Documentation/cgroup-v2.txt

Control Group v2

October, 2015		Tejun Heo <tj@kernel.org>

This is the authoritative documentation on the design, interface and
conventions of cgroup v2.  It describes all userland-visible aspects
of cgroup including core and specific controller behaviors.  All
future changes must be reflected in this document.  Documentation for
v1 is available under Documentation/cgroup-v1/.

CONTENTS

1. Introduction
  1-1. Terminology
  1-2. What is cgroup?
2. Basic Operations
  2-1. Mounting
  2-2. Organizing Processes
  2-3. [Un]populated Notification
  2-4. Controlling Controllers
    2-4-1. Enabling and Disabling
    2-4-2. Top-down Constraint
    2-4-3. No Internal Process Constraint
  2-5. Delegation
    2-5-1. Model of Delegation
    2-5-2. Delegation Containment
  2-6. Guidelines
    2-6-1. Organize Once and Control
    2-6-2. Avoid Name Collisions
3. Resource Distribution Models
  3-1. Weights
  3-2. Limits
  3-3. Protections
  3-4. Allocations
4. Interface Files
  4-1. Format
  4-2. Conventions
  4-3. Core Interface Files
5. Controllers
  5-1. CPU
    5-1-1. CPU Interface Files
  5-2. Memory
    5-2-1. Memory Interface Files
    5-2-2. Usage Guidelines
    5-2-3. Memory Ownership
  5-3. IO
    5-3-1. IO Interface Files
    5-3-2. Writeback
  5-4. PID
    5-4-1. PID Interface Files
  5-5. RDMA
    5-5-1. RDMA Interface Files
  5-6. Misc
    5-6-1. perf_event
6. Namespace
  6-1. Basics
  6-2. The Root and Views
  6-3. Migration and setns(2)
  6-4. Interaction with Other Namespaces
P. Information on Kernel Programming
  P-1. Filesystem Support for Writeback
D. Deprecated v1 Core Features
R. Issues with v1 and Rationales for v2
  R-1. Multiple Hierarchies
  R-2. Thread Granularity
  R-3. Competition Between Inner Nodes and Threads
  R-4. Other Interface Issues
  R-5. Controller Issues and Remedies
    R-5-1. Memory


1. Introduction

1-1. Terminology

"cgroup" stands for "control group" and is never capitalized.  The
singular form is used to designate the whole feature and also as a
qualifier as in "cgroup controllers".  When explicitly referring to
multiple individual control groups, the plural form "cgroups" is used.


1-2. What is cgroup?

cgroup is a mechanism to organize processes hierarchically and
distribute system resources along the hierarchy in a controlled and
configurable manner.

cgroup is largely composed of two parts - the core and controllers.
cgroup core is primarily responsible for hierarchically organizing
processes.  A cgroup controller is usually responsible for
distributing a specific type of system resource along the hierarchy
although there are utility controllers which serve purposes other than
resource distribution.

cgroups form a tree structure and every process in the system belongs
to one and only one cgroup.  All threads of a process belong to the
same cgroup.  On creation, all processes are put in the cgroup that
the parent process belongs to at the time.  A process can be migrated
to another cgroup.  Migration of a process doesn't affect already
existing descendant processes.

Following certain structural constraints, controllers may be enabled or
disabled selectively on a cgroup.  All controller behaviors are
hierarchical - if a controller is enabled on a cgroup, it affects all
processes which belong to the cgroups consisting the inclusive
sub-hierarchy of the cgroup.  When a controller is enabled on a nested
cgroup, it always restricts the resource distribution further.  The
restrictions set closer to the root in the hierarchy can not be
overridden from further away.


2. Basic Operations

2-1. Mounting

Unlike v1, cgroup v2 has only single hierarchy.  The cgroup v2
hierarchy can be mounted with the following mount command.

  # mount -t cgroup2 none $MOUNT_POINT

cgroup2 filesystem has the magic number 0x63677270 ("cgrp").  All
controllers which support v2 and are not bound to a v1 hierarchy are
automatically bound to the v2 hierarchy and show up at the root.
Controllers which are not in active use in the v2 hierarchy can be
bound to other hierarchies.  This allows mixing v2 hierarchy with the
legacy v1 multiple hierarchies in a fully backward compatible way.

A controller can be moved across hierarchies only after the controller
is no longer referenced in its current hierarchy.  Because per-cgroup
controller states are destroyed asynchronously and controllers may
have lingering references, a controller may not show up immediately on
the v2 hierarchy after the final umount of the previous hierarchy.
Similarly, a controller should be fully disabled to be moved out of
the unified hierarchy and it may take some time for the disabled
controller to become available for other hierarchies; furthermore, due
to inter-controller dependencies, other controllers may need to be
disabled too.

While useful for development and manual configurations, moving
controllers dynamically between the v2 and other hierarchies is
strongly discouraged for production use.  It is recommended to decide
the hierarchies and controller associations before starting using the
controllers after system boot.

During transition to v2, system management software might still
automount the v1 cgroup filesystem and so hijack all controllers
during boot, before manual intervention is possible. To make testing
and experimenting easier, the kernel parameter cgroup_no_v1= allows
disabling controllers in v1 and make them always available in v2.


2-2. Organizing Processes

Initially, only the root cgroup exists to which all processes belong.
A child cgroup can be created by creating a sub-directory.

  # mkdir $CGROUP_NAME

A given cgroup may have multiple child cgroups forming a tree
structure.  Each cgroup has a read-writable interface file
"cgroup.procs".  When read, it lists the PIDs of all processes which
belong to the cgroup one-per-line.  The PIDs are not ordered and the
same PID may show up more than once if the process got moved to
another cgroup and then back or the PID got recycled while reading.

A process can be migrated into a cgroup by writing its PID to the
target cgroup's "cgroup.procs" file.  Only one process can be migrated
on a single write(2) call.  If a process is composed of multiple
threads, writing the PID of any thread migrates all threads of the
process.

When a process forks a child process, the new process is born into the
cgroup that the forking process belongs to at the time of the
operation.  After exit, a process stays associated with the cgroup
that it belonged to at the time of exit until it's reaped; however, a
zombie process does not appear in "cgroup.procs" and thus can't be
moved to another cgroup.

A cgroup which doesn't have any children or live processes can be
destroyed by removing the directory.  Note that a cgroup which doesn't
have any children and is associated only with zombie processes is
considered empty and can be removed.

  # rmdir $CGROUP_NAME

"/proc/$PID/cgroup" lists a process's cgroup membership.  If legacy
cgroup is in use in the system, this file may contain multiple lines,
one for each hierarchy.  The entry for cgroup v2 is always in the
format "0::$PATH".

