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 [Whenever any new section is added to this document, please also add 19 an entry here.] 20 21 1. Introduction 22 1-1. Terminology 23 1-2. What is cgroup? 24 2. Basic Operations 25 2-1. Mounting 26 2-2. Organizing Processes and Threads 27 2-2-1. Processes 28 2-2-2. Threads 29 2-3. [Un]populated Notification 30 2-4. Controlling Controllers 31 2-4-1. Availability 32 2-4-2. Enabling and Disabling 33 2-4-3. Top-down Constraint 34 2-4-4. No Internal Process Constraint 35 2-5. Delegation 36 2-5-1. Model of Delegation 37 2-5-2. Delegation Containment 38 2-6. Guidelines 39 2-6-1. Organize Once and Control 40 2-6-2. Avoid Name Collisions 41 3. Resource Distribution Models 42 3-1. Weights 43 3-2. Limits 44 3-3. Protections 45 3-4. Allocations 46 4. Interface Files 47 4-1. Format 48 4-2. Conventions 49 4-3. Core Interface Files 50 5. Controllers 51 5-1. CPU 52 5-1-1. CPU Interface Files 53 5-2. Memory 54 5-2-1. Memory Interface Files 55 5-2-2. Usage Guidelines 56 5-2-3. Reclaim Protection 57 5-2-4. Memory Ownership 58 5-3. IO 59 5-3-1. IO Interface Files 60 5-3-2. Writeback 61 5-3-3. IO Latency 62 5-3-3-1. How IO Latency Throttling Works 63 5-3-3-2. IO Latency Interface Files 64 5-3-4. IO Priority 65 5-4. PID 66 5-4-1. PID Interface Files 67 5-5. Cpuset 68 5.5-1. Cpuset Interface Files 69 5-6. Device controller 70 5-7. RDMA 71 5-7-1. RDMA Interface Files 72 5-8. DMEM 73 5-8-1. DMEM Interface Files 74 5-9. HugeTLB 75 5.9-1. HugeTLB Interface Files 76 5-10. Misc 77 5.10-1 Misc Interface Files 78 5.10-2 Migration and Ownership 79 5-11. Others 80 5-11-1. perf_event 81 5-N. Non-normative information 82 5-N-1. CPU controller root cgroup process behaviour 83 5-N-2. IO controller root cgroup process behaviour 84 6. Namespace 85 6-1. Basics 86 6-2. The Root and Views 87 6-3. Migration and setns(2) 88 6-4. Interaction with Other Namespaces 89 P. Information on Kernel Programming 90 P-1. Filesystem Support for Writeback 91 D. Deprecated v1 Core Features 92 R. Issues with v1 and Rationales for v2 93 R-1. Multiple Hierarchies 94 R-2. Thread Granularity 95 R-3. Competition Between Inner Nodes and Threads 96 R-4. Other Interface Issues 97 R-5. Controller Issues and Remedies 98 R-5-1. Memory 99 100 101Introduction 102============ 103 104Terminology 105----------- 106 107"cgroup" stands for "control group" and is never capitalized. The 108singular form is used to designate the whole feature and also as a 109qualifier as in "cgroup controllers". When explicitly referring to 110multiple individual control groups, the plural form "cgroups" is used. 111 112 113What is cgroup? 114--------------- 115 116cgroup is a mechanism to organize processes hierarchically and 117distribute system resources along the hierarchy in a controlled and 118configurable manner. 119 120cgroup is largely composed of two parts - the core and controllers. 121cgroup core is primarily responsible for hierarchically organizing 122processes. A cgroup controller is usually responsible for 123distributing a specific type of system resource along the hierarchy 124although there are utility controllers which serve purposes other than 125resource distribution. 126 127cgroups form a tree structure and every process in the system belongs 128to one and only one cgroup. All threads of a process belong to the 129same cgroup. On creation, all processes are put in the cgroup that 130the parent process belongs to at the time. A process can be migrated 131to another cgroup. Migration of a process doesn't affect already 132existing descendant processes. 133 134Following certain structural constraints, controllers may be enabled or 135disabled selectively on a cgroup. All controller behaviors are 136hierarchical - if a controller is enabled on a cgroup, it affects all 137processes which belong to the cgroups consisting the inclusive 138sub-hierarchy of the cgroup. When a controller is enabled on a nested 139cgroup, it always restricts the resource distribution further. The 140restrictions set closer to the root in the hierarchy can not be 141overridden from further away. 142 143 144Basic Operations 145================ 146 147Mounting 148-------- 149 150Unlike v1, cgroup v2 has only single hierarchy. The cgroup v2 151hierarchy can be mounted with the following mount command:: 152 153 # mount -t cgroup2 none $MOUNT_POINT 154 155cgroup2 filesystem has the magic number 0x63677270 ("cgrp"). All 156controllers which support v2 and are not bound to a v1 hierarchy are 157automatically bound to the v2 hierarchy and show up at the root. 158Controllers which are not in active use in the v2 hierarchy can be 159bound to other hierarchies. This allows mixing v2 hierarchy with the 160legacy v1 multiple hierarchies in a fully backward compatible way. 161 162A controller can be moved across hierarchies only after the controller 163is no longer referenced in its current hierarchy. Because per-cgroup 164controller states are destroyed asynchronously and controllers may 165have lingering references, a controller may not show up immediately on 166the v2 hierarchy after the final umount of the previous hierarchy. 167Similarly, a controller should be fully disabled to be moved out of 168the unified hierarchy and it may take some time for the disabled 169controller to become available for other hierarchies; furthermore, due 170to inter-controller dependencies, other controllers may need to be 171disabled too. 172 173While useful for development and manual configurations, moving 174controllers dynamically between the v2 and other hierarchies is 175strongly discouraged for production use. It is recommended to decide 176the hierarchies and controller associations before starting using the 177controllers after system boot. 178 179During transition to v2, system management software might still 180automount the v1 cgroup filesystem and so hijack all controllers 181during boot, before manual intervention is possible. To make testing 182and experimenting easier, the kernel parameter cgroup_no_v1= allows 183disabling controllers in v1 and make them always available in v2. 184 185cgroup v2 currently supports the following mount options. 186 187 nsdelegate 188 Consider cgroup namespaces as delegation boundaries. This 189 option is system wide and can only be set on mount or modified 190 through remount from the init namespace. The mount option is 191 ignored on non-init namespace mounts. Please refer to the 192 Delegation section for details. 193 194 favordynmods 195 Reduce the latencies of dynamic cgroup modifications such as 196 task migrations and controller on/offs at the cost of making 197 hot path operations such as forks and exits more expensive. 198 The static usage pattern of creating a cgroup, enabling 199 controllers, and then seeding it with CLONE_INTO_CGROUP is 200 not affected by this option. 201 202 memory_localevents 203 Only populate memory.events with data for the current cgroup, 204 and not any subtrees. This is legacy behaviour, the default 205 behaviour without this option is to include subtree counts. 206 This option is system wide and can only be set on mount or 207 modified through remount from the init namespace. The mount 208 option is ignored on non-init namespace mounts. 209 210 memory_recursiveprot 211 Recursively apply memory.min and memory.low protection to 212 entire subtrees, without requiring explicit downward 213 propagation into leaf cgroups. This allows protecting entire 214 subtrees from one another, while retaining free competition 215 within those subtrees. This should have been the default 216 behavior but is a mount-option to avoid regressing setups 217 relying on the original semantics (e.g. specifying bogusly 218 high 'bypass' protection values at higher tree levels). 219 220 memory_hugetlb_accounting 221 Count HugeTLB memory usage towards the cgroup's overall 222 memory usage for the memory controller (for the purpose of 223 statistics reporting and memory protection). This is a new 224 behavior that could regress existing setups, so it must be 225 explicitly opted in with this mount option. 226 227 A few caveats to keep in mind: 228 229 * There is no HugeTLB pool management involved in the memory 230 controller. The pre-allocated pool does not belong to anyone. 231 Specifically, when a new HugeTLB folio is allocated to 232 the pool, it is not accounted for from the perspective of the 233 memory controller. It is only charged to a cgroup when it is 234 actually used (for e.g at page fault time). Host memory 235 overcommit management has to consider this when configuring 236 hard limits. In general, HugeTLB pool management should be 237 done via other mechanisms (such as the HugeTLB controller). 238 * Failure to charge a HugeTLB folio to the memory controller 239 results in SIGBUS. This could happen even if the HugeTLB pool 240 still has pages available (but the cgroup limit is hit and 241 reclaim attempt fails). 242 * Charging HugeTLB memory towards the memory controller affects 243 memory protection and reclaim dynamics. Any userspace tuning 244 (of low, min limits for e.g) needs to take this into account. 245 * HugeTLB pages utilized while this option is not selected 246 will not be tracked by the memory controller (even if cgroup 247 v2 is remounted later on). 248 249 pids_localevents 250 The option restores v1-like behavior of pids.events:max, that is only 251 local (inside cgroup proper) fork failures are counted. Without this 252 option pids.events.max represents any pids.max enforcemnt across 253 cgroup's subtree. 254 255 256 257Organizing Processes and Threads 258-------------------------------- 259 260Processes 261~~~~~~~~~ 262 263Initially, only the root cgroup exists to which all processes belong. 264A child cgroup can be created by creating a sub-directory:: 265 266 # mkdir $CGROUP_NAME 267 268A given cgroup may have multiple child cgroups forming a tree 269structure. Each cgroup has a read-writable interface file 270"cgroup.procs". When read, it lists the PIDs of all processes which 271belong to the cgroup one-per-line. The PIDs are not ordered and the 272same PID may show up more than once if the process got moved to 273another cgroup and then back or the PID got recycled while reading. 274 275A process can be migrated into a cgroup by writing its PID to the 276target cgroup's "cgroup.procs" file. Only one process can be migrated 277on a single write(2) call. If a process is composed of multiple 278threads, writing the PID of any thread migrates all threads of the 279process. 280 281When a process forks a child process, the new process is born into the 282cgroup that the forking process belongs to at the time of the 283operation. After exit, a process stays associated with the cgroup 284that it belonged to at the time of exit until it's reaped; however, a 285zombie process does not appear in "cgroup.procs" and thus can't be 286moved to another cgroup. 287 288A cgroup which doesn't have any children or live processes can be 289destroyed by removing the directory. Note that a cgroup which doesn't 290have any children and is associated only with zombie processes is 291considered empty and can be removed:: 292 293 # rmdir $CGROUP_NAME 294 295"/proc/$PID/cgroup" lists a process's cgroup membership. If legacy 296cgroup is in use in the system, this file may contain multiple lines, 297one for each hierarchy. The entry for cgroup v2 is always in the 298format "0::$PATH":: 299 300 # cat /proc/842/cgroup 301 ... 302 0::/test-cgroup/test-cgroup-nested 303 304If the process becomes a zombie and the cgroup it was associated with 305is removed subsequently, " (deleted)" is appended to the path:: 306 307 # cat /proc/842/cgroup 308 ... 309 0::/test-cgroup/test-cgroup-nested (deleted) 310 311 312Threads 313~~~~~~~ 314 315cgroup v2 supports thread granularity for a subset of controllers to 316support use cases requiring hierarchical resource distribution across 317the threads of a group of processes. By default, all threads of a 318process belong to the same cgroup, which also serves as the resource 319domain to host resource consumptions which are not specific to a 320process or thread. The thread mode allows threads to be spread across 321a subtree while still maintaining the common resource domain for them. 322 323Controllers which support thread mode are called threaded controllers. 324The ones which don't are called domain controllers. 325 326Marking a cgroup threaded makes it join the resource domain of its 327parent as a threaded cgroup. The parent may be another threaded 328cgroup whose resource domain is further up in the hierarchy. The root 329of a threaded subtree, that is, the nearest ancestor which is not 330threaded, is called threaded domain or thread root interchangeably and 331serves as the resource domain for the entire subtree. 332 333Inside a threaded subtree, threads of a process can be put in 334different cgroups and are not subject to the no internal process 335constraint - threaded controllers can be enabled on non-leaf cgroups 336whether they have threads in them or not. 337 338As the threaded domain cgroup hosts all the domain resource 339consumptions of the subtree, it is considered to have internal 340resource consumptions whether there are processes in it or not and 341can't have populated child cgroups which aren't threaded. Because the 342root cgroup is not subject to no internal process constraint, it can 343serve both as a threaded domain and a parent to domain cgroups. 344 345The current operation mode or type of the cgroup is shown in the 346"cgroup.type" file which indicates whether the cgroup is a normal 347domain, a domain which is serving as the domain of a threaded subtree, 348or a threaded cgroup. 349 350On creation, a cgroup is always a domain cgroup and can be made 351threaded by writing "threaded" to the "cgroup.type" file. The 352operation is single direction:: 353 354 # echo threaded > cgroup.type 355 356Once threaded, the cgroup can't be made a domain again. To enable the 357thread mode, the following conditions must be met. 358 359- As the cgroup will join the parent's resource domain. The parent 360 must either be a valid (threaded) domain or a threaded cgroup. 361 362- When the parent is an unthreaded domain, it must not have any domain 363 controllers enabled or populated domain children. The root is 364 exempt from this requirement. 365 366Topology-wise, a cgroup can be in an invalid state. Please consider 367the following topology:: 368 369 A (threaded domain) - B (threaded) - C (domain, just created) 370 371C is created as a domain but isn't connected to a parent which can 372host child domains. C can't be used until it is turned into a 373threaded cgroup. "cgroup.type" file will report "domain (invalid)" in 374these cases. Operations which fail due to invalid topology use 375EOPNOTSUPP as the errno. 