1================ 2Control Group v2 3================ 4 5:Date: October, 2015 6:Author: Tejun Heo <tj@kernel.org> 7 8This is the authoritative documentation on the design, interface and 9conventions of cgroup v2. It describes all userland-visible aspects 10of cgroup including core and specific controller behaviors. All 11future changes must be reflected in this document. Documentation for 12v1 is available under :ref:`Documentation/admin-guide/cgroup-v1/index.rst <cgroup-v1>`. 13 14.. CONTENTS 15 16 1. Introduction 17 1-1. Terminology 18 1-2. What is cgroup? 19 2. Basic Operations 20 2-1. Mounting 21 2-2. Organizing Processes and Threads 22 2-2-1. Processes 23 2-2-2. Threads 24 2-3. [Un]populated Notification 25 2-4. Controlling Controllers 26 2-4-1. Enabling and Disabling 27 2-4-2. Top-down Constraint 28 2-4-3. No Internal Process Constraint 29 2-5. Delegation 30 2-5-1. Model of Delegation 31 2-5-2. Delegation Containment 32 2-6. Guidelines 33 2-6-1. Organize Once and Control 34 2-6-2. Avoid Name Collisions 35 3. Resource Distribution Models 36 3-1. Weights 37 3-2. Limits 38 3-3. Protections 39 3-4. Allocations 40 4. Interface Files 41 4-1. Format 42 4-2. Conventions 43 4-3. Core Interface Files 44 5. Controllers 45 5-1. CPU 46 5-1-1. CPU Interface Files 47 5-2. Memory 48 5-2-1. Memory Interface Files 49 5-2-2. Usage Guidelines 50 5-2-3. Memory Ownership 51 5-3. IO 52 5-3-1. IO Interface Files 53 5-3-2. Writeback 54 5-3-3. IO Latency 55 5-3-3-1. How IO Latency Throttling Works 56 5-3-3-2. IO Latency Interface Files 57 5-3-4. IO Priority 58 5-4. PID 59 5-4-1. PID Interface Files 60 5-5. Cpuset 61 5.5-1. Cpuset Interface Files 62 5-6. Device 63 5-7. RDMA 64 5-7-1. RDMA Interface Files 65 5-8. HugeTLB 66 5.8-1. HugeTLB Interface Files 67 5-8. Misc 68 5-8-1. perf_event 69 5-N. Non-normative information 70 5-N-1. CPU controller root cgroup process behaviour 71 5-N-2. IO controller root cgroup process behaviour 72 6. Namespace 73 6-1. Basics 74 6-2. The Root and Views 75 6-3. Migration and setns(2) 76 6-4. Interaction with Other Namespaces 77 P. Information on Kernel Programming 78 P-1. Filesystem Support for Writeback 79 D. Deprecated v1 Core Features 80 R. Issues with v1 and Rationales for v2 81 R-1. Multiple Hierarchies 82 R-2. Thread Granularity 83 R-3. Competition Between Inner Nodes and Threads 84 R-4. Other Interface Issues 85 R-5. Controller Issues and Remedies 86 R-5-1. Memory 87 88 89Introduction 90============ 91 92Terminology 93----------- 94 95"cgroup" stands for "control group" and is never capitalized. The 96singular form is used to designate the whole feature and also as a 97qualifier as in "cgroup controllers". When explicitly referring to 98multiple individual control groups, the plural form "cgroups" is used. 99 100 101What is cgroup? 102--------------- 103 104cgroup is a mechanism to organize processes hierarchically and 105distribute system resources along the hierarchy in a controlled and 106configurable manner. 107 108cgroup is largely composed of two parts - the core and controllers. 109cgroup core is primarily responsible for hierarchically organizing 110processes. A cgroup controller is usually responsible for 111distributing a specific type of system resource along the hierarchy 112although there are utility controllers which serve purposes other than 113resource distribution. 114 115cgroups form a tree structure and every process in the system belongs 116to one and only one cgroup. All threads of a process belong to the 117same cgroup. On creation, all processes are put in the cgroup that 118the parent process belongs to at the time. A process can be migrated 119to another cgroup. Migration of a process doesn't affect already 120existing descendant processes. 121 122Following certain structural constraints, controllers may be enabled or 123disabled selectively on a cgroup. All controller behaviors are 124hierarchical - if a controller is enabled on a cgroup, it affects all 125processes which belong to the cgroups consisting the inclusive 126sub-hierarchy of the cgroup. When a controller is enabled on a nested 127cgroup, it always restricts the resource distribution further. The 128restrictions set closer to the root in the hierarchy can not be 129overridden from further away. 130 131 132Basic Operations 133================ 134 135Mounting 136-------- 137 138Unlike v1, cgroup v2 has only single hierarchy. The cgroup v2 139hierarchy can be mounted with the following mount command:: 140 141 # mount -t cgroup2 none $MOUNT_POINT 142 143cgroup2 filesystem has the magic number 0x63677270 ("cgrp"). All 144controllers which support v2 and are not bound to a v1 hierarchy are 145automatically bound to the v2 hierarchy and show up at the root. 146Controllers which are not in active use in the v2 hierarchy can be 147bound to other hierarchies. This allows mixing v2 hierarchy with the 148legacy v1 multiple hierarchies in a fully backward compatible way. 149 150A controller can be moved across hierarchies only after the controller 151is no longer referenced in its current hierarchy. Because per-cgroup 152controller states are destroyed asynchronously and controllers may 153have lingering references, a controller may not show up immediately on 154the v2 hierarchy after the final umount of the previous hierarchy. 155Similarly, a controller should be fully disabled to be moved out of 156the unified hierarchy and it may take some time for the disabled 157controller to become available for other hierarchies; furthermore, due 158to inter-controller dependencies, other controllers may need to be 159disabled too. 160 161While useful for development and manual configurations, moving 162controllers dynamically between the v2 and other hierarchies is 163strongly discouraged for production use. It is recommended to decide 164the hierarchies and controller associations before starting using the 165controllers after system boot. 166 167During transition to v2, system management software might still 168automount the v1 cgroup filesystem and so hijack all controllers 169during boot, before manual intervention is possible. To make testing 170and experimenting easier, the kernel parameter cgroup_no_v1= allows 171disabling controllers in v1 and make them always available in v2. 172 173cgroup v2 currently supports the following mount options. 174 175 nsdelegate 176 177 Consider cgroup namespaces as delegation boundaries. This 178 option is system wide and can only be set on mount or modified 179 through remount from the init namespace. The mount option is 180 ignored on non-init namespace mounts. Please refer to the 181 Delegation section for details. 182 183 memory_localevents 184 185 Only populate memory.events with data for the current cgroup, 186 and not any subtrees. This is legacy behaviour, the default 187 behaviour without this option is to include subtree counts. 188 This option is system wide and can only be set on mount or 189 modified through remount from the init namespace. The mount 190 option is ignored on non-init namespace mounts. 191 192 memory_recursiveprot 193 194 Recursively apply memory.min and memory.low protection to 195 entire subtrees, without requiring explicit downward 196 propagation into leaf cgroups. This allows protecting entire 197 subtrees from one another, while retaining free competition 198 within those subtrees. This should have been the default 199 behavior but is a mount-option to avoid regressing setups 200 relying on the original semantics (e.g. specifying bogusly 201 high 'bypass' protection values at higher tree levels). 202 203 204Organizing Processes and Threads 205-------------------------------- 206 207Processes 208~~~~~~~~~ 209 210Initially, only the root cgroup exists to which all processes belong. 211A child cgroup can be created by creating a sub-directory:: 212 213 # mkdir $CGROUP_NAME 214 215A given cgroup may have multiple child cgroups forming a tree 216structure. Each cgroup has a read-writable interface file 217"cgroup.procs". When read, it lists the PIDs of all processes which 218belong to the cgroup one-per-line. The PIDs are not ordered and the 219same PID may show up more than once if the process got moved to 220another cgroup and then back or the PID got recycled while reading. 221 222A process can be migrated into a cgroup by writing its PID to the 223target cgroup's "cgroup.procs" file. Only one process can be migrated 224on a single write(2) call. If a process is composed of multiple 225threads, writing the PID of any thread migrates all threads of the 226process. 227 228When a process forks a child process, the new process is born into the 229cgroup that the forking process belongs to at the time of the 230operation. After exit, a process stays associated with the cgroup 231that it belonged to at the time of exit until it's reaped; however, a 232zombie process does not appear in "cgroup.procs" and thus can't be 233moved to another cgroup. 234 235A cgroup which doesn't have any children or live processes can be 236destroyed by removing the directory. Note that a cgroup which doesn't 237have any children and is associated only with zombie processes is 238considered empty and can be removed:: 239 240 # rmdir $CGROUP_NAME 241 242"/proc/$PID/cgroup" lists a process's cgroup membership. If legacy 243cgroup is in use in the system, this file may contain multiple lines, 244one for each hierarchy. The entry for cgroup v2 is always in the 245format "0::$PATH":: 246 247 # cat /proc/842/cgroup 248 ... 249 0::/test-cgroup/test-cgroup-nested 250 251If the process becomes a zombie and the cgroup it was associated with 252is removed subsequently, " (deleted)" is appended to the path:: 253 254 # cat /proc/842/cgroup 255 ... 256 0::/test-cgroup/test-cgroup-nested (deleted) 257 258 259Threads 260~~~~~~~ 261 262cgroup v2 supports thread granularity for a subset of controllers to 263support use cases requiring hierarchical resource distribution across 264the threads of a group of processes. By default, all threads of a 265process belong to the same cgroup, which also serves as the resource 266domain to host resource consumptions which are not specific to a 267process or thread. The thread mode allows threads to be spread across 268a subtree while still maintaining the common resource domain for them. 269 270Controllers which support thread mode are called threaded controllers. 271The ones which don't are called domain controllers. 272 273Marking a cgroup threaded makes it join the resource domain of its 274parent as a threaded cgroup. The parent may be another threaded 275cgroup whose resource domain is further up in the hierarchy. The root 276of a threaded subtree, that is, the nearest ancestor which is not 277threaded, is called threaded domain or thread root interchangeably and 278serves as the resource domain for the entire subtree. 279 280Inside a threaded subtree, threads of a process can be put in 281different cgroups and are not subject to the no internal process 282constraint - threaded controllers can be enabled on non-leaf cgroups 283whether they have threads in them or not. 284 285As the threaded domain cgroup hosts all the domain resource 286consumptions of the subtree, it is considered to have internal 287resource consumptions whether there are processes in it or not and 288can't have populated child cgroups which aren't threaded. Because the 289root cgroup is not subject to no internal process constraint, it can 290serve both as a threaded domain and a parent to domain cgroups. 291 292The current operation mode or type of the cgroup is shown in the 293"cgroup.type" file which indicates whether the cgroup is a normal 294domain, a domain which is serving as the domain of a threaded subtree, 295or a threaded cgroup. 296 297On creation, a cgroup is always a domain cgroup and can be made 298threaded by writing "threaded" to the "cgroup.type" file. The 299operation is single direction:: 300 301 # echo threaded > cgroup.type 302 303Once threaded, the cgroup can't be made a domain again. To enable the 304thread mode, the following conditions must be met. 305 306- As the cgroup will join the parent's resource domain. The parent 307 must either be a valid (threaded) domain or a threaded cgroup. 308 309- When the parent is an unthreaded domain, it must not have any domain 310 controllers enabled or populated domain children. The root is 311 exempt from this requirement. 312 313Topology-wise, a cgroup can be in an invalid state. Please consider 314the following topology:: 315 316 A (threaded domain) - B (threaded) - C (domain, just created) 317 318C is created as a domain but isn't connected to a parent which can 319host child domains. C can't be used until it is turned into a 320threaded cgroup. "cgroup.type" file will report "domain (invalid)" in 321these cases. Operations which fail due to invalid topology use 322EOPNOTSUPP as the errno. 323 324A domain cgroup is turned into a threaded domain when one of its child 325cgroup becomes threaded or threaded controllers are enabled in the 326"cgroup.subtree_control" file while there are processes in the cgroup. 327A threaded domain reverts to a normal domain when the conditions 328clear. 329 330When read, "cgroup.threads" contains the list of the thread IDs of all 331threads in the cgroup. Except that the operations are per-thread 332instead of per-process, "cgroup.threads" has the same format and 333behaves the same way as "cgroup.procs". While "cgroup.threads" can be 334written to in any cgroup, as it can only move threads inside the same 335threaded domain, its operations are confined inside each threaded 336subtree. 