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