376 377A domain cgroup is turned into a threaded domain when one of its child 378cgroup becomes threaded or threaded controllers are enabled in the 379"cgroup.subtree_control" file while there are processes in the cgroup. 380A threaded domain reverts to a normal domain when the conditions 381clear. 382 383When read, "cgroup.threads" contains the list of the thread IDs of all 384threads in the cgroup. Except that the operations are per-thread 385instead of per-process, "cgroup.threads" has the same format and 386behaves the same way as "cgroup.procs". While "cgroup.threads" can be 387written to in any cgroup, as it can only move threads inside the same 388threaded domain, its operations are confined inside each threaded 389subtree. 390 391The threaded domain cgroup serves as the resource domain for the whole 392subtree, and, while the threads can be scattered across the subtree, 393all the processes are considered to be in the threaded domain cgroup. 394"cgroup.procs" in a threaded domain cgroup contains the PIDs of all 395processes in the subtree and is not readable in the subtree proper. 396However, "cgroup.procs" can be written to from anywhere in the subtree 397to migrate all threads of the matching process to the cgroup. 398 399Only threaded controllers can be enabled in a threaded subtree. When 400a threaded controller is enabled inside a threaded subtree, it only 401accounts for and controls resource consumptions associated with the 402threads in the cgroup and its descendants. All consumptions which 403aren't tied to a specific thread belong to the threaded domain cgroup. 404 405Because a threaded subtree is exempt from no internal process 406constraint, a threaded controller must be able to handle competition 407between threads in a non-leaf cgroup and its child cgroups. Each 408threaded controller defines how such competitions are handled. 409 410Currently, the following controllers are threaded and can be enabled 411in a threaded cgroup:: 412 413- cpu 414- cpuset 415- perf_event 416- pids 417 418[Un]populated Notification 419-------------------------- 420 421Each non-root cgroup has a "cgroup.events" file which contains 422"populated" field indicating whether the cgroup's sub-hierarchy has 423live processes in it. Its value is 0 if there is no live process in 424the cgroup and its descendants; otherwise, 1. poll and [id]notify 425events are triggered when the value changes. This can be used, for 426example, to start a clean-up operation after all processes of a given 427sub-hierarchy have exited. The populated state updates and 428notifications are recursive. Consider the following sub-hierarchy 429where the numbers in the parentheses represent the numbers of processes 430in each cgroup:: 431 432 A(4) - B(0) - C(1) 433 \ D(0) 434 435A, B and C's "populated" fields would be 1 while D's 0. After the one 436process in C exits, B and C's "populated" fields would flip to "0" and 437file modified events will be generated on the "cgroup.events" files of 438both cgroups. 439 440 441Controlling Controllers 442----------------------- 443 444Availability 445~~~~~~~~~~~~ 446 447A controller is available in a cgroup when it is supported by the kernel (i.e., 448compiled in, not disabled and not attached to a v1 hierarchy) and listed in the 449"cgroup.controllers" file. Availability means the controller's interface files 450are exposed in the cgroup’s directory, allowing the distribution of the target 451resource to be observed or controlled within that cgroup. 452 453Enabling and Disabling 454~~~~~~~~~~~~~~~~~~~~~~ 455 456Each cgroup has a "cgroup.controllers" file which lists all 457controllers available for the cgroup to enable:: 458 459 # cat cgroup.controllers 460 cpu io memory 461 462No controller is enabled by default. Controllers can be enabled and 463disabled by writing to the "cgroup.subtree_control" file:: 464 465 # echo "+cpu +memory -io" > cgroup.subtree_control 466 467Only controllers which are listed in "cgroup.controllers" can be 468enabled. When multiple operations are specified as above, either they 469all succeed or fail. If multiple operations on the same controller 470are specified, the last one is effective. 471 472Enabling a controller in a cgroup indicates that the distribution of 473the target resource across its immediate children will be controlled. 474Consider the following sub-hierarchy. The enabled controllers are 475listed in parentheses:: 476 477 A(cpu,memory) - B(memory) - C() 478 \ D() 479 480As A has "cpu" and "memory" enabled, A will control the distribution 481of CPU cycles and memory to its children, in this case, B. As B has 482"memory" enabled but not "CPU", C and D will compete freely on CPU 483cycles but their division of memory available to B will be controlled. 484 485As a controller regulates the distribution of the target resource to 486the cgroup's children, enabling it creates the controller's interface 487files in the child cgroups. In the above example, enabling "cpu" on B 488would create the "cpu." prefixed controller interface files in C and 489D. Likewise, disabling "memory" from B would remove the "memory." 490prefixed controller interface files from C and D. This means that the 491controller interface files - anything which doesn't start with 492"cgroup." are owned by the parent rather than the cgroup itself. 493 494 495Top-down Constraint 496~~~~~~~~~~~~~~~~~~~ 497 498Resources are distributed top-down and a cgroup can further distribute 499a resource only if the resource has been distributed to it from the 500parent. This means that all non-root "cgroup.subtree_control" files 501can only contain controllers which are enabled in the parent's 502"cgroup.subtree_control" file. A controller can be enabled only if 503the parent has the controller enabled and a controller can't be 504disabled if one or more children have it enabled. 505 506 507No Internal Process Constraint 508~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~ 509 510Non-root cgroups can distribute domain resources to their children 511only when they don't have any processes of their own. In other words, 512only domain cgroups which don't contain any processes can have domain 513controllers enabled in their "cgroup.subtree_control" files. 514 515This guarantees that, when a domain controller is looking at the part 516of the hierarchy which has it enabled, processes are always only on 517the leaves. This rules out situations where child cgroups compete 518against internal processes of the parent. 519 520The root cgroup is exempt from this restriction. Root contains 521processes and anonymous resource consumption which can't be associated 522with any other cgroups and requires special treatment from most 523controllers. How resource consumption in the root cgroup is governed 524is up to each controller (for more information on this topic please 525refer to the Non-normative information section in the Controllers 526chapter). 527 528Note that the restriction doesn't get in the way if there is no 529enabled controller in the cgroup's "cgroup.subtree_control". This is 530important as otherwise it wouldn't be possible to create children of a 531populated cgroup. To control resource distribution of a cgroup, the 532cgroup must create children and transfer all its processes to the 533children before enabling controllers in its "cgroup.subtree_control" 534file. 535 536 537Delegation 538---------- 539 540Model of Delegation 541~~~~~~~~~~~~~~~~~~~ 542 543A cgroup can be delegated in two ways. First, to a less privileged 544user by granting write access of the directory and its "cgroup.procs", 545"cgroup.threads" and "cgroup.subtree_control" files to the user. 546Second, if the "nsdelegate" mount option is set, automatically to a 547cgroup namespace on namespace creation. 548 549Because the resource control interface files in a given directory 550control the distribution of the parent's resources, the delegatee 551shouldn't be allowed to write to them. For the first method, this is 552achieved by not granting access to these files. For the second, files 553outside the namespace should be hidden from the delegatee by the means 554of at least mount namespacing, and the kernel rejects writes to all 555files on a namespace root from inside the cgroup namespace, except for 556those files listed in "/sys/kernel/cgroup/delegate" (including 557"cgroup.procs", "cgroup.threads", "cgroup.subtree_control", etc.). 558 559The end results are equivalent for both delegation types. Once 560delegated, the user can build sub-hierarchy under the directory, 561organize processes inside it as it sees fit and further distribute the 562resources it received from the parent. The limits and other settings 563of all resource controllers are hierarchical and regardless of what 564happens in the delegated sub-hierarchy, nothing can escape the 565resource restrictions imposed by the parent. 566 567Currently, cgroup doesn't impose any restrictions on the number of 568cgroups in or nesting depth of a delegated sub-hierarchy; however, 569this may be limited explicitly in the future. 570 571 572Delegation Containment 573~~~~~~~~~~~~~~~~~~~~~~ 574 575A delegated sub-hierarchy is contained in the sense that processes 576can't be moved into or out of the sub-hierarchy by the delegatee. 577 578For delegations to a less privileged user, this is achieved by 579requiring the following conditions for a process with a non-root euid 580to migrate a target process into a cgroup by writing its PID to the 581"cgroup.procs" file. 582 583- The writer must have write access to the "cgroup.procs" file. 584 585- The writer must have write access to the "cgroup.procs" file of the 586 common ancestor of the source and destination cgroups. 587 588The above two constraints ensure that while a delegatee may migrate 589processes around freely in the delegated sub-hierarchy it can't pull 590in from or push out to outside the sub-hierarchy. 591 592For an example, let's assume cgroups C0 and C1 have been delegated to 593user U0 who created C00, C01 under C0 and C10 under C1 as follows and 594all processes under C0 and C1 belong to U0:: 595 596 ~~~~~~~~~~~~~ - C0 - C00 597 ~ cgroup ~ \ C01 598 ~ hierarchy ~ 599 ~~~~~~~~~~~~~ - C1 - C10 600 601Let's also say U0 wants to write the PID of a process which is 602currently in C10 into "C00/cgroup.procs". U0 has write access to the 603file; however, the common ancestor of the source cgroup C10 and the 604destination cgroup C00 is above the points of delegation and U0 would 605not have write access to its "cgroup.procs" files and thus the write 606will be denied with -EACCES. 607 608For delegations to namespaces, containment is achieved by requiring 609that both the source and destination cgroups are reachable from the 610namespace of the process which is attempting the migration. If either 611is not reachable, the migration is rejected with -ENOENT. 612 613 614Guidelines 615---------- 616 617Organize Once and Control 618~~~~~~~~~~~~~~~~~~~~~~~~~ 619 620Migrating a process across cgroups is a relatively expensive operation 621and stateful resources such as memory are not moved together with the 622process. This is an explicit design decision as there often exist 623inherent trade-offs between migration and various hot paths in terms 624of synchronization cost. 625 626As such, migrating processes across cgroups frequently as a means to 627apply different resource restrictions is discouraged. A workload 628should be assigned to a cgroup according to the system's logical and 629resource structure once on start-up. Dynamic adjustments to resource 630distribution can be made by changing controller configuration through 631the interface files. 632 633 634Avoid Name Collisions 635~~~~~~~~~~~~~~~~~~~~~ 636 637Interface files for a cgroup and its children cgroups occupy the same 638directory and it is possible to create children cgroups which collide 639with interface files. 640 641All cgroup core interface files are prefixed with "cgroup." and each 642controller's interface files are prefixed with the controller name and 643a dot. A controller's name is composed of lower case alphabets and 644'_'s but never begins with an '_' so it can be used as the prefix 645character for collision avoidance. Also, interface file names won't 646start or end with terms which are often used in categorizing workloads 647such as job, service, slice, unit or workload. 648 649cgroup doesn't do anything to prevent name collisions and it's the 650user's responsibility to avoid them. 651 652 653Resource Distribution Models 654============================ 655 656cgroup controllers implement several resource distribution schemes 657depending on the resource type and expected use cases. This section 658describes major schemes in use along with their expected behaviors. 659 660 661Weights 662------- 663 664A parent's resource is distributed by adding up the weights of all 665active children and giving each the fraction matching the ratio of its 666weight against the sum. As only children which can make use of the 667resource at the moment participate in the distribution, this is 668work-conserving. Due to the dynamic nature, this model is usually 669used for stateless resources. 670 671All weights are in the range [1, 10000] with the default at 100. This 672allows symmetric multiplicative biases in both directions at fine 673enough granularity while staying in the intuitive range. 674 675As long as the weight is in range, all configuration combinations are 676valid and there is no reason to reject configuration changes or 677process migrations. 678 679"cpu.weight" proportionally distributes CPU cycles to active children 680and is an example of this type. 681 682 683.. _cgroupv2-limits-distributor: 684 685Limits 686------ 687 688A child can only consume up to the configured amount of the resource. 689Limits can be over-committed - the sum of the limits of children can 690exceed the amount of resource available to the parent. 691 692Limits are in the range [0, max] and defaults to "max", which is noop. 693 694As limits can be over-committed, all configuration combinations are 695valid and there is no reason to reject configuration changes or 696process migrations. 697 698"io.max" limits the maximum BPS and/or IOPS that a cgroup can consume 699on an IO device and is an example of this type. 700 701.. _cgroupv2-protections-distributor: 702 703Protections 704----------- 705 706A cgroup is protected up to the configured amount of the resource 707as long as the usages of all its ancestors are under their 708protected levels. Protections can be hard guarantees or best effort 709soft boundaries. Protections can also be over-committed in which case 710only up to the amount available to the parent is protected among 711children. 