blob: ff49cf901148d895b765800ec6ddb79c0e38ed53 [file] [log] [blame]
Tejun Heo6c292092015-11-16 11:13:34 -05001
2Control Group v2
3
4October, 2015 Tejun Heo <tj@kernel.org>
5
6This is the authoritative documentation on the design, interface and
7conventions of cgroup v2. It describes all userland-visible aspects
8of cgroup including core and specific controller behaviors. All
9future changes must be reflected in this document. Documentation for
W. Trevor King9a2ddda2016-01-27 13:01:52 -080010v1 is available under Documentation/cgroup-v1/.
Tejun Heo6c292092015-11-16 11:13:34 -050011
12CONTENTS
13
141. Introduction
15 1-1. Terminology
16 1-2. What is cgroup?
172. Basic Operations
18 2-1. Mounting
19 2-2. Organizing Processes
20 2-3. [Un]populated Notification
21 2-4. Controlling Controllers
22 2-4-1. Enabling and Disabling
23 2-4-2. Top-down Constraint
24 2-4-3. No Internal Process Constraint
25 2-5. Delegation
26 2-5-1. Model of Delegation
27 2-5-2. Delegation Containment
28 2-6. Guidelines
29 2-6-1. Organize Once and Control
30 2-6-2. Avoid Name Collisions
313. Resource Distribution Models
32 3-1. Weights
33 3-2. Limits
34 3-3. Protections
35 3-4. Allocations
364. Interface Files
37 4-1. Format
38 4-2. Conventions
39 4-3. Core Interface Files
405. Controllers
41 5-1. CPU
42 5-1-1. CPU Interface Files
43 5-2. Memory
44 5-2-1. Memory Interface Files
45 5-2-2. Usage Guidelines
46 5-2-3. Memory Ownership
47 5-3. IO
48 5-3-1. IO Interface Files
49 5-3-2. Writeback
50P. Information on Kernel Programming
51 P-1. Filesystem Support for Writeback
52D. Deprecated v1 Core Features
53R. Issues with v1 and Rationales for v2
54 R-1. Multiple Hierarchies
55 R-2. Thread Granularity
56 R-3. Competition Between Inner Nodes and Threads
57 R-4. Other Interface Issues
58 R-5. Controller Issues and Remedies
59 R-5-1. Memory
60
61
621. Introduction
63
641-1. Terminology
65
66"cgroup" stands for "control group" and is never capitalized. The
67singular form is used to designate the whole feature and also as a
68qualifier as in "cgroup controllers". When explicitly referring to
69multiple individual control groups, the plural form "cgroups" is used.
70
71
721-2. What is cgroup?
73
74cgroup is a mechanism to organize processes hierarchically and
75distribute system resources along the hierarchy in a controlled and
76configurable manner.
77
78cgroup is largely composed of two parts - the core and controllers.
79cgroup core is primarily responsible for hierarchically organizing
80processes. A cgroup controller is usually responsible for
81distributing a specific type of system resource along the hierarchy
82although there are utility controllers which serve purposes other than
83resource distribution.
84
85cgroups form a tree structure and every process in the system belongs
86to one and only one cgroup. All threads of a process belong to the
87same cgroup. On creation, all processes are put in the cgroup that
88the parent process belongs to at the time. A process can be migrated
89to another cgroup. Migration of a process doesn't affect already
90existing descendant processes.
91
92Following certain structural constraints, controllers may be enabled or
93disabled selectively on a cgroup. All controller behaviors are
94hierarchical - if a controller is enabled on a cgroup, it affects all
95processes which belong to the cgroups consisting the inclusive
96sub-hierarchy of the cgroup. When a controller is enabled on a nested
97cgroup, it always restricts the resource distribution further. The
98restrictions set closer to the root in the hierarchy can not be
99overridden from further away.
100
101
1022. Basic Operations
103
1042-1. Mounting
105
106Unlike v1, cgroup v2 has only single hierarchy. The cgroup v2
107hierarchy can be mounted with the following mount command.
108
109 # mount -t cgroup2 none $MOUNT_POINT
110
111cgroup2 filesystem has the magic number 0x63677270 ("cgrp"). All
112controllers which support v2 and are not bound to a v1 hierarchy are
113automatically bound to the v2 hierarchy and show up at the root.
114Controllers which are not in active use in the v2 hierarchy can be
115bound to other hierarchies. This allows mixing v2 hierarchy with the
116legacy v1 multiple hierarchies in a fully backward compatible way.
117
118A controller can be moved across hierarchies only after the controller
119is no longer referenced in its current hierarchy. Because per-cgroup
120controller states are destroyed asynchronously and controllers may
121have lingering references, a controller may not show up immediately on
122the v2 hierarchy after the final umount of the previous hierarchy.
123Similarly, a controller should be fully disabled to be moved out of
124the unified hierarchy and it may take some time for the disabled
125controller to become available for other hierarchies; furthermore, due
126to inter-controller dependencies, other controllers may need to be
127disabled too.
128
129While useful for development and manual configurations, moving
130controllers dynamically between the v2 and other hierarchies is
131strongly discouraged for production use. It is recommended to decide
132the hierarchies and controller associations before starting using the
133controllers after system boot.
134
135
1362-2. Organizing Processes
137
138Initially, only the root cgroup exists to which all processes belong.
139A child cgroup can be created by creating a sub-directory.
140
141 # mkdir $CGROUP_NAME
142
143A given cgroup may have multiple child cgroups forming a tree
144structure. Each cgroup has a read-writable interface file
145"cgroup.procs". When read, it lists the PIDs of all processes which
146belong to the cgroup one-per-line. The PIDs are not ordered and the
147same PID may show up more than once if the process got moved to
148another cgroup and then back or the PID got recycled while reading.
149
150A process can be migrated into a cgroup by writing its PID to the
151target cgroup's "cgroup.procs" file. Only one process can be migrated
152on a single write(2) call. If a process is composed of multiple
153threads, writing the PID of any thread migrates all threads of the
154process.
155
156When a process forks a child process, the new process is born into the
157cgroup that the forking process belongs to at the time of the
158operation. After exit, a process stays associated with the cgroup
159that it belonged to at the time of exit until it's reaped; however, a
160zombie process does not appear in "cgroup.procs" and thus can't be
161moved to another cgroup.