337 338The threaded domain cgroup serves as the resource domain for the whole 339subtree, and, while the threads can be scattered across the subtree, 340all the processes are considered to be in the threaded domain cgroup. 341"cgroup.procs" in a threaded domain cgroup contains the PIDs of all 342processes in the subtree and is not readable in the subtree proper. 343However, "cgroup.procs" can be written to from anywhere in the subtree 344to migrate all threads of the matching process to the cgroup. 345 346Only threaded controllers can be enabled in a threaded subtree. When 347a threaded controller is enabled inside a threaded subtree, it only 348accounts for and controls resource consumptions associated with the 349threads in the cgroup and its descendants. All consumptions which 350aren't tied to a specific thread belong to the threaded domain cgroup. 351 352Because a threaded subtree is exempt from no internal process 353constraint, a threaded controller must be able to handle competition 354between threads in a non-leaf cgroup and its child cgroups. Each 355threaded controller defines how such competitions are handled. 356 357 358[Un]populated Notification 359-------------------------- 360 361Each non-root cgroup has a "cgroup.events" file which contains 362"populated" field indicating whether the cgroup's sub-hierarchy has 363live processes in it. Its value is 0 if there is no live process in 364the cgroup and its descendants; otherwise, 1. poll and [id]notify 365events are triggered when the value changes. This can be used, for 366example, to start a clean-up operation after all processes of a given 367sub-hierarchy have exited. The populated state updates and 368notifications are recursive. Consider the following sub-hierarchy 369where the numbers in the parentheses represent the numbers of processes 370in each cgroup:: 371 372 A(4) - B(0) - C(1) 373 \ D(0) 374 375A, B and C's "populated" fields would be 1 while D's 0. After the one 376process in C exits, B and C's "populated" fields would flip to "0" and 377file modified events will be generated on the "cgroup.events" files of 378both cgroups. 379 380 381Controlling Controllers 382----------------------- 383 384Enabling and Disabling 385~~~~~~~~~~~~~~~~~~~~~~ 386 387Each cgroup has a "cgroup.controllers" file which lists all 388controllers available for the cgroup to enable:: 389 390 # cat cgroup.controllers 391 cpu io memory 392 393No controller is enabled by default. Controllers can be enabled and 394disabled by writing to the "cgroup.subtree_control" file:: 395 396 # echo "+cpu +memory -io" > cgroup.subtree_control 397 398Only controllers which are listed in "cgroup.controllers" can be 399enabled. When multiple operations are specified as above, either they 400all succeed or fail. If multiple operations on the same controller 401are specified, the last one is effective. 402 403Enabling a controller in a cgroup indicates that the distribution of 404the target resource across its immediate children will be controlled. 405Consider the following sub-hierarchy. The enabled controllers are 406listed in parentheses:: 407 408 A(cpu,memory) - B(memory) - C() 409 \ D() 410 411As A has "cpu" and "memory" enabled, A will control the distribution 412of CPU cycles and memory to its children, in this case, B. As B has 413"memory" enabled but not "CPU", C and D will compete freely on CPU 414cycles but their division of memory available to B will be controlled. 415 416As a controller regulates the distribution of the target resource to 417the cgroup's children, enabling it creates the controller's interface 418files in the child cgroups. In the above example, enabling "cpu" on B 419would create the "cpu." prefixed controller interface files in C and 420D. Likewise, disabling "memory" from B would remove the "memory." 421prefixed controller interface files from C and D. This means that the 422controller interface files - anything which doesn't start with 423"cgroup." are owned by the parent rather than the cgroup itself. 424 425 426Top-down Constraint 427~~~~~~~~~~~~~~~~~~~ 428 429Resources are distributed top-down and a cgroup can further distribute 430a resource only if the resource has been distributed to it from the 431parent. This means that all non-root "cgroup.subtree_control" files 432can only contain controllers which are enabled in the parent's 433"cgroup.subtree_control" file. A controller can be enabled only if 434the parent has the controller enabled and a controller can't be 435disabled if one or more children have it enabled. 436 437 438No Internal Process Constraint 439~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~ 440 441Non-root cgroups can distribute domain resources to their children 442only when they don't have any processes of their own. In other words, 443only domain cgroups which don't contain any processes can have domain 444controllers enabled in their "cgroup.subtree_control" files. 445 446This guarantees that, when a domain controller is looking at the part 447of the hierarchy which has it enabled, processes are always only on 448the leaves. This rules out situations where child cgroups compete 449against internal processes of the parent. 450 451The root cgroup is exempt from this restriction. Root contains 452processes and anonymous resource consumption which can't be associated 453with any other cgroups and requires special treatment from most 454controllers. How resource consumption in the root cgroup is governed 455is up to each controller (for more information on this topic please 456refer to the Non-normative information section in the Controllers 457chapter). 458 459Note that the restriction doesn't get in the way if there is no 460enabled controller in the cgroup's "cgroup.subtree_control". This is 461important as otherwise it wouldn't be possible to create children of a 462populated cgroup. To control resource distribution of a cgroup, the 463cgroup must create children and transfer all its processes to the 464children before enabling controllers in its "cgroup.subtree_control" 465file. 466 467 468Delegation 469---------- 470 471Model of Delegation 472~~~~~~~~~~~~~~~~~~~ 473 474A cgroup can be delegated in two ways. First, to a less privileged 475user by granting write access of the directory and its "cgroup.procs", 476"cgroup.threads" and "cgroup.subtree_control" files to the user. 477Second, if the "nsdelegate" mount option is set, automatically to a 478cgroup namespace on namespace creation. 479 480Because the resource control interface files in a given directory 481control the distribution of the parent's resources, the delegatee 482shouldn't be allowed to write to them. For the first method, this is 483achieved by not granting access to these files. For the second, the 484kernel rejects writes to all files other than "cgroup.procs" and 485"cgroup.subtree_control" on a namespace root from inside the 486namespace. 487 488The end results are equivalent for both delegation types. Once 489delegated, the user can build sub-hierarchy under the directory, 490organize processes inside it as it sees fit and further distribute the 491resources it received from the parent. The limits and other settings 492of all resource controllers are hierarchical and regardless of what 493happens in the delegated sub-hierarchy, nothing can escape the 494resource restrictions imposed by the parent. 495 496Currently, cgroup doesn't impose any restrictions on the number of 497cgroups in or nesting depth of a delegated sub-hierarchy; however, 498this may be limited explicitly in the future. 499 500 501Delegation Containment 502~~~~~~~~~~~~~~~~~~~~~~ 503 504A delegated sub-hierarchy is contained in the sense that processes 505can't be moved into or out of the sub-hierarchy by the delegatee. 506 507For delegations to a less privileged user, this is achieved by 508requiring the following conditions for a process with a non-root euid 509to migrate a target process into a cgroup by writing its PID to the 510"cgroup.procs" file. 511 512- The writer must have write access to the "cgroup.procs" file. 513 514- The writer must have write access to the "cgroup.procs" file of the 515 common ancestor of the source and destination cgroups. 516 517The above two constraints ensure that while a delegatee may migrate 518processes around freely in the delegated sub-hierarchy it can't pull 519in from or push out to outside the sub-hierarchy. 520 521For an example, let's assume cgroups C0 and C1 have been delegated to 522user U0 who created C00, C01 under C0 and C10 under C1 as follows and 523all processes under C0 and C1 belong to U0:: 524 525 ~~~~~~~~~~~~~ - C0 - C00 526 ~ cgroup ~ \ C01 527 ~ hierarchy ~ 528 ~~~~~~~~~~~~~ - C1 - C10 529 530Let's also say U0 wants to write the PID of a process which is 531currently in C10 into "C00/cgroup.procs". U0 has write access to the 532file; however, the common ancestor of the source cgroup C10 and the 533destination cgroup C00 is above the points of delegation and U0 would 534not have write access to its "cgroup.procs" files and thus the write 535will be denied with -EACCES. 536 537For delegations to namespaces, containment is achieved by requiring 538that both the source and destination cgroups are reachable from the 539namespace of the process which is attempting the migration. If either 540is not reachable, the migration is rejected with -ENOENT. 541 542 543Guidelines 544---------- 545 546Organize Once and Control 547~~~~~~~~~~~~~~~~~~~~~~~~~ 548 549Migrating a process across cgroups is a relatively expensive operation 550and stateful resources such as memory are not moved together with the 551process. This is an explicit design decision as there often exist 552inherent trade-offs between migration and various hot paths in terms 553of synchronization cost. 554 555As such, migrating processes across cgroups frequently as a means to 556apply different resource restrictions is discouraged. A workload 557should be assigned to a cgroup according to the system's logical and 558resource structure once on start-up. Dynamic adjustments to resource 559distribution can be made by changing controller configuration through 560the interface files. 561 562 563Avoid Name Collisions 564~~~~~~~~~~~~~~~~~~~~~ 565 566Interface files for a cgroup and its children cgroups occupy the same 567directory and it is possible to create children cgroups which collide 568with interface files. 569 570All cgroup core interface files are prefixed with "cgroup." and each 571controller's interface files are prefixed with the controller name and 572a dot. A controller's name is composed of lower case alphabets and 573'_'s but never begins with an '_' so it can be used as the prefix 574character for collision avoidance. Also, interface file names won't 575start or end with terms which are often used in categorizing workloads 576such as job, service, slice, unit or workload. 577 578cgroup doesn't do anything to prevent name collisions and it's the 579user's responsibility to avoid them. 580 581 582Resource Distribution Models 583============================ 584 585cgroup controllers implement several resource distribution schemes 586depending on the resource type and expected use cases. This section 587describes major schemes in use along with their expected behaviors. 588 589 590Weights 591------- 592 593A parent's resource is distributed by adding up the weights of all 594active children and giving each the fraction matching the ratio of its 595weight against the sum. As only children which can make use of the 596resource at the moment participate in the distribution, this is 597work-conserving. Due to the dynamic nature, this model is usually 598used for stateless resources. 599 600All weights are in the range [1, 10000] with the default at 100. This 601allows symmetric multiplicative biases in both directions at fine 602enough granularity while staying in the intuitive range. 603 604As long as the weight is in range, all configuration combinations are 605valid and there is no reason to reject configuration changes or 606process migrations. 607 608"cpu.weight" proportionally distributes CPU cycles to active children 609and is an example of this type. 610 611 612Limits 613------ 614 615A child can only consume upto the configured amount of the resource. 616Limits can be over-committed - the sum of the limits of children can 617exceed the amount of resource available to the parent. 618 619Limits are in the range [0, max] and defaults to "max", which is noop. 620 621As limits can be over-committed, all configuration combinations are 622valid and there is no reason to reject configuration changes or 623process migrations. 624 625"io.max" limits the maximum BPS and/or IOPS that a cgroup can consume 626on an IO device and is an example of this type. 627 628 629Protections 630----------- 631 632A cgroup is protected upto the configured amount of the resource 633as long as the usages of all its ancestors are under their 634protected levels. Protections can be hard guarantees or best effort 635soft boundaries. Protections can also be over-committed in which case 636only upto the amount available to the parent is protected among 637children. 638 639Protections are in the range [0, max] and defaults to 0, which is 640noop. 641 642As protections can be over-committed, all configuration combinations 643are valid and there is no reason to reject configuration changes or 644process migrations. 645 646"memory.low" implements best-effort memory protection and is an 647example of this type. 648 649 650Allocations 651----------- 652 653A cgroup is exclusively allocated a certain amount of a finite 654resource. Allocations can't be over-committed - the sum of the 655allocations of children can not exceed the amount of resource 656available to the parent. 