712 713Protections are in the range [0, max] and defaults to 0, which is 714noop. 715 716As protections can be over-committed, all configuration combinations 717are valid and there is no reason to reject configuration changes or 718process migrations. 719 720"memory.low" implements best-effort memory protection and is an 721example of this type. 722 723 724Allocations 725----------- 726 727A cgroup is exclusively allocated a certain amount of a finite 728resource. Allocations can't be over-committed - the sum of the 729allocations of children can not exceed the amount of resource 730available to the parent. 731 732Allocations are in the range [0, max] and defaults to 0, which is no 733resource. 734 735As allocations can't be over-committed, some configuration 736combinations are invalid and should be rejected. Also, if the 737resource is mandatory for execution of processes, process migrations 738may be rejected. 739 740 741Interface Files 742=============== 743 744Format 745------ 746 747All interface files should be in one of the following formats whenever 748possible:: 749 750 New-line separated values 751 (when only one value can be written at once) 752 753 VAL0\n 754 VAL1\n 755 ... 756 757 Space separated values 758 (when read-only or multiple values can be written at once) 759 760 VAL0 VAL1 ...\n 761 762 Flat keyed 763 764 KEY0 VAL0\n 765 KEY1 VAL1\n 766 ... 767 768 Nested keyed 769 770 KEY0 SUB_KEY0=VAL00 SUB_KEY1=VAL01... 771 KEY1 SUB_KEY0=VAL10 SUB_KEY1=VAL11... 772 ... 773 774For a writable file, the format for writing should generally match 775reading; however, controllers may allow omitting later fields or 776implement restricted shortcuts for most common use cases. 777 778For both flat and nested keyed files, only the values for a single key 779can be written at a time. For nested keyed files, the sub key pairs 780may be specified in any order and not all pairs have to be specified. 781 782 783Conventions 784----------- 785 786- Settings for a single feature should be contained in a single file. 787 788- The root cgroup should be exempt from resource control and thus 789 shouldn't have resource control interface files. 790 791- The default time unit is microseconds. If a different unit is ever 792 used, an explicit unit suffix must be present. 793 794- A parts-per quantity should use a percentage decimal with at least 795 two digit fractional part - e.g. 13.40. 796 797- If a controller implements weight based resource distribution, its 798 interface file should be named "weight" and have the range [1, 799 10000] with 100 as the default. The values are chosen to allow 800 enough and symmetric bias in both directions while keeping it 801 intuitive (the default is 100%). 802 803- If a controller implements an absolute resource guarantee and/or 804 limit, the interface files should be named "min" and "max" 805 respectively. If a controller implements best effort resource 806 guarantee and/or limit, the interface files should be named "low" 807 and "high" respectively. 808 809 In the above four control files, the special token "max" should be 810 used to represent upward infinity for both reading and writing. 811 812- If a setting has a configurable default value and keyed specific 813 overrides, the default entry should be keyed with "default" and 814 appear as the first entry in the file. 815 816 The default value can be updated by writing either "default $VAL" or 817 "$VAL". 818 819 When writing to update a specific override, "default" can be used as 820 the value to indicate removal of the override. Override entries 821 with "default" as the value must not appear when read. 822 823 For example, a setting which is keyed by major:minor device numbers 824 with integer values may look like the following:: 825 826 # cat cgroup-example-interface-file 827 default 150 828 8:0 300 829 830 The default value can be updated by:: 831 832 # echo 125 > cgroup-example-interface-file 833 834 or:: 835 836 # echo "default 125" > cgroup-example-interface-file 837 838 An override can be set by:: 839 840 # echo "8:16 170" > cgroup-example-interface-file 841 842 and cleared by:: 843 844 # echo "8:0 default" > cgroup-example-interface-file 845 # cat cgroup-example-interface-file 846 default 125 847 8:16 170 848 849- For events which are not very high frequency, an interface file 850 "events" should be created which lists event key value pairs. 851 Whenever a notifiable event happens, file modified event should be 852 generated on the file. 853 854 855Core Interface Files 856-------------------- 857 858All cgroup core files are prefixed with "cgroup." 859 860 cgroup.type 861 A read-write single value file which exists on non-root 862 cgroups. 863 864 When read, it indicates the current type of the cgroup, which 865 can be one of the following values. 866 867 - "domain" : A normal valid domain cgroup. 868 869 - "domain threaded" : A threaded domain cgroup which is 870 serving as the root of a threaded subtree. 871 872 - "domain invalid" : A cgroup which is in an invalid state. 873 It can't be populated or have controllers enabled. It may 874 be allowed to become a threaded cgroup. 875 876 - "threaded" : A threaded cgroup which is a member of a 877 threaded subtree. 878 879 A cgroup can be turned into a threaded cgroup by writing 880 "threaded" to this file. 881 882 cgroup.procs 883 A read-write new-line separated values file which exists on 884 all cgroups. 885 886 When read, it lists the PIDs of all processes which belong to 887 the cgroup one-per-line. The PIDs are not ordered and the 888 same PID may show up more than once if the process got moved 889 to another cgroup and then back or the PID got recycled while 890 reading. 891 892 A PID can be written to migrate the process associated with 893 the PID to the cgroup. The writer should match all of the 894 following conditions. 895 896 - It must have write access to the "cgroup.procs" file. 897 898 - It must have write access to the "cgroup.procs" file of the 899 common ancestor of the source and destination cgroups. 900 901 When delegating a sub-hierarchy, write access to this file 902 should be granted along with the containing directory. 903 904 In a threaded cgroup, reading this file fails with EOPNOTSUPP 905 as all the processes belong to the thread root. Writing is 906 supported and moves every thread of the process to the cgroup. 907 908 cgroup.threads 909 A read-write new-line separated values file which exists on 910 all cgroups. 911 912 When read, it lists the TIDs of all threads which belong to 913 the cgroup one-per-line. The TIDs are not ordered and the 914 same TID may show up more than once if the thread got moved to 915 another cgroup and then back or the TID got recycled while 916 reading. 917 918 A TID can be written to migrate the thread associated with the 919 TID to the cgroup. The writer should match all of the 920 following conditions. 921 922 - It must have write access to the "cgroup.threads" file. 923 924 - The cgroup that the thread is currently in must be in the 925 same resource domain as the destination cgroup. 926 927 - It must have write access to the "cgroup.procs" file of the 928 common ancestor of the source and destination cgroups. 929 930 When delegating a sub-hierarchy, write access to this file 931 should be granted along with the containing directory. 932 933 cgroup.controllers 934 A read-only space separated values file which exists on all 935 cgroups. 936 937 It shows space separated list of all controllers available to 938 the cgroup. The controllers are not ordered. 939 940 cgroup.subtree_control 941 A read-write space separated values file which exists on all 942 cgroups. Starts out empty. 943 944 When read, it shows space separated list of the controllers 945 which are enabled to control resource distribution from the 946 cgroup to its children. 947 948 Space separated list of controllers prefixed with '+' or '-' 949 can be written to enable or disable controllers. A controller 950 name prefixed with '+' enables the controller and '-' 951 disables. If a controller appears more than once on the list, 952 the last one is effective. When multiple enable and disable 953 operations are specified, either all succeed or all fail. 954 955 cgroup.events 956 A read-only flat-keyed file which exists on non-root cgroups. 957 The following entries are defined. Unless specified 958 otherwise, a value change in this file generates a file 959 modified event. 960 961 populated 962 1 if the cgroup or its descendants contains any live 963 processes; otherwise, 0. 964 frozen 965 1 if the cgroup is frozen; otherwise, 0. 966 967 cgroup.max.descendants 968 A read-write single value files. The default is "max". 969 970 Maximum allowed number of descent cgroups. 971 If the actual number of descendants is equal or larger, 972 an attempt to create a new cgroup in the hierarchy will fail. 973 974 cgroup.max.depth 975 A read-write single value files. The default is "max". 976 977 Maximum allowed descent depth below the current cgroup. 978 If the actual descent depth is equal or larger, 979 an attempt to create a new child cgroup will fail. 980 981 cgroup.stat 982 A read-only flat-keyed file with the following entries: 983 984 nr_descendants 985 Total number of visible descendant cgroups. 986 987 nr_dying_descendants 988 Total number of dying descendant cgroups. A cgroup becomes 989 dying after being deleted by a user. The cgroup will remain 990 in dying state for some time undefined time (which can depend 991 on system load) before being completely destroyed. 992 993 A process can't enter a dying cgroup under any circumstances, 994 a dying cgroup can't revive. 995 996 A dying cgroup can consume system resources not exceeding 997 limits, which were active at the moment of cgroup deletion. 998 999 nr_subsys_<cgroup_subsys> 1000 Total number of live cgroup subsystems (e.g memory 1001 cgroup) at and beneath the current cgroup. 1002 1003 nr_dying_subsys_<cgroup_subsys> 1004 Total number of dying cgroup subsystems (e.g. memory 1005 cgroup) at and beneath the current cgroup. 1006 1007 cgroup.stat.local 1008 A read-only flat-keyed file which exists in non-root cgroups. 1009 The following entry is defined: 1010 1011 frozen_usec 1012 Cumulative time that this cgroup has spent between freezing and 1013 thawing, regardless of whether by self or ancestor groups. 1014 NB: (not) reaching "frozen" state is not accounted here. 1015 1016 Using the following ASCII representation of a cgroup's freezer 1017 state, :: 1018 1019 1 _____ 1020 frozen 0 __/ \__ 1021 ab cd 1022 1023 the duration being measured is the span between a and c. 1024 1025 cgroup.freeze 1026 A read-write single value file which exists on non-root cgroups. 1027 Allowed values are "0" and "1". The default is "0". 1028 1029 Writing "1" to the file causes freezing of the cgroup and all 1030 descendant cgroups. This means that all belonging processes will 1031 be stopped and will not run until the cgroup will be explicitly 1032 unfrozen. Freezing of the cgroup may take some time; when this action 1033 is completed, the "frozen" value in the cgroup.events control file 1034 will be updated to "1" and the corresponding notification will be 1035 issued. 1036 1037 A cgroup can be frozen either by its own settings, or by settings 1038 of any ancestor cgroups. If any of ancestor cgroups is frozen, the 1039 cgroup will remain frozen. 1040 1041 Processes in the frozen cgroup can be killed by a fatal signal. 1042 They also can enter and leave a frozen cgroup: either by an explicit 1043 move by a user, or if freezing of the cgroup races with fork(). 1044 If a process is moved to a frozen cgroup, it stops. If a process is 1045 moved out of a frozen cgroup, it becomes running. 1046 1047 Frozen status of a cgroup doesn't affect any cgroup tree operations: 1048 it's possible to delete a frozen (and empty) cgroup, as well as 1049 create new sub-cgroups. 1050 1051 cgroup.kill 1052 A write-only single value file which exists in non-root cgroups. 1053 The only allowed value is "1". 1054 1055 Writing "1" to the file causes the cgroup and all descendant cgroups to 1056 be killed. This means that all processes located in the affected cgroup 1057 tree will be killed via SIGKILL. 1058 1059 Killing a cgroup tree will deal with concurrent forks appropriately and 1060 is protected against migrations. 1061 1062 In a threaded cgroup, writing this file fails with EOPNOTSUPP as 1063 killing cgroups is a process directed operation, i.e. it affects 1064 the whole thread-group. 1065 1066 cgroup.pressure 1067 A read-write single value file that allowed values are "0" and "1". 1068 The default is "1". 1069 1070 Writing "0" to the file will disable the cgroup PSI accounting. 1071 Writing "1" to the file will re-enable the cgroup PSI accounting. 1072 1073 This control attribute is not hierarchical, so disable or enable PSI 1074 accounting in a cgroup does not affect PSI accounting in descendants 1075 and doesn't need pass enablement via ancestors from root. 1076 1077 The reason this control attribute exists is that PSI accounts stalls for 1078 each cgroup separately and aggregates it at each level of the hierarchy. 1079 This may cause non-negligible overhead for some workloads when under 1080 deep level of the hierarchy, in which case this control attribute can 1081 be used to disable PSI accounting in the non-leaf cgroups. 1082 1083 irq.pressure 1084 A read-write nested-keyed file. 1085 1086 Shows pressure stall information for IRQ/SOFTIRQ. See 1087 :ref:`Documentation/accounting/psi.rst <psi>` for details. 1088 1089Controllers 1090=========== 1091 1092.. _cgroup-v2-cpu: 1093 1094CPU 1095--- 1096 1097The "cpu" controllers regulates distribution of CPU cycles. This 1098controller implements weight and absolute bandwidth limit models for 1099normal scheduling policy and absolute bandwidth allocation model for 1100realtime scheduling policy. 1101 1102In all the above models, cycles distribution is defined only on a temporal 1103base and it does not account for the frequency at which tasks are executed. 