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