1.. _cgroup-v2:
2
3================
4Control Group v2
5================
6
7:Date: October, 2015
8:Author: Tejun Heo <tj@kernel.org>
9
10This is the authoritative documentation on the design, interface and
11conventions of cgroup v2.  It describes all userland-visible aspects
12of cgroup including core and specific controller behaviors.  All
13future changes must be reflected in this document.  Documentation for
14v1 is available under :ref:`Documentation/admin-guide/cgroup-v1/index.rst <cgroup-v1>`.
15
16.. CONTENTS
17
18   1. Introduction
19     1-1. Terminology
20     1-2. What is cgroup?
21   2. Basic Operations
22     2-1. Mounting
23     2-2. Organizing Processes and Threads
24       2-2-1. Processes
25       2-2-2. Threads
26     2-3. [Un]populated Notification
27     2-4. Controlling Controllers
28       2-4-1. Enabling and Disabling
29       2-4-2. Top-down Constraint
30       2-4-3. No Internal Process Constraint
31     2-5. Delegation
32       2-5-1. Model of Delegation
33       2-5-2. Delegation Containment
34     2-6. Guidelines
35       2-6-1. Organize Once and Control
36       2-6-2. Avoid Name Collisions
37   3. Resource Distribution Models
38     3-1. Weights
39     3-2. Limits
40     3-3. Protections
41     3-4. Allocations
42   4. Interface Files
43     4-1. Format
44     4-2. Conventions
45     4-3. Core Interface Files
46   5. Controllers
47     5-1. CPU
48       5-1-1. CPU Interface Files
49     5-2. Memory
50       5-2-1. Memory Interface Files
51       5-2-2. Usage Guidelines
52       5-2-3. Memory Ownership
53     5-3. IO
54       5-3-1. IO Interface Files
55       5-3-2. Writeback
56       5-3-3. IO Latency
57         5-3-3-1. How IO Latency Throttling Works
58         5-3-3-2. IO Latency Interface Files
59       5-3-4. IO Priority
60     5-4. PID
61       5-4-1. PID Interface Files
62     5-5. Cpuset
63       5.5-1. Cpuset Interface Files
64     5-6. Device
65     5-7. RDMA
66       5-7-1. RDMA Interface Files
67     5-8. HugeTLB
68       5.8-1. HugeTLB Interface Files
69     5-9. Misc
70       5.9-1 Miscellaneous cgroup Interface Files
71       5.9-2 Migration and Ownership
72     5-10. Others
73       5-10-1. perf_event
74     5-N. Non-normative information
75       5-N-1. CPU controller root cgroup process behaviour
76       5-N-2. IO controller root cgroup process behaviour
77   6. Namespace
78     6-1. Basics
79     6-2. The Root and Views
80     6-3. Migration and setns(2)
81     6-4. Interaction with Other Namespaces
82   P. Information on Kernel Programming
83     P-1. Filesystem Support for Writeback
84   D. Deprecated v1 Core Features
85   R. Issues with v1 and Rationales for v2
86     R-1. Multiple Hierarchies
87     R-2. Thread Granularity
88     R-3. Competition Between Inner Nodes and Threads
89     R-4. Other Interface Issues
90     R-5. Controller Issues and Remedies
91       R-5-1. Memory
92
93
94Introduction
95============
96
97Terminology
98-----------
99
100"cgroup" stands for "control group" and is never capitalized.  The
101singular form is used to designate the whole feature and also as a
102qualifier as in "cgroup controllers".  When explicitly referring to
103multiple individual control groups, the plural form "cgroups" is used.
104
105
106What is cgroup?
107---------------
108
109cgroup is a mechanism to organize processes hierarchically and
110distribute system resources along the hierarchy in a controlled and
111configurable manner.
112
113cgroup is largely composed of two parts - the core and controllers.
114cgroup core is primarily responsible for hierarchically organizing
115processes.  A cgroup controller is usually responsible for
116distributing a specific type of system resource along the hierarchy
117although there are utility controllers which serve purposes other than
118resource distribution.
119
120cgroups form a tree structure and every process in the system belongs
121to one and only one cgroup.  All threads of a process belong to the
122same cgroup.  On creation, all processes are put in the cgroup that
123the parent process belongs to at the time.  A process can be migrated
124to another cgroup.  Migration of a process doesn't affect already
125existing descendant processes.
126
127Following certain structural constraints, controllers may be enabled or
128disabled selectively on a cgroup.  All controller behaviors are
129hierarchical - if a controller is enabled on a cgroup, it affects all
130processes which belong to the cgroups consisting the inclusive
131sub-hierarchy of the cgroup.  When a controller is enabled on a nested
132cgroup, it always restricts the resource distribution further.  The
133restrictions set closer to the root in the hierarchy can not be
134overridden from further away.
135
136
137Basic Operations
138================
139
140Mounting
141--------
142
143Unlike v1, cgroup v2 has only single hierarchy.  The cgroup v2
144hierarchy can be mounted with the following mount command::
145
146  # mount -t cgroup2 none $MOUNT_POINT
147
148cgroup2 filesystem has the magic number 0x63677270 ("cgrp").  All
149controllers which support v2 and are not bound to a v1 hierarchy are
150automatically bound to the v2 hierarchy and show up at the root.
151Controllers which are not in active use in the v2 hierarchy can be
152bound to other hierarchies.  This allows mixing v2 hierarchy with the
153legacy v1 multiple hierarchies in a fully backward compatible way.
154
155A controller can be moved across hierarchies only after the controller
156is no longer referenced in its current hierarchy.  Because per-cgroup
157controller states are destroyed asynchronously and controllers may
158have lingering references, a controller may not show up immediately on
159the v2 hierarchy after the final umount of the previous hierarchy.
160Similarly, a controller should be fully disabled to be moved out of
161the unified hierarchy and it may take some time for the disabled
162controller to become available for other hierarchies; furthermore, due
163to inter-controller dependencies, other controllers may need to be
164disabled too.
165
166While useful for development and manual configurations, moving
167controllers dynamically between the v2 and other hierarchies is
168strongly discouraged for production use.  It is recommended to decide
169the hierarchies and controller associations before starting using the
170controllers after system boot.
171
172During transition to v2, system management software might still
173automount the v1 cgroup filesystem and so hijack all controllers
174during boot, before manual intervention is possible. To make testing
175and experimenting easier, the kernel parameter cgroup_no_v1= allows
176disabling controllers in v1 and make them always available in v2.
177
178cgroup v2 currently supports the following mount options.
179
180  nsdelegate
181	Consider cgroup namespaces as delegation boundaries.  This
182	option is system wide and can only be set on mount or modified
183	through remount from the init namespace.  The mount option is
184	ignored on non-init namespace mounts.  Please refer to the
185	Delegation section for details.
186
187  favordynmods
188        Reduce the latencies of dynamic cgroup modifications such as
189        task migrations and controller on/offs at the cost of making
190        hot path operations such as forks and exits more expensive.
191        The static usage pattern of creating a cgroup, enabling
192        controllers, and then seeding it with CLONE_INTO_CGROUP is
193        not affected by this option.
194
195  memory_localevents
196        Only populate memory.events with data for the current cgroup,
197        and not any subtrees. This is legacy behaviour, the default
198        behaviour without this option is to include subtree counts.
199        This option is system wide and can only be set on mount or
200        modified through remount from the init namespace. The mount
201        option is ignored on non-init namespace mounts.
202
203  memory_recursiveprot
204        Recursively apply memory.min and memory.low protection to
205        entire subtrees, without requiring explicit downward
206        propagation into leaf cgroups.  This allows protecting entire
207        subtrees from one another, while retaining free competition
208        within those subtrees.  This should have been the default
209        behavior but is a mount-option to avoid regressing setups
210        relying on the original semantics (e.g. specifying bogusly
211        high 'bypass' protection values at higher tree levels).
212
213
214Organizing Processes and Threads
215--------------------------------
216
217Processes
218~~~~~~~~~
219
220Initially, only the root cgroup exists to which all processes belong.
221A child cgroup can be created by creating a sub-directory::
222
223  # mkdir $CGROUP_NAME
224
225A given cgroup may have multiple child cgroups forming a tree
226structure.  Each cgroup has a read-writable interface file
227"cgroup.procs".  When read, it lists the PIDs of all processes which
228belong to the cgroup one-per-line.  The PIDs are not ordered and the
229same PID may show up more than once if the process got moved to
230another cgroup and then back or the PID got recycled while reading.
231
232A process can be migrated into a cgroup by writing its PID to the
233target cgroup's "cgroup.procs" file.  Only one process can be migrated
234on a single write(2) call.  If a process is composed of multiple
235threads, writing the PID of any thread migrates all threads of the
236process.
237
238When a process forks a child process, the new process is born into the
239cgroup that the forking process belongs to at the time of the
240operation.  After exit, a process stays associated with the cgroup
241that it belonged to at the time of exit until it's reaped; however, a
242zombie process does not appear in "cgroup.procs" and thus can't be
243moved to another cgroup.
244
245A cgroup which doesn't have any children or live processes can be
246destroyed by removing the directory.  Note that a cgroup which doesn't
247have any children and is associated only with zombie processes is
248considered empty and can be removed::
249
250  # rmdir $CGROUP_NAME
251
252"/proc/$PID/cgroup" lists a process's cgroup membership.  If legacy
253cgroup is in use in the system, this file may contain multiple lines,
254one for each hierarchy.  The entry for cgroup v2 is always in the
255format "0::$PATH"::
256
257  # cat /proc/842/cgroup
258  ...
259  0::/test-cgroup/test-cgroup-nested
260
261If the process becomes a zombie and the cgroup it was associated with
262is removed subsequently, " (deleted)" is appended to the path::
263
264  # cat /proc/842/cgroup
265  ...
266  0::/test-cgroup/test-cgroup-nested (deleted)
267
268
269Threads
270~~~~~~~
271
272cgroup v2 supports thread granularity for a subset of controllers to
273support use cases requiring hierarchical resource distribution across
274the threads of a group of processes.  By default, all threads of a
275process belong to the same cgroup, which also serves as the resource
276domain to host resource consumptions which are not specific to a
277process or thread.  The thread mode allows threads to be spread across
278a subtree while still maintaining the common resource domain for them.
279
280Controllers which support thread mode are called threaded controllers.
281The ones which don't are called domain controllers.
282
283Marking a cgroup threaded makes it join the resource domain of its
284parent as a threaded cgroup.  The parent may be another threaded
285cgroup whose resource domain is further up in the hierarchy.  The root
286of a threaded subtree, that is, the nearest ancestor which is not
287threaded, is called threaded domain or thread root interchangeably and
288serves as the resource domain for the entire subtree.
289
290Inside a threaded subtree, threads of a process can be put in
291different cgroups and are not subject to the no internal process
292constraint - threaded controllers can be enabled on non-leaf cgroups
293whether they have threads in them or not.
294
295As the threaded domain cgroup hosts all the domain resource
296consumptions of the subtree, it is considered to have internal
297resource consumptions whether there are processes in it or not and
298can't have populated child cgroups which aren't threaded.  Because the
299root cgroup is not subject to no internal process constraint, it can
300serve both as a threaded domain and a parent to domain cgroups.
301
302The current operation mode or type of the cgroup is shown in the
303"cgroup.type" file which indicates whether the cgroup is a normal
304domain, a domain which is serving as the domain of a threaded subtree,
305or a threaded cgroup.
306
307On creation, a cgroup is always a domain cgroup and can be made
308threaded by writing "threaded" to the "cgroup.type" file.  The
309operation is single direction::
310
311  # echo threaded > cgroup.type
312
313Once threaded, the cgroup can't be made a domain again.  To enable the
314thread mode, the following conditions must be met.
315
316- As the cgroup will join the parent's resource domain.  The parent
317  must either be a valid (threaded) domain or a threaded cgroup.
318
319- When the parent is an unthreaded domain, it must not have any domain
320  controllers enabled or populated domain children.  The root is
321  exempt from this requirement.
322
323Topology-wise, a cgroup can be in an invalid state.  Please consider
324the following topology::
325
326  A (threaded domain) - B (threaded) - C (domain, just created)
327
328C is created as a domain but isn't connected to a parent which can
329host child domains.  C can't be used until it is turned into a
330threaded cgroup.  "cgroup.type" file will report "domain (invalid)" in
331these cases.  Operations which fail due to invalid topology use
332EOPNOTSUPP as the errno.
333
334A domain cgroup is turned into a threaded domain when one of its child
335cgroup becomes threaded or threaded controllers are enabled in the
336"cgroup.subtree_control" file while there are processes in the cgroup.
337A threaded domain reverts to a normal domain when the conditions
338clear.
339
340When read, "cgroup.threads" contains the list of the thread IDs of all
341threads in the cgroup.  Except that the operations are per-thread
342instead of per-process, "cgroup.threads" has the same format and
343behaves the same way as "cgroup.procs".  While "cgroup.threads" can be
344written to in any cgroup, as it can only move threads inside the same
345threaded domain, its operations are confined inside each threaded
346subtree.
347
348The threaded domain cgroup serves as the resource domain for the whole
349subtree, and, while the threads can be scattered across the subtree,
350all the processes are considered to be in the threaded domain cgroup.
351"cgroup.procs" in a threaded domain cgroup contains the PIDs of all
352processes in the subtree and is not readable in the subtree proper.
353However, "cgroup.procs" can be written to from anywhere in the subtree
354to migrate all threads of the matching process to the cgroup.
355
356Only threaded controllers can be enabled in a threaded subtree.  When
357a threaded controller is enabled inside a threaded subtree, it only
358accounts for and controls resource consumptions associated with the
359threads in the cgroup and its descendants.  All consumptions which
360aren't tied to a specific thread belong to the threaded domain cgroup.
361
362Because a threaded subtree is exempt from no internal process
363constraint, a threaded controller must be able to handle competition
364between threads in a non-leaf cgroup and its child cgroups.  Each
365threaded controller defines how such competitions are handled.
