CGCONFIG.CONF(5) CGCONFIG.CONF(5)

NAME cgconfig.conf - libcgroup configuration file

DESCRIPTION cgconfig.conf is a configuration file used by libcgroup to define con- trol groups, their parameters and their mount points. The file con- sists of mount , group and default sections. These sections can be in arbitrary order and all of them are optional. Any line starting with # is considered a comment line and is ignored.

mount section has this form:

mount { <controller> = <path>; ... }

controller Name of the kernel subsystem. The list of subsystems supported by the kernel can be found in /proc/cgroups file. Named hierar- chy can be specified as controller "name=<somename>". Do not forget to use double quotes around this controller name (see examples below).

Libcgroup merges all subsystems mounted to the same directory (see Example 1) and the directory is mounted only once.

path The directory path where the group hierarchy associated to a given controller shall be mounted. The directory is created automatically on cgconfig service startup if it does not exist and is deleted on service shutdown.

If no mount section is specified, no controllers are mounted.

group section has this form:

group <name> { [permissions] <controller> { <param name> = <param value>; ... } ... }

name Name of the control group. It can contain only characters, which are allowed for directory names. The groups form a tree, i.e. a control group can contain zero or more subgroups. Subgroups can be specified using / delimiter.

The root control group is always created automatically in all hierarchies and it is the base of the group hierarchy. It can be explicitly specified in cgconfig.conf by using . as group name. This can be used e.g. to set its permissions, as shown in Example 6.

When the parent control group of a subgroup is not specified it is created automatically.

permissions Permissions of the given control group on mounted filesystem. root has always permission to do anything with the control group. Permissions have the following syntax: perm { task { uid = <task user>; gid = <task group>; fperm = <file permissions> } admin { uid = <admin name>; gid = <admin group>; dperm = <directory permissions> fperm = <file permissions> } }

task user/group Name of the user and the group, which own the tasks file of the control group. Given fperm then specify the file permissions. Please note that the given value is not used as was speci- fied. Instead, current file owner permissions are used as a "umask" for group and others per- missions. For example if fperm = 777 then both group and others will get the same permissions as the file owner.

admin user/group Name of the user and the group which own the rest of control group s files. Given fperm and dperm control file and directory permissions. Again, the given value is masked by the file/directory owner permissions.

Permissions are only apply to the enclosing control group and are not inherited by subgroups. If there is no perm section in the control group definition, root:root is the owner of all files and default file permissions are preserved if fperm resp. dperm are not specified.

controller Name of the kernel subsystem. The section can be empty, default kernel parameters will be used in this case. By specifying con- troller the control group and all its parents are controlled by the specific subsystem. One control group can be controlled by multiple subsystems, even if the subsystems are mounted on dif- ferent directories. Each control group must be controlled by at least one subsystem, so that libcgroup knows in which hierar- chies the control group should be created.

The parameters of the given controller can be modified in the following section enclosed in brackets.

param name Name of the file to set. Each controller can have zero or more parameters.

param value Value which should be written to the file when the con- trol group is created. If it is enclosed in double quotes " , it can contain spaces and other special characters.

If no group section is specified, no groups are created.

default section has this form:

default { perm { task { uid = <task user>; gid = <task group>; fperm = <file permissions> } admin { uid = <admin name>; gid = <admin group>; dperm = <directory permissions> fperm = <file permissions> } } }

Content of the perm section has the same form as in group section. The permissions defined here specify owner and permissions of groups and files of all groups, which do not have explicitly specified their per- missions in their group section.

template section has the same structure as group section. Template name uses the same templates string as cgrules.conf destination tag (see (cgrules.conf (5)). Template definition is used as a control group definition for rules in cgrules.conf (5) with the same destination name. Templates does not use default section settings.

EXAMPLES Example 1 The configuration file:

mount { cpu = /mnt/cgroups/cpu; cpuacct = /mnt/cgroups/cpu; }

creates the hierarchy controlled by two subsystems with no groups inside. It corresponds to the following operations:

mkdir /mnt/cgroups/cpu mount -t cgroup -o cpu,cpuacct cpu /mnt/cgroups/cpu

Example 2 The configuration file:

mount { cpu = /mnt/cgroups/cpu; "name=scheduler" = /mnt/cgroups/cpu; "name=noctrl" = /mnt/cgroups/noctrl; }

group daemons { cpu { cpu.shares = "1000"; } } group test { "name=noctrl" { } } creates two hierarchies. One hierarchy named scheduler controlled by cpu subsystem, with group daemons inside. Second hierarchy is named noctrl without any controller, with group test. It corresponds to fol- lowing operations:

mkdir /mnt/cgroups/cpu mount -t cgroup -o cpu,name=scheduler cpu /mnt/cgroups/cpu mount -t cgroup -o none,name=noctrl none /mnt/cgroups/noctrl

mkdir /mnt/cgroups/cpu/daemons echo 1000 > /mnt/cgroups/cpu/daemons/www/cpu.shares

mkdir /mnt/cgroups/noctrl/tests

The daemons group is created automatically when its first subgroup is created. All its parameters have the default value and only root can access groups files.

