1.. _cgroup-v2: 2 3================ 4Control Group v2 5================ 6 7:Date: October, 2015 8:Author: Tejun Heo <tj@kernel.org> 9 10This is the authoritative documentation on the design, interface and 11conventions of cgroup v2. It describes all userland-visible aspects 12of cgroup including core and specific controller behaviors. All 13future changes must be reflected in this document. Documentation for 14v1 is available under :ref:`Documentation/admin-guide/cgroup-v1/index.rst <cgroup-v1>`. 15 16.. CONTENTS 17 18 1. Introduction 19 1-1. Terminology 20 1-2. What is cgroup? 21 2. Basic Operations 22 2-1. Mounting 23 2-2. Organizing Processes and Threads 24 2-2-1. Processes 25 2-2-2. Threads 26 2-3. [Un]populated Notification 27 2-4. Controlling Controllers 28 2-4-1. Enabling and Disabling 29 2-4-2. Top-down Constraint 30 2-4-3. No Internal Process Constraint 31 2-5. Delegation 32 2-5-1. Model of Delegation 33 2-5-2. Delegation Containment 34 2-6. Guidelines 35 2-6-1. Organize Once and Control 36 2-6-2. Avoid Name Collisions 37 3. Resource Distribution Models 38 3-1. Weights 39 3-2. Limits 40 3-3. Protections 41 3-4. Allocations 42 4. Interface Files 43 4-1. Format 44 4-2. Conventions 45 4-3. Core Interface Files 46 5. Controllers 47 5-1. CPU 48 5-1-1. CPU Interface Files 49 5-2. Memory 50 5-2-1. Memory Interface Files 51 5-2-2. Usage Guidelines 52 5-2-3. Memory Ownership 53 5-3. IO 54 5-3-1. IO Interface Files 55 5-3-2. Writeback 56 5-3-3. IO Latency 57 5-3-3-1. How IO Latency Throttling Works 58 5-3-3-2. IO Latency Interface Files 59 5-3-4. IO Priority 60 5-4. PID 61 5-4-1. PID Interface Files 62 5-5. Cpuset 63 5.5-1. Cpuset Interface Files 64 5-6. Device 65 5-7. RDMA 66 5-7-1. RDMA Interface Files 67 5-8. HugeTLB 68 5.8-1. HugeTLB Interface Files 69 5-9. Misc 70 5.9-1 Miscellaneous cgroup Interface Files 71 5.9-2 Migration and Ownership 72 5-10. Others 73 5-10-1. perf_event 74 5-N. Non-normative information 75 5-N-1. CPU controller root cgroup process behaviour 76 5-N-2. IO controller root cgroup process behaviour 77 6. Namespace 78 6-1. Basics 79 6-2. The Root and Views 80 6-3. Migration and setns(2) 81 6-4. Interaction with Other Namespaces 82 P. Information on Kernel Programming 83 P-1. Filesystem Support for Writeback 84 D. Deprecated v1 Core Features 85 R. Issues with v1 and Rationales for v2 86 R-1. Multiple Hierarchies 87 R-2. Thread Granularity 88 R-3. Competition Between Inner Nodes and Threads 89 R-4. Other Interface Issues 90 R-5. Controller Issues and Remedies 91 R-5-1. Memory 92 93 94Introduction 95============ 96 97Terminology 98----------- 99 100"cgroup" stands for "control group" and is never capitalized. The 101singular form is used to designate the whole feature and also as a 102qualifier as in "cgroup controllers". When explicitly referring to 103multiple individual control groups, the plural form "cgroups" is used. 104 105 106What is cgroup? 107--------------- 108 109cgroup is a mechanism to organize processes hierarchically and 110distribute system resources along the hierarchy in a controlled and 111configurable manner. 112 113cgroup is largely composed of two parts - the core and controllers. 114cgroup core is primarily responsible for hierarchically organizing 115processes. A cgroup controller is usually responsible for 116distributing a specific type of system resource along the hierarchy 117although there are utility controllers which serve purposes other than 118resource distribution. 119 120cgroups form a tree structure and every process in the system belongs 121to one and only one cgroup. All threads of a process belong to the 122same cgroup. On creation, all processes are put in the cgroup that 123the parent process belongs to at the time. A process can be migrated 124to another cgroup. Migration of a process doesn't affect already 125existing descendant processes. 126 127Following certain structural constraints, controllers may be enabled or 128disabled selectively on a cgroup. All controller behaviors are 129hierarchical - if a controller is enabled on a cgroup, it affects all 130processes which belong to the cgroups consisting the inclusive 131sub-hierarchy of the cgroup. When a controller is enabled on a nested 132cgroup, it always restricts the resource distribution further. The 133restrictions set closer to the root in the hierarchy can not be 134overridden from further away. 135 136 137Basic Operations 138================ 139 140Mounting 141-------- 142 143Unlike v1, cgroup v2 has only single hierarchy. The cgroup v2 144hierarchy can be mounted with the following mount command:: 145 146 # mount -t cgroup2 none $MOUNT_POINT 147 148cgroup2 filesystem has the magic number 0x63677270 ("cgrp"). All 149controllers which support v2 and are not bound to a v1 hierarchy are 150automatically bound to the v2 hierarchy and show up at the root. 151Controllers which are not in active use in the v2 hierarchy can be 152bound to other hierarchies. This allows mixing v2 hierarchy with the 153legacy v1 multiple hierarchies in a fully backward compatible way. 154 155A controller can be moved across hierarchies only after the controller 156is no longer referenced in its current hierarchy. Because per-cgroup 157controller states are destroyed asynchronously and controllers may 158have lingering references, a controller may not show up immediately on 159the v2 hierarchy after the final umount of the previous hierarchy. 160Similarly, a controller should be fully disabled to be moved out of 161the unified hierarchy and it may take some time for the disabled 162controller to become available for other hierarchies; furthermore, due 163to inter-controller dependencies, other controllers may need to be 164disabled too. 165 166While useful for development and manual configurations, moving 167controllers dynamically between the v2 and other hierarchies is 168strongly discouraged for production use. It is recommended to decide 169the hierarchies and controller associations before starting using the 170controllers after system boot. 171 172During transition to v2, system management software might still 173automount the v1 cgroup filesystem and so hijack all controllers 174during boot, before manual intervention is possible. To make testing 175and experimenting easier, the kernel parameter cgroup_no_v1= allows 176disabling controllers in v1 and make them always available in v2. 177 178cgroup v2 currently supports the following mount options. 179 180 nsdelegate 181 Consider cgroup namespaces as delegation boundaries. This 182 option is system wide and can only be set on mount or modified 183 through remount from the init namespace. The mount option is 184 ignored on non-init namespace mounts. Please refer to the 185 Delegation section for details. 186 187 memory_localevents 188 Only populate memory.events with data for the current cgroup, 189 and not any subtrees. This is legacy behaviour, the default 190 behaviour without this option is to include subtree counts. 191 This option is system wide and can only be set on mount or 192 modified through remount from the init namespace. The mount 193 option is ignored on non-init namespace mounts. 194 195 memory_recursiveprot 196 Recursively apply memory.min and memory.low protection to 197 entire subtrees, without requiring explicit downward 198 propagation into leaf cgroups. This allows protecting entire 199 subtrees from one another, while retaining free competition 200 within those subtrees. This should have been the default 201 behavior but is a mount-option to avoid regressing setups 202 relying on the original semantics (e.g. specifying bogusly 203 high 'bypass' protection values at higher tree levels). 204 205 206Organizing Processes and Threads 207-------------------------------- 208 209Processes 210~~~~~~~~~ 211 212Initially, only the root cgroup exists to which all processes belong. 213A child cgroup can be created by creating a sub-directory:: 214 215 # mkdir $CGROUP_NAME 216 217A given cgroup may have multiple child cgroups forming a tree 218structure. Each cgroup has a read-writable interface file 219"cgroup.procs". When read, it lists the PIDs of all processes which 220belong to the cgroup one-per-line. The PIDs are not ordered and the 221same PID may show up more than once if the process got moved to 222another cgroup and then back or the PID got recycled while reading. 223 224A process can be migrated into a cgroup by writing its PID to the 225target cgroup's "cgroup.procs" file. Only one process can be migrated 226on a single write(2) call. If a process is composed of multiple 227threads, writing the PID of any thread migrates all threads of the 228process. 229 230When a process forks a child process, the new process is born into the 231cgroup that the forking process belongs to at the time of the 232operation. After exit, a process stays associated with the cgroup 233that it belonged to at the time of exit until it's reaped; however, a 234zombie process does not appear in "cgroup.procs" and thus can't be 235moved to another cgroup. 236 237A cgroup which doesn't have any children or live processes can be 238destroyed by removing the directory. Note that a cgroup which doesn't 239have any children and is associated only with zombie processes is 240considered empty and can be removed:: 241 242 # rmdir $CGROUP_NAME 243 244"/proc/$PID/cgroup" lists a process's cgroup membership. If legacy 245cgroup is in use in the system, this file may contain multiple lines, 246one for each hierarchy. The entry for cgroup v2 is always in the 247format "0::$PATH":: 248 249 # cat /proc/842/cgroup 250 ... 251 0::/test-cgroup/test-cgroup-nested 252 253If the process becomes a zombie and the cgroup it was associated with 254is removed subsequently, " (deleted)" is appended to the path:: 255 256 # cat /proc/842/cgroup 257 ... 258 0::/test-cgroup/test-cgroup-nested (deleted) 259 260 261Threads 262~~~~~~~ 263 264cgroup v2 supports thread granularity for a subset of controllers to 265support use cases requiring hierarchical resource distribution across 266the threads of a group of processes. By default, all threads of a 267process belong to the same cgroup, which also serves as the resource 268domain to host resource consumptions which are not specific to a 269process or thread. The thread mode allows threads to be spread across 270a subtree while still maintaining the common resource domain for them. 271 272Controllers which support thread mode are called threaded controllers. 273The ones which don't are called domain controllers. 274 275Marking a cgroup threaded makes it join the resource domain of its 276parent as a threaded cgroup. The parent may be another threaded 277cgroup whose resource domain is further up in the hierarchy. The root 278of a threaded subtree, that is, the nearest ancestor which is not 279threaded, is called threaded domain or thread root interchangeably and 280serves as the resource domain for the entire subtree. 281 282Inside a threaded subtree, threads of a process can be put in 283different cgroups and are not subject to the no internal process 284constraint - threaded controllers can be enabled on non-leaf cgroups 285whether they have threads in them or not. 286 287As the threaded domain cgroup hosts all the domain resource 288consumptions of the subtree, it is considered to have internal 289resource consumptions whether there are processes in it or not and 290can't have populated child cgroups which aren't threaded. Because the 291root cgroup is not subject to no internal process constraint, it can 292serve both as a threaded domain and a parent to domain cgroups. 293 294The current operation mode or type of the cgroup is shown in the 295"cgroup.type" file which indicates whether the cgroup is a normal 296domain, a domain which is serving as the domain of a threaded subtree, 297or a threaded cgroup. 298 299On creation, a cgroup is always a domain cgroup and can be made 300threaded by writing "threaded" to the "cgroup.type" file. The 301operation is single direction:: 302 303 # echo threaded > cgroup.type 304 305Once threaded, the cgroup can't be made a domain again. To enable the 306thread mode, the following conditions must be met. 307 308- As the cgroup will join the parent's resource domain. The parent 309 must either be a valid (threaded) domain or a threaded cgroup. 310 311- When the parent is an unthreaded domain, it must not have any domain 312 controllers enabled or populated domain children. The root is 313 exempt from this requirement. 314 315Topology-wise, a cgroup can be in an invalid state. Please consider 316the following topology:: 317 318 A (threaded domain) - B (threaded) - C (domain, just created) 319 320C is created as a domain but isn't connected to a parent which can 321host child domains. C can't be used until it is turned into a 322threaded cgroup. "cgroup.type" file will report "domain (invalid)" in 323these cases. Operations which fail due to invalid topology use 324EOPNOTSUPP as the errno. 325 326A domain cgroup is turned into a threaded domain when one of its child 327cgroup becomes threaded or threaded controllers are enabled in the 328"cgroup.subtree_control" file while there are processes in the cgroup. 329A threaded domain reverts to a normal domain when the conditions 330clear. 