366
367
368[Un]populated Notification
369--------------------------
370
371Each non-root cgroup has a "cgroup.events" file which contains
372"populated" field indicating whether the cgroup's sub-hierarchy has
373live processes in it.  Its value is 0 if there is no live process in
374the cgroup and its descendants; otherwise, 1.  poll and [id]notify
375events are triggered when the value changes.  This can be used, for
376example, to start a clean-up operation after all processes of a given
377sub-hierarchy have exited.  The populated state updates and
378notifications are recursive.  Consider the following sub-hierarchy
379where the numbers in the parentheses represent the numbers of processes
380in each cgroup::
381
382  A(4) - B(0) - C(1)
383              \ D(0)
384
385A, B and C's "populated" fields would be 1 while D's 0.  After the one
386process in C exits, B and C's "populated" fields would flip to "0" and
387file modified events will be generated on the "cgroup.events" files of
388both cgroups.
389
390
391Controlling Controllers
392-----------------------
393
394Enabling and Disabling
395~~~~~~~~~~~~~~~~~~~~~~
396
397Each cgroup has a "cgroup.controllers" file which lists all
398controllers available for the cgroup to enable::
399
400  # cat cgroup.controllers
401  cpu io memory
402
403No controller is enabled by default.  Controllers can be enabled and
404disabled by writing to the "cgroup.subtree_control" file::
405
406  # echo "+cpu +memory -io" > cgroup.subtree_control
407
408Only controllers which are listed in "cgroup.controllers" can be
409enabled.  When multiple operations are specified as above, either they
410all succeed or fail.  If multiple operations on the same controller
411are specified, the last one is effective.
412
413Enabling a controller in a cgroup indicates that the distribution of
414the target resource across its immediate children will be controlled.
415Consider the following sub-hierarchy.  The enabled controllers are
416listed in parentheses::
417
418  A(cpu,memory) - B(memory) - C()
419                            \ D()
420
421As A has "cpu" and "memory" enabled, A will control the distribution
422of CPU cycles and memory to its children, in this case, B.  As B has
423"memory" enabled but not "CPU", C and D will compete freely on CPU
424cycles but their division of memory available to B will be controlled.
425
426As a controller regulates the distribution of the target resource to
427the cgroup's children, enabling it creates the controller's interface
428files in the child cgroups.  In the above example, enabling "cpu" on B
429would create the "cpu." prefixed controller interface files in C and
430D.  Likewise, disabling "memory" from B would remove the "memory."
431prefixed controller interface files from C and D.  This means that the
432controller interface files - anything which doesn't start with
433"cgroup." are owned by the parent rather than the cgroup itself.
434
435
436Top-down Constraint
437~~~~~~~~~~~~~~~~~~~
438
439Resources are distributed top-down and a cgroup can further distribute
440a resource only if the resource has been distributed to it from the
441parent.  This means that all non-root "cgroup.subtree_control" files
442can only contain controllers which are enabled in the parent's
443"cgroup.subtree_control" file.  A controller can be enabled only if
444the parent has the controller enabled and a controller can't be
445disabled if one or more children have it enabled.
446
447
448No Internal Process Constraint
449~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
450
451Non-root cgroups can distribute domain resources to their children
452only when they don't have any processes of their own.  In other words,
453only domain cgroups which don't contain any processes can have domain
454controllers enabled in their "cgroup.subtree_control" files.
455
456This guarantees that, when a domain controller is looking at the part
457of the hierarchy which has it enabled, processes are always only on
458the leaves.  This rules out situations where child cgroups compete
459against internal processes of the parent.
460
461The root cgroup is exempt from this restriction.  Root contains
462processes and anonymous resource consumption which can't be associated
463with any other cgroups and requires special treatment from most
464controllers.  How resource consumption in the root cgroup is governed
465is up to each controller (for more information on this topic please
466refer to the Non-normative information section in the Controllers
467chapter).
468
469Note that the restriction doesn't get in the way if there is no
470enabled controller in the cgroup's "cgroup.subtree_control".  This is
471important as otherwise it wouldn't be possible to create children of a
472populated cgroup.  To control resource distribution of a cgroup, the
473cgroup must create children and transfer all its processes to the
474children before enabling controllers in its "cgroup.subtree_control"
475file.
476
477
478Delegation
479----------
480
481Model of Delegation
482~~~~~~~~~~~~~~~~~~~
483
484A cgroup can be delegated in two ways.  First, to a less privileged
485user by granting write access of the directory and its "cgroup.procs",
486"cgroup.threads" and "cgroup.subtree_control" files to the user.
487Second, if the "nsdelegate" mount option is set, automatically to a
488cgroup namespace on namespace creation.
489
490Because the resource control interface files in a given directory
491control the distribution of the parent's resources, the delegatee
492shouldn't be allowed to write to them.  For the first method, this is
493achieved by not granting access to these files.  For the second, the
494kernel rejects writes to all files other than "cgroup.procs" and
495"cgroup.subtree_control" on a namespace root from inside the
496namespace.
497
498The end results are equivalent for both delegation types.  Once
499delegated, the user can build sub-hierarchy under the directory,
500organize processes inside it as it sees fit and further distribute the
501resources it received from the parent.  The limits and other settings
502of all resource controllers are hierarchical and regardless of what
503happens in the delegated sub-hierarchy, nothing can escape the
504resource restrictions imposed by the parent.
505
506Currently, cgroup doesn't impose any restrictions on the number of
507cgroups in or nesting depth of a delegated sub-hierarchy; however,
508this may be limited explicitly in the future.
509
510
511Delegation Containment
512~~~~~~~~~~~~~~~~~~~~~~
513
514A delegated sub-hierarchy is contained in the sense that processes
515can't be moved into or out of the sub-hierarchy by the delegatee.
516
517For delegations to a less privileged user, this is achieved by
518requiring the following conditions for a process with a non-root euid
519to migrate a target process into a cgroup by writing its PID to the
520"cgroup.procs" file.
521
522- The writer must have write access to the "cgroup.procs" file.
523
524- The writer must have write access to the "cgroup.procs" file of the
525  common ancestor of the source and destination cgroups.
526
527The above two constraints ensure that while a delegatee may migrate
528processes around freely in the delegated sub-hierarchy it can't pull
529in from or push out to outside the sub-hierarchy.
530
531For an example, let's assume cgroups C0 and C1 have been delegated to
532user U0 who created C00, C01 under C0 and C10 under C1 as follows and
533all processes under C0 and C1 belong to U0::
534
535  ~~~~~~~~~~~~~ - C0 - C00
536  ~ cgroup    ~      \ C01
537  ~ hierarchy ~
538  ~~~~~~~~~~~~~ - C1 - C10
539
540Let's also say U0 wants to write the PID of a process which is
541currently in C10 into "C00/cgroup.procs".  U0 has write access to the
542file; however, the common ancestor of the source cgroup C10 and the
543destination cgroup C00 is above the points of delegation and U0 would
544not have write access to its "cgroup.procs" files and thus the write
545will be denied with -EACCES.
546
547For delegations to namespaces, containment is achieved by requiring
548that both the source and destination cgroups are reachable from the
549namespace of the process which is attempting the migration.  If either
550is not reachable, the migration is rejected with -ENOENT.
551
552
553Guidelines
554----------
555
556Organize Once and Control
557~~~~~~~~~~~~~~~~~~~~~~~~~
558
559Migrating a process across cgroups is a relatively expensive operation
560and stateful resources such as memory are not moved together with the
561process.  This is an explicit design decision as there often exist
562inherent trade-offs between migration and various hot paths in terms
563of synchronization cost.
564
565As such, migrating processes across cgroups frequently as a means to
566apply different resource restrictions is discouraged.  A workload
567should be assigned to a cgroup according to the system's logical and
568resource structure once on start-up.  Dynamic adjustments to resource
569distribution can be made by changing controller configuration through
570the interface files.
571
572
573Avoid Name Collisions
574~~~~~~~~~~~~~~~~~~~~~
575
576Interface files for a cgroup and its children cgroups occupy the same
577directory and it is possible to create children cgroups which collide
578with interface files.
579
580All cgroup core interface files are prefixed with "cgroup." and each
581controller's interface files are prefixed with the controller name and
582a dot.  A controller's name is composed of lower case alphabets and
583'_'s but never begins with an '_' so it can be used as the prefix
584character for collision avoidance.  Also, interface file names won't
585start or end with terms which are often used in categorizing workloads
586such as job, service, slice, unit or workload.
587
588cgroup doesn't do anything to prevent name collisions and it's the
589user's responsibility to avoid them.
590
591
592Resource Distribution Models
593============================
594
595cgroup controllers implement several resource distribution schemes
596depending on the resource type and expected use cases.  This section
597describes major schemes in use along with their expected behaviors.
598
599
600Weights
601-------
602
603A parent's resource is distributed by adding up the weights of all
604active children and giving each the fraction matching the ratio of its
605weight against the sum.  As only children which can make use of the
606resource at the moment participate in the distribution, this is
607work-conserving.  Due to the dynamic nature, this model is usually
608used for stateless resources.
609
610All weights are in the range [1, 10000] with the default at 100.  This
611allows symmetric multiplicative biases in both directions at fine
612enough granularity while staying in the intuitive range.
613
614As long as the weight is in range, all configuration combinations are
615valid and there is no reason to reject configuration changes or
616process migrations.
617
618"cpu.weight" proportionally distributes CPU cycles to active children
619and is an example of this type.
620
621
622.. _cgroupv2-limits-distributor:
623
624Limits
625------
626
627A child can only consume up to the configured amount of the resource.
628Limits can be over-committed - the sum of the limits of children can
629exceed the amount of resource available to the parent.
630
631Limits are in the range [0, max] and defaults to "max", which is noop.
632
633As limits can be over-committed, all configuration combinations are
634valid and there is no reason to reject configuration changes or
635process migrations.
636
637"io.max" limits the maximum BPS and/or IOPS that a cgroup can consume
638on an IO device and is an example of this type.
639
640.. _cgroupv2-protections-distributor:
641
642Protections
643-----------
644
645A cgroup is protected up to the configured amount of the resource
646as long as the usages of all its ancestors are under their
647protected levels.  Protections can be hard guarantees or best effort
648soft boundaries.  Protections can also be over-committed in which case
649only up to the amount available to the parent is protected among
650children.
651
652Protections are in the range [0, max] and defaults to 0, which is
653noop.
654
655As protections can be over-committed, all configuration combinations
656are valid and there is no reason to reject configuration changes or
657process migrations.
658
659"memory.low" implements best-effort memory protection and is an
660example of this type.
661
662
663Allocations
664-----------
665
666A cgroup is exclusively allocated a certain amount of a finite
667resource.  Allocations can't be over-committed - the sum of the
668allocations of children can not exceed the amount of resource
669available to the parent.
670
671Allocations are in the range [0, max] and defaults to 0, which is no
672resource.
673
674As allocations can't be over-committed, some configuration
675combinations are invalid and should be rejected.  Also, if the
676resource is mandatory for execution of processes, process migrations
677may be rejected.
678
679"cpu.rt.max" hard-allocates realtime slices and is an example of this
680type.
681
682
683Interface Files
684===============
685
686Format
687------
688
689All interface files should be in one of the following formats whenever
690possible::
691
692  New-line separated values
693  (when only one value can be written at once)
694
695	VAL0\n
696	VAL1\n
697	...
698
699  Space separated values
700  (when read-only or multiple values can be written at once)
701
702	VAL0 VAL1 ...\n
703
704  Flat keyed
705
706	KEY0 VAL0\n
707	KEY1 VAL1\n
708	...
709
710  Nested keyed
711
712	KEY0 SUB_KEY0=VAL00 SUB_KEY1=VAL01...
713	KEY1 SUB_KEY0=VAL10 SUB_KEY1=VAL11...
714	...
715
716For a writable file, the format for writing should generally match
717reading; however, controllers may allow omitting later fields or
718implement restricted shortcuts for most common use cases.
719
720For both flat and nested keyed files, only the values for a single key
721can be written at a time.  For nested keyed files, the sub key pairs
722may be specified in any order and not all pairs have to be specified.
723
724
725Conventions
726-----------
727
728- Settings for a single feature should be contained in a single file.
729
730- The root cgroup should be exempt from resource control and thus
731  shouldn't have resource control interface files.
732
733- The default time unit is microseconds.  If a different unit is ever
734  used, an explicit unit suffix must be present.
735
736- A parts-per quantity should use a percentage decimal with at least
737  two digit fractional part - e.g. 13.40.
738
739- If a controller implements weight based resource distribution, its
740  interface file should be named "weight" and have the range [1,
741  10000] with 100 as the default.  The values are chosen to allow
742  enough and symmetric bias in both directions while keeping it
743  intuitive (the default is 100%).
744
745- If a controller implements an absolute resource guarantee and/or
746  limit, the interface files should be named "min" and "max"
747  respectively.  If a controller implements best effort resource
748  guarantee and/or limit, the interface files should be named "low"
749  and "high" respectively.
750
751  In the above four control files, the special token "max" should be
752  used to represent upward infinity for both reading and writing.
753
754- If a setting has a configurable default value and keyed specific
755  overrides, the default entry should be keyed with "default" and
756  appear as the first entry in the file.
757
758  The default value can be updated by writing either "default $VAL" or
759  "$VAL".
760
761  When writing to update a specific override, "default" can be used as
762  the value to indicate removal of the override.  Override entries
763  with "default" as the value must not appear when read.
764
765  For example, a setting which is keyed by major:minor device numbers
766  with integer values may look like the following::
767
768    # cat cgroup-example-interface-file
769    default 150
770    8:0 300
771
772  The default value can be updated by::
773
774    # echo 125 > cgroup-example-interface-file
775
776  or::
777
778    # echo "default 125" > cgroup-example-interface-file
779
780  An override can be set by::
781
782    # echo "8:16 170" > cgroup-example-interface-file
783
784  and cleared by::
785
786    # echo "8:0 default" > cgroup-example-interface-file
787    # cat cgroup-example-interface-file
788    default 125
789    8:16 170
790
791- For events which are not very high frequency, an interface file
792  "events" should be created which lists event key value pairs.