162
163A cgroup which doesn't have any children or live processes can be
164destroyed by removing the directory. Note that a cgroup which doesn't
165have any children and is associated only with zombie processes is
166considered empty and can be removed.
167
168 # rmdir $CGROUP_NAME
169
170"/proc/$PID/cgroup" lists a process's cgroup membership. If legacy
171cgroup is in use in the system, this file may contain multiple lines,
172one for each hierarchy. The entry for cgroup v2 is always in the
173format "0::$PATH".
174
175 # cat /proc/842/cgroup
176 ...
177 0::/test-cgroup/test-cgroup-nested
178
179If the process becomes a zombie and the cgroup it was associated with
180is removed subsequently, " (deleted)" is appended to the path.
181
182 # cat /proc/842/cgroup
183 ...
184 0::/test-cgroup/test-cgroup-nested (deleted)
185
186
1872-3. [Un]populated Notification
188
189Each non-root cgroup has a "cgroup.events" file which contains
190"populated" field indicating whether the cgroup's sub-hierarchy has
191live processes in it. Its value is 0 if there is no live process in
192the cgroup and its descendants; otherwise, 1. poll and [id]notify
193events are triggered when the value changes. This can be used, for
194example, to start a clean-up operation after all processes of a given
195sub-hierarchy have exited. The populated state updates and
196notifications are recursive. Consider the following sub-hierarchy
197where the numbers in the parentheses represent the numbers of processes
198in each cgroup.
199
200 A(4) - B(0) - C(1)
201 \ D(0)
202
203A, B and C's "populated" fields would be 1 while D's 0. After the one
204process in C exits, B and C's "populated" fields would flip to "0" and
205file modified events will be generated on the "cgroup.events" files of
206both cgroups.
207
208
2092-4. Controlling Controllers
210
2112-4-1. Enabling and Disabling
212
213Each cgroup has a "cgroup.controllers" file which lists all
214controllers available for the cgroup to enable.
215
216 # cat cgroup.controllers
217 cpu io memory
218
219No controller is enabled by default. Controllers can be enabled and
220disabled by writing to the "cgroup.subtree_control" file.
221
222 # echo "+cpu +memory -io" > cgroup.subtree_control
223
224Only controllers which are listed in "cgroup.controllers" can be
225enabled. When multiple operations are specified as above, either they
226all succeed or fail. If multiple operations on the same controller
227are specified, the last one is effective.
228
229Enabling a controller in a cgroup indicates that the distribution of
230the target resource across its immediate children will be controlled.
231Consider the following sub-hierarchy. The enabled controllers are
232listed in parentheses.
233
234 A(cpu,memory) - B(memory) - C()
235 \ D()
236
237As A has "cpu" and "memory" enabled, A will control the distribution
238of CPU cycles and memory to its children, in this case, B. As B has
239"memory" enabled but not "CPU", C and D will compete freely on CPU
240cycles but their division of memory available to B will be controlled.
241
242As a controller regulates the distribution of the target resource to
243the cgroup's children, enabling it creates the controller's interface
244files in the child cgroups. In the above example, enabling "cpu" on B
245would create the "cpu." prefixed controller interface files in C and
246D. Likewise, disabling "memory" from B would remove the "memory."
247prefixed controller interface files from C and D. This means that the
248controller interface files - anything which doesn't start with
249"cgroup." are owned by the parent rather than the cgroup itself.
250
251
2522-4-2. Top-down Constraint
253
254Resources are distributed top-down and a cgroup can further distribute
255a resource only if the resource has been distributed to it from the
256parent. This means that all non-root "cgroup.subtree_control" files
257can only contain controllers which are enabled in the parent's
258"cgroup.subtree_control" file. A controller can be enabled only if
259the parent has the controller enabled and a controller can't be
260disabled if one or more children have it enabled.
261
262
2632-4-3. No Internal Process Constraint
264
265Non-root cgroups can only distribute resources to their children when
266they don't have any processes of their own. In other words, only
267cgroups which don't contain any processes can have controllers enabled
268in their "cgroup.subtree_control" files.
269
270This guarantees that, when a controller is looking at the part of the
271hierarchy which has it enabled, processes are always only on the
272leaves. This rules out situations where child cgroups compete against
273internal processes of the parent.
274
275The root cgroup is exempt from this restriction. Root contains
276processes and anonymous resource consumption which can't be associated
277with any other cgroups and requires special treatment from most
278controllers. How resource consumption in the root cgroup is governed
279is up to each controller.
280
281Note that the restriction doesn't get in the way if there is no
282enabled controller in the cgroup's "cgroup.subtree_control". This is
283important as otherwise it wouldn't be possible to create children of a
284populated cgroup. To control resource distribution of a cgroup, the
285cgroup must create children and transfer all its processes to the
286children before enabling controllers in its "cgroup.subtree_control"
287file.
288
289
2902-5. Delegation
291
2922-5-1. Model of Delegation
293
294A cgroup can be delegated to a less privileged user by granting write
295access of the directory and its "cgroup.procs" file to the user. Note
296that resource control interface files in a given directory control the
297distribution of the parent's resources and thus must not be delegated
298along with the directory.
299
300Once delegated, the user can build sub-hierarchy under the directory,
301organize processes as it sees fit and further distribute the resources
302it received from the parent. The limits and other settings of all
303resource controllers are hierarchical and regardless of what happens
304in the delegated sub-hierarchy, nothing can escape the resource
305restrictions imposed by the parent.
306
307Currently, cgroup doesn't impose any restrictions on the number of
308cgroups in or nesting depth of a delegated sub-hierarchy; however,
309this may be limited explicitly in the future.
310
311
3122-5-2. Delegation Containment
313
314A delegated sub-hierarchy is contained in the sense that processes
315can't be moved into or out of the sub-hierarchy by the delegatee. For
316a process with a non-root euid to migrate a target process into a
317cgroup by writing its PID to the "cgroup.procs" file, the following
318conditions must be met.
319
320- The writer's euid must match either uid or suid of the target process.
321
322- The writer must have write access to the "cgroup.procs" file.
323
324- The writer must have write access to the "cgroup.procs" file of the
325 common ancestor of the source and destination cgroups.