  # cat /proc/842/cgroup
  ...
  0::/test-cgroup/test-cgroup-nested

If the process becomes a zombie and the cgroup it was associated with
is removed subsequently, " (deleted)" is appended to the path.

  # cat /proc/842/cgroup
  ...
  0::/test-cgroup/test-cgroup-nested (deleted)


2-3. [Un]populated Notification

Each non-root cgroup has a "cgroup.events" file which contains
"populated" field indicating whether the cgroup's sub-hierarchy has
live processes in it.  Its value is 0 if there is no live process in
the cgroup and its descendants; otherwise, 1.  poll and [id]notify
events are triggered when the value changes.  This can be used, for
example, to start a clean-up operation after all processes of a given
sub-hierarchy have exited.  The populated state updates and
notifications are recursive.  Consider the following sub-hierarchy
where the numbers in the parentheses represent the numbers of processes
in each cgroup.

  A(4) - B(0) - C(1)
              \ D(0)

A, B and C's "populated" fields would be 1 while D's 0.  After the one
process in C exits, B and C's "populated" fields would flip to "0" and
file modified events will be generated on the "cgroup.events" files of
both cgroups.


2-4. Controlling Controllers

2-4-1. Enabling and Disabling

Each cgroup has a "cgroup.controllers" file which lists all
controllers available for the cgroup to enable.

  # cat cgroup.controllers
  cpu io memory

No controller is enabled by default.  Controllers can be enabled and
disabled by writing to the "cgroup.subtree_control" file.

  # echo "+cpu +memory -io" > cgroup.subtree_control

Only controllers which are listed in "cgroup.controllers" can be
enabled.  When multiple operations are specified as above, either they
all succeed or fail.  If multiple operations on the same controller
are specified, the last one is effective.

Enabling a controller in a cgroup indicates that the distribution of
the target resource across its immediate children will be controlled.
Consider the following sub-hierarchy.  The enabled controllers are
listed in parentheses.

  A(cpu,memory) - B(memory) - C()
                            \ D()

As A has "cpu" and "memory" enabled, A will control the distribution
of CPU cycles and memory to its children, in this case, B.  As B has
"memory" enabled but not "CPU", C and D will compete freely on CPU
cycles but their division of memory available to B will be controlled.

As a controller regulates the distribution of the target resource to
the cgroup's children, enabling it creates the controller's interface
files in the child cgroups.  In the above example, enabling "cpu" on B
would create the "cpu." prefixed controller interface files in C and
D.  Likewise, disabling "memory" from B would remove the "memory."
prefixed controller interface files from C and D.  This means that the
controller interface files - anything which doesn't start with
"cgroup." are owned by the parent rather than the cgroup itself.


2-4-2. Top-down Constraint

Resources are distributed top-down and a cgroup can further distribute
a resource only if the resource has been distributed to it from the
parent.  This means that all non-root "cgroup.subtree_control" files
can only contain controllers which are enabled in the parent's
"cgroup.subtree_control" file.  A controller can be enabled only if
the parent has the controller enabled and a controller can't be
disabled if one or more children have it enabled.


2-4-3. No Internal Process Constraint

Non-root cgroups can only distribute resources to their children when
they don't have any processes of their own.  In other words, only
cgroups which don't contain any processes can have controllers enabled
in their "cgroup.subtree_control" files.

This guarantees that, when a controller is looking at the part of the
hierarchy which has it enabled, processes are always only on the
leaves.  This rules out situations where child cgroups compete against
internal processes of the parent.

The root cgroup is exempt from this restriction.  Root contains
processes and anonymous resource consumption which can't be associated
with any other cgroups and requires special treatment from most
controllers.  How resource consumption in the root cgroup is governed
is up to each controller.

Note that the restriction doesn't get in the way if there is no
enabled controller in the cgroup's "cgroup.subtree_control".  This is
important as otherwise it wouldn't be possible to create children of a
populated cgroup.  To control resource distribution of a cgroup, the
cgroup must create children and transfer all its processes to the
children before enabling controllers in its "cgroup.subtree_control"
file.


2-5. Delegation

2-5-1. Model of Delegation

A cgroup can be delegated to a less privileged user by granting write
access of the directory and its "cgroup.procs" file to the user.  Note
that resource control interface files in a given directory control the
distribution of the parent's resources and thus must not be delegated
along with the directory.

Once delegated, the user can build sub-hierarchy under the directory,
organize processes as it sees fit and further distribute the resources
it received from the parent.  The limits and other settings of all
resource controllers are hierarchical and regardless of what happens
in the delegated sub-hierarchy, nothing can escape the resource
restrictions imposed by the parent.

Currently, cgroup doesn't impose any restrictions on the number of
cgroups in or nesting depth of a delegated sub-hierarchy; however,
this may be limited explicitly in the future.


2-5-2. Delegation Containment

A delegated sub-hierarchy is contained in the sense that processes
can't be moved into or out of the sub-hierarchy by the delegatee.  For
a process with a non-root euid to migrate a target process into a
cgroup by writing its PID to the "cgroup.procs" file, the following
conditions must be met.

- The writer must have write access to the "cgroup.procs" file.

- The writer must have write access to the "cgroup.procs" file of the
  common ancestor of the source and destination cgroups.

The above two constraints ensure that while a delegatee may migrate
processes around freely in the delegated sub-hierarchy it can't pull
in from or push out to outside the sub-hierarchy.

For an example, let's assume cgroups C0 and C1 have been delegated to
user U0 who created C00, C01 under C0 and C10 under C1 as follows and
all processes under C0 and C1 belong to U0.

  ~~~~~~~~~~~~~ - C0 - C00
  ~ cgroup    ~      \ C01
  ~ hierarchy ~
  ~~~~~~~~~~~~~ - C1 - C10

Let's also say U0 wants to write the PID of a process which is
currently in C10 into "C00/cgroup.procs".  U0 has write access to the
file; however, the common ancestor of the source cgroup C10 and the
destination cgroup C00 is above the points of delegation and U0 would
not have write access to its "cgroup.procs" files and thus the write
will be denied with -EACCES.


2-6. Guidelines

2-6-1. Organize Once and Control

Migrating a process across cgroups is a relatively expensive operation
and stateful resources such as memory are not moved together with the
process.  This is an explicit design decision as there often exist
inherent trade-offs between migration and various hot paths in terms
of synchronization cost.

As such, migrating processes across cgroups frequently as a means to
apply different resource restrictions is discouraged.  A workload
should be assigned to a cgroup according to the system's logical and
resource structure once on start-up.  Dynamic adjustments to resource
distribution can be made by changing controller configuration through
the interface files.