793  Whenever a notifiable event happens, file modified event should be
794  generated on the file.
795
796
797Core Interface Files
798--------------------
799
800All cgroup core files are prefixed with "cgroup."
801
802  cgroup.type
803	A read-write single value file which exists on non-root
804	cgroups.
805
806	When read, it indicates the current type of the cgroup, which
807	can be one of the following values.
808
809	- "domain" : A normal valid domain cgroup.
810
811	- "domain threaded" : A threaded domain cgroup which is
812          serving as the root of a threaded subtree.
813
814	- "domain invalid" : A cgroup which is in an invalid state.
815	  It can't be populated or have controllers enabled.  It may
816	  be allowed to become a threaded cgroup.
817
818	- "threaded" : A threaded cgroup which is a member of a
819          threaded subtree.
820
821	A cgroup can be turned into a threaded cgroup by writing
822	"threaded" to this file.
823
824  cgroup.procs
825	A read-write new-line separated values file which exists on
826	all cgroups.
827
828	When read, it lists the PIDs of all processes which belong to
829	the cgroup one-per-line.  The PIDs are not ordered and the
830	same PID may show up more than once if the process got moved
831	to another cgroup and then back or the PID got recycled while
832	reading.
833
834	A PID can be written to migrate the process associated with
835	the PID to the cgroup.  The writer should match all of the
836	following conditions.
837
838	- It must have write access to the "cgroup.procs" file.
839
840	- It must have write access to the "cgroup.procs" file of the
841	  common ancestor of the source and destination cgroups.
842
843	When delegating a sub-hierarchy, write access to this file
844	should be granted along with the containing directory.
845
846	In a threaded cgroup, reading this file fails with EOPNOTSUPP
847	as all the processes belong to the thread root.  Writing is
848	supported and moves every thread of the process to the cgroup.
849
850  cgroup.threads
851	A read-write new-line separated values file which exists on
852	all cgroups.
853
854	When read, it lists the TIDs of all threads which belong to
855	the cgroup one-per-line.  The TIDs are not ordered and the
856	same TID may show up more than once if the thread got moved to
857	another cgroup and then back or the TID got recycled while
858	reading.
859
860	A TID can be written to migrate the thread associated with the
861	TID to the cgroup.  The writer should match all of the
862	following conditions.
863
864	- It must have write access to the "cgroup.threads" file.
865
866	- The cgroup that the thread is currently in must be in the
867          same resource domain as the destination cgroup.
868
869	- It must have write access to the "cgroup.procs" file of the
870	  common ancestor of the source and destination cgroups.
871
872	When delegating a sub-hierarchy, write access to this file
873	should be granted along with the containing directory.
874
875  cgroup.controllers
876	A read-only space separated values file which exists on all
877	cgroups.
878
879	It shows space separated list of all controllers available to
880	the cgroup.  The controllers are not ordered.
881
882  cgroup.subtree_control
883	A read-write space separated values file which exists on all
884	cgroups.  Starts out empty.
885
886	When read, it shows space separated list of the controllers
887	which are enabled to control resource distribution from the
888	cgroup to its children.
889
890	Space separated list of controllers prefixed with '+' or '-'
891	can be written to enable or disable controllers.  A controller
892	name prefixed with '+' enables the controller and '-'
893	disables.  If a controller appears more than once on the list,
894	the last one is effective.  When multiple enable and disable
895	operations are specified, either all succeed or all fail.
896
897  cgroup.events
898	A read-only flat-keyed file which exists on non-root cgroups.
899	The following entries are defined.  Unless specified
900	otherwise, a value change in this file generates a file
901	modified event.
902
903	  populated
904		1 if the cgroup or its descendants contains any live
905		processes; otherwise, 0.
906	  frozen
907		1 if the cgroup is frozen; otherwise, 0.
908
909  cgroup.max.descendants
910	A read-write single value files.  The default is "max".
911
912	Maximum allowed number of descent cgroups.
913	If the actual number of descendants is equal or larger,
914	an attempt to create a new cgroup in the hierarchy will fail.
915
916  cgroup.max.depth
917	A read-write single value files.  The default is "max".
918
919	Maximum allowed descent depth below the current cgroup.
920	If the actual descent depth is equal or larger,
921	an attempt to create a new child cgroup will fail.
922
923  cgroup.stat
924	A read-only flat-keyed file with the following entries:
925
926	  nr_descendants
927		Total number of visible descendant cgroups.
928
929	  nr_dying_descendants
930		Total number of dying descendant cgroups. A cgroup becomes
931		dying after being deleted by a user. The cgroup will remain
932		in dying state for some time undefined time (which can depend
933		on system load) before being completely destroyed.
934
935		A process can't enter a dying cgroup under any circumstances,
936		a dying cgroup can't revive.
937
938		A dying cgroup can consume system resources not exceeding
939		limits, which were active at the moment of cgroup deletion.
940
941  cgroup.freeze
942	A read-write single value file which exists on non-root cgroups.
943	Allowed values are "0" and "1". The default is "0".
944
945	Writing "1" to the file causes freezing of the cgroup and all
946	descendant cgroups. This means that all belonging processes will
947	be stopped and will not run until the cgroup will be explicitly
948	unfrozen. Freezing of the cgroup may take some time; when this action
949	is completed, the "frozen" value in the cgroup.events control file
950	will be updated to "1" and the corresponding notification will be
951	issued.
952
953	A cgroup can be frozen either by its own settings, or by settings
954	of any ancestor cgroups. If any of ancestor cgroups is frozen, the
955	cgroup will remain frozen.
956
957	Processes in the frozen cgroup can be killed by a fatal signal.
958	They also can enter and leave a frozen cgroup: either by an explicit
959	move by a user, or if freezing of the cgroup races with fork().
960	If a process is moved to a frozen cgroup, it stops. If a process is
961	moved out of a frozen cgroup, it becomes running.
962
963	Frozen status of a cgroup doesn't affect any cgroup tree operations:
964	it's possible to delete a frozen (and empty) cgroup, as well as
965	create new sub-cgroups.
966
967  cgroup.kill
968	A write-only single value file which exists in non-root cgroups.
969	The only allowed value is "1".
970
971	Writing "1" to the file causes the cgroup and all descendant cgroups to
972	be killed. This means that all processes located in the affected cgroup
973	tree will be killed via SIGKILL.
974
975	Killing a cgroup tree will deal with concurrent forks appropriately and
976	is protected against migrations.
977
978	In a threaded cgroup, writing this file fails with EOPNOTSUPP as
979	killing cgroups is a process directed operation, i.e. it affects
980	the whole thread-group.
981
982  cgroup.pressure
983	A read-write single value file that allowed values are "0" and "1".
984	The default is "1".
985
986	Writing "0" to the file will disable the cgroup PSI accounting.
987	Writing "1" to the file will re-enable the cgroup PSI accounting.
988
989	This control attribute is not hierarchical, so disable or enable PSI
990	accounting in a cgroup does not affect PSI accounting in descendants
991	and doesn't need pass enablement via ancestors from root.
992
993	The reason this control attribute exists is that PSI accounts stalls for
994	each cgroup separately and aggregates it at each level of the hierarchy.
995	This may cause non-negligible overhead for some workloads when under
996	deep level of the hierarchy, in which case this control attribute can
997	be used to disable PSI accounting in the non-leaf cgroups.
998
999  irq.pressure
1000	A read-write nested-keyed file.
1001
1002	Shows pressure stall information for IRQ/SOFTIRQ. See
1003	:ref:`Documentation/accounting/psi.rst <psi>` for details.
1004
1005Controllers
1006===========
1007
1008.. _cgroup-v2-cpu:
1009
1010CPU
1011---
1012
1013The "cpu" controllers regulates distribution of CPU cycles.  This
1014controller implements weight and absolute bandwidth limit models for
1015normal scheduling policy and absolute bandwidth allocation model for
1016realtime scheduling policy.
1017
1018In all the above models, cycles distribution is defined only on a temporal
1019base and it does not account for the frequency at which tasks are executed.
1020The (optional) utilization clamping support allows to hint the schedutil
1021cpufreq governor about the minimum desired frequency which should always be
1022provided by a CPU, as well as the maximum desired frequency, which should not
1023be exceeded by a CPU.
1024
1025WARNING: cgroup2 doesn't yet support control of realtime processes and
1026the cpu controller can only be enabled when all RT processes are in
1027the root cgroup.  Be aware that system management software may already
1028have placed RT processes into nonroot cgroups during the system boot
1029process, and these processes may need to be moved to the root cgroup
1030before the cpu controller can be enabled.
1031
1032
1033CPU Interface Files
1034~~~~~~~~~~~~~~~~~~~
1035
1036All time durations are in microseconds.
1037
1038  cpu.stat
1039	A read-only flat-keyed file.
1040	This file exists whether the controller is enabled or not.
1041
1042	It always reports the following three stats:
1043
1044	- usage_usec
1045	- user_usec
1046	- system_usec
1047
1048	and the following three when the controller is enabled:
1049
1050	- nr_periods
1051	- nr_throttled
1052	- throttled_usec
1053	- nr_bursts
1054	- burst_usec
1055
1056  cpu.weight
1057	A read-write single value file which exists on non-root
1058	cgroups.  The default is "100".
1059
1060	The weight in the range [1, 10000].
1061
1062  cpu.weight.nice
1063	A read-write single value file which exists on non-root
1064	cgroups.  The default is "0".
1065
1066	The nice value is in the range [-20, 19].
1067
1068	This interface file is an alternative interface for
1069	"cpu.weight" and allows reading and setting weight using the
1070	same values used by nice(2).  Because the range is smaller and
1071	granularity is coarser for the nice values, the read value is
1072	the closest approximation of the current weight.
1073
1074  cpu.max
1075	A read-write two value file which exists on non-root cgroups.
1076	The default is "max 100000".
1077
1078	The maximum bandwidth limit.  It's in the following format::
1079
1080	  $MAX $PERIOD
1081
1082	which indicates that the group may consume up to $MAX in each
1083	$PERIOD duration.  "max" for $MAX indicates no limit.  If only
1084	one number is written, $MAX is updated.
1085
1086  cpu.max.burst
1087	A read-write single value file which exists on non-root
1088	cgroups.  The default is "0".
1089
1090	The burst in the range [0, $MAX].
1091
1092  cpu.pressure
1093	A read-write nested-keyed file.
1094
1095	Shows pressure stall information for CPU. See
1096	:ref:`Documentation/accounting/psi.rst <psi>` for details.
1097
1098  cpu.uclamp.min
1099        A read-write single value file which exists on non-root cgroups.
1100        The default is "0", i.e. no utilization boosting.
1101
1102        The requested minimum utilization (protection) as a percentage
1103        rational number, e.g. 12.34 for 12.34%.
1104
1105        This interface allows reading and setting minimum utilization clamp
1106        values similar to the sched_setattr(2). This minimum utilization
1107        value is used to clamp the task specific minimum utilization clamp.
1108
1109        The requested minimum utilization (protection) is always capped by
1110        the current value for the maximum utilization (limit), i.e.
1111        `cpu.uclamp.max`.
1112
1113  cpu.uclamp.max
1114        A read-write single value file which exists on non-root cgroups.
1115        The default is "max". i.e. no utilization capping
1116
1117        The requested maximum utilization (limit) as a percentage rational
1118        number, e.g. 98.76 for 98.76%.
1119
1120        This interface allows reading and setting maximum utilization clamp
1121        values similar to the sched_setattr(2). This maximum utilization
1122        value is used to clamp the task specific maximum utilization clamp.
1123
1124
1125
1126Memory
1127------
1128
1129The "memory" controller regulates distribution of memory.  Memory is
1130stateful and implements both limit and protection models.  Due to the
1131intertwining between memory usage and reclaim pressure and the
1132stateful nature of memory, the distribution model is relatively
1133complex.
1134
1135While not completely water-tight, all major memory usages by a given
1136cgroup are tracked so that the total memory consumption can be
1137accounted and controlled to a reasonable extent.  Currently, the
1138following types of memory usages are tracked.
1139
1140- Userland memory - page cache and anonymous memory.
1141
1142- Kernel data structures such as dentries and inodes.
1143
1144- TCP socket buffers.
1145
1146The above list may expand in the future for better coverage.
1147
1148
1149Memory Interface Files
1150~~~~~~~~~~~~~~~~~~~~~~
1151
1152All memory amounts are in bytes.  If a value which is not aligned to
1153PAGE_SIZE is written, the value may be rounded up to the closest
1154PAGE_SIZE multiple when read back.
1155
1156  memory.current
1157	A read-only single value file which exists on non-root
1158	cgroups.
1159
1160	The total amount of memory currently being used by the cgroup
1161	and its descendants.
1162
1163  memory.min
1164	A read-write single value file which exists on non-root
1165	cgroups.  The default is "0".
1166
1167	Hard memory protection.  If the memory usage of a cgroup
1168	is within its effective min boundary, the cgroup's memory
1169	won't be reclaimed under any conditions. If there is no
1170	unprotected reclaimable memory available, OOM killer
1171	is invoked. Above the effective min boundary (or
1172	effective low boundary if it is higher), pages are reclaimed
1173	proportionally to the overage, reducing reclaim pressure for
1174	smaller overages.
1175
1176	Effective min boundary is limited by memory.min values of
1177	all ancestor cgroups. If there is memory.min overcommitment
1178	(child cgroup or cgroups are requiring more protected memory
1179	than parent will allow), then each child cgroup will get
1180	the part of parent's protection proportional to its
1181	actual memory usage below memory.min.
1182
1183	Putting more memory than generally available under this
1184	protection is discouraged and may lead to constant OOMs.
