Cgroup unified hierarchy April, 2014 Tejun Heo This document describes the changes made by unified hierarchy and their rationales. It will eventually be merged into the main cgroup documentation. CONTENTS 1. Background 2. Basic Operation 2-1. Mounting 2-2. cgroup.subtree_control 2-3. cgroup.controllers 3. Structural Constraints 3-1. Top-down 3-2. No internal tasks 4. Delegation 4-1. Model of delegation 4-2. Common ancestor rule 5. Other Changes 5-1. [Un]populated Notification 5-2. Other Core Changes 5-3. Controller File Conventions 5-3-1. Format 5-3-2. Control Knobs 5-4. Per-Controller Changes 5-4-1. io 5-4-2. cpuset 5-4-3. memory 6. Planned Changes 6-1. CAP for resource control 1. Background cgroup allows an arbitrary number of hierarchies and each hierarchy can host any number of controllers. While this seems to provide a high level of flexibility, it isn't quite 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 can only be used in one. The issue is exacerbated by the fact that controllers can't be moved around once hierarchies are populated. Another issue is that all controllers bound to a hierarchy are forced to have exactly the same view of the hierarchy. It isn't possible to vary the granularity depending on the specific controller. In practice, these issues heavily limit which controllers can be put on the same hierarchy and most configurations resort to putting each controller on its own hierarchy. Only closely related ones, such as the cpu and cpuacct controllers, make sense to put on the same hierarchy. This often means that userland ends up managing multiple similar hierarchies repeating the same steps on each hierarchy whenever a hierarchy management operation is necessary. Unfortunately, support for multiple hierarchies comes at a steep cost. Internal implementation in cgroup core proper is dazzlingly complicated but more importantly the support for multiple hierarchies restricts how cgroup is used in general and what controllers can do. There's no limit on how many hierarchies there may be, which means that a task's cgroup membership can't be described in finite length. The key may contain any varying number of entries and is unlimited in length, which makes it highly awkward to handle and leads to addition of controllers which exist only to identify membership, which in turn exacerbates the original problem. Also, as a controller can't have any expectation regarding what shape of hierarchies other controllers would be on, each controller has to assume that all other controllers are operating on completely orthogonal hierarchies. This makes 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. Unified hierarchy is the next version of cgroup interface. It aims to address the aforementioned issues by having more structure while retaining enough flexibility for most use cases. Various other general and controller-specific interface issues are also addressed in the process. 2. Basic Operation 2-1. Mounting Currently, unified hierarchy can be mounted with the following mount command. Note that this is still under development and scheduled to change soon. mount -t cgroup -o __DEVEL__sane_behavior cgroup $MOUNT_POINT All controllers which support the unified hierarchy and are not bound to other hierarchies are automatically bound to unified hierarchy and show up at the root of it. Controllers which are enabled only in the root of unified hierarchy can be bound to other hierarchies. This allows mixing unified hierarchy with the traditional multiple hierarchies in a fully backward compatible way. For development purposes, the following boot parameter makes all controllers to appear on the unified hierarchy whether supported or not. cgroup__DEVEL__legacy_files_on_dfl 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 unified 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 dependencies among controllers, other controllers may need to be disabled too. While useful for development and manual configurations, dynamically moving controllers between the unified and other hierarchies is strongly discouraged for production use. It is recommended to decide the hierarchies and controller associations before starting using the controllers. 2-2. cgroup.subtree_control All cgroups on unified hierarchy have a "cgroup.subtree_control" file which governs which controllers are enabled on the children of the cgroup. Let's assume a hierarchy like the following. root - A - B - C \ D root's "cgroup.subtree_control" file determines which controllers are enabled on A. A's on B. B's on C and D. This coincides with the fact that controllers on the immediate sub-level are used to distribute the resources of the parent. In fact, it's natural to assume that resource control knobs of a child belong to its parent. Enabling a controller in a "cgroup.subtree_control" file declares that distribution of the respective resources of the cgroup will be controlled. Note that this means that controller enable states are shared among siblings. When read, the file contains a space-separated list of currently enabled controllers. A write to the file should contain a space-separated list of controllers with '+' or '-' prefixed (without the quotes). Controllers prefixed with '+' are enabled and '-' disabled. If a controller is listed multiple times, the last entry wins. The specific operations are executed atomically - either all succeed or fail. 2-3. cgroup.controllers Read-only "cgroup.controllers" file contains a space-separated list of controllers which can be enabled in the cgroup's "cgroup.subtree_control" file. In the root cgroup, this lists controllers which are not bound to other hierarchies and the content changes as controllers are bound to and unbound from other hierarchies. In non-root cgroups, the content of this file equals that of the parent's "cgroup.subtree_control" file as only controllers enabled from the parent can be used in its children. 