2-6-2. Avoid Name Collisions

Interface files for a cgroup and its children cgroups occupy the same
directory and it is possible to create children cgroups which collide
with interface files.

All cgroup core interface files are prefixed with "cgroup." and each
controller's interface files are prefixed with the controller name and
a dot.  A controller's name is composed of lower case alphabets and
'_'s but never begins with an '_' so it can be used as the prefix
character for collision avoidance.  Also, interface file names won't
start or end with terms which are often used in categorizing workloads
such as job, service, slice, unit or workload.

cgroup doesn't do anything to prevent name collisions and it's the
user's responsibility to avoid them.


3. Resource Distribution Models

cgroup controllers implement several resource distribution schemes
depending on the resource type and expected use cases.  This section
describes major schemes in use along with their expected behaviors.


3-1. Weights

A parent's resource is distributed by adding up the weights of all
active children and giving each the fraction matching the ratio of its
weight against the sum.  As only children which can make use of the
resource at the moment participate in the distribution, this is
work-conserving.  Due to the dynamic nature, this model is usually
used for stateless resources.

All weights are in the range [1, 10000] with the default at 100.  This
allows symmetric multiplicative biases in both directions at fine
enough granularity while staying in the intuitive range.

As long as the weight is in range, all configuration combinations are
valid and there is no reason to reject configuration changes or
process migrations.

"cpu.weight" proportionally distributes CPU cycles to active children
and is an example of this type.


3-2. Limits

A child can only consume upto the configured amount of the resource.
Limits can be over-committed - the sum of the limits of children can
exceed the amount of resource available to the parent.

Limits are in the range [0, max] and defaults to "max", which is noop.

As limits can be over-committed, all configuration combinations are
valid and there is no reason to reject configuration changes or
process migrations.

"io.max" limits the maximum BPS and/or IOPS that a cgroup can consume
on an IO device and is an example of this type.


3-3. Protections

A cgroup is protected to be allocated upto the configured amount of
the resource if the usages of all its ancestors are under their
protected levels.  Protections can be hard guarantees or best effort
soft boundaries.  Protections can also be over-committed in which case
only upto the amount available to the parent is protected among
children.

Protections are in the range [0, max] and defaults to 0, which is
noop.

As protections can be over-committed, all configuration combinations
are valid and there is no reason to reject configuration changes or
process migrations.

"memory.low" implements best-effort memory protection and is an
example of this type.


3-4. Allocations

A cgroup is exclusively allocated a certain amount of a finite
resource.  Allocations can't be over-committed - the sum of the
allocations of children can not exceed the amount of resource
available to the parent.

Allocations are in the range [0, max] and defaults to 0, which is no
resource.

As allocations can't be over-committed, some configuration
combinations are invalid and should be rejected.  Also, if the
resource is mandatory for execution of processes, process migrations
may be rejected.

"cpu.rt.max" hard-allocates realtime slices and is an example of this
type.


4. Interface Files

4-1. Format

All interface files should be in one of the following formats whenever
possible.

  New-line separated values
  (when only one value can be written at once)

	VAL0\n
	VAL1\n
	...

  Space separated values
  (when read-only or multiple values can be written at once)

	VAL0 VAL1 ...\n

  Flat keyed

	KEY0 VAL0\n
	KEY1 VAL1\n
	...

  Nested keyed

	KEY0 SUB_KEY0=VAL00 SUB_KEY1=VAL01...
	KEY1 SUB_KEY0=VAL10 SUB_KEY1=VAL11...
	...

For a writable file, the format for writing should generally match
reading; however, controllers may allow omitting later fields or
implement restricted shortcuts for most common use cases.

For both flat and nested keyed files, only the values for a single key
can be written at a time.  For nested keyed files, the sub key pairs
may be specified in any order and not all pairs have to be specified.


4-2. Conventions

- Settings for a single feature should be contained in a single file.

- The root cgroup should be exempt from resource control and thus
  shouldn't have resource control interface files.  Also,
  informational files on the root cgroup which end up showing global
  information available elsewhere shouldn't exist.

- If a controller implements weight based resource distribution, its
  interface file should be named "weight" and have the range [1,
  10000] with 100 as the default.  The values are chosen to allow
  enough and symmetric bias in both directions while keeping it
  intuitive (the default is 100%).

- If a controller implements an absolute resource guarantee and/or
  limit, the interface files should be named "min" and "max"
  respectively.  If a controller implements best effort resource
  guarantee and/or limit, the interface files should be named "low"
  and "high" respectively.

  In the above four control files, the special token "max" should be
  used to represent upward infinity for both reading and writing.

- If a setting has a configurable default value and keyed specific
  overrides, the default entry should be keyed with "default" and
  appear as the first entry in the file.

  The default value can be updated by writing either "default $VAL" or
  "$VAL".

  When writing to update a specific override, "default" can be used as
  the value to indicate removal of the override.  Override entries
  with "default" as the value must not appear when read.

  For example, a setting which is keyed by major:minor device numbers
  with integer values may look like the following.

    # cat cgroup-example-interface-file
    default 150
    8:0 300

  The default value can be updated by

    # echo 125 > cgroup-example-interface-file

  or

    # echo "default 125" > cgroup-example-interface-file

  An override can be set by

    # echo "8:16 170" > cgroup-example-interface-file

  and cleared by

    # echo "8:0 default" > cgroup-example-interface-file
    # cat cgroup-example-interface-file
    default 125
    8:16 170

- For events which are not very high frequency, an interface file
  "events" should be created which lists event key value pairs.
  Whenever a notifiable event happens, file modified event should be
  generated on the file.


4-3. Core Interface Files

All cgroup core files are prefixed with "cgroup."

  cgroup.procs

	A read-write new-line separated values file which exists on
	all cgroups.

	When read, it lists the PIDs of all processes which belong to
	the cgroup one-per-line.  The PIDs are not ordered and the
	same PID may show up more than once if the process got moved
	to another cgroup and then back or the PID got recycled while
	reading.

	A PID can be written to migrate the process associated with
	the PID to the cgroup.  The writer should match all of the
	following conditions.

	- Its euid is either root or must match either uid or suid of
          the target process.

	- It must have write access to the "cgroup.procs" file.

	- It must have write access to the "cgroup.procs" file of the
	  common ancestor of the source and destination cgroups.

	When delegating a sub-hierarchy, write access to this file
	should be granted along with the containing directory.

  cgroup.controllers

	A read-only space separated values file which exists on all
	cgroups.

	It shows space separated list of all controllers available to
	the cgroup.  The controllers are not ordered.

  cgroup.subtree_control

	A read-write space separated values file which exists on all
	cgroups.  Starts out empty.