1185
1186	If a memory cgroup is not populated with processes,
1187	its memory.min is ignored.
1188
1189  memory.low
1190	A read-write single value file which exists on non-root
1191	cgroups.  The default is "0".
1192
1193	Best-effort memory protection.  If the memory usage of a
1194	cgroup is within its effective low boundary, the cgroup's
1195	memory won't be reclaimed unless there is no reclaimable
1196	memory available in unprotected cgroups.
1197	Above the effective low	boundary (or
1198	effective min boundary if it is higher), pages are reclaimed
1199	proportionally to the overage, reducing reclaim pressure for
1200	smaller overages.
1201
1202	Effective low boundary is limited by memory.low values of
1203	all ancestor cgroups. If there is memory.low overcommitment
1204	(child cgroup or cgroups are requiring more protected memory
1205	than parent will allow), then each child cgroup will get
1206	the part of parent's protection proportional to its
1207	actual memory usage below memory.low.
1208
1209	Putting more memory than generally available under this
1210	protection is discouraged.
1211
1212  memory.high
1213	A read-write single value file which exists on non-root
1214	cgroups.  The default is "max".
1215
1216	Memory usage throttle limit.  This is the main mechanism to
1217	control memory usage of a cgroup.  If a cgroup's usage goes
1218	over the high boundary, the processes of the cgroup are
1219	throttled and put under heavy reclaim pressure.
1220
1221	Going over the high limit never invokes the OOM killer and
1222	under extreme conditions the limit may be breached.
1223
1224  memory.max
1225	A read-write single value file which exists on non-root
1226	cgroups.  The default is "max".
1227
1228	Memory usage hard limit.  This is the final protection
1229	mechanism.  If a cgroup's memory usage reaches this limit and
1230	can't be reduced, the OOM killer is invoked in the cgroup.
1231	Under certain circumstances, the usage may go over the limit
1232	temporarily.
1233
1234	In default configuration regular 0-order allocations always
1235	succeed unless OOM killer chooses current task as a victim.
1236
1237	Some kinds of allocations don't invoke the OOM killer.
1238	Caller could retry them differently, return into userspace
1239	as -ENOMEM or silently ignore in cases like disk readahead.
1240
1241	This is the ultimate protection mechanism.  As long as the
1242	high limit is used and monitored properly, this limit's
1243	utility is limited to providing the final safety net.
1244
1245  memory.reclaim
1246	A write-only nested-keyed file which exists for all cgroups.
1247
1248	This is a simple interface to trigger memory reclaim in the
1249	target cgroup.
1250
1251	This file accepts a single key, the number of bytes to reclaim.
1252	No nested keys are currently supported.
1253
1254	Example::
1255
1256	  echo "1G" > memory.reclaim
1257
1258	The interface can be later extended with nested keys to
1259	configure the reclaim behavior. For example, specify the
1260	type of memory to reclaim from (anon, file, ..).
1261
1262	Please note that the kernel can over or under reclaim from
1263	the target cgroup. If less bytes are reclaimed than the
1264	specified amount, -EAGAIN is returned.
1265
1266	Please note that the proactive reclaim (triggered by this
1267	interface) is not meant to indicate memory pressure on the
1268	memory cgroup. Therefore socket memory balancing triggered by
1269	the memory reclaim normally is not exercised in this case.
1270	This means that the networking layer will not adapt based on
1271	reclaim induced by memory.reclaim.
1272
1273  memory.peak
1274	A read-only single value file which exists on non-root
1275	cgroups.
1276
1277	The max memory usage recorded for the cgroup and its
1278	descendants since the creation of the cgroup.
1279
1280  memory.oom.group
1281	A read-write single value file which exists on non-root
1282	cgroups.  The default value is "0".
1283
1284	Determines whether the cgroup should be treated as
1285	an indivisible workload by the OOM killer. If set,
1286	all tasks belonging to the cgroup or to its descendants
1287	(if the memory cgroup is not a leaf cgroup) are killed
1288	together or not at all. This can be used to avoid
1289	partial kills to guarantee workload integrity.
1290
1291	Tasks with the OOM protection (oom_score_adj set to -1000)
1292	are treated as an exception and are never killed.
1293
1294	If the OOM killer is invoked in a cgroup, it's not going
1295	to kill any tasks outside of this cgroup, regardless
1296	memory.oom.group values of ancestor cgroups.
1297
1298  memory.events
1299	A read-only flat-keyed file which exists on non-root cgroups.
1300	The following entries are defined.  Unless specified
1301	otherwise, a value change in this file generates a file
1302	modified event.
1303
1304	Note that all fields in this file are hierarchical and the
1305	file modified event can be generated due to an event down the
1306	hierarchy. For the local events at the cgroup level see
1307	memory.events.local.
1308
1309	  low
1310		The number of times the cgroup is reclaimed due to
1311		high memory pressure even though its usage is under
1312		the low boundary.  This usually indicates that the low
1313		boundary is over-committed.
1314
1315	  high
1316		The number of times processes of the cgroup are
1317		throttled and routed to perform direct memory reclaim
1318		because the high memory boundary was exceeded.  For a
1319		cgroup whose memory usage is capped by the high limit
1320		rather than global memory pressure, this event's
1321		occurrences are expected.
1322
1323	  max
1324		The number of times the cgroup's memory usage was
1325		about to go over the max boundary.  If direct reclaim
1326		fails to bring it down, the cgroup goes to OOM state.
1327
1328	  oom
1329		The number of time the cgroup's memory usage was
1330		reached the limit and allocation was about to fail.
1331
1332		This event is not raised if the OOM killer is not
1333		considered as an option, e.g. for failed high-order
1334		allocations or if caller asked to not retry attempts.
1335
1336	  oom_kill
1337		The number of processes belonging to this cgroup
1338		killed by any kind of OOM killer.
1339
1340          oom_group_kill
1341                The number of times a group OOM has occurred.
1342
1343  memory.events.local
1344	Similar to memory.events but the fields in the file are local
1345	to the cgroup i.e. not hierarchical. The file modified event
1346	generated on this file reflects only the local events.
1347
1348  memory.stat
1349	A read-only flat-keyed file which exists on non-root cgroups.
1350
1351	This breaks down the cgroup's memory footprint into different
1352	types of memory, type-specific details, and other information
1353	on the state and past events of the memory management system.
1354
1355	All memory amounts are in bytes.
1356
1357	The entries are ordered to be human readable, and new entries
1358	can show up in the middle. Don't rely on items remaining in a
1359	fixed position; use the keys to look up specific values!
1360
1361	If the entry has no per-node counter (or not show in the
1362	memory.numa_stat). We use 'npn' (non-per-node) as the tag
1363	to indicate that it will not show in the memory.numa_stat.
1364
1365	  anon
1366		Amount of memory used in anonymous mappings such as
1367		brk(), sbrk(), and mmap(MAP_ANONYMOUS)
1368
1369	  file
1370		Amount of memory used to cache filesystem data,
1371		including tmpfs and shared memory.
1372
1373	  kernel (npn)
1374		Amount of total kernel memory, including
1375		(kernel_stack, pagetables, percpu, vmalloc, slab) in
1376		addition to other kernel memory use cases.
1377
1378	  kernel_stack
1379		Amount of memory allocated to kernel stacks.
1380
1381	  pagetables
1382                Amount of memory allocated for page tables.
1383
1384	  sec_pagetables
1385		Amount of memory allocated for secondary page tables,
1386		this currently includes KVM mmu allocations on x86
1387		and arm64.
1388
1389	  percpu (npn)
1390		Amount of memory used for storing per-cpu kernel
1391		data structures.
1392
1393	  sock (npn)
1394		Amount of memory used in network transmission buffers
1395
1396	  vmalloc (npn)
1397		Amount of memory used for vmap backed memory.
1398
1399	  shmem
1400		Amount of cached filesystem data that is swap-backed,
1401		such as tmpfs, shm segments, shared anonymous mmap()s
1402
1403	  zswap
1404		Amount of memory consumed by the zswap compression backend.
1405
1406	  zswapped
1407		Amount of application memory swapped out to zswap.
1408
1409	  file_mapped
1410		Amount of cached filesystem data mapped with mmap()
1411
1412	  file_dirty
1413		Amount of cached filesystem data that was modified but
1414		not yet written back to disk
1415
1416	  file_writeback
1417		Amount of cached filesystem data that was modified and
1418		is currently being written back to disk
1419
1420	  swapcached
1421		Amount of swap cached in memory. The swapcache is accounted
1422		against both memory and swap usage.
1423
1424	  anon_thp
1425		Amount of memory used in anonymous mappings backed by
1426		transparent hugepages
1427
1428	  file_thp
1429		Amount of cached filesystem data backed by transparent
1430		hugepages
1431
1432	  shmem_thp
1433		Amount of shm, tmpfs, shared anonymous mmap()s backed by
1434		transparent hugepages
1435
1436	  inactive_anon, active_anon, inactive_file, active_file, unevictable
1437		Amount of memory, swap-backed and filesystem-backed,
1438		on the internal memory management lists used by the
1439		page reclaim algorithm.
1440
1441		As these represent internal list state (eg. shmem pages are on anon
1442		memory management lists), inactive_foo + active_foo may not be equal to
1443		the value for the foo counter, since the foo counter is type-based, not
1444		list-based.
1445
1446	  slab_reclaimable
1447		Part of "slab" that might be reclaimed, such as
1448		dentries and inodes.
1449
1450	  slab_unreclaimable
1451		Part of "slab" that cannot be reclaimed on memory
1452		pressure.
1453
1454	  slab (npn)
1455		Amount of memory used for storing in-kernel data
1456		structures.
1457
1458	  workingset_refault_anon
1459		Number of refaults of previously evicted anonymous pages.
1460
1461	  workingset_refault_file
1462		Number of refaults of previously evicted file pages.
1463
1464	  workingset_activate_anon
1465		Number of refaulted anonymous pages that were immediately
1466		activated.
1467
1468	  workingset_activate_file
1469		Number of refaulted file pages that were immediately activated.
1470
1471	  workingset_restore_anon
1472		Number of restored anonymous pages which have been detected as
1473		an active workingset before they got reclaimed.
1474
1475	  workingset_restore_file
1476		Number of restored file pages which have been detected as an
1477		active workingset before they got reclaimed.
1478
1479	  workingset_nodereclaim
1480		Number of times a shadow node has been reclaimed
1481
1482	  pgscan (npn)
1483		Amount of scanned pages (in an inactive LRU list)
1484
1485	  pgsteal (npn)
1486		Amount of reclaimed pages
1487
1488	  pgscan_kswapd (npn)
1489		Amount of scanned pages by kswapd (in an inactive LRU list)
1490
1491	  pgscan_direct (npn)
1492		Amount of scanned pages directly  (in an inactive LRU list)
1493
1494	  pgscan_khugepaged (npn)
1495		Amount of scanned pages by khugepaged  (in an inactive LRU list)
1496
1497	  pgsteal_kswapd (npn)
1498		Amount of reclaimed pages by kswapd
1499
1500	  pgsteal_direct (npn)
1501		Amount of reclaimed pages directly
1502
1503	  pgsteal_khugepaged (npn)
1504		Amount of reclaimed pages by khugepaged
1505
1506	  pgfault (npn)
1507		Total number of page faults incurred
1508
1509	  pgmajfault (npn)
1510		Number of major page faults incurred
1511
1512	  pgrefill (npn)
1513		Amount of scanned pages (in an active LRU list)
1514
1515	  pgactivate (npn)
1516		Amount of pages moved to the active LRU list
1517
1518	  pgdeactivate (npn)
1519		Amount of pages moved to the inactive LRU list
1520
1521	  pglazyfree (npn)
1522		Amount of pages postponed to be freed under memory pressure
1523
1524	  pglazyfreed (npn)
1525		Amount of reclaimed lazyfree pages
1526
1527	  thp_fault_alloc (npn)
1528		Number of transparent hugepages which were allocated to satisfy
1529		a page fault. This counter is not present when CONFIG_TRANSPARENT_HUGEPAGE
1530                is not set.
1531
1532	  thp_collapse_alloc (npn)
1533		Number of transparent hugepages which were allocated to allow
1534		collapsing an existing range of pages. This counter is not
1535		present when CONFIG_TRANSPARENT_HUGEPAGE is not set.
1536
1537  memory.numa_stat
1538	A read-only nested-keyed file which exists on non-root cgroups.
1539
1540	This breaks down the cgroup's memory footprint into different
1541	types of memory, type-specific details, and other information
1542	per node on the state of the memory management system.
1543
1544	This is useful for providing visibility into the NUMA locality
1545	information within an memcg since the pages are allowed to be
1546	allocated from any physical node. One of the use case is evaluating
1547	application performance by combining this information with the
1548	application's CPU allocation.
1549
1550	All memory amounts are in bytes.
1551
1552	The output format of memory.numa_stat is::
1553
1554	  type N0=<bytes in node 0> N1=<bytes in node 1> ...
1555
1556	The entries are ordered to be human readable, and new entries
1557	can show up in the middle. Don't rely on items remaining in a
1558	fixed position; use the keys to look up specific values!
1559
1560	The entries can refer to the memory.stat.
1561
1562  memory.swap.current
1563	A read-only single value file which exists on non-root
1564	cgroups.
1565
1566	The total amount of swap currently being used by the cgroup
1567	and its descendants.
1568
1569  memory.swap.high
1570	A read-write single value file which exists on non-root
1571	cgroups.  The default is "max".
1572
1573	Swap usage throttle limit.  If a cgroup's swap usage exceeds
1574	this limit, all its further allocations will be throttled to
1575	allow userspace to implement custom out-of-memory procedures.
1576
1577	This limit marks a point of no return for the cgroup. It is NOT
1578	designed to manage the amount of swapping a workload does
1579	during regular operation. Compare to memory.swap.max, which
1580	prohibits swapping past a set amount, but lets the cgroup
1581	continue unimpeded as long as other memory can be reclaimed.