1104The (optional) utilization clamping support allows to hint the schedutil 1105cpufreq governor about the minimum desired frequency which should always be 1106provided by a CPU, as well as the maximum desired frequency, which should not 1107be exceeded by a CPU. 1108 1109WARNING: cgroup2 cpu controller doesn't yet support the (bandwidth) control of 1110realtime processes. For a kernel built with the CONFIG_RT_GROUP_SCHED option 1111enabled for group scheduling of realtime processes, the cpu controller can only 1112be enabled when all RT processes are in the root cgroup. Be aware that system 1113management software may already have placed RT processes into non-root cgroups 1114during the system boot process, and these processes may need to be moved to the 1115root cgroup before the cpu controller can be enabled with a 1116CONFIG_RT_GROUP_SCHED enabled kernel. 1117 1118With CONFIG_RT_GROUP_SCHED disabled, this limitation does not apply and some of 1119the interface files either affect realtime processes or account for them. See 1120the following section for details. Only the cpu controller is affected by 1121CONFIG_RT_GROUP_SCHED. Other controllers can be used for the resource control of 1122realtime processes irrespective of CONFIG_RT_GROUP_SCHED. 1123 1124 1125CPU Interface Files 1126~~~~~~~~~~~~~~~~~~~ 1127 1128The interaction of a process with the cpu controller depends on its scheduling 1129policy and the underlying scheduler. From the point of view of the cpu controller, 1130processes can be categorized as follows: 1131 1132* Processes under the fair-class scheduler 1133* Processes under a BPF scheduler with the ``cgroup_set_weight`` callback 1134* Everything else: ``SCHED_{FIFO,RR,DEADLINE}`` and processes under a BPF scheduler 1135 without the ``cgroup_set_weight`` callback 1136 1137For details on when a process is under the fair-class scheduler or a BPF scheduler, 1138check out :ref:`Documentation/scheduler/sched-ext.rst <sched-ext>`. 1139 1140For each of the following interface files, the above categories 1141will be referred to. All time durations are in microseconds. 1142 1143 cpu.stat 1144 A read-only flat-keyed file. 1145 This file exists whether the controller is enabled or not. 1146 1147 It always reports the following three stats, which account for all the 1148 processes in the cgroup (including those in descendant cgroups): 1149 1150 - usage_usec 1151 - user_usec 1152 - system_usec 1153 1154 and the following five when the controller is enabled, which account for 1155 only the processes under the fair-class scheduler: 1156 1157 - nr_periods 1158 - nr_throttled 1159 - throttled_usec 1160 - nr_bursts 1161 - burst_usec 1162 1163 Note that the above five CFS bandwidth stats are non-hierarchical; 1164 they only account for throttling caused by this cgroup's own bandwidth 1165 limit, not including throttling inherited from ancestor cgroups. 1166 1167 cpu.stat.local 1168 A read-only flat-keyed file. 1169 This file exists whether the controller is enabled or not. 1170 1171 It reports the following stat when the controller is enabled: 1172 1173 - throttled_usec 1174 1175 Unlike the ``throttled_usec`` reported by ``cpu.stat`` which 1176 accounts for throttling caused by this cgroup's own CFS 1177 bandwidth limit, ``cpu.stat.local`` reports the actual 1178 throttling time incurred by this cgroup's own runqueues, 1179 which may include throttling inherited from ancestor 1180 cgroup bandwidth limits. 1181 1182 When the controller is not enabled, this stat is not reported. 1183 1184 cpu.weight 1185 A read-write single value file which exists on non-root 1186 cgroups. The default is "100". 1187 1188 For non idle groups (cpu.idle = 0), the weight is in the 1189 range [1, 10000]. 1190 1191 If the cgroup has been configured to be SCHED_IDLE (cpu.idle = 1), 1192 then the weight will show as a 0. 1193 1194 This file affects only processes under the fair-class scheduler and a BPF 1195 scheduler with the ``cgroup_set_weight`` callback depending on what the 1196 callback actually does. 1197 1198 cpu.weight.nice 1199 A read-write single value file which exists on non-root 1200 cgroups. The default is "0". 1201 1202 The nice value is in the range [-20, 19]. 1203 1204 This interface file is an alternative interface for 1205 "cpu.weight" and allows reading and setting weight using the 1206 same values used by nice(2). Because the range is smaller and 1207 granularity is coarser for the nice values, the read value is 1208 the closest approximation of the current weight. 1209 1210 This file affects only processes under the fair-class scheduler and a BPF 1211 scheduler with the ``cgroup_set_weight`` callback depending on what the 1212 callback actually does. 1213 1214 cpu.max 1215 A read-write two value file which exists on non-root cgroups. 1216 The default is "max 100000". 1217 1218 The maximum bandwidth limit. It's in the following format:: 1219 1220 $MAX $PERIOD 1221 1222 which indicates that the group may consume up to $MAX in each 1223 $PERIOD duration. "max" for $MAX indicates no limit. If only 1224 one number is written, $MAX is updated. 1225 1226 This file affects only processes under the fair-class scheduler. 1227 1228 cpu.max.burst 1229 A read-write single value file which exists on non-root 1230 cgroups. The default is "0". 1231 1232 The burst in the range [0, $MAX]. 1233 1234 This file affects only processes under the fair-class scheduler. 1235 1236 cpu.pressure 1237 A read-write nested-keyed file. 1238 1239 Shows pressure stall information for CPU. See 1240 :ref:`Documentation/accounting/psi.rst <psi>` for details. 1241 1242 This file accounts for all the processes in the cgroup. 1243 1244 cpu.uclamp.min 1245 A read-write single value file which exists on non-root cgroups. 1246 The default is "0", i.e. no utilization boosting. 1247 1248 The requested minimum utilization (protection) as a percentage 1249 rational number, e.g. 12.34 for 12.34%. 1250 1251 This interface allows reading and setting minimum utilization clamp 1252 values similar to the sched_setattr(2). This minimum utilization 1253 value is used to clamp the task specific minimum utilization clamp, 1254 including those of realtime processes. 1255 1256 The requested minimum utilization (protection) is always capped by 1257 the current value for the maximum utilization (limit), i.e. 1258 `cpu.uclamp.max`. 1259 1260 This file affects all the processes in the cgroup. 1261 1262 cpu.uclamp.max 1263 A read-write single value file which exists on non-root cgroups. 1264 The default is "max". i.e. no utilization capping 1265 1266 The requested maximum utilization (limit) as a percentage rational 1267 number, e.g. 98.76 for 98.76%. 1268 1269 This interface allows reading and setting maximum utilization clamp 1270 values similar to the sched_setattr(2). This maximum utilization 1271 value is used to clamp the task specific maximum utilization clamp, 1272 including those of realtime processes. 1273 1274 This file affects all the processes in the cgroup. 1275 1276 cpu.idle 1277 A read-write single value file which exists on non-root cgroups. 1278 The default is 0. 1279 1280 This is the cgroup analog of the per-task SCHED_IDLE sched policy. 1281 Setting this value to a 1 will make the scheduling policy of the 1282 cgroup SCHED_IDLE. The threads inside the cgroup will retain their 1283 own relative priorities, but the cgroup itself will be treated as 1284 very low priority relative to its peers. 1285 1286 This file affects only processes under the fair-class scheduler. 1287 1288Memory 1289------ 1290 1291The "memory" controller regulates distribution of memory. Memory is 1292stateful and implements both limit and protection models. Due to the 1293intertwining between memory usage and reclaim pressure and the 1294stateful nature of memory, the distribution model is relatively 1295complex. 1296 1297While not completely water-tight, all major memory usages by a given 1298cgroup are tracked so that the total memory consumption can be 1299accounted and controlled to a reasonable extent. Currently, the 1300following types of memory usages are tracked. 1301 1302- Userland memory - page cache and anonymous memory. 1303 1304- Kernel data structures such as dentries and inodes. 1305 1306- TCP socket buffers. 1307 1308The above list may expand in the future for better coverage. 1309 1310 1311Memory Interface Files 1312~~~~~~~~~~~~~~~~~~~~~~ 1313 1314All memory amounts are in bytes. If a value which is not aligned to 1315PAGE_SIZE is written, the value may be rounded up to the closest 1316PAGE_SIZE multiple when read back. 1317 1318 memory.current 1319 A read-only single value file which exists on non-root 1320 cgroups. 1321 1322 The total amount of memory currently being used by the cgroup 1323 and its descendants. 1324 1325 memory.min 1326 A read-write single value file which exists on non-root 1327 cgroups. The default is "0". 1328 1329 Hard memory protection. If the memory usage of a cgroup 1330 is within its effective min boundary, the cgroup's memory 1331 won't be reclaimed under any conditions. If there is no 1332 unprotected reclaimable memory available, OOM killer 1333 is invoked. Above the effective min boundary (or 1334 effective low boundary if it is higher), pages are reclaimed 1335 proportionally to the overage, reducing reclaim pressure for 1336 smaller overages. 1337 1338 Effective min boundary is limited by memory.min values of 1339 ancestor cgroups. If there is memory.min overcommitment 1340 (child cgroup or cgroups are requiring more protected memory 1341 than parent will allow), then each child cgroup will get 1342 the part of parent's protection proportional to its 1343 actual memory usage below memory.min. 1344 1345 Putting more memory than generally available under this 1346 protection is discouraged and may lead to constant OOMs. 1347 1348 memory.low 1349 A read-write single value file which exists on non-root 1350 cgroups. The default is "0". 1351 1352 Best-effort memory protection. If the memory usage of a 1353 cgroup is within its effective low boundary, the cgroup's 1354 memory won't be reclaimed unless there is no reclaimable 1355 memory available in unprotected cgroups. 1356 Above the effective low boundary (or 1357 effective min boundary if it is higher), pages are reclaimed 1358 proportionally to the overage, reducing reclaim pressure for 1359 smaller overages. 1360 1361 Effective low boundary is limited by memory.low values of 1362 ancestor cgroups. If there is memory.low overcommitment 1363 (child cgroup or cgroups are requiring more protected memory 1364 than parent will allow), then each child cgroup will get 1365 the part of parent's protection proportional to its 1366 actual memory usage below memory.low. 1367 1368 Putting more memory than generally available under this 1369 protection is discouraged. 1370 1371 memory.high 1372 A read-write single value file which exists on non-root 1373 cgroups. The default is "max". 1374 1375 Memory usage throttle limit. If a cgroup's usage goes 1376 over the high boundary, the processes of the cgroup are 1377 throttled and put under heavy reclaim pressure. 1378 1379 Going over the high limit never invokes the OOM killer and 1380 under extreme conditions the limit may be breached. The high 1381 limit should be used in scenarios where an external process 1382 monitors the limited cgroup to alleviate heavy reclaim 1383 pressure. 1384 1385 If memory.high is opened with O_NONBLOCK then the synchronous 1386 reclaim is bypassed. This is useful for admin processes that 1387 need to dynamically adjust the job's memory limits without 1388 expending their own CPU resources on memory reclamation. The 1389 job will trigger the reclaim and/or get throttled on its 1390 next charge request. 1391 1392 Please note that with O_NONBLOCK, there is a chance that the 1393 target memory cgroup may take indefinite amount of time to 1394 reduce usage below the limit due to delayed charge request or 1395 busy-hitting its memory to slow down reclaim. 1396 1397 memory.max 1398 A read-write single value file which exists on non-root 1399 cgroups. The default is "max". 1400 1401 Memory usage hard limit. This is the main mechanism to limit 1402 memory usage of a cgroup. If a cgroup's memory usage reaches 1403 this limit and can't be reduced, the OOM killer is invoked in 1404 the cgroup. Under certain circumstances, the usage may go 1405 over the limit temporarily. 1406 1407 In default configuration regular 0-order allocations always 1408 succeed unless OOM killer chooses current task as a victim. 1409 1410 Some kinds of allocations don't invoke the OOM killer. 1411 Caller could retry them differently, return into userspace 1412 as -ENOMEM or silently ignore in cases like disk readahead. 1413 1414 If memory.max is opened with O_NONBLOCK, then the synchronous 1415 reclaim and oom-kill are bypassed. This is useful for admin 1416 processes that need to dynamically adjust the job's memory limits 1417 without expending their own CPU resources on memory reclamation. 1418 The job will trigger the reclaim and/or oom-kill on its next 1419 charge request. 1420 1421 Please note that with O_NONBLOCK, there is a chance that the 1422 target memory cgroup may take indefinite amount of time to 1423 reduce usage below the limit due to delayed charge request or 1424 busy-hitting its memory to slow down reclaim. 1425 1426 memory.reclaim 1427 A write-only nested-keyed file which exists for all cgroups. 1428 1429 This is a simple interface to trigger memory reclaim in the 1430 target cgroup. 1431 1432 Example:: 1433 1434 echo "1G" > memory.reclaim 1435 1436 Please note that the kernel can over or under reclaim from 1437 the target cgroup. If less bytes are reclaimed than the 1438 specified amount, -EAGAIN is returned. 1439 1440 Please note that the proactive reclaim (triggered by this 1441 interface) is not meant to indicate memory pressure on the 1442 memory cgroup. Therefore socket memory balancing triggered by 1443 the memory reclaim normally is not exercised in this case. 1444 This means that the networking layer will not adapt based on 1445 reclaim induced by memory.reclaim. 1446 1447The following nested keys are defined. 1448 1449 ========== ================================ 1450 swappiness Swappiness value to reclaim with 1451 ========== ================================ 1452 1453 Specifying a swappiness value instructs the kernel to perform 1454 the reclaim with that swappiness value. Note that this has the 1455 same semantics as vm.swappiness applied to memcg reclaim with 1456 all the existing limitations and potential future extensions. 1457 1458 The valid range for swappiness is [0-200, max], setting 1459 swappiness=max exclusively reclaims anonymous memory. 1460 1461 memory.peak 1462 A read-write single value file which exists on non-root cgroups. 1463 1464 The max memory usage recorded for the cgroup and its descendants since 1465 either the creation of the cgroup or the most recent reset for that FD. 1466 1467 A write of any non-empty string to this file resets it to the 1468 current memory usage for subsequent reads through the same 1469 file descriptor. 