	When read, it shows space separated list of the controllers
	which are enabled to control resource distribution from the
	cgroup to its children.

	Space separated list of controllers prefixed with '+' or '-'
	can be written to enable or disable controllers.  A controller
	name prefixed with '+' enables the controller and '-'
	disables.  If a controller appears more than once on the list,
	the last one is effective.  When multiple enable and disable
	operations are specified, either all succeed or all fail.

  cgroup.events

	A read-only flat-keyed file which exists on non-root cgroups.
	The following entries are defined.  Unless specified
	otherwise, a value change in this file generates a file
	modified event.

	  populated

		1 if the cgroup or its descendants contains any live
		processes; otherwise, 0.


5. Controllers

5-1. CPU

[NOTE: The interface for the cpu controller hasn't been merged yet]

The "cpu" controllers regulates distribution of CPU cycles.  This
controller implements weight and absolute bandwidth limit models for
normal scheduling policy and absolute bandwidth allocation model for
realtime scheduling policy.


5-1-1. CPU Interface Files

All time durations are in microseconds.

  cpu.stat

	A read-only flat-keyed file which exists on non-root cgroups.

	It reports the following six stats.

	  usage_usec
	  user_usec
	  system_usec
	  nr_periods
	  nr_throttled
	  throttled_usec

  cpu.weight

	A read-write single value file which exists on non-root
	cgroups.  The default is "100".

	The weight in the range [1, 10000].

  cpu.max

	A read-write two value file which exists on non-root cgroups.
	The default is "max 100000".

	The maximum bandwidth limit.  It's in the following format.

	  $MAX $PERIOD

	which indicates that the group may consume upto $MAX in each
	$PERIOD duration.  "max" for $MAX indicates no limit.  If only
	one number is written, $MAX is updated.

  cpu.rt.max

  [NOTE: The semantics of this file is still under discussion and the
   interface hasn't been merged yet]

	A read-write two value file which exists on all cgroups.
	The default is "0 100000".

	The maximum realtime runtime allocation.  Over-committing
	configurations are disallowed and process migrations are
	rejected if not enough bandwidth is available.  It's in the
	following format.

	  $MAX $PERIOD

	which indicates that the group may consume upto $MAX in each
	$PERIOD duration.  If only one number is written, $MAX is
	updated.


5-2. Memory

The "memory" controller regulates distribution of memory.  Memory is
stateful and implements both limit and protection models.  Due to the
intertwining between memory usage and reclaim pressure and the
stateful nature of memory, the distribution model is relatively
complex.

While not completely water-tight, all major memory usages by a given
cgroup are tracked so that the total memory consumption can be
accounted and controlled to a reasonable extent.  Currently, the
following types of memory usages are tracked.

- Userland memory - page cache and anonymous memory.

- Kernel data structures such as dentries and inodes.

- TCP socket buffers.

The above list may expand in the future for better coverage.


5-2-1. Memory Interface Files

All memory amounts are in bytes.  If a value which is not aligned to
PAGE_SIZE is written, the value may be rounded up to the closest
PAGE_SIZE multiple when read back.

  memory.current

	A read-only single value file which exists on non-root
	cgroups.

	The total amount of memory currently being used by the cgroup
	and its descendants.

  memory.low

	A read-write single value file which exists on non-root
	cgroups.  The default is "0".

	Best-effort memory protection.  If the memory usages of a
	cgroup and all its ancestors are below their low boundaries,
	the cgroup's memory won't be reclaimed unless memory can be
	reclaimed from unprotected cgroups.

	Putting more memory than generally available under this
	protection is discouraged.

  memory.high

	A read-write single value file which exists on non-root
	cgroups.  The default is "max".

	Memory usage throttle limit.  This is the main mechanism to
	control memory usage of a cgroup.  If a cgroup's usage goes
	over the high boundary, the processes of the cgroup are
	throttled and put under heavy reclaim pressure.

	Going over the high limit never invokes the OOM killer and
	under extreme conditions the limit may be breached.

  memory.max

	A read-write single value file which exists on non-root
	cgroups.  The default is "max".

	Memory usage hard limit.  This is the final protection
	mechanism.  If a cgroup's memory usage reaches this limit and
	can't be reduced, the OOM killer is invoked in the cgroup.
	Under certain circumstances, the usage may go over the limit
	temporarily.

	This is the ultimate protection mechanism.  As long as the
	high limit is used and monitored properly, this limit's
	utility is limited to providing the final safety net.

  memory.events

	A read-only flat-keyed file which exists on non-root cgroups.
	The following entries are defined.  Unless specified
	otherwise, a value change in this file generates a file
	modified event.

	  low

		The number of times the cgroup is reclaimed due to
		high memory pressure even though its usage is under
		the low boundary.  This usually indicates that the low
		boundary is over-committed.

	  high

		The number of times processes of the cgroup are
		throttled and routed to perform direct memory reclaim
		because the high memory boundary was exceeded.  For a
		cgroup whose memory usage is capped by the high limit
		rather than global memory pressure, this event's
		occurrences are expected.

	  max

		The number of times the cgroup's memory usage was
		about to go over the max boundary.  If direct reclaim
		fails to bring it down, the OOM killer is invoked.

	  oom

		The number of times the OOM killer has been invoked in
		the cgroup.  This may not exactly match the number of
		processes killed but should generally be close.

  memory.stat

	A read-only flat-keyed file which exists on non-root cgroups.

	This breaks down the cgroup's memory footprint into different
	types of memory, type-specific details, and other information
	on the state and past events of the memory management system.

	All memory amounts are in bytes.

	The entries are ordered to be human readable, and new entries
	can show up in the middle. Don't rely on items remaining in a
	fixed position; use the keys to look up specific values!

	  anon

		Amount of memory used in anonymous mappings such as
		brk(), sbrk(), and mmap(MAP_ANONYMOUS)

	  file

		Amount of memory used to cache filesystem data,
		including tmpfs and shared memory.

	  kernel_stack

		Amount of memory allocated to kernel stacks.

	  slab

		Amount of memory used for storing in-kernel data
		structures.

	  sock

		Amount of memory used in network transmission buffers

	  shmem

		Amount of cached filesystem data that is swap-backed,
		such as tmpfs, shm segments, shared anonymous mmap()s

	  file_mapped

		Amount of cached filesystem data mapped with mmap()

	  file_dirty

		Amount of cached filesystem data that was modified but
		not yet written back to disk

	  file_writeback

		Amount of cached filesystem data that was modified and
		is currently being written back to disk

	  inactive_anon
	  active_anon
	  inactive_file
	  active_file
	  unevictable

		Amount of memory, swap-backed and filesystem-backed,
		on the internal memory management lists used by the
		page reclaim algorithm

	  slab_reclaimable

		Part of "slab" that might be reclaimed, such as
		dentries and inodes.

	  slab_unreclaimable

		Part of "slab" that cannot be reclaimed on memory
		pressure.

	  pgfault

		Total number of page faults incurred

	  pgmajfault

		Number of major page faults incurred

  memory.swap.current

	A read-only single value file which exists on non-root
	cgroups.