3. Structural Constraints 3-1. Top-down As it doesn't make sense to nest control of an uncontrolled resource, 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. 3-2. No internal tasks One long-standing issue that cgroup faces is the competition between tasks belonging to the parent cgroup and its children cgroups. This is inherently nasty as two different types of entities compete and there is no agreed-upon obvious way to handle it. Different controllers are doing different things. The cpu controller considers tasks and cgroups as equivalents and maps nice levels to cgroup weights. This works for some cases but falls flat when children should be allocated specific ratios of CPU cycles and the number of internal tasks fluctuates - the ratios constantly change as the number of competing entities fluctuates. There also are other issues. The mapping from nice level to weight isn't obvious or universal, and there are various other knobs which simply aren't available for tasks. The io controller implicitly creates a hidden leaf node for each cgroup to host the tasks. The hidden leaf has its own copies of all the knobs with "leaf_" prefixed. While this allows equivalent control over internal tasks, it's with serious drawbacks. It always adds an extra layer of nesting which may not be necessary, makes the interface messy and significantly complicates the implementation. The memory controller currently doesn't have a way to control what happens between internal tasks and child cgroups and the behavior is not clearly defined. There have been attempts to add ad-hoc behaviors and knobs to tailor the behavior to specific workloads. Continuing this direction will lead to problems which will be extremely difficult to resolve in the long term. Multiple controllers struggle with internal tasks and came up with different ways to deal with it; unfortunately, all the approaches in use now are severely flawed and, furthermore, the widely different behaviors make cgroup as whole highly inconsistent. It is clear that this is something which needs to be addressed from cgroup core proper in a uniform way so that controllers don't need to worry about it and cgroup as a whole shows a consistent and logical behavior. To achieve that, unified hierarchy enforces the following structural constraint: Except for the root, only cgroups which don't contain any task may have controllers enabled in their "cgroup.subtree_control" files. Combined with other properties, this guarantees that, when a controller is looking at the part of the hierarchy which has it enabled, tasks are always only on the leaves. This rules out situations where child cgroups compete against internal tasks of the parent. There are two things to note. Firstly, the root cgroup is exempt from the restriction. Root contains tasks and anonymous resource consumption which can't be associated with any other cgroup and requires special treatment from most controllers. How resource consumption in the root cgroup is governed is up to each controller. Secondly, the restriction doesn't take effect if there is no enabled controller in the cgroup's "cgroup.subtree_control" file. 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 tasks to the children before enabling controllers in its "cgroup.subtree_control" file. 4. Delegation 4-1. 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 the resource control knobs in a given directory concern the resources of the parent 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 got 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 in the future be limited explicitly. 4-2. Common ancestor rule On the unified hierarchy, to write to a "cgroup.procs" file, in addition to the usual write permission to the file and uid match, the writer must also have write access to the "cgroup.procs" file of the common ancestor of the source and destination cgroups. This prevents delegatees from smuggling processes across disjoint sub-hierarchies. Let's say cgroups C0 and C1 have been delegated to user U0 who created C00, C01 under C0 and C10 under C1 as follows. ~~~~~~~~~~~~~ - C0 - C00 ~ cgroup ~ \ C01 ~ hierarchy ~ ~~~~~~~~~~~~~ - C1 - C10 C0 and C1 are separate entities in terms of resource distribution regardless of their relative positions in the hierarchy. The resources the processes under C0 are entitled to are controlled by C0's ancestors and may be completely different from C1. It's clear that the intention of delegating C0 to U0 is allowing U0 to organize the processes under C0 and further control the distribution of C0's resources. On traditional hierarchies, if a task has write access to "tasks" or "cgroup.procs" file of a cgroup and its uid agrees with the target, it can move the target to the cgroup. In the above example, U0 will not only be able to move processes in each sub-hierarchy but also across the two sub-hierarchies, effectively allowing it to violate the organizational and resource restrictions implied by the hierarchical structure above C0 and C1. On the unified hierarchy, let's say U0 wants to write the pid of a process which has a matching uid and is currently in C10 into "C00/cgroup.procs". U0 obviously has write access to the file and migration permission on the process; 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" and thus be denied with -EACCES. 