326
327The above three constraints ensure that while a delegatee may migrate
328processes around freely in the delegated sub-hierarchy it can't pull
329in from or push out to outside the sub-hierarchy.
330
331For an example, let's assume cgroups C0 and C1 have been delegated to
332user U0 who created C00, C01 under C0 and C10 under C1 as follows and
333all processes under C0 and C1 belong to U0.
334
335 ~~~~~~~~~~~~~ - C0 - C00
336 ~ cgroup ~ \ C01
337 ~ hierarchy ~
338 ~~~~~~~~~~~~~ - C1 - C10
339
340Let's also say U0 wants to write the PID of a process which is
341currently in C10 into "C00/cgroup.procs". U0 has write access to the
342file and uid match on the process; however, the common ancestor of the
343source cgroup C10 and the destination cgroup C00 is above the points
344of delegation and U0 would not have write access to its "cgroup.procs"
345files and thus the write will be denied with -EACCES.
346
347
3482-6. Guidelines
349
3502-6-1. Organize Once and Control
351
352Migrating a process across cgroups is a relatively expensive operation
353and stateful resources such as memory are not moved together with the
354process. This is an explicit design decision as there often exist
355inherent trade-offs between migration and various hot paths in terms
356of synchronization cost.
357
358As such, migrating processes across cgroups frequently as a means to
359apply different resource restrictions is discouraged. A workload
360should be assigned to a cgroup according to the system's logical and
361resource structure once on start-up. Dynamic adjustments to resource
362distribution can be made by changing controller configuration through
363the interface files.
364
365
3662-6-2. Avoid Name Collisions
367
368Interface files for a cgroup and its children cgroups occupy the same
369directory and it is possible to create children cgroups which collide
370with interface files.
371
372All cgroup core interface files are prefixed with "cgroup." and each
373controller's interface files are prefixed with the controller name and
374a dot. A controller's name is composed of lower case alphabets and
375'_'s but never begins with an '_' so it can be used as the prefix
376character for collision avoidance. Also, interface file names won't
377start or end with terms which are often used in categorizing workloads
378such as job, service, slice, unit or workload.
379
380cgroup doesn't do anything to prevent name collisions and it's the
381user's responsibility to avoid them.
382
383
3843. Resource Distribution Models
385
386cgroup controllers implement several resource distribution schemes
387depending on the resource type and expected use cases. This section
388describes major schemes in use along with their expected behaviors.
389
390
3913-1. Weights
392
393A parent's resource is distributed by adding up the weights of all
394active children and giving each the fraction matching the ratio of its
395weight against the sum. As only children which can make use of the
396resource at the moment participate in the distribution, this is
397work-conserving. Due to the dynamic nature, this model is usually
398used for stateless resources.
399
400All weights are in the range [1, 10000] with the default at 100. This
401allows symmetric multiplicative biases in both directions at fine
402enough granularity while staying in the intuitive range.
403
404As long as the weight is in range, all configuration combinations are
405valid and there is no reason to reject configuration changes or
406process migrations.
407
408"cpu.weight" proportionally distributes CPU cycles to active children
409and is an example of this type.
410
411
4123-2. Limits
413
414A child can only consume upto the configured amount of the resource.
415Limits can be over-committed - the sum of the limits of children can
416exceed the amount of resource available to the parent.
417
418Limits are in the range [0, max] and defaults to "max", which is noop.
419
420As limits can be over-committed, all configuration combinations are
421valid and there is no reason to reject configuration changes or
422process migrations.
423
424"io.max" limits the maximum BPS and/or IOPS that a cgroup can consume
425on an IO device and is an example of this type.
426
427
4283-3. Protections
429
430A cgroup is protected to be allocated upto the configured amount of
431the resource if the usages of all its ancestors are under their
432protected levels. Protections can be hard guarantees or best effort
433soft boundaries. Protections can also be over-committed in which case
434only upto the amount available to the parent is protected among
435children.
436
437Protections are in the range [0, max] and defaults to 0, which is
438noop.
439
440As protections can be over-committed, all configuration combinations
441are valid and there is no reason to reject configuration changes or
442process migrations.
443
444"memory.low" implements best-effort memory protection and is an
445example of this type.
446
447
4483-4. Allocations
449
450A cgroup is exclusively allocated a certain amount of a finite
451resource. Allocations can't be over-committed - the sum of the
452allocations of children can not exceed the amount of resource
453available to the parent.
454
455Allocations are in the range [0, max] and defaults to 0, which is no
456resource.
457
458As allocations can't be over-committed, some configuration
459combinations are invalid and should be rejected. Also, if the
460resource is mandatory for execution of processes, process migrations
461may be rejected.
462
463"cpu.rt.max" hard-allocates realtime slices and is an example of this
464type.
465
466
4674. Interface Files
468
4694-1. Format
470
471All interface files should be in one of the following formats whenever
472possible.
473
474 New-line separated values
475 (when only one value can be written at once)
476
477 VAL0\n
478 VAL1\n
479 ...
480
481 Space separated values
482 (when read-only or multiple values can be written at once)
483
484 VAL0 VAL1 ...\n
485
486 Flat keyed
487
488 KEY0 VAL0\n
489 KEY1 VAL1\n
490 ...
491
492 Nested keyed
493
494 KEY0 SUB_KEY0=VAL00 SUB_KEY1=VAL01...
495 KEY1 SUB_KEY0=VAL10 SUB_KEY1=VAL11...
496 ...
497
498For a writable file, the format for writing should generally match
499reading; however, controllers may allow omitting later fields or
500implement restricted shortcuts for most common use cases.
501
502For both flat and nested keyed files, only the values for a single key
503can be written at a time. For nested keyed files, the sub key pairs
504may be specified in any order and not all pairs have to be specified.
505
506
5074-2. Conventions
508
509- Settings for a single feature should be contained in a single file.
510
511- The root cgroup should be exempt from resource control and thus
512 shouldn't have resource control interface files. Also,
513 informational files on the root cgroup which end up showing global
514 information available elsewhere shouldn't exist.
515
516- If a controller implements weight based resource distribution, its
517 interface file should be named "weight" and have the range [1,
518 10000] with 100 as the default. The values are chosen to allow
519 enough and symmetric bias in both directions while keeping it
520 intuitive (the default is 100%).