Since both cpuacct and cpu subsystems are mounted to the same direc- tory, all groups are implicitly controlled also by cpuacct subsystem, even if there is no cpuacct section in any of the groups.

Example 3 The configuration file:

mount { cpu = /mnt/cgroups/cpu; cpuacct = /mnt/cgroups/cpu; }

group daemons/www { perm { task { uid = root; gid = webmaster; fperm = 770; } admin { uid = root; gid = root; dperm = 775; fperm = 744; } } cpu { cpu.shares = "1000"; } }

group daemons/ftp { perm { task { uid = root; gid = ftpmaster; fperm = 774; } admin { uid = root; gid = root; dperm = 755; fperm = 700; } } cpu { cpu.shares = "500"; } } creates the hierarchy controlled by two subsystems with one group and two subgroups inside, setting one parameter. It corresponds to the following operations (except for file permissions which are little bit trickier to emulate via chmod):

mkdir /mnt/cgroups/cpu mount -t cgroup -o cpu,cpuacct cpu /mnt/cgroups/cpu

mkdir /mnt/cgroups/cpu/daemons

mkdir /mnt/cgroups/cpu/daemons/www chown root:root /mnt/cgroups/cpu/daemons/www/* chown root:webmaster /mnt/cgroups/cpu/daemons/www/tasks echo 1000 > /mnt/cgroups/cpu/daemons/www/cpu.shares

# + chmod the files so the result looks like: # ls -la /mnt/cgroups/cpu/daemons/www/ # admin.dperm = 755: # drwxr-xr-x. 2 root webmaster 0 Jun 16 11:51 . # # admin.fperm = 744: # --w-------. 1 root webmaster 0 Jun 16 11:51 cgroup.event_control # -r--r--r--. 1 root webmaster 0 Jun 16 11:51 cgroup.procs # -r--r--r--. 1 root webmaster 0 Jun 16 11:51 cpuacct.stat # -rw-r--r--. 1 root webmaster 0 Jun 16 11:51 cpuacct.usage # -r--r--r--. 1 root webmaster 0 Jun 16 11:51 cpuacct.usage_percpu # -rw-r--r--. 1 root webmaster 0 Jun 16 11:51 cpu.rt_period_us # -rw-r--r--. 1 root webmaster 0 Jun 16 11:51 cpu.rt_runtime_us # -rw-r--r--. 1 root webmaster 0 Jun 16 11:51 cpu.shares # -rw-r--r--. 1 root webmaster 0 Jun 16 11:51 notify_on_release # # tasks.fperm = 770 # -rw-rw----. 1 root webmaster 0 Jun 16 11:51 tasks

mkdir /mnt/cgroups/cpu/daemons/ftp chown root:root /mnt/cgroups/cpu/daemons/ftp/* chown root:ftpmaster /mnt/cgroups/cpu/daemons/ftp/tasks echo 500 > /mnt/cgroups/cpu/daemons/ftp/cpu.shares

# + chmod the files so the result looks like: # ls -la /mnt/cgroups/cpu/daemons/ftp/ # admin.dperm = 755: # drwxr-xr-x. 2 root ftpmaster 0 Jun 16 11:51 . # # admin.fperm = 700: # --w-------. 1 root ftpmaster 0 Jun 16 11:51 cgroup.event_control # -r--------. 1 root ftpmaster 0 Jun 16 11:51 cgroup.procs # -r--------. 1 root ftpmaster 0 Jun 16 11:51 cpuacct.stat # -rw-------. 1 root ftpmaster 0 Jun 16 11:51 cpuacct.usage # -r--------. 1 root ftpmaster 0 Jun 16 11:51 cpuacct.usage_percpu # -rw-------. 1 root ftpmaster 0 Jun 16 11:51 cpu.rt_period_us # -rw-------. 1 root ftpmaster 0 Jun 16 11:51 cpu.rt_runtime_us # -rw-------. 1 root ftpmaster 0 Jun 16 11:51 cpu.shares # -rw-------. 1 root ftpmaster 0 Jun 16 11:51 notify_on_release # # tasks.fperm = 774: # -rw-rw-r--. 1 root ftpmaster 0 Jun 16 11:51 tasks

The daemons group is created automatically when its first subgroup is created. All its parameters have the default value and only root can access the groups files.

