About Kernel Documentation Linux Kernel Contact Linux Resources Linux Blog

Documentation / cgroup-v2.txt




Custom Search

Based on kernel version 4.13.3. Page generated on 2017-09-23 13:54 EST.

1	================
2	Control Group v2
3	================
4	
5	:Date: October, 2015
6	:Author: Tejun Heo <tj@kernel.org>
7	
8	This is the authoritative documentation on the design, interface and
9	conventions of cgroup v2.  It describes all userland-visible aspects
10	of cgroup including core and specific controller behaviors.  All
11	future changes must be reflected in this document.  Documentation for
12	v1 is available under Documentation/cgroup-v1/.
13	
14	.. CONTENTS
15	
16	   1. Introduction
17	     1-1. Terminology
18	     1-2. What is cgroup?
19	   2. Basic Operations
20	     2-1. Mounting
21	     2-2. Organizing Processes
22	     2-3. [Un]populated Notification
23	     2-4. Controlling Controllers
24	       2-4-1. Enabling and Disabling
25	       2-4-2. Top-down Constraint
26	       2-4-3. No Internal Process Constraint
27	     2-5. Delegation
28	       2-5-1. Model of Delegation
29	       2-5-2. Delegation Containment
30	     2-6. Guidelines
31	       2-6-1. Organize Once and Control
32	       2-6-2. Avoid Name Collisions
33	   3. Resource Distribution Models
34	     3-1. Weights
35	     3-2. Limits
36	     3-3. Protections
37	     3-4. Allocations
38	   4. Interface Files
39	     4-1. Format
40	     4-2. Conventions
41	     4-3. Core Interface Files
42	   5. Controllers
43	     5-1. CPU
44	       5-1-1. CPU Interface Files
45	     5-2. Memory
46	       5-2-1. Memory Interface Files
47	       5-2-2. Usage Guidelines
48	       5-2-3. Memory Ownership
49	     5-3. IO
50	       5-3-1. IO Interface Files
51	       5-3-2. Writeback
52	     5-4. PID
53	       5-4-1. PID Interface Files
54	     5-5. RDMA
55	       5-5-1. RDMA Interface Files
56	     5-6. Misc
57	       5-6-1. perf_event
58	   6. Namespace
59	     6-1. Basics
60	     6-2. The Root and Views
61	     6-3. Migration and setns(2)
62	     6-4. Interaction with Other Namespaces
63	   P. Information on Kernel Programming
64	     P-1. Filesystem Support for Writeback
65	   D. Deprecated v1 Core Features
66	   R. Issues with v1 and Rationales for v2
67	     R-1. Multiple Hierarchies
68	     R-2. Thread Granularity
69	     R-3. Competition Between Inner Nodes and Threads
70	     R-4. Other Interface Issues
71	     R-5. Controller Issues and Remedies
72	       R-5-1. Memory
73	
74	
75	Introduction
76	============
77	
78	Terminology
79	-----------
80	
81	"cgroup" stands for "control group" and is never capitalized.  The
82	singular form is used to designate the whole feature and also as a
83	qualifier as in "cgroup controllers".  When explicitly referring to
84	multiple individual control groups, the plural form "cgroups" is used.
85	
86	
87	What is cgroup?
88	---------------
89	
90	cgroup is a mechanism to organize processes hierarchically and
91	distribute system resources along the hierarchy in a controlled and
92	configurable manner.
93	
94	cgroup is largely composed of two parts - the core and controllers.
95	cgroup core is primarily responsible for hierarchically organizing
96	processes.  A cgroup controller is usually responsible for
97	distributing a specific type of system resource along the hierarchy
98	although there are utility controllers which serve purposes other than
99	resource distribution.
100	
101	cgroups form a tree structure and every process in the system belongs
102	to one and only one cgroup.  All threads of a process belong to the
103	same cgroup.  On creation, all processes are put in the cgroup that
104	the parent process belongs to at the time.  A process can be migrated
105	to another cgroup.  Migration of a process doesn't affect already
106	existing descendant processes.
107	
108	Following certain structural constraints, controllers may be enabled or
109	disabled selectively on a cgroup.  All controller behaviors are
110	hierarchical - if a controller is enabled on a cgroup, it affects all
111	processes which belong to the cgroups consisting the inclusive
112	sub-hierarchy of the cgroup.  When a controller is enabled on a nested
113	cgroup, it always restricts the resource distribution further.  The
114	restrictions set closer to the root in the hierarchy can not be
115	overridden from further away.
116	
117	
118	Basic Operations
119	================
120	
121	Mounting
122	--------
123	
124	Unlike v1, cgroup v2 has only single hierarchy.  The cgroup v2
125	hierarchy can be mounted with the following mount command::
126	
127	  # mount -t cgroup2 none $MOUNT_POINT
128	
129	cgroup2 filesystem has the magic number 0x63677270 ("cgrp").  All
130	controllers which support v2 and are not bound to a v1 hierarchy are
131	automatically bound to the v2 hierarchy and show up at the root.
132	Controllers which are not in active use in the v2 hierarchy can be
133	bound to other hierarchies.  This allows mixing v2 hierarchy with the
134	legacy v1 multiple hierarchies in a fully backward compatible way.
135	
136	A controller can be moved across hierarchies only after the controller
137	is no longer referenced in its current hierarchy.  Because per-cgroup
138	controller states are destroyed asynchronously and controllers may
139	have lingering references, a controller may not show up immediately on
140	the v2 hierarchy after the final umount of the previous hierarchy.
141	Similarly, a controller should be fully disabled to be moved out of
142	the unified hierarchy and it may take some time for the disabled
143	controller to become available for other hierarchies; furthermore, due
144	to inter-controller dependencies, other controllers may need to be
145	disabled too.
146	
147	While useful for development and manual configurations, moving
148	controllers dynamically between the v2 and other hierarchies is
149	strongly discouraged for production use.  It is recommended to decide
150	the hierarchies and controller associations before starting using the
151	controllers after system boot.
152	
153	During transition to v2, system management software might still
154	automount the v1 cgroup filesystem and so hijack all controllers
155	during boot, before manual intervention is possible. To make testing
156	and experimenting easier, the kernel parameter cgroup_no_v1= allows
157	disabling controllers in v1 and make them always available in v2.
158	
159	cgroup v2 currently supports the following mount options.
160	
161	  nsdelegate
162	
163		Consider cgroup namespaces as delegation boundaries.  This
164		option is system wide and can only be set on mount or modified
165		through remount from the init namespace.  The mount option is
166		ignored on non-init namespace mounts.  Please refer to the
167		Delegation section for details.
168	
169	
170	Organizing Processes
171	--------------------
172	
173	Initially, only the root cgroup exists to which all processes belong.
174	A child cgroup can be created by creating a sub-directory::
175	
176	  # mkdir $CGROUP_NAME
177	
178	A given cgroup may have multiple child cgroups forming a tree
179	structure.  Each cgroup has a read-writable interface file
180	"cgroup.procs".  When read, it lists the PIDs of all processes which
181	belong to the cgroup one-per-line.  The PIDs are not ordered and the
182	same PID may show up more than once if the process got moved to
183	another cgroup and then back or the PID got recycled while reading.
184	
185	A process can be migrated into a cgroup by writing its PID to the
186	target cgroup's "cgroup.procs" file.  Only one process can be migrated
187	on a single write(2) call.  If a process is composed of multiple
188	threads, writing the PID of any thread migrates all threads of the
189	process.
190	
191	When a process forks a child process, the new process is born into the
192	cgroup that the forking process belongs to at the time of the
193	operation.  After exit, a process stays associated with the cgroup
194	that it belonged to at the time of exit until it's reaped; however, a
195	zombie process does not appear in "cgroup.procs" and thus can't be
196	moved to another cgroup.
197	
198	A cgroup which doesn't have any children or live processes can be
199	destroyed by removing the directory.  Note that a cgroup which doesn't
200	have any children and is associated only with zombie processes is
201	considered empty and can be removed::
202	
203	  # rmdir $CGROUP_NAME
204	
205	"/proc/$PID/cgroup" lists a process's cgroup membership.  If legacy
206	cgroup is in use in the system, this file may contain multiple lines,
207	one for each hierarchy.  The entry for cgroup v2 is always in the
208	format "0::$PATH"::
209	
210	  # cat /proc/842/cgroup
211	  ...