	The total amount of swap currently being used by the cgroup
	and its descendants.

  memory.swap.max

	A read-write single value file which exists on non-root
	cgroups.  The default is "max".

	Swap usage hard limit.  If a cgroup's swap usage reaches this
	limit, anonymous meomry of the cgroup will not be swapped out.


5-2-2. Usage Guidelines

"memory.high" is the main mechanism to control memory usage.
Over-committing on high limit (sum of high limits > available memory)
and letting global memory pressure to distribute memory according to
usage is a viable strategy.

Because breach of the high limit doesn't trigger the OOM killer but
throttles the offending cgroup, a management agent has ample
opportunities to monitor and take appropriate actions such as granting
more memory or terminating the workload.

Determining whether a cgroup has enough memory is not trivial as
memory usage doesn't indicate whether the workload can benefit from
more memory.  For example, a workload which writes data received from
network to a file can use all available memory but can also operate as
performant with a small amount of memory.  A measure of memory
pressure - how much the workload is being impacted due to lack of
memory - is necessary to determine whether a workload needs more
memory; unfortunately, memory pressure monitoring mechanism isn't
implemented yet.


5-2-3. Memory Ownership

A memory area is charged to the cgroup which instantiated it and stays
charged to the cgroup until the area is released.  Migrating a process
to a different cgroup doesn't move the memory usages that it
instantiated while in the previous cgroup to the new cgroup.

A memory area may be used by processes belonging to different cgroups.
To which cgroup the area will be charged is in-deterministic; however,
over time, the memory area is likely to end up in a cgroup which has
enough memory allowance to avoid high reclaim pressure.

If a cgroup sweeps a considerable amount of memory which is expected
to be accessed repeatedly by other cgroups, it may make sense to use
POSIX_FADV_DONTNEED to relinquish the ownership of memory areas
belonging to the affected files to ensure correct memory ownership.


5-3. IO

The "io" controller regulates the distribution of IO resources.  This
controller implements both weight based and absolute bandwidth or IOPS
limit distribution; however, weight based distribution is available
only if cfq-iosched is in use and neither scheme is available for
blk-mq devices.


5-3-1. IO Interface Files

  io.stat

	A read-only nested-keyed file which exists on non-root
	cgroups.

	Lines are keyed by $MAJ:$MIN device numbers and not ordered.
	The following nested keys are defined.

	  rbytes	Bytes read
	  wbytes	Bytes written
	  rios		Number of read IOs
	  wios		Number of write IOs

	An example read output follows.

	  8:16 rbytes=1459200 wbytes=314773504 rios=192 wios=353
	  8:0 rbytes=90430464 wbytes=299008000 rios=8950 wios=1252

  io.weight

	A read-write flat-keyed file which exists on non-root cgroups.
	The default is "default 100".

	The first line is the default weight applied to devices
	without specific override.  The rest are overrides keyed by
	$MAJ:$MIN device numbers and not ordered.  The weights are in
	the range [1, 10000] and specifies the relative amount IO time
	the cgroup can use in relation to its siblings.

	The default weight can be updated by writing either "default
	$WEIGHT" or simply "$WEIGHT".  Overrides can be set by writing
	"$MAJ:$MIN $WEIGHT" and unset by writing "$MAJ:$MIN default".

	An example read output follows.

	  default 100
	  8:16 200
	  8:0 50

  io.max

	A read-write nested-keyed file which exists on non-root
	cgroups.

	BPS and IOPS based IO limit.  Lines are keyed by $MAJ:$MIN
	device numbers and not ordered.  The following nested keys are
	defined.

	  rbps		Max read bytes per second
	  wbps		Max write bytes per second
	  riops		Max read IO operations per second
	  wiops		Max write IO operations per second

	When writing, any number of nested key-value pairs can be
	specified in any order.  "max" can be specified as the value
	to remove a specific limit.  If the same key is specified
	multiple times, the outcome is undefined.

	BPS and IOPS are measured in each IO direction and IOs are
	delayed if limit is reached.  Temporary bursts are allowed.

	Setting read limit at 2M BPS and write at 120 IOPS for 8:16.

	  echo "8:16 rbps=2097152 wiops=120" > io.max

	Reading returns the following.

	  8:16 rbps=2097152 wbps=max riops=max wiops=120

	Write IOPS limit can be removed by writing the following.

	  echo "8:16 wiops=max" > io.max

	Reading now returns the following.

	  8:16 rbps=2097152 wbps=max riops=max wiops=max


5-3-2. Writeback

Page cache is dirtied through buffered writes and shared mmaps and
written asynchronously to the backing filesystem by the writeback
mechanism.  Writeback sits between the memory and IO domains and
regulates the proportion of dirty memory by balancing dirtying and
write IOs.

The io controller, in conjunction with the memory controller,
implements control of page cache writeback IOs.  The memory controller
defines the memory domain that dirty memory ratio is calculated and
maintained for and the io controller defines the io domain which
writes out dirty pages for the memory domain.  Both system-wide and
per-cgroup dirty memory states are examined and the more restrictive
of the two is enforced.

cgroup writeback requires explicit support from the underlying
filesystem.  Currently, cgroup writeback is implemented on ext2, ext4
and btrfs.  On other filesystems, all writeback IOs are attributed to
the root cgroup.

There are inherent differences in memory and writeback management
which affects how cgroup ownership is tracked.  Memory is tracked per
page while writeback per inode.  For the purpose of writeback, an
inode is assigned to a cgroup and all IO requests to write dirty pages
from the inode are attributed to that cgroup.

As cgroup ownership for memory is tracked per page, there can be pages
which are associated with different cgroups than the one the inode is
associated with.  These are called foreign pages.  The writeback
constantly keeps track of foreign pages and, if a particular foreign
cgroup becomes the majority over a certain period of time, switches
the ownership of the inode to that cgroup.

While this model is enough for most use cases where a given inode is
mostly dirtied by a single cgroup even when the main writing cgroup
changes over time, use cases where multiple cgroups write to a single
inode simultaneously are not supported well.  In such circumstances, a
significant portion of IOs are likely to be attributed incorrectly.
As memory controller assigns page ownership on the first use and
doesn't update it until the page is released, even if writeback
strictly follows page ownership, multiple cgroups dirtying overlapping
areas wouldn't work as expected.  It's recommended to avoid such usage
patterns.