521
522- If a controller implements an absolute resource guarantee and/or
523 limit, the interface files should be named "min" and "max"
524 respectively. If a controller implements best effort resource
525 guarantee and/or limit, the interface files should be named "low"
526 and "high" respectively.
527
528 In the above four control files, the special token "max" should be
529 used to represent upward infinity for both reading and writing.
530
531- If a setting has a configurable default value and keyed specific
532 overrides, the default entry should be keyed with "default" and
533 appear as the first entry in the file.
534
535 The default value can be updated by writing either "default $VAL" or
536 "$VAL".
537
538 When writing to update a specific override, "default" can be used as
539 the value to indicate removal of the override. Override entries
540 with "default" as the value must not appear when read.
541
542 For example, a setting which is keyed by major:minor device numbers
543 with integer values may look like the following.
544
545 # cat cgroup-example-interface-file
546 default 150
547 8:0 300
548
549 The default value can be updated by
550
551 # echo 125 > cgroup-example-interface-file
552
553 or
554
555 # echo "default 125" > cgroup-example-interface-file
556
557 An override can be set by
558
559 # echo "8:16 170" > cgroup-example-interface-file
560
561 and cleared by
562
563 # echo "8:0 default" > cgroup-example-interface-file
564 # cat cgroup-example-interface-file
565 default 125
566 8:16 170
567
568- For events which are not very high frequency, an interface file
569 "events" should be created which lists event key value pairs.
570 Whenever a notifiable event happens, file modified event should be
571 generated on the file.
572
573
5744-3. Core Interface Files
575
576All cgroup core files are prefixed with "cgroup."
577
578 cgroup.procs
579
580 A read-write new-line separated values file which exists on
581 all cgroups.
582
583 When read, it lists the PIDs of all processes which belong to
584 the cgroup one-per-line. The PIDs are not ordered and the
585 same PID may show up more than once if the process got moved
586 to another cgroup and then back or the PID got recycled while
587 reading.
588
589 A PID can be written to migrate the process associated with
590 the PID to the cgroup. The writer should match all of the
591 following conditions.
592
593 - Its euid is either root or must match either uid or suid of
594 the target process.
595
596 - It must have write access to the "cgroup.procs" file.
597
598 - It must have write access to the "cgroup.procs" file of the
599 common ancestor of the source and destination cgroups.
600
601 When delegating a sub-hierarchy, write access to this file
602 should be granted along with the containing directory.
603
604 cgroup.controllers
605
606 A read-only space separated values file which exists on all
607 cgroups.
608
609 It shows space separated list of all controllers available to
610 the cgroup. The controllers are not ordered.
611
612 cgroup.subtree_control
613
614 A read-write space separated values file which exists on all
615 cgroups. Starts out empty.
616
617 When read, it shows space separated list of the controllers
618 which are enabled to control resource distribution from the
619 cgroup to its children.
620
621 Space separated list of controllers prefixed with '+' or '-'
622 can be written to enable or disable controllers. A controller
623 name prefixed with '+' enables the controller and '-'
624 disables. If a controller appears more than once on the list,
625 the last one is effective. When multiple enable and disable
626 operations are specified, either all succeed or all fail.
627
628 cgroup.events
629
630 A read-only flat-keyed file which exists on non-root cgroups.
631 The following entries are defined. Unless specified
632 otherwise, a value change in this file generates a file
633 modified event.
634
635 populated
636
637 1 if the cgroup or its descendants contains any live
638 processes; otherwise, 0.
639
640
6415. Controllers
642
6435-1. CPU
644
645[NOTE: The interface for the cpu controller hasn't been merged yet]
646
647The "cpu" controllers regulates distribution of CPU cycles. This
648controller implements weight and absolute bandwidth limit models for
649normal scheduling policy and absolute bandwidth allocation model for
650realtime scheduling policy.
651
652
6535-1-1. CPU Interface Files
654
655All time durations are in microseconds.
656
657 cpu.stat
658
659 A read-only flat-keyed file which exists on non-root cgroups.
660
661 It reports the following six stats.
662
663 usage_usec
664 user_usec
665 system_usec
666 nr_periods
667 nr_throttled
668 throttled_usec
669
670 cpu.weight
671
672 A read-write single value file which exists on non-root
673 cgroups. The default is "100".
674
675 The weight in the range [1, 10000].
676
677 cpu.max
678
679 A read-write two value file which exists on non-root cgroups.
680 The default is "max 100000".
681
682 The maximum bandwidth limit. It's in the following format.
683
684 $MAX $PERIOD
685
686 which indicates that the group may consume upto $MAX in each
687 $PERIOD duration. "max" for $MAX indicates no limit. If only
688 one number is written, $MAX is updated.
689
690 cpu.rt.max
691
692 [NOTE: The semantics of this file is still under discussion and the
693 interface hasn't been merged yet]
694
695 A read-write two value file which exists on all cgroups.
696 The default is "0 100000".
697
698 The maximum realtime runtime allocation. Over-committing
699 configurations are disallowed and process migrations are
700 rejected if not enough bandwidth is available. It's in the
701 following format.
702
703 $MAX $PERIOD
704
705 which indicates that the group may consume upto $MAX in each
706 $PERIOD duration. If only one number is written, $MAX is
707 updated.
708
709
7105-2. Memory
711
712The "memory" controller regulates distribution of memory. Memory is
713stateful and implements both limit and protection models. Due to the
714intertwining between memory usage and reclaim pressure and the
715stateful nature of memory, the distribution model is relatively
716complex.
717
718While not completely water-tight, all major memory usages by a given
719cgroup are tracked so that the total memory consumption can be
720accounted and controlled to a reasonable extent. Currently, the
721following types of memory usages are tracked.
722
723- Userland memory - page cache and anonymous memory.
724
725- Kernel data structures such as dentries and inodes.
726
727- TCP socket buffers.
728
729The above list may expand in the future for better coverage.
730
731
7325-2-1. Memory Interface Files
733
734All memory amounts are in bytes. If a value which is not aligned to
735PAGE_SIZE is written, the value may be rounded up to the closest
736PAGE_SIZE multiple when read back.
737
738 memory.current
739
740 A read-only single value file which exists on non-root
741 cgroups.