331 332When read, "cgroup.threads" contains the list of the thread IDs of all 333threads in the cgroup. Except that the operations are per-thread 334instead of per-process, "cgroup.threads" has the same format and 335behaves the same way as "cgroup.procs". While "cgroup.threads" can be 336written to in any cgroup, as it can only move threads inside the same 337threaded domain, its operations are confined inside each threaded 338subtree. 339 340The threaded domain cgroup serves as the resource domain for the whole 341subtree, and, while the threads can be scattered across the subtree, 342all the processes are considered to be in the threaded domain cgroup. 343"cgroup.procs" in a threaded domain cgroup contains the PIDs of all 344processes in the subtree and is not readable in the subtree proper. 345However, "cgroup.procs" can be written to from anywhere in the subtree 346to migrate all threads of the matching process to the cgroup. 347 348Only threaded controllers can be enabled in a threaded subtree. When 349a threaded controller is enabled inside a threaded subtree, it only 350accounts for and controls resource consumptions associated with the 351threads in the cgroup and its descendants. All consumptions which 352aren't tied to a specific thread belong to the threaded domain cgroup. 353 354Because a threaded subtree is exempt from no internal process 355constraint, a threaded controller must be able to handle competition 356between threads in a non-leaf cgroup and its child cgroups. Each 357threaded controller defines how such competitions are handled. 358 359 360[Un]populated Notification 361-------------------------- 362 363Each non-root cgroup has a "cgroup.events" file which contains 364"populated" field indicating whether the cgroup's sub-hierarchy has 365live processes in it. Its value is 0 if there is no live process in 366the cgroup and its descendants; otherwise, 1. poll and [id]notify 367events are triggered when the value changes. This can be used, for 368example, to start a clean-up operation after all processes of a given 369sub-hierarchy have exited. The populated state updates and 370notifications are recursive. Consider the following sub-hierarchy 371where the numbers in the parentheses represent the numbers of processes 372in each cgroup:: 373 374 A(4) - B(0) - C(1) 375 \ D(0) 376 377A, B and C's "populated" fields would be 1 while D's 0. After the one 378process in C exits, B and C's "populated" fields would flip to "0" and 379file modified events will be generated on the "cgroup.events" files of 380both cgroups. 381 382 383Controlling Controllers 384----------------------- 385 386Enabling and Disabling 387~~~~~~~~~~~~~~~~~~~~~~ 388 389Each cgroup has a "cgroup.controllers" file which lists all 390controllers available for the cgroup to enable:: 391 392 # cat cgroup.controllers 393 cpu io memory 394 395No controller is enabled by default. Controllers can be enabled and 396disabled by writing to the "cgroup.subtree_control" file:: 397 398 # echo "+cpu +memory -io" > cgroup.subtree_control 399 400Only controllers which are listed in "cgroup.controllers" can be 401enabled. When multiple operations are specified as above, either they 402all succeed or fail. If multiple operations on the same controller 403are specified, the last one is effective. 404 405Enabling a controller in a cgroup indicates that the distribution of 406the target resource across its immediate children will be controlled. 407Consider the following sub-hierarchy. The enabled controllers are 408listed in parentheses:: 409 410 A(cpu,memory) - B(memory) - C() 411 \ D() 412 413As A has "cpu" and "memory" enabled, A will control the distribution 414of CPU cycles and memory to its children, in this case, B. As B has 415"memory" enabled but not "CPU", C and D will compete freely on CPU 416cycles but their division of memory available to B will be controlled. 417 418As a controller regulates the distribution of the target resource to 419the cgroup's children, enabling it creates the controller's interface 420files in the child cgroups. In the above example, enabling "cpu" on B 421would create the "cpu." prefixed controller interface files in C and 422D. Likewise, disabling "memory" from B would remove the "memory." 423prefixed controller interface files from C and D. This means that the 424controller interface files - anything which doesn't start with 425"cgroup." are owned by the parent rather than the cgroup itself. 426 427 428Top-down Constraint 429~~~~~~~~~~~~~~~~~~~ 430 431Resources are distributed top-down and a cgroup can further distribute 432a resource only if the resource has been distributed to it from the 433parent. This means that all non-root "cgroup.subtree_control" files 434can only contain controllers which are enabled in the parent's 435"cgroup.subtree_control" file. A controller can be enabled only if 436the parent has the controller enabled and a controller can't be 437disabled if one or more children have it enabled. 438 439 440No Internal Process Constraint 441~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~ 442 443Non-root cgroups can distribute domain resources to their children 444only when they don't have any processes of their own. In other words, 445only domain cgroups which don't contain any processes can have domain 446controllers enabled in their "cgroup.subtree_control" files. 447 448This guarantees that, when a domain controller is looking at the part 449of the hierarchy which has it enabled, processes are always only on 450the leaves. This rules out situations where child cgroups compete 451against internal processes of the parent. 452 453The root cgroup is exempt from this restriction. Root contains 454processes and anonymous resource consumption which can't be associated 455with any other cgroups and requires special treatment from most 456controllers. How resource consumption in the root cgroup is governed 457is up to each controller (for more information on this topic please 458refer to the Non-normative information section in the Controllers 459chapter). 460 461Note that the restriction doesn't get in the way if there is no 462enabled controller in the cgroup's "cgroup.subtree_control". This is 463important as otherwise it wouldn't be possible to create children of a 464populated cgroup. To control resource distribution of a cgroup, the 465cgroup must create children and transfer all its processes to the 466children before enabling controllers in its "cgroup.subtree_control" 467file. 468 469 470Delegation 471---------- 472 473Model of Delegation 474~~~~~~~~~~~~~~~~~~~ 475 476A cgroup can be delegated in two ways. First, to a less privileged 477user by granting write access of the directory and its "cgroup.procs", 478"cgroup.threads" and "cgroup.subtree_control" files to the user. 479Second, if the "nsdelegate" mount option is set, automatically to a 480cgroup namespace on namespace creation. 481 482Because the resource control interface files in a given directory 483control the distribution of the parent's resources, the delegatee 484shouldn't be allowed to write to them. For the first method, this is 485achieved by not granting access to these files. For the second, the 486kernel rejects writes to all files other than "cgroup.procs" and 487"cgroup.subtree_control" on a namespace root from inside the 488namespace. 489 490The end results are equivalent for both delegation types. Once 491delegated, the user can build sub-hierarchy under the directory, 492organize processes inside it as it sees fit and further distribute the 493resources it received from the parent. The limits and other settings 494of all resource controllers are hierarchical and regardless of what 495happens in the delegated sub-hierarchy, nothing can escape the 496resource restrictions imposed by the parent. 497 498Currently, cgroup doesn't impose any restrictions on the number of 499cgroups in or nesting depth of a delegated sub-hierarchy; however, 500this may be limited explicitly in the future. 501 502 503Delegation Containment 504~~~~~~~~~~~~~~~~~~~~~~ 505 506A delegated sub-hierarchy is contained in the sense that processes 507can't be moved into or out of the sub-hierarchy by the delegatee. 508 509For delegations to a less privileged user, this is achieved by 510requiring the following conditions for a process with a non-root euid 511to migrate a target process into a cgroup by writing its PID to the 512"cgroup.procs" file. 513 514- The writer must have write access to the "cgroup.procs" file. 515 516- The writer must have write access to the "cgroup.procs" file of the 517 common ancestor of the source and destination cgroups. 518 519The above two constraints ensure that while a delegatee may migrate 520processes around freely in the delegated sub-hierarchy it can't pull 521in from or push out to outside the sub-hierarchy. 522 523For an example, let's assume cgroups C0 and C1 have been delegated to 524user U0 who created C00, C01 under C0 and C10 under C1 as follows and 525all processes under C0 and C1 belong to U0:: 526 527 ~~~~~~~~~~~~~ - C0 - C00 528 ~ cgroup ~ \ C01 529 ~ hierarchy ~ 530 ~~~~~~~~~~~~~ - C1 - C10 531 532Let's also say U0 wants to write the PID of a process which is 533currently in C10 into "C00/cgroup.procs". U0 has write access to the 534file; however, the common ancestor of the source cgroup C10 and the 535destination cgroup C00 is above the points of delegation and U0 would 536not have write access to its "cgroup.procs" files and thus the write 537will be denied with -EACCES. 538 539For delegations to namespaces, containment is achieved by requiring 540that both the source and destination cgroups are reachable from the 541namespace of the process which is attempting the migration. If either 542is not reachable, the migration is rejected with -ENOENT. 543 544 545Guidelines 546---------- 547 548Organize Once and Control 549~~~~~~~~~~~~~~~~~~~~~~~~~ 550 551Migrating a process across cgroups is a relatively expensive operation 552and stateful resources such as memory are not moved together with the 553process. This is an explicit design decision as there often exist 554inherent trade-offs between migration and various hot paths in terms 555of synchronization cost. 556 557As such, migrating processes across cgroups frequently as a means to 558apply different resource restrictions is discouraged. A workload 559should be assigned to a cgroup according to the system's logical and 560resource structure once on start-up. Dynamic adjustments to resource 561distribution can be made by changing controller configuration through 562the interface files. 563 564 565Avoid Name Collisions 566~~~~~~~~~~~~~~~~~~~~~ 567 568Interface files for a cgroup and its children cgroups occupy the same 569directory and it is possible to create children cgroups which collide 570with interface files. 571 572All cgroup core interface files are prefixed with "cgroup." and each 573controller's interface files are prefixed with the controller name and 574a dot. A controller's name is composed of lower case alphabets and 575'_'s but never begins with an '_' so it can be used as the prefix 576character for collision avoidance. Also, interface file names won't 577start or end with terms which are often used in categorizing workloads 578such as job, service, slice, unit or workload. 579 580cgroup doesn't do anything to prevent name collisions and it's the 581user's responsibility to avoid them. 582 583 584Resource Distribution Models 585============================ 586 587cgroup controllers implement several resource distribution schemes 588depending on the resource type and expected use cases. This section 589describes major schemes in use along with their expected behaviors. 590 591 592Weights 593------- 594 595A parent's resource is distributed by adding up the weights of all 596active children and giving each the fraction matching the ratio of its 597weight against the sum. As only children which can make use of the 598resource at the moment participate in the distribution, this is 599work-conserving. Due to the dynamic nature, this model is usually 600used for stateless resources. 601 602All weights are in the range [1, 10000] with the default at 100. This 603allows symmetric multiplicative biases in both directions at fine 604enough granularity while staying in the intuitive range. 605 606As long as the weight is in range, all configuration combinations are 607valid and there is no reason to reject configuration changes or 608process migrations. 609 610"cpu.weight" proportionally distributes CPU cycles to active children 611and is an example of this type. 612 613 614Limits 615------ 616 617A child can only consume upto the configured amount of the resource. 618Limits can be over-committed - the sum of the limits of children can 619exceed the amount of resource available to the parent. 620 621Limits are in the range [0, max] and defaults to "max", which is noop. 622 623As limits can be over-committed, all configuration combinations are 624valid and there is no reason to reject configuration changes or 625process migrations. 626 627"io.max" limits the maximum BPS and/or IOPS that a cgroup can consume 628on an IO device and is an example of this type. 629 630 631Protections 632----------- 633 634A cgroup is protected upto the configured amount of the resource 635as long as the usages of all its ancestors are under their 636protected levels. Protections can be hard guarantees or best effort 637soft boundaries. Protections can also be over-committed in which case 638only upto the amount available to the parent is protected among 639children. 