212	  0::/test-cgroup/test-cgroup-nested
213	
214	If the process becomes a zombie and the cgroup it was associated with
215	is removed subsequently, " (deleted)" is appended to the path::
216	
217	  # cat /proc/842/cgroup
218	  ...
219	  0::/test-cgroup/test-cgroup-nested (deleted)
220	
221	
222	[Un]populated Notification
223	--------------------------
224	
225	Each non-root cgroup has a "cgroup.events" file which contains
226	"populated" field indicating whether the cgroup's sub-hierarchy has
227	live processes in it.  Its value is 0 if there is no live process in
228	the cgroup and its descendants; otherwise, 1.  poll and [id]notify
229	events are triggered when the value changes.  This can be used, for
230	example, to start a clean-up operation after all processes of a given
231	sub-hierarchy have exited.  The populated state updates and
232	notifications are recursive.  Consider the following sub-hierarchy
233	where the numbers in the parentheses represent the numbers of processes
234	in each cgroup::
235	
236	  A(4) - B(0) - C(1)
237	              \ D(0)
238	
239	A, B and C's "populated" fields would be 1 while D's 0.  After the one
240	process in C exits, B and C's "populated" fields would flip to "0" and
241	file modified events will be generated on the "cgroup.events" files of
242	both cgroups.
243	
244	
245	Controlling Controllers
246	-----------------------
247	
248	Enabling and Disabling
249	~~~~~~~~~~~~~~~~~~~~~~
250	
251	Each cgroup has a "cgroup.controllers" file which lists all
252	controllers available for the cgroup to enable::
253	
254	  # cat cgroup.controllers
255	  cpu io memory
256	
257	No controller is enabled by default.  Controllers can be enabled and
258	disabled by writing to the "cgroup.subtree_control" file::
259	
260	  # echo "+cpu +memory -io" > cgroup.subtree_control
261	
262	Only controllers which are listed in "cgroup.controllers" can be
263	enabled.  When multiple operations are specified as above, either they
264	all succeed or fail.  If multiple operations on the same controller
265	are specified, the last one is effective.
266	
267	Enabling a controller in a cgroup indicates that the distribution of
268	the target resource across its immediate children will be controlled.
269	Consider the following sub-hierarchy.  The enabled controllers are
270	listed in parentheses::
271	
272	  A(cpu,memory) - B(memory) - C()
273	                            \ D()
274	
275	As A has "cpu" and "memory" enabled, A will control the distribution
276	of CPU cycles and memory to its children, in this case, B.  As B has
277	"memory" enabled but not "CPU", C and D will compete freely on CPU
278	cycles but their division of memory available to B will be controlled.
279	
280	As a controller regulates the distribution of the target resource to
281	the cgroup's children, enabling it creates the controller's interface
282	files in the child cgroups.  In the above example, enabling "cpu" on B
283	would create the "cpu." prefixed controller interface files in C and
284	D.  Likewise, disabling "memory" from B would remove the "memory."
285	prefixed controller interface files from C and D.  This means that the
286	controller interface files - anything which doesn't start with
287	"cgroup." are owned by the parent rather than the cgroup itself.
288	
289	
290	Top-down Constraint
291	~~~~~~~~~~~~~~~~~~~
292	
293	Resources are distributed top-down and a cgroup can further distribute
294	a resource only if the resource has been distributed to it from the
295	parent.  This means that all non-root "cgroup.subtree_control" files
296	can only contain controllers which are enabled in the parent's
297	"cgroup.subtree_control" file.  A controller can be enabled only if
298	the parent has the controller enabled and a controller can't be
299	disabled if one or more children have it enabled.
300	
301	
302	No Internal Process Constraint
303	~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
304	
305	Non-root cgroups can only distribute resources to their children when
306	they don't have any processes of their own.  In other words, only
307	cgroups which don't contain any processes can have controllers enabled
308	in their "cgroup.subtree_control" files.
309	
310	This guarantees that, when a controller is looking at the part of the
311	hierarchy which has it enabled, processes are always only on the
312	leaves.  This rules out situations where child cgroups compete against
313	internal processes of the parent.
314	
315	The root cgroup is exempt from this restriction.  Root contains
316	processes and anonymous resource consumption which can't be associated
317	with any other cgroups and requires special treatment from most
318	controllers.  How resource consumption in the root cgroup is governed
319	is up to each controller.
320	
321	Note that the restriction doesn't get in the way if there is no
322	enabled controller in the cgroup's "cgroup.subtree_control".  This is
323	important as otherwise it wouldn't be possible to create children of a
324	populated cgroup.  To control resource distribution of a cgroup, the
325	cgroup must create children and transfer all its processes to the
326	children before enabling controllers in its "cgroup.subtree_control"
327	file.
328	
329	
330	Delegation
331	----------
332	
333	Model of Delegation
334	~~~~~~~~~~~~~~~~~~~
335	
336	A cgroup can be delegated in two ways.  First, to a less privileged
337	user by granting write access of the directory and its "cgroup.procs"
338	and "cgroup.subtree_control" files to the user.  Second, if the
339	"nsdelegate" mount option is set, automatically to a cgroup namespace
340	on namespace creation.
341	
342	Because the resource control interface files in a given directory
343	control the distribution of the parent's resources, the delegatee
344	shouldn't be allowed to write to them.  For the first method, this is
345	achieved by not granting access to these files.  For the second, the
346	kernel rejects writes to all files other than "cgroup.procs" and
347	"cgroup.subtree_control" on a namespace root from inside the
348	namespace.
349	
350	The end results are equivalent for both delegation types.  Once
351	delegated, the user can build sub-hierarchy under the directory,
352	organize processes inside it as it sees fit and further distribute the
353	resources it received from the parent.  The limits and other settings
354	of all resource controllers are hierarchical and regardless of what
355	happens in the delegated sub-hierarchy, nothing can escape the
356	resource restrictions imposed by the parent.
357	
358	Currently, cgroup doesn't impose any restrictions on the number of
359	cgroups in or nesting depth of a delegated sub-hierarchy; however,
360	this may be limited explicitly in the future.
361	
362	
363	Delegation Containment
364	~~~~~~~~~~~~~~~~~~~~~~
365	
366	A delegated sub-hierarchy is contained in the sense that processes
367	can't be moved into or out of the sub-hierarchy by the delegatee.
368	
369	For delegations to a less privileged user, this is achieved by
370	requiring the following conditions for a process with a non-root euid
371	to migrate a target process into a cgroup by writing its PID to the
372	"cgroup.procs" file.
373	
374	- The writer must have write access to the "cgroup.procs" file.
375	
376	- The writer must have write access to the "cgroup.procs" file of the
377	  common ancestor of the source and destination cgroups.
378	
379	The above two constraints ensure that while a delegatee may migrate
380	processes around freely in the delegated sub-hierarchy it can't pull
381	in from or push out to outside the sub-hierarchy.
382	
383	For an example, let's assume cgroups C0 and C1 have been delegated to
384	user U0 who created C00, C01 under C0 and C10 under C1 as follows and
385	all processes under C0 and C1 belong to U0::
386	
387	  ~~~~~~~~~~~~~ - C0 - C00
388	  ~ cgroup    ~      \ C01
389	  ~ hierarchy ~
390	  ~~~~~~~~~~~~~ - C1 - C10
391	
392	Let's also say U0 wants to write the PID of a process which is
393	currently in C10 into "C00/cgroup.procs".  U0 has write access to the
394	file; however, the common ancestor of the source cgroup C10 and the
395	destination cgroup C00 is above the points of delegation and U0 would
396	not have write access to its "cgroup.procs" files and thus the write
397	will be denied with -EACCES.
398	
399	For delegations to namespaces, containment is achieved by requiring
400	that both the source and destination cgroups are reachable from the
401	namespace of the process which is attempting the migration.  If either
402	is not reachable, the migration is rejected with -ENOENT.
403	
404	
405	Guidelines
406	----------
407	
408	Organize Once and Control
409	~~~~~~~~~~~~~~~~~~~~~~~~~
410	
411	Migrating a process across cgroups is a relatively expensive operation
412	and stateful resources such as memory are not moved together with the
413	process.  This is an explicit design decision as there often exist
414	inherent trade-offs between migration and various hot paths in terms
415	of synchronization cost.