1582
1583	Healthy workloads are not expected to reach this limit.
1584
1585  memory.swap.peak
1586	A read-only single value file which exists on non-root
1587	cgroups.
1588
1589	The max swap usage recorded for the cgroup and its
1590	descendants since the creation of the cgroup.
1591
1592  memory.swap.max
1593	A read-write single value file which exists on non-root
1594	cgroups.  The default is "max".
1595
1596	Swap usage hard limit.  If a cgroup's swap usage reaches this
1597	limit, anonymous memory of the cgroup will not be swapped out.
1598
1599  memory.swap.events
1600	A read-only flat-keyed file which exists on non-root cgroups.
1601	The following entries are defined.  Unless specified
1602	otherwise, a value change in this file generates a file
1603	modified event.
1604
1605	  high
1606		The number of times the cgroup's swap usage was over
1607		the high threshold.
1608
1609	  max
1610		The number of times the cgroup's swap usage was about
1611		to go over the max boundary and swap allocation
1612		failed.
1613
1614	  fail
1615		The number of times swap allocation failed either
1616		because of running out of swap system-wide or max
1617		limit.
1618
1619	When reduced under the current usage, the existing swap
1620	entries are reclaimed gradually and the swap usage may stay
1621	higher than the limit for an extended period of time.  This
1622	reduces the impact on the workload and memory management.
1623
1624  memory.zswap.current
1625	A read-only single value file which exists on non-root
1626	cgroups.
1627
1628	The total amount of memory consumed by the zswap compression
1629	backend.
1630
1631  memory.zswap.max
1632	A read-write single value file which exists on non-root
1633	cgroups.  The default is "max".
1634
1635	Zswap usage hard limit. If a cgroup's zswap pool reaches this
1636	limit, it will refuse to take any more stores before existing
1637	entries fault back in or are written out to disk.
1638
1639  memory.pressure
1640	A read-only nested-keyed file.
1641
1642	Shows pressure stall information for memory. See
1643	:ref:`Documentation/accounting/psi.rst <psi>` for details.
1644
1645
1646Usage Guidelines
1647~~~~~~~~~~~~~~~~
1648
1649"memory.high" is the main mechanism to control memory usage.
1650Over-committing on high limit (sum of high limits > available memory)
1651and letting global memory pressure to distribute memory according to
1652usage is a viable strategy.
1653
1654Because breach of the high limit doesn't trigger the OOM killer but
1655throttles the offending cgroup, a management agent has ample
1656opportunities to monitor and take appropriate actions such as granting
1657more memory or terminating the workload.
1658
1659Determining whether a cgroup has enough memory is not trivial as
1660memory usage doesn't indicate whether the workload can benefit from
1661more memory.  For example, a workload which writes data received from
1662network to a file can use all available memory but can also operate as
1663performant with a small amount of memory.  A measure of memory
1664pressure - how much the workload is being impacted due to lack of
1665memory - is necessary to determine whether a workload needs more
1666memory; unfortunately, memory pressure monitoring mechanism isn't
1667implemented yet.
1668
1669
1670Memory Ownership
1671~~~~~~~~~~~~~~~~
1672
1673A memory area is charged to the cgroup which instantiated it and stays
1674charged to the cgroup until the area is released.  Migrating a process
1675to a different cgroup doesn't move the memory usages that it
1676instantiated while in the previous cgroup to the new cgroup.
1677
1678A memory area may be used by processes belonging to different cgroups.
1679To which cgroup the area will be charged is in-deterministic; however,
1680over time, the memory area is likely to end up in a cgroup which has
1681enough memory allowance to avoid high reclaim pressure.
1682
1683If a cgroup sweeps a considerable amount of memory which is expected
1684to be accessed repeatedly by other cgroups, it may make sense to use
1685POSIX_FADV_DONTNEED to relinquish the ownership of memory areas
1686belonging to the affected files to ensure correct memory ownership.
1687
1688
1689IO
1690--
1691
1692The "io" controller regulates the distribution of IO resources.  This
1693controller implements both weight based and absolute bandwidth or IOPS
1694limit distribution; however, weight based distribution is available
1695only if cfq-iosched is in use and neither scheme is available for
1696blk-mq devices.
1697
1698
1699IO Interface Files
1700~~~~~~~~~~~~~~~~~~
1701
1702  io.stat
1703	A read-only nested-keyed file.
1704
1705	Lines are keyed by $MAJ:$MIN device numbers and not ordered.
1706	The following nested keys are defined.
1707
1708	  ======	=====================
1709	  rbytes	Bytes read
1710	  wbytes	Bytes written
1711	  rios		Number of read IOs
1712	  wios		Number of write IOs
1713	  dbytes	Bytes discarded
1714	  dios		Number of discard IOs
1715	  ======	=====================
1716
1717	An example read output follows::
1718
1719	  8:16 rbytes=1459200 wbytes=314773504 rios=192 wios=353 dbytes=0 dios=0
1720	  8:0 rbytes=90430464 wbytes=299008000 rios=8950 wios=1252 dbytes=50331648 dios=3021
1721
1722  io.cost.qos
1723	A read-write nested-keyed file which exists only on the root
1724	cgroup.
1725
1726	This file configures the Quality of Service of the IO cost
1727	model based controller (CONFIG_BLK_CGROUP_IOCOST) which
1728	currently implements "io.weight" proportional control.  Lines
1729	are keyed by $MAJ:$MIN device numbers and not ordered.  The
1730	line for a given device is populated on the first write for
1731	the device on "io.cost.qos" or "io.cost.model".  The following
1732	nested keys are defined.
1733
1734	  ======	=====================================
1735	  enable	Weight-based control enable
1736	  ctrl		"auto" or "user"
1737	  rpct		Read latency percentile    [0, 100]
1738	  rlat		Read latency threshold
1739	  wpct		Write latency percentile   [0, 100]
1740	  wlat		Write latency threshold
1741	  min		Minimum scaling percentage [1, 10000]
1742	  max		Maximum scaling percentage [1, 10000]
1743	  ======	=====================================
1744
1745	The controller is disabled by default and can be enabled by
1746	setting "enable" to 1.  "rpct" and "wpct" parameters default
1747	to zero and the controller uses internal device saturation
1748	state to adjust the overall IO rate between "min" and "max".
1749
1750	When a better control quality is needed, latency QoS
1751	parameters can be configured.  For example::
1752
1753	  8:16 enable=1 ctrl=auto rpct=95.00 rlat=75000 wpct=95.00 wlat=150000 min=50.00 max=150.0
1754
1755	shows that on sdb, the controller is enabled, will consider
1756	the device saturated if the 95th percentile of read completion
1757	latencies is above 75ms or write 150ms, and adjust the overall
1758	IO issue rate between 50% and 150% accordingly.
1759
1760	The lower the saturation point, the better the latency QoS at
1761	the cost of aggregate bandwidth.  The narrower the allowed
1762	adjustment range between "min" and "max", the more conformant
1763	to the cost model the IO behavior.  Note that the IO issue
1764	base rate may be far off from 100% and setting "min" and "max"
1765	blindly can lead to a significant loss of device capacity or
1766	control quality.  "min" and "max" are useful for regulating
1767	devices which show wide temporary behavior changes - e.g. a
1768	ssd which accepts writes at the line speed for a while and
1769	then completely stalls for multiple seconds.
1770
1771	When "ctrl" is "auto", the parameters are controlled by the
1772	kernel and may change automatically.  Setting "ctrl" to "user"
1773	or setting any of the percentile and latency parameters puts
1774	it into "user" mode and disables the automatic changes.  The
1775	automatic mode can be restored by setting "ctrl" to "auto".
1776
1777  io.cost.model
1778	A read-write nested-keyed file which exists only on the root
1779	cgroup.
1780
1781	This file configures the cost model of the IO cost model based
1782	controller (CONFIG_BLK_CGROUP_IOCOST) which currently
1783	implements "io.weight" proportional control.  Lines are keyed
1784	by $MAJ:$MIN device numbers and not ordered.  The line for a
1785	given device is populated on the first write for the device on
1786	"io.cost.qos" or "io.cost.model".  The following nested keys
1787	are defined.
1788
1789	  =====		================================
1790	  ctrl		"auto" or "user"
1791	  model		The cost model in use - "linear"
1792	  =====		================================
1793
1794	When "ctrl" is "auto", the kernel may change all parameters
1795	dynamically.  When "ctrl" is set to "user" or any other
1796	parameters are written to, "ctrl" become "user" and the
1797	automatic changes are disabled.
1798
1799	When "model" is "linear", the following model parameters are
1800	defined.
1801
1802	  =============	========================================
1803	  [r|w]bps	The maximum sequential IO throughput
1804	  [r|w]seqiops	The maximum 4k sequential IOs per second
1805	  [r|w]randiops	The maximum 4k random IOs per second
1806	  =============	========================================
1807
1808	From the above, the builtin linear model determines the base
1809	costs of a sequential and random IO and the cost coefficient
1810	for the IO size.  While simple, this model can cover most
1811	common device classes acceptably.
1812
1813	The IO cost model isn't expected to be accurate in absolute
1814	sense and is scaled to the device behavior dynamically.
1815
1816	If needed, tools/cgroup/iocost_coef_gen.py can be used to
1817	generate device-specific coefficients.
1818
1819  io.weight
1820	A read-write flat-keyed file which exists on non-root cgroups.
1821	The default is "default 100".
1822
1823	The first line is the default weight applied to devices
1824	without specific override.  The rest are overrides keyed by
1825	$MAJ:$MIN device numbers and not ordered.  The weights are in
1826	the range [1, 10000] and specifies the relative amount IO time
1827	the cgroup can use in relation to its siblings.
1828
1829	The default weight can be updated by writing either "default
1830	$WEIGHT" or simply "$WEIGHT".  Overrides can be set by writing
1831	"$MAJ:$MIN $WEIGHT" and unset by writing "$MAJ:$MIN default".
1832
1833	An example read output follows::
1834
1835	  default 100
1836	  8:16 200
1837	  8:0 50
1838
1839  io.max
1840	A read-write nested-keyed file which exists on non-root
1841	cgroups.
1842
1843	BPS and IOPS based IO limit.  Lines are keyed by $MAJ:$MIN
1844	device numbers and not ordered.  The following nested keys are
1845	defined.
1846
1847	  =====		==================================
1848	  rbps		Max read bytes per second
1849	  wbps		Max write bytes per second
1850	  riops		Max read IO operations per second
1851	  wiops		Max write IO operations per second
1852	  =====		==================================
1853
1854	When writing, any number of nested key-value pairs can be
1855	specified in any order.  "max" can be specified as the value
1856	to remove a specific limit.  If the same key is specified
1857	multiple times, the outcome is undefined.
1858
1859	BPS and IOPS are measured in each IO direction and IOs are
1860	delayed if limit is reached.  Temporary bursts are allowed.
1861
1862	Setting read limit at 2M BPS and write at 120 IOPS for 8:16::
1863
1864	  echo "8:16 rbps=2097152 wiops=120" > io.max
1865
1866	Reading returns the following::
1867
1868	  8:16 rbps=2097152 wbps=max riops=max wiops=120
1869
1870	Write IOPS limit can be removed by writing the following::
1871
1872	  echo "8:16 wiops=max" > io.max
1873
1874	Reading now returns the following::
1875
1876	  8:16 rbps=2097152 wbps=max riops=max wiops=max
1877
1878  io.pressure
1879	A read-only nested-keyed file.
1880
1881	Shows pressure stall information for IO. See
1882	:ref:`Documentation/accounting/psi.rst <psi>` for details.
1883
1884
1885Writeback
1886~~~~~~~~~
1887
1888Page cache is dirtied through buffered writes and shared mmaps and
1889written asynchronously to the backing filesystem by the writeback
1890mechanism.  Writeback sits between the memory and IO domains and
1891regulates the proportion of dirty memory by balancing dirtying and
1892write IOs.
1893
1894The io controller, in conjunction with the memory controller,
1895implements control of page cache writeback IOs.  The memory controller
1896defines the memory domain that dirty memory ratio is calculated and
1897maintained for and the io controller defines the io domain which
1898writes out dirty pages for the memory domain.  Both system-wide and
1899per-cgroup dirty memory states are examined and the more restrictive
1900of the two is enforced.
1901
1902cgroup writeback requires explicit support from the underlying
1903filesystem.  Currently, cgroup writeback is implemented on ext2, ext4,
1904btrfs, f2fs, and xfs.  On other filesystems, all writeback IOs are
1905attributed to the root cgroup.
1906
1907There are inherent differences in memory and writeback management
1908which affects how cgroup ownership is tracked.  Memory is tracked per
1909page while writeback per inode.  For the purpose of writeback, an
1910inode is assigned to a cgroup and all IO requests to write dirty pages
1911from the inode are attributed to that cgroup.
1912
1913As cgroup ownership for memory is tracked per page, there can be pages
1914which are associated with different cgroups than the one the inode is
1915associated with.  These are called foreign pages.  The writeback
1916constantly keeps track of foreign pages and, if a particular foreign
1917cgroup becomes the majority over a certain period of time, switches
1918the ownership of the inode to that cgroup.
1919
1920While this model is enough for most use cases where a given inode is
1921mostly dirtied by a single cgroup even when the main writing cgroup
1922changes over time, use cases where multiple cgroups write to a single
1923inode simultaneously are not supported well.  In such circumstances, a
1924significant portion of IOs are likely to be attributed incorrectly.
1925As memory controller assigns page ownership on the first use and
1926doesn't update it until the page is released, even if writeback
1927strictly follows page ownership, multiple cgroups dirtying overlapping
1928areas wouldn't work as expected.  It's recommended to avoid such usage
1929patterns.
1930
1931The sysctl knobs which affect writeback behavior are applied to cgroup
1932writeback as follows.