1470 1471 memory.oom.group 1472 A read-write single value file which exists on non-root 1473 cgroups. The default value is "0". 1474 1475 Determines whether the cgroup should be treated as 1476 an indivisible workload by the OOM killer. If set, 1477 all tasks belonging to the cgroup or to its descendants 1478 (if the memory cgroup is not a leaf cgroup) are killed 1479 together or not at all. This can be used to avoid 1480 partial kills to guarantee workload integrity. 1481 1482 Tasks with the OOM protection (oom_score_adj set to -1000) 1483 are treated as an exception and are never killed. 1484 1485 If the OOM killer is invoked in a cgroup, it's not going 1486 to kill any tasks outside of this cgroup, regardless 1487 memory.oom.group values of ancestor cgroups. 1488 1489 memory.events 1490 A read-only flat-keyed file which exists on non-root cgroups. 1491 The following entries are defined. Unless specified 1492 otherwise, a value change in this file generates a file 1493 modified event. 1494 1495 Note that all fields in this file are hierarchical and the 1496 file modified event can be generated due to an event down the 1497 hierarchy. For the local events at the cgroup level see 1498 memory.events.local. 1499 1500 low 1501 The number of times the cgroup is reclaimed due to 1502 high memory pressure even though its usage is under 1503 the low boundary. This usually indicates that the low 1504 boundary is over-committed. 1505 1506 high 1507 The number of times processes of the cgroup are 1508 throttled and routed to perform direct memory reclaim 1509 because the high memory boundary was exceeded. For a 1510 cgroup whose memory usage is capped by the high limit 1511 rather than global memory pressure, this event's 1512 occurrences are expected. 1513 1514 max 1515 The number of times the cgroup's memory usage was 1516 about to go over the max boundary. If direct reclaim 1517 fails to bring it down, the cgroup goes to OOM state. 1518 1519 oom 1520 The number of time the cgroup's memory usage was 1521 reached the limit and allocation was about to fail. 1522 1523 This event is not raised if the OOM killer is not 1524 considered as an option, e.g. for failed high-order 1525 allocations or if caller asked to not retry attempts. 1526 1527 oom_kill 1528 The number of processes belonging to this cgroup 1529 killed by any kind of OOM killer. 1530 1531 oom_group_kill 1532 The number of times a group OOM has occurred. 1533 1534 sock_throttled 1535 The number of times network sockets associated with 1536 this cgroup are throttled. 1537 1538 memory.events.local 1539 Similar to memory.events but the fields in the file are local 1540 to the cgroup i.e. not hierarchical. The file modified event 1541 generated on this file reflects only the local events. 1542 1543 memory.stat 1544 A read-only flat-keyed file which exists on non-root cgroups. 1545 1546 This breaks down the cgroup's memory footprint into different 1547 types of memory, type-specific details, and other information 1548 on the state and past events of the memory management system. 1549 1550 All memory amounts are in bytes. 1551 1552 The entries are ordered to be human readable, and new entries 1553 can show up in the middle. Don't rely on items remaining in a 1554 fixed position; use the keys to look up specific values! 1555 1556 If the entry has no per-node counter (or not show in the 1557 memory.numa_stat). We use 'npn' (non-per-node) as the tag 1558 to indicate that it will not show in the memory.numa_stat. 1559 1560 anon 1561 Amount of memory used in anonymous mappings such as 1562 brk(), sbrk(), and mmap(MAP_ANONYMOUS). Note that 1563 some kernel configurations might account complete larger 1564 allocations (e.g., THP) if only some, but not all the 1565 memory of such an allocation is mapped anymore. 1566 1567 file 1568 Amount of memory used to cache filesystem data, 1569 including tmpfs and shared memory. 1570 1571 kernel (npn) 1572 Amount of total kernel memory, including 1573 (kernel_stack, pagetables, percpu, vmalloc, slab) in 1574 addition to other kernel memory use cases. 1575 1576 kernel_stack 1577 Amount of memory allocated to kernel stacks. 1578 1579 pagetables 1580 Amount of memory allocated for page tables. 1581 1582 sec_pagetables 1583 Amount of memory allocated for secondary page tables, 1584 this currently includes KVM mmu allocations on x86 1585 and arm64 and IOMMU page tables. 1586 1587 percpu (npn) 1588 Amount of memory used for storing per-cpu kernel 1589 data structures. 1590 1591 sock (npn) 1592 Amount of memory used in network transmission buffers 1593 1594 vmalloc 1595 Amount of memory used for vmap backed memory. 1596 1597 shmem 1598 Amount of cached filesystem data that is swap-backed, 1599 such as tmpfs, shm segments, shared anonymous mmap()s 1600 1601 zswap 1602 Amount of memory consumed by the zswap compression backend. 1603 1604 zswapped 1605 Amount of application memory swapped out to zswap. 1606 1607 file_mapped 1608 Amount of cached filesystem data mapped with mmap(). Note 1609 that some kernel configurations might account complete 1610 larger allocations (e.g., THP) if only some, but not 1611 not all the memory of such an allocation is mapped. 1612 1613 file_dirty 1614 Amount of cached filesystem data that was modified but 1615 not yet written back to disk 1616 1617 file_writeback 1618 Amount of cached filesystem data that was modified and 1619 is currently being written back to disk 1620 1621 swapcached 1622 Amount of swap cached in memory. The swapcache is accounted 1623 against both memory and swap usage. 1624 1625 anon_thp 1626 Amount of memory used in anonymous mappings backed by 1627 transparent hugepages 1628 1629 file_thp 1630 Amount of cached filesystem data backed by transparent 1631 hugepages 1632 1633 shmem_thp 1634 Amount of shm, tmpfs, shared anonymous mmap()s backed by 1635 transparent hugepages 1636 1637 inactive_anon, active_anon, inactive_file, active_file, unevictable 1638 Amount of memory, swap-backed and filesystem-backed, 1639 on the internal memory management lists used by the 1640 page reclaim algorithm. 1641 1642 As these represent internal list state (eg. shmem pages are on anon 1643 memory management lists), inactive_foo + active_foo may not be equal to 1644 the value for the foo counter, since the foo counter is type-based, not 1645 list-based. 1646 1647 slab_reclaimable 1648 Part of "slab" that might be reclaimed, such as 1649 dentries and inodes. 1650 1651 slab_unreclaimable 1652 Part of "slab" that cannot be reclaimed on memory 1653 pressure. 1654 1655 slab (npn) 1656 Amount of memory used for storing in-kernel data 1657 structures. 1658 1659 workingset_refault_anon 1660 Number of refaults of previously evicted anonymous pages. 1661 1662 workingset_refault_file 1663 Number of refaults of previously evicted file pages. 1664 1665 workingset_activate_anon 1666 Number of refaulted anonymous pages that were immediately 1667 activated. 1668 1669 workingset_activate_file 1670 Number of refaulted file pages that were immediately activated. 1671 1672 workingset_restore_anon 1673 Number of restored anonymous pages which have been detected as 1674 an active workingset before they got reclaimed. 1675 1676 workingset_restore_file 1677 Number of restored file pages which have been detected as an 1678 active workingset before they got reclaimed. 1679 1680 workingset_nodereclaim 1681 Number of times a shadow node has been reclaimed 1682 1683 pswpin (npn) 1684 Number of pages swapped into memory 1685 1686 pswpout (npn) 1687 Number of pages swapped out of memory 1688 1689 pgscan (npn) 1690 Amount of scanned pages (in an inactive LRU list) 1691 1692 pgsteal (npn) 1693 Amount of reclaimed pages 1694 1695 pgscan_kswapd (npn) 1696 Amount of scanned pages by kswapd (in an inactive LRU list) 1697 1698 pgscan_direct (npn) 1699 Amount of scanned pages directly (in an inactive LRU list) 1700 1701 pgscan_khugepaged (npn) 1702 Amount of scanned pages by khugepaged (in an inactive LRU list) 1703 1704 pgscan_proactive (npn) 1705 Amount of scanned pages proactively (in an inactive LRU list) 1706 1707 pgsteal_kswapd (npn) 1708 Amount of reclaimed pages by kswapd 1709 1710 pgsteal_direct (npn) 1711 Amount of reclaimed pages directly 1712 1713 pgsteal_khugepaged (npn) 1714 Amount of reclaimed pages by khugepaged 1715 1716 pgsteal_proactive (npn) 1717 Amount of reclaimed pages proactively 1718 1719 pgfault (npn) 1720 Total number of page faults incurred 1721 1722 pgmajfault (npn) 1723 Number of major page faults incurred 1724 1725 pgrefill (npn) 1726 Amount of scanned pages (in an active LRU list) 1727 1728 pgactivate (npn) 1729 Amount of pages moved to the active LRU list 1730 1731 pgdeactivate (npn) 1732 Amount of pages moved to the inactive LRU list 1733 1734 pglazyfree (npn) 1735 Amount of pages postponed to be freed under memory pressure 1736 1737 pglazyfreed (npn) 1738 Amount of reclaimed lazyfree pages 1739 1740 swpin_zero 1741 Number of pages swapped into memory and filled with zero, where I/O 1742 was optimized out because the page content was detected to be zero 1743 during swapout. 1744 1745 swpout_zero 1746 Number of zero-filled pages swapped out with I/O skipped due to the 1747 content being detected as zero. 1748 1749 zswpin 1750 Number of pages moved in to memory from zswap. 1751 1752 zswpout 1753 Number of pages moved out of memory to zswap. 1754 1755 zswpwb 1756 Number of pages written from zswap to swap. 1757 1758 zswap_incomp 1759 Amount of memory used by incompressible pages currently stored in zswap 1760 without compression. These pages could not be compressed to 1761 a size smaller than PAGE_SIZE, so they are stored as-is. 1762 1763 thp_fault_alloc (npn) 1764 Number of transparent hugepages which were allocated to satisfy 1765 a page fault. This counter is not present when CONFIG_TRANSPARENT_HUGEPAGE 1766 is not set. 1767 1768 thp_collapse_alloc (npn) 1769 Number of transparent hugepages which were allocated to allow 1770 collapsing an existing range of pages. This counter is not 1771 present when CONFIG_TRANSPARENT_HUGEPAGE is not set. 1772 1773 thp_swpout (npn) 1774 Number of transparent hugepages which are swapout in one piece 1775 without splitting. 1776 1777 thp_swpout_fallback (npn) 1778 Number of transparent hugepages which were split before swapout. 1779 Usually because failed to allocate some continuous swap space 1780 for the huge page. 1781 1782 numa_pages_migrated (npn) 1783 Number of pages migrated by NUMA balancing. 1784 1785 numa_pte_updates (npn) 1786 Number of pages whose page table entries are modified by 1787 NUMA balancing to produce NUMA hinting faults on access. 1788 1789 numa_hint_faults (npn) 1790 Number of NUMA hinting faults. 1791 1792 pgdemote_kswapd 1793 Number of pages demoted by kswapd. 1794 1795 pgdemote_direct 1796 Number of pages demoted directly. 1797 1798 pgdemote_khugepaged 1799 Number of pages demoted by khugepaged. 1800 1801 pgdemote_proactive 1802 Number of pages demoted by proactively. 1803 1804 hugetlb 1805 Amount of memory used by hugetlb pages. This metric only shows 1806 up if hugetlb usage is accounted for in memory.current (i.e. 1807 cgroup is mounted with the memory_hugetlb_accounting option). 1808 1809 memory.numa_stat 1810 A read-only nested-keyed file which exists on non-root cgroups. 1811 1812 This breaks down the cgroup's memory footprint into different 1813 types of memory, type-specific details, and other information 1814 per node on the state of the memory management system. 1815 1816 This is useful for providing visibility into the NUMA locality 1817 information within an memcg since the pages are allowed to be 1818 allocated from any physical node. One of the use case is evaluating 1819 application performance by combining this information with the 1820 application's CPU allocation. 1821 1822 All memory amounts are in bytes. 1823 1824 The output format of memory.numa_stat is:: 1825 1826 type N0=<bytes in node 0> N1=<bytes in node 1> ... 1827 1828 The entries are ordered to be human readable, and new entries 1829 can show up in the middle. Don't rely on items remaining in a 1830 fixed position; use the keys to look up specific values! 1831 1832 The entries can refer to the memory.stat. 1833 1834 memory.swap.current 1835 A read-only single value file which exists on non-root 1836 cgroups. 1837 1838 The total amount of swap currently being used by the cgroup 1839 and its descendants. 1840 1841 memory.swap.high 1842 A read-write single value file which exists on non-root 1843 cgroups. The default is "max". 1844 1845 Swap usage throttle limit. If a cgroup's swap usage exceeds 1846 this limit, all its further allocations will be throttled to 1847 allow userspace to implement custom out-of-memory procedures. 1848 1849 This limit marks a point of no return for the cgroup. It is NOT 1850 designed to manage the amount of swapping a workload does 1851 during regular operation. Compare to memory.swap.max, which 1852 prohibits swapping past a set amount, but lets the cgroup 1853 continue unimpeded as long as other memory can be reclaimed. 1854 1855 Healthy workloads are not expected to reach this limit. 1856 1857 memory.swap.peak 1858 A read-write single value file which exists on non-root cgroups. 1859 1860 The max swap usage recorded for the cgroup and its descendants since 1861 the creation of the cgroup or the most recent reset for that FD. 1862 1863 A write of any non-empty string to this file resets it to the 1864 current memory usage for subsequent reads through the same 1865 file descriptor. 1866 1867 memory.swap.max 1868 A read-write single value file which exists on non-root 1869 cgroups. The default is "max". 1870 1871 Swap usage hard limit. If a cgroup's swap usage reaches this 1872 limit, anonymous memory of the cgroup will not be swapped out. 1873 1874 memory.swap.events 1875 A read-only flat-keyed file which exists on non-root cgroups. 1876 The following entries are defined. Unless specified 1877 otherwise, a value change in this file generates a file 1878 modified event. 1879 1880 high 1881 The number of times the cgroup's swap usage was over 1882 the high threshold. 1883 1884 max 1885 The number of times the cgroup's swap usage was about 1886 to go over the max boundary and swap allocation 1887 failed. 1888 1889 fail 1890 The number of times swap allocation failed either 1891 because of running out of swap system-wide or max 1892 limit. 1893 1894 When reduced under the current usage, the existing swap 1895 entries are reclaimed gradually and the swap usage may stay 1896 higher than the limit for an extended period of time. This 1897 reduces the impact on the workload and memory management. 