cgroup.  The parent may be another threaded
261cgroup whose resource domain is further up in the hierarchy.  The root
262of a threaded subtree, that is, the nearest ancestor which is not
263threaded, is called threaded domain or thread root interchangeably and
264serves as the resource domain for the entire subtree.
265
266Inside a threaded subtree, threads of a process can be put in
267different cgroups and are not subject to the no internal process
268constraint - threaded controllers can be enabled on non-leaf cgroups
269whether they have threads in them or not.
270
271As the threaded domain cgroup hosts all the domain resource
272consumptions of the subtree, it is considered to have internal
273resource consumptions whether there are processes in it or not and
274can't have populated child cgroups which aren't threaded.  Because the
275root cgroup is not subject to no internal process constraint, it can
276serve both as a threaded domain and a parent to domain cgroups.
277
278The current operation mode or type of the cgroup is shown in the
279"cgroup.type" file which indicates whether the cgroup is a normal
280domain, a domain which is serving as the domain of a threaded subtree,
281or a threaded cgroup.
282
283On creation, a cgroup is always a domain cgroup and can be made
284threaded by writing "threaded" to the "cgroup.type" file.  The
285operation is single direction::
286
287  # echo threaded > cgroup.type
288
289Once threaded, the cgroup can't be made a domain again.  To enable the
290thread mode, the following conditions must be met.
291
292- As the cgroup will join the parent's resource domain.  The parent
293  must either be a valid (threaded) domain or a threaded cgroup.
294
295- When the parent is an unthreaded domain, it must not have any domain
296  controllers enabled or populated domain children.  The root is
297  exempt from this requirement.
298
299Topology-wise, a cgroup can be in an invalid state.  Please consider
300the following topology::
301
302  A (threaded domain) - B (threaded) - C (domain, just created)
303
304C is created as a domain but isn't connected to a parent which can
305host child domains.  C can't be used until it is turned into a
306threaded cgroup.  "cgroup.type" file will report "domain (invalid)" in
307these cases.  Operations which fail due to invalid topology use
308EOPNOTSUPP as the errno.
309
310A domain cgroup is turned into a threaded domain when one of its child
311cgroup becomes threaded or threaded controllers are enabled in the
312"cgroup.subtree_control" file while there are processes in the cgroup.
313A threaded domain reverts to a normal domain when the conditions
314clear.
315
316When read, "cgroup.threads" contains the list of the thread IDs of all
317threads in the cgroup.  Except that the operations are per-thread
318instead of per-process, "cgroup.threads" has the same format and
319behaves the same way as "cgroup.procs".  While "cgroup.threads" can be
320written to in any cgroup, as it can only move threads inside the same
321threaded domain, its operations are confined inside each threaded
322subtree.
323
324The threaded domain cgroup serves as the resource domain for the whole
325subtree, and, while the threads can be scattered across the subtree,
326all the processes are considered to be in the threaded domain cgroup.
327"cgroup.procs" in a threaded domain cgroup contains the PIDs of all
328processes in the subtree and is not readable in the subtree proper.
329However, "cgroup.procs" can be written to from anywhere in the subtree
330to migrate all threads of the matching process to the cgroup.
331
332Only threaded controllers can be enabled in a threaded subtree.  When
333a threaded controller is enabled inside a threaded subtree, it only
334accounts for and controls resource consumptions associated with the
335threads in the cgroup and its descendants.  All consumptions which
336aren't tied to a specific thread belong to the threaded domain cgroup.
337
338Because a threaded subtree is exempt from no internal process
339constraint, a threaded controller must be able to handle competition
340between threads in a non-leaf cgroup and its child cgroups.  Each
341threaded controller defines how such competitions are handled.
342
343
344[Un]populated Notification
345--------------------------
346
347Each non-root cgroup has a "cgroup.events" file which contains
348"populated" field indicating whether the cgroup's sub-hierarchy has
349live processes in it.  Its value is 0 if there is no live process in
350the cgroup and its descendants; otherwise, 1.  poll and [id]notify
351events are triggered when the value changes.  This can be used, for
352example, to start a clean-up operation after all processes of a given
353sub-hierarchy have exited.  The populated state updates and
354notifications are recursive.  Consider the following sub-hierarchy
355where the numbers in the parentheses represent the numbers of processes
356in each cgroup::
357
358  A(4) - B(0) - C(1)
359              \ D(0)
360
361A, B and C's "populated" fields would be 1 while D's 0.  After the one
362process in C exits, B and C's "populated" fields would flip to "0" and
363file modified events will be generated on the "cgroup.events" files of
364both cgroups.
365
366
367Controlling Controllers
368-----------------------
369
370Enabling and Disabling
371~~~~~~~~~~~~~~~~~~~~~~
372
373Each cgroup has a "cgroup.controllers" file which lists all
374controllers available for the cgroup to enable::
375
376  # cat cgroup.controllers
377  cpu io memory
378
379No controller is enabled by default.  Controllers can be enabled and
380disabled by writing to the "cgroup.subtree_control" file::
381
382  # echo "+cpu +memory -io" > cgroup.subtree_control
383
384Only controllers which are listed in "cgroup.controllers" can be
385enabled.  When multiple operations are specified as above, either they
386all succeed or fail.  If multiple operations on the same controller
387are specified, the last one is effective.
388
389Enabling a controller in a cgroup indicates that the distribution of
390the target resource across its immediate children will be controlled.
391Consider the following sub-hierarchy.  The enabled controllers are
392listed in parentheses::
393
394  A(cpu,memory) - B(memory) - C()
395                            \ D()
396
397As A has "cpu" and "memory" enabled, A will control the distribution
398of CPU cycles and memory to its children, in this case, B.  As B has
399"memory" enabled but not "CPU", C and D will compete freely on CPU
400cycles but their division of memory available to B will be controlled.
401
402As a controller regulates the distribution of the target resource to
403the cgroup's children, enabling it creates the controller's interface
404files in the child cgroups.  In the above example, enabling "cpu" on B
405would create the "cpu." prefixed controller interface files in C and
406D.  Likewise, disabling "memory" from B would remove the "memory."
407prefixed controller interface files from C and D.  This means that the
408controller interface files - anything which doesn't start with
409"cgroup." are owned by the parent rather than the cgroup itself.
410
411
412Top-down Constraint
413~~~~~~~~~~~~~~~~~~~
414
415Resources are distributed top-down and a cgroup can further distribute
416a resource only if the resource has been distributed to it from the
417parent.  This means that all non-root "cgroup.subtree_control" files
418can only contain controllers which are enabled in the parent's
419"cgroup.subtree_control" file.  A controller can be enabled only if
420the parent has the controller enabled and a controller can't be
421disabled if one or more children have it enabled.
422
423
424No Internal Process Constraint
425~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
426
427Non-root cgroups can distribute domain resources to their children
428only when they don't have any processes of their own.  In other words,
429only domain cgroups which don't contain any processes can have domain
430controllers enabled in their "cgroup.subtree_control" files.
431
432This guarantees that, when a domain controller is looking at the part
433of the hierarchy which has it enabled, processes are always only on
434the leaves.  This rules out situations where child cgroups compete
435against internal processes of the parent.
436
437The root cgroup is exempt from this restriction.  Root contains
438processes and anonymous resource consumption which can't be associated
439with any other cgroups and requires special treatment from most
440controllers.  How resource consumption in the root cgroup is governed
441is up to each controller (for more information on this topic please
442refer to the Non-normative information section in the Controllers
443chapter).
444
445Note that the restriction doesn't get in the way if there is no
446enabled controller in the cgroup's "cgroup.subtree_control".  This is
447important as otherwise it wouldn't be possible to create children of a
448populated cgroup.  To control resource distribution of a cgroup, the
449cgroup must create children and transfer all its processes to the
450children before enabling controllers in its "cgroup.subtree_control"
451file.
452
453
454Delegation
455----------
456
457Model of Delegation
458~~~~~~~~~~~~~~~~~~~
459
460A cgroup can be delegated in two ways.  First, to a less privileged
461user by granting write access of the directory and its "cgroup.procs",
462"cgroup.threads" and "cgroup.subtree_control" files to the user.
463Second, if the "nsdelegate" mount option is set, automatically to a
464cgroup namespace on namespace creation.
465
466Because the resource control interface files in a given directory
467control the distribution of the parent's resources, the delegatee
468shouldn't be allowed to write to them.  For the first method, this is
469achieved by not granting access to these files.  For the second, the
470kernel rejects writes to all files other than "cgroup.procs" and
471"cgroup.subtree_control" on a namespace root from inside the
472namespace.
473
474The end results are equivalent for both delegation types.  Once
475delegated, the user can build sub-hierarchy under the directory,
476organize processes inside it as it sees fit and further distribute the
477resources it received from the parent.  The limits and other settings
478of all resource controllers are hierarchical and regardless of what
479happens in the delegated sub-hierarchy, nothing can escape the
480resource restrictions imposed by the parent.
481
482Currently, cgroup doesn't impose any restrictions on the number of
483cgroups in or nesting depth of a delegated sub-hierarchy; however,
484this may be limited explicitly in the future.
485
486
487Delegation Containment
488~~~~~~~~~~~~~~~~~~~~~~
489
490A delegated sub-hierarchy is contained in the sense that processes
491can't be moved into or out of the sub-hierarchy by the delegatee.
492
493For delegations to a less privileged user, this is achieved by
494requiring the following conditions for a process with a non-root euid
495to migrate a target process into a cgroup by writing its PID to the
496"cgroup.procs" file.
497
498- The writer must have write access to the "cgroup.procs" file.
499
500- The writer must have write access to the "cgroup.procs" file of the
501  common ancestor of the source and destination cgroups.
502
503The above two constraints ensure that while a delegatee may migrate
504processes around freely in the delegated sub-hierarchy it can't pull
505in from or push out to outside the sub-hierarchy.
506
507For an example, let's assume cgroups C0 and C1 have been delegated to
508user U0 who created C00, C01 under C0 and C10 under C1 as follows and
509all processes under C0 and C1 belong to U0::
510
511  ~~~~~~~~~~~~~ - C0 - C00
512  ~ cgroup    ~      \ C01
513  ~ hierarchy ~
514  ~~~~~~~~~~~~~ - C1 - C10
515
516Let's also say U0 wants to write the PID of a process which is
517currently in C10 into "C00/cgroup.procs".  U0 has write access to the
518file; however, the common ancestor of the source cgroup C10 and the
519destination cgroup C00 is above the points of delegation and U0 would
520not have write access to its "cgroup.procs" files and thus the write
521will be denied with -EACCES.