1933
1934  vm.dirty_background_ratio, vm.dirty_ratio
1935	These ratios apply the same to cgroup writeback with the
1936	amount of available memory capped by limits imposed by the
1937	memory controller and system-wide clean memory.
1938
1939  vm.dirty_background_bytes, vm.dirty_bytes
1940	For cgroup writeback, this is calculated into ratio against
1941	total available memory and applied the same way as
1942	vm.dirty[_background]_ratio.
1943
1944
1945IO Latency
1946~~~~~~~~~~
1947
1948This is a cgroup v2 controller for IO workload protection.  You provide a group
1949with a latency target, and if the average latency exceeds that target the
1950controller will throttle any peers that have a lower latency target than the
1951protected workload.
1952
1953The limits are only applied at the peer level in the hierarchy.  This means that
1954in the diagram below, only groups A, B, and C will influence each other, and
1955groups D and F will influence each other.  Group G will influence nobody::
1956
1957			[root]
1958		/	   |		\
1959		A	   B		C
1960	       /  \        |
1961	      D    F	   G
1962
1963
1964So the ideal way to configure this is to set io.latency in groups A, B, and C.
1965Generally you do not want to set a value lower than the latency your device
1966supports.  Experiment to find the value that works best for your workload.
1967Start at higher than the expected latency for your device and watch the
1968avg_lat value in io.stat for your workload group to get an idea of the
1969latency you see during normal operation.  Use the avg_lat value as a basis for
1970your real setting, setting at 10-15% higher than the value in io.stat.
1971
1972How IO Latency Throttling Works
1973~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
1974
1975io.latency is work conserving; so as long as everybody is meeting their latency
1976target the controller doesn't do anything.  Once a group starts missing its
1977target it begins throttling any peer group that has a higher target than itself.
1978This throttling takes 2 forms:
1979
1980- Queue depth throttling.  This is the number of outstanding IO's a group is
1981  allowed to have.  We will clamp down relatively quickly, starting at no limit
1982  and going all the way down to 1 IO at a time.
1983
1984- Artificial delay induction.  There are certain types of IO that cannot be
1985  throttled without possibly adversely affecting higher priority groups.  This
1986  includes swapping and metadata IO.  These types of IO are allowed to occur
1987  normally, however they are "charged" to the originating group.  If the
1988  originating group is being throttled you will see the use_delay and delay
1989  fields in io.stat increase.  The delay value is how many microseconds that are
1990  being added to any process that runs in this group.  Because this number can
1991  grow quite large if there is a lot of swapping or metadata IO occurring we
1992  limit the individual delay events to 1 second at a time.
1993
1994Once the victimized group starts meeting its latency target again it will start
1995unthrottling any peer groups that were throttled previously.  If the victimized
1996group simply stops doing IO the global counter will unthrottle appropriately.
1997
1998IO Latency Interface Files
1999~~~~~~~~~~~~~~~~~~~~~~~~~~
2000
2001  io.latency
2002	This takes a similar format as the other controllers.
2003
2004		"MAJOR:MINOR target=<target time in microseconds>"
2005
2006  io.stat
2007	If the controller is enabled you will see extra stats in io.stat in
2008	addition to the normal ones.
2009
2010	  depth
2011		This is the current queue depth for the group.
2012
2013	  avg_lat
2014		This is an exponential moving average with a decay rate of 1/exp
2015		bound by the sampling interval.  The decay rate interval can be
2016		calculated by multiplying the win value in io.stat by the
2017		corresponding number of samples based on the win value.
2018
2019	  win
2020		The sampling window size in milliseconds.  This is the minimum
2021		duration of time between evaluation events.  Windows only elapse
2022		with IO activity.  Idle periods extend the most recent window.
2023
2024IO Priority
2025~~~~~~~~~~~
2026
2027A single attribute controls the behavior of the I/O priority cgroup policy,
2028namely the blkio.prio.class attribute. The following values are accepted for
2029that attribute:
2030
2031  no-change
2032	Do not modify the I/O priority class.
2033
2034  none-to-rt
2035	For requests that do not have an I/O priority class (NONE),
2036	change the I/O priority class into RT. Do not modify
2037	the I/O priority class of other requests.
2038
2039  restrict-to-be
2040	For requests that do not have an I/O priority class or that have I/O
2041	priority class RT, change it into BE. Do not modify the I/O priority
2042	class of requests that have priority class IDLE.
2043
2044  idle
2045	Change the I/O priority class of all requests into IDLE, the lowest
2046	I/O priority class.
2047
2048The following numerical values are associated with the I/O priority policies:
2049
2050+-------------+---+
2051| no-change   | 0 |
2052+-------------+---+
2053| none-to-rt  | 1 |
2054+-------------+---+
2055| rt-to-be    | 2 |
2056+-------------+---+
2057| all-to-idle | 3 |
2058+-------------+---+
2059
2060The numerical value that corresponds to each I/O priority class is as follows:
2061
2062+-------------------------------+---+
2063| IOPRIO_CLASS_NONE             | 0 |
2064+-------------------------------+---+
2065| IOPRIO_CLASS_RT (real-time)   | 1 |
2066+-------------------------------+---+
2067| IOPRIO_CLASS_BE (best effort) | 2 |
2068+-------------------------------+---+
2069| IOPRIO_CLASS_IDLE             | 3 |
2070+-------------------------------+---+
2071
2072The algorithm to set the I/O priority class for a request is as follows:
2073
2074- Translate the I/O priority class policy into a number.
2075- Change the request I/O priority class into the maximum of the I/O priority
2076  class policy number and the numerical I/O priority class.
2077
2078PID
2079---
2080
2081The process number controller is used to allow a cgroup to stop any
2082new tasks from being fork()'d or clone()'d after a specified limit is
2083reached.
2084
2085The number of tasks in a cgroup can be exhausted in ways which other
2086controllers cannot prevent, thus warranting its own controller.  For
2087example, a fork bomb is likely to exhaust the number of tasks before
2088hitting memory restrictions.
2089
2090Note that PIDs used in this controller refer to TIDs, process IDs as
2091used by the kernel.
2092
2093
2094PID Interface Files
2095~~~~~~~~~~~~~~~~~~~
2096
2097  pids.max
2098	A read-write single value file which exists on non-root
2099	cgroups.  The default is "max".
2100
2101	Hard limit of number of processes.
2102
2103  pids.current
2104	A read-only single value file which exists on all cgroups.
2105
2106	The number of processes currently in the cgroup and its
2107	descendants.
2108
2109Organisational operations are not blocked by cgroup policies, so it is
2110possible to have pids.current > pids.max.  This can be done by either
2111setting the limit to be smaller than pids.current, or attaching enough
2112processes to the cgroup such that pids.current is larger than
2113pids.max.  However, it is not possible to violate a cgroup PID policy
2114through fork() or clone(). These will return -EAGAIN if the creation
2115of a new process would cause a cgroup policy to be violated.
2116
2117
2118Cpuset
2119------
2120
2121The "cpuset" controller provides a mechanism for constraining
2122the CPU and memory node placement of tasks to only the resources
2123specified in the cpuset interface files in a task's current cgroup.
2124This is especially valuable on large NUMA systems where placing jobs
2125on properly sized subsets of the systems with careful processor and
2126memory placement to reduce cross-node memory access and contention
2127can improve overall system performance.
2128
2129The "cpuset" controller is hierarchical.  That means the controller
2130cannot use CPUs or memory nodes not allowed in its parent.
2131
2132
2133Cpuset Interface Files
2134~~~~~~~~~~~~~~~~~~~~~~
2135
2136  cpuset.cpus
2137	A read-write multiple values file which exists on non-root
2138	cpuset-enabled cgroups.
2139
2140	It lists the requested CPUs to be used by tasks within this
2141	cgroup.  The actual list of CPUs to be granted, however, is
2142	subjected to constraints imposed by its parent and can differ
2143	from the requested CPUs.
2144
2145	The CPU numbers are comma-separated numbers or ranges.
2146	For example::
2147
2148	  # cat cpuset.cpus
2149	  0-4,6,8-10
2150
2151	An empty value indicates that the cgroup is using the same
2152	setting as the nearest cgroup ancestor with a non-empty
2153	"cpuset.cpus" or all the available CPUs if none is found.
2154
2155	The value of "cpuset.cpus" stays constant until the next update
2156	and won't be affected by any CPU hotplug events.
2157
2158  cpuset.cpus.effective
2159	A read-only multiple values file which exists on all
2160	cpuset-enabled cgroups.
2161
2162	It lists the onlined CPUs that are actually granted to this
2163	cgroup by its parent.  These CPUs are allowed to be used by
2164	tasks within the current cgroup.
2165
2166	If "cpuset.cpus" is empty, the "cpuset.cpus.effective" file shows
2167	all the CPUs from the parent cgroup that can be available to
2168	be used by this cgroup.  Otherwise, it should be a subset of
2169	"cpuset.cpus" unless none of the CPUs listed in "cpuset.cpus"
2170	can be granted.  In this case, it will be treated just like an
2171	empty "cpuset.cpus".
2172
2173	Its value will be affected by CPU hotplug events.
2174
2175  cpuset.mems
2176	A read-write multiple values file which exists on non-root
2177	cpuset-enabled cgroups.
2178
2179	It lists the requested memory nodes to be used by tasks within
2180	this cgroup.  The actual list of memory nodes granted, however,
2181	is subjected to constraints imposed by its parent and can differ
2182	from the requested memory nodes.
2183
2184	The memory node numbers are comma-separated numbers or ranges.
2185	For example::
2186
2187	  # cat cpuset.mems
2188	  0-1,3
2189
2190	An empty value indicates that the cgroup is using the same
2191	setting as the nearest cgroup ancestor with a non-empty
2192	"cpuset.mems" or all the available memory nodes if none
2193	is found.
2194
2195	The value of "cpuset.mems" stays constant until the next update
2196	and won't be affected by any memory nodes hotplug events.
2197
2198	Setting a non-empty value to "cpuset.mems" causes memory of
2199	tasks within the cgroup to be migrated to the designated nodes if
2200	they are currently using memory outside of the designated nodes.
2201
2202	There is a cost for this memory migration.  The migration
2203	may not be complete and some memory pages may be left behind.
2204	So it is recommended that "cpuset.mems" should be set properly
2205	before spawning new tasks into the cpuset.  Even if there is
2206	a need to change "cpuset.mems" with active tasks, it shouldn't
2207	be done frequently.
2208
2209  cpuset.mems.effective
2210	A read-only multiple values file which exists on all
2211	cpuset-enabled cgroups.
2212
2213	It lists the onlined memory nodes that are actually granted to
2214	this cgroup by its parent. These memory nodes are allowed to
2215	be used by tasks within the current cgroup.
2216
2217	If "cpuset.mems" is empty, it shows all the memory nodes from the
2218	parent cgroup that will be available to be used by this cgroup.
2219	Otherwise, it should be a subset of "cpuset.mems" unless none of
2220	the memory nodes listed in "cpuset.mems" can be granted.  In this
2221	case, it will be treated just like an empty "cpuset.mems".
2222
2223	Its value will be affected by memory nodes hotplug events.
2224
2225  cpuset.cpus.partition
2226	A read-write single value file which exists on non-root
2227	cpuset-enabled cgroups.  This flag is owned by the parent cgroup
2228	and is not delegatable.
2229
2230	It accepts only the following input values when written to.
2231
2232	  ==========	=====================================
2233	  "member"	Non-root member of a partition
2234	  "root"	Partition root
2235	  "isolated"	Partition root without load balancing
2236	  ==========	=====================================
2237
2238	The root cgroup is always a partition root and its state
2239	cannot be changed.  All other non-root cgroups start out as
2240	"member".
2241
2242	When set to "root", the current cgroup is the root of a new
2243	partition or scheduling domain that comprises itself and all
2244	its descendants except those that are separate partition roots
2245	themselves and their descendants.
2246
2247	When set to "isolated", the CPUs in that partition root will
2248	be in an isolated state without any load balancing from the
2249	scheduler.  Tasks placed in such a partition with multiple
2250	CPUs should be carefully distributed and bound to each of the
2251	individual CPUs for optimal performance.
2252
2253	The value shown in "cpuset.cpus.effective" of a partition root
2254	is the CPUs that the partition root can dedicate to a potential
2255	new child partition root. The new child subtracts available
2256	CPUs from its parent "cpuset.cpus.effective".
2257
2258	A partition root ("root" or "isolated") can be in one of the
2259	two possible states - valid or invalid.  An invalid partition
2260	root is in a degraded state where some state information may
2261	be retained, but behaves more like a "member".
2262
2263	All possible state transitions among "member", "root" and
2264	"isolated" are allowed.
2265
2266	On read, the "cpuset.cpus.partition" file can show the following
2267	values.
2268
2269	  =============================	=====================================
2270	  "member"			Non-root member of a partition
2271	  "root"			Partition root
2272	  "isolated"			Partition root without load balancing
2273	  "root invalid (<reason>)"	Invalid partition root
2274	  "isolated invalid (<reason>)"	Invalid isolated partition root
2275	  =============================	=====================================
2276
2277	In the case of an invalid partition root, a descriptive string on
2278	why the partition is invalid is included within parentheses.
2279
2280	For a partition root to become valid, the following conditions
2281	must be met.
2282
2283	1) The "cpuset.cpus" is exclusive with its siblings , i.e. they
2284	   are not shared by any of its siblings (exclusivity rule).
2285	2) The parent cgroup is a valid partition root.
2286	3) The "cpuset.cpus" is not empty and must contain at least
2287	   one of the CPUs from parent's "cpuset.cpus", i.e. they overlap.
2288	4) The "cpuset.cpus.effective" cannot be empty unless there is
2289	   no task associated with this partition.
2290
2291	External events like hotplug or changes to "cpuset.cpus" can
2292	cause a valid partition root to become invalid and vice versa.