416	
417	As such, migrating processes across cgroups frequently as a means to
418	apply different resource restrictions is discouraged.  A workload
419	should be assigned to a cgroup according to the system's logical and
420	resource structure once on start-up.  Dynamic adjustments to resource
421	distribution can be made by changing controller configuration through
422	the interface files.
423	
424	
425	Avoid Name Collisions
426	~~~~~~~~~~~~~~~~~~~~~
427	
428	Interface files for a cgroup and its children cgroups occupy the same
429	directory and it is possible to create children cgroups which collide
430	with interface files.
431	
432	All cgroup core interface files are prefixed with "cgroup." and each
433	controller's interface files are prefixed with the controller name and
434	a dot.  A controller's name is composed of lower case alphabets and
435	'_'s but never begins with an '_' so it can be used as the prefix
436	character for collision avoidance.  Also, interface file names won't
437	start or end with terms which are often used in categorizing workloads
438	such as job, service, slice, unit or workload.
439	
440	cgroup doesn't do anything to prevent name collisions and it's the
441	user's responsibility to avoid them.
442	
443	
444	Resource Distribution Models
445	============================
446	
447	cgroup controllers implement several resource distribution schemes
448	depending on the resource type and expected use cases.  This section
449	describes major schemes in use along with their expected behaviors.
450	
451	
452	Weights
453	-------
454	
455	A parent's resource is distributed by adding up the weights of all
456	active children and giving each the fraction matching the ratio of its
457	weight against the sum.  As only children which can make use of the
458	resource at the moment participate in the distribution, this is
459	work-conserving.  Due to the dynamic nature, this model is usually
460	used for stateless resources.
461	
462	All weights are in the range [1, 10000] with the default at 100.  This
463	allows symmetric multiplicative biases in both directions at fine
464	enough granularity while staying in the intuitive range.
465	
466	As long as the weight is in range, all configuration combinations are
467	valid and there is no reason to reject configuration changes or
468	process migrations.
469	
470	"cpu.weight" proportionally distributes CPU cycles to active children
471	and is an example of this type.
472	
473	
474	Limits
475	------
476	
477	A child can only consume upto the configured amount of the resource.
478	Limits can be over-committed - the sum of the limits of children can
479	exceed the amount of resource available to the parent.
480	
481	Limits are in the range [0, max] and defaults to "max", which is noop.
482	
483	As limits can be over-committed, all configuration combinations are
484	valid and there is no reason to reject configuration changes or
485	process migrations.
486	
487	"io.max" limits the maximum BPS and/or IOPS that a cgroup can consume
488	on an IO device and is an example of this type.
489	
490	
491	Protections
492	-----------
493	
494	A cgroup is protected to be allocated upto the configured amount of
495	the resource if the usages of all its ancestors are under their
496	protected levels.  Protections can be hard guarantees or best effort
497	soft boundaries.  Protections can also be over-committed in which case
498	only upto the amount available to the parent is protected among
499	children.
500	
501	Protections are in the range [0, max] and defaults to 0, which is
502	noop.
503	
504	As protections can be over-committed, all configuration combinations
505	are valid and there is no reason to reject configuration changes or
506	process migrations.
507	
508	"memory.low" implements best-effort memory protection and is an
509	example of this type.
510	
511	
512	Allocations
513	-----------
514	
515	A cgroup is exclusively allocated a certain amount of a finite
516	resource.  Allocations can't be over-committed - the sum of the
517	allocations of children can not exceed the amount of resource
518	available to the parent.
519	
520	Allocations are in the range [0, max] and defaults to 0, which is no
521	resource.
522	
523	As allocations can't be over-committed, some configuration
524	combinations are invalid and should be rejected.  Also, if the
525	resource is mandatory for execution of processes, process migrations
526	may be rejected.
527	
528	"cpu.rt.max" hard-allocates realtime slices and is an example of this
529	type.
530	
531	
532	Interface Files
533	===============
534	
535	Format
536	------
537	
538	All interface files should be in one of the following formats whenever
539	possible::
540	
541	  New-line separated values
542	  (when only one value can be written at once)
543	
544		VAL0\n
545		VAL1\n
546		...
547	
548	  Space separated values
549	  (when read-only or multiple values can be written at once)
550	
551		VAL0 VAL1 ...\n
552	
553	  Flat keyed
554	
555		KEY0 VAL0\n
556		KEY1 VAL1\n
557		...
558	
559	  Nested keyed
560	
561		KEY0 SUB_KEY0=VAL00 SUB_KEY1=VAL01...
562		KEY1 SUB_KEY0=VAL10 SUB_KEY1=VAL11...
563		...
564	
565	For a writable file, the format for writing should generally match
566	reading; however, controllers may allow omitting later fields or
567	implement restricted shortcuts for most common use cases.
568	
569	For both flat and nested keyed files, only the values for a single key
570	can be written at a time.  For nested keyed files, the sub key pairs
571	may be specified in any order and not all pairs have to be specified.
572	
573	
574	Conventions
575	-----------
576	
577	- Settings for a single feature should be contained in a single file.
578	
579	- The root cgroup should be exempt from resource control and thus
580	  shouldn't have resource control interface files.  Also,
581	  informational files on the root cgroup which end up showing global
582	  information available elsewhere shouldn't exist.
583	
584	- If a controller implements weight based resource distribution, its
585	  interface file should be named "weight" and have the range [1,
586	  10000] with 100 as the default.  The values are chosen to allow
587	  enough and symmetric bias in both directions while keeping it
588	  intuitive (the default is 100%).
589	
590	- If a controller implements an absolute resource guarantee and/or
591	  limit, the interface files should be named "min" and "max"
592	  respectively.  If a controller implements best effort resource
593	  guarantee and/or limit, the interface files should be named "low"
594	  and "high" respectively.
595	
596	  In the above four control files, the special token "max" should be
597	  used to represent upward infinity for both reading and writing.
598	
599	- If a setting has a configurable default value and keyed specific
600	  overrides, the default entry should be keyed with "default" and
601	  appear as the first entry in the file.
602	
603	  The default value can be updated by writing either "default $VAL" or
604	  "$VAL".
605	
606	  When writing to update a specific override, "default" can be used as
607	  the value to indicate removal of the override.  Override entries
608	  with "default" as the value must not appear when read.
609	
610	  For example, a setting which is keyed by major:minor device numbers
611	  with integer values may look like the following::
612	
613	    # cat cgroup-example-interface-file
614	    default 150
615	    8:0 300
616	
617	  The default value can be updated by::
618	
619	    # echo 125 > cgroup-example-interface-file
620	
621	  or::
622	
623	    # echo "default 125" > cgroup-example-interface-file
624	
625	  An override can be set by::
626	
627	    # echo "8:16 170" > cgroup-example-interface-file
628	
629	  and cleared by::
630	
631	    # echo "8:0 default" > cgroup-example-interface-file
632	    # cat cgroup-example-interface-file
633	    default 125
634	    8:16 170
635	
636	- For events which are not very high frequency, an interface file
637	  "events" should be created which lists event key value pairs.
638	  Whenever a notifiable event happens, file modified event should be
639	  generated on the file.
640	
641	
642	Core Interface Files
643	--------------------
644	
645	All cgroup core files are prefixed with "cgroup."
646	
647	  cgroup.procs
648		A read-write new-line separated values file which exists on
649		all cgroups.
650	
651		When read, it lists the PIDs of all processes which belong to
652		the cgroup one-per-line.  The PIDs are not ordered and the
653		same PID may show up more than once if the process got moved
654		to another cgroup and then back or the PID got recycled while
655		reading.
656	
657		A PID can be written to migrate the process associated with
658		the PID to the cgroup.  The writer should match all of the
659		following conditions.
660	
661		- Its euid is either root or must match either uid or suid of
662	          the target process.
663	
664		- It must have write access to the "cgroup.procs" file.
665	
666		- It must have write access to the "cgroup.procs" file of the
667		  common ancestor of the source and destination cgroups.
668	
669		When delegating a sub-hierarchy, write access to this file
670		should be granted along with the containing directory.
671	
672	  cgroup.controllers
673		A read-only space separated values file which exists on all
674		cgroups.
675	
676		It shows space separated list of all controllers available to
677		the cgroup.  The controllers are not ordered.
678	
679	  cgroup.subtree_control
680		A read-write space separated values file which exists on all
681		cgroups.  Starts out empty.
682	
683		When read, it shows space separated list of the controllers
684		which are enabled to control resource distribution from the
685		cgroup to its children.