1898 1899 memory.zswap.current 1900 A read-only single value file which exists on non-root 1901 cgroups. 1902 1903 The total amount of memory consumed by the zswap compression 1904 backend. 1905 1906 memory.zswap.max 1907 A read-write single value file which exists on non-root 1908 cgroups. The default is "max". 1909 1910 Zswap usage hard limit. If a cgroup's zswap pool reaches this 1911 limit, it will refuse to take any more stores before existing 1912 entries fault back in or are written out to disk. 1913 1914 memory.zswap.writeback 1915 A read-write single value file. The default value is "1". 1916 Note that this setting is hierarchical, i.e. the writeback would be 1917 implicitly disabled for child cgroups if the upper hierarchy 1918 does so. 1919 1920 When this is set to 0, all swapping attempts to swapping devices 1921 are disabled. This included both zswap writebacks, and swapping due 1922 to zswap store failures. If the zswap store failures are recurring 1923 (for e.g if the pages are incompressible), users can observe 1924 reclaim inefficiency after disabling writeback (because the same 1925 pages might be rejected again and again). 1926 1927 Note that this is subtly different from setting memory.swap.max to 1928 0, as it still allows for pages to be written to the zswap pool. 1929 This setting has no effect if zswap is disabled, and swapping 1930 is allowed unless memory.swap.max is set to 0. 1931 1932 memory.pressure 1933 A read-write nested-keyed file. 1934 1935 Shows pressure stall information for memory. See 1936 :ref:`Documentation/accounting/psi.rst <psi>` for details. 1937 1938 1939Usage Guidelines 1940~~~~~~~~~~~~~~~~ 1941 1942"memory.high" is the main mechanism to control memory usage. 1943Over-committing on high limit (sum of high limits > available memory) 1944and letting global memory pressure to distribute memory according to 1945usage is a viable strategy. 1946 1947Because breach of the high limit doesn't trigger the OOM killer but 1948throttles the offending cgroup, a management agent has ample 1949opportunities to monitor and take appropriate actions such as granting 1950more memory or terminating the workload. 1951 1952Determining whether a cgroup has enough memory is not trivial as 1953memory usage doesn't indicate whether the workload can benefit from 1954more memory. For example, a workload which writes data received from 1955network to a file can use all available memory but can also operate as 1956performant with a small amount of memory. A measure of memory 1957pressure - how much the workload is being impacted due to lack of 1958memory - is necessary to determine whether a workload needs more 1959memory; unfortunately, memory pressure monitoring mechanism isn't 1960implemented yet. 1961 1962Reclaim Protection 1963~~~~~~~~~~~~~~~~~~ 1964 1965The protection configured with "memory.low" or "memory.min" applies relatively 1966to the target of the reclaim (i.e. any of memory cgroup limits, proactive 1967memory.reclaim or global reclaim apparently located in the root cgroup). 1968The protection value configured for B applies unchanged to the reclaim 1969targeting A (i.e. caused by competition with the sibling E):: 1970 1971 root - ... - A - B - C 1972 \ ` D 1973 ` E 1974 1975When the reclaim targets ancestors of A, the effective protection of B is 1976capped by the protection value configured for A (and any other intermediate 1977ancestors between A and the target). 1978 1979To express indifference about relative sibling protection, it is suggested to 1980use memory_recursiveprot. Configuring all descendants of a parent with finite 1981protection to "max" works but it may unnecessarily skew memory.events:low 1982field. 1983 1984Memory Ownership 1985~~~~~~~~~~~~~~~~ 1986 1987A memory area is charged to the cgroup which instantiated it and stays 1988charged to the cgroup until the area is released. Migrating a process 1989to a different cgroup doesn't move the memory usages that it 1990instantiated while in the previous cgroup to the new cgroup. 1991 1992A memory area may be used by processes belonging to different cgroups. 1993To which cgroup the area will be charged is in-deterministic; however, 1994over time, the memory area is likely to end up in a cgroup which has 1995enough memory allowance to avoid high reclaim pressure. 1996 1997If a cgroup sweeps a considerable amount of memory which is expected 1998to be accessed repeatedly by other cgroups, it may make sense to use 1999POSIX_FADV_DONTNEED to relinquish the ownership of memory areas 2000belonging to the affected files to ensure correct memory ownership. 2001 2002 2003IO 2004-- 2005 2006The "io" controller regulates the distribution of IO resources. This 2007controller implements both weight based and absolute bandwidth or IOPS 2008limit distribution. Absolute BPS and IOPS limits are enforced by 2009blk-throttle and apply to all devices, while weight based proportional 2010distribution is provided by the iocost cost model controller 2011(CONFIG_BLK_CGROUP_IOCOST) and, when the BFQ I/O scheduler is in use 2012for a device, by BFQ's own cgroup support. Latency-based protection 2013(CONFIG_BLK_CGROUP_IOLATENCY) and I/O priority assignment 2014(CONFIG_BLK_CGROUP_IOPRIO) are also available. 2015 2016 2017IO Interface Files 2018~~~~~~~~~~~~~~~~~~ 2019 2020 io.stat 2021 A read-only nested-keyed file. 2022 2023 Lines are keyed by $MAJ:$MIN device numbers and not ordered. 2024 The following nested keys are defined. 2025 2026 ====== ===================== 2027 rbytes Bytes read 2028 wbytes Bytes written 2029 rios Number of read IOs 2030 wios Number of write IOs 2031 dbytes Bytes discarded 2032 dios Number of discard IOs 2033 ====== ===================== 2034 2035 An example read output follows:: 2036 2037 8:16 rbytes=1459200 wbytes=314773504 rios=192 wios=353 dbytes=0 dios=0 2038 8:0 rbytes=90430464 wbytes=299008000 rios=8950 wios=1252 dbytes=50331648 dios=3021 2039 2040 io.cost.qos 2041 A read-write nested-keyed file which exists only on the root 2042 cgroup. 2043 2044 This file configures the Quality of Service of the IO cost 2045 model based controller (CONFIG_BLK_CGROUP_IOCOST) which 2046 currently implements "io.weight" proportional control. Lines 2047 are keyed by $MAJ:$MIN device numbers and not ordered. The 2048 line for a given device is populated on the first write for 2049 the device on "io.cost.qos" or "io.cost.model". The following 2050 nested keys are defined. 2051 2052 ====== ===================================== 2053 enable Weight-based control enable 2054 ctrl "auto" or "user" 2055 rpct Read latency percentile [0, 100] 2056 rlat Read latency threshold 2057 wpct Write latency percentile [0, 100] 2058 wlat Write latency threshold 2059 min Minimum scaling percentage [1, 10000] 2060 max Maximum scaling percentage [1, 10000] 2061 ====== ===================================== 2062 2063 The controller is disabled by default and can be enabled by 2064 setting "enable" to 1. "rpct" and "wpct" parameters default 2065 to zero and the controller uses internal device saturation 2066 state to adjust the overall IO rate between "min" and "max". 2067 2068 When a better control quality is needed, latency QoS 2069 parameters can be configured. For example:: 2070 2071 8:16 enable=1 ctrl=auto rpct=95.00 rlat=75000 wpct=95.00 wlat=150000 min=50.00 max=150.0 2072 2073 shows that on sdb, the controller is enabled, will consider 2074 the device saturated if the 95th percentile of read completion 2075 latencies is above 75ms or write 150ms, and adjust the overall 2076 IO issue rate between 50% and 150% accordingly. 2077 2078 The lower the saturation point, the better the latency QoS at 2079 the cost of aggregate bandwidth. The narrower the allowed 2080 adjustment range between "min" and "max", the more conformant 2081 to the cost model the IO behavior. Note that the IO issue 2082 base rate may be far off from 100% and setting "min" and "max" 2083 blindly can lead to a significant loss of device capacity or 2084 control quality. "min" and "max" are useful for regulating 2085 devices which show wide temporary behavior changes - e.g. a 2086 ssd which accepts writes at the line speed for a while and 2087 then completely stalls for multiple seconds. 2088 2089 When "ctrl" is "auto", the parameters are controlled by the 2090 kernel and may change automatically. Setting "ctrl" to "user" 2091 or setting any of the percentile and latency parameters puts 2092 it into "user" mode and disables the automatic changes. The 2093 automatic mode can be restored by setting "ctrl" to "auto". 2094 2095 io.cost.model 2096 A read-write nested-keyed file which exists only on the root 2097 cgroup. 2098 2099 This file configures the cost model of the IO cost model based 2100 controller (CONFIG_BLK_CGROUP_IOCOST) which currently 2101 implements "io.weight" proportional control. Lines are keyed 2102 by $MAJ:$MIN device numbers and not ordered. The line for a 2103 given device is populated on the first write for the device on 2104 "io.cost.qos" or "io.cost.model". The following nested keys 2105 are defined. 2106 2107 ===== ================================ 2108 ctrl "auto" or "user" 2109 model The cost model in use - "linear" 2110 ===== ================================ 2111 2112 When "ctrl" is "auto", the kernel may change all parameters 2113 dynamically. When "ctrl" is set to "user" or any other 2114 parameters are written to, "ctrl" become "user" and the 2115 automatic changes are disabled. 2116 2117 When "model" is "linear", the following model parameters are 2118 defined. 2119 2120 ============= ======================================== 2121 [r|w]bps The maximum sequential IO throughput 2122 [r|w]seqiops The maximum 4k sequential IOs per second 2123 [r|w]randiops The maximum 4k random IOs per second 2124 ============= ======================================== 2125 2126 From the above, the builtin linear model determines the base 2127 costs of a sequential and random IO and the cost coefficient 2128 for the IO size. While simple, this model can cover most 2129 common device classes acceptably. 2130 2131 The IO cost model isn't expected to be accurate in absolute 2132 sense and is scaled to the device behavior dynamically. 2133 2134 If needed, tools/cgroup/iocost_coef_gen.py can be used to 2135 generate device-specific coefficients. 2136 2137 io.weight 2138 A read-write flat-keyed file which exists on non-root cgroups. 2139 The default is "default 100". 2140 2141 The first line is the default weight applied to devices 2142 without specific override. The rest are overrides keyed by 2143 $MAJ:$MIN device numbers and not ordered. The weights are in 2144 the range [1, 10000] and specifies the relative amount IO time 2145 the cgroup can use in relation to its siblings. 2146 2147 The default weight can be updated by writing either "default 2148 $WEIGHT" or simply "$WEIGHT". Overrides can be set by writing 2149 "$MAJ:$MIN $WEIGHT" and unset by writing "$MAJ:$MIN default". 2150 2151 An example read output follows:: 2152 2153 default 100 2154 8:16 200 2155 8:0 50 2156 2157 io.max 2158 A read-write nested-keyed file which exists on non-root 2159 cgroups. 2160 2161 BPS and IOPS based IO limit. Lines are keyed by $MAJ:$MIN 2162 device numbers and not ordered. The following nested keys are 2163 defined. 2164 2165 ===== ================================== 2166 rbps Max read bytes per second 2167 wbps Max write bytes per second 2168 riops Max read IO operations per second 2169 wiops Max write IO operations per second 2170 ===== ================================== 2171 2172 When writing, any number of nested key-value pairs can be 2173 specified in any order. "max" can be specified as the value 2174 to remove a specific limit. If the same key is specified 2175 multiple times, the outcome is undefined. 2176 2177 BPS and IOPS are measured in each IO direction and IOs are 2178 delayed if limit is reached. Temporary bursts are allowed. 2179 2180 Setting read limit at 2M BPS and write at 120 IOPS for 8:16:: 2181 2182 echo "8:16 rbps=2097152 wiops=120" > io.max 2183 2184 Reading returns the following:: 2185 2186 8:16 rbps=2097152 wbps=max riops=max wiops=120 2187 2188 Write IOPS limit can be removed by writing the following:: 2189 2190 echo "8:16 wiops=max" > io.max 2191 2192 Reading now returns the following:: 2193 2194 8:16 rbps=2097152 wbps=max riops=max wiops=max 2195 2196 io.pressure 2197 A read-write nested-keyed file. 2198 2199 Shows pressure stall information for IO. See 2200 :ref:`Documentation/accounting/psi.rst <psi>` for details. 2201 2202 2203Writeback 2204~~~~~~~~~ 2205 2206Page cache is dirtied through buffered writes and shared mmaps and 2207written asynchronously to the backing filesystem by the writeback 2208mechanism. Writeback sits between the memory and IO domains and 2209regulates the proportion of dirty memory by balancing dirtying and 2210write IOs. 2211 2212The io controller, in conjunction with the memory controller, 2213implements control of page cache writeback IOs. The memory controller 2214defines the memory domain that dirty memory ratio is calculated and 2215maintained for and the io controller defines the io domain which 2216writes out dirty pages for the memory domain. Both system-wide and 2217per-cgroup dirty memory states are examined and the more restrictive 2218of the two is enforced. 2219 2220cgroup writeback requires explicit support from the underlying 2221filesystem. Currently, cgroup writeback is implemented on ext2, ext4, 2222btrfs, f2fs, and xfs. On other filesystems, all writeback IOs are 2223attributed to the root cgroup. 2224 2225There are inherent differences in memory and writeback management 2226which affects how cgroup ownership is tracked. Memory is tracked per 2227page while writeback per inode. For the purpose of writeback, an 2228inode is assigned to a cgroup and all IO requests to write dirty pages 2229from the inode are attributed to that cgroup. 2230 2231As cgroup ownership for memory is tracked per page, there can be pages 2232which are associated with different cgroups than the one the inode is 2233associated with. These are called foreign pages. The writeback 2234constantly keeps track of foreign pages and, if a particular foreign 2235cgroup becomes the majority over a certain period of time, switches 2236the ownership of the inode to that cgroup. 2237 2238While this model is enough for most use cases where a given inode is 2239mostly dirtied by a single cgroup even when the main writing cgroup 2240changes over time, use cases where multiple cgroups write to a single 2241inode simultaneously are not supported well. In such circumstances, a 2242significant portion of IOs are likely to be attributed incorrectly. 