640 641Protections are in the range [0, max] and defaults to 0, which is 642noop. 643 644As protections can be over-committed, all configuration combinations 645are valid and there is no reason to reject configuration changes or 646process migrations. 647 648"memory.low" implements best-effort memory protection and is an 649example of this type. 650 651 652Allocations 653----------- 654 655A cgroup is exclusively allocated a certain amount of a finite 656resource. Allocations can't be over-committed - the sum of the 657allocations of children can not exceed the amount of resource 658available to the parent. 659 660Allocations are in the range [0, max] and defaults to 0, which is no 661resource. 662 663As allocations can't be over-committed, some configuration 664combinations are invalid and should be rejected. Also, if the 665resource is mandatory for execution of processes, process migrations 666may be rejected. 667 668"cpu.rt.max" hard-allocates realtime slices and is an example of this 669type. 670 671 672Interface Files 673=============== 674 675Format 676------ 677 678All interface files should be in one of the following formats whenever 679possible:: 680 681 New-line separated values 682 (when only one value can be written at once) 683 684 VAL0\n 685 VAL1\n 686 ... 687 688 Space separated values 689 (when read-only or multiple values can be written at once) 690 691 VAL0 VAL1 ...\n 692 693 Flat keyed 694 695 KEY0 VAL0\n 696 KEY1 VAL1\n 697 ... 698 699 Nested keyed 700 701 KEY0 SUB_KEY0=VAL00 SUB_KEY1=VAL01... 702 KEY1 SUB_KEY0=VAL10 SUB_KEY1=VAL11... 703 ... 704 705For a writable file, the format for writing should generally match 706reading; however, controllers may allow omitting later fields or 707implement restricted shortcuts for most common use cases. 708 709For both flat and nested keyed files, only the values for a single key 710can be written at a time. For nested keyed files, the sub key pairs 711may be specified in any order and not all pairs have to be specified. 712 713 714Conventions 715----------- 716 717- Settings for a single feature should be contained in a single file. 718 719- The root cgroup should be exempt from resource control and thus 720 shouldn't have resource control interface files. 721 722- The default time unit is microseconds. If a different unit is ever 723 used, an explicit unit suffix must be present. 724 725- A parts-per quantity should use a percentage decimal with at least 726 two digit fractional part - e.g. 13.40. 727 728- If a controller implements weight based resource distribution, its 729 interface file should be named "weight" and have the range [1, 730 10000] with 100 as the default. The values are chosen to allow 731 enough and symmetric bias in both directions while keeping it 732 intuitive (the default is 100%). 733 734- If a controller implements an absolute resource guarantee and/or 735 limit, the interface files should be named "min" and "max" 736 respectively. If a controller implements best effort resource 737 guarantee and/or limit, the interface files should be named "low" 738 and "high" respectively. 739 740 In the above four control files, the special token "max" should be 741 used to represent upward infinity for both reading and writing. 742 743- If a setting has a configurable default value and keyed specific 744 overrides, the default entry should be keyed with "default" and 745 appear as the first entry in the file. 746 747 The default value can be updated by writing either "default $VAL" or 748 "$VAL". 749 750 When writing to update a specific override, "default" can be used as 751 the value to indicate removal of the override. Override entries 752 with "default" as the value must not appear when read. 753 754 For example, a setting which is keyed by major:minor device numbers 755 with integer values may look like the following:: 756 757 # cat cgroup-example-interface-file 758 default 150 759 8:0 300 760 761 The default value can be updated by:: 762 763 # echo 125 > cgroup-example-interface-file 764 765 or:: 766 767 # echo "default 125" > cgroup-example-interface-file 768 769 An override can be set by:: 770 771 # echo "8:16 170" > cgroup-example-interface-file 772 773 and cleared by:: 774 775 # echo "8:0 default" > cgroup-example-interface-file 776 # cat cgroup-example-interface-file 777 default 125 778 8:16 170 779 780- For events which are not very high frequency, an interface file 781 "events" should be created which lists event key value pairs. 782 Whenever a notifiable event happens, file modified event should be 783 generated on the file. 784 785 786Core Interface Files 787-------------------- 788 789All cgroup core files are prefixed with "cgroup." 790 791 cgroup.type 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. 955 956 cgroup.kill 957 A write-only single value file which exists in non-root cgroups. 958 The only allowed value is "1". 959 960 Writing "1" to the file causes the cgroup and all descendant cgroups to 961 be killed. This means that all processes located in the affected cgroup 962 tree will be killed via SIGKILL. 963 964 Killing a cgroup tree will deal with concurrent forks appropriately and 965 is protected against migrations. 966 967 In a threaded cgroup, writing this file fails with EOPNOTSUPP as 968 killing cgroups is a process directed operation, i.e. it affects 969 the whole thread-group. 970 971Controllers 972=========== 973 974.. _cgroup-v2-cpu: 975 976CPU 977--- 978 979The "cpu" controllers regulates distribution of CPU cycles. This 980controller implements weight and absolute bandwidth limit models for 981normal scheduling policy and absolute bandwidth allocation model for 982realtime scheduling policy. 983 984In all the above models, cycles distribution is defined only on a temporal 985base and it does not account for the frequency at which tasks are executed. 986The (optional) utilization clamping support allows to hint the schedutil 987cpufreq governor about the minimum desired frequency which should always be 988provided by a CPU, as well as the maximum desired frequency, which should not 989be exceeded by a CPU. 990 991WARNING: cgroup2 doesn't yet support control of realtime processes and 992the cpu controller can only be enabled when all RT processes are in 993the root cgroup. Be aware that system management software may already 994have placed RT processes into nonroot cgroups during the system boot 995process, and these processes may need to be moved to the root cgroup 996before the cpu controller can be enabled. 997 998 999CPU Interface Files 1000~~~~~~~~~~~~~~~~~~~ 1001 1002All time durations are in microseconds. 1003 1004 cpu.stat 1005 A read-only flat-keyed file. 1006 This file exists whether the controller is enabled or not. 1007 1008 It always reports the following three stats: 1009 1010 - usage_usec 1011 - user_usec 1012 - system_usec 1013 1014 and the following three when the controller is enabled: 1015 1016 - nr_periods 1017 - nr_throttled 1018 - throttled_usec 1019 1020 cpu.weight 1021 A read-write single value file which exists on non-root 1022 cgroups. The default is "100". 1023 1024 The weight in the range [1, 10000]. 1025 1026 cpu.weight.nice 1027 A read-write single value file which exists on non-root 1028 cgroups. The default is "0". 1029 1030 The nice value is in the range [-20, 19]. 1031 1032 This interface file is an alternative interface for 1033 "cpu.weight" and allows reading and setting weight using the 1034 same values used by nice(2). Because the range is smaller and 1035 granularity is coarser for the nice values, the read value is 1036 the closest approximation of the current weight. 1037 1038 cpu.max 1039 A read-write two value file which exists on non-root cgroups. 1040 The default is "max 100000". 1041 1042 The maximum bandwidth limit. It's in the following format:: 1043 1044 $MAX $PERIOD 1045 1046 which indicates that the group may consume upto $MAX in each 1047 $PERIOD duration. "max" for $MAX indicates no limit. If only 1048 one number is written, $MAX is updated. 1049 1050 cpu.pressure 1051 A read-write nested-keyed file. 1052 1053 Shows pressure stall information for CPU. See 1054 :ref:`Documentation/accounting/psi.rst <psi>` for details. 1055 1056 cpu.uclamp.min 1057 A read-write single value file which exists on non-root cgroups. 1058 The default is "0", i.e. no utilization boosting. 1059 1060 The requested minimum utilization (protection) as a percentage 1061 rational number, e.g. 12.34 for 12.34%. 1062 1063 This interface allows reading and setting minimum utilization clamp 1064 values similar to the sched_setattr(2). This minimum utilization 1065 value is used to clamp the task specific minimum utilization clamp. 1066 1067 The requested minimum utilization (protection) is always capped by 1068 the current value for the maximum utilization (limit), i.e. 1069 `cpu.uclamp.max`. 1070 1071 cpu.uclamp.max 1072 A read-write single value file which exists on non-root cgroups. 1073 The default is "max". i.e. no utilization capping 1074 1075 The requested maximum utilization (limit) as a percentage rational 1076 number, e.g. 98.76 for 98.76%. 1077 1078 This interface allows reading and setting maximum utilization clamp 1079 values similar to the sched_setattr(2). This maximum utilization 1080 value is used to clamp the task specific maximum utilization clamp. 1081 1082 1083 1084Memory 1085------ 1086 1087The "memory" controller regulates distribution of memory. Memory is 1088stateful and implements both limit and protection models. Due to the 1089intertwining between memory usage and reclaim pressure and the 1090stateful nature of memory, the distribution model is relatively 1091complex. 1092 1093While not completely water-tight, all major memory usages by a given 1094cgroup are tracked so that the total memory consumption can be 1095accounted and controlled to a reasonable extent. Currently, the 1096following types of memory usages are tracked. 1097 1098- Userland memory - page cache and anonymous memory. 1099 1100- Kernel data structures such as dentries and inodes. 1101 1102- TCP socket buffers. 1103 1104The above list may expand in the future for better coverage. 1105 1106 1107Memory Interface Files 1108~~~~~~~~~~~~~~~~~~~~~~ 1109 1110All memory amounts are in bytes. If a value which is not aligned to 1111PAGE_SIZE is written, the value may be rounded up to the closest 1112PAGE_SIZE multiple when read back. 1113 1114 memory.current 1115 A read-only single value file which exists on non-root 1116 cgroups. 1117 1118 The total amount of memory currently being used by the cgroup 1119 and its descendants. 1120 1121 memory.min 1122 A read-write single value file which exists on non-root 1123 cgroups. The default is "0". 1124 1125 Hard memory protection. If the memory usage of a cgroup 1126 is within its effective min boundary, the cgroup's memory 1127 won't be reclaimed under any conditions. If there is no 1128 unprotected reclaimable memory available, OOM killer 1129 is invoked. Above the effective min boundary (or 1130 effective low boundary if it is higher), pages are reclaimed 1131 proportionally to the overage, reducing reclaim pressure for 1132 smaller overages. 1133 1134 Effective min boundary is limited by memory.min values of 1135 all ancestor cgroups. If there is memory.min overcommitment 1136 (child cgroup or cgroups are requiring more protected memory 1137 than parent will allow), then each child cgroup will get 1138 the part of parent's protection proportional to its 1139 actual memory usage below memory.min. 1140 1141 Putting more memory than generally available under this 1142 protection is discouraged and may lead to constant OOMs. 1143 1144 If a memory cgroup is not populated with processes, 1145 its memory.min is ignored. 1146 1147 memory.low 1148 A read-write single value file which exists on non-root 1149 cgroups. The default is "0". 1150 1151 Best-effort memory protection. If the memory usage of a 1152 cgroup is within its effective low boundary, the cgroup's 1153 memory won't be reclaimed unless there is no reclaimable 1154 memory available in unprotected cgroups. 1155 Above the effective low boundary (or 1156 effective min boundary if it is higher), pages are reclaimed 1157 proportionally to the overage, reducing reclaim pressure for 1158 smaller overages. 1159 1160 Effective low boundary is limited by memory.low values of 1161 all ancestor cgroups. If there is memory.low overcommitment 1162 (child cgroup or cgroups are requiring more protected memory 1163 than parent will allow), then each child cgroup will get 1164 the part of parent's protection proportional to its 1165 actual memory usage below memory.low. 1166 1167 Putting more memory than generally available under this 1168 protection is discouraged. 1169 1170 memory.high 1171 A read-write single value file which exists on non-root 1172 cgroups. The default is "max". 1173 1174 Memory usage throttle limit. This is the main mechanism to 1175 control memory usage of a cgroup. If a cgroup's usage goes 1176 over the high boundary, the processes of the cgroup are 1177 throttled and put under heavy reclaim pressure. 