686	
687		Space separated list of controllers prefixed with '+' or '-'
688		can be written to enable or disable controllers.  A controller
689		name prefixed with '+' enables the controller and '-'
690		disables.  If a controller appears more than once on the list,
691		the last one is effective.  When multiple enable and disable
692		operations are specified, either all succeed or all fail.
693	
694	  cgroup.events
695		A read-only flat-keyed file which exists on non-root cgroups.
696		The following entries are defined.  Unless specified
697		otherwise, a value change in this file generates a file
698		modified event.
699	
700		  populated
701			1 if the cgroup or its descendants contains any live
702			processes; otherwise, 0.
703	
704	
705	Controllers
706	===========
707	
708	CPU
709	---
710	
711	.. note::
712	
713		The interface for the cpu controller hasn't been merged yet
714	
715	The "cpu" controllers regulates distribution of CPU cycles.  This
716	controller implements weight and absolute bandwidth limit models for
717	normal scheduling policy and absolute bandwidth allocation model for
718	realtime scheduling policy.
719	
720	
721	CPU Interface Files
722	~~~~~~~~~~~~~~~~~~~
723	
724	All time durations are in microseconds.
725	
726	  cpu.stat
727		A read-only flat-keyed file which exists on non-root cgroups.
728	
729		It reports the following six stats:
730	
731		- usage_usec
732		- user_usec
733		- system_usec
734		- nr_periods
735		- nr_throttled
736		- throttled_usec
737	
738	  cpu.weight
739		A read-write single value file which exists on non-root
740		cgroups.  The default is "100".
741	
742		The weight in the range [1, 10000].
743	
744	  cpu.max
745		A read-write two value file which exists on non-root cgroups.
746		The default is "max 100000".
747	
748		The maximum bandwidth limit.  It's in the following format::
749	
750		  $MAX $PERIOD
751	
752		which indicates that the group may consume upto $MAX in each
753		$PERIOD duration.  "max" for $MAX indicates no limit.  If only
754		one number is written, $MAX is updated.
755	
756	  cpu.rt.max
757		.. note::
758	
759		   The semantics of this file is still under discussion and the
760		   interface hasn't been merged yet
761	
762		A read-write two value file which exists on all cgroups.
763		The default is "0 100000".
764	
765		The maximum realtime runtime allocation.  Over-committing
766		configurations are disallowed and process migrations are
767		rejected if not enough bandwidth is available.  It's in the
768		following format::
769	
770		  $MAX $PERIOD
771	
772		which indicates that the group may consume upto $MAX in each
773		$PERIOD duration.  If only one number is written, $MAX is
774		updated.
775	
776	
777	Memory
778	------
779	
780	The "memory" controller regulates distribution of memory.  Memory is
781	stateful and implements both limit and protection models.  Due to the
782	intertwining between memory usage and reclaim pressure and the
783	stateful nature of memory, the distribution model is relatively
784	complex.
785	
786	While not completely water-tight, all major memory usages by a given
787	cgroup are tracked so that the total memory consumption can be
788	accounted and controlled to a reasonable extent.  Currently, the
789	following types of memory usages are tracked.
790	
791	- Userland memory - page cache and anonymous memory.
792	
793	- Kernel data structures such as dentries and inodes.
794	
795	- TCP socket buffers.
796	
797	The above list may expand in the future for better coverage.
798	
799	
800	Memory Interface Files
801	~~~~~~~~~~~~~~~~~~~~~~
802	
803	All memory amounts are in bytes.  If a value which is not aligned to
804	PAGE_SIZE is written, the value may be rounded up to the closest
805	PAGE_SIZE multiple when read back.
806	
807	  memory.current
808		A read-only single value file which exists on non-root
809		cgroups.
810	
811		The total amount of memory currently being used by the cgroup
812		and its descendants.
813	
814	  memory.low
815		A read-write single value file which exists on non-root
816		cgroups.  The default is "0".
817	
818		Best-effort memory protection.  If the memory usages of a
819		cgroup and all its ancestors are below their low boundaries,
820		the cgroup's memory won't be reclaimed unless memory can be
821		reclaimed from unprotected cgroups.
822	
823		Putting more memory than generally available under this
824		protection is discouraged.
825	
826	  memory.high
827		A read-write single value file which exists on non-root
828		cgroups.  The default is "max".
829	
830		Memory usage throttle limit.  This is the main mechanism to
831		control memory usage of a cgroup.  If a cgroup's usage goes
832		over the high boundary, the processes of the cgroup are
833		throttled and put under heavy reclaim pressure.
834	
835		Going over the high limit never invokes the OOM killer and
836		under extreme conditions the limit may be breached.
837	
838	  memory.max
839		A read-write single value file which exists on non-root
840		cgroups.  The default is "max".
841	
842		Memory usage hard limit.  This is the final protection
843		mechanism.  If a cgroup's memory usage reaches this limit and
844		can't be reduced, the OOM killer is invoked in the cgroup.
845		Under certain circumstances, the usage may go over the limit
846		temporarily.
847	
848		This is the ultimate protection mechanism.  As long as the
849		high limit is used and monitored properly, this limit's
850		utility is limited to providing the final safety net.
851	
852	  memory.events
853		A read-only flat-keyed file which exists on non-root cgroups.
854		The following entries are defined.  Unless specified
855		otherwise, a value change in this file generates a file
856		modified event.
857	
858		  low
859			The number of times the cgroup is reclaimed due to
860			high memory pressure even though its usage is under
861			the low boundary.  This usually indicates that the low
862			boundary is over-committed.
863	
864		  high
865			The number of times processes of the cgroup are
866			throttled and routed to perform direct memory reclaim
867			because the high memory boundary was exceeded.  For a
868			cgroup whose memory usage is capped by the high limit
869			rather than global memory pressure, this event's
870			occurrences are expected.
871	
872		  max
873			The number of times the cgroup's memory usage was
874			about to go over the max boundary.  If direct reclaim
875			fails to bring it down, the cgroup goes to OOM state.
876	
877		  oom
878			The number of time the cgroup's memory usage was
879			reached the limit and allocation was about to fail.
880	
881			Depending on context result could be invocation of OOM
882			killer and retrying allocation or failing alloction.
883	
884			Failed allocation in its turn could be returned into
885			userspace as -ENOMEM or siletly ignored in cases like
886			disk readahead.  For now OOM in memory cgroup kills
887			tasks iff shortage has happened inside page fault.
888	
889		  oom_kill
890			The number of processes belonging to this cgroup
891			killed by any kind of OOM killer.
892	
893	  memory.stat
894		A read-only flat-keyed file which exists on non-root cgroups.
895	
896		This breaks down the cgroup's memory footprint into different
897		types of memory, type-specific details, and other information
898		on the state and past events of the memory management system.
899	
900		All memory amounts are in bytes.
901	
902		The entries are ordered to be human readable, and new entries
903		can show up in the middle. Don't rely on items remaining in a
904		fixed position; use the keys to look up specific values!
905	
906		  anon
907			Amount of memory used in anonymous mappings such as
908			brk(), sbrk(), and mmap(MAP_ANONYMOUS)
909	
910		  file
911			Amount of memory used to cache filesystem data,
912			including tmpfs and shared memory.
913	
914		  kernel_stack
915			Amount of memory allocated to kernel stacks.
916	
917		  slab
918			Amount of memory used for storing in-kernel data
919			structures.
920	
921		  sock
922			Amount of memory used in network transmission buffers
923	
924		  shmem
925			Amount of cached filesystem data that is swap-backed,
926			such as tmpfs, shm segments, shared anonymous mmap()s
927	
928		  file_mapped
929			Amount of cached filesystem data mapped with mmap()
930	
931		  file_dirty
932			Amount of cached filesystem data that was modified but
933			not yet written back to disk
934	
935		  file_writeback
936			Amount of cached filesystem data that was modified and
937			is currently being written back to disk
938	
939		  inactive_anon, active_anon, inactive_file, active_file, unevictable
940			Amount of memory, swap-backed and filesystem-backed,
941			on the internal memory management lists used by the
942			page reclaim algorithm
943	
944		  slab_reclaimable
945			Part of "slab" that might be reclaimed, such as
946			dentries and inodes.
947	
948		  slab_unreclaimable
949			Part of "slab" that cannot be reclaimed on memory
950			pressure.