742
743 The total amount of memory currently being used by the cgroup
744 and its descendants.
745
746 memory.low
747
748 A read-write single value file which exists on non-root
749 cgroups. The default is "0".
750
751 Best-effort memory protection. If the memory usages of a
752 cgroup and all its ancestors are below their low boundaries,
753 the cgroup's memory won't be reclaimed unless memory can be
754 reclaimed from unprotected cgroups.
755
756 Putting more memory than generally available under this
757 protection is discouraged.
758
759 memory.high
760
761 A read-write single value file which exists on non-root
762 cgroups. The default is "max".
763
764 Memory usage throttle limit. This is the main mechanism to
765 control memory usage of a cgroup. If a cgroup's usage goes
766 over the high boundary, the processes of the cgroup are
767 throttled and put under heavy reclaim pressure.
768
769 Going over the high limit never invokes the OOM killer and
770 under extreme conditions the limit may be breached.
771
772 memory.max
773
774 A read-write single value file which exists on non-root
775 cgroups. The default is "max".
776
777 Memory usage hard limit. This is the final protection
778 mechanism. If a cgroup's memory usage reaches this limit and
779 can't be reduced, the OOM killer is invoked in the cgroup.
780 Under certain circumstances, the usage may go over the limit
781 temporarily.
782
783 This is the ultimate protection mechanism. As long as the
784 high limit is used and monitored properly, this limit's
785 utility is limited to providing the final safety net.
786
787 memory.events
788
789 A read-only flat-keyed file which exists on non-root cgroups.
790 The following entries are defined. Unless specified
791 otherwise, a value change in this file generates a file
792 modified event.
793
794 low
795
796 The number of times the cgroup is reclaimed due to
797 high memory pressure even though its usage is under
798 the low boundary. This usually indicates that the low
799 boundary is over-committed.
800
801 high
802
803 The number of times processes of the cgroup are
804 throttled and routed to perform direct memory reclaim
805 because the high memory boundary was exceeded. For a
806 cgroup whose memory usage is capped by the high limit
807 rather than global memory pressure, this event's
808 occurrences are expected.
809
810 max
811
812 The number of times the cgroup's memory usage was
813 about to go over the max boundary. If direct reclaim
814 fails to bring it down, the OOM killer is invoked.
815
816 oom
817
818 The number of times the OOM killer has been invoked in
819 the cgroup. This may not exactly match the number of
820 processes killed but should generally be close.
821
Johannes Weiner587d9f72016-01-20 15:03:19 -0800822 memory.stat
823
824 A read-only flat-keyed file which exists on non-root cgroups.
825
826 This breaks down the cgroup's memory footprint into different
827 types of memory, type-specific details, and other information
828 on the state and past events of the memory management system.
829
830 All memory amounts are in bytes.
831
832 The entries are ordered to be human readable, and new entries
833 can show up in the middle. Don't rely on items remaining in a
834 fixed position; use the keys to look up specific values!
835
836 anon
837
838 Amount of memory used in anonymous mappings such as
839 brk(), sbrk(), and mmap(MAP_ANONYMOUS)
840
841 file
842
843 Amount of memory used to cache filesystem data,
844 including tmpfs and shared memory.
845
Johannes Weiner4758e192016-02-02 16:57:41 -0800846 sock
847
848 Amount of memory used in network transmission buffers
849
Johannes Weiner587d9f72016-01-20 15:03:19 -0800850 file_mapped
851
852 Amount of cached filesystem data mapped with mmap()
853
854 file_dirty
855
856 Amount of cached filesystem data that was modified but
857 not yet written back to disk
858
859 file_writeback
860
861 Amount of cached filesystem data that was modified and
862 is currently being written back to disk
863
864 inactive_anon
865 active_anon
866 inactive_file
867 active_file
868 unevictable
869
870 Amount of memory, swap-backed and filesystem-backed,
871 on the internal memory management lists used by the
872 page reclaim algorithm
873
874 pgfault
875
876 Total number of page faults incurred
877
878 pgmajfault
879
880 Number of major page faults incurred
881
Vladimir Davydov3e24b192016-01-20 15:03:13 -0800882 memory.swap.current
883
884 A read-only single value file which exists on non-root
885 cgroups.
886
887 The total amount of swap currently being used by the cgroup
888 and its descendants.
889
890 memory.swap.max
891
892 A read-write single value file which exists on non-root
893 cgroups. The default is "max".
894
895 Swap usage hard limit. If a cgroup's swap usage reaches this
896 limit, anonymous meomry of the cgroup will not be swapped out.
897
Tejun Heo6c292092015-11-16 11:13:34 -0500898
8995-2-2. General Usage
900
901"memory.high" is the main mechanism to control memory usage.
902Over-committing on high limit (sum of high limits > available memory)
903and letting global memory pressure to distribute memory according to
904usage is a viable strategy.
905
906Because breach of the high limit doesn't trigger the OOM killer but
907throttles the offending cgroup, a management agent has ample
908opportunities to monitor and take appropriate actions such as granting
909more memory or terminating the workload.
910
911Determining whether a cgroup has enough memory is not trivial as
912memory usage doesn't indicate whether the workload can benefit from
913more memory. For example, a workload which writes data received from
914network to a file can use all available memory but can also operate as
915performant with a small amount of memory. A measure of memory
916pressure - how much the workload is being impacted due to lack of
917memory - is necessary to determine whether a workload needs more
918memory; unfortunately, memory pressure monitoring mechanism isn't
919implemented yet.
920
921
9225-2-3. Memory Ownership
923
924A memory area is charged to the cgroup which instantiated it and stays
925charged to the cgroup until the area is released. Migrating a process
926to a different cgroup doesn't move the memory usages that it
927instantiated while in the previous cgroup to the new cgroup.
928
929A memory area may be used by processes belonging to different cgroups.
930To which cgroup the area will be charged is in-deterministic; however,
931over time, the memory area is likely to end up in a cgroup which has
932enough memory allowance to avoid high reclaim pressure.
933
934If a cgroup sweeps a considerable amount of memory which is expected
935to be accessed repeatedly by other cgroups, it may make sense to use
936POSIX_FADV_DONTNEED to relinquish the ownership of memory areas
937belonging to the affected files to ensure correct memory ownership.