The sysctl knobs which affect writeback behavior are applied to cgroup
writeback as follows.

  vm.dirty_background_ratio
  vm.dirty_ratio

	These ratios apply the same to cgroup writeback with the
	amount of available memory capped by limits imposed by the
	memory controller and system-wide clean memory.

  vm.dirty_background_bytes
  vm.dirty_bytes

	For cgroup writeback, this is calculated into ratio against
	total available memory and applied the same way as
	vm.dirty[_background]_ratio.


5-4. PID

The process number controller is used to allow a cgroup to stop any
new tasks from being fork()'d or clone()'d after a specified limit is
reached.

The number of tasks in a cgroup can be exhausted in ways which other
controllers cannot prevent, thus warranting its own controller.  For
example, a fork bomb is likely to exhaust the number of tasks before
hitting memory restrictions.

Note that PIDs used in this controller refer to TIDs, process IDs as
used by the kernel.


5-4-1. PID Interface Files

  pids.max

	A read-write single value file which exists on non-root
	cgroups.  The default is "max".

	Hard limit of number of processes.

  pids.current

	A read-only single value file which exists on all cgroups.

	The number of processes currently in the cgroup and its
	descendants.

Organisational operations are not blocked by cgroup policies, so it is
possible to have pids.current > pids.max.  This can be done by either
setting the limit to be smaller than pids.current, or attaching enough
processes to the cgroup such that pids.current is larger than
pids.max.  However, it is not possible to violate a cgroup PID policy
through fork() or clone(). These will return -EAGAIN if the creation
of a new process would cause a cgroup policy to be violated.


5-5. RDMA

The "rdma" controller regulates the distribution and accounting of
of RDMA resources.

5-5-1. RDMA Interface Files

  rdma.max
	A readwrite nested-keyed file that exists for all the cgroups
	except root that describes current configured resource limit
	for a RDMA/IB device.

	Lines are keyed by device name and are not ordered.
	Each line contains space separated resource name and its configured
	limit that can be distributed.

	The following nested keys are defined.

	  hca_handle	Maximum number of HCA Handles
	  hca_object 	Maximum number of HCA Objects

	An example for mlx4 and ocrdma device follows.

	  mlx4_0 hca_handle=2 hca_object=2000
	  ocrdma1 hca_handle=3 hca_object=max

  rdma.current
	A read-only file that describes current resource usage.
	It exists for all the cgroup except root.

	An example for mlx4 and ocrdma device follows.

	  mlx4_0 hca_handle=1 hca_object=20
	  ocrdma1 hca_handle=1 hca_object=23


5-6. Misc

5-6-1. perf_event

perf_event controller, if not mounted on a legacy hierarchy, is
automatically enabled on the v2 hierarchy so that perf events can
always be filtered by cgroup v2 path.  The controller can still be
moved to a legacy hierarchy after v2 hierarchy is populated.


6. Namespace

6-1. Basics

cgroup namespace provides a mechanism to virtualize the view of the
"/proc/$PID/cgroup" file and cgroup mounts.  The CLONE_NEWCGROUP clone
flag can be used with clone(2) and unshare(2) to create a new cgroup
namespace.  The process running inside the cgroup namespace will have
its "/proc/$PID/cgroup" output restricted to cgroupns root.  The
cgroupns root is the cgroup of the process at the time of creation of
the cgroup namespace.

Without cgroup namespace, the "/proc/$PID/cgroup" file shows the
complete path of the cgroup of a process.  In a container setup where
a set of cgroups and namespaces are intended to isolate processes the
"/proc/$PID/cgroup" file may leak potential system level information
to the isolated processes.  For Example:

  # cat /proc/self/cgroup
  0::/batchjobs/container_id1

The path '/batchjobs/container_id1' can be considered as system-data
and undesirable to expose to the isolated processes.  cgroup namespace
can be used to restrict visibility of this path.  For example, before
creating a cgroup namespace, one would see:

  # ls -l /proc/self/ns/cgroup
  lrwxrwxrwx 1 root root 0 2014-07-15 10:37 /proc/self/ns/cgroup -> cgroup:[4026531835]
  # cat /proc/self/cgroup
  0::/batchjobs/container_id1

After unsharing a new namespace, the view changes.

  # ls -l /proc/self/ns/cgroup
  lrwxrwxrwx 1 root root 0 2014-07-15 10:35 /proc/self/ns/cgroup -> cgroup:[4026532183]
  # cat /proc/self/cgroup
  0::/

When some thread from a multi-threaded process unshares its cgroup
namespace, the new cgroupns gets applied to the entire process (all
the threads).  This is natural for the v2 hierarchy; however, for the
legacy hierarchies, this may be unexpected.

A cgroup namespace is alive as long as there are processes inside or
mounts pinning it.  When the last usage goes away, the cgroup
namespace is destroyed.  The cgroupns root and the actual cgroups
remain.


6-2. The Root and Views

The 'cgroupns root' for a cgroup namespace is the cgroup in which the
process calling unshare(2) is running.  For example, if a process in
/batchjobs/container_id1 cgroup calls unshare, cgroup
/batchjobs/container_id1 becomes the cgroupns root.  For the
init_cgroup_ns, this is the real root ('/') cgroup.

The cgroupns root cgroup does not change even if the namespace creator
process later moves to a different cgroup.

  # ~/unshare -c # unshare cgroupns in some cgroup
  # cat /proc/self/cgroup
  0::/
  # mkdir sub_cgrp_1
  # echo 0 > sub_cgrp_1/cgroup.procs
  # cat /proc/self/cgroup
  0::/sub_cgrp_1

Each process gets its namespace-specific view of "/proc/$PID/cgroup"

Processes running inside the cgroup namespace will be able to see
cgroup paths (in /proc/self/cgroup) only inside their root cgroup.
From within an unshared cgroupns:

  # sleep 100000 &
  [1] 7353
  # echo 7353 > sub_cgrp_1/cgroup.procs
  # cat /proc/7353/cgroup
  0::/sub_cgrp_1

From the initial cgroup namespace, the real cgroup path will be
visible:

  $ cat /proc/7353/cgroup
  0::/batchjobs/container_id1/sub_cgrp_1

From a sibling cgroup namespace (that is, a namespace rooted at a
different cgroup), the cgroup path relative to its own cgroup
namespace root will be shown.  For instance, if PID 7353's cgroup
namespace root is at '/batchjobs/container_id2', then it will see

  # cat /proc/7353/cgroup
  0::/../container_id2/sub_cgrp_1

Note that the relative path always starts with '/' to indicate that
its relative to the cgroup namespace root of the caller.