951	
952		  pgfault
953			Total number of page faults incurred
954	
955		  pgmajfault
956			Number of major page faults incurred
957	
958		  workingset_refault
959	
960			Number of refaults of previously evicted pages
961	
962		  workingset_activate
963	
964			Number of refaulted pages that were immediately activated
965	
966		  workingset_nodereclaim
967	
968			Number of times a shadow node has been reclaimed
969	
970		  pgrefill
971	
972			Amount of scanned pages (in an active LRU list)
973	
974		  pgscan
975	
976			Amount of scanned pages (in an inactive LRU list)
977	
978		  pgsteal
979	
980			Amount of reclaimed pages
981	
982		  pgactivate
983	
984			Amount of pages moved to the active LRU list
985	
986		  pgdeactivate
987	
988			Amount of pages moved to the inactive LRU lis
989	
990		  pglazyfree
991	
992			Amount of pages postponed to be freed under memory pressure
993	
994		  pglazyfreed
995	
996			Amount of reclaimed lazyfree pages
997	
998	  memory.swap.current
999		A read-only single value file which exists on non-root
1000		cgroups.
1001	
1002		The total amount of swap currently being used by the cgroup
1003		and its descendants.
1004	
1005	  memory.swap.max
1006		A read-write single value file which exists on non-root
1007		cgroups.  The default is "max".
1008	
1009		Swap usage hard limit.  If a cgroup's swap usage reaches this
1010		limit, anonymous meomry of the cgroup will not be swapped out.
1011	
1012	
1013	Usage Guidelines
1014	~~~~~~~~~~~~~~~~
1015	
1016	"memory.high" is the main mechanism to control memory usage.
1017	Over-committing on high limit (sum of high limits > available memory)
1018	and letting global memory pressure to distribute memory according to
1019	usage is a viable strategy.
1020	
1021	Because breach of the high limit doesn't trigger the OOM killer but
1022	throttles the offending cgroup, a management agent has ample
1023	opportunities to monitor and take appropriate actions such as granting
1024	more memory or terminating the workload.
1025	
1026	Determining whether a cgroup has enough memory is not trivial as
1027	memory usage doesn't indicate whether the workload can benefit from
1028	more memory.  For example, a workload which writes data received from
1029	network to a file can use all available memory but can also operate as
1030	performant with a small amount of memory.  A measure of memory
1031	pressure - how much the workload is being impacted due to lack of
1032	memory - is necessary to determine whether a workload needs more
1033	memory; unfortunately, memory pressure monitoring mechanism isn't
1034	implemented yet.
1035	
1036	
1037	Memory Ownership
1038	~~~~~~~~~~~~~~~~
1039	
1040	A memory area is charged to the cgroup which instantiated it and stays
1041	charged to the cgroup until the area is released.  Migrating a process
1042	to a different cgroup doesn't move the memory usages that it
1043	instantiated while in the previous cgroup to the new cgroup.
1044	
1045	A memory area may be used by processes belonging to different cgroups.
1046	To which cgroup the area will be charged is in-deterministic; however,
1047	over time, the memory area is likely to end up in a cgroup which has
1048	enough memory allowance to avoid high reclaim pressure.
1049	
1050	If a cgroup sweeps a considerable amount of memory which is expected
1051	to be accessed repeatedly by other cgroups, it may make sense to use
1052	POSIX_FADV_DONTNEED to relinquish the ownership of memory areas
1053	belonging to the affected files to ensure correct memory ownership.
1054	
1055	
1056	IO
1057	--
1058	
1059	The "io" controller regulates the distribution of IO resources.  This
1060	controller implements both weight based and absolute bandwidth or IOPS
1061	limit distribution; however, weight based distribution is available
1062	only if cfq-iosched is in use and neither scheme is available for
1063	blk-mq devices.
1064	
1065	
1066	IO Interface Files
1067	~~~~~~~~~~~~~~~~~~
1068	
1069	  io.stat
1070		A read-only nested-keyed file which exists on non-root
1071		cgroups.
1072	
1073		Lines are keyed by $MAJ:$MIN device numbers and not ordered.
1074		The following nested keys are defined.
1075	
1076		  ======	===================
1077		  rbytes	Bytes read
1078		  wbytes	Bytes written
1079		  rios		Number of read IOs
1080		  wios		Number of write IOs
1081		  ======	===================
1082	
1083		An example read output follows:
1084	
1085		  8:16 rbytes=1459200 wbytes=314773504 rios=192 wios=353
1086		  8:0 rbytes=90430464 wbytes=299008000 rios=8950 wios=1252
1087	
1088	  io.weight
1089		A read-write flat-keyed file which exists on non-root cgroups.
1090		The default is "default 100".
1091	
1092		The first line is the default weight applied to devices
1093		without specific override.  The rest are overrides keyed by
1094		$MAJ:$MIN device numbers and not ordered.  The weights are in
1095		the range [1, 10000] and specifies the relative amount IO time
1096		the cgroup can use in relation to its siblings.
1097	
1098		The default weight can be updated by writing either "default
1099		$WEIGHT" or simply "$WEIGHT".  Overrides can be set by writing
1100		"$MAJ:$MIN $WEIGHT" and unset by writing "$MAJ:$MIN default".
1101	
1102		An example read output follows::
1103	
1104		  default 100
1105		  8:16 200
1106		  8:0 50
1107	
1108	  io.max
1109		A read-write nested-keyed file which exists on non-root
1110		cgroups.
1111	
1112		BPS and IOPS based IO limit.  Lines are keyed by $MAJ:$MIN
1113		device numbers and not ordered.  The following nested keys are
1114		defined.
1115	
1116		  =====		==================================
1117		  rbps		Max read bytes per second
1118		  wbps		Max write bytes per second
1119		  riops		Max read IO operations per second
1120		  wiops		Max write IO operations per second
1121		  =====		==================================
1122	
1123		When writing, any number of nested key-value pairs can be
1124		specified in any order.  "max" can be specified as the value
1125		to remove a specific limit.  If the same key is specified
1126		multiple times, the outcome is undefined.
1127	
1128		BPS and IOPS are measured in each IO direction and IOs are
1129		delayed if limit is reached.  Temporary bursts are allowed.
1130	
1131		Setting read limit at 2M BPS and write at 120 IOPS for 8:16::
1132	
1133		  echo "8:16 rbps=2097152 wiops=120" > io.max
1134	
1135		Reading returns the following::
1136	
1137		  8:16 rbps=2097152 wbps=max riops=max wiops=120
1138	
1139		Write IOPS limit can be removed by writing the following::
1140	
1141		  echo "8:16 wiops=max" > io.max
1142	
1143		Reading now returns the following::
1144	
1145		  8:16 rbps=2097152 wbps=max riops=max wiops=max
1146	
1147	
1148	Writeback
1149	~~~~~~~~~
1150	
1151	Page cache is dirtied through buffered writes and shared mmaps and
1152	written asynchronously to the backing filesystem by the writeback
1153	mechanism.  Writeback sits between the memory and IO domains and
1154	regulates the proportion of dirty memory by balancing dirtying and
1155	write IOs.
1156	
1157	The io controller, in conjunction with the memory controller,
1158	implements control of page cache writeback IOs.  The memory controller
1159	defines the memory domain that dirty memory ratio is calculated and
1160	maintained for and the io controller defines the io domain which
1161	writes out dirty pages for the memory domain.  Both system-wide and
1162	per-cgroup dirty memory states are examined and the more restrictive
1163	of the two is enforced.
1164	
1165	cgroup writeback requires explicit support from the underlying
1166	filesystem.  Currently, cgroup writeback is implemented on ext2, ext4
1167	and btrfs.  On other filesystems, all writeback IOs are attributed to
1168	the root cgroup.
1169	
1170	There are inherent differences in memory and writeback management
1171	which affects how cgroup ownership is tracked.  Memory is tracked per
1172	page while writeback per inode.  For the purpose of writeback, an
1173	inode is assigned to a cgroup and all IO requests to write dirty pages
1174	from the inode are attributed to that cgroup.
1175	
1176	As cgroup ownership for memory is tracked per page, there can be pages
1177	which are associated with different cgroups than the one the inode is
1178	associated with.  These are called foreign pages.  The writeback
1179	constantly keeps track of foreign pages and, if a particular foreign
1180	cgroup becomes the majority over a certain period of time, switches
1181	the ownership of the inode to that cgroup.