1178 1179 Going over the high limit never invokes the OOM killer and 1180 under extreme conditions the limit may be breached. 1181 1182 memory.max 1183 A read-write single value file which exists on non-root 1184 cgroups. The default is "max". 1185 1186 Memory usage hard limit. This is the final protection 1187 mechanism. If a cgroup's memory usage reaches this limit and 1188 can't be reduced, the OOM killer is invoked in the cgroup. 1189 Under certain circumstances, the usage may go over the limit 1190 temporarily. 1191 1192 In default configuration regular 0-order allocations always 1193 succeed unless OOM killer chooses current task as a victim. 1194 1195 Some kinds of allocations don't invoke the OOM killer. 1196 Caller could retry them differently, return into userspace 1197 as -ENOMEM or silently ignore in cases like disk readahead. 1198 1199 This is the ultimate protection mechanism. As long as the 1200 high limit is used and monitored properly, this limit's 1201 utility is limited to providing the final safety net. 1202 1203 memory.oom.group 1204 A read-write single value file which exists on non-root 1205 cgroups. The default value is "0". 1206 1207 Determines whether the cgroup should be treated as 1208 an indivisible workload by the OOM killer. If set, 1209 all tasks belonging to the cgroup or to its descendants 1210 (if the memory cgroup is not a leaf cgroup) are killed 1211 together or not at all. This can be used to avoid 1212 partial kills to guarantee workload integrity. 1213 1214 Tasks with the OOM protection (oom_score_adj set to -1000) 1215 are treated as an exception and are never killed. 1216 1217 If the OOM killer is invoked in a cgroup, it's not going 1218 to kill any tasks outside of this cgroup, regardless 1219 memory.oom.group values of ancestor cgroups. 1220 1221 memory.events 1222 A read-only flat-keyed file which exists on non-root cgroups. 1223 The following entries are defined. Unless specified 1224 otherwise, a value change in this file generates a file 1225 modified event. 1226 1227 Note that all fields in this file are hierarchical and the 1228 file modified event can be generated due to an event down the 1229 hierarchy. For for the local events at the cgroup level see 1230 memory.events.local. 1231 1232 low 1233 The number of times the cgroup is reclaimed due to 1234 high memory pressure even though its usage is under 1235 the low boundary. This usually indicates that the low 1236 boundary is over-committed. 1237 1238 high 1239 The number of times processes of the cgroup are 1240 throttled and routed to perform direct memory reclaim 1241 because the high memory boundary was exceeded. For a 1242 cgroup whose memory usage is capped by the high limit 1243 rather than global memory pressure, this event's 1244 occurrences are expected. 1245 1246 max 1247 The number of times the cgroup's memory usage was 1248 about to go over the max boundary. If direct reclaim 1249 fails to bring it down, the cgroup goes to OOM state. 1250 1251 oom 1252 The number of time the cgroup's memory usage was 1253 reached the limit and allocation was about to fail. 1254 1255 This event is not raised if the OOM killer is not 1256 considered as an option, e.g. for failed high-order 1257 allocations or if caller asked to not retry attempts. 1258 1259 oom_kill 1260 The number of processes belonging to this cgroup 1261 killed by any kind of OOM killer. 1262 1263 memory.events.local 1264 Similar to memory.events but the fields in the file are local 1265 to the cgroup i.e. not hierarchical. The file modified event 1266 generated on this file reflects only the local events. 1267 1268 memory.stat 1269 A read-only flat-keyed file which exists on non-root cgroups. 1270 1271 This breaks down the cgroup's memory footprint into different 1272 types of memory, type-specific details, and other information 1273 on the state and past events of the memory management system. 1274 1275 All memory amounts are in bytes. 1276 1277 The entries are ordered to be human readable, and new entries 1278 can show up in the middle. Don't rely on items remaining in a 1279 fixed position; use the keys to look up specific values! 1280 1281 If the entry has no per-node counter (or not show in the 1282 memory.numa_stat). We use 'npn' (non-per-node) as the tag 1283 to indicate that it will not show in the memory.numa_stat. 1284 1285 anon 1286 Amount of memory used in anonymous mappings such as 1287 brk(), sbrk(), and mmap(MAP_ANONYMOUS) 1288 1289 file 1290 Amount of memory used to cache filesystem data, 1291 including tmpfs and shared memory. 1292 1293 kernel_stack 1294 Amount of memory allocated to kernel stacks. 1295 1296 pagetables 1297 Amount of memory allocated for page tables. 1298 1299 percpu (npn) 1300 Amount of memory used for storing per-cpu kernel 1301 data structures. 1302 1303 sock (npn) 1304 Amount of memory used in network transmission buffers 1305 1306 shmem 1307 Amount of cached filesystem data that is swap-backed, 1308 such as tmpfs, shm segments, shared anonymous mmap()s 1309 1310 file_mapped 1311 Amount of cached filesystem data mapped with mmap() 1312 1313 file_dirty 1314 Amount of cached filesystem data that was modified but 1315 not yet written back to disk 1316 1317 file_writeback 1318 Amount of cached filesystem data that was modified and 1319 is currently being written back to disk 1320 1321 swapcached 1322 Amount of swap cached in memory. The swapcache is accounted 1323 against both memory and swap usage. 1324 1325 anon_thp 1326 Amount of memory used in anonymous mappings backed by 1327 transparent hugepages 1328 1329 file_thp 1330 Amount of cached filesystem data backed by transparent 1331 hugepages 1332 1333 shmem_thp 1334 Amount of shm, tmpfs, shared anonymous mmap()s backed by 1335 transparent hugepages 1336 1337 inactive_anon, active_anon, inactive_file, active_file, unevictable 1338 Amount of memory, swap-backed and filesystem-backed, 1339 on the internal memory management lists used by the 1340 page reclaim algorithm. 1341 1342 As these represent internal list state (eg. shmem pages are on anon 1343 memory management lists), inactive_foo + active_foo may not be equal to 1344 the value for the foo counter, since the foo counter is type-based, not 1345 list-based. 1346 1347 slab_reclaimable 1348 Part of "slab" that might be reclaimed, such as 1349 dentries and inodes. 1350 1351 slab_unreclaimable 1352 Part of "slab" that cannot be reclaimed on memory 1353 pressure. 1354 1355 slab (npn) 1356 Amount of memory used for storing in-kernel data 1357 structures. 1358 1359 workingset_refault_anon 1360 Number of refaults of previously evicted anonymous pages. 1361 1362 workingset_refault_file 1363 Number of refaults of previously evicted file pages. 1364 1365 workingset_activate_anon 1366 Number of refaulted anonymous pages that were immediately 1367 activated. 1368 1369 workingset_activate_file 1370 Number of refaulted file pages that were immediately activated. 1371 1372 workingset_restore_anon 1373 Number of restored anonymous pages which have been detected as 1374 an active workingset before they got reclaimed. 1375 1376 workingset_restore_file 1377 Number of restored file pages which have been detected as an 1378 active workingset before they got reclaimed. 1379 1380 workingset_nodereclaim 1381 Number of times a shadow node has been reclaimed 1382 1383 pgfault (npn) 1384 Total number of page faults incurred 1385 1386 pgmajfault (npn) 1387 Number of major page faults incurred 1388 1389 pgrefill (npn) 1390 Amount of scanned pages (in an active LRU list) 1391 1392 pgscan (npn) 1393 Amount of scanned pages (in an inactive LRU list) 1394 1395 pgsteal (npn) 1396 Amount of reclaimed pages 1397 1398 pgactivate (npn) 1399 Amount of pages moved to the active LRU list 1400 1401 pgdeactivate (npn) 1402 Amount of pages moved to the inactive LRU list 1403 1404 pglazyfree (npn) 1405 Amount of pages postponed to be freed under memory pressure 1406 1407 pglazyfreed (npn) 1408 Amount of reclaimed lazyfree pages 1409 1410 thp_fault_alloc (npn) 1411 Number of transparent hugepages which were allocated to satisfy 1412 a page fault. This counter is not present when CONFIG_TRANSPARENT_HUGEPAGE 1413 is not set. 1414 1415 thp_collapse_alloc (npn) 1416 Number of transparent hugepages which were allocated to allow 1417 collapsing an existing range of pages. This counter is not 1418 present when CONFIG_TRANSPARENT_HUGEPAGE is not set. 1419 1420 memory.numa_stat 1421 A read-only nested-keyed file which exists on non-root cgroups. 1422 1423 This breaks down the cgroup's memory footprint into different 1424 types of memory, type-specific details, and other information 1425 per node on the state of the memory management system. 1426 1427 This is useful for providing visibility into the NUMA locality 1428 information within an memcg since the pages are allowed to be 1429 allocated from any physical node. One of the use case is evaluating 1430 application performance by combining this information with the 1431 application's CPU allocation. 1432 1433 All memory amounts are in bytes. 1434 1435 The output format of memory.numa_stat is:: 1436 1437 type N0=<bytes in node 0> N1=<bytes in node 1> ... 1438 1439 The entries are ordered to be human readable, and new entries 1440 can show up in the middle. Don't rely on items remaining in a 1441 fixed position; use the keys to look up specific values! 1442 1443 The entries can refer to the memory.stat. 1444 1445 memory.swap.current 1446 A read-only single value file which exists on non-root 1447 cgroups. 1448 1449 The total amount of swap currently being used by the cgroup 1450 and its descendants. 1451 1452 memory.swap.high 1453 A read-write single value file which exists on non-root 1454 cgroups. The default is "max". 1455 1456 Swap usage throttle limit. If a cgroup's swap usage exceeds 1457 this limit, all its further allocations will be throttled to 1458 allow userspace to implement custom out-of-memory procedures. 1459 1460 This limit marks a point of no return for the cgroup. It is NOT 1461 designed to manage the amount of swapping a workload does 1462 during regular operation. Compare to memory.swap.max, which 1463 prohibits swapping past a set amount, but lets the cgroup 1464 continue unimpeded as long as other memory can be reclaimed. 1465 1466 Healthy workloads are not expected to reach this limit. 1467 1468 memory.swap.max 1469 A read-write single value file which exists on non-root 1470 cgroups. The default is "max". 1471 1472 Swap usage hard limit. If a cgroup's swap usage reaches this 1473 limit, anonymous memory of the cgroup will not be swapped out. 1474 1475 memory.swap.events 1476 A read-only flat-keyed file which exists on non-root cgroups. 1477 The following entries are defined. Unless specified 1478 otherwise, a value change in this file generates a file 1479 modified event. 1480 1481 high 1482 The number of times the cgroup's swap usage was over 1483 the high threshold. 1484 1485 max 1486 The number of times the cgroup's swap usage was about 1487 to go over the max boundary and swap allocation 1488 failed. 1489 1490 fail 1491 The number of times swap allocation failed either 1492 because of running out of swap system-wide or max 1493 limit. 1494 1495 When reduced under the current usage, the existing swap 1496 entries are reclaimed gradually and the swap usage may stay 1497 higher than the limit for an extended period of time. This 1498 reduces the impact on the workload and memory management. 1499 1500 memory.pressure 1501 A read-only nested-keyed file. 1502 1503 Shows pressure stall information for memory. See 1504 :ref:`Documentation/accounting/psi.rst <psi>` for details. 1505 1506 1507Usage Guidelines 1508~~~~~~~~~~~~~~~~ 1509 1510"memory.high" is the main mechanism to control memory usage. 1511Over-committing on high limit (sum of high limits > available memory) 1512and letting global memory pressure to distribute memory according to 1513usage is a viable strategy. 1514 1515Because breach of the high limit doesn't trigger the OOM killer but 1516throttles the offending cgroup, a management agent has ample 1517opportunities to monitor and take appropriate actions such as granting 1518more memory or terminating the workload. 1519 1520Determining whether a cgroup has enough memory is not trivial as 1521memory usage doesn't indicate whether the workload can benefit from 1522more memory. For example, a workload which writes data received from 1523network to a file can use all available memory but can also operate as 1524performant with a small amount of memory. A measure of memory 1525pressure - how much the workload is being impacted due to lack of 1526memory - is necessary to determine whether a workload needs more 1527memory; unfortunately, memory pressure monitoring mechanism isn't 1528implemented yet. 1529 1530 1531Memory Ownership 1532~~~~~~~~~~~~~~~~ 1533 1534A memory area is charged to the cgroup which instantiated it and stays 1535charged to the cgroup until the area is released. Migrating a process 1536to a different cgroup doesn't move the memory usages that it 1537instantiated while in the previous cgroup to the new cgroup. 1538 1539A memory area may be used by processes belonging to different cgroups. 1540To which cgroup the area will be charged is in-deterministic; however, 1541over time, the memory area is likely to end up in a cgroup which has 1542enough memory allowance to avoid high reclaim pressure. 