522
523For delegations to namespaces, containment is achieved by requiring
524that both the source and destination cgroups are reachable from the
525namespace of the process which is attempting the migration.  If either
526is not reachable, the migration is rejected with -ENOENT.
527
528
529Guidelines
530----------
531
532Organize Once and Control
533~~~~~~~~~~~~~~~~~~~~~~~~~
534
535Migrating a process across cgroups is a relatively expensive operation
536and stateful resources such as memory are not moved together with the
537process.  This is an explicit design decision as there often exist
538inherent trade-offs between migration and various hot paths in terms
539of synchronization cost.
540
541As such, migrating processes across cgroups frequently as a means to
542apply different resource restrictions is discouraged.  A workload
543should be assigned to a cgroup according to the system's logical and
544resource structure once on start-up.  Dynamic adjustments to resource
545distribution can be made by changing controller configuration through
546the interface files.
547
548
549Avoid Name Collisions
550~~~~~~~~~~~~~~~~~~~~~
551
552Interface files for a cgroup and its children cgroups occupy the same
553directory and it is possible to create children cgroups which collide
554with interface files.
555
556All cgroup core interface files are prefixed with "cgroup." and each
557controller's interface files are prefixed with the controller name and
558a dot.  A controller's name is composed of lower case alphabets and
559'_'s but never begins with an '_' so it can be used as the prefix
560character for collision avoidance.  Also, interface file names won't
561start or end with terms which are often used in categorizing workloads
562such as job, service, slice, unit or workload.
563
564cgroup doesn't do anything to prevent name collisions and it's the
565user's responsibility to avoid them.
566
567
568Resource Distribution Models
569============================
570
571cgroup controllers implement several resource distribution schemes
572depending on the resource type and expected use cases.  This section
573describes major schemes in use along with their expected behaviors.
574
575
576Weights
577-------
578
579A parent's resource is distributed by adding up the weights of all
580active children and giving each the fraction matching the ratio of its
581weight against the sum.  As only children which can make use of the
582resource at the moment participate in the distribution, this is
583work-conserving.  Due to the dynamic nature, this model is usually
584used for stateless resources.
585
586All weights are in the range [1, 10000] with the default at 100.  This
587allows symmetric multiplicative biases in both directions at fine
588enough granularity while staying in the intuitive range.
589
590As long as the weight is in range, all configuration combinations are
591valid and there is no reason to reject configuration changes or
592process migrations.
593
594"cpu.weight" proportionally distributes CPU cycles to active children
595and is an example of this type.
596
597
598Limits
599------
600
601A child can only consume upto the configured amount of the resource.
602Limits can be over-committed - the sum of the limits of children can
603exceed the amount of resource available to the parent.
604
605Limits are in the range [0, max] and defaults to "max", which is noop.
606
607As limits can be over-committed, all configuration combinations are
608valid and there is no reason to reject configuration changes or
609process migrations.
610
611"io.max" limits the maximum BPS and/or IOPS that a cgroup can consume
612on an IO device and is an example of this type.
613
614
615Protections
616-----------
617
618A cgroup is protected to be allocated upto the configured amount of
619the resource if the usages of all its ancestors are under their
620protected levels.  Protections can be hard guarantees or best effort
621soft boundaries.  Protections can also be over-committed in which case
622only upto the amount available to the parent is protected among
623children.
624
625Protections are in the range [0, max] and defaults to 0, which is
626noop.
627
628As protections can be over-committed, all configuration combinations
629are valid and there is no reason to reject configuration changes or
630process migrations.
631
632"memory.low" implements best-effort memory protection and is an
633example of this type.
634
635
636Allocations
637-----------
638
639A cgroup is exclusively allocated a certain amount of a finite
640resource.  Allocations can't be over-committed - the sum of the
641allocations of children can not exceed the amount of resource
642available to the parent.
643
644Allocations are in the range [0, max] and defaults to 0, which is no
645resource.
646
647As allocations can't be over-committed, some configuration
648combinations are invalid and should be rejected.  Also, if the
649resource is mandatory for execution of processes, process migrations
650may be rejected.
651
652"cpu.rt.max" hard-allocates realtime slices and is an example of this
653type.
654
655
656Interface Files
657===============
658
659Format
660------
661
662All interface files should be in one of the following formats whenever
663possible::
664
665  New-line separated values
666  (when only one value can be written at once)
667
668	VAL0\n
669	VAL1\n
670	...
671
672  Space separated values
673  (when read-only or multiple values can be written at once)
674
675	VAL0 VAL1 ...\n
676
677  Flat keyed
678
679	KEY0 VAL0\n
680	KEY1 VAL1\n
681	...
682
683  Nested keyed
684
685	KEY0 SUB_KEY0=VAL00 SUB_KEY1=VAL01...
686	KEY1 SUB_KEY0=VAL10 SUB_KEY1=VAL11...
687	...
688
689For a writable file, the format for writing should generally match
690reading; however, controllers may allow omitting later fields or
691implement restricted shortcuts for most common use cases.
692
693For both flat and nested keyed files, only the values for a single key
694can be written at a time.  For nested keyed files, the sub key pairs
695may be specified in any order and not all pairs have to be specified.
696
697
698Conventions
699-----------
700
701- Settings for a single feature should be contained in a single file.
702
703- The root cgroup should be exempt from resource control and thus
704  shouldn't have resource control interface files.  Also,
705  informational files on the root cgroup which end up showing global
706  information available elsewhere shouldn't exist.
707
708- The default time unit is microseconds.  If a different unit is ever
709  used, an explicit unit suffix must be present.
710
711- A parts-per quantity should use a percentage decimal with at least
712  two digit fractional part - e.g. 13.40.
713
714- If a controller implements weight based resource distribution, its
715  interface file should be named "weight" and have the range [1,
716  10000] with 100 as the default.  The values are chosen to allow
717  enough and symmetric bias in both directions while keeping it
718  intuitive (the default is 100%).
719
720- If a controller implements an absolute resource guarantee and/or
721  limit, the interface files should be named "min" and "max"
722  respectively.  If a controller implements best effort resource
723  guarantee and/or limit, the interface files should be named "low"
724  and "high" respectively.
725
726  In the above four control files, the special token "max" should be
727  used to represent upward infinity for both reading and writing.
728
729- If a setting has a configurable default value and keyed specific
730  overrides, the default entry should be keyed with "default" and
731  appear as the first entry in the file.
732
733  The default value can be updated by writing either "default $VAL" or
734  "$VAL".
735
736  When writing to update a specific override, "default" can be used as
737  the value to indicate removal of the override.  Override entries
738  with "default" as the value must not appear when read.
739
740  For example, a setting which is keyed by major:minor device numbers
741  with integer values may look like the following::
742
743    # cat cgroup-example-interface-file
744    default 150
745    8:0 300
746
747  The default value can be updated by::
748
749    # echo 125 > cgroup-example-interface-file
750
751  or::
752
753    # echo "default 125" > cgroup-example-interface-file
754
755  An override can be set by::
756
757    # echo "8:16 170" > cgroup-example-interface-file
758
759  and cleared by::
760
761    # echo "8:0 default" > cgroup-example-interface-file
762    # cat cgroup-example-interface-file
763    default 125
764    8:16 170
765
766- For events which are not very high frequency, an interface file
767  "events" should be created which lists event key value pairs.
768  Whenever a notifiable event happens, file modified event should be
769  generated on the file.
770
771
772Core Interface Files
773--------------------
774
775All cgroup core files are prefixed with "cgroup."
776
777  cgroup.type
778
779	A read-write single value file which exists on non-root
780	cgroups.
781
782	When read, it indicates the current type of the cgroup, which
783	can be one of the following values.
784
785	- "domain" : A normal valid domain cgroup.
786
787	- "domain threaded" : A threaded domain cgroup which is
788          serving as the root of a threaded subtree.
789
790	- "domain invalid" : A cgroup which is in an invalid state.
791	  It can't be populated or have controllers enabled.  It may
792	  be allowed to become a threaded cgroup.
793
794	- "threaded" : A threaded cgroup which is a member of a
795          threaded subtree.
796
797	A cgroup can be turned into a threaded cgroup by writing
798	"threaded" to this file.
799
800  cgroup.procs
801	A read-write new-line separated values file which exists on
802	all cgroups.
803
804	When read, it lists the PIDs of all processes which belong to
805	the cgroup one-per-line.  The PIDs are not ordered and the
806	same PID may show up more than once if the process got moved
807	to another cgroup and then back or the PID got recycled while
808	reading.
809
810	A PID can be written to migrate the process associated with
811	the PID to the cgroup.  The writer should match all of the
812	following conditions.
813
814	- It must have write access to the "cgroup.procs" file.
815
816	- It must have write access to the "cgroup.procs" file of the
817	  common ancestor of the source and destination cgroups.
818
819	When delegating a sub-hierarchy, write access to this file
820	should be granted along with the containing directory.
821
822	In a threaded cgroup, reading this file fails with EOPNOTSUPP
823	as all the processes belong to the thread root.  Writing is
824	supported and moves every thread of the process to the cgroup.
825
826  cgroup.threads
827	A read-write new-line separated values file which exists on
828	all cgroups.
829
830	When read, it lists the TIDs of all threads which belong to
831	the cgroup one-per-line.  The TIDs are not ordered and the
832	same TID may show up more than once if the thread got moved to
833	another cgroup and then back or the TID got recycled while
834	reading.
835
836	A TID can be written to migrate the thread associated with the
837	TID to the cgroup.  The writer should match all of the
838	following conditions.
839
840	- It must have write access to the "cgroup.threads" file.
841
842	- The cgroup that the thread is currently in must be in the
843          same resource domain as the destination cgroup.
844
845	- It must have write access to the "cgroup.procs" file of the
846	  common ancestor of the source and destination cgroups.
847
848	When delegating a sub-hierarchy, write access to this file
849	should be granted along with the containing directory.
850
851  cgroup.controllers
852	A read-only space separated values file which exists on all
853	cgroups.
854
855	It shows space separated list of all controllers available to
856	the cgroup.  The controllers are not ordered.
857
858  cgroup.subtree_control
859	A read-write space separated values file which exists on all
860	cgroups.  Starts out empty.
861
862	When read, it shows space separated list of the controllers
863	which are enabled to control resource distribution from the
864	cgroup to its children.
865
866	Space separated list of controllers prefixed with '+' or '-'
867	can be written to enable or disable controllers.  A controller
868	name prefixed with '+' enables the controller and '-'
869	disables.  If a controller appears more than once on the list,
870	the last one is effective.  When multiple enable and disable
871	operations are specified, either all succeed or all fail.
872
873  cgroup.events
874	A read-only flat-keyed file which exists on non-root cgroups.
875	The following entries are defined.  Unless specified
876	otherwise, a value change in this file generates a file
877	modified event.