1182	
1183	While this model is enough for most use cases where a given inode is
1184	mostly dirtied by a single cgroup even when the main writing cgroup
1185	changes over time, use cases where multiple cgroups write to a single
1186	inode simultaneously are not supported well.  In such circumstances, a
1187	significant portion of IOs are likely to be attributed incorrectly.
1188	As memory controller assigns page ownership on the first use and
1189	doesn't update it until the page is released, even if writeback
1190	strictly follows page ownership, multiple cgroups dirtying overlapping
1191	areas wouldn't work as expected.  It's recommended to avoid such usage
1192	patterns.
1193	
1194	The sysctl knobs which affect writeback behavior are applied to cgroup
1195	writeback as follows.
1196	
1197	  vm.dirty_background_ratio, vm.dirty_ratio
1198		These ratios apply the same to cgroup writeback with the
1199		amount of available memory capped by limits imposed by the
1200		memory controller and system-wide clean memory.
1201	
1202	  vm.dirty_background_bytes, vm.dirty_bytes
1203		For cgroup writeback, this is calculated into ratio against
1204		total available memory and applied the same way as
1205		vm.dirty[_background]_ratio.
1206	
1207	
1208	PID
1209	---
1210	
1211	The process number controller is used to allow a cgroup to stop any
1212	new tasks from being fork()'d or clone()'d after a specified limit is
1213	reached.
1214	
1215	The number of tasks in a cgroup can be exhausted in ways which other
1216	controllers cannot prevent, thus warranting its own controller.  For
1217	example, a fork bomb is likely to exhaust the number of tasks before
1218	hitting memory restrictions.
1219	
1220	Note that PIDs used in this controller refer to TIDs, process IDs as
1221	used by the kernel.
1222	
1223	
1224	PID Interface Files
1225	~~~~~~~~~~~~~~~~~~~
1226	
1227	  pids.max
1228		A read-write single value file which exists on non-root
1229		cgroups.  The default is "max".
1230	
1231		Hard limit of number of processes.
1232	
1233	  pids.current
1234		A read-only single value file which exists on all cgroups.
1235	
1236		The number of processes currently in the cgroup and its
1237		descendants.
1238	
1239	Organisational operations are not blocked by cgroup policies, so it is
1240	possible to have pids.current > pids.max.  This can be done by either
1241	setting the limit to be smaller than pids.current, or attaching enough
1242	processes to the cgroup such that pids.current is larger than
1243	pids.max.  However, it is not possible to violate a cgroup PID policy
1244	through fork() or clone(). These will return -EAGAIN if the creation
1245	of a new process would cause a cgroup policy to be violated.
1246	
1247	
1248	RDMA
1249	----
1250	
1251	The "rdma" controller regulates the distribution and accounting of
1252	of RDMA resources.
1253	
1254	RDMA Interface Files
1255	~~~~~~~~~~~~~~~~~~~~
1256	
1257	  rdma.max
1258		A readwrite nested-keyed file that exists for all the cgroups
1259		except root that describes current configured resource limit
1260		for a RDMA/IB device.
1261	
1262		Lines are keyed by device name and are not ordered.
1263		Each line contains space separated resource name and its configured
1264		limit that can be distributed.
1265	
1266		The following nested keys are defined.
1267	
1268		  ==========	=============================
1269		  hca_handle	Maximum number of HCA Handles
1270		  hca_object 	Maximum number of HCA Objects
1271		  ==========	=============================
1272	
1273		An example for mlx4 and ocrdma device follows::
1274	
1275		  mlx4_0 hca_handle=2 hca_object=2000
1276		  ocrdma1 hca_handle=3 hca_object=max
1277	
1278	  rdma.current
1279		A read-only file that describes current resource usage.
1280		It exists for all the cgroup except root.
1281	
1282		An example for mlx4 and ocrdma device follows::
1283	
1284		  mlx4_0 hca_handle=1 hca_object=20
1285		  ocrdma1 hca_handle=1 hca_object=23
1286	
1287	
1288	Misc
1289	----
1290	
1291	perf_event
1292	~~~~~~~~~~
1293	
1294	perf_event controller, if not mounted on a legacy hierarchy, is
1295	automatically enabled on the v2 hierarchy so that perf events can
1296	always be filtered by cgroup v2 path.  The controller can still be
1297	moved to a legacy hierarchy after v2 hierarchy is populated.
1298	
1299	
1300	Namespace
1301	=========
1302	
1303	Basics
1304	------
1305	
1306	cgroup namespace provides a mechanism to virtualize the view of the
1307	"/proc/$PID/cgroup" file and cgroup mounts.  The CLONE_NEWCGROUP clone
1308	flag can be used with clone(2) and unshare(2) to create a new cgroup
1309	namespace.  The process running inside the cgroup namespace will have
1310	its "/proc/$PID/cgroup" output restricted to cgroupns root.  The
1311	cgroupns root is the cgroup of the process at the time of creation of
1312	the cgroup namespace.
1313	
1314	Without cgroup namespace, the "/proc/$PID/cgroup" file shows the
1315	complete path of the cgroup of a process.  In a container setup where
1316	a set of cgroups and namespaces are intended to isolate processes the
1317	"/proc/$PID/cgroup" file may leak potential system level information
1318	to the isolated processes.  For Example::
1319	
1320	  # cat /proc/self/cgroup
1321	  0::/batchjobs/container_id1
1322	
1323	The path '/batchjobs/container_id1' can be considered as system-data
1324	and undesirable to expose to the isolated processes.  cgroup namespace
1325	can be used to restrict visibility of this path.  For example, before
1326	creating a cgroup namespace, one would see::
1327	
1328	  # ls -l /proc/self/ns/cgroup
1329	  lrwxrwxrwx 1 root root 0 2014-07-15 10:37 /proc/self/ns/cgroup -> cgroup:[4026531835]
1330	  # cat /proc/self/cgroup
1331	  0::/batchjobs/container_id1
1332	
1333	After unsharing a new namespace, the view changes::
1334	
1335	  # ls -l /proc/self/ns/cgroup
1336	  lrwxrwxrwx 1 root root 0 2014-07-15 10:35 /proc/self/ns/cgroup -> cgroup:[4026532183]
1337	  # cat /proc/self/cgroup
1338	  0::/
1339	
1340	When some thread from a multi-threaded process unshares its cgroup
1341	namespace, the new cgroupns gets applied to the entire process (all
1342	the threads).  This is natural for the v2 hierarchy; however, for the
1343	legacy hierarchies, this may be unexpected.
1344	
1345	A cgroup namespace is alive as long as there are processes inside or
1346	mounts pinning it.  When the last usage goes away, the cgroup
1347	namespace is destroyed.  The cgroupns root and the actual cgroups
1348	remain.
1349	
1350	
1351	The Root and Views
1352	------------------
1353	
1354	The 'cgroupns root' for a cgroup namespace is the cgroup in which the
1355	process calling unshare(2) is running.  For example, if a process in
1356	/batchjobs/container_id1 cgroup calls unshare, cgroup
1357	/batchjobs/container_id1 becomes the cgroupns root.  For the
1358	init_cgroup_ns, this is the real root ('/') cgroup.
1359	
1360	The cgroupns root cgroup does not change even if the namespace creator
1361	process later moves to a different cgroup::
1362	
1363	  # ~/unshare -c # unshare cgroupns in some cgroup
1364	  # cat /proc/self/cgroup
1365	  0::/
1366	  # mkdir sub_cgrp_1
1367	  # echo 0 > sub_cgrp_1/cgroup.procs
1368	  # cat /proc/self/cgroup
1369	  0::/sub_cgrp_1
1370	
1371	Each process gets its namespace-specific view of "/proc/$PID/cgroup"
1372	
1373	Processes running inside the cgroup namespace will be able to see
1374	cgroup paths (in /proc/self/cgroup) only inside their root cgroup.
1375	From within an unshared cgroupns::
1376	
1377	  # sleep 100000 &
1378	  [1] 7353
1379	  # echo 7353 > sub_cgrp_1/cgroup.procs
1380	  # cat /proc/7353/cgroup
1381	  0::/sub_cgrp_1
1382	
1383	From the initial cgroup namespace, the real cgroup path will be
1384	visible::
1385	
1386	  $ cat /proc/7353/cgroup
1387	  0::/batchjobs/container_id1/sub_cgrp_1
1388	
1389	From a sibling cgroup namespace (that is, a namespace rooted at a
1390	different cgroup), the cgroup path relative to its own cgroup
1391	namespace root will be shown.  For instance, if PID 7353's cgroup
1392	namespace root is at '/batchjobs/container_id2', then it will see::
1393	
1394	  # cat /proc/7353/cgroup
1395	  0::/../container_id2/sub_cgrp_1
1396	
1397	Note that the relative path always starts with '/' to indicate that
1398	its relative to the cgroup namespace root of the caller.