6-3. Migration and setns(2)

Processes inside a cgroup namespace can move into and out of the
namespace root if they have proper access to external cgroups.  For
example, from inside a namespace with cgroupns root at
/batchjobs/container_id1, and assuming that the global hierarchy is
still accessible inside cgroupns:

  # cat /proc/7353/cgroup
  0::/sub_cgrp_1
  # echo 7353 > batchjobs/container_id2/cgroup.procs
  # cat /proc/7353/cgroup
  0::/../container_id2

Note that this kind of setup is not encouraged.  A task inside cgroup
namespace should only be exposed to its own cgroupns hierarchy.

setns(2) to another cgroup namespace is allowed when:

(a) the process has CAP_SYS_ADMIN against its current user namespace
(b) the process has CAP_SYS_ADMIN against the target cgroup
    namespace's userns

No implicit cgroup changes happen with attaching to another cgroup
namespace.  It is expected that the someone moves the attaching
process under the target cgroup namespace root.


6-4. Interaction with Other Namespaces

Namespace specific cgroup hierarchy can be mounted by a process
running inside a non-init cgroup namespace.

  # mount -t cgroup2 none $MOUNT_POINT

This will mount the unified cgroup hierarchy with cgroupns root as the
filesystem root.  The process needs CAP_SYS_ADMIN against its user and
mount namespaces.

The virtualization of /proc/self/cgroup file combined with restricting
the view of cgroup hierarchy by namespace-private cgroupfs mount
provides a properly isolated cgroup view inside the container.


P. Information on Kernel Programming

This section contains kernel programming information in the areas
where interacting with cgroup is necessary.  cgroup core and
controllers are not covered.


P-1. Filesystem Support for Writeback

A filesystem can support cgroup writeback by updating
address_space_operations->writepage[s]() to annotate bio's using the
following two functions.

  wbc_init_bio(@wbc, @bio)

	Should be called for each bio carrying writeback data and
	associates the bio with the inode's owner cgroup.  Can be
	called anytime between bio allocation and submission.

  wbc_account_io(@wbc, @page, @bytes)

	Should be called for each data segment being written out.
	While this function doesn't care exactly when it's called
	during the writeback session, it's the easiest and most
	natural to call it as data segments are added to a bio.

With writeback bio's annotated, cgroup support can be enabled per
super_block by setting SB_I_CGROUPWB in ->s_iflags.  This allows for
selective disabling of cgroup writeback support which is helpful when
certain filesystem features, e.g. journaled data mode, are
incompatible.

wbc_init_bio() binds the specified bio to its cgroup.  Depending on
the configuration, the bio may be executed at a lower priority and if
the writeback session is holding shared resources, e.g. a journal
entry, may lead to priority inversion.  There is no one easy solution
for the problem.  Filesystems can try to work around specific problem
cases by skipping wbc_init_bio() or using bio_associate_blkcg()
directly.


D. Deprecated v1 Core Features

- Multiple hierarchies including named ones are not supported.

- All mount options and remounting are not supported.

- The "tasks" file is removed and "cgroup.procs" is not sorted.

- "cgroup.clone_children" is removed.

- /proc/cgroups is meaningless for v2.  Use "cgroup.controllers" file
  at the root instead.


R. Issues with v1 and Rationales for v2

R-1. Multiple Hierarchies

cgroup v1 allowed an arbitrary number of hierarchies and each
hierarchy could host any number of controllers.  While this seemed to
provide a high level of flexibility, it wasn't useful in practice.

For example, as there is only one instance of each controller, utility
type controllers such as freezer which can be useful in all
hierarchies could only be used in one.  The issue is exacerbated by
the fact that controllers couldn't be moved to another hierarchy once
hierarchies were populated.  Another issue was that all controllers
bound to a hierarchy were forced to have exactly the same view of the
hierarchy.  It wasn't possible to vary the granularity depending on
the specific controller.

In practice, these issues heavily limited which controllers could be
put on the same hierarchy and most configurations resorted to putting
each controller on its own hierarchy.  Only closely related ones, such
as the cpu and cpuacct controllers, made sense to be put on the same
hierarchy.  This often meant that userland ended up managing multiple
similar hierarchies repeating the same steps on each hierarchy
whenever a hierarchy management operation was necessary.

Furthermore, support for multiple hierarchies came at a steep cost.
It greatly complicated cgroup core implementation but more importantly
the support for multiple hierarchies restricted how cgroup could be
used in general and what controllers was able to do.

There was no limit on how many hierarchies there might be, which meant
that a thread's cgroup membership couldn't be described in finite
length.  The key might contain any number of entries and was unlimited
in length, which made it highly awkward to manipulate and led to
addition of controllers which existed only to identify membership,
which in turn exacerbated the original problem of proliferating number
of hierarchies.

Also, as a controller couldn't have any expectation regarding the
topologies of hierarchies other controllers might be on, each
controller had to assume that all other controllers were attached to
completely orthogonal hierarchies.  This made it impossible, or at
least very cumbersome, for controllers to cooperate with each other.

In most use cases, putting controllers on hierarchies which are
completely orthogonal to each other isn't necessary.  What usually is
called for is the ability to have differing levels of granularity
depending on the specific controller.  In other words, hierarchy may
be collapsed from leaf towards root when viewed from specific
controllers.  For example, a given configuration might not care about
how memory is distributed beyond a certain level while still wanting
to control how CPU cycles are distributed.


R-2. Thread Granularity

cgroup v1 allowed threads of a process to belong to different cgroups.
This didn't make sense for some controllers and those controllers
ended up implementing different ways to ignore such situations but
much more importantly it blurred the line between API exposed to
individual applications and system management interface.

Generally, in-process knowledge is available only to the process
itself; thus, unlike service-level organization of processes,
categorizing threads of a process requires active participation from
the application which owns the target process.

cgroup v1 had an ambiguously defined delegation model which got abused
in combination with thread granularity.  cgroups were delegated to
individual applications so that they can create and manage their own
sub-hierarchies and control resource distributions along them.  This
effectively raised cgroup to the status of a syscall-like API exposed
to lay programs.

First of all, cgroup has a fundamentally inadequate interface to be
exposed this way.  For a process to access its own knobs, it has to
extract the path on the target hierarchy from /proc/self/cgroup,
construct the path by appending the name of the knob to the path, open
and then read and/or write to it.  This is not only extremely clunky
and unusual but also inherently racy.  There is no conventional way to
define transaction across the required steps and nothing can guarantee
that the process would actually be operating on its own sub-hierarchy.

cgroup controllers implemented a number of knobs which would never be
accepted as public APIs because they were just adding control knobs to
system-management pseudo filesystem.  cgroup ended up with interface
knobs which were not properly abstracted or refined and directly
revealed kernel internal details.  These knobs got exposed to
individual applications through the ill-defined delegation mechanism
effectively abusing cgroup as a shortcut to implementing public APIs
without going through the required scrutiny.