5. Other Changes 5-1. [Un]populated Notification cgroup users often need a way to determine when a cgroup's subhierarchy becomes empty so that it can be cleaned up. cgroup currently provides release_agent for it; unfortunately, this mechanism is riddled with issues. - It delivers events by forking and execing a userland binary specified as the release_agent. This is a long deprecated method of notification delivery. It's extremely heavy, slow and cumbersome to integrate with larger infrastructure. - There is single monitoring point at the root. There's no way to delegate management of a subtree. - The event isn't recursive. It triggers when a cgroup doesn't have any tasks or child cgroups. Events for internal nodes trigger only after all children are removed. This again makes it impossible to delegate management of a subtree. - Events are filtered from the kernel side. A "notify_on_release" file is used to subscribe to or suppress release events. This is unnecessarily complicated and probably done this way because event delivery itself was expensive. Unified hierarchy implements an interface file "cgroup.populated" which can be used to monitor whether the cgroup's subhierarchy has tasks in it or not. Its value is 0 if there is no task in the cgroup and its descendants; otherwise, 1. poll and [id]notify events are triggered when the value changes. This is significantly lighter and simpler and trivially allows delegating management of subhierarchy - subhierarchy monitoring can block further propagation simply by putting itself or another process in the subhierarchy and monitor events that it's interested in from there without interfering with monitoring higher in the tree. In unified hierarchy, the release_agent mechanism is no longer supported and the interface files "release_agent" and "notify_on_release" do not exist. 5-2. Other Core Changes - None of the mount options is allowed. - remount is disallowed. - rename(2) is disallowed. - The "tasks" file is removed. Everything should at process granularity. Use the "cgroup.procs" file instead. - The "cgroup.procs" file is not sorted. pids will be unique unless they got recycled in-between reads. - The "cgroup.clone_children" file is removed. 5-3. Controller File Conventions 5-3-1. Format In general, all controller files should be in one of the following formats whenever possible. - Values only files VAL0 VAL1...\n - Flat keyed files KEY0 VAL0\n KEY1 VAL1\n ... - Nested keyed files KEY0 SUB_KEY0=VAL00 SUB_KEY1=VAL01... KEY1 SUB_KEY0=VAL10 SUB_KEY1=VAL11... ... For a writeable 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. 5-3-2. Control Knobs - Settings for a single feature should generally be implemented in a single file. - In general, the root cgroup should be exempt from resource control and thus shouldn't have resource control knobs. - If a controller implements ratio based resource distribution, the control knob should be named "weight" and have the range [1, 10000] and 100 should be the default value. 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 control knobs should be named "min" and "max" respectively. If a controller implements best effort resource gurantee and/or limit, the control knobs 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 configurable default value and specific overrides, the default settings should be keyed with "default" and appear as the first entry in the file. Specific entries can use "default" as its value to indicate inheritance of the default value. 5-4. Per-Controller Changes 5-4-1. io - blkio is renamed to io. The interface is overhauled anyway. The new name is more in line with the other two major controllers, cpu and memory, and better suited given that it may be used for cgroup writeback without involving block layer. - Everything including stat is always hierarchical making separate recursive stat files pointless and, as no internal node can have tasks, leaf weights are meaningless. The operation model is simplified and the interface is overhauled accordingly. io.stat The stat file. The reported stats are from the point where bio's are issued to request_queue. The stats are counted independent of which policies are enabled. Each line in the file follows the following format. More fields may later be added at the end. $MAJ:$MIN rbytes=$RBYTES wbytes=$WBYTES rios=$RIOS wrios=$WIOS io.weight The weight setting, currently only available and effective if cfq-iosched is in use for the target device. The weight is between 1 and 10000 and defaults to 100. The first line always contains the