1399	
1400	
1401	Migration and setns(2)
1402	----------------------
1403	
1404	Processes inside a cgroup namespace can move into and out of the
1405	namespace root if they have proper access to external cgroups.  For
1406	example, from inside a namespace with cgroupns root at
1407	/batchjobs/container_id1, and assuming that the global hierarchy is
1408	still accessible inside cgroupns::
1409	
1410	  # cat /proc/7353/cgroup
1411	  0::/sub_cgrp_1
1412	  # echo 7353 > batchjobs/container_id2/cgroup.procs
1413	  # cat /proc/7353/cgroup
1414	  0::/../container_id2
1415	
1416	Note that this kind of setup is not encouraged.  A task inside cgroup
1417	namespace should only be exposed to its own cgroupns hierarchy.
1418	
1419	setns(2) to another cgroup namespace is allowed when:
1420	
1421	(a) the process has CAP_SYS_ADMIN against its current user namespace
1422	(b) the process has CAP_SYS_ADMIN against the target cgroup
1423	    namespace's userns
1424	
1425	No implicit cgroup changes happen with attaching to another cgroup
1426	namespace.  It is expected that the someone moves the attaching
1427	process under the target cgroup namespace root.
1428	
1429	
1430	Interaction with Other Namespaces
1431	---------------------------------
1432	
1433	Namespace specific cgroup hierarchy can be mounted by a process
1434	running inside a non-init cgroup namespace::
1435	
1436	  # mount -t cgroup2 none $MOUNT_POINT
1437	
1438	This will mount the unified cgroup hierarchy with cgroupns root as the
1439	filesystem root.  The process needs CAP_SYS_ADMIN against its user and
1440	mount namespaces.
1441	
1442	The virtualization of /proc/self/cgroup file combined with restricting
1443	the view of cgroup hierarchy by namespace-private cgroupfs mount
1444	provides a properly isolated cgroup view inside the container.
1445	
1446	
1447	Information on Kernel Programming
1448	=================================
1449	
1450	This section contains kernel programming information in the areas
1451	where interacting with cgroup is necessary.  cgroup core and
1452	controllers are not covered.
1453	
1454	
1455	Filesystem Support for Writeback
1456	--------------------------------
1457	
1458	A filesystem can support cgroup writeback by updating
1459	address_space_operations->writepage[s]() to annotate bio's using the
1460	following two functions.
1461	
1462	  wbc_init_bio(@wbc, @bio)
1463		Should be called for each bio carrying writeback data and
1464		associates the bio with the inode's owner cgroup.  Can be
1465		called anytime between bio allocation and submission.
1466	
1467	  wbc_account_io(@wbc, @page, @bytes)
1468		Should be called for each data segment being written out.
1469		While this function doesn't care exactly when it's called
1470		during the writeback session, it's the easiest and most
1471		natural to call it as data segments are added to a bio.
1472	
1473	With writeback bio's annotated, cgroup support can be enabled per
1474	super_block by setting SB_I_CGROUPWB in ->s_iflags.  This allows for
1475	selective disabling of cgroup writeback support which is helpful when
1476	certain filesystem features, e.g. journaled data mode, are
1477	incompatible.
1478	
1479	wbc_init_bio() binds the specified bio to its cgroup.  Depending on
1480	the configuration, the bio may be executed at a lower priority and if
1481	the writeback session is holding shared resources, e.g. a journal
1482	entry, may lead to priority inversion.  There is no one easy solution
1483	for the problem.  Filesystems can try to work around specific problem
1484	cases by skipping wbc_init_bio() or using bio_associate_blkcg()
1485	directly.
1486	
1487	
1488	Deprecated v1 Core Features
1489	===========================
1490	
1491	- Multiple hierarchies including named ones are not supported.
1492	
1493	- All v1 mount options are not supported.
1494	
1495	- The "tasks" file is removed and "cgroup.procs" is not sorted.
1496	
1497	- "cgroup.clone_children" is removed.
1498	
1499	- /proc/cgroups is meaningless for v2.  Use "cgroup.controllers" file
1500	  at the root instead.
1501	
1502	
1503	Issues with v1 and Rationales for v2
1504	====================================
1505	
1506	Multiple Hierarchies
1507	--------------------
1508	
1509	cgroup v1 allowed an arbitrary number of hierarchies and each
1510	hierarchy could host any number of controllers.  While this seemed to
1511	provide a high level of flexibility, it wasn't useful in practice.
1512	
1513	For example, as there is only one instance of each controller, utility
1514	type controllers such as freezer which can be useful in all
1515	hierarchies could only be used in one.  The issue is exacerbated by
1516	the fact that controllers couldn't be moved to another hierarchy once
1517	hierarchies were populated.  Another issue was that all controllers
1518	bound to a hierarchy were forced to have exactly the same view of the
1519	hierarchy.  It wasn't possible to vary the granularity depending on
1520	the specific controller.
1521	
1522	In practice, these issues heavily limited which controllers could be
1523	put on the same hierarchy and most configurations resorted to putting
1524	each controller on its own hierarchy.  Only closely related ones, such
1525	as the cpu and cpuacct controllers, made sense to be put on the same
1526	hierarchy.  This often meant that userland ended up managing multiple
1527	similar hierarchies repeating the same steps on each hierarchy
1528	whenever a hierarchy management operation was necessary.
1529	
1530	Furthermore, support for multiple hierarchies came at a steep cost.
1531	It greatly complicated cgroup core implementation but more importantly
1532	the support for multiple hierarchies restricted how cgroup could be
1533	used in general and what controllers was able to do.
1534	
1535	There was no limit on how many hierarchies there might be, which meant
1536	that a thread's cgroup membership couldn't be described in finite
1537	length.  The key might contain any number of entries and was unlimited
1538	in length, which made it highly awkward to manipulate and led to
1539	addition of controllers which existed only to identify membership,
1540	which in turn exacerbated the original problem of proliferating number
1541	of hierarchies.
1542	
1543	Also, as a controller couldn't have any expectation regarding the
1544	topologies of hierarchies other controllers might be on, each
1545	controller had to assume that all other controllers were attached to
1546	completely orthogonal hierarchies.  This made it impossible, or at
1547	least very cumbersome, for controllers to cooperate with each other.
1548	
1549	In most use cases, putting controllers on hierarchies which are
1550	completely orthogonal to each other isn't necessary.  What usually is
1551	called for is the ability to have differing levels of granularity
1552	depending on the specific controller.  In other words, hierarchy may
1553	be collapsed from leaf towards root when viewed from specific
1554	controllers.  For example, a given configuration might not care about
1555	how memory is distributed beyond a certain level while still wanting
1556	to control how CPU cycles are distributed.
1557	
1558	
1559	Thread Granularity
1560	------------------
1561	
1562	cgroup v1 allowed threads of a process to belong to different cgroups.
1563	This didn't make sense for some controllers and those controllers
1564	ended up implementing different ways to ignore such situations but
1565	much more importantly it blurred the line between API exposed to
1566	individual applications and system management interface.
1567	
1568	Generally, in-process knowledge is available only to the process
1569	itself; thus, unlike service-level organization of processes,
1570	categorizing threads of a process requires active participation from
1571	the application which owns the target process.
1572	
1573	cgroup v1 had an ambiguously defined delegation model which got abused
1574	in combination with thread granularity.  cgroups were delegated to
1575	individual applications so that they can create and manage their own
1576	sub-hierarchies and control resource distributions along them.  This
1577	effectively raised cgroup to the status of a syscall-like API exposed
1578	to lay programs.
1579	
1580	First of all, cgroup has a fundamentally inadequate interface to be
1581	exposed this way.  For a process to access its own knobs, it has to
1582	extract the path on the target hierarchy from /proc/self/cgroup,
1583	construct the path by appending the name of the knob to the path, open
1584	and then read and/or write to it.  This is not only extremely clunky
1585	and unusual but also inherently racy.  There is no conventional way to
1586	define transaction across the required steps and nothing can guarantee
1587	that the process would actually be operating on its own sub-hierarchy.