878
879	  populated
880		1 if the cgroup or its descendants contains any live
881		processes; otherwise, 0.
882	  frozen
883		1 if the cgroup is frozen; otherwise, 0.
884
885  cgroup.max.descendants
886	A read-write single value files.  The default is "max".
887
888	Maximum allowed number of descent cgroups.
889	If the actual number of descendants is equal or larger,
890	an attempt to create a new cgroup in the hierarchy will fail.
891
892  cgroup.max.depth
893	A read-write single value files.  The default is "max".
894
895	Maximum allowed descent depth below the current cgroup.
896	If the actual descent depth is equal or larger,
897	an attempt to create a new child cgroup will fail.
898
899  cgroup.stat
900	A read-only flat-keyed file with the following entries:
901
902	  nr_descendants
903		Total number of visible descendant cgroups.
904
905	  nr_dying_descendants
906		Total number of dying descendant cgroups. A cgroup becomes
907		dying after being deleted by a user. The cgroup will remain
908		in dying state for some time undefined time (which can depend
909		on system load) before being completely destroyed.
910
911		A process can't enter a dying cgroup under any circumstances,
912		a dying cgroup can't revive.
913
914		A dying cgroup can consume system resources not exceeding
915		limits, which were active at the moment of cgroup deletion.
916
917  cgroup.freeze
918	A read-write single value file which exists on non-root cgroups.
919	Allowed values are "0" and "1". The default is "0".
920
921	Writing "1" to the file causes freezing of the cgroup and all
922	descendant cgroups. This means that all belonging processes will
923	be stopped and will not run until the cgroup will be explicitly
924	unfrozen. Freezing of the cgroup may take some time; when this action
925	is completed, the "frozen" value in the cgroup.events control file
926	will be updated to "1" and the corresponding notification will be
927	issued.
928
929	A cgroup can be frozen either by its own settings, or by settings
930	of any ancestor cgroups. If any of ancestor cgroups is frozen, the
931	cgroup will remain frozen.
932
933	Processes in the frozen cgroup can be killed by a fatal signal.
934	They also can enter and leave a frozen cgroup: either by an explicit
935	move by a user, or if freezing of the cgroup races with fork().
936	If a process is moved to a frozen cgroup, it stops. If a process is
937	moved out of a frozen cgroup, it becomes running.
938
939	Frozen status of a cgroup doesn't affect any cgroup tree operations:
940	it's possible to delete a frozen (and empty) cgroup, as well as
941	create new sub-cgroups.
942
943Controllers
944===========
945
946CPU
947---
948
949The "cpu" controllers regulates distribution of CPU cycles.  This
950controller implements weight and absolute bandwidth limit models for
951normal scheduling policy and absolute bandwidth allocation model for
952realtime scheduling policy.
953
954In all the above models, cycles distribution is defined only on a temporal
955base and it does not account for the frequency at which tasks are executed.
956The (optional) utilization clamping support allows to hint the schedutil
957cpufreq governor about the minimum desired frequency which should always be
958provided by a CPU, as well as the maximum desired frequency, which should not
959be exceeded by a CPU.
960
961WARNING: cgroup2 doesn't yet support control of realtime processes and
962the cpu controller can only be enabled when all RT processes are in
963the root cgroup.  Be aware that system management software may already
964have placed RT processes into nonroot cgroups during the system boot
965process, and these processes may need to be moved to the root cgroup
966before the cpu controller can be enabled.
967
968
969CPU Interface Files
970~~~~~~~~~~~~~~~~~~~
971
972All time durations are in microseconds.
973
974  cpu.stat
975	A read-only flat-keyed file which exists on non-root cgroups.
976	This file exists whether the controller is enabled or not.
977
978	It always reports the following three stats:
979
980	- usage_usec
981	- user_usec
982	- system_usec
983
984	and the following three when the controller is enabled:
985
986	- nr_periods
987	- nr_throttled
988	- throttled_usec
989
990  cpu.weight
991	A read-write single value file which exists on non-root
992	cgroups.  The default is "100".
993
994	The weight in the range [1, 10000].
995
996  cpu.weight.nice
997	A read-write single value file which exists on non-root
998	cgroups.  The default is "0".
999
1000	The nice value is in the range [-20, 19].
1001
1002	This interface file is an alternative interface for
1003	"cpu.weight" and allows reading and setting weight using the
1004	same values used by nice(2).  Because the range is smaller and
1005	granularity is coarser for the nice values, the read value is
1006	the closest approximation of the current weight.
1007
1008  cpu.max
1009	A read-write two value file which exists on non-root cgroups.
1010	The default is "max 100000".
1011
1012	The maximum bandwidth limit.  It's in the following format::
1013
1014	  $MAX $PERIOD
1015
1016	which indicates that the group may consume upto $MAX in each
1017	$PERIOD duration.  "max" for $MAX indicates no limit.  If only
1018	one number is written, $MAX is updated.
1019
1020  cpu.pressure
1021	A read-only nested-key file which exists on non-root cgroups.
1022
1023	Shows pressure stall information for CPU. See
1024	Documentation/accounting/psi.rst for details.
1025
1026  cpu.uclamp.min
1027        A read-write single value file which exists on non-root cgroups.
1028        The default is "0", i.e. no utilization boosting.
1029
1030        The requested minimum utilization (protection) as a percentage
1031        rational number, e.g. 12.34 for 12.34%.
1032
1033        This interface allows reading and setting minimum utilization clamp
1034        values similar to the sched_setattr(2). This minimum utilization
1035        value is used to clamp the task specific minimum utilization clamp.
1036
1037        The requested minimum utilization (protection) is always capped by
1038        the current value for the maximum utilization (limit), i.e.
1039        `cpu.uclamp.max`.
1040
1041  cpu.uclamp.max
1042        A read-write single value file which exists on non-root cgroups.
1043        The default is "max". i.e. no utilization capping
1044
1045        The requested maximum utilization (limit) as a percentage rational
1046        number, e.g. 98.76 for 98.76%.
1047
1048        This interface allows reading and setting maximum utilization clamp
1049        values similar to the sched_setattr(2). This maximum utilization
1050        value is used to clamp the task specific maximum utilization clamp.
1051
1052
1053
1054Memory
1055------
1056
1057The "memory" controller regulates distribution of memory.  Memory is
1058stateful and implements both limit and protection models.  Due to the
1059intertwining between memory usage and reclaim pressure and the
1060stateful nature of memory, the distribution model is relatively
1061complex.
1062
1063While not completely water-tight, all major memory usages by a given
1064cgroup are tracked so that the total memory consumption can be
1065accounted and controlled to a reasonable extent.  Currently, the
1066following types of memory usages are tracked.
1067
1068- Userland memory - page cache and anonymous memory.
1069
1070- Kernel data structures such as dentries and inodes.
1071
1072- TCP socket buffers.
1073
1074The above list may expand in the future for better coverage.
1075
1076
1077Memory Interface Files
1078~~~~~~~~~~~~~~~~~~~~~~
1079
1080All memory amounts are in bytes.  If a value which is not aligned to
1081PAGE_SIZE is written, the value may be rounded up to the closest
1082PAGE_SIZE multiple when read back.
1083
1084  memory.current
1085	A read-only single value file which exists on non-root
1086	cgroups.
1087
1088	The total amount of memory currently being used by the cgroup
1089	and its descendants.
1090
1091  memory.min
1092	A read-write single value file which exists on non-root
1093	cgroups.  The default is "0".
1094
1095	Hard memory protection.  If the memory usage of a cgroup
1096	is within its effective min boundary, the cgroup's memory
1097	won't be reclaimed under any conditions. If there is no
1098	unprotected reclaimable memory available, OOM killer
1099	is invoked.
1100
1101       Effective min boundary is limited by memory.min values of
1102	all ancestor cgroups. If there is memory.min overcommitment
1103	(child cgroup or cgroups are requiring more protected memory
1104	than parent will allow), then each child cgroup will get
1105	the part of parent's protection proportional to its
1106	actual memory usage below memory.min.
1107
1108	Putting more memory than generally available under this
1109	protection is discouraged and may lead to constant OOMs.
1110
1111	If a memory cgroup is not populated with processes,
1112	its memory.min is ignored.
1113
1114  memory.low
1115	A read-write single value file which exists on non-root
1116	cgroups.  The default is "0".
1117
1118	Best-effort memory protection.  If the memory usage of a
1119	cgroup is within its effective low boundary, the cgroup's
1120	memory won't be reclaimed unless memory can be reclaimed
1121	from unprotected cgroups.
1122
1123	Effective low boundary is limited by memory.low values of
1124	all ancestor cgroups. If there is memory.low overcommitment
1125	(child cgroup or cgroups are requiring more protected memory
1126	than parent will allow), then each child cgroup will get
1127	the part of parent's protection proportional to its
1128	actual memory usage below memory.low.
1129
1130	Putting more memory than generally available under this
1131	protection is discouraged.
1132
1133  memory.high
1134	A read-write single value file which exists on non-root
1135	cgroups.  The default is "max".
1136
1137	Memory usage throttle limit.  This is the main mechanism to
1138	control memory usage of a cgroup.  If a cgroup's usage goes
1139	over the high boundary, the processes of the cgroup are
1140	throttled and put under heavy reclaim pressure.
1141
1142	Going over the high limit never invokes the OOM killer and
1143	under extreme conditions the limit may be breached.
1144
1145  memory.max
1146	A read-write single value file which exists on non-root
1147	cgroups.  The default is "max".
1148
1149	Memory usage hard limit.  This is the final protection
1150	mechanism.  If a cgroup's memory usage reaches this limit and
1151	can't be reduced, the OOM killer is invoked in the cgroup.
1152	Under certain circumstances, the usage may go over the limit
1153	temporarily.
1154
1155	This is the ultimate protection mechanism.  As long as the
1156	high limit is used and monitored properly, this limit's
1157	utility is limited to providing the final safety net.
1158
1159  memory.oom.group
1160	A read-write single value file which exists on non-root
1161	cgroups.  The default value is "0".
1162
1163	Determines whether the cgroup should be treated as
1164	an indivisible workload by the OOM killer. If set,
1165	all tasks belonging to the cgroup or to its descendants
1166	(if the memory cgroup is not a leaf cgroup) are killed
1167	together or not at all. This can be used to avoid
1168	partial kills to guarantee workload integrity.
1169
1170	Tasks with the OOM protection (oom_score_adj set to -1000)
1171	are treated as an exception and are never killed.
1172
1173	If the OOM killer is invoked in a cgroup, it's not going
1174	to kill any tasks outside of this cgroup, regardless
1175	memory.oom.group values of ancestor cgroups.