657 658Allocations are in the range [0, max] and defaults to 0, which is no 659resource. 660 661As allocations can't be over-committed, some configuration 662combinations are invalid and should be rejected. Also, if the 663resource is mandatory for execution of processes, process migrations 664may be rejected. 665 666"cpu.rt.max" hard-allocates realtime slices and is an example of this 667type. 668 669 670Interface Files 671=============== 672 673Format 674------ 675 676All interface files should be in one of the following formats whenever 677possible:: 678 679 New-line separated values 680 (when only one value can be written at once) 681 682 VAL0\n 683 VAL1\n 684 ... 685 686 Space separated values 687 (when read-only or multiple values can be written at once) 688 689 VAL0 VAL1 ...\n 690 691 Flat keyed 692 693 KEY0 VAL0\n 694 KEY1 VAL1\n 695 ... 696 697 Nested keyed 698 699 KEY0 SUB_KEY0=VAL00 SUB_KEY1=VAL01... 700 KEY1 SUB_KEY0=VAL10 SUB_KEY1=VAL11... 701 ... 702 703For a writable file, the format for writing should generally match 704reading; however, controllers may allow omitting later fields or 705implement restricted shortcuts for most common use cases. 706 707For both flat and nested keyed files, only the values for a single key 708can be written at a time. For nested keyed files, the sub key pairs 709may be specified in any order and not all pairs have to be specified. 710 711 712Conventions 713----------- 714 715- Settings for a single feature should be contained in a single file. 716 717- The root cgroup should be exempt from resource control and thus 718 shouldn't have resource control interface files. 719 720- The default time unit is microseconds. If a different unit is ever 721 used, an explicit unit suffix must be present. 722 723- A parts-per quantity should use a percentage decimal with at least 724 two digit fractional part - e.g. 13.40. 725 726- If a controller implements weight based resource distribution, its 727 interface file should be named "weight" and have the range [1, 728 10000] with 100 as the default. The values are chosen to allow 729 enough and symmetric bias in both directions while keeping it 730 intuitive (the default is 100%). 731 732- If a controller implements an absolute resource guarantee and/or 733 limit, the interface files should be named "min" and "max" 734 respectively. If a controller implements best effort resource 735 guarantee and/or limit, the interface files should be named "low" 736 and "high" respectively. 737 738 In the above four control files, the special token "max" should be 739 used to represent upward infinity for both reading and writing. 740 741- If a setting has a configurable default value and keyed specific 742 overrides, the default entry should be keyed with "default" and 743 appear as the first entry in the file. 744 745 The default value can be updated by writing either "default $VAL" or 746 "$VAL". 747 748 When writing to update a specific override, "default" can be used as 749 the value to indicate removal of the override. Override entries 750 with "default" as the value must not appear when read. 751 752 For example, a setting which is keyed by major:minor device numbers 753 with integer values may look like the following:: 754 755 # cat cgroup-example-interface-file 756 default 150 757 8:0 300 758 759 The default value can be updated by:: 760 761 # echo 125 > cgroup-example-interface-file 762 763 or:: 764 765 # echo "default 125" > cgroup-example-interface-file 766 767 An override can be set by:: 768 769 # echo "8:16 170" > cgroup-example-interface-file 770 771 and cleared by:: 772 773 # echo "8:0 default" > cgroup-example-interface-file 774 # cat cgroup-example-interface-file 775 default 125 776 8:16 170 777 778- For events which are not very high frequency, an interface file 779 "events" should be created which lists event key value pairs. 780 Whenever a notifiable event happens, file modified event should be 781 generated on the file. 782 783 784Core Interface Files 785-------------------- 786 787All cgroup core files are prefixed with "cgroup." 788 789 cgroup.type 790 791 A read-write single value file which exists on non-root 792 cgroups. 793 794 When read, it indicates the current type of the cgroup, which 795 can be one of the following values. 796 797 - "domain" : A normal valid domain cgroup. 798 799 - "domain threaded" : A threaded domain cgroup which is 800 serving as the root of a threaded subtree. 801 802 - "domain invalid" : A cgroup which is in an invalid state. 803 It can't be populated or have controllers enabled. It may 804 be allowed to become a threaded cgroup. 805 806 - "threaded" : A threaded cgroup which is a member of a 807 threaded subtree. 808 809 A cgroup can be turned into a threaded cgroup by writing 810 "threaded" to this file. 811 812 cgroup.procs 813 A read-write new-line separated values file which exists on 814 all cgroups. 815 816 When read, it lists the PIDs of all processes which belong to 817 the cgroup one-per-line. The PIDs are not ordered and the 818 same PID may show up more than once if the process got moved 819 to another cgroup and then back or the PID got recycled while 820 reading. 821 822 A PID can be written to migrate the process associated with 823 the PID to the cgroup. The writer should match all of the 824 following conditions. 825 826 - It must have write access to the "cgroup.procs" file. 827 828 - It must have write access to the "cgroup.procs" file of the 829 common ancestor of the source and destination cgroups. 830 831 When delegating a sub-hierarchy, write access to this file 832 should be granted along with the containing directory. 833 834 In a threaded cgroup, reading this file fails with EOPNOTSUPP 835 as all the processes belong to the thread root. Writing is 836 supported and moves every thread of the process to the cgroup. 837 838 cgroup.threads 839 A read-write new-line separated values file which exists on 840 all cgroups. 841 842 When read, it lists the TIDs of all threads which belong to 843 the cgroup one-per-line. The TIDs are not ordered and the 844 same TID may show up more than once if the thread got moved to 845 another cgroup and then back or the TID got recycled while 846 reading. 847 848 A TID can be written to migrate the thread associated with the 849 TID to the cgroup. The writer should match all of the 850 following conditions. 851 852 - It must have write access to the "cgroup.threads" file. 853 854 - The cgroup that the thread is currently in must be in the 855 same resource domain as the destination cgroup. 856 857 - It must have write access to the "cgroup.procs" file of the 858 common ancestor of the source and destination cgroups. 859 860 When delegating a sub-hierarchy, write access to this file 861 should be granted along with the containing directory. 862 863 cgroup.controllers 864 A read-only space separated values file which exists on all 865 cgroups. 866 867 It shows space separated list of all controllers available to 868 the cgroup. The controllers are not ordered. 869 870 cgroup.subtree_control 871 A read-write space separated values file which exists on all 872 cgroups. Starts out empty. 873 874 When read, it shows space separated list of the controllers 875 which are enabled to control resource distribution from the 876 cgroup to its children. 877 878 Space separated list of controllers prefixed with '+' or '-' 879 can be written to enable or disable controllers. A controller 880 name prefixed with '+' enables the controller and '-' 881 disables. If a controller appears more than once on the list, 882 the last one is effective. When multiple enable and disable 883 operations are specified, either all succeed or all fail. 884 885 cgroup.events 886 A read-only flat-keyed file which exists on non-root cgroups. 887 The following entries are defined. Unless specified 888 otherwise, a value change in this file generates a file 889 modified event. 890 891 populated 892 1 if the cgroup or its descendants contains any live 893 processes; otherwise, 0. 894 frozen 895 1 if the cgroup is frozen; otherwise, 0. 896 897 cgroup.max.descendants 898 A read-write single value files. The default is "max". 899 900 Maximum allowed number of descent cgroups. 901 If the actual number of descendants is equal or larger, 902 an attempt to create a new cgroup in the hierarchy will fail. 903 904 cgroup.max.depth 905 A read-write single value files. The default is "max". 906 907 Maximum allowed descent depth below the current cgroup. 908 If the actual descent depth is equal or larger, 909 an attempt to create a new child cgroup will fail. 910 911 cgroup.stat 912 A read-only flat-keyed file with the following entries: 913 914 nr_descendants 915 Total number of visible descendant cgroups. 916 917 nr_dying_descendants 918 Total number of dying descendant cgroups. A cgroup becomes 919 dying after being deleted by a user. The cgroup will remain 920 in dying state for some time undefined time (which can depend 921 on system load) before being completely destroyed. 922 923 A process can't enter a dying cgroup under any circumstances, 924 a dying cgroup can't revive. 925 926 A dying cgroup can consume system resources not exceeding 927 limits, which were active at the moment of cgroup deletion. 928 929 cgroup.freeze 930 A read-write single value file which exists on non-root cgroups. 931 Allowed values are "0" and "1". The default is "0". 932 933 Writing "1" to the file causes freezing of the cgroup and all 934 descendant cgroups. This means that all belonging processes will 935 be stopped and will not run until the cgroup will be explicitly 936 unfrozen. Freezing of the cgroup may take some time; when this action 937 is completed, the "frozen" value in the cgroup.events control file 938 will be updated to "1" and the corresponding notification will be 939 issued. 940 941 A cgroup can be frozen either by its own settings, or by settings 942 of any ancestor cgroups. If any of ancestor cgroups is frozen, the 943 cgroup will remain frozen. 944 945 Processes in the frozen cgroup can be killed by a fatal signal. 946 They also can enter and leave a frozen cgroup: either by an explicit 947 move by a user, or if freezing of the cgroup races with fork(). 948 If a process is moved to a frozen cgroup, it stops. If a process is 949 moved out of a frozen cgroup, it becomes running. 950 951 Frozen status of a cgroup doesn't affect any cgroup tree operations: 952 it's possible to delete a frozen (and empty) cgroup, as well as 953 create new sub-cgroups. 954 955Controllers 956=========== 957 958CPU 959--- 960 961The "cpu" controllers regulates distribution of CPU cycles. This 962controller implements weight and absolute bandwidth limit models for 963normal scheduling policy and absolute bandwidth allocation model for 964realtime scheduling policy. 965 966In all the above models, cycles distribution is defined only on a temporal 967base and it does not account for the frequency at which tasks are executed. 968The (optional) utilization clamping support allows to hint the schedutil 969cpufreq governor about the minimum desired frequency which should always be 970provided by a CPU, as well as the maximum desired frequency, which should not 971be exceeded by a CPU. 972 973WARNING: cgroup2 doesn't yet support control of realtime processes and 974the cpu controller can only be enabled when all RT processes are in 975the root cgroup. Be aware that system management software may already 976have placed RT processes into nonroot cgroups during the system boot 977process, and these processes may need to be moved to the root cgroup 978before the cpu controller can be enabled. 979 980 981CPU Interface Files 982~~~~~~~~~~~~~~~~~~~ 983 984All time durations are in microseconds. 985 986 cpu.stat 987 A read-only flat-keyed file. 988 This file exists whether the controller is enabled or not. 989 990 It always reports the following three stats: 991 992 - usage_usec 993 - user_usec 994 - system_usec 995 996 and the following three when the controller is enabled: 997 998 - nr_periods 999 - nr_throttled 1000 - throttled_usec 1001 1002 cpu.weight 1003 A read-write single value file which exists on non-root 1004 cgroups. The default is "100". 1005 1006 The weight in the range [1, 10000]. 1007 1008 cpu.weight.nice 1009 A read-write single value file which exists on non-root 1010 cgroups. The default is "0". 1011 1012 The nice value is in the range [-20, 19]. 1013 1014 This interface file is an alternative interface for 1015 "cpu.weight" and allows reading and setting weight using the 1016 same values used by nice(2). Because the range is smaller and 1017 granularity is coarser for the nice values, the read value is 1018 the closest approximation of the current weight. 1019 1020 cpu.max 1021 A read-write two value file which exists on non-root cgroups. 1022 The default is "max 100000". 1023 1024 The maximum bandwidth limit. It's in the following format:: 1025 1026 $MAX $PERIOD 1027 1028 which indicates that the group may consume upto $MAX in each 1029 $PERIOD duration. "max" for $MAX indicates no limit. If only 1030 one number is written, $MAX is updated. 1031 1032 cpu.pressure 1033 A read-only nested-key file which exists on non-root cgroups. 1034 1035 Shows pressure stall information for CPU. See 1036 :ref:`Documentation/accounting/psi.rst <psi>` for details. 1037 1038 cpu.uclamp.min 1039 A read-write single value file which exists on non-root cgroups. 1040 The default is "0", i.e. no utilization boosting. 1041 1042 The requested minimum utilization (protection) as a percentage 1043 rational number, e.g. 12.34 for 12.34%. 1044 1045 This interface allows reading and setting minimum utilization clamp 1046 values similar to the sched_setattr(2). This minimum utilization 1047 value is used to clamp the task specific minimum utilization clamp. 