938
939
9405-3. IO
941
942The "io" controller regulates the distribution of IO resources. This
943controller implements both weight based and absolute bandwidth or IOPS
944limit distribution; however, weight based distribution is available
945only if cfq-iosched is in use and neither scheme is available for
946blk-mq devices.
947
948
9495-3-1. IO Interface Files
950
951 io.stat
952
953 A read-only nested-keyed file which exists on non-root
954 cgroups.
955
956 Lines are keyed by $MAJ:$MIN device numbers and not ordered.
957 The following nested keys are defined.
958
959 rbytes Bytes read
960 wbytes Bytes written
961 rios Number of read IOs
962 wios Number of write IOs
963
964 An example read output follows.
965
966 8:16 rbytes=1459200 wbytes=314773504 rios=192 wios=353
967 8:0 rbytes=90430464 wbytes=299008000 rios=8950 wios=1252
968
969 io.weight
970
971 A read-write flat-keyed file which exists on non-root cgroups.
972 The default is "default 100".
973
974 The first line is the default weight applied to devices
975 without specific override. The rest are overrides keyed by
976 $MAJ:$MIN device numbers and not ordered. The weights are in
977 the range [1, 10000] and specifies the relative amount IO time
978 the cgroup can use in relation to its siblings.
979
980 The default weight can be updated by writing either "default
981 $WEIGHT" or simply "$WEIGHT". Overrides can be set by writing
982 "$MAJ:$MIN $WEIGHT" and unset by writing "$MAJ:$MIN default".
983
984 An example read output follows.
985
986 default 100
987 8:16 200
988 8:0 50
989
990 io.max
991
992 A read-write nested-keyed file which exists on non-root
993 cgroups.
994
995 BPS and IOPS based IO limit. Lines are keyed by $MAJ:$MIN
996 device numbers and not ordered. The following nested keys are
997 defined.
998
999 rbps Max read bytes per second
1000 wbps Max write bytes per second
1001 riops Max read IO operations per second
1002 wiops Max write IO operations per second
1003
1004 When writing, any number of nested key-value pairs can be
1005 specified in any order. "max" can be specified as the value
1006 to remove a specific limit. If the same key is specified
1007 multiple times, the outcome is undefined.
1008
1009 BPS and IOPS are measured in each IO direction and IOs are
1010 delayed if limit is reached. Temporary bursts are allowed.
1011
1012 Setting read limit at 2M BPS and write at 120 IOPS for 8:16.
1013
1014 echo "8:16 rbps=2097152 wiops=120" > io.max
1015
1016 Reading returns the following.
1017
1018 8:16 rbps=2097152 wbps=max riops=max wiops=120
1019
1020 Write IOPS limit can be removed by writing the following.
1021
1022 echo "8:16 wiops=max" > io.max
1023
1024 Reading now returns the following.
1025
1026 8:16 rbps=2097152 wbps=max riops=max wiops=max
1027
1028
10295-3-2. Writeback
1030
1031Page cache is dirtied through buffered writes and shared mmaps and
1032written asynchronously to the backing filesystem by the writeback
1033mechanism. Writeback sits between the memory and IO domains and
1034regulates the proportion of dirty memory by balancing dirtying and
1035write IOs.
1036
1037The io controller, in conjunction with the memory controller,
1038implements control of page cache writeback IOs. The memory controller
1039defines the memory domain that dirty memory ratio is calculated and
1040maintained for and the io controller defines the io domain which
1041writes out dirty pages for the memory domain. Both system-wide and
1042per-cgroup dirty memory states are examined and the more restrictive
1043of the two is enforced.
1044
1045cgroup writeback requires explicit support from the underlying
1046filesystem. Currently, cgroup writeback is implemented on ext2, ext4
1047and btrfs. On other filesystems, all writeback IOs are attributed to
1048the root cgroup.
1049
1050There are inherent differences in memory and writeback management
1051which affects how cgroup ownership is tracked. Memory is tracked per
1052page while writeback per inode. For the purpose of writeback, an
1053inode is assigned to a cgroup and all IO requests to write dirty pages
1054from the inode are attributed to that cgroup.
1055
1056As cgroup ownership for memory is tracked per page, there can be pages
1057which are associated with different cgroups than the one the inode is
1058associated with. These are called foreign pages. The writeback
1059constantly keeps track of foreign pages and, if a particular foreign
1060cgroup becomes the majority over a certain period of time, switches
1061the ownership of the inode to that cgroup.
1062
1063While this model is enough for most use cases where a given inode is
1064mostly dirtied by a single cgroup even when the main writing cgroup
1065changes over time, use cases where multiple cgroups write to a single
1066inode simultaneously are not supported well. In such circumstances, a
1067significant portion of IOs are likely to be attributed incorrectly.
1068As memory controller assigns page ownership on the first use and
1069doesn't update it until the page is released, even if writeback
1070strictly follows page ownership, multiple cgroups dirtying overlapping
1071areas wouldn't work as expected. It's recommended to avoid such usage
1072patterns.
1073
1074The sysctl knobs which affect writeback behavior are applied to cgroup
1075writeback as follows.
1076
1077 vm.dirty_background_ratio
1078 vm.dirty_ratio
1079
1080 These ratios apply the same to cgroup writeback with the
1081 amount of available memory capped by limits imposed by the
1082 memory controller and system-wide clean memory.
1083
1084 vm.dirty_background_bytes
1085 vm.dirty_bytes
1086
1087 For cgroup writeback, this is calculated into ratio against
1088 total available memory and applied the same way as
1089 vm.dirty[_background]_ratio.
1090
1091
1092P. Information on Kernel Programming
1093
1094This section contains kernel programming information in the areas
1095where interacting with cgroup is necessary. cgroup core and
1096controllers are not covered.
1097
1098
1099P-1. Filesystem Support for Writeback
1100
1101A filesystem can support cgroup writeback by updating
1102address_space_operations->writepage[s]() to annotate bio's using the
1103following two functions.
1104
1105 wbc_init_bio(@wbc, @bio)
1106
1107 Should be called for each bio carrying writeback data and
1108 associates the bio with the inode's owner cgroup. Can be
1109 called anytime between bio allocation and submission.