1176
1177  memory.events
1178	A read-only flat-keyed file which exists on non-root cgroups.
1179	The following entries are defined.  Unless specified
1180	otherwise, a value change in this file generates a file
1181	modified event.
1182
1183	Note that all fields in this file are hierarchical and the
1184	file modified event can be generated due to an event down the
1185	hierarchy. For for the local events at the cgroup level see
1186	memory.events.local.
1187
1188	  low
1189		The number of times the cgroup is reclaimed due to
1190		high memory pressure even though its usage is under
1191		the low boundary.  This usually indicates that the low
1192		boundary is over-committed.
1193
1194	  high
1195		The number of times processes of the cgroup are
1196		throttled and routed to perform direct memory reclaim
1197		because the high memory boundary was exceeded.  For a
1198		cgroup whose memory usage is capped by the high limit
1199		rather than global memory pressure, this event's
1200		occurrences are expected.
1201
1202	  max
1203		The number of times the cgroup's memory usage was
1204		about to go over the max boundary.  If direct reclaim
1205		fails to bring it down, the cgroup goes to OOM state.
1206
1207	  oom
1208		The number of time the cgroup's memory usage was
1209		reached the limit and allocation was about to fail.
1210
1211		Depending on context result could be invocation of OOM
1212		killer and retrying allocation or failing allocation.
1213
1214		Failed allocation in its turn could be returned into
1215		userspace as -ENOMEM or silently ignored in cases like
1216		disk readahead.  For now OOM in memory cgroup kills
1217		tasks iff shortage has happened inside page fault.
1218
1219		This event is not raised if the OOM killer is not
1220		considered as an option, e.g. for failed high-order
1221		allocations.
1222
1223	  oom_kill
1224		The number of processes belonging to this cgroup
1225		killed by any kind of OOM killer.
1226
1227  memory.events.local
1228	Similar to memory.events but the fields in the file are local
1229	to the cgroup i.e. not hierarchical. The file modified event
1230	generated on this file reflects only the local events.
1231
1232  memory.stat
1233	A read-only flat-keyed file which exists on non-root cgroups.
1234
1235	This breaks down the cgroup's memory footprint into different
1236	types of memory, type-specific details, and other information
1237	on the state and past events of the memory management system.
1238
1239	All memory amounts are in bytes.
1240
1241	The entries are ordered to be human readable, and new entries
1242	can show up in the middle. Don't rely on items remaining in a
1243	fixed position; use the keys to look up specific values!
1244
1245	  anon
1246		Amount of memory used in anonymous mappings such as
1247		brk(), sbrk(), and mmap(MAP_ANONYMOUS)
1248
1249	  file
1250		Amount of memory used to cache filesystem data,
1251		including tmpfs and shared memory.
1252
1253	  kernel_stack
1254		Amount of memory allocated to kernel stacks.
1255
1256	  slab
1257		Amount of memory used for storing in-kernel data
1258		structures.
1259
1260	  sock
1261		Amount of memory used in network transmission buffers
1262
1263	  shmem
1264		Amount of cached filesystem data that is swap-backed,
1265		such as tmpfs, shm segments, shared anonymous mmap()s
1266
1267	  file_mapped
1268		Amount of cached filesystem data mapped with mmap()
1269
1270	  file_dirty
1271		Amount of cached filesystem data that was modified but
1272		not yet written back to disk
1273
1274	  file_writeback
1275		Amount of cached filesystem data that was modified and
1276		is currently being written back to disk
1277
1278	  anon_thp
1279		Amount of memory used in anonymous mappings backed by
1280		transparent hugepages
1281
1282	  inactive_anon, active_anon, inactive_file, active_file, unevictable
1283		Amount of memory, swap-backed and filesystem-backed,
1284		on the internal memory management lists used by the
1285		page reclaim algorithm
1286
1287	  slab_reclaimable
1288		Part of "slab" that might be reclaimed, such as
1289		dentries and inodes.
1290
1291	  slab_unreclaimable
1292		Part of "slab" that cannot be reclaimed on memory
1293		pressure.
1294
1295	  pgfault
1296		Total number of page faults incurred
1297
1298	  pgmajfault
1299		Number of major page faults incurred
1300
1301	  workingset_refault
1302
1303		Number of refaults of previously evicted pages
1304
1305	  workingset_activate
1306
1307		Number of refaulted pages that were immediately activated
1308
1309	  workingset_nodereclaim
1310
1311		Number of times a shadow node has been reclaimed
1312
1313	  pgrefill
1314
1315		Amount of scanned pages (in an active LRU list)
1316
1317	  pgscan
1318
1319		Amount of scanned pages (in an inactive LRU list)
1320
1321	  pgsteal
1322
1323		Amount of reclaimed pages
1324
1325	  pgactivate
1326
1327		Amount of pages moved to the active LRU list
1328
1329	  pgdeactivate
1330
1331		Amount of pages moved to the inactive LRU lis
1332
1333	  pglazyfree
1334
1335		Amount of pages postponed to be freed under memory pressure
1336
1337	  pglazyfreed
1338
1339		Amount of reclaimed lazyfree pages
1340
1341	  thp_fault_alloc
1342
1343		Number of transparent hugepages which were allocated to satisfy
1344		a page fault, including COW faults. This counter is not present
1345		when CONFIG_TRANSPARENT_HUGEPAGE is not set.
1346
1347	  thp_collapse_alloc
1348
1349		Number of transparent hugepages which were allocated to allow
1350		collapsing an existing range of pages. This counter is not
1351		present when CONFIG_TRANSPARENT_HUGEPAGE is not set.
1352
1353  memory.swap.current
1354	A read-only single value file which exists on non-root
1355	cgroups.
1356
1357	The total amount of swap currently being used by the cgroup
1358	and its descendants.
1359
1360  memory.swap.max
1361	A read-write single value file which exists on non-root
1362	cgroups.  The default is "max".
1363
1364	Swap usage hard limit.  If a cgroup's swap usage reaches this
1365	limit, anonymous memory of the cgroup will not be swapped out.
1366
1367  memory.swap.events
1368	A read-only flat-keyed file which exists on non-root cgroups.
1369	The following entries are defined.  Unless specified
1370	otherwise, a value change in this file generates a file
1371	modified event.
1372
1373	  max
1374		The number of times the cgroup's swap usage was about
1375		to go over the max boundary and swap allocation
1376		failed.
1377
1378	  fail
1379		The number of times swap allocation failed either
1380		because of running out of swap system-wide or max
1381		limit.
1382
1383	When reduced under the current usage, the existing swap
1384	entries are reclaimed gradually and the swap usage may stay
1385	higher than the limit for an extended period of time.  This
1386	reduces the impact on the workload and memory management.
1387
1388  memory.pressure
1389	A read-only nested-key file which exists on non-root cgroups.
1390
1391	Shows pressure stall information for memory. See
1392	Documentation/accounting/psi.rst for details.
1393
1394
1395Usage Guidelines
1396~~~~~~~~~~~~~~~~
1397
1398"memory.high" is the main mechanism to control memory usage.
1399Over-committing on high limit (sum of high limits > available memory)
1400and letting global memory pressure to distribute memory according to
1401usage is a viable strategy.
1402
1403Because breach of the high limit doesn't trigger the OOM killer but
1404throttles the offending cgroup, a management agent has ample
1405opportunities to monitor and take appropriate actions such as granting
1406more memory or terminating the workload.
1407
1408Determining whether a cgroup has enough memory is not trivial as
1409memory usage doesn't indicate whether the workload can benefit from
1410more memory.  For example, a workload which writes data received from
1411network to a file can use all available memory but can also operate as
1412performant with a small amount of memory.  A measure of memory
1413pressure - how much the workload is being impacted due to lack of
1414memory - is necessary to determine whether a workload needs more
1415memory; unfortunately, memory pressure monitoring mechanism isn't
1416implemented yet.
1417
1418
1419Memory Ownership
1420~~~~~~~~~~~~~~~~
1421
1422A memory area is charged to the cgroup which instantiated it and stays
1423charged to the cgroup until the area is released.  Migrating a process
1424to a different cgroup doesn't move the memory usages that it
1425instantiated while in the previous cgroup to the new cgroup.
1426
1427A memory area may be used by processes belonging to different cgroups.
1428To which cgroup the area will be charged is in-deterministic; however,
1429over time, the memory area is likely to end up in a cgroup which has
1430enough memory allowance to avoid high reclaim pressure.
1431
1432If a cgroup sweeps a considerable amount of memory which is expected
1433to be accessed repeatedly by other cgroups, it may make sense to use
1434POSIX_FADV_DONTNEED to relinquish the ownership of memory areas
1435belonging to the affected files to ensure correct memory ownership.
1436
1437
1438IO
1439--
1440
1441The "io" controller regulates the distribution of IO resources.  This
1442controller implements both weight based and absolute bandwidth or IOPS
1443limit distribution; however, weight based distribution is available
1444only if cfq-iosched is in use and neither scheme is available for
1445blk-mq devices.
1446
1447
1448IO Interface Files
1449~~~~~~~~~~~~~~~~~~
1450
1451  io.stat
1452	A read-only nested-keyed file which exists on non-root
1453	cgroups.
1454
1455	Lines are keyed by $MAJ:$MIN device numbers and not ordered.
1456	The following nested keys are defined.
1457
1458	  ======	=====================
1459	  rbytes	Bytes read
1460	  wbytes	Bytes written
1461	  rios		Number of read IOs
1462	  wios		Number of write IOs
1463	  dbytes	Bytes discarded
1464	  dios		Number of discard IOs
1465	  ======	=====================
1466
1467	An example read output follows:
1468
1469	  8:16 rbytes=1459200 wbytes=314773504 rios=192 wios=353 dbytes=0 dios=0
1470	  8:0 rbytes=90430464 wbytes=299008000 rios=8950 wios=1252 dbytes=50331648 dios=3021
1471
1472  io.cost.qos
1473	A read-write nested-keyed file with exists only on the root
1474	cgroup.
1475
1476	This file configures the Quality of Service of the IO cost
1477	model based controller (CONFIG_BLK_CGROUP_IOCOST) which
1478	currently implements "io.weight" proportional control.  Lines
1479	are keyed by $MAJ:$MIN device numbers and not ordered.  The
1480	line for a given device is populated on the first write for
1481	the device on "io.cost.qos" or "io.cost.model".  The following
1482	nested keys are defined.