1588	
1589	cgroup controllers implemented a number of knobs which would never be
1590	accepted as public APIs because they were just adding control knobs to
1591	system-management pseudo filesystem.  cgroup ended up with interface
1592	knobs which were not properly abstracted or refined and directly
1593	revealed kernel internal details.  These knobs got exposed to
1594	individual applications through the ill-defined delegation mechanism
1595	effectively abusing cgroup as a shortcut to implementing public APIs
1596	without going through the required scrutiny.
1597	
1598	This was painful for both userland and kernel.  Userland ended up with
1599	misbehaving and poorly abstracted interfaces and kernel exposing and
1600	locked into constructs inadvertently.
1601	
1602	
1603	Competition Between Inner Nodes and Threads
1604	-------------------------------------------
1605	
1606	cgroup v1 allowed threads to be in any cgroups which created an
1607	interesting problem where threads belonging to a parent cgroup and its
1608	children cgroups competed for resources.  This was nasty as two
1609	different types of entities competed and there was no obvious way to
1610	settle it.  Different controllers did different things.
1611	
1612	The cpu controller considered threads and cgroups as equivalents and
1613	mapped nice levels to cgroup weights.  This worked for some cases but
1614	fell flat when children wanted to be allocated specific ratios of CPU
1615	cycles and the number of internal threads fluctuated - the ratios
1616	constantly changed as the number of competing entities fluctuated.
1617	There also were other issues.  The mapping from nice level to weight
1618	wasn't obvious or universal, and there were various other knobs which
1619	simply weren't available for threads.
1620	
1621	The io controller implicitly created a hidden leaf node for each
1622	cgroup to host the threads.  The hidden leaf had its own copies of all
1623	the knobs with ``leaf_`` prefixed.  While this allowed equivalent
1624	control over internal threads, it was with serious drawbacks.  It
1625	always added an extra layer of nesting which wouldn't be necessary
1626	otherwise, made the interface messy and significantly complicated the
1627	implementation.
1628	
1629	The memory controller didn't have a way to control what happened
1630	between internal tasks and child cgroups and the behavior was not
1631	clearly defined.  There were attempts to add ad-hoc behaviors and
1632	knobs to tailor the behavior to specific workloads which would have
1633	led to problems extremely difficult to resolve in the long term.
1634	
1635	Multiple controllers struggled with internal tasks and came up with
1636	different ways to deal with it; unfortunately, all the approaches were
1637	severely flawed and, furthermore, the widely different behaviors
1638	made cgroup as a whole highly inconsistent.
1639	
1640	This clearly is a problem which needs to be addressed from cgroup core
1641	in a uniform way.
1642	
1643	
1644	Other Interface Issues
1645	----------------------
1646	
1647	cgroup v1 grew without oversight and developed a large number of
1648	idiosyncrasies and inconsistencies.  One issue on the cgroup core side
1649	was how an empty cgroup was notified - a userland helper binary was
1650	forked and executed for each event.  The event delivery wasn't
1651	recursive or delegatable.  The limitations of the mechanism also led
1652	to in-kernel event delivery filtering mechanism further complicating
1653	the interface.
1654	
1655	Controller interfaces were problematic too.  An extreme example is
1656	controllers completely ignoring hierarchical organization and treating
1657	all cgroups as if they were all located directly under the root
1658	cgroup.  Some controllers exposed a large amount of inconsistent
1659	implementation details to userland.
1660	
1661	There also was no consistency across controllers.  When a new cgroup
1662	was created, some controllers defaulted to not imposing extra
1663	restrictions while others disallowed any resource usage until
1664	explicitly configured.  Configuration knobs for the same type of
1665	control used widely differing naming schemes and formats.  Statistics
1666	and information knobs were named arbitrarily and used different
1667	formats and units even in the same controller.
1668	
1669	cgroup v2 establishes common conventions where appropriate and updates
1670	controllers so that they expose minimal and consistent interfaces.
1671	
1672	
1673	Controller Issues and Remedies
1674	------------------------------
1675	
1676	Memory
1677	~~~~~~
1678	
1679	The original lower boundary, the soft limit, is defined as a limit
1680	that is per default unset.  As a result, the set of cgroups that
1681	global reclaim prefers is opt-in, rather than opt-out.  The costs for
1682	optimizing these mostly negative lookups are so high that the
1683	implementation, despite its enormous size, does not even provide the
1684	basic desirable behavior.  First off, the soft limit has no
1685	hierarchical meaning.  All configured groups are organized in a global
1686	rbtree and treated like equal peers, regardless where they are located
1687	in the hierarchy.  This makes subtree delegation impossible.  Second,
1688	the soft limit reclaim pass is so aggressive that it not just
1689	introduces high allocation latencies into the system, but also impacts
1690	system performance due to overreclaim, to the point where the feature
1691	becomes self-defeating.
1692	
1693	The memory.low boundary on the other hand is a top-down allocated
1694	reserve.  A cgroup enjoys reclaim protection when it and all its
1695	ancestors are below their low boundaries, which makes delegation of
1696	subtrees possible.  Secondly, new cgroups have no reserve per default
1697	and in the common case most cgroups are eligible for the preferred
1698	reclaim pass.  This allows the new low boundary to be efficiently
1699	implemented with just a minor addition to the generic reclaim code,
1700	without the need for out-of-band data structures and reclaim passes.
1701	Because the generic reclaim code considers all cgroups except for the
1702	ones running low in the preferred first reclaim pass, overreclaim of
1703	individual groups is eliminated as well, resulting in much better
1704	overall workload performance.
1705	
1706	The original high boundary, the hard limit, is defined as a strict
1707	limit that can not budge, even if the OOM killer has to be called.
1708	But this generally goes against the goal of making the most out of the
1709	available memory.  The memory consumption of workloads varies during
1710	runtime, and that requires users to overcommit.  But doing that with a
1711	strict upper limit requires either a fairly accurate prediction of the
1712	working set size or adding slack to the limit.  Since working set size
1713	estimation is hard and error prone, and getting it wrong results in
1714	OOM kills, most users tend to err on the side of a looser limit and
1715	end up wasting precious resources.
1716	
1717	The memory.high boundary on the other hand can be set much more
1718	conservatively.  When hit, it throttles allocations by forcing them
1719	into direct reclaim to work off the excess, but it never invokes the
1720	OOM killer.  As a result, a high boundary that is chosen too
1721	aggressively will not terminate the processes, but instead it will
1722	lead to gradual performance degradation.  The user can monitor this
1723	and make corrections until the minimal memory footprint that still
1724	gives acceptable performance is found.
1725	
1726	In extreme cases, with many concurrent allocations and a complete
1727	breakdown of reclaim progress within the group, the high boundary can
1728	be exceeded.  But even then it's mostly better to satisfy the
1729	allocation from the slack available in other groups or the rest of the
1730	system than killing the group.  Otherwise, memory.max is there to
1731	limit this type of spillover and ultimately contain buggy or even
1732	malicious applications.
1733	
1734	Setting the original memory.limit_in_bytes below the current usage was
1735	subject to a race condition, where concurrent charges could cause the
1736	limit setting to fail. memory.max on the other hand will first set the
1737	limit to prevent new charges, and then reclaim and OOM kill until the
1738	new limit is met - or the task writing to memory.max is killed.
1739	
1740	The combined memory+swap accounting and limiting is replaced by real
1741	control over swap space.
1742	
1743	The main argument for a combined memory+swap facility in the original
1744	cgroup design was that global or parental pressure would always be
1745	able to swap all anonymous memory of a child group, regardless of the
1746	child's own (possibly untrusted) configuration.  However, untrusted
1747	groups can sabotage swapping by other means - such as referencing its
1748	anonymous memory in a tight loop - and an admin can not assume full
1749	swappability when overcommitting untrusted jobs.
1750	
1751	For trusted jobs, on the other hand, a combined counter is not an
1752	intuitive userspace interface, and it flies in the face of the idea
1753	that cgroup controllers should account and limit specific physical
1754	resources.  Swap space is a resource like all others in the system,
1755	and that's why unified hierarchy allows distributing it separately.
Hide Line Numbers
About Kernel Documentation Linux Kernel Contact Linux Resources Linux Blog

Information is copyright its respective author. All material is available from the Linux Kernel Source distributed under a GPL License. This page is provided as a free service by mjmwired.net.