default weight in the following format to use when per-device setting is missing. default $WEIGHT Subsequent lines list per-device weights of the following format. $MAJ:$MIN $WEIGHT Writing "$WEIGHT" or "default $WEIGHT" changes the default setting. Writing "$MAJ:$MIN $WEIGHT" sets per-device weight while "$MAJ:$MIN default" clears it. This file is available only on non-root cgroups. io.max The maximum bandwidth and/or iops setting, only available if blk-throttle is enabled. The file is of the following format. $MAJ:$MIN rbps=$RBPS wbps=$WBPS riops=$RIOPS wiops=$WIOPS ${R|W}BPS are read/write bytes per second and ${R|W}IOPS are read/write IOs per second. "max" indicates no limit. Writing to the file follows the same format but the individual settings may be omitted or specified in any order. This file is available only on non-root cgroups. 5-4-2. cpuset - Tasks are kept in empty cpusets after hotplug and take on the masks of the nearest non-empty ancestor, instead of being moved to it. - A task can be moved into an empty cpuset, and again it takes on the masks of the nearest non-empty ancestor. 5-4-3. memory - use_hierarchy is on by default and the cgroup file for the flag is not created. - 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. - The original control file names are unwieldy and inconsistent in many different ways. For example, the upper boundary hit count is exported in the memory.failcnt file, but an OOM event count has to be manually counted by listening to memory.oom_control events, and lower boundary / soft limit events have to be counted by first setting a threshold for that value and then counting those events. Also, usage and limit files encode their units in the filename. That makes the filenames very long, even though this is not information that a user needs to be reminded of every time they type out those names. To address these naming issues, as well as to signal clearly that the new interface carries a new configuration model, the naming conventions in it necessarily differ from the old interface. - The original limit files indicate the state of an unset limit with a Very High Number, and a configured limit can be unset by echoing -1 into those files. But that very high number is implementation and architecture dependent and not very descriptive. And while -1 can be understood as an underflow into the highest possible value, -2 or -10M etc. do not work, so it's not consistent. memory.low, memory.high, and memory.max will use the string "max" to indicate and set the highest possible value. 6. Planned Changes 6-1. CAP for resource control Unified hierarchy will require one of the capabilities(7), which is yet to be decided, for all resource control related knobs. Process organization operations - creation of sub-cgroups and migration of processes in sub-hierarchies may be delegated by changing the ownership and/or permissions on the cgroup directory and "cgroup.procs" interface file; however, all operations which affect resource control - writes to a "cgroup.subtree_control" file or any controller-specific knobs - will require an explicit CAP privilege. This, in part, is to prevent the cgroup interface from being inadvertently promoted to programmable API used by non-privileged binaries. cgroup exposes various aspects of the system in ways which aren't properly abstracted for direct consumption by regular programs. This is an administration interface much closer to sysctl knobs than system calls. Even the basic access model, being filesystem path based, isn't suitable for direct consumption. There's no way to access "my cgroup" in a race-free way or make multiple operations atomic against migration to another cgroup. Another aspect is that, for better or for worse, the cgroup interface goes through far less scrutiny than regular interfaces for unprivileged userland. The upside is that cgroup is able to expose useful features which may not be suitable for general consumption in a reasonable time frame. It provides a relatively short path between internal details and userland-visible interface. Of course, this shortcut comes with high risk. We go through what we go through for general kernel APIs for good reasons. It may end up leaking internal details in a way which can exert significant pain by locking the kernel into a contract that can't be maintained in a reasonable manner. Also, due to the specific nature, cgroup and its controllers don't tend to attract attention from a wide scope of developers. cgroup's short history is already fraught with severely mis-designed interfaces, unnecessary commitments to and exposing of internal details, broken and dangerous implementations of various features. Keeping cgroup as an administration interface is both advantageous for its role and imperative given its nature. Some of the cgroup features may make sense for unprivileged access. If deemed justified, those must be further abstracted and implemented as a different interface, be it a system call or process-private filesystem, and survive through the scrutiny that any interface for general consumption is required to go through. Requiring CAP is not a complete solution but should serve as a significant deterrent against spraying cgroup usages in non-privileged programs.