1483
1484	  ======	=====================================
1485	  enable	Weight-based control enable
1486	  ctrl		"auto" or "user"
1487	  rpct		Read latency percentile    [0, 100]
1488	  rlat		Read latency threshold
1489	  wpct		Write latency percentile   [0, 100]
1490	  wlat		Write latency threshold
1491	  min		Minimum scaling percentage [1, 10000]
1492	  max		Maximum scaling percentage [1, 10000]
1493	  ======	=====================================
1494
1495	The controller is disabled by default and can be enabled by
1496	setting "enable" to 1.  "rpct" and "wpct" parameters default
1497	to zero and the controller uses internal device saturation
1498	state to adjust the overall IO rate between "min" and "max".
1499
1500	When a better control quality is needed, latency QoS
1501	parameters can be configured.  For example::
1502
1503	  8:16 enable=1 ctrl=auto rpct=95.00 rlat=75000 wpct=95.00 wlat=150000 min=50.00 max=150.0
1504
1505	shows that on sdb, the controller is enabled, will consider
1506	the device saturated if the 95th percentile of read completion
1507	latencies is above 75ms or write 150ms, and adjust the overall
1508	IO issue rate between 50% and 150% accordingly.
1509
1510	The lower the saturation point, the better the latency QoS at
1511	the cost of aggregate bandwidth.  The narrower the allowed
1512	adjustment range between "min" and "max", the more conformant
1513	to the cost model the IO behavior.  Note that the IO issue
1514	base rate may be far off from 100% and setting "min" and "max"
1515	blindly can lead to a significant loss of device capacity or
1516	control quality.  "min" and "max" are useful for regulating
1517	devices which show wide temporary behavior changes - e.g. a
1518	ssd which accepts writes at the line speed for a while and
1519	then completely stalls for multiple seconds.
1520
1521	When "ctrl" is "auto", the parameters are controlled by the
1522	kernel and may change automatically.  Setting "ctrl" to "user"
1523	or setting any of the percentile and latency parameters puts
1524	it into "user" mode and disables the automatic changes.  The
1525	automatic mode can be restored by setting "ctrl" to "auto".
1526
1527  io.cost.model
1528	A read-write nested-keyed file with exists only on the root
1529	cgroup.
1530
1531	This file configures the cost model of the IO cost model based
1532	controller (CONFIG_BLK_CGROUP_IOCOST) which currently
1533	implements "io.weight" proportional control.  Lines are keyed
1534	by $MAJ:$MIN device numbers and not ordered.  The line for a
1535	given device is populated on the first write for the device on
1536	"io.cost.qos" or "io.cost.model".  The following nested keys
1537	are defined.
1538
1539	  =====		================================
1540	  ctrl		"auto" or "user"
1541	  model		The cost model in use - "linear"
1542	  =====		================================
1543
1544	When "ctrl" is "auto", the kernel may change all parameters
1545	dynamically.  When "ctrl" is set to "user" or any other
1546	parameters are written to, "ctrl" become "user" and the
1547	automatic changes are disabled.
1548
1549	When "model" is "linear", the following model parameters are
1550	defined.
1551
1552	  =============	========================================
1553	  [r|w]bps	The maximum sequential IO throughput
1554	  [r|w]seqiops	The maximum 4k sequential IOs per second
1555	  [r|w]randiops	The maximum 4k random IOs per second
1556	  =============	========================================
1557
1558	From the above, the builtin linear model determines the base
1559	costs of a sequential and random IO and the cost coefficient
1560	for the IO size.  While simple, this model can cover most
1561	common device classes acceptably.
1562
1563	The IO cost model isn't expected to be accurate in absolute
1564	sense and is scaled to the device behavior dynamically.
1565
1566	If needed, tools/cgroup/iocost_coef_gen.py can be used to
1567	generate device-specific coefficients.
1568
1569  io.weight
1570	A read-write flat-keyed file which exists on non-root cgroups.
1571	The default is "default 100".
1572
1573	The first line is the default weight applied to devices
1574	without specific override.  The rest are overrides keyed by
1575	$MAJ:$MIN device numbers and not ordered.  The weights are in
1576	the range [1, 10000] and specifies the relative amount IO time
1577	the cgroup can use in relation to its siblings.
1578
1579	The default weight can be updated by writing either "default
1580	$WEIGHT" or simply "$WEIGHT".  Overrides can be set by writing
1581	"$MAJ:$MIN $WEIGHT" and unset by writing "$MAJ:$MIN default".
1582
1583	An example read output follows::
1584
1585	  default 100
1586	  8:16 200
1587	  8:0 50
1588
1589  io.max
1590	A read-write nested-keyed file which exists on non-root
1591	cgroups.
1592
1593	BPS and IOPS based IO limit.  Lines are keyed by $MAJ:$MIN
1594	device numbers and not ordered.  The following nested keys are
1595	defined.
1596
1597	  =====		==================================
1598	  rbps		Max read bytes per second
1599	  wbps		Max write bytes per second
1600	  riops		Max read IO operations per second
1601	  wiops		Max write IO operations per second
1602	  =====		==================================
1603
1604	When writing, any number of nested key-value pairs can be
1605	specified in any order.  "max" can be specified as the value
1606	to remove a specific limit.  If the same key is specified
1607	multiple times, the outcome is undefined.
1608
1609	BPS and IOPS are measured in each IO direction and IOs are
1610	delayed if limit is reached.  Temporary bursts are allowed.
1611
1612	Setting read limit at 2M BPS and write at 120 IOPS for 8:16::
1613
1614	  echo "8:16 rbps=2097152 wiops=120" > io.max
1615
1616	Reading returns the following::
1617
1618	  8:16 rbps=2097152 wbps=max riops=max wiops=120
1619
1620	Write IOPS limit can be removed by writing the following::
1621
1622	  echo "8:16 wiops=max" > io.max
1623
1624	Reading now returns the following::
1625
1626	  8:16 rbps=2097152 wbps=max riops=max wiops=max
1627
1628  io.pressure
1629	A read-only nested-key file which exists on non-root cgroups.
1630
1631	Shows pressure stall information for IO. See
1632	Documentation/accounting/psi.rst for details.
1633
1634
1635Writeback
1636~~~~~~~~~
1637
1638Page cache is dirtied through buffered writes and shared mmaps and
1639written asynchronously to the backing filesystem by the writeback
1640mechanism.  Writeback sits between the memory and IO domains and
1641regulates the proportion of dirty memory by balancing dirtying and
1642write IOs.
1643
1644The io controller, in conjunction with the memory controller,
1645implements control of page cache writeback IOs.  The memory controller
1646defines the memory domain that dirty memory ratio is calculated and
1647maintained for and the io controller defines the io domain which
1648writes out dirty pages for the memory domain.  Both system-wide and
1649per-cgroup dirty memory states are examined and the more restrictive
1650of the two is enforced.
1651
1652cgroup writeback requires explicit support from the underlying
1653filesystem.  Currently, cgroup writeback is implemented on ext2, ext4
1654and btrfs.  On other filesystems, all writeback IOs are attributed to
1655the root cgroup.
1656
1657There are inherent differences in memory and writeback management
1658which affects how cgroup ownership is tracked.  Memory is tracked per
1659page while writeback per inode.  For the purpose of writeback, an
1660inode is assigned to a cgroup and all IO requests to write dirty pages
1661from the inode are attributed to that cgroup.
1662
1663As cgroup ownership for memory is tracked per page, there can be pages
1664which are associated with different cgroups than the one the inode is
1665associated with.  These are called foreign pages.  The writeback
1666constantly keeps track of foreign pages and, if a particular foreign
1667cgroup becomes the majority over a certain period of time, switches
1668the ownership of the inode to that cgroup.
1669
1670While this model is enough for most use cases where a given inode is
1671mostly dirtied by a single cgroup even when the main writing cgroup
1672changes over time, use cases where multiple cgroups write to a single
1673inode simultaneously are not supported well.  In such circumstances, a
1674significant portion of IOs are likely to be attributed incorrectly.
1675As memory controller assigns page ownership on the first use and
1676doesn't update it until the page is released, even if writeback
1677strictly follows page ownership, multiple cgroups dirtying overlapping
1678areas wouldn't work as expected.  It's recommended to avoid such usage
1679patterns.
1680
1681The sysctl knobs which affect writeback behavior are applied to cgroup
1682writeback as follows.
1683
1684  vm.dirty_background_ratio, vm.dirty_ratio
1685	These ratios apply the same to cgroup writeback with the
1686	amount of available memory capped by limits imposed by the
1687	memory controller and system-wide clean memory.
1688
1689  vm.dirty_background_bytes, vm.dirty_bytes
1690	For cgroup writeback, this is calculated into ratio against
1691	total available memory and applied the same way as
1692	vm.dirty[_background]_ratio.
1693
1694
1695IO Latency
1696~~~~~~~~~~
1697
1698This is a cgroup v2 controller for IO workload protection.  You provide a group
1699with a latency target, and if the average latency exceeds that target the
1700controller will throttle any peers that have a lower latency target than the
1701protected workload.
1702
1703The limits are only applied at the peer level in the hierarchy.  This means that
1704in the diagram below, only groups A, B, and C will influence each other, and
1705groups D and F will influence each other.  Group G will influence nobody::
1706
1707			[root]
1708		/	   |		\
1709		A	   B		C
1710	       /  \        |
1711	      D    F	   G
1712
1713
1714So the ideal way to configure this is to set io.latency in groups A, B, and C.
1715Generally you do not want to set a value lower than the latency your device
1716supports.  Experiment to find the value that works best for your workload.
1717Start at higher than the expected latency for your device and watch the
1718avg_lat value in io.stat for your workload group to get an idea of the
1719latency you see during normal operation.  Use the avg_lat value as a basis for
1720your real setting, setting at 10-15% higher than the value in io.stat.
1721
1722How IO Latency Throttling Works
1723~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
1724
1725io.latency is work conserving; so as long as everybody is meeting their latency
1726target the controller doesn't do anything.  Once a group starts missing its
1727target it begins throttling any peer group that has a higher target than itself.
1728This throttling takes 2 forms:
1729
1730- Queue depth throttling.  This is the number of outstanding IO's a group is
1731  allowed to have.  We will clamp down relatively quickly, starting at no limit
1732  and going all the way down to 1 IO at a time.
1733
1734- Artificial delay induction.  There are certain types of IO that cannot be
1735  throttled without possibly adversely affecting higher priority groups.  This
1736  includes swapping and metadata IO.  These types of IO are allowed to occur
1737  normally, however they are "charged" to the originating group.  If the
1738  originating group is being throttled you will see the use_delay and delay
1739  fields in io.stat increase.  The delay value is how many microseconds that are
1740  being added to any process that runs in this group.  Because this number can
1741  grow quite large if there is a lot of swapping or metadata IO occurring we
1742  limit the individual delay events to 1 second at a time.