2293	Note that a task cannot be moved to a cgroup with empty
2294	"cpuset.cpus.effective".
2295
2296	For a valid partition root with the sibling cpu exclusivity
2297	rule enabled, changes made to "cpuset.cpus" that violate the
2298	exclusivity rule will invalidate the partition as well as its
2299	sibling partitions with conflicting cpuset.cpus values. So
2300	care must be taking in changing "cpuset.cpus".
2301
2302	A valid non-root parent partition may distribute out all its CPUs
2303	to its child partitions when there is no task associated with it.
2304
2305	Care must be taken to change a valid partition root to
2306	"member" as all its child partitions, if present, will become
2307	invalid causing disruption to tasks running in those child
2308	partitions. These inactivated partitions could be recovered if
2309	their parent is switched back to a partition root with a proper
2310	set of "cpuset.cpus".
2311
2312	Poll and inotify events are triggered whenever the state of
2313	"cpuset.cpus.partition" changes.  That includes changes caused
2314	by write to "cpuset.cpus.partition", cpu hotplug or other
2315	changes that modify the validity status of the partition.
2316	This will allow user space agents to monitor unexpected changes
2317	to "cpuset.cpus.partition" without the need to do continuous
2318	polling.
2319
2320
2321Device controller
2322-----------------
2323
2324Device controller manages access to device files. It includes both
2325creation of new device files (using mknod), and access to the
2326existing device files.
2327
2328Cgroup v2 device controller has no interface files and is implemented
2329on top of cgroup BPF. To control access to device files, a user may
2330create bpf programs of type BPF_PROG_TYPE_CGROUP_DEVICE and attach
2331them to cgroups with BPF_CGROUP_DEVICE flag. On an attempt to access a
2332device file, corresponding BPF programs will be executed, and depending
2333on the return value the attempt will succeed or fail with -EPERM.
2334
2335A BPF_PROG_TYPE_CGROUP_DEVICE program takes a pointer to the
2336bpf_cgroup_dev_ctx structure, which describes the device access attempt:
2337access type (mknod/read/write) and device (type, major and minor numbers).
2338If the program returns 0, the attempt fails with -EPERM, otherwise it
2339succeeds.
2340
2341An example of BPF_PROG_TYPE_CGROUP_DEVICE program may be found in
2342tools/testing/selftests/bpf/progs/dev_cgroup.c in the kernel source tree.
2343
2344
2345RDMA
2346----
2347
2348The "rdma" controller regulates the distribution and accounting of
2349RDMA resources.
2350
2351RDMA Interface Files
2352~~~~~~~~~~~~~~~~~~~~
2353
2354  rdma.max
2355	A readwrite nested-keyed file that exists for all the cgroups
2356	except root that describes current configured resource limit
2357	for a RDMA/IB device.
2358
2359	Lines are keyed by device name and are not ordered.
2360	Each line contains space separated resource name and its configured
2361	limit that can be distributed.
2362
2363	The following nested keys are defined.
2364
2365	  ==========	=============================
2366	  hca_handle	Maximum number of HCA Handles
2367	  hca_object 	Maximum number of HCA Objects
2368	  ==========	=============================
2369
2370	An example for mlx4 and ocrdma device follows::
2371
2372	  mlx4_0 hca_handle=2 hca_object=2000
2373	  ocrdma1 hca_handle=3 hca_object=max
2374
2375  rdma.current
2376	A read-only file that describes current resource usage.
2377	It exists for all the cgroup except root.
2378
2379	An example for mlx4 and ocrdma device follows::
2380
2381	  mlx4_0 hca_handle=1 hca_object=20
2382	  ocrdma1 hca_handle=1 hca_object=23
2383
2384HugeTLB
2385-------
2386
2387The HugeTLB controller allows to limit the HugeTLB usage per control group and
2388enforces the controller limit during page fault.
2389
2390HugeTLB Interface Files
2391~~~~~~~~~~~~~~~~~~~~~~~
2392
2393  hugetlb.<hugepagesize>.current
2394	Show current usage for "hugepagesize" hugetlb.  It exists for all
2395	the cgroup except root.
2396
2397  hugetlb.<hugepagesize>.max
2398	Set/show the hard limit of "hugepagesize" hugetlb usage.
2399	The default value is "max".  It exists for all the cgroup except root.
2400
2401  hugetlb.<hugepagesize>.events
2402	A read-only flat-keyed file which exists on non-root cgroups.
2403
2404	  max
2405		The number of allocation failure due to HugeTLB limit
2406
2407  hugetlb.<hugepagesize>.events.local
2408	Similar to hugetlb.<hugepagesize>.events but the fields in the file
2409	are local to the cgroup i.e. not hierarchical. The file modified event
2410	generated on this file reflects only the local events.
2411
2412  hugetlb.<hugepagesize>.numa_stat
2413	Similar to memory.numa_stat, it shows the numa information of the
2414        hugetlb pages of <hugepagesize> in this cgroup.  Only active in
2415        use hugetlb pages are included.  The per-node values are in bytes.
2416
2417Misc
2418----
2419
2420The Miscellaneous cgroup provides the resource limiting and tracking
2421mechanism for the scalar resources which cannot be abstracted like the other
2422cgroup resources. Controller is enabled by the CONFIG_CGROUP_MISC config
2423option.
2424
2425A resource can be added to the controller via enum misc_res_type{} in the
2426include/linux/misc_cgroup.h file and the corresponding name via misc_res_name[]
2427in the kernel/cgroup/misc.c file. Provider of the resource must set its
2428capacity prior to using the resource by calling misc_cg_set_capacity().
2429
2430Once a capacity is set then the resource usage can be updated using charge and
2431uncharge APIs. All of the APIs to interact with misc controller are in
2432include/linux/misc_cgroup.h.
2433
2434Misc Interface Files
2435~~~~~~~~~~~~~~~~~~~~
2436
2437Miscellaneous controller provides 3 interface files. If two misc resources (res_a and res_b) are registered then:
2438
2439  misc.capacity
2440        A read-only flat-keyed file shown only in the root cgroup.  It shows
2441        miscellaneous scalar resources available on the platform along with
2442        their quantities::
2443
2444	  $ cat misc.capacity
2445	  res_a 50
2446	  res_b 10
2447
2448  misc.current
2449        A read-only flat-keyed file shown in the non-root cgroups.  It shows
2450        the current usage of the resources in the cgroup and its children.::
2451
2452	  $ cat misc.current
2453	  res_a 3
2454	  res_b 0
2455
2456  misc.max
2457        A read-write flat-keyed file shown in the non root cgroups. Allowed
2458        maximum usage of the resources in the cgroup and its children.::
2459
2460	  $ cat misc.max
2461	  res_a max
2462	  res_b 4
2463
2464	Limit can be set by::
2465
2466	  # echo res_a 1 > misc.max
2467
2468	Limit can be set to max by::
2469
2470	  # echo res_a max > misc.max
2471
2472        Limits can be set higher than the capacity value in the misc.capacity
2473        file.
2474
2475  misc.events
2476	A read-only flat-keyed file which exists on non-root cgroups. The
2477	following entries are defined. Unless specified otherwise, a value
2478	change in this file generates a file modified event. All fields in
2479	this file are hierarchical.
2480
2481	  max
2482		The number of times the cgroup's resource usage was
2483		about to go over the max boundary.
2484
2485Migration and Ownership
2486~~~~~~~~~~~~~~~~~~~~~~~
2487
2488A miscellaneous scalar resource is charged to the cgroup in which it is used
2489first, and stays charged to that cgroup until that resource is freed. Migrating
2490a process to a different cgroup does not move the charge to the destination
2491cgroup where the process has moved.
2492
2493Others
2494------
2495
2496perf_event
2497~~~~~~~~~~
2498
2499perf_event controller, if not mounted on a legacy hierarchy, is
2500automatically enabled on the v2 hierarchy so that perf events can
2501always be filtered by cgroup v2 path.  The controller can still be
2502moved to a legacy hierarchy after v2 hierarchy is populated.
2503
2504
2505Non-normative information
2506-------------------------
2507
2508This section contains information that isn't considered to be a part of
2509the stable kernel API and so is subject to change.
2510
2511
2512CPU controller root cgroup process behaviour
2513~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
2514
2515When distributing CPU cycles in the root cgroup each thread in this
2516cgroup is treated as if it was hosted in a separate child cgroup of the
2517root cgroup. This child cgroup weight is dependent on its thread nice
2518level.
2519
2520For details of this mapping see sched_prio_to_weight array in
2521kernel/sched/core.c file (values from this array should be scaled
2522appropriately so the neutral - nice 0 - value is 100 instead of 1024).
2523
2524
2525IO controller root cgroup process behaviour
2526~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
2527
2528Root cgroup processes are hosted in an implicit leaf child node.
2529When distributing IO resources this implicit child node is taken into
2530account as if it was a normal child cgroup of the root cgroup with a
2531weight value of 200.
2532
2533
2534Namespace
2535=========
2536
2537Basics
2538------
2539
2540cgroup namespace provides a mechanism to virtualize the view of the
2541"/proc/$PID/cgroup" file and cgroup mounts.  The CLONE_NEWCGROUP clone
2542flag can be used with clone(2) and unshare(2) to create a new cgroup
2543namespace.  The process running inside the cgroup namespace will have
2544its "/proc/$PID/cgroup" output restricted to cgroupns root.  The
2545cgroupns root is the cgroup of the process at the time of creation of
2546the cgroup namespace.
2547
2548Without cgroup namespace, the "/proc/$PID/cgroup" file shows the
2549complete path of the cgroup of a process.  In a container setup where
2550a set of cgroups and namespaces are intended to isolate processes the
2551"/proc/$PID/cgroup" file may leak potential system level information
2552to the isolated processes.  For example::
2553
2554  # cat /proc/self/cgroup
2555  0::/batchjobs/container_id1
2556
2557The path '/batchjobs/container_id1' can be considered as system-data
2558and undesirable to expose to the isolated processes.  cgroup namespace
2559can be used to restrict visibility of this path.  For example, before
2560creating a cgroup namespace, one would see::
2561
2562  # ls -l /proc/self/ns/cgroup
2563  lrwxrwxrwx 1 root root 0 2014-07-15 10:37 /proc/self/ns/cgroup -> cgroup:[4026531835]
2564  # cat /proc/self/cgroup
2565  0::/batchjobs/container_id1
2566
2567After unsharing a new namespace, the view changes::
2568
2569  # ls -l /proc/self/ns/cgroup
2570  lrwxrwxrwx 1 root root 0 2014-07-15 10:35 /proc/self/ns/cgroup -> cgroup:[4026532183]
2571  # cat /proc/self/cgroup
2572  0::/
2573
2574When some thread from a multi-threaded process unshares its cgroup
2575namespace, the new cgroupns gets applied to the entire process (all
2576the threads).  This is natural for the v2 hierarchy; however, for the
2577legacy hierarchies, this may be unexpected.
2578
2579A cgroup namespace is alive as long as there are processes inside or
2580mounts pinning it.  When the last usage goes away, the cgroup
2581namespace is destroyed.  The cgroupns root and the actual cgroups
2582remain.
2583
2584
2585The Root and Views
2586------------------
2587
2588The 'cgroupns root' for a cgroup namespace is the cgroup in which the
2589process calling unshare(2) is running.  For example, if a process in
2590/batchjobs/container_id1 cgroup calls unshare, cgroup
2591/batchjobs/container_id1 becomes the cgroupns root.  For the
2592init_cgroup_ns, this is the real root ('/') cgroup.
2593
2594The cgroupns root cgroup does not change even if the namespace creator
2595process later moves to a different cgroup::
2596
2597  # ~/unshare -c # unshare cgroupns in some cgroup
2598  # cat /proc/self/cgroup
2599  0::/
2600  # mkdir sub_cgrp_1
2601  # echo 0 > sub_cgrp_1/cgroup.procs
2602  # cat /proc/self/cgroup
2603  0::/sub_cgrp_1
2604
2605Each process gets its namespace-specific view of "/proc/$PID/cgroup"
2606
2607Processes running inside the cgroup namespace will be able to see
2608cgroup paths (in /proc/self/cgroup) only inside their root cgroup.
2609From within an unshared cgroupns::
2610
2611  # sleep 100000 &
2612  [1] 7353
2613  # echo 7353 > sub_cgrp_1/cgroup.procs
2614  # cat /proc/7353/cgroup
2615  0::/sub_cgrp_1
2616
2617From the initial cgroup namespace, the real cgroup path will be
2618visible::
2619
2620  $ cat /proc/7353/cgroup
2621  0::/batchjobs/container_id1/sub_cgrp_1
2622
2623From a sibling cgroup namespace (that is, a namespace rooted at a
2624different cgroup), the cgroup path relative to its own cgroup
2625namespace root will be shown.  For instance, if PID 7353's cgroup
2626namespace root is at '/batchjobs/container_id2', then it will see::
2627
2628  # cat /proc/7353/cgroup
2629  0::/../container_id2/sub_cgrp_1
2630
2631Note that the relative path always starts with '/' to indicate that
2632its relative to the cgroup namespace root of the caller.
2633
2634
2635Migration and setns(2)
2636----------------------
2637
2638Processes inside a cgroup namespace can move into and out of the
2639namespace root if they have proper access to external cgroups.  For
2640example, from inside a namespace with cgroupns root at
2641/batchjobs/container_id1, and assuming that the global hierarchy is
2642still accessible inside cgroupns::
2643
2644  # cat /proc/7353/cgroup
2645  0::/sub_cgrp_1
2646  # echo 7353 > batchjobs/container_id2/cgroup.procs
2647  # cat /proc/7353/cgroup
2648  0::/../container_id2
2649
2650Note that this kind of setup is not encouraged.  A task inside cgroup
2651namespace should only be exposed to its own cgroupns hierarchy.