2243As memory controller assigns page ownership on the first use and 2244doesn't update it until the page is released, even if writeback 2245strictly follows page ownership, multiple cgroups dirtying overlapping 2246areas wouldn't work as expected. It's recommended to avoid such usage 2247patterns. 2248 2249The sysctl knobs which affect writeback behavior are applied to cgroup 2250writeback as follows. 2251 2252 vm.dirty_background_ratio, vm.dirty_ratio 2253 These ratios apply the same to cgroup writeback with the 2254 amount of available memory capped by limits imposed by the 2255 memory controller and system-wide clean memory. 2256 2257 vm.dirty_background_bytes, vm.dirty_bytes 2258 For cgroup writeback, this is calculated into ratio against 2259 total available memory and applied the same way as 2260 vm.dirty[_background]_ratio. 2261 2262 2263IO Latency 2264~~~~~~~~~~ 2265 2266This is a cgroup v2 controller for IO workload protection. You provide a group 2267with a latency target, and if the group misses its target the controller will 2268throttle any peers that have a lower latency target than the protected 2269workload. How a miss is detected depends on the device: on rotational devices 2270the average latency over the window must exceed the target, while on 2271non-rotational devices a miss is counted once enough of the IOs in the window 2272individually exceed the target. 2273 2274The limits are only applied at the peer level in the hierarchy. This means that 2275in the diagram below, only groups A, B, and C will influence each other, and 2276groups D and F will influence each other. Group G will influence nobody:: 2277 2278 [root] 2279 / | \ 2280 A B C 2281 / \ | 2282 D F G 2283 2284 2285So the ideal way to configure this is to set io.latency in groups A, B, and C. 2286Generally you do not want to set a value lower than the latency your device 2287supports. Experiment to find the value that works best for your workload. 2288Start at higher than the expected latency for your device and, with 2289blkcg_debug_stats enabled, observe io.stat for your workload group to get an 2290idea of the latency you see during normal operation. On rotational devices, 2291use the avg_lat value as a basis for your real setting, setting it 10-15% 2292higher. On non-rotational devices io.stat reports no average latency; set 2293the target based on your device and use the missed/total fields to verify it 2294is being met. 2295 2296How IO Latency Throttling Works 2297~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~ 2298 2299io.latency is work conserving; so as long as everybody is meeting their latency 2300target the controller doesn't do anything. Once a group starts missing its 2301target it begins throttling any peer group that has a higher target than itself. 2302This throttling takes 2 forms: 2303 2304- Queue depth throttling. This is the number of outstanding IO's a group is 2305 allowed to have. We will clamp down relatively quickly, starting at no limit 2306 and going all the way down to 1 IO at a time. 2307 2308- Artificial delay induction. There are certain types of IO that cannot be 2309 throttled without possibly adversely affecting higher priority groups. This 2310 includes swapping and metadata IO. These types of IO are allowed to occur 2311 normally, however they are "charged" to the originating group. If the 2312 originating group is being throttled you will see the use_delay and delay_nsec 2313 fields in io.stat increase. The delay_nsec value is how many nanoseconds that 2314 are being added to any process that runs in this group. Because this number can 2315 grow quite large if there is a lot of swapping or metadata IO occurring we 2316 limit the individual delay events to 1 second at a time. 2317 2318Once the victimized group starts meeting its latency target again it will start 2319unthrottling any peer groups that were throttled previously. If the victimized 2320group simply stops doing IO the global counter will unthrottle appropriately. 2321 2322IO Latency Interface Files 2323~~~~~~~~~~~~~~~~~~~~~~~~~~ 2324 2325 io.latency 2326 This takes a similar format as the other controllers. 2327 2328 "MAJOR:MINOR target=<target time in microseconds>" 2329 2330 io.stat 2331 If the controller is enabled you will see extra stats in io.stat in 2332 addition to the normal ones. These debug stats are only emitted when 2333 the blkcg_debug_stats module parameter is enabled (it is disabled by 2334 default). 2335 2336 The reported latency fields depend on the device. Rotational devices 2337 report avg_lat and win; non-rotational devices report missed and total 2338 instead. missed and total are live counters for the current window and 2339 may change between reads. 2340 2341 depth 2342 This is the current queue depth for the group. 2343 2344 avg_lat 2345 (Rotational devices only.) This is an exponential moving 2346 average with a decay rate of 1/exp bound by the sampling 2347 interval. The decay rate interval can be calculated by 2348 multiplying the win value in io.stat by the corresponding number 2349 of samples based on the win value. 2350 2351 win 2352 (Rotational devices only.) The sampling window size in 2353 milliseconds. This is the minimum duration of time between 2354 evaluation events. Windows only elapse with IO activity. Idle 2355 periods extend the most recent window. 2356 2357 missed 2358 (Non-rotational devices only.) The number of IOs in the 2359 current window whose latency exceeded the target. A group is 2360 considered to be missing its target once missed reaches a 2361 certain ratio of total. 2362 2363 total 2364 (Non-rotational devices only.) The total number of IOs 2365 accounted in the current window. 2366 2367IO Priority 2368~~~~~~~~~~~ 2369 2370A single attribute controls the behavior of the I/O priority cgroup policy, 2371namely the io.prio.class attribute. The following values are accepted for 2372that attribute: 2373 2374 no-change 2375 Do not modify the I/O priority class. 2376 2377 promote-to-rt 2378 For requests that have a non-RT I/O priority class, change it into RT. 2379 Also change the priority level of these requests to 4. Do not modify 2380 the I/O priority of requests that have priority class RT. 2381 2382 restrict-to-be 2383 For requests that do not have an I/O priority class or that have I/O 2384 priority class RT, change it into BE. Also change the priority level 2385 of these requests to 0. Do not modify the I/O priority class of 2386 requests that have priority class IDLE. 2387 2388 idle 2389 Change the I/O priority class of all requests into IDLE, the lowest 2390 I/O priority class. 2391 2392 none-to-rt 2393 Deprecated. Just an alias for promote-to-rt. 2394 2395The following numerical values are associated with the I/O priority policies: 2396 2397+----------------+---+ 2398| no-change | 0 | 2399+----------------+---+ 2400| promote-to-rt | 1 | 2401+----------------+---+ 2402| restrict-to-be | 2 | 2403+----------------+---+ 2404| idle | 3 | 2405+----------------+---+ 2406 2407The numerical value that corresponds to each I/O priority class is as follows: 2408 2409+-------------------------------+---+ 2410| IOPRIO_CLASS_NONE | 0 | 2411+-------------------------------+---+ 2412| IOPRIO_CLASS_RT (real-time) | 1 | 2413+-------------------------------+---+ 2414| IOPRIO_CLASS_BE (best effort) | 2 | 2415+-------------------------------+---+ 2416| IOPRIO_CLASS_IDLE | 3 | 2417+-------------------------------+---+ 2418 2419The algorithm to set the I/O priority class for a request is as follows: 2420 2421- If I/O priority class policy is promote-to-rt, change the request I/O 2422 priority class to IOPRIO_CLASS_RT and change the request I/O priority 2423 level to 4. 2424- If I/O priority class policy is not promote-to-rt, translate the I/O priority 2425 class policy into a number, then change the request I/O priority class 2426 into the maximum of the I/O priority class policy number and the numerical 2427 I/O priority class. 2428 2429PID 2430--- 2431 2432The process number controller is used to allow a cgroup to stop any 2433new tasks from being fork()'d or clone()'d after a specified limit is 2434reached. 2435 2436The number of tasks in a cgroup can be exhausted in ways which other 2437controllers cannot prevent, thus warranting its own controller. For 2438example, a fork bomb is likely to exhaust the number of tasks before 2439hitting memory restrictions. 2440 2441Note that PIDs used in this controller refer to TIDs, process IDs as 2442used by the kernel. 2443 2444 2445PID Interface Files 2446~~~~~~~~~~~~~~~~~~~ 2447 2448 pids.max 2449 A read-write single value file which exists on non-root 2450 cgroups. The default is "max". 2451 2452 Hard limit of number of processes. 2453 2454 pids.current 2455 A read-only single value file which exists on non-root cgroups. 2456 2457 The number of processes currently in the cgroup and its 2458 descendants. 2459 2460 pids.peak 2461 A read-only single value file which exists on non-root cgroups. 2462 2463 The maximum value that the number of processes in the cgroup and its 2464 descendants has ever reached. 2465 2466 pids.events 2467 A read-only flat-keyed file which exists on non-root cgroups. Unless 2468 specified otherwise, a value change in this file generates a file 2469 modified event. The following entries are defined. 2470 2471 max 2472 The number of times the cgroup's total number of processes hit the pids.max 2473 limit (see also pids_localevents). 2474 2475 pids.events.local 2476 Similar to pids.events but the fields in the file are local 2477 to the cgroup i.e. not hierarchical. The file modified event 2478 generated on this file reflects only the local events. 2479 2480Organisational operations are not blocked by cgroup policies, so it is 2481possible to have pids.current > pids.max. This can be done by either 2482setting the limit to be smaller than pids.current, or attaching enough 2483processes to the cgroup such that pids.current is larger than 2484pids.max. However, it is not possible to violate a cgroup PID policy 2485through fork() or clone(). These will return -EAGAIN if the creation 2486of a new process would cause a cgroup policy to be violated. 2487 2488 2489Cpuset 2490------ 2491 2492The "cpuset" controller provides a mechanism for constraining 2493the CPU and memory node placement of tasks to only the resources 2494specified in the cpuset interface files in a task's current cgroup. 2495This is especially valuable on large NUMA systems where placing jobs 2496on properly sized subsets of the systems with careful processor and 2497memory placement to reduce cross-node memory access and contention 2498can improve overall system performance. 2499 2500The "cpuset" controller is hierarchical. That means the controller 2501cannot use CPUs or memory nodes not allowed in its parent. 2502 2503 2504Cpuset Interface Files 2505~~~~~~~~~~~~~~~~~~~~~~ 2506 2507 cpuset.cpus 2508 A read-write multiple values file which exists on non-root 2509 cpuset-enabled cgroups. 2510 2511 It lists the requested CPUs to be used by tasks within this 2512 cgroup. The actual list of CPUs to be granted, however, is 2513 subjected to constraints imposed by its parent and can differ 2514 from the requested CPUs. 2515 2516 The CPU numbers are comma-separated numbers or ranges. 2517 For example:: 2518 2519 # cat cpuset.cpus 2520 0-4,6,8-10 2521 2522 An empty value indicates that the cgroup is using the same 2523 setting as the nearest cgroup ancestor with a non-empty 2524 "cpuset.cpus" or all the available CPUs if none is found. 2525 2526 The value of "cpuset.cpus" stays constant until the next update 2527 and won't be affected by any CPU hotplug events. 2528 2529 cpuset.cpus.effective 2530 A read-only multiple values file which exists on all 2531 cpuset-enabled cgroups. 2532 2533 It lists the onlined CPUs that are actually granted to this 2534 cgroup by its parent. These CPUs are allowed to be used by 2535 tasks within the current cgroup. 2536 2537 If "cpuset.cpus" is empty, the "cpuset.cpus.effective" file shows 2538 all the CPUs from the parent cgroup that can be available to 2539 be used by this cgroup. Otherwise, it should be a subset of 2540 "cpuset.cpus" unless none of the CPUs listed in "cpuset.cpus" 2541 can be granted. In this case, it will be treated just like an 2542 empty "cpuset.cpus". 2543 2544 Its value will be affected by CPU hotplug events. 2545 2546 cpuset.mems 2547 A read-write multiple values file which exists on non-root 2548 cpuset-enabled cgroups. 2549 2550 It lists the requested memory nodes to be used by tasks within 2551 this cgroup. The actual list of memory nodes granted, however, 2552 is subjected to constraints imposed by its parent and can differ 2553 from the requested memory nodes. 2554 2555 The memory node numbers are comma-separated numbers or ranges. 2556 For example:: 2557 2558 # cat cpuset.mems 2559 0-1,3 2560 2561 An empty value indicates that the cgroup is using the same 2562 setting as the nearest cgroup ancestor with a non-empty 2563 "cpuset.mems" or all the available memory nodes if none 2564 is found. 2565 2566 The value of "cpuset.mems" stays constant until the next update 2567 and won't be affected by any memory nodes hotplug events. 2568 2569 Setting a non-empty value to "cpuset.mems" causes memory of 2570 tasks within the cgroup to be migrated to the designated nodes if 2571 they are currently using memory outside of the designated nodes. 2572 2573 There is a cost for this memory migration. The migration 2574 may not be complete and some memory pages may be left behind. 2575 So it is recommended that "cpuset.mems" should be set properly 2576 before spawning new tasks into the cpuset. Even if there is 2577 a need to change "cpuset.mems" with active tasks, it shouldn't 2578 be done frequently. 2579 2580 For a multithreaded process, the threadgroup leader is 2581 considered the owner of the group's memory. Memory policy 2582 rebinding and migration will only happen with respect to the 2583 threadgroup leader. To avoid unexpected results, non-leading 2584 threads shouldn't be put into another cgroup whose "cpuset.mems" 2585 doesn't fully overlap that of the threadgroup leader. 2586 2587 cpuset.mems.effective 2588 A read-only multiple values file which exists on all 2589 cpuset-enabled cgroups. 2590 2591 It lists the onlined memory nodes that are actually granted to 2592 this cgroup by its parent. These memory nodes are allowed to 2593 be used by tasks within the current cgroup. 2594 2595 If "cpuset.mems" is empty, it shows all the memory nodes from the 2596 parent cgroup that will be available to be used by this cgroup. 