1543 1544If a cgroup sweeps a considerable amount of memory which is expected 1545to be accessed repeatedly by other cgroups, it may make sense to use 1546POSIX_FADV_DONTNEED to relinquish the ownership of memory areas 1547belonging to the affected files to ensure correct memory ownership. 1548 1549 1550IO 1551-- 1552 1553The "io" controller regulates the distribution of IO resources. This 1554controller implements both weight based and absolute bandwidth or IOPS 1555limit distribution; however, weight based distribution is available 1556only if cfq-iosched is in use and neither scheme is available for 1557blk-mq devices. 1558 1559 1560IO Interface Files 1561~~~~~~~~~~~~~~~~~~ 1562 1563 io.stat 1564 A read-only nested-keyed file. 1565 1566 Lines are keyed by $MAJ:$MIN device numbers and not ordered. 1567 The following nested keys are defined. 1568 1569 ====== ===================== 1570 rbytes Bytes read 1571 wbytes Bytes written 1572 rios Number of read IOs 1573 wios Number of write IOs 1574 dbytes Bytes discarded 1575 dios Number of discard IOs 1576 ====== ===================== 1577 1578 An example read output follows:: 1579 1580 8:16 rbytes=1459200 wbytes=314773504 rios=192 wios=353 dbytes=0 dios=0 1581 8:0 rbytes=90430464 wbytes=299008000 rios=8950 wios=1252 dbytes=50331648 dios=3021 1582 1583 io.cost.qos 1584 A read-write nested-keyed file which exists only on the root 1585 cgroup. 1586 1587 This file configures the Quality of Service of the IO cost 1588 model based controller (CONFIG_BLK_CGROUP_IOCOST) which 1589 currently implements "io.weight" proportional control. Lines 1590 are keyed by $MAJ:$MIN device numbers and not ordered. The 1591 line for a given device is populated on the first write for 1592 the device on "io.cost.qos" or "io.cost.model". The following 1593 nested keys are defined. 1594 1595 ====== ===================================== 1596 enable Weight-based control enable 1597 ctrl "auto" or "user" 1598 rpct Read latency percentile [0, 100] 1599 rlat Read latency threshold 1600 wpct Write latency percentile [0, 100] 1601 wlat Write latency threshold 1602 min Minimum scaling percentage [1, 10000] 1603 max Maximum scaling percentage [1, 10000] 1604 ====== ===================================== 1605 1606 The controller is disabled by default and can be enabled by 1607 setting "enable" to 1. "rpct" and "wpct" parameters default 1608 to zero and the controller uses internal device saturation 1609 state to adjust the overall IO rate between "min" and "max". 1610 1611 When a better control quality is needed, latency QoS 1612 parameters can be configured. For example:: 1613 1614 8:16 enable=1 ctrl=auto rpct=95.00 rlat=75000 wpct=95.00 wlat=150000 min=50.00 max=150.0 1615 1616 shows that on sdb, the controller is enabled, will consider 1617 the device saturated if the 95th percentile of read completion 1618 latencies is above 75ms or write 150ms, and adjust the overall 1619 IO issue rate between 50% and 150% accordingly. 1620 1621 The lower the saturation point, the better the latency QoS at 1622 the cost of aggregate bandwidth. The narrower the allowed 1623 adjustment range between "min" and "max", the more conformant 1624 to the cost model the IO behavior. Note that the IO issue 1625 base rate may be far off from 100% and setting "min" and "max" 1626 blindly can lead to a significant loss of device capacity or 1627 control quality. "min" and "max" are useful for regulating 1628 devices which show wide temporary behavior changes - e.g. a 1629 ssd which accepts writes at the line speed for a while and 1630 then completely stalls for multiple seconds. 1631 1632 When "ctrl" is "auto", the parameters are controlled by the 1633 kernel and may change automatically. Setting "ctrl" to "user" 1634 or setting any of the percentile and latency parameters puts 1635 it into "user" mode and disables the automatic changes. The 1636 automatic mode can be restored by setting "ctrl" to "auto". 1637 1638 io.cost.model 1639 A read-write nested-keyed file which exists only on the root 1640 cgroup. 1641 1642 This file configures the cost model of the IO cost model based 1643 controller (CONFIG_BLK_CGROUP_IOCOST) which currently 1644 implements "io.weight" proportional control. Lines are keyed 1645 by $MAJ:$MIN device numbers and not ordered. The line for a 1646 given device is populated on the first write for the device on 1647 "io.cost.qos" or "io.cost.model". The following nested keys 1648 are defined. 1649 1650 ===== ================================ 1651 ctrl "auto" or "user" 1652 model The cost model in use - "linear" 1653 ===== ================================ 1654 1655 When "ctrl" is "auto", the kernel may change all parameters 1656 dynamically. When "ctrl" is set to "user" or any other 1657 parameters are written to, "ctrl" become "user" and the 1658 automatic changes are disabled. 1659 1660 When "model" is "linear", the following model parameters are 1661 defined. 1662 1663 ============= ======================================== 1664 [r|w]bps The maximum sequential IO throughput 1665 [r|w]seqiops The maximum 4k sequential IOs per second 1666 [r|w]randiops The maximum 4k random IOs per second 1667 ============= ======================================== 1668 1669 From the above, the builtin linear model determines the base 1670 costs of a sequential and random IO and the cost coefficient 1671 for the IO size. While simple, this model can cover most 1672 common device classes acceptably. 1673 1674 The IO cost model isn't expected to be accurate in absolute 1675 sense and is scaled to the device behavior dynamically. 1676 1677 If needed, tools/cgroup/iocost_coef_gen.py can be used to 1678 generate device-specific coefficients. 1679 1680 io.weight 1681 A read-write flat-keyed file which exists on non-root cgroups. 1682 The default is "default 100". 1683 1684 The first line is the default weight applied to devices 1685 without specific override. The rest are overrides keyed by 1686 $MAJ:$MIN device numbers and not ordered. The weights are in 1687 the range [1, 10000] and specifies the relative amount IO time 1688 the cgroup can use in relation to its siblings. 1689 1690 The default weight can be updated by writing either "default 1691 $WEIGHT" or simply "$WEIGHT". Overrides can be set by writing 1692 "$MAJ:$MIN $WEIGHT" and unset by writing "$MAJ:$MIN default". 1693 1694 An example read output follows:: 1695 1696 default 100 1697 8:16 200 1698 8:0 50 1699 1700 io.max 1701 A read-write nested-keyed file which exists on non-root 1702 cgroups. 1703 1704 BPS and IOPS based IO limit. Lines are keyed by $MAJ:$MIN 1705 device numbers and not ordered. The following nested keys are 1706 defined. 1707 1708 ===== ================================== 1709 rbps Max read bytes per second 1710 wbps Max write bytes per second 1711 riops Max read IO operations per second 1712 wiops Max write IO operations per second 1713 ===== ================================== 1714 1715 When writing, any number of nested key-value pairs can be 1716 specified in any order. "max" can be specified as the value 1717 to remove a specific limit. If the same key is specified 1718 multiple times, the outcome is undefined. 1719 1720 BPS and IOPS are measured in each IO direction and IOs are 1721 delayed if limit is reached. Temporary bursts are allowed. 1722 1723 Setting read limit at 2M BPS and write at 120 IOPS for 8:16:: 1724 1725 echo "8:16 rbps=2097152 wiops=120" > io.max 1726 1727 Reading returns the following:: 1728 1729 8:16 rbps=2097152 wbps=max riops=max wiops=120 1730 1731 Write IOPS limit can be removed by writing the following:: 1732 1733 echo "8:16 wiops=max" > io.max 1734 1735 Reading now returns the following:: 1736 1737 8:16 rbps=2097152 wbps=max riops=max wiops=max 1738 1739 io.pressure 1740 A read-only nested-keyed file. 1741 1742 Shows pressure stall information for IO. See 1743 :ref:`Documentation/accounting/psi.rst <psi>` for details. 1744 1745 1746Writeback 1747~~~~~~~~~ 1748 1749Page cache is dirtied through buffered writes and shared mmaps and 1750written asynchronously to the backing filesystem by the writeback 1751mechanism. Writeback sits between the memory and IO domains and 1752regulates the proportion of dirty memory by balancing dirtying and 1753write IOs. 1754 1755The io controller, in conjunction with the memory controller, 1756implements control of page cache writeback IOs. The memory controller 1757defines the memory domain that dirty memory ratio is calculated and 1758maintained for and the io controller defines the io domain which 1759writes out dirty pages for the memory domain. Both system-wide and 1760per-cgroup dirty memory states are examined and the more restrictive 1761of the two is enforced. 1762 1763cgroup writeback requires explicit support from the underlying 1764filesystem. Currently, cgroup writeback is implemented on ext2, ext4, 1765btrfs, f2fs, and xfs. On other filesystems, all writeback IOs are 1766attributed to the root cgroup. 1767 1768There are inherent differences in memory and writeback management 1769which affects how cgroup ownership is tracked. Memory is tracked per 1770page while writeback per inode. For the purpose of writeback, an 1771inode is assigned to a cgroup and all IO requests to write dirty pages 1772from the inode are attributed to that cgroup. 1773 1774As cgroup ownership for memory is tracked per page, there can be pages 1775which are associated with different cgroups than the one the inode is 1776associated with. These are called foreign pages. The writeback 1777constantly keeps track of foreign pages and, if a particular foreign 1778cgroup becomes the majority over a certain period of time, switches 1779the ownership of the inode to that cgroup. 1780 1781While this model is enough for most use cases where a given inode is 1782mostly dirtied by a single cgroup even when the main writing cgroup 1783changes over time, use cases where multiple cgroups write to a single 1784inode simultaneously are not supported well. In such circumstances, a 1785significant portion of IOs are likely to be attributed incorrectly. 1786As memory controller assigns page ownership on the first use and 1787doesn't update it until the page is released, even if writeback 1788strictly follows page ownership, multiple cgroups dirtying overlapping 1789areas wouldn't work as expected. It's recommended to avoid such usage 1790patterns. 1791 1792The sysctl knobs which affect writeback behavior are applied to cgroup 1793writeback as follows. 1794 1795 vm.dirty_background_ratio, vm.dirty_ratio 1796 These ratios apply the same to cgroup writeback with the 1797 amount of available memory capped by limits imposed by the 1798 memory controller and system-wide clean memory. 1799 1800 vm.dirty_background_bytes, vm.dirty_bytes 1801 For cgroup writeback, this is calculated into ratio against 1802 total available memory and applied the same way as 1803 vm.dirty[_background]_ratio. 1804 1805 1806IO Latency 1807~~~~~~~~~~ 1808 1809This is a cgroup v2 controller for IO workload protection. You provide a group 1810with a latency target, and if the average latency exceeds that target the 1811controller will throttle any peers that have a lower latency target than the 1812protected workload. 1813 1814The limits are only applied at the peer level in the hierarchy. This means that 1815in the diagram below, only groups A, B, and C will influence each other, and 1816groups D and F will influence each other. Group G will influence nobody:: 1817 1818 [root] 1819 / | \ 1820 A B C 1821 / \ | 1822 D F G 1823 1824 1825So the ideal way to configure this is to set io.latency in groups A, B, and C. 1826Generally you do not want to set a value lower than the latency your device 1827supports. Experiment to find the value that works best for your workload. 1828Start at higher than the expected latency for your device and watch the 1829avg_lat value in io.stat for your workload group to get an idea of the 1830latency you see during normal operation. Use the avg_lat value as a basis for 1831your real setting, setting at 10-15% higher than the value in io.stat. 1832 1833How IO Latency Throttling Works 1834~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~ 1835 1836io.latency is work conserving; so as long as everybody is meeting their latency 1837target the controller doesn't do anything. Once a group starts missing its 1838target it begins throttling any peer group that has a higher target than itself. 1839This throttling takes 2 forms: 1840 1841- Queue depth throttling. This is the number of outstanding IO's a group is 1842 allowed to have. We will clamp down relatively quickly, starting at no limit 1843 and going all the way down to 1 IO at a time. 1844 1845- Artificial delay induction. There are certain types of IO that cannot be 1846 throttled without possibly adversely affecting higher priority groups. This 1847 includes swapping and metadata IO. These types of IO are allowed to occur 1848 normally, however they are "charged" to the originating group. If the 1849 originating group is being throttled you will see the use_delay and delay 1850 fields in io.stat increase. The delay value is how many microseconds that are 1851 being added to any process that runs in this group. Because this number can 1852 grow quite large if there is a lot of swapping or metadata IO occurring we 1853 limit the individual delay events to 1 second at a time. 1854 1855Once the victimized group starts meeting its latency target again it will start 1856unthrottling any peer groups that were throttled previously. If the victimized 1857group simply stops doing IO the global counter will unthrottle appropriately. 