1110
1111 wbc_account_io(@wbc, @page, @bytes)
1112
1113 Should be called for each data segment being written out.
1114 While this function doesn't care exactly when it's called
1115 during the writeback session, it's the easiest and most
1116 natural to call it as data segments are added to a bio.
1117
1118With writeback bio's annotated, cgroup support can be enabled per
1119super_block by setting SB_I_CGROUPWB in ->s_iflags. This allows for
1120selective disabling of cgroup writeback support which is helpful when
1121certain filesystem features, e.g. journaled data mode, are
1122incompatible.
1123
1124wbc_init_bio() binds the specified bio to its cgroup. Depending on
1125the configuration, the bio may be executed at a lower priority and if
1126the writeback session is holding shared resources, e.g. a journal
1127entry, may lead to priority inversion. There is no one easy solution
1128for the problem. Filesystems can try to work around specific problem
1129cases by skipping wbc_init_bio() or using bio_associate_blkcg()
1130directly.
1131
1132
1133D. Deprecated v1 Core Features
1134
1135- Multiple hierarchies including named ones are not supported.
1136
1137- All mount options and remounting are not supported.
1138
1139- The "tasks" file is removed and "cgroup.procs" is not sorted.
1140
1141- "cgroup.clone_children" is removed.
1142
1143- /proc/cgroups is meaningless for v2. Use "cgroup.controllers" file
1144 at the root instead.
1145
1146
1147R. Issues with v1 and Rationales for v2
1148
1149R-1. Multiple Hierarchies
1150
1151cgroup v1 allowed an arbitrary number of hierarchies and each
1152hierarchy could host any number of controllers. While this seemed to
1153provide a high level of flexibility, it wasn't useful in practice.
1154
1155For example, as there is only one instance of each controller, utility
1156type controllers such as freezer which can be useful in all
1157hierarchies could only be used in one. The issue is exacerbated by
1158the fact that controllers couldn't be moved to another hierarchy once
1159hierarchies were populated. Another issue was that all controllers
1160bound to a hierarchy were forced to have exactly the same view of the
1161hierarchy. It wasn't possible to vary the granularity depending on
1162the specific controller.
1163
1164In practice, these issues heavily limited which controllers could be
1165put on the same hierarchy and most configurations resorted to putting
1166each controller on its own hierarchy. Only closely related ones, such
1167as the cpu and cpuacct controllers, made sense to be put on the same
1168hierarchy. This often meant that userland ended up managing multiple
1169similar hierarchies repeating the same steps on each hierarchy
1170whenever a hierarchy management operation was necessary.
1171
1172Furthermore, support for multiple hierarchies came at a steep cost.
1173It greatly complicated cgroup core implementation but more importantly
1174the support for multiple hierarchies restricted how cgroup could be
1175used in general and what controllers was able to do.
1176
1177There was no limit on how many hierarchies there might be, which meant
1178that a thread's cgroup membership couldn't be described in finite
1179length. The key might contain any number of entries and was unlimited
1180in length, which made it highly awkward to manipulate and led to
1181addition of controllers which existed only to identify membership,
1182which in turn exacerbated the original problem of proliferating number
1183of hierarchies.
1184
1185Also, as a controller couldn't have any expectation regarding the
1186topologies of hierarchies other controllers might be on, each
1187controller had to assume that all other controllers were attached to
1188completely orthogonal hierarchies. This made it impossible, or at
1189least very cumbersome, for controllers to cooperate with each other.
1190
1191In most use cases, putting controllers on hierarchies which are
1192completely orthogonal to each other isn't necessary. What usually is
1193called for is the ability to have differing levels of granularity
1194depending on the specific controller. In other words, hierarchy may
1195be collapsed from leaf towards root when viewed from specific
1196controllers. For example, a given configuration might not care about
1197how memory is distributed beyond a certain level while still wanting
1198to control how CPU cycles are distributed.
1199
1200
1201R-2. Thread Granularity
1202
1203cgroup v1 allowed threads of a process to belong to different cgroups.
1204This didn't make sense for some controllers and those controllers
1205ended up implementing different ways to ignore such situations but
1206much more importantly it blurred the line between API exposed to
1207individual applications and system management interface.
1208
1209Generally, in-process knowledge is available only to the process
1210itself; thus, unlike service-level organization of processes,
1211categorizing threads of a process requires active participation from
1212the application which owns the target process.
1213
1214cgroup v1 had an ambiguously defined delegation model which got abused
1215in combination with thread granularity. cgroups were delegated to
1216individual applications so that they can create and manage their own
1217sub-hierarchies and control resource distributions along them. This
1218effectively raised cgroup to the status of a syscall-like API exposed
1219to lay programs.
1220
1221First of all, cgroup has a fundamentally inadequate interface to be
1222exposed this way. For a process to access its own knobs, it has to
1223extract the path on the target hierarchy from /proc/self/cgroup,
1224construct the path by appending the name of the knob to the path, open
1225and then read and/or write to it. This is not only extremely clunky
1226and unusual but also inherently racy. There is no conventional way to
1227define transaction across the required steps and nothing can guarantee
1228that the process would actually be operating on its own sub-hierarchy.
1229
1230cgroup controllers implemented a number of knobs which would never be
1231accepted as public APIs because they were just adding control knobs to
1232system-management pseudo filesystem. cgroup ended up with interface
1233knobs which were not properly abstracted or refined and directly
1234revealed kernel internal details. These knobs got exposed to
1235individual applications through the ill-defined delegation mechanism
1236effectively abusing cgroup as a shortcut to implementing public APIs
1237without going through the required scrutiny.
1238
1239This was painful for both userland and kernel. Userland ended up with
1240misbehaving and poorly abstracted interfaces and kernel exposing and
1241locked into constructs inadvertently.
1242
1243
1244R-3. Competition Between Inner Nodes and Threads
1245
1246cgroup v1 allowed threads to be in any cgroups which created an
1247interesting problem where threads belonging to a parent cgroup and its
1248children cgroups competed for resources. This was nasty as two
1249different types of entities competed and there was no obvious way to
1250settle it. Different controllers did different things.