1743
1744Once the victimized group starts meeting its latency target again it will start
1745unthrottling any peer groups that were throttled previously.  If the victimized
1746group simply stops doing IO the global counter will unthrottle appropriately.
1747
1748IO Latency Interface Files
1749~~~~~~~~~~~~~~~~~~~~~~~~~~
1750
1751  io.latency
1752	This takes a similar format as the other controllers.
1753
1754		"MAJOR:MINOR target=<target time in microseconds"
1755
1756  io.stat
1757	If the controller is enabled you will see extra stats in io.stat in
1758	addition to the normal ones.
1759
1760	  depth
1761		This is the current queue depth for the group.
1762
1763	  avg_lat
1764		This is an exponential moving average with a decay rate of 1/exp
1765		bound by the sampling interval.  The decay rate interval can be
1766		calculated by multiplying the win value in io.stat by the
1767		corresponding number of samples based on the win value.
1768
1769	  win
1770		The sampling window size in milliseconds.  This is the minimum
1771		duration of time between evaluation events.  Windows only elapse
1772		with IO activity.  Idle periods extend the most recent window.
1773
1774PID
1775---
1776
1777The process number controller is used to allow a cgroup to stop any
1778new tasks from being fork()'d or clone()'d after a specified limit is
1779reached.
1780
1781The number of tasks in a cgroup can be exhausted in ways which other
1782controllers cannot prevent, thus warranting its own controller.  For
1783example, a fork bomb is likely to exhaust the number of tasks before
1784hitting memory restrictions.
1785
1786Note that PIDs used in this controller refer to TIDs, process IDs as
1787used by the kernel.
1788
1789
1790PID Interface Files
1791~~~~~~~~~~~~~~~~~~~
1792
1793  pids.max
1794	A read-write single value file which exists on non-root
1795	cgroups.  The default is "max".
1796
1797	Hard limit of number of processes.
1798
1799  pids.current
1800	A read-only single value file which exists on all cgroups.
1801
1802	The number of processes currently in the cgroup and its
1803	descendants.
1804
1805Organisational operations are not blocked by cgroup policies, so it is
1806possible to have pids.current > pids.max.  This can be done by either
1807setting the limit to be smaller than pids.current, or attaching enough
1808processes to the cgroup such that pids.current is larger than
1809pids.max.  However, it is not possible to violate a cgroup PID policy
1810through fork() or clone(). These will return -EAGAIN if the creation
1811of a new process would cause a cgroup policy to be violated.
1812
1813
1814Cpuset
1815------
1816
1817The "cpuset" controller provides a mechanism for constraining
1818the CPU and memory node placement of tasks to only the resources
1819specified in the cpuset interface files in a task's current cgroup.
1820This is especially valuable on large NUMA systems where placing jobs
1821on properly sized subsets of the systems with careful processor and
1822memory placement to reduce cross-node memory access and contention
1823can improve overall system performance.
1824
1825The "cpuset" controller is hierarchical.  That means the controller
1826cannot use CPUs or memory nodes not allowed in its parent.
1827
1828
1829Cpuset Interface Files
1830~~~~~~~~~~~~~~~~~~~~~~
1831
1832  cpuset.cpus
1833	A read-write multiple values file which exists on non-root
1834	cpuset-enabled cgroups.
1835
1836	It lists the requested CPUs to be used by tasks within this
1837	cgroup.  The actual list of CPUs to be granted, however, is
1838	subjected to constraints imposed by its parent and can differ
1839	from the requested CPUs.
1840
1841	The CPU numbers are comma-separated numbers or ranges.
1842	For example:
1843
1844	  # cat cpuset.cpus
1845	  0-4,6,8-10
1846
1847	An empty value indicates that the cgroup is using the same
1848	setting as the nearest cgroup ancestor with a non-empty
1849	"cpuset.cpus" or all the available CPUs if none is found.
1850
1851	The value of "cpuset.cpus" stays constant until the next update
1852	and won't be affected by any CPU hotplug events.
1853
1854  cpuset.cpus.effective
1855	A read-only multiple values file which exists on all
1856	cpuset-enabled cgroups.
1857
1858	It lists the onlined CPUs that are actually granted to this
1859	cgroup by its parent.  These CPUs are allowed to be used by
1860	tasks within the current cgroup.
1861
1862	If "cpuset.cpus" is empty, the "cpuset.cpus.effective" file shows
1863	all the CPUs from the parent cgroup that can be available to
1864	be used by this cgroup.  Otherwise, it should be a subset of
1865	"cpuset.cpus" unless none of the CPUs listed in "cpuset.cpus"
1866	can be granted.  In this case, it will be treated just like an
1867	empty "cpuset.cpus".
1868
1869	Its value will be affected by CPU hotplug events.
1870
1871  cpuset.mems
1872	A read-write multiple values file which exists on non-root
1873	cpuset-enabled cgroups.
1874
1875	It lists the requested memory nodes to be used by tasks within
1876	this cgroup.  The actual list of memory nodes granted, however,
1877	is subjected to constraints imposed by its parent and can differ
1878	from the requested memory nodes.
1879
1880	The memory node numbers are comma-separated numbers or ranges.
1881	For example:
1882
1883	  # cat cpuset.mems
1884	  0-1,3
1885
1886	An empty value indicates that the cgroup is using the same
1887	setting as the nearest cgroup ancestor with a non-empty
1888	"cpuset.mems" or all the available memory nodes if none
1889	is found.
1890
1891	The value of "cpuset.mems" stays constant until the next update
1892	and won't be affected by any memory nodes hotplug events.
1893
1894  cpuset.mems.effective
1895	A read-only multiple values file which exists on all
1896	cpuset-enabled cgroups.
1897
1898	It lists the onlined memory nodes that are actually granted to
1899	this cgroup by its parent. These memory nodes are allowed to
1900	be used by tasks within the current cgroup.
1901
1902	If "cpuset.mems" is empty, it shows all the memory nodes from the
1903	parent cgroup that will be available to be used by this cgroup.
1904	Otherwise, it should be a subset of "cpuset.mems" unless none of
1905	the memory nodes listed in "cpuset.mems" can be granted.  In this
1906	case, it will be treated just like an empty "cpuset.mems".
1907
1908	Its value will be affected by memory nodes hotplug events.
1909
1910  cpuset.cpus.partition
1911	A read-write single value file which exists on non-root
1912	cpuset-enabled cgroups.  This flag is owned by the parent cgroup
1913	and is not delegatable.
1914
1915        It accepts only the following input values when written to.
1916
1917        "root"   - a paritition root
1918        "member" - a non-root member of a partition
1919
1920	When set to be a partition root, the current cgroup is the
1921	root of a new partition or scheduling domain that comprises
1922	itself and all its descendants except those that are separate
1923	partition roots themselves and their descendants.  The root
1924	cgroup is always a partition root.
1925
1926	There are constraints on where a partition root can be set.
1927	It can only be set in a cgroup if all the following conditions
1928	are true.
1929
1930	1) The "cpuset.cpus" is not empty and the list of CPUs are
1931	   exclusive, i.e. they are not shared by any of its siblings.
1932	2) The parent cgroup is a partition root.
1933	3) The "cpuset.cpus" is also a proper subset of the parent's
1934	   "cpuset.cpus.effective".
1935	4) There is no child cgroups with cpuset enabled.  This is for
1936	   eliminating corner cases that have to be handled if such a
1937	   condition is allowed.
1938
1939	Setting it to partition root will take the CPUs away from the
1940	effective CPUs of the parent cgroup.  Once it is set, this
1941	file cannot be reverted back to "member" if there are any child
1942	cgroups with cpuset enabled.
1943
1944	A parent partition cannot distribute all its CPUs to its
1945	child partitions.  There must be at least one cpu left in the
1946	parent partition.
1947
1948	Once becoming a partition root, changes to "cpuset.cpus" is
1949	generally allowed as long as the first condition above is true,
1950	the change will not take away all the CPUs from the parent
1951	partition and the new "cpuset.cpus" value is a superset of its
1952	children's "cpuset.cpus" values.
1953
1954	Sometimes, external factors like changes to ancestors'
1955	"cpuset.cpus" or cpu hotplug can cause the state of the partition
1956	root to change.  On read, the "cpuset.sched.partition" file
1957	can show the following values.
1958
1959	"member"       Non-root member of a partition
1960	"root"         Partition root
1961	"root invalid" Invalid partition root
1962
1963	It is a partition root if the first 2 partition root conditions
1964	above are true and at least one CPU from "cpuset.cpus" is
1965	granted by the parent cgroup.
1966
1967	A partition root can become invalid if none of CPUs requested
1968	in "cpuset.cpus" can be granted by the parent cgroup or the
1969	parent cgroup is no longer a partition root itself.  In this
1970	case, it is not a real partition even though the restriction
1971	of the first partition root condition above will still apply.
1972	The cpu affinity of all the tasks in the cgroup will then be
1973	associated with CPUs in the nearest ancestor partition.
1974
1975	An invalid partition root can be transitioned back to a
1976	real partition root if at least one of the requested CPUs
1977	can now be granted by its parent.  In this case, the cpu
1978	affinity of all the tasks in the formerly invalid partition
1979	will be associated to the CPUs of the newly formed partition.
1980	Changing the partition state of an invalid partition root to
1981	"member" is always allowed even if child cpusets are present.
1982
1983
1984Device controller
1985-----------------
1986
1987Device controller manages access to device files. It includes both
1988creation of new device files (using mknod), and access to the
1989existing device files.
1990
1991Cgroup v2 device controller has no interface files and is implemented
1992on top of cgroup BPF. To control access to device files, a user may
1993create bpf programs of the BPF_CGROUP_DEVICE type and attach them
1994to cgroups. On an attempt to access a device file, corresponding
1995BPF programs will be executed, and depending on the return value
1996the attempt will succeed or fail with -EPERM.
1997
1998A BPF_CGROUP_DEVICE program takes a pointer to the bpf_cgroup_dev_ctx
1999structure, which describes the device access attempt: access type
2000(mknod/read/write) and device (type, major and minor numbers).
2001If the program returns 0, the attempt fails with -EPERM, otherwise
2002it succeeds.
2003
2004An example of BPF_CGROUP_DEVICE program may be found in the kernel
2005source tree in the tools/testing/selftests/bpf/dev_cgroup.c file.
2006
2007
2008RDMA
2009----
2010
2011The "rdma" controller regulates the distribution and accounting of
2012of RDMA resources.
2013
2014RDMA Interface Files
2015~~~~~~~~~~~~~~~~~~~~
2016
2017  rdma.max
2018	A readwrite nested-keyed file that exists for all the cgroups
2019	except root that describes current configured resource limit
2020	for a RDMA/IB device.