This was painful for both userland and kernel.  Userland ended up with
misbehaving and poorly abstracted interfaces and kernel exposing and
locked into constructs inadvertently.


R-3. Competition Between Inner Nodes and Threads

cgroup v1 allowed threads to be in any cgroups which created an
interesting problem where threads belonging to a parent cgroup and its
children cgroups competed for resources.  This was nasty as two
different types of entities competed and there was no obvious way to
settle it.  Different controllers did different things.

The cpu controller considered threads and cgroups as equivalents and
mapped nice levels to cgroup weights.  This worked for some cases but
fell flat when children wanted to be allocated specific ratios of CPU
cycles and the number of internal threads fluctuated - the ratios
constantly changed as the number of competing entities fluctuated.
There also were other issues.  The mapping from nice level to weight
wasn't obvious or universal, and there were various other knobs which
simply weren't available for threads.

The io controller implicitly created a hidden leaf node for each
cgroup to host the threads.  The hidden leaf had its own copies of all
the knobs with "leaf_" prefixed.  While this allowed equivalent
control over internal threads, it was with serious drawbacks.  It
always added an extra layer of nesting which wouldn't be necessary
otherwise, made the interface messy and significantly complicated the
implementation.

The memory controller didn't have a way to control what happened
between internal tasks and child cgroups and the behavior was not
clearly defined.  There were attempts to add ad-hoc behaviors and
knobs to tailor the behavior to specific workloads which would have
led to problems extremely difficult to resolve in the long term.

Multiple controllers struggled with internal tasks and came up with
different ways to deal with it; unfortunately, all the approaches were
severely flawed and, furthermore, the widely different behaviors
made cgroup as a whole highly inconsistent.

This clearly is a problem which needs to be addressed from cgroup core
in a uniform way.


R-4. Other Interface Issues

cgroup v1 grew without oversight and developed a large number of
idiosyncrasies and inconsistencies.  One issue on the cgroup core side
was how an empty cgroup was notified - a userland helper binary was
forked and executed for each event.  The event delivery wasn't
recursive or delegatable.  The limitations of the mechanism also led
to in-kernel event delivery filtering mechanism further complicating
the interface.

Controller interfaces were problematic too.  An extreme example is
controllers completely ignoring hierarchical organization and treating
all cgroups as if they were all located directly under the root
cgroup.  Some controllers exposed a large amount of inconsistent
implementation details to userland.

There also was no consistency across controllers.  When a new cgroup
was created, some controllers defaulted to not imposing extra
restrictions while others disallowed any resource usage until
explicitly configured.  Configuration knobs for the same type of
control used widely differing naming schemes and formats.  Statistics
and information knobs were named arbitrarily and used different
formats and units even in the same controller.

cgroup v2 establishes common conventions where appropriate and updates
controllers so that they expose minimal and consistent interfaces.


R-5. Controller Issues and Remedies

R-5-1. Memory

The original lower boundary, the soft limit, is defined as a limit
that is per default unset.  As a result, the set of cgroups that
global reclaim prefers is opt-in, rather than opt-out.  The costs for
optimizing these mostly negative lookups are so high that the
implementation, despite its enormous size, does not even provide the
basic desirable behavior.  First off, the soft limit has no
hierarchical meaning.  All configured groups are organized in a global
rbtree and treated like equal peers, regardless where they are located
in the hierarchy.  This makes subtree delegation impossible.  Second,
the soft limit reclaim pass is so aggressive that it not just
introduces high allocation latencies into the system, but also impacts
system performance due to overreclaim, to the point where the feature
becomes self-defeating.

The memory.low boundary on the other hand is a top-down allocated
reserve.  A cgroup enjoys reclaim protection when it and all its
ancestors are below their low boundaries, which makes delegation of
subtrees possible.  Secondly, new cgroups have no reserve per default
and in the common case most cgroups are eligible for the preferred
reclaim pass.  This allows the new low boundary to be efficiently
implemented with just a minor addition to the generic reclaim code,
without the need for out-of-band data structures and reclaim passes.
Because the generic reclaim code considers all cgroups except for the
ones running low in the preferred first reclaim pass, overreclaim of
individual groups is eliminated as well, resulting in much better
overall workload performance.

The original high boundary, the hard limit, is defined as a strict
limit that can not budge, even if the OOM killer has to be called.
But this generally goes against the goal of making the most out of the
available memory.  The memory consumption of workloads varies during
runtime, and that requires users to overcommit.  But doing that with a
strict upper limit requires either a fairly accurate prediction of the
working set size or adding slack to the limit.  Since working set size
estimation is hard and error prone, and getting it wrong results in
OOM kills, most users tend to err on the side of a looser limit and
end up wasting precious resources.

The memory.high boundary on the other hand can be set much more
conservatively.  When hit, it throttles allocations by forcing them
into direct reclaim to work off the excess, but it never invokes the
OOM killer.  As a result, a high boundary that is chosen too
aggressively will not terminate the processes, but instead it will
lead to gradual performance degradation.  The user can monitor this
and make corrections until the minimal memory footprint that still
gives acceptable performance is found.

In extreme cases, with many concurrent allocations and a complete
breakdown of reclaim progress within the group, the high boundary can
be exceeded.  But even then it's mostly better to satisfy the
allocation from the slack available in other groups or the rest of the
system than killing the group.  Otherwise, memory.max is there to
limit this type of spillover and ultimately contain buggy or even
malicious applications.

Setting the original memory.limit_in_bytes below the current usage was
subject to a race condition, where concurrent charges could cause the
limit setting to fail. memory.max on the other hand will first set the
limit to prevent new charges, and then reclaim and OOM kill until the
new limit is met - or the task writing to memory.max is killed.

The combined memory+swap accounting and limiting is replaced by real
control over swap space.

The main argument for a combined memory+swap facility in the original
cgroup design was that global or parental pressure would always be
able to swap all anonymous memory of a child group, regardless of the
child's own (possibly untrusted) configuration.  However, untrusted
groups can sabotage swapping by other means - such as referencing its
anonymous memory in a tight loop - and an admin can not assume full
swappability when overcommitting untrusted jobs.

For trusted jobs, on the other hand, a combined counter is not an
intuitive userspace interface, and it flies in the face of the idea
that cgroup controllers should account and limit specific physical
resources.  Swap space is a resource like all others in the system,
and that's why unified hierarchy allows distributing it separately.