1048 1049 The requested minimum utilization (protection) is always capped by 1050 the current value for the maximum utilization (limit), i.e. 1051 `cpu.uclamp.max`. 1052 1053 cpu.uclamp.max 1054 A read-write single value file which exists on non-root cgroups. 1055 The default is "max". i.e. no utilization capping 1056 1057 The requested maximum utilization (limit) as a percentage rational 1058 number, e.g. 98.76 for 98.76%. 1059 1060 This interface allows reading and setting maximum utilization clamp 1061 values similar to the sched_setattr(2). This maximum utilization 1062 value is used to clamp the task specific maximum utilization clamp. 1063 1064 1065 1066Memory 1067------ 1068 1069The "memory" controller regulates distribution of memory. Memory is 1070stateful and implements both limit and protection models. Due to the 1071intertwining between memory usage and reclaim pressure and the 1072stateful nature of memory, the distribution model is relatively 1073complex. 1074 1075While not completely water-tight, all major memory usages by a given 1076cgroup are tracked so that the total memory consumption can be 1077accounted and controlled to a reasonable extent. Currently, the 1078following types of memory usages are tracked. 1079 1080- Userland memory - page cache and anonymous memory. 1081 1082- Kernel data structures such as dentries and inodes. 1083 1084- TCP socket buffers. 1085 1086The above list may expand in the future for better coverage. 1087 1088 1089Memory Interface Files 1090~~~~~~~~~~~~~~~~~~~~~~ 1091 1092All memory amounts are in bytes. If a value which is not aligned to 1093PAGE_SIZE is written, the value may be rounded up to the closest 1094PAGE_SIZE multiple when read back. 1095 1096 memory.current 1097 A read-only single value file which exists on non-root 1098 cgroups. 1099 1100 The total amount of memory currently being used by the cgroup 1101 and its descendants. 1102 1103 memory.min 1104 A read-write single value file which exists on non-root 1105 cgroups. The default is "0". 1106 1107 Hard memory protection. If the memory usage of a cgroup 1108 is within its effective min boundary, the cgroup's memory 1109 won't be reclaimed under any conditions. If there is no 1110 unprotected reclaimable memory available, OOM killer 1111 is invoked. Above the effective min boundary (or 1112 effective low boundary if it is higher), pages are reclaimed 1113 proportionally to the overage, reducing reclaim pressure for 1114 smaller overages. 1115 1116 Effective min boundary is limited by memory.min values of 1117 all ancestor cgroups. If there is memory.min overcommitment 1118 (child cgroup or cgroups are requiring more protected memory 1119 than parent will allow), then each child cgroup will get 1120 the part of parent's protection proportional to its 1121 actual memory usage below memory.min. 1122 1123 Putting more memory than generally available under this 1124 protection is discouraged and may lead to constant OOMs. 1125 1126 If a memory cgroup is not populated with processes, 1127 its memory.min is ignored. 1128 1129 memory.low 1130 A read-write single value file which exists on non-root 1131 cgroups. The default is "0". 1132 1133 Best-effort memory protection. If the memory usage of a 1134 cgroup is within its effective low boundary, the cgroup's 1135 memory won't be reclaimed unless there is no reclaimable 1136 memory available in unprotected cgroups. 1137 Above the effective low boundary (or 1138 effective min boundary if it is higher), pages are reclaimed 1139 proportionally to the overage, reducing reclaim pressure for 1140 smaller overages. 1141 1142 Effective low boundary is limited by memory.low values of 1143 all ancestor cgroups. If there is memory.low overcommitment 1144 (child cgroup or cgroups are requiring more protected memory 1145 than parent will allow), then each child cgroup will get 1146 the part of parent's protection proportional to its 1147 actual memory usage below memory.low. 1148 1149 Putting more memory than generally available under this 1150 protection is discouraged. 1151 1152 memory.high 1153 A read-write single value file which exists on non-root 1154 cgroups. The default is "max". 1155 1156 Memory usage throttle limit. This is the main mechanism to 1157 control memory usage of a cgroup. If a cgroup's usage goes 1158 over the high boundary, the processes of the cgroup are 1159 throttled and put under heavy reclaim pressure. 1160 1161 Going over the high limit never invokes the OOM killer and 1162 under extreme conditions the limit may be breached. 1163 1164 memory.max 1165 A read-write single value file which exists on non-root 1166 cgroups. The default is "max". 1167 1168 Memory usage hard limit. This is the final protection 1169 mechanism. If a cgroup's memory usage reaches this limit and 1170 can't be reduced, the OOM killer is invoked in the cgroup. 1171 Under certain circumstances, the usage may go over the limit 1172 temporarily. 1173 1174 In default configuration regular 0-order allocations always 1175 succeed unless OOM killer chooses current task as a victim. 1176 1177 Some kinds of allocations don't invoke the OOM killer. 1178 Caller could retry them differently, return into userspace 1179 as -ENOMEM or silently ignore in cases like disk readahead. 1180 1181 This is the ultimate protection mechanism. As long as the 1182 high limit is used and monitored properly, this limit's 1183 utility is limited to providing the final safety net. 1184 1185 memory.oom.group 1186 A read-write single value file which exists on non-root 1187 cgroups. The default value is "0". 1188 1189 Determines whether the cgroup should be treated as 1190 an indivisible workload by the OOM killer. If set, 1191 all tasks belonging to the cgroup or to its descendants 1192 (if the memory cgroup is not a leaf cgroup) are killed 1193 together or not at all. This can be used to avoid 1194 partial kills to guarantee workload integrity. 1195 1196 Tasks with the OOM protection (oom_score_adj set to -1000) 1197 are treated as an exception and are never killed. 1198 1199 If the OOM killer is invoked in a cgroup, it's not going 1200 to kill any tasks outside of this cgroup, regardless 1201 memory.oom.group values of ancestor cgroups. 1202 1203 memory.events 1204 A read-only flat-keyed file which exists on non-root cgroups. 1205 The following entries are defined. Unless specified 1206 otherwise, a value change in this file generates a file 1207 modified event. 1208 1209 Note that all fields in this file are hierarchical and the 1210 file modified event can be generated due to an event down the 1211 hierarchy. For for the local events at the cgroup level see 1212 memory.events.local. 1213 1214 low 1215 The number of times the cgroup is reclaimed due to 1216 high memory pressure even though its usage is under 1217 the low boundary. This usually indicates that the low 1218 boundary is over-committed. 1219 1220 high 1221 The number of times processes of the cgroup are 1222 throttled and routed to perform direct memory reclaim 1223 because the high memory boundary was exceeded. For a 1224 cgroup whose memory usage is capped by the high limit 1225 rather than global memory pressure, this event's 1226 occurrences are expected. 1227 1228 max 1229 The number of times the cgroup's memory usage was 1230 about to go over the max boundary. If direct reclaim 1231 fails to bring it down, the cgroup goes to OOM state. 1232 1233 oom 1234 The number of time the cgroup's memory usage was 1235 reached the limit and allocation was about to fail. 1236 1237 This event is not raised if the OOM killer is not 1238 considered as an option, e.g. for failed high-order 1239 allocations or if caller asked to not retry attempts. 1240 1241 oom_kill 1242 The number of processes belonging to this cgroup 1243 killed by any kind of OOM killer. 1244 1245 memory.events.local 1246 Similar to memory.events but the fields in the file are local 1247 to the cgroup i.e. not hierarchical. The file modified event 1248 generated on this file reflects only the local events. 1249 1250 memory.stat 1251 A read-only flat-keyed file which exists on non-root cgroups. 1252 1253 This breaks down the cgroup's memory footprint into different 1254 types of memory, type-specific details, and other information 1255 on the state and past events of the memory management system. 1256 1257 All memory amounts are in bytes. 1258 1259 The entries are ordered to be human readable, and new entries 1260 can show up in the middle. Don't rely on items remaining in a 1261 fixed position; use the keys to look up specific values! 1262 1263 If the entry has no per-node counter(or not show in the 1264 mempry.numa_stat). We use 'npn'(non-per-node) as the tag 1265 to indicate that it will not show in the mempry.numa_stat. 1266 1267 anon 1268 Amount of memory used in anonymous mappings such as 1269 brk(), sbrk(), and mmap(MAP_ANONYMOUS) 1270 1271 file 1272 Amount of memory used to cache filesystem data, 1273 including tmpfs and shared memory. 1274 1275 kernel_stack 1276 Amount of memory allocated to kernel stacks. 1277 1278 percpu(npn) 1279 Amount of memory used for storing per-cpu kernel 1280 data structures. 1281 1282 sock(npn) 1283 Amount of memory used in network transmission buffers 1284 1285 shmem 1286 Amount of cached filesystem data that is swap-backed, 1287 such as tmpfs, shm segments, shared anonymous mmap()s 1288 1289 file_mapped 1290 Amount of cached filesystem data mapped with mmap() 1291 1292 file_dirty 1293 Amount of cached filesystem data that was modified but 1294 not yet written back to disk 1295 1296 file_writeback 1297 Amount of cached filesystem data that was modified and 1298 is currently being written back to disk 1299 1300 anon_thp 1301 Amount of memory used in anonymous mappings backed by 1302 transparent hugepages 1303 1304 inactive_anon, active_anon, inactive_file, active_file, unevictable 1305 Amount of memory, swap-backed and filesystem-backed, 1306 on the internal memory management lists used by the 1307 page reclaim algorithm. 1308 1309 As these represent internal list state (eg. shmem pages are on anon 1310 memory management lists), inactive_foo + active_foo may not be equal to 1311 the value for the foo counter, since the foo counter is type-based, not 1312 list-based. 1313 1314 slab_reclaimable 1315 Part of "slab" that might be reclaimed, such as 1316 dentries and inodes. 1317 1318 slab_unreclaimable 1319 Part of "slab" that cannot be reclaimed on memory 1320 pressure. 1321 1322 slab(npn) 1323 Amount of memory used for storing in-kernel data 1324 structures. 1325 1326 workingset_refault_anon 1327 Number of refaults of previously evicted anonymous pages. 1328 1329 workingset_refault_file 1330 Number of refaults of previously evicted file pages. 1331 1332 workingset_activate_anon 1333 Number of refaulted anonymous pages that were immediately 1334 activated. 1335 1336 workingset_activate_file 1337 Number of refaulted file pages that were immediately activated. 1338 1339 workingset_restore_anon 1340 Number of restored anonymous pages which have been detected as 1341 an active workingset before they got reclaimed. 1342 1343 workingset_restore_file 1344 Number of restored file pages which have been detected as an 1345 active workingset before they got reclaimed. 1346 1347 workingset_nodereclaim 1348 Number of times a shadow node has been reclaimed 1349 1350 pgfault(npn) 1351 Total number of page faults incurred 1352 1353 pgmajfault(npn) 1354 Number of major page faults incurred 1355 1356 pgrefill(npn) 1357 Amount of scanned pages (in an active LRU list) 1358 1359 pgscan(npn) 1360 Amount of scanned pages (in an inactive LRU list) 1361 1362 pgsteal(npn) 1363 Amount of reclaimed pages 1364 1365 pgactivate(npn) 1366 Amount of pages moved to the active LRU list 1367 1368 pgdeactivate(npn) 1369 Amount of pages moved to the inactive LRU list 1370 1371 pglazyfree(npn) 1372 Amount of pages postponed to be freed under memory pressure 1373 1374 pglazyfreed(npn) 1375 Amount of reclaimed lazyfree pages 1376 1377 thp_fault_alloc(npn) 1378 Number of transparent hugepages which were allocated to satisfy 1379 a page fault. This counter is not present when CONFIG_TRANSPARENT_HUGEPAGE 1380 is not set. 1381 1382 thp_collapse_alloc(npn) 1383 Number of transparent hugepages which were allocated to allow 1384 collapsing an existing range of pages. This counter is not 1385 present when CONFIG_TRANSPARENT_HUGEPAGE is not set. 1386 1387 memory.numa_stat 1388 A read-only nested-keyed file which exists on non-root cgroups. 1389 1390 This breaks down the cgroup's memory footprint into different 1391 types of memory, type-specific details, and other information 1392 per node on the state of the memory management system. 1393 1394 This is useful for providing visibility into the NUMA locality 1395 information within an memcg since the pages are allowed to be 1396 allocated from any physical node. One of the use case is evaluating 1397 application performance by combining this information with the 1398 application's CPU allocation. 1399 1400 All memory amounts are in bytes. 1401 1402 The output format of memory.numa_stat is:: 1403 1404 type N0=<bytes in node 0> N1=<bytes in node 1> ... 1405 1406 The entries are ordered to be human readable, and new entries 1407 can show up in the middle. Don't rely on items remaining in a 1408 fixed position; use the keys to look up specific values! 1409 1410 The entries can refer to the memory.stat. 1411 1412 memory.swap.current 1413 A read-only single value file which exists on non-root 1414 cgroups. 