2021
2022	Lines are keyed by device name and are not ordered.
2023	Each line contains space separated resource name and its configured
2024	limit that can be distributed.
2025
2026	The following nested keys are defined.
2027
2028	  ==========	=============================
2029	  hca_handle	Maximum number of HCA Handles
2030	  hca_object 	Maximum number of HCA Objects
2031	  ==========	=============================
2032
2033	An example for mlx4 and ocrdma device follows::
2034
2035	  mlx4_0 hca_handle=2 hca_object=2000
2036	  ocrdma1 hca_handle=3 hca_object=max
2037
2038  rdma.current
2039	A read-only file that describes current resource usage.
2040	It exists for all the cgroup except root.
2041
2042	An example for mlx4 and ocrdma device follows::
2043
2044	  mlx4_0 hca_handle=1 hca_object=20
2045	  ocrdma1 hca_handle=1 hca_object=23
2046
2047
2048Misc
2049----
2050
2051perf_event
2052~~~~~~~~~~
2053
2054perf_event controller, if not mounted on a legacy hierarchy, is
2055automatically enabled on the v2 hierarchy so that perf events can
2056always be filtered by cgroup v2 path.  The controller can still be
2057moved to a legacy hierarchy after v2 hierarchy is populated.
2058
2059
2060Non-normative information
2061-------------------------
2062
2063This section contains information that isn't considered to be a part of
2064the stable kernel API and so is subject to change.
2065
2066
2067CPU controller root cgroup process behaviour
2068~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
2069
2070When distributing CPU cycles in the root cgroup each thread in this
2071cgroup is treated as if it was hosted in a separate child cgroup of the
2072root cgroup. This child cgroup weight is dependent on its thread nice
2073level.
2074
2075For details of this mapping see sched_prio_to_weight array in
2076kernel/sched/core.c file (values from this array should be scaled
2077appropriately so the neutral - nice 0 - value is 100 instead of 1024).
2078
2079
2080IO controller root cgroup process behaviour
2081~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
2082
2083Root cgroup processes are hosted in an implicit leaf child node.
2084When distributing IO resources this implicit child node is taken into
2085account as if it was a normal child cgroup of the root cgroup with a
2086weight value of 200.
2087
2088
2089Namespace
2090=========
2091
2092Basics
2093------
2094
2095cgroup namespace provides a mechanism to virtualize the view of the
2096"/proc/$PID/cgroup" file and cgroup mounts.  The CLONE_NEWCGROUP clone
2097flag can be used with clone(2) and unshare(2) to create a new cgroup
2098namespace.  The process running inside the cgroup namespace will have
2099its "/proc/$PID/cgroup" output restricted to cgroupns root.  The
2100cgroupns root is the cgroup of the process at the time of creation of
2101the cgroup namespace.
2102
2103Without cgroup namespace, the "/proc/$PID/cgroup" file shows the
2104complete path of the cgroup of a process.  In a container setup where
2105a set of cgroups and namespaces are intended to isolate processes the
2106"/proc/$PID/cgroup" file may leak potential system level information
2107to the isolated processes.  For Example::
2108
2109  # cat /proc/self/cgroup
2110  0::/batchjobs/container_id1
2111
2112The path '/batchjobs/container_id1' can be considered as system-data
2113and undesirable to expose to the isolated processes.  cgroup namespace
2114can be used to restrict visibility of this path.  For example, before
2115creating a cgroup namespace, one would see::
2116
2117  # ls -l /proc/self/ns/cgroup
2118  lrwxrwxrwx 1 root root 0 2014-07-15 10:37 /proc/self/ns/cgroup -> cgroup:[4026531835]
2119  # cat /proc/self/cgroup
2120  0::/batchjobs/container_id1
2121
2122After unsharing a new namespace, the view changes::
2123
2124  # ls -l /proc/self/ns/cgroup
2125  lrwxrwxrwx 1 root root 0 2014-07-15 10:35 /proc/self/ns/cgroup -> cgroup:[4026532183]
2126  # cat /proc/self/cgroup
2127  0::/
2128
2129When some thread from a multi-threaded process unshares its cgroup
2130namespace, the new cgroupns gets applied to the entire process (all
2131the threads).  This is natural for the v2 hierarchy; however, for the
2132legacy hierarchies, this may be unexpected.
2133
2134A cgroup namespace is alive as long as there are processes inside or
2135mounts pinning it.  When the last usage goes away, the cgroup
2136namespace is destroyed.  The cgroupns root and the actual cgroups
2137remain.
2138
2139
2140The Root and Views
2141------------------
2142
2143The 'cgroupns root' for a cgroup namespace is the cgroup in which the
2144process calling unshare(2) is running.  For example, if a process in
2145/batchjobs/container_id1 cgroup calls unshare, cgroup
2146/batchjobs/container_id1 becomes the cgroupns root.  For the
2147init_cgroup_ns, this is the real root ('/') cgroup.
2148
2149The cgroupns root cgroup does not change even if the namespace creator
2150process later moves to a different cgroup::
2151
2152  # ~/unshare -c # unshare cgroupns in some cgroup
2153  # cat /proc/self/cgroup
2154  0::/
2155  # mkdir sub_cgrp_1
2156  # echo 0 > sub_cgrp_1/cgroup.procs
2157  # cat /proc/self/cgroup
2158  0::/sub_cgrp_1
2159
2160Each process gets its namespace-specific view of "/proc/$PID/cgroup"
2161
2162Processes running inside the cgroup namespace will be able to see
2163cgroup paths (in /proc/self/cgroup) only inside their root cgroup.
2164From within an unshared cgroupns::
2165
2166  # sleep 100000 &
2167  [1] 7353
2168  # echo 7353 > sub_cgrp_1/cgroup.procs
2169  # cat /proc/7353/cgroup
2170  0::/sub_cgrp_1
2171
2172From the initial cgroup namespace, the real cgroup path will be
2173visible::
2174
2175  $ cat /proc/7353/cgroup
2176  0::/batchjobs/container_id1/sub_cgrp_1
2177
2178From a sibling cgroup namespace (that is, a namespace rooted at a
2179different cgroup), the cgroup path relative to its own cgroup
2180namespace root will be shown.  For instance, if PID 7353's cgroup
2181namespace root is at '/batchjobs/container_id2', then it will see::
2182
2183  # cat /proc/7353/cgroup
2184  0::/../container_id2/sub_cgrp_1
2185
2186Note that the relative path always starts with '/' to indicate that
2187its relative to the cgroup namespace root of the caller.
2188
2189
2190Migration and setns(2)
2191----------------------
2192
2193Processes inside a cgroup namespace can move into and out of the
2194namespace root if they have proper access to external cgroups.  For
2195example, from inside a namespace with cgroupns root at
2196/batchjobs/container_id1, and assuming that the global hierarchy is
2197still accessible inside cgroupns::
2198
2199  # cat /proc/7353/cgroup
2200  0::/sub_cgrp_1
2201  # echo 7353 > batchjobs/container_id2/cgroup.procs
2202  # cat /proc/7353/cgroup
2203  0::/../container_id2
2204
2205Note that this kind of setup is not encouraged.  A task inside cgroup
2206namespace should only be exposed to its own cgroupns hierarchy.
2207
2208setns(2) to another cgroup namespace is allowed when:
2209
2210(a) the process has CAP_SYS_ADMIN against its current user namespace
2211(b) the process has CAP_SYS_ADMIN against the target cgroup
2212    namespace's userns
2213
2214No implicit cgroup changes happen with attaching to another cgroup
2215namespace.  It is expected that the someone moves the attaching
2216process under the target cgroup namespace root.
2217
2218
2219Interaction with Other Namespaces
2220---------------------------------
2221
2222Namespace specific cgroup hierarchy can be mounted by a process
2223running inside a non-init cgroup namespace::
2224
2225  # mount -t cgroup2 none $MOUNT_POINT
2226
2227This will mount the unified cgroup hierarchy with cgroupns root as the
2228filesystem root.  The process needs CAP_SYS_ADMIN against its user and
2229mount namespaces.
2230
2231The virtualization of /proc/self/cgroup file combined with restricting
2232the view of cgroup hierarchy by namespace-private cgroupfs mount
2233provides a properly isolated cgroup view inside the container.
2234
2235
2236Information on Kernel Programming
2237=================================
2238
2239This section contains kernel programming information in the areas
2240where interacting with cgroup is necessary.  cgroup core and
2241controllers are not covered.
2242
2243
2244Filesystem Support for Writeback
2245--------------------------------
2246
2247A filesystem can support cgroup writeback by updating
2248address_space_operations->writepage[s]() to annotate bio's using the
2249following two functions.
2250
2251  wbc_init_bio(@wbc, @bio)
2252	Should be called for each bio carrying writeback data and
2253	associates the bio with the inode's owner cgroup and the
2254	corresponding request queue.  This must be called after
2255	a queue (device) has been associated with the bio and
2256	before submission.
2257
2258  wbc_account_cgroup_owner(@wbc, @page, @bytes)
2259	Should be called for each data segment being written out.
2260	While this function doesn't care exactly when it's called
2261	during the writeback session, it's the easiest and most
2262	natural to call it as data segments are added to a bio.
2263
2264With writeback bio's annotated, cgroup support can be enabled per
2265super_block by setting SB_I_CGROUPWB in ->s_iflags.  This allows for
2266selective disabling of cgroup writeback support which is helpful when
2267certain filesystem features, e.g. journaled data mode, are
2268incompatible.
2269
2270wbc_init_bio() binds the specified bio to its cgroup.  Depending on
2271the configuration, the bio may be executed at a lower priority and if
2272the writeback session is holding shared resources, e.g. a journal
2273entry, may lead to priority inversion.  There is no one easy solution
2274for the problem.  Filesystems can try to work around specific problem
2275cases by skipping wbc_init_bio() and using bio_associate_blkg()
2276directly.
2277
2278
2279Deprecated v1 Core Features
2280===========================
2281
2282- Multiple hierarchies including named ones are not supported.
2283
2284- All v1 mount options are not supported.
2285
2286- The "tasks" file is removed and "cgroup.procs" is not sorted.
2287
2288- "cgroup.clone_children" is removed.
2289
2290- /proc/cgroups is meaningless for v2.  Use "cgroup.controllers" file
2291  at the root instead.
2292
2293
2294Issues with v1 and Rationales for v2
2295====================================
2296
2297Multiple Hierarchies
2298--------------------
2299
2300cgroup v1 allowed an arbitrary number of hierarchies and each
2301hierarchy could host any number of controllers.  While this seemed to
2302provide a high level of flexibility, it wasn't useful in practice.