1858 1859IO Latency Interface Files 1860~~~~~~~~~~~~~~~~~~~~~~~~~~ 1861 1862 io.latency 1863 This takes a similar format as the other controllers. 1864 1865 "MAJOR:MINOR target=<target time in microseconds" 1866 1867 io.stat 1868 If the controller is enabled you will see extra stats in io.stat in 1869 addition to the normal ones. 1870 1871 depth 1872 This is the current queue depth for the group. 1873 1874 avg_lat 1875 This is an exponential moving average with a decay rate of 1/exp 1876 bound by the sampling interval. The decay rate interval can be 1877 calculated by multiplying the win value in io.stat by the 1878 corresponding number of samples based on the win value. 1879 1880 win 1881 The sampling window size in milliseconds. This is the minimum 1882 duration of time between evaluation events. Windows only elapse 1883 with IO activity. Idle periods extend the most recent window. 1884 1885IO Priority 1886~~~~~~~~~~~ 1887 1888A single attribute controls the behavior of the I/O priority cgroup policy, 1889namely the blkio.prio.class attribute. The following values are accepted for 1890that attribute: 1891 1892 no-change 1893 Do not modify the I/O priority class. 1894 1895 none-to-rt 1896 For requests that do not have an I/O priority class (NONE), 1897 change the I/O priority class into RT. Do not modify 1898 the I/O priority class of other requests. 1899 1900 restrict-to-be 1901 For requests that do not have an I/O priority class or that have I/O 1902 priority class RT, change it into BE. Do not modify the I/O priority 1903 class of requests that have priority class IDLE. 1904 1905 idle 1906 Change the I/O priority class of all requests into IDLE, the lowest 1907 I/O priority class. 1908 1909The following numerical values are associated with the I/O priority policies: 1910 1911+-------------+---+ 1912| no-change | 0 | 1913+-------------+---+ 1914| none-to-rt | 1 | 1915+-------------+---+ 1916| rt-to-be | 2 | 1917+-------------+---+ 1918| all-to-idle | 3 | 1919+-------------+---+ 1920 1921The numerical value that corresponds to each I/O priority class is as follows: 1922 1923+-------------------------------+---+ 1924| IOPRIO_CLASS_NONE | 0 | 1925+-------------------------------+---+ 1926| IOPRIO_CLASS_RT (real-time) | 1 | 1927+-------------------------------+---+ 1928| IOPRIO_CLASS_BE (best effort) | 2 | 1929+-------------------------------+---+ 1930| IOPRIO_CLASS_IDLE | 3 | 1931+-------------------------------+---+ 1932 1933The algorithm to set the I/O priority class for a request is as follows: 1934 1935- Translate the I/O priority class policy into a number. 1936- Change the request I/O priority class into the maximum of the I/O priority 1937 class policy number and the numerical I/O priority class. 1938 1939PID 1940--- 1941 1942The process number controller is used to allow a cgroup to stop any 1943new tasks from being fork()'d or clone()'d after a specified limit is 1944reached. 1945 1946The number of tasks in a cgroup can be exhausted in ways which other 1947controllers cannot prevent, thus warranting its own controller. For 1948example, a fork bomb is likely to exhaust the number of tasks before 1949hitting memory restrictions. 1950 1951Note that PIDs used in this controller refer to TIDs, process IDs as 1952used by the kernel. 1953 1954 1955PID Interface Files 1956~~~~~~~~~~~~~~~~~~~ 1957 1958 pids.max 1959 A read-write single value file which exists on non-root 1960 cgroups. The default is "max". 1961 1962 Hard limit of number of processes. 1963 1964 pids.current 1965 A read-only single value file which exists on all cgroups. 1966 1967 The number of processes currently in the cgroup and its 1968 descendants. 1969 1970Organisational operations are not blocked by cgroup policies, so it is 1971possible to have pids.current > pids.max. This can be done by either 1972setting the limit to be smaller than pids.current, or attaching enough 1973processes to the cgroup such that pids.current is larger than 1974pids.max. However, it is not possible to violate a cgroup PID policy 1975through fork() or clone(). These will return -EAGAIN if the creation 1976of a new process would cause a cgroup policy to be violated. 1977 1978 1979Cpuset 1980------ 1981 1982The "cpuset" controller provides a mechanism for constraining 1983the CPU and memory node placement of tasks to only the resources 1984specified in the cpuset interface files in a task's current cgroup. 1985This is especially valuable on large NUMA systems where placing jobs 1986on properly sized subsets of the systems with careful processor and 1987memory placement to reduce cross-node memory access and contention 1988can improve overall system performance. 1989 1990The "cpuset" controller is hierarchical. That means the controller 1991cannot use CPUs or memory nodes not allowed in its parent. 1992 1993 1994Cpuset Interface Files 1995~~~~~~~~~~~~~~~~~~~~~~ 1996 1997 cpuset.cpus 1998 A read-write multiple values file which exists on non-root 1999 cpuset-enabled cgroups. 2000 2001 It lists the requested CPUs to be used by tasks within this 2002 cgroup. The actual list of CPUs to be granted, however, is 2003 subjected to constraints imposed by its parent and can differ 2004 from the requested CPUs. 2005 2006 The CPU numbers are comma-separated numbers or ranges. 2007 For example:: 2008 2009 # cat cpuset.cpus 2010 0-4,6,8-10 2011 2012 An empty value indicates that the cgroup is using the same 2013 setting as the nearest cgroup ancestor with a non-empty 2014 "cpuset.cpus" or all the available CPUs if none is found. 2015 2016 The value of "cpuset.cpus" stays constant until the next update 2017 and won't be affected by any CPU hotplug events. 2018 2019 cpuset.cpus.effective 2020 A read-only multiple values file which exists on all 2021 cpuset-enabled cgroups. 2022 2023 It lists the onlined CPUs that are actually granted to this 2024 cgroup by its parent. These CPUs are allowed to be used by 2025 tasks within the current cgroup. 2026 2027 If "cpuset.cpus" is empty, the "cpuset.cpus.effective" file shows 2028 all the CPUs from the parent cgroup that can be available to 2029 be used by this cgroup. Otherwise, it should be a subset of 2030 "cpuset.cpus" unless none of the CPUs listed in "cpuset.cpus" 2031 can be granted. In this case, it will be treated just like an 2032 empty "cpuset.cpus". 2033 2034 Its value will be affected by CPU hotplug events. 2035 2036 cpuset.mems 2037 A read-write multiple values file which exists on non-root 2038 cpuset-enabled cgroups. 2039 2040 It lists the requested memory nodes to be used by tasks within 2041 this cgroup. The actual list of memory nodes granted, however, 2042 is subjected to constraints imposed by its parent and can differ 2043 from the requested memory nodes. 2044 2045 The memory node numbers are comma-separated numbers or ranges. 2046 For example:: 2047 2048 # cat cpuset.mems 2049 0-1,3 2050 2051 An empty value indicates that the cgroup is using the same 2052 setting as the nearest cgroup ancestor with a non-empty 2053 "cpuset.mems" or all the available memory nodes if none 2054 is found. 2055 2056 The value of "cpuset.mems" stays constant until the next update 2057 and won't be affected by any memory nodes hotplug events. 2058 2059 cpuset.mems.effective 2060 A read-only multiple values file which exists on all 2061 cpuset-enabled cgroups. 2062 2063 It lists the onlined memory nodes that are actually granted to 2064 this cgroup by its parent. These memory nodes are allowed to 2065 be used by tasks within the current cgroup. 2066 2067 If "cpuset.mems" is empty, it shows all the memory nodes from the 2068 parent cgroup that will be available to be used by this cgroup. 2069 Otherwise, it should be a subset of "cpuset.mems" unless none of 2070 the memory nodes listed in "cpuset.mems" can be granted. In this 2071 case, it will be treated just like an empty "cpuset.mems". 2072 2073 Its value will be affected by memory nodes hotplug events. 2074 2075 cpuset.cpus.partition 2076 A read-write single value file which exists on non-root 2077 cpuset-enabled cgroups. This flag is owned by the parent cgroup 2078 and is not delegatable. 2079 2080 It accepts only the following input values when written to. 2081 2082 ======== ================================ 2083 "root" a partition root 2084 "member" a non-root member of a partition 2085 ======== ================================ 2086 2087 When set to be a partition root, the current cgroup is the 2088 root of a new partition or scheduling domain that comprises 2089 itself and all its descendants except those that are separate 2090 partition roots themselves and their descendants. The root 2091 cgroup is always a partition root. 2092 2093 There are constraints on where a partition root can be set. 2094 It can only be set in a cgroup if all the following conditions 2095 are true. 2096 2097 1) The "cpuset.cpus" is not empty and the list of CPUs are 2098 exclusive, i.e. they are not shared by any of its siblings. 2099 2) The parent cgroup is a partition root. 2100 3) The "cpuset.cpus" is also a proper subset of the parent's 2101 "cpuset.cpus.effective". 2102 4) There is no child cgroups with cpuset enabled. This is for 2103 eliminating corner cases that have to be handled if such a 2104 condition is allowed. 2105 2106 Setting it to partition root will take the CPUs away from the 2107 effective CPUs of the parent cgroup. Once it is set, this 2108 file cannot be reverted back to "member" if there are any child 2109 cgroups with cpuset enabled. 2110 2111 A parent partition cannot distribute all its CPUs to its 2112 child partitions. There must be at least one cpu left in the 2113 parent partition. 2114 2115 Once becoming a partition root, changes to "cpuset.cpus" is 2116 generally allowed as long as the first condition above is true, 2117 the change will not take away all the CPUs from the parent 2118 partition and the new "cpuset.cpus" value is a superset of its 2119 children's "cpuset.cpus" values. 2120 2121 Sometimes, external factors like changes to ancestors' 2122 "cpuset.cpus" or cpu hotplug can cause the state of the partition 2123 root to change. On read, the "cpuset.sched.partition" file 2124 can show the following values. 2125 2126 ============== ============================== 2127 "member" Non-root member of a partition 2128 "root" Partition root 2129 "root invalid" Invalid partition root 2130 ============== ============================== 2131 2132 It is a partition root if the first 2 partition root conditions 2133 above are true and at least one CPU from "cpuset.cpus" is 2134 granted by the parent cgroup. 2135 2136 A partition root can become invalid if none of CPUs requested 2137 in "cpuset.cpus" can be granted by the parent cgroup or the 2138 parent cgroup is no longer a partition root itself. In this 2139 case, it is not a real partition even though the restriction 2140 of the first partition root condition above will still apply. 2141 The cpu affinity of all the tasks in the cgroup will then be 2142 associated with CPUs in the nearest ancestor partition. 2143 2144 An invalid partition root can be transitioned back to a 2145 real partition root if at least one of the requested CPUs 2146 can now be granted by its parent. In this case, the cpu 2147 affinity of all the tasks in the formerly invalid partition 2148 will be associated to the CPUs of the newly formed partition. 2149 Changing the partition state of an invalid partition root to 2150 "member" is always allowed even if child cpusets are present. 2151 2152 2153Device controller 2154----------------- 2155 2156Device controller manages access to device files. It includes both 2157creation of new device files (using mknod), and access to the 2158existing device files. 2159 2160Cgroup v2 device controller has no interface files and is implemented 2161on top of cgroup BPF. To control access to device files, a user may 2162create bpf programs of the BPF_CGROUP_DEVICE type and attach them 2163to cgroups. On an attempt to access a device file, corresponding 2164BPF programs will be executed, and depending on the return value 2165the attempt will succeed or fail with -EPERM. 2166 2167A BPF_CGROUP_DEVICE program takes a pointer to the bpf_cgroup_dev_ctx 2168structure, which describes the device access attempt: access type 2169(mknod/read/write) and device (type, major and minor numbers). 2170If the program returns 0, the attempt fails with -EPERM, otherwise 2171it succeeds. 2172 2173An example of BPF_CGROUP_DEVICE program may be found in the kernel 2174source tree in the tools/testing/selftests/bpf/progs/dev_cgroup.c file. 2175 2176 2177RDMA 2178---- 2179 2180The "rdma" controller regulates the distribution and accounting of 2181RDMA resources. 2182 2183RDMA Interface Files 2184~~~~~~~~~~~~~~~~~~~~ 2185 2186 rdma.max 2187 A readwrite nested-keyed file that exists for all the cgroups 2188 except root that describes current configured resource limit 2189 for a RDMA/IB device. 2190 2191 Lines are keyed by device name and are not ordered. 