1415 1416 The total amount of swap currently being used by the cgroup 1417 and its descendants. 1418 1419 memory.swap.high 1420 A read-write single value file which exists on non-root 1421 cgroups. The default is "max". 1422 1423 Swap usage throttle limit. If a cgroup's swap usage exceeds 1424 this limit, all its further allocations will be throttled to 1425 allow userspace to implement custom out-of-memory procedures. 1426 1427 This limit marks a point of no return for the cgroup. It is NOT 1428 designed to manage the amount of swapping a workload does 1429 during regular operation. Compare to memory.swap.max, which 1430 prohibits swapping past a set amount, but lets the cgroup 1431 continue unimpeded as long as other memory can be reclaimed. 1432 1433 Healthy workloads are not expected to reach this limit. 1434 1435 memory.swap.max 1436 A read-write single value file which exists on non-root 1437 cgroups. The default is "max". 1438 1439 Swap usage hard limit. If a cgroup's swap usage reaches this 1440 limit, anonymous memory of the cgroup will not be swapped out. 1441 1442 memory.swap.events 1443 A read-only flat-keyed file which exists on non-root cgroups. 1444 The following entries are defined. Unless specified 1445 otherwise, a value change in this file generates a file 1446 modified event. 1447 1448 high 1449 The number of times the cgroup's swap usage was over 1450 the high threshold. 1451 1452 max 1453 The number of times the cgroup's swap usage was about 1454 to go over the max boundary and swap allocation 1455 failed. 1456 1457 fail 1458 The number of times swap allocation failed either 1459 because of running out of swap system-wide or max 1460 limit. 1461 1462 When reduced under the current usage, the existing swap 1463 entries are reclaimed gradually and the swap usage may stay 1464 higher than the limit for an extended period of time. This 1465 reduces the impact on the workload and memory management. 1466 1467 memory.pressure 1468 A read-only nested-key file which exists on non-root cgroups. 1469 1470 Shows pressure stall information for memory. See 1471 :ref:`Documentation/accounting/psi.rst <psi>` for details. 1472 1473 1474Usage Guidelines 1475~~~~~~~~~~~~~~~~ 1476 1477"memory.high" is the main mechanism to control memory usage. 1478Over-committing on high limit (sum of high limits > available memory) 1479and letting global memory pressure to distribute memory according to 1480usage is a viable strategy. 1481 1482Because breach of the high limit doesn't trigger the OOM killer but 1483throttles the offending cgroup, a management agent has ample 1484opportunities to monitor and take appropriate actions such as granting 1485more memory or terminating the workload. 1486 1487Determining whether a cgroup has enough memory is not trivial as 1488memory usage doesn't indicate whether the workload can benefit from 1489more memory. For example, a workload which writes data received from 1490network to a file can use all available memory but can also operate as 1491performant with a small amount of memory. A measure of memory 1492pressure - how much the workload is being impacted due to lack of 1493memory - is necessary to determine whether a workload needs more 1494memory; unfortunately, memory pressure monitoring mechanism isn't 1495implemented yet. 1496 1497 1498Memory Ownership 1499~~~~~~~~~~~~~~~~ 1500 1501A memory area is charged to the cgroup which instantiated it and stays 1502charged to the cgroup until the area is released. Migrating a process 1503to a different cgroup doesn't move the memory usages that it 1504instantiated while in the previous cgroup to the new cgroup. 1505 1506A memory area may be used by processes belonging to different cgroups. 1507To which cgroup the area will be charged is in-deterministic; however, 1508over time, the memory area is likely to end up in a cgroup which has 1509enough memory allowance to avoid high reclaim pressure. 1510 1511If a cgroup sweeps a considerable amount of memory which is expected 1512to be accessed repeatedly by other cgroups, it may make sense to use 1513POSIX_FADV_DONTNEED to relinquish the ownership of memory areas 1514belonging to the affected files to ensure correct memory ownership. 1515 1516 1517IO 1518-- 1519 1520The "io" controller regulates the distribution of IO resources. This 1521controller implements both weight based and absolute bandwidth or IOPS 1522limit distribution; however, weight based distribution is available 1523only if cfq-iosched is in use and neither scheme is available for 1524blk-mq devices. 1525 1526 1527IO Interface Files 1528~~~~~~~~~~~~~~~~~~ 1529 1530 io.stat 1531 A read-only nested-keyed file. 1532 1533 Lines are keyed by $MAJ:$MIN device numbers and not ordered. 1534 The following nested keys are defined. 1535 1536 ====== ===================== 1537 rbytes Bytes read 1538 wbytes Bytes written 1539 rios Number of read IOs 1540 wios Number of write IOs 1541 dbytes Bytes discarded 1542 dios Number of discard IOs 1543 ====== ===================== 1544 1545 An example read output follows:: 1546 1547 8:16 rbytes=1459200 wbytes=314773504 rios=192 wios=353 dbytes=0 dios=0 1548 8:0 rbytes=90430464 wbytes=299008000 rios=8950 wios=1252 dbytes=50331648 dios=3021 1549 1550 io.cost.qos 1551 A read-write nested-keyed file with exists only on the root 1552 cgroup. 1553 1554 This file configures the Quality of Service of the IO cost 1555 model based controller (CONFIG_BLK_CGROUP_IOCOST) which 1556 currently implements "io.weight" proportional control. Lines 1557 are keyed by $MAJ:$MIN device numbers and not ordered. The 1558 line for a given device is populated on the first write for 1559 the device on "io.cost.qos" or "io.cost.model". The following 1560 nested keys are defined. 1561 1562 ====== ===================================== 1563 enable Weight-based control enable 1564 ctrl "auto" or "user" 1565 rpct Read latency percentile [0, 100] 1566 rlat Read latency threshold 1567 wpct Write latency percentile [0, 100] 1568 wlat Write latency threshold 1569 min Minimum scaling percentage [1, 10000] 1570 max Maximum scaling percentage [1, 10000] 1571 ====== ===================================== 1572 1573 The controller is disabled by default and can be enabled by 1574 setting "enable" to 1. "rpct" and "wpct" parameters default 1575 to zero and the controller uses internal device saturation 1576 state to adjust the overall IO rate between "min" and "max". 1577 1578 When a better control quality is needed, latency QoS 1579 parameters can be configured. For example:: 1580 1581 8:16 enable=1 ctrl=auto rpct=95.00 rlat=75000 wpct=95.00 wlat=150000 min=50.00 max=150.0 1582 1583 shows that on sdb, the controller is enabled, will consider 1584 the device saturated if the 95th percentile of read completion 1585 latencies is above 75ms or write 150ms, and adjust the overall 1586 IO issue rate between 50% and 150% accordingly. 1587 1588 The lower the saturation point, the better the latency QoS at 1589 the cost of aggregate bandwidth. The narrower the allowed 1590 adjustment range between "min" and "max", the more conformant 1591 to the cost model the IO behavior. Note that the IO issue 1592 base rate may be far off from 100% and setting "min" and "max" 1593 blindly can lead to a significant loss of device capacity or 1594 control quality. "min" and "max" are useful for regulating 1595 devices which show wide temporary behavior changes - e.g. a 1596 ssd which accepts writes at the line speed for a while and 1597 then completely stalls for multiple seconds. 1598 1599 When "ctrl" is "auto", the parameters are controlled by the 1600 kernel and may change automatically. Setting "ctrl" to "user" 1601 or setting any of the percentile and latency parameters puts 1602 it into "user" mode and disables the automatic changes. The 1603 automatic mode can be restored by setting "ctrl" to "auto". 1604 1605 io.cost.model 1606 A read-write nested-keyed file with exists only on the root 1607 cgroup. 1608 1609 This file configures the cost model of the IO cost model based 1610 controller (CONFIG_BLK_CGROUP_IOCOST) which currently 1611 implements "io.weight" proportional control. Lines are keyed 1612 by $MAJ:$MIN device numbers and not ordered. The line for a 1613 given device is populated on the first write for the device on 1614 "io.cost.qos" or "io.cost.model". The following nested keys 1615 are defined. 1616 1617 ===== ================================ 1618 ctrl "auto" or "user" 1619 model The cost model in use - "linear" 1620 ===== ================================ 1621 1622 When "ctrl" is "auto", the kernel may change all parameters 1623 dynamically. When "ctrl" is set to "user" or any other 1624 parameters are written to, "ctrl" become "user" and the 1625 automatic changes are disabled. 1626 1627 When "model" is "linear", the following model parameters are 1628 defined. 1629 1630 ============= ======================================== 1631 [r|w]bps The maximum sequential IO throughput 1632 [r|w]seqiops The maximum 4k sequential IOs per second 1633 [r|w]randiops The maximum 4k random IOs per second 1634 ============= ======================================== 1635 1636 From the above, the builtin linear model determines the base 1637 costs of a sequential and random IO and the cost coefficient 1638 for the IO size. While simple, this model can cover most 1639 common device classes acceptably. 1640 1641 The IO cost model isn't expected to be accurate in absolute 1642 sense and is scaled to the device behavior dynamically. 1643 1644 If needed, tools/cgroup/iocost_coef_gen.py can be used to 1645 generate device-specific coefficients. 1646 1647 io.weight 1648 A read-write flat-keyed file which exists on non-root cgroups. 1649 The default is "default 100". 1650 1651 The first line is the default weight applied to devices 1652 without specific override. The rest are overrides keyed by 1653 $MAJ:$MIN device numbers and not ordered. The weights are in 1654 the range [1, 10000] and specifies the relative amount IO time 1655 the cgroup can use in relation to its siblings. 1656 1657 The default weight can be updated by writing either "default 1658 $WEIGHT" or simply "$WEIGHT". Overrides can be set by writing 1659 "$MAJ:$MIN $WEIGHT" and unset by writing "$MAJ:$MIN default". 1660 1661 An example read output follows:: 1662 1663 default 100 1664 8:16 200 1665 8:0 50 1666 1667 io.max 1668 A read-write nested-keyed file which exists on non-root 1669 cgroups. 1670 1671 BPS and IOPS based IO limit. Lines are keyed by $MAJ:$MIN 1672 device numbers and not ordered. The following nested keys are 1673 defined. 1674 1675 ===== ================================== 1676 rbps Max read bytes per second 1677 wbps Max write bytes per second 1678 riops Max read IO operations per second 1679 wiops Max write IO operations per second 1680 ===== ================================== 1681 1682 When writing, any number of nested key-value pairs can be 1683 specified in any order. "max" can be specified as the value 1684 to remove a specific limit. If the same key is specified 1685 multiple times, the outcome is undefined. 1686 1687 BPS and IOPS are measured in each IO direction and IOs are 1688 delayed if limit is reached. Temporary bursts are allowed. 1689 1690 Setting read limit at 2M BPS and write at 120 IOPS for 8:16:: 1691 1692 echo "8:16 rbps=2097152 wiops=120" > io.max 1693 1694 Reading returns the following:: 1695 1696 8:16 rbps=2097152 wbps=max riops=max wiops=120 1697 1698 Write IOPS limit can be removed by writing the following:: 1699 1700 echo "8:16 wiops=max" > io.max 1701 1702 Reading now returns the following:: 1703 1704 8:16 rbps=2097152 wbps=max riops=max wiops=max 1705 1706 io.pressure 1707 A read-only nested-key file which exists on non-root cgroups. 1708 1709 Shows pressure stall information for IO. See 1710 :ref:`Documentation/accounting/psi.rst <psi>` for details. 1711 1712 1713Writeback 1714~~~~~~~~~ 1715 1716Page cache is dirtied through buffered writes and shared mmaps and 1717written asynchronously to the backing filesystem by the writeback 1718mechanism. Writeback sits between the memory and IO domains and 1719regulates the proportion of dirty memory by balancing dirtying and 1720write IOs. 1721 1722The io controller, in conjunction with the memory controller, 1723implements control of page cache writeback IOs. The memory controller 1724defines the memory domain that dirty memory ratio is calculated and 1725maintained for and the io controller defines the io domain which 1726writes out dirty pages for the memory domain. Both system-wide and 1727per-cgroup dirty memory states are examined and the more restrictive 1728of the two is enforced. 1729 1730cgroup writeback requires explicit support from the underlying 1731filesystem. Currently, cgroup writeback is implemented on ext2, ext4, 1732btrfs, f2fs, and xfs. On other filesystems, all writeback IOs are 1733attributed to the root cgroup. 1734 1735There are inherent differences in memory and writeback management 1736which affects how cgroup ownership is tracked. Memory is tracked per 1737page while writeback per inode. For the purpose of writeback, an 1738inode is assigned to a cgroup and all IO requests to write dirty pages 1739from the inode are attributed to that cgroup. 1740 1741As cgroup ownership for memory is tracked per page, there can be pages 1742which are associated with different cgroups than the one the inode is 1743associated with. These are called foreign pages. The writeback 1744constantly keeps track of foreign pages and, if a particular foreign 1745cgroup becomes the majority over a certain period of time, switches 1746the ownership of the inode to that cgroup. 