2303
2304For example, as there is only one instance of each controller, utility
2305type controllers such as freezer which can be useful in all
2306hierarchies could only be used in one.  The issue is exacerbated by
2307the fact that controllers couldn't be moved to another hierarchy once
2308hierarchies were populated.  Another issue was that all controllers
2309bound to a hierarchy were forced to have exactly the same view of the
2310hierarchy.  It wasn't possible to vary the granularity depending on
2311the specific controller.
2312
2313In practice, these issues heavily limited which controllers could be
2314put on the same hierarchy and most configurations resorted to putting
2315each controller on its own hierarchy.  Only closely related ones, such
2316as the cpu and cpuacct controllers, made sense to be put on the same
2317hierarchy.  This often meant that userland ended up managing multiple
2318similar hierarchies repeating the same steps on each hierarchy
2319whenever a hierarchy management operation was necessary.
2320
2321Furthermore, support for multiple hierarchies came at a steep cost.
2322It greatly complicated cgroup core implementation but more importantly
2323the support for multiple hierarchies restricted how cgroup could be
2324used in general and what controllers was able to do.
2325
2326There was no limit on how many hierarchies there might be, which meant
2327that a thread's cgroup membership couldn't be described in finite
2328length.  The key might contain any number of entries and was unlimited
2329in length, which made it highly awkward to manipulate and led to
2330addition of controllers which existed only to identify membership,
2331which in turn exacerbated the original problem of proliferating number
2332of hierarchies.
2333
2334Also, as a controller couldn't have any expectation regarding the
2335topologies of hierarchies other controllers might be on, each
2336controller had to assume that all other controllers were attached to
2337completely orthogonal hierarchies.  This made it impossible, or at
2338least very cumbersome, for controllers to cooperate with each other.
2339
2340In most use cases, putting controllers on hierarchies which are
2341completely orthogonal to each other isn't necessary.  What usually is
2342called for is the ability to have differing levels of granularity
2343depending on the specific controller.  In other words, hierarchy may
2344be collapsed from leaf towards root when viewed from specific
2345controllers.  For example, a given configuration might not care about
2346how memory is distributed beyond a certain level while still wanting
2347to control how CPU cycles are distributed.
2348
2349
2350Thread Granularity
2351------------------
2352
2353cgroup v1 allowed threads of a process to belong to different cgroups.
2354This didn't make sense for some controllers and those controllers
2355ended up implementing different ways to ignore such situations but
2356much more importantly it blurred the line between API exposed to
2357individual applications and system management interface.
2358
2359Generally, in-process knowledge is available only to the process
2360itself; thus, unlike service-level organization of processes,
2361categorizing threads of a process requires active participation from
2362the application which owns the target process.
2363
2364cgroup v1 had an ambiguously defined delegation model which got abused
2365in combination with thread granularity.  cgroups were delegated to
2366individual applications so that they can create and manage their own
2367sub-hierarchies and control resource distributions along them.  This
2368effectively raised cgroup to the status of a syscall-like API exposed
2369to lay programs.
2370
2371First of all, cgroup has a fundamentally inadequate interface to be
2372exposed this way.  For a process to access its own knobs, it has to
2373extract the path on the target hierarchy from /proc/self/cgroup,
2374construct the path by appending the name of the knob to the path, open
2375and then read and/or write to it.  This is not only extremely clunky
2376and unusual but also inherently racy.  There is no conventional way to
2377define transaction across the required steps and nothing can guarantee
2378that the process would actually be operating on its own sub-hierarchy.
2379
2380cgroup controllers implemented a number of knobs which would never be
2381accepted as public APIs because they were just adding control knobs to
2382system-management pseudo filesystem.  cgroup ended up with interface
2383knobs which were not properly abstracted or refined and directly
2384revealed kernel internal details.  These knobs got exposed to
2385individual applications through the ill-defined delegation mechanism
2386effectively abusing cgroup as a shortcut to implementing public APIs
2387without going through the required scrutiny.
2388
2389This was painful for both userland and kernel.  Userland ended up with
2390misbehaving and poorly abstracted interfaces and kernel exposing and
2391locked into constructs inadvertently.
2392
2393
2394Competition Between Inner Nodes and Threads
2395-------------------------------------------
2396
2397cgroup v1 allowed threads to be in any cgroups which created an
2398interesting problem where threads belonging to a parent cgroup and its
2399children cgroups competed for resources.  This was nasty as two
2400different types of entities competed and there was no obvious way to
2401settle it.  Different controllers did different things.
2402
2403The cpu controller considered threads and cgroups as equivalents and
2404mapped nice levels to cgroup weights.  This worked for some cases but
2405fell flat when children wanted to be allocated specific ratios of CPU
2406cycles and the number of internal threads fluctuated - the ratios
2407constantly changed as the number of competing entities fluctuated.
2408There also were other issues.  The mapping from nice level to weight
2409wasn't obvious or universal, and there were various other knobs which
2410simply weren't available for threads.
2411
2412The io controller implicitly created a hidden leaf node for each
2413cgroup to host the threads.  The hidden leaf had its own copies of all
2414the knobs with ``leaf_`` prefixed.  While this allowed equivalent
2415control over internal threads, it was with serious drawbacks.  It
2416always added an extra layer of nesting which wouldn't be necessary
2417otherwise, made the interface messy and significantly complicated the
2418implementation.
2419
2420The memory controller didn't have a way to control what happened
2421between internal tasks and child cgroups and the behavior was not
2422clearly defined.  There were attempts to add ad-hoc behaviors and
2423knobs to tailor the behavior to specific workloads which would have
2424led to problems extremely difficult to resolve in the long term.
2425
2426Multiple controllers struggled with internal tasks and came up with
2427different ways to deal with it; unfortunately, all the approaches were
2428severely flawed and, furthermore, the widely different behaviors
2429made cgroup as a whole highly inconsistent.
2430
2431This clearly is a problem which needs to be addressed from cgroup core
2432in a uniform way.
2433
2434
2435Other Interface Issues
2436----------------------
2437
2438cgroup v1 grew without oversight and developed a large number of
2439idiosyncrasies and inconsistencies.  One issue on the cgroup core side
2440was how an empty cgroup was notified - a userland helper binary was
2441forked and executed for each event.  The event delivery wasn't
2442recursive or delegatable.  The limitations of the mechanism also led
2443to in-kernel event delivery filtering mechanism further complicating
2444the interface.
2445
2446Controller interfaces were problematic too.  An extreme example is
2447controllers completely ignoring hierarchical organization and treating
2448all cgroups as if they were all located directly under the root
2449cgroup.  Some controllers exposed a large amount of inconsistent
2450implementation details to userland.
2451
2452There also was no consistency across controllers.  When a new cgroup
2453was created, some controllers defaulted to not imposing extra
2454restrictions while others disallowed any resource usage until
2455explicitly configured.  Configuration knobs for the same type of
2456control used widely differing naming schemes and formats.  Statistics
2457and information knobs were named arbitrarily and used different
2458formats and units even in the same controller.
2459
2460cgroup v2 establishes common conventions where appropriate and updates
2461controllers so that they expose minimal and consistent interfaces.
2462
2463
2464Controller Issues and Remedies
2465------------------------------
2466
2467Memory
2468~~~~~~
2469
2470The original lower boundary, the soft limit, is defined as a limit
2471that is per default unset.  As a result, the set of cgroups that
2472global reclaim prefers is opt-in, rather than opt-out.  The costs for
2473optimizing these mostly negative lookups are so high that the
2474implementation, despite its enormous size, does not even provide the
2475basic desirable behavior.  First off, the soft limit has no
2476hierarchical meaning.  All configured groups are organized in a global
2477rbtree and treated like equal peers, regardless where they are located
2478in the hierarchy.  This makes subtree delegation impossible.  Second,
2479the soft limit reclaim pass is so aggressive that it not just
2480introduces high allocation latencies into the system, but also impacts
2481system performance due to overreclaim, to the point where the feature
2482becomes self-defeating.
2483
2484The memory.low boundary on the other hand is a top-down allocated
2485reserve.  A cgroup enjoys reclaim protection when it's within its low,
2486which makes delegation of subtrees possible.
2487
2488The original high boundary, the hard limit, is defined as a strict
2489limit that can not budge, even if the OOM killer has to be called.
2490But this generally goes against the goal of making the most out of the
2491available memory.  The memory consumption of workloads varies during
2492runtime, and that requires users to overcommit.  But doing that with a
2493strict upper limit requires either a fairly accurate prediction of the
2494working set size or adding slack to the limit.  Since working set size
2495estimation is hard and error prone, and getting it wrong results in
2496OOM kills, most users tend to err on the side of a looser limit and
2497end up wasting precious resources.
2498
2499The memory.high boundary on the other hand can be set much more
2500conservatively.  When hit, it throttles allocations by forcing them
2501into direct reclaim to work off the excess, but it never invokes the
2502OOM killer.  As a result, a high boundary that is chosen too
2503aggressively will not terminate the processes, but instead it will
2504lead to gradual performance degradation.  The user can monitor this
2505and make corrections until the minimal memory footprint that still
2506gives acceptable performance is found.
2507
2508In extreme cases, with many concurrent allocations and a complete
2509breakdown of reclaim progress within the group, the high boundary can
2510be exceeded.  But even then it's mostly better to satisfy the
2511allocation from the slack available in other groups or the rest of the
2512system than killing the group.  Otherwise, memory.max is there to
2513limit this type of spillover and ultimately contain buggy or even
2514malicious applications.
2515
2516Setting the original memory.limit_in_bytes below the current usage was
2517subject to a race condition, where concurrent charges could cause the
2518limit setting to fail. memory.max on the other hand will first set the
2519limit to prevent new charges, and then reclaim and OOM kill until the
2520new limit is met - or the task writing to memory.max is killed.
2521
2522The combined memory+swap accounting and limiting is replaced by real
2523control over swap space.
2524
2525The main argument for a combined memory+swap facility in the original
2526cgroup design was that global or parental pressure would always be
2527able to swap all anonymous memory of a child group, regardless of the
2528child's own (possibly untrusted) configuration.  However, untrusted
2529groups can sabotage swapping by other means - such as referencing its
2530anonymous memory in a tight loop - and an admin can not assume full
2531swappability when overcommitting untrusted jobs.
2532
2533For trusted jobs, on the other hand, a combined counter is not an
2534intuitive userspace interface, and it flies in the face of the idea
2535that cgroup controllers should account and limit specific physical
2536resources.  Swap space is a resource like all others in the system,
2537and that's why unified hierarchy allows distributing it separately.
2538