2192 Each line contains space separated resource name and its configured 2193 limit that can be distributed. 2194 2195 The following nested keys are defined. 2196 2197 ========== ============================= 2198 hca_handle Maximum number of HCA Handles 2199 hca_object Maximum number of HCA Objects 2200 ========== ============================= 2201 2202 An example for mlx4 and ocrdma device follows:: 2203 2204 mlx4_0 hca_handle=2 hca_object=2000 2205 ocrdma1 hca_handle=3 hca_object=max 2206 2207 rdma.current 2208 A read-only file that describes current resource usage. 2209 It exists for all the cgroup except root. 2210 2211 An example for mlx4 and ocrdma device follows:: 2212 2213 mlx4_0 hca_handle=1 hca_object=20 2214 ocrdma1 hca_handle=1 hca_object=23 2215 2216HugeTLB 2217------- 2218 2219The HugeTLB controller allows to limit the HugeTLB usage per control group and 2220enforces the controller limit during page fault. 2221 2222HugeTLB Interface Files 2223~~~~~~~~~~~~~~~~~~~~~~~ 2224 2225 hugetlb.<hugepagesize>.current 2226 Show current usage for "hugepagesize" hugetlb. It exists for all 2227 the cgroup except root. 2228 2229 hugetlb.<hugepagesize>.max 2230 Set/show the hard limit of "hugepagesize" hugetlb usage. 2231 The default value is "max". It exists for all the cgroup except root. 2232 2233 hugetlb.<hugepagesize>.events 2234 A read-only flat-keyed file which exists on non-root cgroups. 2235 2236 max 2237 The number of allocation failure due to HugeTLB limit 2238 2239 hugetlb.<hugepagesize>.events.local 2240 Similar to hugetlb.<hugepagesize>.events but the fields in the file 2241 are local to the cgroup i.e. not hierarchical. The file modified event 2242 generated on this file reflects only the local events. 2243 2244Misc 2245---- 2246 2247The Miscellaneous cgroup provides the resource limiting and tracking 2248mechanism for the scalar resources which cannot be abstracted like the other 2249cgroup resources. Controller is enabled by the CONFIG_CGROUP_MISC config 2250option. 2251 2252A resource can be added to the controller via enum misc_res_type{} in the 2253include/linux/misc_cgroup.h file and the corresponding name via misc_res_name[] 2254in the kernel/cgroup/misc.c file. Provider of the resource must set its 2255capacity prior to using the resource by calling misc_cg_set_capacity(). 2256 2257Once a capacity is set then the resource usage can be updated using charge and 2258uncharge APIs. All of the APIs to interact with misc controller are in 2259include/linux/misc_cgroup.h. 2260 2261Misc Interface Files 2262~~~~~~~~~~~~~~~~~~~~ 2263 2264Miscellaneous controller provides 3 interface files. If two misc resources (res_a and res_b) are registered then: 2265 2266 misc.capacity 2267 A read-only flat-keyed file shown only in the root cgroup. It shows 2268 miscellaneous scalar resources available on the platform along with 2269 their quantities:: 2270 2271 $ cat misc.capacity 2272 res_a 50 2273 res_b 10 2274 2275 misc.current 2276 A read-only flat-keyed file shown in the non-root cgroups. It shows 2277 the current usage of the resources in the cgroup and its children.:: 2278 2279 $ cat misc.current 2280 res_a 3 2281 res_b 0 2282 2283 misc.max 2284 A read-write flat-keyed file shown in the non root cgroups. Allowed 2285 maximum usage of the resources in the cgroup and its children.:: 2286 2287 $ cat misc.max 2288 res_a max 2289 res_b 4 2290 2291 Limit can be set by:: 2292 2293 # echo res_a 1 > misc.max 2294 2295 Limit can be set to max by:: 2296 2297 # echo res_a max > misc.max 2298 2299 Limits can be set higher than the capacity value in the misc.capacity 2300 file. 2301 2302Migration and Ownership 2303~~~~~~~~~~~~~~~~~~~~~~~ 2304 2305A miscellaneous scalar resource is charged to the cgroup in which it is used 2306first, and stays charged to that cgroup until that resource is freed. Migrating 2307a process to a different cgroup does not move the charge to the destination 2308cgroup where the process has moved. 2309 2310Others 2311------ 2312 2313perf_event 2314~~~~~~~~~~ 2315 2316perf_event controller, if not mounted on a legacy hierarchy, is 2317automatically enabled on the v2 hierarchy so that perf events can 2318always be filtered by cgroup v2 path. The controller can still be 2319moved to a legacy hierarchy after v2 hierarchy is populated. 2320 2321 2322Non-normative information 2323------------------------- 2324 2325This section contains information that isn't considered to be a part of 2326the stable kernel API and so is subject to change. 2327 2328 2329CPU controller root cgroup process behaviour 2330~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~ 2331 2332When distributing CPU cycles in the root cgroup each thread in this 2333cgroup is treated as if it was hosted in a separate child cgroup of the 2334root cgroup. This child cgroup weight is dependent on its thread nice 2335level. 2336 2337For details of this mapping see sched_prio_to_weight array in 2338kernel/sched/core.c file (values from this array should be scaled 2339appropriately so the neutral - nice 0 - value is 100 instead of 1024). 2340 2341 2342IO controller root cgroup process behaviour 2343~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~ 2344 2345Root cgroup processes are hosted in an implicit leaf child node. 2346When distributing IO resources this implicit child node is taken into 2347account as if it was a normal child cgroup of the root cgroup with a 2348weight value of 200. 2349 2350 2351Namespace 2352========= 2353 2354Basics 2355------ 2356 2357cgroup namespace provides a mechanism to virtualize the view of the 2358"/proc/$PID/cgroup" file and cgroup mounts. The CLONE_NEWCGROUP clone 2359flag can be used with clone(2) and unshare(2) to create a new cgroup 2360namespace. The process running inside the cgroup namespace will have 2361its "/proc/$PID/cgroup" output restricted to cgroupns root. The 2362cgroupns root is the cgroup of the process at the time of creation of 2363the cgroup namespace. 2364 2365Without cgroup namespace, the "/proc/$PID/cgroup" file shows the 2366complete path of the cgroup of a process. In a container setup where 2367a set of cgroups and namespaces are intended to isolate processes the 2368"/proc/$PID/cgroup" file may leak potential system level information 2369to the isolated processes. For example:: 2370 2371 # cat /proc/self/cgroup 2372 0::/batchjobs/container_id1 2373 2374The path '/batchjobs/container_id1' can be considered as system-data 2375and undesirable to expose to the isolated processes. cgroup namespace 2376can be used to restrict visibility of this path. For example, before 2377creating a cgroup namespace, one would see:: 2378 2379 # ls -l /proc/self/ns/cgroup 2380 lrwxrwxrwx 1 root root 0 2014-07-15 10:37 /proc/self/ns/cgroup -> cgroup:[4026531835] 2381 # cat /proc/self/cgroup 2382 0::/batchjobs/container_id1 2383 2384After unsharing a new namespace, the view changes:: 2385 2386 # ls -l /proc/self/ns/cgroup 2387 lrwxrwxrwx 1 root root 0 2014-07-15 10:35 /proc/self/ns/cgroup -> cgroup:[4026532183] 2388 # cat /proc/self/cgroup 2389 0::/ 2390 2391When some thread from a multi-threaded process unshares its cgroup 2392namespace, the new cgroupns gets applied to the entire process (all 2393the threads). This is natural for the v2 hierarchy; however, for the 2394legacy hierarchies, this may be unexpected. 2395 2396A cgroup namespace is alive as long as there are processes inside or 2397mounts pinning it. When the last usage goes away, the cgroup 2398namespace is destroyed. The cgroupns root and the actual cgroups 2399remain. 2400 2401 2402The Root and Views 2403------------------ 2404 2405The 'cgroupns root' for a cgroup namespace is the cgroup in which the 2406process calling unshare(2) is running. For example, if a process in 2407/batchjobs/container_id1 cgroup calls unshare, cgroup 2408/batchjobs/container_id1 becomes the cgroupns root. For the 2409init_cgroup_ns, this is the real root ('/') cgroup. 2410 2411The cgroupns root cgroup does not change even if the namespace creator 2412process later moves to a different cgroup:: 2413 2414 # ~/unshare -c # unshare cgroupns in some cgroup 2415 # cat /proc/self/cgroup 2416 0::/ 2417 # mkdir sub_cgrp_1 2418 # echo 0 > sub_cgrp_1/cgroup.procs 2419 # cat /proc/self/cgroup 2420 0::/sub_cgrp_1 2421 2422Each process gets its namespace-specific view of "/proc/$PID/cgroup" 2423 2424Processes running inside the cgroup namespace will be able to see 2425cgroup paths (in /proc/self/cgroup) only inside their root cgroup. 2426From within an unshared cgroupns:: 2427 2428 # sleep 100000 & 2429 [1] 7353 2430 # echo 7353 > sub_cgrp_1/cgroup.procs 2431 # cat /proc/7353/cgroup 2432 0::/sub_cgrp_1 2433 2434From the initial cgroup namespace, the real cgroup path will be 2435visible:: 2436 2437 $ cat /proc/7353/cgroup 2438 0::/batchjobs/container_id1/sub_cgrp_1 2439 2440From a sibling cgroup namespace (that is, a namespace rooted at a 2441different cgroup), the cgroup path relative to its own cgroup 2442namespace root will be shown. For instance, if PID 7353's cgroup 2443namespace root is at '/batchjobs/container_id2', then it will see:: 2444 2445 # cat /proc/7353/cgroup 2446 0::/../container_id2/sub_cgrp_1 2447 2448Note that the relative path always starts with '/' to indicate that 2449its relative to the cgroup namespace root of the caller. 2450 2451 2452Migration and setns(2) 2453---------------------- 2454 2455Processes inside a cgroup namespace can move into and out of the 2456namespace root if they have proper access to external cgroups. For 2457example, from inside a namespace with cgroupns root at 2458/batchjobs/container_id1, and assuming that the global hierarchy is 2459still accessible inside cgroupns:: 2460 2461 # cat /proc/7353/cgroup 2462 0::/sub_cgrp_1 2463 # echo 7353 > batchjobs/container_id2/cgroup.procs 2464 # cat /proc/7353/cgroup 2465 0::/../container_id2 2466 2467Note that this kind of setup is not encouraged. A task inside cgroup 2468namespace should only be exposed to its own cgroupns hierarchy. 2469 2470setns(2) to another cgroup namespace is allowed when: 2471 2472(a) the process has CAP_SYS_ADMIN against its current user namespace 2473(b) the process has CAP_SYS_ADMIN against the target cgroup 2474 namespace's userns 2475 2476No implicit cgroup changes happen with attaching to another cgroup 2477namespace. It is expected that the someone moves the attaching 2478process under the target cgroup namespace root. 2479 2480 2481Interaction with Other Namespaces 2482--------------------------------- 2483 2484Namespace specific cgroup hierarchy can be mounted by a process 2485running inside a non-init cgroup namespace:: 2486 2487 # mount -t cgroup2 none $MOUNT_POINT 2488 2489This will mount the unified cgroup hierarchy with cgroupns root as the 2490filesystem root. The process needs CAP_SYS_ADMIN against its user and 2491mount namespaces. 2492 2493The virtualization of /proc/self/cgroup file combined with restricting 2494the view of cgroup hierarchy by namespace-private cgroupfs mount 2495provides a properly isolated cgroup view inside the container. 2496 2497 2498Information on Kernel Programming 2499================================= 2500 2501This section contains kernel programming information in the areas 2502where interacting with cgroup is necessary. cgroup core and 2503controllers are not covered. 2504 2505 2506Filesystem Support for Writeback 2507-------------------------------- 2508 2509A filesystem can support cgroup writeback by updating 2510address_space_operations->writepage[s]() to annotate bio's using the 2511following two functions. 2512 2513 wbc_init_bio(@wbc, @bio) 2514 Should be called for each bio carrying writeback data and 2515 associates the bio with the inode's owner cgroup and the 2516 corresponding request queue. This must be called after 2517 a queue (device) has been associated with the bio and 2518 before submission. 2519 2520 wbc_account_cgroup_owner(@wbc, @page, @bytes) 2521 Should be called for each data segment being written out. 2522 While this function doesn't care exactly when it's called 2523 during the writeback session, it's the easiest and most 2524 natural to call it as data segments are added to a bio. 2525 2526With writeback bio's annotated, cgroup support can be enabled per 2527super_block by setting SB_I_CGROUPWB in ->s_iflags. This allows for 2528selective disabling of cgroup writeback support which is helpful when 2529certain filesystem features, e.g. journaled data mode, are 2530incompatible. 2531 2532wbc_init_bio() binds the specified bio to its cgroup. Depending on 2533the configuration, the bio may be executed at a lower priority and if 2534the writeback session is holding shared resources, e.g. a journal 2535entry, may lead to priority inversion. There is no one easy solution 2536for the problem. Filesystems can try to work around specific problem 2537cases by skipping wbc_init_bio() and using bio_associate_blkg() 2538directly. 2539 2540 2541Deprecated v1 Core Features 2542=========================== 2543 2544- Multiple hierarchies including named ones are not supported. 2545 2546- All v1 mount options are not supported. 2547 2548- The "tasks" file is removed and "cgroup.procs" is not sorted. 2549 2550- "cgroup.clone_children" is removed. 2551 2552- /proc/cgroups is meaningless for v2. Use "cgroup.controllers" file 2553 at the root instead. 