2652
2653setns(2) to another cgroup namespace is allowed when:
2654
2655(a) the process has CAP_SYS_ADMIN against its current user namespace
2656(b) the process has CAP_SYS_ADMIN against the target cgroup
2657    namespace's userns
2658
2659No implicit cgroup changes happen with attaching to another cgroup
2660namespace.  It is expected that the someone moves the attaching
2661process under the target cgroup namespace root.
2662
2663
2664Interaction with Other Namespaces
2665---------------------------------
2666
2667Namespace specific cgroup hierarchy can be mounted by a process
2668running inside a non-init cgroup namespace::
2669
2670  # mount -t cgroup2 none $MOUNT_POINT
2671
2672This will mount the unified cgroup hierarchy with cgroupns root as the
2673filesystem root.  The process needs CAP_SYS_ADMIN against its user and
2674mount namespaces.
2675
2676The virtualization of /proc/self/cgroup file combined with restricting
2677the view of cgroup hierarchy by namespace-private cgroupfs mount
2678provides a properly isolated cgroup view inside the container.
2679
2680
2681Information on Kernel Programming
2682=================================
2683
2684This section contains kernel programming information in the areas
2685where interacting with cgroup is necessary.  cgroup core and
2686controllers are not covered.
2687
2688
2689Filesystem Support for Writeback
2690--------------------------------
2691
2692A filesystem can support cgroup writeback by updating
2693address_space_operations->writepage[s]() to annotate bio's using the
2694following two functions.
2695
2696  wbc_init_bio(@wbc, @bio)
2697	Should be called for each bio carrying writeback data and
2698	associates the bio with the inode's owner cgroup and the
2699	corresponding request queue.  This must be called after
2700	a queue (device) has been associated with the bio and
2701	before submission.
2702
2703  wbc_account_cgroup_owner(@wbc, @page, @bytes)
2704	Should be called for each data segment being written out.
2705	While this function doesn't care exactly when it's called
2706	during the writeback session, it's the easiest and most
2707	natural to call it as data segments are added to a bio.
2708
2709With writeback bio's annotated, cgroup support can be enabled per
2710super_block by setting SB_I_CGROUPWB in ->s_iflags.  This allows for
2711selective disabling of cgroup writeback support which is helpful when
2712certain filesystem features, e.g. journaled data mode, are
2713incompatible.
2714
2715wbc_init_bio() binds the specified bio to its cgroup.  Depending on
2716the configuration, the bio may be executed at a lower priority and if
2717the writeback session is holding shared resources, e.g. a journal
2718entry, may lead to priority inversion.  There is no one easy solution
2719for the problem.  Filesystems can try to work around specific problem
2720cases by skipping wbc_init_bio() and using bio_associate_blkg()
2721directly.
2722
2723
2724Deprecated v1 Core Features
2725===========================
2726
2727- Multiple hierarchies including named ones are not supported.
2728
2729- All v1 mount options are not supported.
2730
2731- The "tasks" file is removed and "cgroup.procs" is not sorted.
2732
2733- "cgroup.clone_children" is removed.
2734
2735- /proc/cgroups is meaningless for v2.  Use "cgroup.controllers" file
2736  at the root instead.
2737
2738
2739Issues with v1 and Rationales for v2
2740====================================
2741
2742Multiple Hierarchies
2743--------------------
2744
2745cgroup v1 allowed an arbitrary number of hierarchies and each
2746hierarchy could host any number of controllers.  While this seemed to
2747provide a high level of flexibility, it wasn't useful in practice.
2748
2749For example, as there is only one instance of each controller, utility
2750type controllers such as freezer which can be useful in all
2751hierarchies could only be used in one.  The issue is exacerbated by
2752the fact that controllers couldn't be moved to another hierarchy once
2753hierarchies were populated.  Another issue was that all controllers
2754bound to a hierarchy were forced to have exactly the same view of the
2755hierarchy.  It wasn't possible to vary the granularity depending on
2756the specific controller.
2757
2758In practice, these issues heavily limited which controllers could be
2759put on the same hierarchy and most configurations resorted to putting
2760each controller on its own hierarchy.  Only closely related ones, such
2761as the cpu and cpuacct controllers, made sense to be put on the same
2762hierarchy.  This often meant that userland ended up managing multiple
2763similar hierarchies repeating the same steps on each hierarchy
2764whenever a hierarchy management operation was necessary.
2765
2766Furthermore, support for multiple hierarchies came at a steep cost.
2767It greatly complicated cgroup core implementation but more importantly
2768the support for multiple hierarchies restricted how cgroup could be
2769used in general and what controllers was able to do.
2770
2771There was no limit on how many hierarchies there might be, which meant
2772that a thread's cgroup membership couldn't be described in finite
2773length.  The key might contain any number of entries and was unlimited
2774in length, which made it highly awkward to manipulate and led to
2775addition of controllers which existed only to identify membership,
2776which in turn exacerbated the original problem of proliferating number
2777of hierarchies.
2778
2779Also, as a controller couldn't have any expectation regarding the
2780topologies of hierarchies other controllers might be on, each
2781controller had to assume that all other controllers were attached to
2782completely orthogonal hierarchies.  This made it impossible, or at
2783least very cumbersome, for controllers to cooperate with each other.
2784
2785In most use cases, putting controllers on hierarchies which are
2786completely orthogonal to each other isn't necessary.  What usually is
2787called for is the ability to have differing levels of granularity
2788depending on the specific controller.  In other words, hierarchy may
2789be collapsed from leaf towards root when viewed from specific
2790controllers.  For example, a given configuration might not care about
2791how memory is distributed beyond a certain level while still wanting
2792to control how CPU cycles are distributed.
2793
2794
2795Thread Granularity
2796------------------
2797
2798cgroup v1 allowed threads of a process to belong to different cgroups.
2799This didn't make sense for some controllers and those controllers
2800ended up implementing different ways to ignore such situations but
2801much more importantly it blurred the line between API exposed to
2802individual applications and system management interface.
2803
2804Generally, in-process knowledge is available only to the process
2805itself; thus, unlike service-level organization of processes,
2806categorizing threads of a process requires active participation from
2807the application which owns the target process.
2808
2809cgroup v1 had an ambiguously defined delegation model which got abused
2810in combination with thread granularity.  cgroups were delegated to
2811individual applications so that they can create and manage their own
2812sub-hierarchies and control resource distributions along them.  This
2813effectively raised cgroup to the status of a syscall-like API exposed
2814to lay programs.
2815
2816First of all, cgroup has a fundamentally inadequate interface to be
2817exposed this way.  For a process to access its own knobs, it has to
2818extract the path on the target hierarchy from /proc/self/cgroup,
2819construct the path by appending the name of the knob to the path, open
2820and then read and/or write to it.  This is not only extremely clunky
2821and unusual but also inherently racy.  There is no conventional way to
2822define transaction across the required steps and nothing can guarantee
2823that the process would actually be operating on its own sub-hierarchy.
2824
2825cgroup controllers implemented a number of knobs which would never be
2826accepted as public APIs because they were just adding control knobs to
2827system-management pseudo filesystem.  cgroup ended up with interface
2828knobs which were not properly abstracted or refined and directly
2829revealed kernel internal details.  These knobs got exposed to
2830individual applications through the ill-defined delegation mechanism
2831effectively abusing cgroup as a shortcut to implementing public APIs
2832without going through the required scrutiny.
2833
2834This was painful for both userland and kernel.  Userland ended up with
2835misbehaving and poorly abstracted interfaces and kernel exposing and
2836locked into constructs inadvertently.
2837
2838
2839Competition Between Inner Nodes and Threads
2840-------------------------------------------
2841
2842cgroup v1 allowed threads to be in any cgroups which created an
2843interesting problem where threads belonging to a parent cgroup and its
2844children cgroups competed for resources.  This was nasty as two
2845different types of entities competed and there was no obvious way to
2846settle it.  Different controllers did different things.
2847
2848The cpu controller considered threads and cgroups as equivalents and
2849mapped nice levels to cgroup weights.  This worked for some cases but
2850fell flat when children wanted to be allocated specific ratios of CPU
2851cycles and the number of internal threads fluctuated - the ratios
2852constantly changed as the number of competing entities fluctuated.
2853There also were other issues.  The mapping from nice level to weight
2854wasn't obvious or universal, and there were various other knobs which
2855simply weren't available for threads.
2856
2857The io controller implicitly created a hidden leaf node for each
2858cgroup to host the threads.  The hidden leaf had its own copies of all
2859the knobs with ``leaf_`` prefixed.  While this allowed equivalent
2860control over internal threads, it was with serious drawbacks.  It
2861always added an extra layer of nesting which wouldn't be necessary
2862otherwise, made the interface messy and significantly complicated the
2863implementation.
2864
2865The memory controller didn't have a way to control what happened
2866between internal tasks and child cgroups and the behavior was not
2867clearly defined.  There were attempts to add ad-hoc behaviors and
2868knobs to tailor the behavior to specific workloads which would have
2869led to problems extremely difficult to resolve in the long term.
2870
2871Multiple controllers struggled with internal tasks and came up with
2872different ways to deal with it; unfortunately, all the approaches were
2873severely flawed and, furthermore, the widely different behaviors
2874made cgroup as a whole highly inconsistent.
2875
2876This clearly is a problem which needs to be addressed from cgroup core
2877in a uniform way.
2878
2879
2880Other Interface Issues
2881----------------------
2882
2883cgroup v1 grew without oversight and developed a large number of
2884idiosyncrasies and inconsistencies.  One issue on the cgroup core side
2885was how an empty cgroup was notified - a userland helper binary was
2886forked and executed for each event.  The event delivery wasn't
2887recursive or delegatable.  The limitations of the mechanism also led
2888to in-kernel event delivery filtering mechanism further complicating
2889the interface.
2890
2891Controller interfaces were problematic too.  An extreme example is
2892controllers completely ignoring hierarchical organization and treating
2893all cgroups as if they were all located directly under the root
2894cgroup.  Some controllers exposed a large amount of inconsistent
2895implementation details to userland.
2896
2897There also was no consistency across controllers.  When a new cgroup
2898was created, some controllers defaulted to not imposing extra
2899restrictions while others disallowed any resource usage until
2900explicitly configured.  Configuration knobs for the same type of
2901control used widely differing naming schemes and formats.  Statistics
2902and information knobs were named arbitrarily and used different
2903formats and units even in the same controller.
2904
2905cgroup v2 establishes common conventions where appropriate and updates
2906controllers so that they expose minimal and consistent interfaces.
2907
2908
2909Controller Issues and Remedies
2910------------------------------
2911
2912Memory
2913~~~~~~
2914
2915The original lower boundary, the soft limit, is defined as a limit
2916that is per default unset.  As a result, the set of cgroups that
2917global reclaim prefers is opt-in, rather than opt-out.  The costs for
2918optimizing these mostly negative lookups are so high that the
2919implementation, despite its enormous size, does not even provide the
2920basic desirable behavior.  First off, the soft limit has no
2921hierarchical meaning.  All configured groups are organized in a global
2922rbtree and treated like equal peers, regardless where they are located
2923in the hierarchy.  This makes subtree delegation impossible.  Second,
2924the soft limit reclaim pass is so aggressive that it not just
2925introduces high allocation latencies into the system, but also impacts
2926system performance due to overreclaim, to the point where the feature
2927becomes self-defeating.
2928
2929The memory.low boundary on the other hand is a top-down allocated
2930reserve.  A cgroup enjoys reclaim protection when it's within its
2931effective low, which makes delegation of subtrees possible. It also
2932enjoys having reclaim pressure proportional to its overage when
2933above its effective low.
2934
2935The original high boundary, the hard limit, is defined as a strict
2936limit that can not budge, even if the OOM killer has to be called.
2937But this generally goes against the goal of making the most out of the
2938available memory.  The memory consumption of workloads varies during
2939runtime, and that requires users to overcommit.  But doing that with a
2940strict upper limit requires either a fairly accurate prediction of the
2941working set size or adding slack to the limit.  Since working set size
2942estimation is hard and error prone, and getting it wrong results in
2943OOM kills, most users tend to err on the side of a looser limit and
2944end up wasting precious resources.
2945
2946The memory.high boundary on the other hand can be set much more
2947conservatively.  When hit, it throttles allocations by forcing them
2948into direct reclaim to work off the excess, but it never invokes the
2949OOM killer.  As a result, a high boundary that is chosen too
2950aggressively will not terminate the processes, but instead it will
2951lead to gradual performance degradation.  The user can monitor this
2952and make corrections until the minimal memory footprint that still
2953gives acceptable performance is found.
2954
2955In extreme cases, with many concurrent allocations and a complete
2956breakdown of reclaim progress within the group, the high boundary can
2957be exceeded.  But even then it's mostly better to satisfy the
2958allocation from the slack available in other groups or the rest of the
2959system than killing the group.  Otherwise, memory.max is there to
2960limit this type of spillover and ultimately contain buggy or even
2961malicious applications.
2962
2963Setting the original memory.limit_in_bytes below the current usage was
2964subject to a race condition, where concurrent charges could cause the
2965limit setting to fail. memory.max on the other hand will first set the
2966limit to prevent new charges, and then reclaim and OOM kill until the
2967new limit is met - or the task writing to memory.max is killed.
2968
2969The combined memory+swap accounting and limiting is replaced by real
2970control over swap space.
2971
2972The main argument for a combined memory+swap facility in the original
2973cgroup design was that global or parental pressure would always be
2974able to swap all anonymous memory of a child group, regardless of the
2975child's own (possibly untrusted) configuration.  However, untrusted
2976groups can sabotage swapping by other means - such as referencing its
2977anonymous memory in a tight loop - and an admin can not assume full
2978swappability when overcommitting untrusted jobs.
2979
2980For trusted jobs, on the other hand, a combined counter is not an
2981intuitive userspace interface, and it flies in the face of the idea
2982that cgroup controllers should account and limit specific physical
2983resources.  Swap space is a resource like all others in the system,
2984and that's why unified hierarchy allows distributing it separately.
2985