1747 1748While this model is enough for most use cases where a given inode is 1749mostly dirtied by a single cgroup even when the main writing cgroup 1750changes over time, use cases where multiple cgroups write to a single 1751inode simultaneously are not supported well. In such circumstances, a 1752significant portion of IOs are likely to be attributed incorrectly. 1753As memory controller assigns page ownership on the first use and 1754doesn't update it until the page is released, even if writeback 1755strictly follows page ownership, multiple cgroups dirtying overlapping 1756areas wouldn't work as expected. It's recommended to avoid such usage 1757patterns. 1758 1759The sysctl knobs which affect writeback behavior are applied to cgroup 1760writeback as follows. 1761 1762 vm.dirty_background_ratio, vm.dirty_ratio 1763 These ratios apply the same to cgroup writeback with the 1764 amount of available memory capped by limits imposed by the 1765 memory controller and system-wide clean memory. 1766 1767 vm.dirty_background_bytes, vm.dirty_bytes 1768 For cgroup writeback, this is calculated into ratio against 1769 total available memory and applied the same way as 1770 vm.dirty[_background]_ratio. 1771 1772 1773IO Latency 1774~~~~~~~~~~ 1775 1776This is a cgroup v2 controller for IO workload protection. You provide a group 1777with a latency target, and if the average latency exceeds that target the 1778controller will throttle any peers that have a lower latency target than the 1779protected workload. 1780 1781The limits are only applied at the peer level in the hierarchy. This means that 1782in the diagram below, only groups A, B, and C will influence each other, and 1783groups D and F will influence each other. Group G will influence nobody:: 1784 1785 [root] 1786 / | \ 1787 A B C 1788 / \ | 1789 D F G 1790 1791 1792So the ideal way to configure this is to set io.latency in groups A, B, and C. 1793Generally you do not want to set a value lower than the latency your device 1794supports. Experiment to find the value that works best for your workload. 1795Start at higher than the expected latency for your device and watch the 1796avg_lat value in io.stat for your workload group to get an idea of the 1797latency you see during normal operation. Use the avg_lat value as a basis for 1798your real setting, setting at 10-15% higher than the value in io.stat. 1799 1800How IO Latency Throttling Works 1801~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~ 1802 1803io.latency is work conserving; so as long as everybody is meeting their latency 1804target the controller doesn't do anything. Once a group starts missing its 1805target it begins throttling any peer group that has a higher target than itself. 1806This throttling takes 2 forms: 1807 1808- Queue depth throttling. This is the number of outstanding IO's a group is 1809 allowed to have. We will clamp down relatively quickly, starting at no limit 1810 and going all the way down to 1 IO at a time. 1811 1812- Artificial delay induction. There are certain types of IO that cannot be 1813 throttled without possibly adversely affecting higher priority groups. This 1814 includes swapping and metadata IO. These types of IO are allowed to occur 1815 normally, however they are "charged" to the originating group. If the 1816 originating group is being throttled you will see the use_delay and delay 1817 fields in io.stat increase. The delay value is how many microseconds that are 1818 being added to any process that runs in this group. Because this number can 1819 grow quite large if there is a lot of swapping or metadata IO occurring we 1820 limit the individual delay events to 1 second at a time. 1821 1822Once the victimized group starts meeting its latency target again it will start 1823unthrottling any peer groups that were throttled previously. If the victimized 1824group simply stops doing IO the global counter will unthrottle appropriately. 1825 1826IO Latency Interface Files 1827~~~~~~~~~~~~~~~~~~~~~~~~~~ 1828 1829 io.latency 1830 This takes a similar format as the other controllers. 1831 1832 "MAJOR:MINOR target=<target time in microseconds" 1833 1834 io.stat 1835 If the controller is enabled you will see extra stats in io.stat in 1836 addition to the normal ones. 1837 1838 depth 1839 This is the current queue depth for the group. 1840 1841 avg_lat 1842 This is an exponential moving average with a decay rate of 1/exp 1843 bound by the sampling interval. The decay rate interval can be 1844 calculated by multiplying the win value in io.stat by the 1845 corresponding number of samples based on the win value. 1846 1847 win 1848 The sampling window size in milliseconds. This is the minimum 1849 duration of time between evaluation events. Windows only elapse 1850 with IO activity. Idle periods extend the most recent window. 1851 1852IO Priority 1853~~~~~~~~~~~ 1854 1855A single attribute controls the behavior of the I/O priority cgroup policy, 1856namely the blkio.prio.class attribute. The following values are accepted for 1857that attribute: 1858 1859 no-change 1860 Do not modify the I/O priority class. 1861 1862 none-to-rt 1863 For requests that do not have an I/O priority class (NONE), 1864 change the I/O priority class into RT. Do not modify 1865 the I/O priority class of other requests. 1866 1867 restrict-to-be 1868 For requests that do not have an I/O priority class or that have I/O 1869 priority class RT, change it into BE. Do not modify the I/O priority 1870 class of requests that have priority class IDLE. 1871 1872 idle 1873 Change the I/O priority class of all requests into IDLE, the lowest 1874 I/O priority class. 1875 1876The following numerical values are associated with the I/O priority policies: 1877 1878+-------------+---+ 1879| no-change | 0 | 1880+-------------+---+ 1881| none-to-rt | 1 | 1882+-------------+---+ 1883| rt-to-be | 2 | 1884+-------------+---+ 1885| all-to-idle | 3 | 1886+-------------+---+ 1887 1888The numerical value that corresponds to each I/O priority class is as follows: 1889 1890+-------------------------------+---+ 1891| IOPRIO_CLASS_NONE | 0 | 1892+-------------------------------+---+ 1893| IOPRIO_CLASS_RT (real-time) | 1 | 1894+-------------------------------+---+ 1895| IOPRIO_CLASS_BE (best effort) | 2 | 1896+-------------------------------+---+ 1897| IOPRIO_CLASS_IDLE | 3 | 1898+-------------------------------+---+ 1899 1900The algorithm to set the I/O priority class for a request is as follows: 1901 1902- Translate the I/O priority class policy into a number. 1903- Change the request I/O priority class into the maximum of the I/O priority 1904 class policy number and the numerical I/O priority class. 1905 1906PID 1907--- 1908 1909The process number controller is used to allow a cgroup to stop any 1910new tasks from being fork()'d or clone()'d after a specified limit is 1911reached. 1912 1913The number of tasks in a cgroup can be exhausted in ways which other 1914controllers cannot prevent, thus warranting its own controller. For 1915example, a fork bomb is likely to exhaust the number of tasks before 1916hitting memory restrictions. 1917 1918Note that PIDs used in this controller refer to TIDs, process IDs as 1919used by the kernel. 1920 1921 1922PID Interface Files 1923~~~~~~~~~~~~~~~~~~~ 1924 1925 pids.max 1926 A read-write single value file which exists on non-root 1927 cgroups. The default is "max". 1928 1929 Hard limit of number of processes. 1930 1931 pids.current 1932 A read-only single value file which exists on all cgroups. 1933 1934 The number of processes currently in the cgroup and its 1935 descendants. 1936 1937Organisational operations are not blocked by cgroup policies, so it is 1938possible to have pids.current > pids.max. This can be done by either 1939setting the limit to be smaller than pids.current, or attaching enough 1940processes to the cgroup such that pids.current is larger than 1941pids.max. However, it is not possible to violate a cgroup PID policy 1942through fork() or clone(). These will return -EAGAIN if the creation 1943of a new process would cause a cgroup policy to be violated. 1944 1945 1946Cpuset 1947------ 1948 1949The "cpuset" controller provides a mechanism for constraining 1950the CPU and memory node placement of tasks to only the resources 1951specified in the cpuset interface files in a task's current cgroup. 1952This is especially valuable on large NUMA systems where placing jobs 1953on properly sized subsets of the systems with careful processor and 1954memory placement to reduce cross-node memory access and contention 1955can improve overall system performance. 1956 1957The "cpuset" controller is hierarchical. That means the controller 1958cannot use CPUs or memory nodes not allowed in its parent. 1959 1960 1961Cpuset Interface Files 1962~~~~~~~~~~~~~~~~~~~~~~ 1963 1964 cpuset.cpus 1965 A read-write multiple values file which exists on non-root 1966 cpuset-enabled cgroups. 1967 1968 It lists the requested CPUs to be used by tasks within this 1969 cgroup. The actual list of CPUs to be granted, however, is 1970 subjected to constraints imposed by its parent and can differ 1971 from the requested CPUs. 1972 1973 The CPU numbers are comma-separated numbers or ranges. 1974 For example:: 1975 1976 # cat cpuset.cpus 1977 0-4,6,8-10 1978 1979 An empty value indicates that the cgroup is using the same 1980 setting as the nearest cgroup ancestor with a non-empty 1981 "cpuset.cpus" or all the available CPUs if none is found. 1982 1983 The value of "cpuset.cpus" stays constant until the next update 1984 and won't be affected by any CPU hotplug events. 1985 1986 cpuset.cpus.effective 1987 A read-only multiple values file which exists on all 1988 cpuset-enabled cgroups. 1989 1990 It lists the onlined CPUs that are actually granted to this 1991 cgroup by its parent. These CPUs are allowed to be used by 1992 tasks within the current cgroup. 1993 1994 If "cpuset.cpus" is empty, the "cpuset.cpus.effective" file shows 1995 all the CPUs from the parent cgroup that can be available to 1996 be used by this cgroup. Otherwise, it should be a subset of 1997 "cpuset.cpus" unless none of the CPUs listed in "cpuset.cpus" 1998 can be granted. In this case, it will be treated just like an 1999 empty "cpuset.cpus". 2000 2001 Its value will be affected by CPU hotplug events. 2002 2003 cpuset.mems 2004 A read-write multiple values file which exists on non-root 2005 cpuset-enabled cgroups. 2006 2007 It lists the requested memory nodes to be used by tasks within 2008 this cgroup. The actual list of memory nodes granted, however, 2009 is subjected to constraints imposed by its parent and can differ 2010 from the requested memory nodes. 2011 2012 The memory node numbers are comma-separated numbers or ranges. 2013 For example:: 2014 2015 # cat cpuset.mems 2016 0-1,3 2017 2018 An empty value indicates that the cgroup is using the same 2019 setting as the nearest cgroup ancestor with a non-empty 2020 "cpuset.mems" or all the available memory nodes if none 2021 is found. 2022 2023 The value of "cpuset.mems" stays constant until the next update 2024 and won't be affected by any memory nodes hotplug events. 2025 2026 cpuset.mems.effective 2027 A read-only multiple values file which exists on all 2028 cpuset-enabled cgroups. 2029 2030 It lists the onlined memory nodes that are actually granted to 2031 this cgroup by its parent. These memory nodes are allowed to 2032 be used by tasks within the current cgroup. 2033 2034 If "cpuset.mems" is empty, it shows all the memory nodes from the 2035 parent cgroup that will be available to be used by this cgroup. 2036 Otherwise, it should be a subset of "cpuset.mems" unless none of 2037 the memory nodes listed in "cpuset.mems" can be granted. In this 2038 case, it will be treated just like an empty "cpuset.mems". 2039 2040 Its value will be affected by memory nodes hotplug events. 2041 2042 cpuset.cpus.partition 2043 A read-write single value file which exists on non-root 2044 cpuset-enabled cgroups. This flag is owned by the parent cgroup 2045 and is not delegatable. 2046 2047 It accepts only the following input values when written to. 2048 2049 "root" - a partition root 2050 "member" - a non-root member of a partition 2051 2052 When set to be a partition root, the current cgroup is the 2053 root of a new partition or scheduling domain that comprises 2054 itself and all its descendants except those that are separate 2055 partition roots themselves and their descendants. The root 2056 cgroup is always a partition root. 2057 2058 There are constraints on where a partition root can be set. 2059 It can only be set in a cgroup if all the following conditions 2060 are true. 2061 2062 1) The "cpuset.cpus" is not empty and the list of CPUs are 2063 exclusive, i.e. they are not shared by any of its siblings. 2064 2) The parent cgroup is a partition root. 2065 3) The "cpuset.cpus" is also a proper subset of the parent's 2066 "cpuset.cpus.effective". 2067 4) There is no child cgroups with cpuset enabled. This is for 2068 eliminating corner cases that have to be handled if such a 2069 condition is allowed. 2070 2071 Setting it to partition root will take the CPUs away from the 2072 effective CPUs of the parent cgroup. Once it is set, this 2073 file cannot be reverted back to "member" if there are any child 2074 cgroups with cpuset enabled. 2075 2076 A parent partition cannot distribute all its CPUs to its 2077 child partitions. There must be at least one cpu left in the 2078 parent partition. 2079 2080 Once becoming a partition root, changes to "cpuset.cpus" is 2081 generally allowed as long as the first condition above is true, 2082 the change will not take away all the CPUs from the parent 2083 partition and the new "cpuset.cpus" value is a superset of its 2084 children's "cpuset.cpus" values. 