2554 2555 2556Issues with v1 and Rationales for v2 2557==================================== 2558 2559Multiple Hierarchies 2560-------------------- 2561 2562cgroup v1 allowed an arbitrary number of hierarchies and each 2563hierarchy could host any number of controllers. While this seemed to 2564provide a high level of flexibility, it wasn't useful in practice. 2565 2566For example, as there is only one instance of each controller, utility 2567type controllers such as freezer which can be useful in all 2568hierarchies could only be used in one. The issue is exacerbated by 2569the fact that controllers couldn't be moved to another hierarchy once 2570hierarchies were populated. Another issue was that all controllers 2571bound to a hierarchy were forced to have exactly the same view of the 2572hierarchy. It wasn't possible to vary the granularity depending on 2573the specific controller. 2574 2575In practice, these issues heavily limited which controllers could be 2576put on the same hierarchy and most configurations resorted to putting 2577each controller on its own hierarchy. Only closely related ones, such 2578as the cpu and cpuacct controllers, made sense to be put on the same 2579hierarchy. This often meant that userland ended up managing multiple 2580similar hierarchies repeating the same steps on each hierarchy 2581whenever a hierarchy management operation was necessary. 2582 2583Furthermore, support for multiple hierarchies came at a steep cost. 2584It greatly complicated cgroup core implementation but more importantly 2585the support for multiple hierarchies restricted how cgroup could be 2586used in general and what controllers was able to do. 2587 2588There was no limit on how many hierarchies there might be, which meant 2589that a thread's cgroup membership couldn't be described in finite 2590length. The key might contain any number of entries and was unlimited 2591in length, which made it highly awkward to manipulate and led to 2592addition of controllers which existed only to identify membership, 2593which in turn exacerbated the original problem of proliferating number 2594of hierarchies. 2595 2596Also, as a controller couldn't have any expectation regarding the 2597topologies of hierarchies other controllers might be on, each 2598controller had to assume that all other controllers were attached to 2599completely orthogonal hierarchies. This made it impossible, or at 2600least very cumbersome, for controllers to cooperate with each other. 2601 2602In most use cases, putting controllers on hierarchies which are 2603completely orthogonal to each other isn't necessary. What usually is 2604called for is the ability to have differing levels of granularity 2605depending on the specific controller. In other words, hierarchy may 2606be collapsed from leaf towards root when viewed from specific 2607controllers. For example, a given configuration might not care about 2608how memory is distributed beyond a certain level while still wanting 2609to control how CPU cycles are distributed. 2610 2611 2612Thread Granularity 2613------------------ 2614 2615cgroup v1 allowed threads of a process to belong to different cgroups. 2616This didn't make sense for some controllers and those controllers 2617ended up implementing different ways to ignore such situations but 2618much more importantly it blurred the line between API exposed to 2619individual applications and system management interface. 2620 2621Generally, in-process knowledge is available only to the process 2622itself; thus, unlike service-level organization of processes, 2623categorizing threads of a process requires active participation from 2624the application which owns the target process. 2625 2626cgroup v1 had an ambiguously defined delegation model which got abused 2627in combination with thread granularity. cgroups were delegated to 2628individual applications so that they can create and manage their own 2629sub-hierarchies and control resource distributions along them. This 2630effectively raised cgroup to the status of a syscall-like API exposed 2631to lay programs. 2632 2633First of all, cgroup has a fundamentally inadequate interface to be 2634exposed this way. For a process to access its own knobs, it has to 2635extract the path on the target hierarchy from /proc/self/cgroup, 2636construct the path by appending the name of the knob to the path, open 2637and then read and/or write to it. This is not only extremely clunky 2638and unusual but also inherently racy. There is no conventional way to 2639define transaction across the required steps and nothing can guarantee 2640that the process would actually be operating on its own sub-hierarchy. 2641 2642cgroup controllers implemented a number of knobs which would never be 2643accepted as public APIs because they were just adding control knobs to 2644system-management pseudo filesystem. cgroup ended up with interface 2645knobs which were not properly abstracted or refined and directly 2646revealed kernel internal details. These knobs got exposed to 2647individual applications through the ill-defined delegation mechanism 2648effectively abusing cgroup as a shortcut to implementing public APIs 2649without going through the required scrutiny. 2650 2651This was painful for both userland and kernel. Userland ended up with 2652misbehaving and poorly abstracted interfaces and kernel exposing and 2653locked into constructs inadvertently. 2654 2655 2656Competition Between Inner Nodes and Threads 2657------------------------------------------- 2658 2659cgroup v1 allowed threads to be in any cgroups which created an 2660interesting problem where threads belonging to a parent cgroup and its 2661children cgroups competed for resources. This was nasty as two 2662different types of entities competed and there was no obvious way to 2663settle it. Different controllers did different things. 2664 2665The cpu controller considered threads and cgroups as equivalents and 2666mapped nice levels to cgroup weights. This worked for some cases but 2667fell flat when children wanted to be allocated specific ratios of CPU 2668cycles and the number of internal threads fluctuated - the ratios 2669constantly changed as the number of competing entities fluctuated. 2670There also were other issues. The mapping from nice level to weight 2671wasn't obvious or universal, and there were various other knobs which 2672simply weren't available for threads. 2673 2674The io controller implicitly created a hidden leaf node for each 2675cgroup to host the threads. The hidden leaf had its own copies of all 2676the knobs with ``leaf_`` prefixed. While this allowed equivalent 2677control over internal threads, it was with serious drawbacks. It 2678always added an extra layer of nesting which wouldn't be necessary 2679otherwise, made the interface messy and significantly complicated the 2680implementation. 2681 2682The memory controller didn't have a way to control what happened 2683between internal tasks and child cgroups and the behavior was not 2684clearly defined. There were attempts to add ad-hoc behaviors and 2685knobs to tailor the behavior to specific workloads which would have 2686led to problems extremely difficult to resolve in the long term. 2687 2688Multiple controllers struggled with internal tasks and came up with 2689different ways to deal with it; unfortunately, all the approaches were 2690severely flawed and, furthermore, the widely different behaviors 2691made cgroup as a whole highly inconsistent. 2692 2693This clearly is a problem which needs to be addressed from cgroup core 2694in a uniform way. 2695 2696 2697Other Interface Issues 2698---------------------- 2699 2700cgroup v1 grew without oversight and developed a large number of 2701idiosyncrasies and inconsistencies. One issue on the cgroup core side 2702was how an empty cgroup was notified - a userland helper binary was 2703forked and executed for each event. The event delivery wasn't 2704recursive or delegatable. The limitations of the mechanism also led 2705to in-kernel event delivery filtering mechanism further complicating 2706the interface. 2707 2708Controller interfaces were problematic too. An extreme example is 2709controllers completely ignoring hierarchical organization and treating 2710all cgroups as if they were all located directly under the root 2711cgroup. Some controllers exposed a large amount of inconsistent 2712implementation details to userland. 2713 2714There also was no consistency across controllers. When a new cgroup 2715was created, some controllers defaulted to not imposing extra 2716restrictions while others disallowed any resource usage until 2717explicitly configured. Configuration knobs for the same type of 2718control used widely differing naming schemes and formats. Statistics 2719and information knobs were named arbitrarily and used different 2720formats and units even in the same controller. 2721 2722cgroup v2 establishes common conventions where appropriate and updates 2723controllers so that they expose minimal and consistent interfaces. 2724 2725 2726Controller Issues and Remedies 2727------------------------------ 2728 2729Memory 2730~~~~~~ 2731 2732The original lower boundary, the soft limit, is defined as a limit 2733that is per default unset. As a result, the set of cgroups that 2734global reclaim prefers is opt-in, rather than opt-out. The costs for 2735optimizing these mostly negative lookups are so high that the 2736implementation, despite its enormous size, does not even provide the 2737basic desirable behavior. First off, the soft limit has no 2738hierarchical meaning. All configured groups are organized in a global 2739rbtree and treated like equal peers, regardless where they are located 2740in the hierarchy. This makes subtree delegation impossible. Second, 2741the soft limit reclaim pass is so aggressive that it not just 2742introduces high allocation latencies into the system, but also impacts 2743system performance due to overreclaim, to the point where the feature 2744becomes self-defeating. 2745 2746The memory.low boundary on the other hand is a top-down allocated 2747reserve. A cgroup enjoys reclaim protection when it's within its 2748effective low, which makes delegation of subtrees possible. It also 2749enjoys having reclaim pressure proportional to its overage when 2750above its effective low. 2751 2752The original high boundary, the hard limit, is defined as a strict 2753limit that can not budge, even if the OOM killer has to be called. 2754But this generally goes against the goal of making the most out of the 2755available memory. The memory consumption of workloads varies during 2756runtime, and that requires users to overcommit. But doing that with a 2757strict upper limit requires either a fairly accurate prediction of the 2758working set size or adding slack to the limit. Since working set size 2759estimation is hard and error prone, and getting it wrong results in 2760OOM kills, most users tend to err on the side of a looser limit and 2761end up wasting precious resources. 2762 2763The memory.high boundary on the other hand can be set much more 2764conservatively. When hit, it throttles allocations by forcing them 2765into direct reclaim to work off the excess, but it never invokes the 2766OOM killer. As a result, a high boundary that is chosen too 2767aggressively will not terminate the processes, but instead it will 2768lead to gradual performance degradation. The user can monitor this 2769and make corrections until the minimal memory footprint that still 2770gives acceptable performance is found. 2771 2772In extreme cases, with many concurrent allocations and a complete 2773breakdown of reclaim progress within the group, the high boundary can 2774be exceeded. But even then it's mostly better to satisfy the 2775allocation from the slack available in other groups or the rest of the 2776system than killing the group. Otherwise, memory.max is there to 2777limit this type of spillover and ultimately contain buggy or even 2778malicious applications. 2779 2780Setting the original memory.limit_in_bytes below the current usage was 2781subject to a race condition, where concurrent charges could cause the 2782limit setting to fail. memory.max on the other hand will first set the 2783limit to prevent new charges, and then reclaim and OOM kill until the 2784new limit is met - or the task writing to memory.max is killed. 2785 2786The combined memory+swap accounting and limiting is replaced by real 2787control over swap space. 2788 2789The main argument for a combined memory+swap facility in the original 2790cgroup design was that global or parental pressure would always be 2791able to swap all anonymous memory of a child group, regardless of the 2792child's own (possibly untrusted) configuration. However, untrusted 2793groups can sabotage swapping by other means - such as referencing its 2794anonymous memory in a tight loop - and an admin can not assume full 2795swappability when overcommitting untrusted jobs. 2796 2797For trusted jobs, on the other hand, a combined counter is not an 2798intuitive userspace interface, and it flies in the face of the idea 2799that cgroup controllers should account and limit specific physical 2800resources. Swap space is a resource like all others in the system, 2801and that's why unified hierarchy allows distributing it separately. 2802