1251
1252The cpu controller considered threads and cgroups as equivalents and
1253mapped nice levels to cgroup weights. This worked for some cases but
1254fell flat when children wanted to be allocated specific ratios of CPU
1255cycles and the number of internal threads fluctuated - the ratios
1256constantly changed as the number of competing entities fluctuated.
1257There also were other issues. The mapping from nice level to weight
1258wasn't obvious or universal, and there were various other knobs which
1259simply weren't available for threads.
1260
1261The io controller implicitly created a hidden leaf node for each
1262cgroup to host the threads. The hidden leaf had its own copies of all
1263the knobs with "leaf_" prefixed. While this allowed equivalent
1264control over internal threads, it was with serious drawbacks. It
1265always added an extra layer of nesting which wouldn't be necessary
1266otherwise, made the interface messy and significantly complicated the
1267implementation.
1268
1269The memory controller didn't have a way to control what happened
1270between internal tasks and child cgroups and the behavior was not
1271clearly defined. There were attempts to add ad-hoc behaviors and
1272knobs to tailor the behavior to specific workloads which would have
1273led to problems extremely difficult to resolve in the long term.
1274
1275Multiple controllers struggled with internal tasks and came up with
1276different ways to deal with it; unfortunately, all the approaches were
1277severely flawed and, furthermore, the widely different behaviors
1278made cgroup as a whole highly inconsistent.
1279
1280This clearly is a problem which needs to be addressed from cgroup core
1281in a uniform way.
1282
1283
1284R-4. Other Interface Issues
1285
1286cgroup v1 grew without oversight and developed a large number of
1287idiosyncrasies and inconsistencies. One issue on the cgroup core side
1288was how an empty cgroup was notified - a userland helper binary was
1289forked and executed for each event. The event delivery wasn't
1290recursive or delegatable. The limitations of the mechanism also led
1291to in-kernel event delivery filtering mechanism further complicating
1292the interface.
1293
1294Controller interfaces were problematic too. An extreme example is
1295controllers completely ignoring hierarchical organization and treating
1296all cgroups as if they were all located directly under the root
1297cgroup. Some controllers exposed a large amount of inconsistent
1298implementation details to userland.
1299
1300There also was no consistency across controllers. When a new cgroup
1301was created, some controllers defaulted to not imposing extra
1302restrictions while others disallowed any resource usage until
1303explicitly configured. Configuration knobs for the same type of
1304control used widely differing naming schemes and formats. Statistics
1305and information knobs were named arbitrarily and used different
1306formats and units even in the same controller.
1307
1308cgroup v2 establishes common conventions where appropriate and updates
1309controllers so that they expose minimal and consistent interfaces.
1310
1311
1312R-5. Controller Issues and Remedies
1313
1314R-5-1. Memory
1315
1316The original lower boundary, the soft limit, is defined as a limit
1317that is per default unset. As a result, the set of cgroups that
1318global reclaim prefers is opt-in, rather than opt-out. The costs for
1319optimizing these mostly negative lookups are so high that the
1320implementation, despite its enormous size, does not even provide the
1321basic desirable behavior. First off, the soft limit has no
1322hierarchical meaning. All configured groups are organized in a global
1323rbtree and treated like equal peers, regardless where they are located
1324in the hierarchy. This makes subtree delegation impossible. Second,
1325the soft limit reclaim pass is so aggressive that it not just
1326introduces high allocation latencies into the system, but also impacts
1327system performance due to overreclaim, to the point where the feature
1328becomes self-defeating.
1329
1330The memory.low boundary on the other hand is a top-down allocated
1331reserve. A cgroup enjoys reclaim protection when it and all its
1332ancestors are below their low boundaries, which makes delegation of
1333subtrees possible. Secondly, new cgroups have no reserve per default
1334and in the common case most cgroups are eligible for the preferred
1335reclaim pass. This allows the new low boundary to be efficiently
1336implemented with just a minor addition to the generic reclaim code,
1337without the need for out-of-band data structures and reclaim passes.
1338Because the generic reclaim code considers all cgroups except for the
1339ones running low in the preferred first reclaim pass, overreclaim of
1340individual groups is eliminated as well, resulting in much better
1341overall workload performance.
1342
1343The original high boundary, the hard limit, is defined as a strict
1344limit that can not budge, even if the OOM killer has to be called.
1345But this generally goes against the goal of making the most out of the
1346available memory. The memory consumption of workloads varies during
1347runtime, and that requires users to overcommit. But doing that with a
1348strict upper limit requires either a fairly accurate prediction of the
1349working set size or adding slack to the limit. Since working set size
1350estimation is hard and error prone, and getting it wrong results in
1351OOM kills, most users tend to err on the side of a looser limit and
1352end up wasting precious resources.
1353
1354The memory.high boundary on the other hand can be set much more
1355conservatively. When hit, it throttles allocations by forcing them
1356into direct reclaim to work off the excess, but it never invokes the
1357OOM killer. As a result, a high boundary that is chosen too
1358aggressively will not terminate the processes, but instead it will
1359lead to gradual performance degradation. The user can monitor this
1360and make corrections until the minimal memory footprint that still
1361gives acceptable performance is found.
1362
1363In extreme cases, with many concurrent allocations and a complete
1364breakdown of reclaim progress within the group, the high boundary can
1365be exceeded. But even then it's mostly better to satisfy the
1366allocation from the slack available in other groups or the rest of the
1367system than killing the group. Otherwise, memory.max is there to
1368limit this type of spillover and ultimately contain buggy or even
1369malicious applications.
Vladimir Davydov3e24b192016-01-20 15:03:13 -08001370
1371The combined memory+swap accounting and limiting is replaced by real
1372control over swap space.
1373
1374The main argument for a combined memory+swap facility in the original
1375cgroup design was that global or parental pressure would always be
1376able to swap all anonymous memory of a child group, regardless of the
1377child's own (possibly untrusted) configuration. However, untrusted
1378groups can sabotage swapping by other means - such as referencing its
1379anonymous memory in a tight loop - and an admin can not assume full
1380swappability when overcommitting untrusted jobs.
1381
1382For trusted jobs, on the other hand, a combined counter is not an
1383intuitive userspace interface, and it flies in the face of the idea
1384that cgroup controllers should account and limit specific physical
1385resources. Swap space is a resource like all others in the system,
1386and that's why unified hierarchy allows distributing it separately.