2085 2086 Sometimes, external factors like changes to ancestors' 2087 "cpuset.cpus" or cpu hotplug can cause the state of the partition 2088 root to change. On read, the "cpuset.sched.partition" file 2089 can show the following values. 2090 2091 "member" Non-root member of a partition 2092 "root" Partition root 2093 "root invalid" Invalid partition root 2094 2095 It is a partition root if the first 2 partition root conditions 2096 above are true and at least one CPU from "cpuset.cpus" is 2097 granted by the parent cgroup. 2098 2099 A partition root can become invalid if none of CPUs requested 2100 in "cpuset.cpus" can be granted by the parent cgroup or the 2101 parent cgroup is no longer a partition root itself. In this 2102 case, it is not a real partition even though the restriction 2103 of the first partition root condition above will still apply. 2104 The cpu affinity of all the tasks in the cgroup will then be 2105 associated with CPUs in the nearest ancestor partition. 2106 2107 An invalid partition root can be transitioned back to a 2108 real partition root if at least one of the requested CPUs 2109 can now be granted by its parent. In this case, the cpu 2110 affinity of all the tasks in the formerly invalid partition 2111 will be associated to the CPUs of the newly formed partition. 2112 Changing the partition state of an invalid partition root to 2113 "member" is always allowed even if child cpusets are present. 2114 2115 2116Device controller 2117----------------- 2118 2119Device controller manages access to device files. It includes both 2120creation of new device files (using mknod), and access to the 2121existing device files. 2122 2123Cgroup v2 device controller has no interface files and is implemented 2124on top of cgroup BPF. To control access to device files, a user may 2125create bpf programs of the BPF_CGROUP_DEVICE type and attach them 2126to cgroups. On an attempt to access a device file, corresponding 2127BPF programs will be executed, and depending on the return value 2128the attempt will succeed or fail with -EPERM. 2129 2130A BPF_CGROUP_DEVICE program takes a pointer to the bpf_cgroup_dev_ctx 2131structure, which describes the device access attempt: access type 2132(mknod/read/write) and device (type, major and minor numbers). 2133If the program returns 0, the attempt fails with -EPERM, otherwise 2134it succeeds. 2135 2136An example of BPF_CGROUP_DEVICE program may be found in the kernel 2137source tree in the tools/testing/selftests/bpf/dev_cgroup.c file. 2138 2139 2140RDMA 2141---- 2142 2143The "rdma" controller regulates the distribution and accounting of 2144RDMA resources. 2145 2146RDMA Interface Files 2147~~~~~~~~~~~~~~~~~~~~ 2148 2149 rdma.max 2150 A readwrite nested-keyed file that exists for all the cgroups 2151 except root that describes current configured resource limit 2152 for a RDMA/IB device. 2153 2154 Lines are keyed by device name and are not ordered. 2155 Each line contains space separated resource name and its configured 2156 limit that can be distributed. 2157 2158 The following nested keys are defined. 2159 2160 ========== ============================= 2161 hca_handle Maximum number of HCA Handles 2162 hca_object Maximum number of HCA Objects 2163 ========== ============================= 2164 2165 An example for mlx4 and ocrdma device follows:: 2166 2167 mlx4_0 hca_handle=2 hca_object=2000 2168 ocrdma1 hca_handle=3 hca_object=max 2169 2170 rdma.current 2171 A read-only file that describes current resource usage. 2172 It exists for all the cgroup except root. 2173 2174 An example for mlx4 and ocrdma device follows:: 2175 2176 mlx4_0 hca_handle=1 hca_object=20 2177 ocrdma1 hca_handle=1 hca_object=23 2178 2179HugeTLB 2180------- 2181 2182The HugeTLB controller allows to limit the HugeTLB usage per control group and 2183enforces the controller limit during page fault. 2184 2185HugeTLB Interface Files 2186~~~~~~~~~~~~~~~~~~~~~~~ 2187 2188 hugetlb.<hugepagesize>.current 2189 Show current usage for "hugepagesize" hugetlb. It exists for all 2190 the cgroup except root. 2191 2192 hugetlb.<hugepagesize>.max 2193 Set/show the hard limit of "hugepagesize" hugetlb usage. 2194 The default value is "max". It exists for all the cgroup except root. 2195 2196 hugetlb.<hugepagesize>.events 2197 A read-only flat-keyed file which exists on non-root cgroups. 2198 2199 max 2200 The number of allocation failure due to HugeTLB limit 2201 2202 hugetlb.<hugepagesize>.events.local 2203 Similar to hugetlb.<hugepagesize>.events but the fields in the file 2204 are local to the cgroup i.e. not hierarchical. The file modified event 2205 generated on this file reflects only the local events. 2206 2207Misc 2208---- 2209 2210perf_event 2211~~~~~~~~~~ 2212 2213perf_event controller, if not mounted on a legacy hierarchy, is 2214automatically enabled on the v2 hierarchy so that perf events can 2215always be filtered by cgroup v2 path. The controller can still be 2216moved to a legacy hierarchy after v2 hierarchy is populated. 2217 2218 2219Non-normative information 2220------------------------- 2221 2222This section contains information that isn't considered to be a part of 2223the stable kernel API and so is subject to change. 2224 2225 2226CPU controller root cgroup process behaviour 2227~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~ 2228 2229When distributing CPU cycles in the root cgroup each thread in this 2230cgroup is treated as if it was hosted in a separate child cgroup of the 2231root cgroup. This child cgroup weight is dependent on its thread nice 2232level. 2233 2234For details of this mapping see sched_prio_to_weight array in 2235kernel/sched/core.c file (values from this array should be scaled 2236appropriately so the neutral - nice 0 - value is 100 instead of 1024). 2237 2238 2239IO controller root cgroup process behaviour 2240~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~ 2241 2242Root cgroup processes are hosted in an implicit leaf child node. 2243When distributing IO resources this implicit child node is taken into 2244account as if it was a normal child cgroup of the root cgroup with a 2245weight value of 200. 2246 2247 2248Namespace 2249========= 2250 2251Basics 2252------ 2253 2254cgroup namespace provides a mechanism to virtualize the view of the 2255"/proc/$PID/cgroup" file and cgroup mounts. The CLONE_NEWCGROUP clone 2256flag can be used with clone(2) and unshare(2) to create a new cgroup 2257namespace. The process running inside the cgroup namespace will have 2258its "/proc/$PID/cgroup" output restricted to cgroupns root. The 2259cgroupns root is the cgroup of the process at the time of creation of 2260the cgroup namespace. 2261 2262Without cgroup namespace, the "/proc/$PID/cgroup" file shows the 2263complete path of the cgroup of a process. In a container setup where 2264a set of cgroups and namespaces are intended to isolate processes the 2265"/proc/$PID/cgroup" file may leak potential system level information 2266to the isolated processes. For Example:: 2267 2268 # cat /proc/self/cgroup 2269 0::/batchjobs/container_id1 2270 2271The path '/batchjobs/container_id1' can be considered as system-data 2272and undesirable to expose to the isolated processes. cgroup namespace 2273can be used to restrict visibility of this path. For example, before 2274creating a cgroup namespace, one would see:: 2275 2276 # ls -l /proc/self/ns/cgroup 2277 lrwxrwxrwx 1 root root 0 2014-07-15 10:37 /proc/self/ns/cgroup -> cgroup:[4026531835] 2278 # cat /proc/self/cgroup 2279 0::/batchjobs/container_id1 2280 2281After unsharing a new namespace, the view changes:: 2282 2283 # ls -l /proc/self/ns/cgroup 2284 lrwxrwxrwx 1 root root 0 2014-07-15 10:35 /proc/self/ns/cgroup -> cgroup:[4026532183] 2285 # cat /proc/self/cgroup 2286 0::/ 2287 2288When some thread from a multi-threaded process unshares its cgroup 2289namespace, the new cgroupns gets applied to the entire process (all 2290the threads). This is natural for the v2 hierarchy; however, for the 2291legacy hierarchies, this may be unexpected. 2292 2293A cgroup namespace is alive as long as there are processes inside or 2294mounts pinning it. When the last usage goes away, the cgroup 2295namespace is destroyed. The cgroupns root and the actual cgroups 2296remain. 2297 2298 2299The Root and Views 2300------------------ 2301 2302The 'cgroupns root' for a cgroup namespace is the cgroup in which the 2303process calling unshare(2) is running. For example, if a process in 2304/batchjobs/container_id1 cgroup calls unshare, cgroup 2305/batchjobs/container_id1 becomes the cgroupns root. For the 2306init_cgroup_ns, this is the real root ('/') cgroup. 2307 2308The cgroupns root cgroup does not change even if the namespace creator 2309process later moves to a different cgroup:: 2310 2311 # ~/unshare -c # unshare cgroupns in some cgroup 2312 # cat /proc/self/cgroup 2313 0::/ 2314 # mkdir sub_cgrp_1 2315 # echo 0 > sub_cgrp_1/cgroup.procs 2316 # cat /proc/self/cgroup 2317 0::/sub_cgrp_1 2318 2319Each process gets its namespace-specific view of "/proc/$PID/cgroup" 2320 2321Processes running inside the cgroup namespace will be able to see 2322cgroup paths (in /proc/self/cgroup) only inside their root cgroup. 2323From within an unshared cgroupns:: 2324 2325 # sleep 100000 & 2326 [1] 7353 2327 # echo 7353 > sub_cgrp_1/cgroup.procs 2328 # cat /proc/7353/cgroup 2329 0::/sub_cgrp_1 2330 2331From the initial cgroup namespace, the real cgroup path will be 2332visible:: 2333 2334 $ cat /proc/7353/cgroup 2335 0::/batchjobs/container_id1/sub_cgrp_1 2336 2337From a sibling cgroup namespace (that is, a namespace rooted at a 2338different cgroup), the cgroup path relative to its own cgroup 2339namespace root will be shown. For instance, if PID 7353's cgroup 2340namespace root is at '/batchjobs/container_id2', then it will see:: 2341 2342 # cat /proc/7353/cgroup 2343 0::/../container_id2/sub_cgrp_1 2344 2345Note that the relative path always starts with '/' to indicate that 2346its relative to the cgroup namespace root of the caller. 2347 2348 2349Migration and setns(2) 2350---------------------- 2351 2352Processes inside a cgroup namespace can move into and out of the 2353namespace root if they have proper access to external cgroups. For 2354example, from inside a namespace with cgroupns root at 2355/batchjobs/container_id1, and assuming that the global hierarchy is 2356still accessible inside cgroupns:: 2357 2358 # cat /proc/7353/cgroup 2359 0::/sub_cgrp_1 2360 # echo 7353 > batchjobs/container_id2/cgroup.procs 2361 # cat /proc/7353/cgroup 2362 0::/../container_id2 2363 2364Note that this kind of setup is not encouraged. A task inside cgroup 2365namespace should only be exposed to its own cgroupns hierarchy. 2366 2367setns(2) to another cgroup namespace is allowed when: 2368 2369(a) the process has CAP_SYS_ADMIN against its current user namespace 2370(b) the process has CAP_SYS_ADMIN against the target cgroup 2371 namespace's userns 2372 2373No implicit cgroup changes happen with attaching to another cgroup 2374namespace. It is expected that the someone moves the attaching 2375process under the target cgroup namespace root. 2376 2377 2378Interaction with Other Namespaces 2379--------------------------------- 2380 2381Namespace specific cgroup hierarchy can be mounted by a process 2382running inside a non-init cgroup namespace:: 2383 2384 # mount -t cgroup2 none $MOUNT_POINT 2385 2386This will mount the unified cgroup hierarchy with cgroupns root as the 2387filesystem root. The process needs CAP_SYS_ADMIN against its user and 2388mount namespaces. 2389 2390The virtualization of /proc/self/cgroup file combined with restricting 2391the view of cgroup hierarchy by namespace-private cgroupfs mount 2392provides a properly isolated cgroup view inside the container. 2393 2394 2395Information on Kernel Programming 2396================================= 2397 2398This section contains kernel programming information in the areas 2399where interacting with cgroup is necessary. cgroup core and 2400controllers are not covered. 2401 2402 2403Filesystem Support for Writeback 2404-------------------------------- 2405 2406A filesystem can support cgroup writeback by updating 2407address_space_operations->writepage[s]() to annotate bio's using the 2408following two functions. 2409 2410 wbc_init_bio(@wbc, @bio) 2411 Should be called for each bio carrying writeback data and 2412 associates the bio with the inode's owner cgroup and the 2413 corresponding request queue. This must be called after 2414 a queue (device) has been associated with the bio and 2415 before submission. 2416 2417 wbc_account_cgroup_owner(@wbc, @page, @bytes) 2418 Should be called for each data segment being written out. 2419 While this function doesn't care exactly when it's called 2420 during the writeback session, it's the easiest and most 2421 natural to call it as data segments are added to a bio. 2422 2423With writeback bio's annotated, cgroup support can be enabled per 2424super_block by setting SB_I_CGROUPWB in ->s_iflags. This allows for 2425selective disabling of cgroup writeback support which is helpful when 2426certain filesystem features, e.g. journaled data mode, are 2427incompatible. 2428 2429wbc_init_bio() binds the specified bio to its cgroup. Depending on 2430the configuration, the bio may be executed at a lower priority and if 2431the writeback session is holding shared resources, e.g. a journal 2432entry, may lead to priority inversion. There is no one easy solution 2433for the problem. Filesystems can try to work around specific problem 2434cases by skipping wbc_init_bio() and using bio_associate_blkg() 2435directly. 