Since both cpuacct and cpu subsystems are mounted to the same direc- tory, all groups are implicitly also controlled by the cpuacct subsys- tem, even if there is no cpuacct section in any of the groups.

Example 4 The configuration file:

mount { cpu = /mnt/cgroups/cpu; cpuacct = /mnt/cgroups/cpuacct; }

group daemons { cpuacct{ } cpu { } } creates two hierarchies and one common group in both of them. It cor- responds to the following operations:

mkdir /mnt/cgroups/cpu mkdir /mnt/cgroups/cpuacct mount -t cgroup -o cpu cpu /mnt/cgroups/cpu mount -t cgroup -o cpuacct cpuacct /mnt/cgroups/cpuacct

mkdir /mnt/cgroups/cpu/daemons mkdir /mnt/cgroups/cpuacct/daemons

In fact there are two groups created. One in the cpuacct hierarchy, the second in the cpu hierarchy. These two groups have nothing in common and can contain different subgroups and different tasks.

Example 5 The configuration file:

mount { cpu = /mnt/cgroups/cpu; cpuacct = /mnt/cgroups/cpuacct; }

group daemons { cpuacct{ } }

group daemons/www { cpu { cpu.shares = "1000"; } }

group daemons/ftp { cpu { cpu.shares = "500"; } } creates two hierarchies with few groups inside. One of the groups is created in both hierarchies.

It corresponds to the following operations:

mkdir /mnt/cgroups/cpu mkdir /mnt/cgroups/cpuacct mount -t cgroup -o cpu cpu /mnt/cgroups/cpu mount -t cgroup -o cpuacct cpuacct /mnt/cgroups/cpuacct

mkdir /mnt/cgroups/cpuacct/daemons mkdir /mnt/cgroups/cpu/daemons mkdir /mnt/cgroups/cpu/daemons/www echo 1000 > /mnt/cgroups/cpu/daemons/www/cpu.shares mkdir /mnt/cgroups/cpu/daemons/ftp echo 500 > /mnt/cgroups/cpu/daemons/ftp/cpu.shares Group daemons is created in both hierarchies. In the cpuacct hierarchy the group is explicitly mentioned in the configuration file. In the cpu hierarchy the group is created implicitly when www is created there. These two groups have nothing in common, for example they do not share processes and subgroups. Groups www and ftp are created only in the cpu hierarchy and are not controlled by the cpuacct subsystem.

Example 6 The configuration file:

mount { cpu = /mnt/cgroups/cpu; cpuacct = /mnt/cgroups/cpu; }

group . { perm { task { uid = root; gid = operator; } admin { uid = root; gid = operator; } } cpu { } }

group daemons { perm { task { uid = root; gid = daemonmaster; } admin { uid = root; gid = operator; } } cpu { } } creates the hierarchy controlled by two subsystems with one group hav- ing some special permissions. It corresponds to the following opera- tions:

mkdir /mnt/cgroups/cpu mount -t cgroup -o cpu,cpuacct cpu /mnt/cgroups/cpu

chown root:operator /mnt/cgroups/cpu/* chown root:operator /mnt/cgroups/cpu/tasks

mkdir /mnt/cgroups/cpu/daemons chown root:operator /mnt/cgroups/cpu/daemons/* chown root:daemonmaster /mnt/cgroups/cpu/daemons/tasks

Users which are members of the operator group are allowed to administer the control groups, i.e. create new control groups and move processes between these groups without having root privileges.

Members of the daemonmaster group can move processes to the daemons control group, but they can not move the process out of the group. Only the operator or root can do that.

Example 7 The configuration file:

mount { cpu = /mnt/cgroups/cpu; cpuacct = /mnt/cgroups/cpuacct; }

group students { cpuacct{ } cpu { } }

template students/%u { cpuacct{ } cpu { } }

mkdir /mnt/cgroups/cpu/daemons mkdir /mnt/cgroups/cpuacct/daemons

The situation is the similar as in Example 4. The only difference is template, which is used if some rule uses "/students/%u" as a destina- tion.

RECOMMENDATIONS Keep hierarchies separated Having multiple hierarchies is perfectly valid and can be useful in various scenarios. To keeps things clean, do not create one group in multiple hierarchies. Examples 4 and 5 show how unreadable and confus- ing it can be, especially when reading somebody elses configuration file.

Explicit is better than implicit libcgroup can implicitly create groups which are needed for the cre- ation of configured subgroups. This may be useful and save some typing in simple scenarios. When it comes to multiple hierarchies, its better to explicitly specify all groups and all controllers related to them.

FILES /etc/cgconfig.conf default libcgroup configuration file

SEE ALSO cgconfigparser (8)

BUGS Parameter values must be single strings without spaces. Parsing of quoted strings is not implemented.

CGCONFIG.CONF(5)