2436 2437 2438Deprecated v1 Core Features 2439=========================== 2440 2441- Multiple hierarchies including named ones are not supported. 2442 2443- All v1 mount options are not supported. 2444 2445- The "tasks" file is removed and "cgroup.procs" is not sorted. 2446 2447- "cgroup.clone_children" is removed. 2448 2449- /proc/cgroups is meaningless for v2. Use "cgroup.controllers" file 2450 at the root instead. 2451 2452 2453Issues with v1 and Rationales for v2 2454==================================== 2455 2456Multiple Hierarchies 2457-------------------- 2458 2459cgroup v1 allowed an arbitrary number of hierarchies and each 2460hierarchy could host any number of controllers. While this seemed to 2461provide a high level of flexibility, it wasn't useful in practice. 2462 2463For example, as there is only one instance of each controller, utility 2464type controllers such as freezer which can be useful in all 2465hierarchies could only be used in one. The issue is exacerbated by 2466the fact that controllers couldn't be moved to another hierarchy once 2467hierarchies were populated. Another issue was that all controllers 2468bound to a hierarchy were forced to have exactly the same view of the 2469hierarchy. It wasn't possible to vary the granularity depending on 2470the specific controller. 2471 2472In practice, these issues heavily limited which controllers could be 2473put on the same hierarchy and most configurations resorted to putting 2474each controller on its own hierarchy. Only closely related ones, such 2475as the cpu and cpuacct controllers, made sense to be put on the same 2476hierarchy. This often meant that userland ended up managing multiple 2477similar hierarchies repeating the same steps on each hierarchy 2478whenever a hierarchy management operation was necessary. 2479 2480Furthermore, support for multiple hierarchies came at a steep cost. 2481It greatly complicated cgroup core implementation but more importantly 2482the support for multiple hierarchies restricted how cgroup could be 2483used in general and what controllers was able to do. 2484 2485There was no limit on how many hierarchies there might be, which meant 2486that a thread's cgroup membership couldn't be described in finite 2487length. The key might contain any number of entries and was unlimited 2488in length, which made it highly awkward to manipulate and led to 2489addition of controllers which existed only to identify membership, 2490which in turn exacerbated the original problem of proliferating number 2491of hierarchies. 2492 2493Also, as a controller couldn't have any expectation regarding the 2494topologies of hierarchies other controllers might be on, each 2495controller had to assume that all other controllers were attached to 2496completely orthogonal hierarchies. This made it impossible, or at 2497least very cumbersome, for controllers to cooperate with each other. 2498 2499In most use cases, putting controllers on hierarchies which are 2500completely orthogonal to each other isn't necessary. What usually is 2501called for is the ability to have differing levels of granularity 2502depending on the specific controller. In other words, hierarchy may 2503be collapsed from leaf towards root when viewed from specific 2504controllers. For example, a given configuration might not care about 2505how memory is distributed beyond a certain level while still wanting 2506to control how CPU cycles are distributed. 2507 2508 2509Thread Granularity 2510------------------ 2511 2512cgroup v1 allowed threads of a process to belong to different cgroups. 2513This didn't make sense for some controllers and those controllers 2514ended up implementing different ways to ignore such situations but 2515much more importantly it blurred the line between API exposed to 2516individual applications and system management interface. 2517 2518Generally, in-process knowledge is available only to the process 2519itself; thus, unlike service-level organization of processes, 2520categorizing threads of a process requires active participation from 2521the application which owns the target process. 2522 2523cgroup v1 had an ambiguously defined delegation model which got abused 2524in combination with thread granularity. cgroups were delegated to 2525individual applications so that they can create and manage their own 2526sub-hierarchies and control resource distributions along them. This 2527effectively raised cgroup to the status of a syscall-like API exposed 2528to lay programs. 2529 2530First of all, cgroup has a fundamentally inadequate interface to be 2531exposed this way. For a process to access its own knobs, it has to 2532extract the path on the target hierarchy from /proc/self/cgroup, 2533construct the path by appending the name of the knob to the path, open 2534and then read and/or write to it. This is not only extremely clunky 2535and unusual but also inherently racy. There is no conventional way to 2536define transaction across the required steps and nothing can guarantee 2537that the process would actually be operating on its own sub-hierarchy. 2538 2539cgroup controllers implemented a number of knobs which would never be 2540accepted as public APIs because they were just adding control knobs to 2541system-management pseudo filesystem. cgroup ended up with interface 2542knobs which were not properly abstracted or refined and directly 2543revealed kernel internal details. These knobs got exposed to 2544individual applications through the ill-defined delegation mechanism 2545effectively abusing cgroup as a shortcut to implementing public APIs 2546without going through the required scrutiny. 2547 2548This was painful for both userland and kernel. Userland ended up with 2549misbehaving and poorly abstracted interfaces and kernel exposing and 2550locked into constructs inadvertently. 2551 2552 2553Competition Between Inner Nodes and Threads 2554------------------------------------------- 2555 2556cgroup v1 allowed threads to be in any cgroups which created an 2557interesting problem where threads belonging to a parent cgroup and its 2558children cgroups competed for resources. This was nasty as two 2559different types of entities competed and there was no obvious way to 2560settle it. Different controllers did different things. 2561 2562The cpu controller considered threads and cgroups as equivalents and 2563mapped nice levels to cgroup weights. This worked for some cases but 2564fell flat when children wanted to be allocated specific ratios of CPU 2565cycles and the number of internal threads fluctuated - the ratios 2566constantly changed as the number of competing entities fluctuated. 2567There also were other issues. The mapping from nice level to weight 2568wasn't obvious or universal, and there were various other knobs which 2569simply weren't available for threads. 2570 2571The io controller implicitly created a hidden leaf node for each 2572cgroup to host the threads. The hidden leaf had its own copies of all 2573the knobs with ``leaf_`` prefixed. While this allowed equivalent 2574control over internal threads, it was with serious drawbacks. It 2575always added an extra layer of nesting which wouldn't be necessary 2576otherwise, made the interface messy and significantly complicated the 2577implementation. 2578 2579The memory controller didn't have a way to control what happened 2580between internal tasks and child cgroups and the behavior was not 2581clearly defined. There were attempts to add ad-hoc behaviors and 2582knobs to tailor the behavior to specific workloads which would have 2583led to problems extremely difficult to resolve in the long term. 2584 2585Multiple controllers struggled with internal tasks and came up with 2586different ways to deal with it; unfortunately, all the approaches were 2587severely flawed and, furthermore, the widely different behaviors 2588made cgroup as a whole highly inconsistent. 2589 2590This clearly is a problem which needs to be addressed from cgroup core 2591in a uniform way. 2592 2593 2594Other Interface Issues 2595---------------------- 2596 2597cgroup v1 grew without oversight and developed a large number of 2598idiosyncrasies and inconsistencies. One issue on the cgroup core side 2599was how an empty cgroup was notified - a userland helper binary was 2600forked and executed for each event. The event delivery wasn't 2601recursive or delegatable. The limitations of the mechanism also led 2602to in-kernel event delivery filtering mechanism further complicating 2603the interface. 2604 2605Controller interfaces were problematic too. An extreme example is 2606controllers completely ignoring hierarchical organization and treating 2607all cgroups as if they were all located directly under the root 2608cgroup. Some controllers exposed a large amount of inconsistent 2609implementation details to userland. 2610 2611There also was no consistency across controllers. When a new cgroup 2612was created, some controllers defaulted to not imposing extra 2613restrictions while others disallowed any resource usage until 2614explicitly configured. Configuration knobs for the same type of 2615control used widely differing naming schemes and formats. Statistics 2616and information knobs were named arbitrarily and used different 2617formats and units even in the same controller. 2618 2619cgroup v2 establishes common conventions where appropriate and updates 2620controllers so that they expose minimal and consistent interfaces. 2621 2622 2623Controller Issues and Remedies 2624------------------------------ 2625 2626Memory 2627~~~~~~ 2628 2629The original lower boundary, the soft limit, is defined as a limit 2630that is per default unset. As a result, the set of cgroups that 2631global reclaim prefers is opt-in, rather than opt-out. The costs for 2632optimizing these mostly negative lookups are so high that the 2633implementation, despite its enormous size, does not even provide the 2634basic desirable behavior. First off, the soft limit has no 2635hierarchical meaning. All configured groups are organized in a global 2636rbtree and treated like equal peers, regardless where they are located 2637in the hierarchy. This makes subtree delegation impossible. Second, 2638the soft limit reclaim pass is so aggressive that it not just 2639introduces high allocation latencies into the system, but also impacts 2640system performance due to overreclaim, to the point where the feature 2641becomes self-defeating. 2642 2643The memory.low boundary on the other hand is a top-down allocated 2644reserve. A cgroup enjoys reclaim protection when it's within its 2645effective low, which makes delegation of subtrees possible. It also 2646enjoys having reclaim pressure proportional to its overage when 2647above its effective low. 2648 2649The original high boundary, the hard limit, is defined as a strict 2650limit that can not budge, even if the OOM killer has to be called. 2651But this generally goes against the goal of making the most out of the 2652available memory. The memory consumption of workloads varies during 2653runtime, and that requires users to overcommit. But doing that with a 2654strict upper limit requires either a fairly accurate prediction of the 2655working set size or adding slack to the limit. Since working set size 2656estimation is hard and error prone, and getting it wrong results in 2657OOM kills, most users tend to err on the side of a looser limit and 2658end up wasting precious resources. 2659 2660The memory.high boundary on the other hand can be set much more 2661conservatively. When hit, it throttles allocations by forcing them 2662into direct reclaim to work off the excess, but it never invokes the 2663OOM killer. As a result, a high boundary that is chosen too 2664aggressively will not terminate the processes, but instead it will 2665lead to gradual performance degradation. The user can monitor this 2666and make corrections until the minimal memory footprint that still 2667gives acceptable performance is found. 2668 2669In extreme cases, with many concurrent allocations and a complete 2670breakdown of reclaim progress within the group, the high boundary can 2671be exceeded. But even then it's mostly better to satisfy the 2672allocation from the slack available in other groups or the rest of the 2673system than killing the group. Otherwise, memory.max is there to 2674limit this type of spillover and ultimately contain buggy or even 2675malicious applications. 2676 2677Setting the original memory.limit_in_bytes below the current usage was 2678subject to a race condition, where concurrent charges could cause the 2679limit setting to fail. memory.max on the other hand will first set the 2680limit to prevent new charges, and then reclaim and OOM kill until the 2681new limit is met - or the task writing to memory.max is killed. 2682 2683The combined memory+swap accounting and limiting is replaced by real 2684control over swap space. 2685 2686The main argument for a combined memory+swap facility in the original 2687cgroup design was that global or parental pressure would always be 2688able to swap all anonymous memory of a child group, regardless of the 2689child's own (possibly untrusted) configuration. However, untrusted 2690groups can sabotage swapping by other means - such as referencing its 2691anonymous memory in a tight loop - and an admin can not assume full 2692swappability when overcommitting untrusted jobs. 2693 2694For trusted jobs, on the other hand, a combined counter is not an 2695intuitive userspace interface, and it flies in the face of the idea 2696that cgroup controllers should account and limit specific physical 2697resources. Swap space is a resource like all others in the system, 2698and that's why unified hierarchy allows distributing it separately. 2699