2597 Otherwise, it should be a subset of "cpuset.mems" unless none of 2598 the memory nodes listed in "cpuset.mems" can be granted. In this 2599 case, it will be treated just like an empty "cpuset.mems". 2600 2601 Its value will be affected by memory nodes hotplug events. 2602 2603 cpuset.cpus.exclusive 2604 A read-write multiple values file which exists on non-root 2605 cpuset-enabled cgroups. 2606 2607 It lists all the exclusive CPUs that are allowed to be used 2608 to create a new cpuset partition. Its value is not used 2609 unless the cgroup becomes a valid partition root. See the 2610 "cpuset.cpus.partition" section below for a description of what 2611 a cpuset partition is. 2612 2613 When the cgroup becomes a partition root, the actual exclusive 2614 CPUs that are allocated to that partition are listed in 2615 "cpuset.cpus.exclusive.effective" which may be different 2616 from "cpuset.cpus.exclusive". If "cpuset.cpus.exclusive" 2617 has previously been set, "cpuset.cpus.exclusive.effective" 2618 is always a subset of it. 2619 2620 Users can manually set it to a value that is different from 2621 "cpuset.cpus". One constraint in setting it is that the list of 2622 CPUs must be exclusive with respect to "cpuset.cpus.exclusive" 2623 and "cpuset.cpus.exclusive.effective" of its siblings. Another 2624 constraint is that it cannot be a superset of "cpuset.cpus" 2625 of its sibling in order to leave at least one CPU available to 2626 that sibling when the exclusive CPUs are taken away. 2627 2628 For a parent cgroup, any one of its exclusive CPUs can only 2629 be distributed to at most one of its child cgroups. Having an 2630 exclusive CPU appearing in two or more of its child cgroups is 2631 not allowed (the exclusivity rule). A value that violates the 2632 exclusivity rule will be rejected with a write error. 2633 2634 The root cgroup is a partition root and all its available CPUs 2635 are in its exclusive CPU set. 2636 2637 cpuset.cpus.exclusive.effective 2638 A read-only multiple values file which exists on all non-root 2639 cpuset-enabled cgroups. 2640 2641 This file shows the effective set of exclusive CPUs that 2642 can be used to create a partition root. The content 2643 of this file will always be a subset of its parent's 2644 "cpuset.cpus.exclusive.effective" if its parent is not the root 2645 cgroup. It will also be a subset of "cpuset.cpus.exclusive" 2646 if it is set. This file should only be non-empty if either 2647 "cpuset.cpus.exclusive" is set or when the current cpuset is 2648 a valid partition root. 2649 2650 cpuset.cpus.isolated 2651 A read-only and root cgroup only multiple values file. 2652 2653 This file shows the set of all isolated CPUs used in existing 2654 isolated partitions. It will be empty if no isolated partition 2655 is created. 2656 2657 cpuset.cpus.partition 2658 A read-write single value file which exists on non-root 2659 cpuset-enabled cgroups. This flag is owned by the parent cgroup 2660 and is not delegatable. 2661 2662 It accepts only the following input values when written to. 2663 2664 ========== ===================================== 2665 "member" Non-root member of a partition 2666 "root" Partition root 2667 "isolated" Partition root without load balancing 2668 ========== ===================================== 2669 2670 A cpuset partition is a collection of cpuset-enabled cgroups with 2671 a partition root at the top of the hierarchy and its descendants 2672 except those that are separate partition roots themselves and 2673 their descendants. A partition has exclusive access to the 2674 set of exclusive CPUs allocated to it. Other cgroups outside 2675 of that partition cannot use any CPUs in that set. 2676 2677 There are two types of partitions - local and remote. A local 2678 partition is one whose parent cgroup is also a valid partition 2679 root. A remote partition is one whose parent cgroup is not a 2680 valid partition root itself. 2681 2682 Writing to "cpuset.cpus.exclusive" is optional for the creation 2683 of a local partition as its "cpuset.cpus.exclusive" file will 2684 assume an implicit value that is the same as "cpuset.cpus" if it 2685 is not set. Writing the proper "cpuset.cpus.exclusive" values 2686 down the cgroup hierarchy before the target partition root is 2687 mandatory for the creation of a remote partition. 2688 2689 Not all the CPUs requested in "cpuset.cpus.exclusive" can be 2690 used to form a new partition. Only those that were present 2691 in its parent's "cpuset.cpus.exclusive.effective" control 2692 file can be used. For partitions created without setting 2693 "cpuset.cpus.exclusive", exclusive CPUs specified in sibling's 2694 "cpuset.cpus.exclusive" or "cpuset.cpus.exclusive.effective" 2695 also cannot be used. 2696 2697 Currently, a remote partition cannot be created under a local 2698 partition. All the ancestors of a remote partition root except 2699 the root cgroup cannot be a partition root. 2700 2701 The root cgroup is always a partition root and its state cannot 2702 be changed. All other non-root cgroups start out as "member". 2703 Even though the "cpuset.cpus.exclusive*" and "cpuset.cpus" 2704 control files are not present in the root cgroup, they are 2705 implicitly the same as the "/sys/devices/system/cpu/possible" 2706 sysfs file. 2707 2708 When set to "root", the current cgroup is the root of a new 2709 partition or scheduling domain. The set of exclusive CPUs is 2710 determined by the value of its "cpuset.cpus.exclusive.effective". 2711 2712 When set to "isolated", the CPUs in that partition will be in 2713 an isolated state without any load balancing from the scheduler 2714 and excluded from the unbound workqueues. Tasks placed in such 2715 a partition with multiple CPUs should be carefully distributed 2716 and bound to each of the individual CPUs for optimal performance. 2717 2718 A partition root ("root" or "isolated") can be in one of the 2719 two possible states - valid or invalid. An invalid partition 2720 root is in a degraded state where some state information may 2721 be retained, but behaves more like a "member". 2722 2723 All possible state transitions among "member", "root" and 2724 "isolated" are allowed. 2725 2726 On read, the "cpuset.cpus.partition" file can show the following 2727 values. 2728 2729 ============================= ===================================== 2730 "member" Non-root member of a partition 2731 "root" Partition root 2732 "isolated" Partition root without load balancing 2733 "root invalid (<reason>)" Invalid partition root 2734 "isolated invalid (<reason>)" Invalid isolated partition root 2735 ============================= ===================================== 2736 2737 In the case of an invalid partition root, a descriptive string on 2738 why the partition is invalid is included within parentheses. 2739 2740 For a local partition root to be valid, the following conditions 2741 must be met. 2742 2743 1) The parent cgroup is a valid partition root. 2744 2) The "cpuset.cpus.exclusive.effective" file cannot be empty, 2745 though it may contain offline CPUs. 2746 3) The "cpuset.cpus.effective" cannot be empty unless there is 2747 no task associated with this partition. 2748 2749 For a remote partition root to be valid, all the above conditions 2750 except the first one must be met. 2751 2752 External events like hotplug or changes to "cpuset.cpus" or 2753 "cpuset.cpus.exclusive" can cause a valid partition root to 2754 become invalid and vice versa. Note that a task cannot be 2755 moved to a cgroup with empty "cpuset.cpus.effective". 2756 2757 A valid non-root parent partition may distribute out all its CPUs 2758 to its child local partitions when there is no task associated 2759 with it. 2760 2761 Care must be taken to change a valid partition root to "member" 2762 as all its child local partitions, if present, will become 2763 invalid causing disruption to tasks running in those child 2764 partitions. These inactivated partitions could be recovered if 2765 their parent is switched back to a partition root with a proper 2766 value in "cpuset.cpus" or "cpuset.cpus.exclusive". 2767 2768 Poll and inotify events are triggered whenever the state of 2769 "cpuset.cpus.partition" changes. That includes changes caused 2770 by write to "cpuset.cpus.partition", cpu hotplug or other 2771 changes that modify the validity status of the partition. 2772 This will allow user space agents to monitor unexpected changes 2773 to "cpuset.cpus.partition" without the need to do continuous 2774 polling. 2775 2776 A user can pre-configure certain CPUs to an isolated state 2777 with load balancing disabled at boot time with the "isolcpus" 2778 kernel boot command line option. If those CPUs are to be put 2779 into a partition, they have to be used in an isolated partition. 2780 2781 2782Device controller 2783----------------- 2784 2785Device controller manages access to device files. It includes both 2786creation of new device files (using mknod), and access to the 2787existing device files. 2788 2789Cgroup v2 device controller has no interface files and is implemented 2790on top of cgroup BPF. To control access to device files, a user may 2791create bpf programs of type BPF_PROG_TYPE_CGROUP_DEVICE and attach 2792them to cgroups with BPF_CGROUP_DEVICE flag. On an attempt to access a 2793device file, corresponding BPF programs will be executed, and depending 2794on the return value the attempt will succeed or fail with -EPERM. 2795 2796A BPF_PROG_TYPE_CGROUP_DEVICE program takes a pointer to the 2797bpf_cgroup_dev_ctx structure, which describes the device access attempt: 2798access type (mknod/read/write) and device (type, major and minor numbers). 2799If the program returns 0, the attempt fails with -EPERM, otherwise it 2800succeeds. 2801 2802An example of BPF_PROG_TYPE_CGROUP_DEVICE program may be found in 2803tools/testing/selftests/bpf/progs/dev_cgroup.c in the kernel source tree. 2804 2805 2806RDMA 2807---- 2808 2809The "rdma" controller regulates the distribution and accounting of 2810RDMA resources. 2811 2812RDMA devices from all network namespaces are listed. Each line starts with 2813the device name. If more than one device has the same name, ``index=N`` 2814follows the name, where ``N`` is the system-wide RDMA device index, unique 2815among registered devices. 2816 2817RDMA Interface Files 2818~~~~~~~~~~~~~~~~~~~~ 2819 2820 rdma.max 2821 A readwrite nested-keyed file that exists for all the cgroups 2822 except root that describes current configured resource limit 2823 for a RDMA/IB device. 2824 2825 Lines are keyed by device name and are not ordered. A write may 2826 include ``index=N`` after the device name. The index is optional 2827 when the name is globally unique. If multiple devices have that 2828 name, the index is required and a write without it fails with 2829 ``-ENOTUNIQ``. 2830 Each line contains space separated resource name and its configured 2831 limit that can be distributed. 2832 2833 The following nested keys are defined. 2834 2835 ========== ============================= 2836 hca_handle Maximum number of HCA Handles 2837 hca_object Maximum number of HCA Objects 2838 ========== ============================= 2839 2840 An example for mlx4 and ocrdma device follows:: 2841 2842 mlx4_0 hca_handle=2 hca_object=2000 2843 ocrdma1 hca_handle=3 hca_object=max 2844 2845 For devices with duplicate names, select the device by index:: 2846 2847 echo "rxe0 index=5 hca_handle=2" > rdma.max 2848 2849 rdma.current 2850 A read-only file that describes current resource usage. 2851 It exists for all the cgroup except root. 2852 2853 An example for mlx4 and ocrdma device follows:: 2854 2855 mlx4_0 hca_handle=1 hca_object=20 2856 ocrdma1 hca_handle=1 hca_object=23 2857 2858 rdma.peak 2859 A read-only nested-keyed file that exists for all the cgroups 2860 except root. It shows the historical high watermark of 2861 resource usage per device since the cgroup was created. 2862 2863 An example for mlx4 and ocrdma device follows:: 2864 2865 mlx4_0 hca_handle=1 hca_object=20 2866 ocrdma1 hca_handle=0 hca_object=23 2867 2868 rdma.events 2869 A read-only nested-keyed file which exists on non-root 2870 cgroups. The following nested keys are defined. 2871 2872 max 2873 The number of times a process in this cgroup or its 2874 descendants attempted an RDMA resource allocation that 2875 was rejected because a rdma.max limit in the subtree 2876 was reached. This is a hierarchical counter: the event 2877 is propagated upward to all ancestor cgroups. A value 2878 change in this file generates a file modified event. 2879 2880 alloc_fail 2881 The number of RDMA resource allocation attempts that 2882 originated in this cgroup or its descendants and failed 2883 due to a rdma.max limit being reached. This is a 2884 hierarchical counter propagated upward. 2885 2886 An example for mlx4 device follows:: 2887 2888 mlx4_0 hca_handle.max=5 hca_handle.alloc_fail=3 hca_object.max=0 hca_object.alloc_fail=0 2889 2890 rdma.events.local 2891 Similar to rdma.events but the fields in the file are local 2892 to the cgroup i.e. not hierarchical. The file modified event 2893 generated on this file reflects only the local events. 2894 2895 The following nested keys are defined. 2896 2897 max 2898 The number of times a process in this cgroup or its 2899 descendants attempted an RDMA resource allocation that 2900 was rejected because this cgroup's own rdma.max limit 2901 was reached. 2902 alloc_fail 2903 The number of RDMA resource allocation attempts 2904 originating from this cgroup that failed due to this 2905 cgroup's or an ancestor's rdma.max limit. 2906 2907 An example for mlx4 device follows:: 2908 2909 mlx4_0 hca_handle.max=5 hca_handle.alloc_fail=0 hca_object.max=0 hca_object.alloc_fail=0 2910 2911DMEM 2912---- 2913 2914The "dmem" controller regulates the distribution and accounting of 2915device memory regions. Because each memory region may have its own page size, 2916which does not have to be equal to the system page size, the units are always bytes. 2917 2918DMEM Interface Files 2919~~~~~~~~~~~~~~~~~~~~ 2920 2921 dmem.max, dmem.min, dmem.low 2922 A readwrite nested-keyed file that exists for all the cgroups 2923 except root that describes current configured resource limit 2924 for a region. 2925 2926 An example for xe follows:: 2927 2928 drm/0000:03:00.0/vram0 1073741824 2929 drm/0000:03:00.0/stolen max 2930 2931 The semantics are the same as for the memory cgroup controller, and are 2932 calculated in the same way. 2933 2934 dmem.peak 2935 A read-only nested-keyed file that exists on non-root cgroups. 2936 2937 The max device memory usage recorded for the cgroup and its 2938 descendants since the creation of the cgroup for each region. 2939 2940 dmem.capacity 2941 A read-only file that describes maximum region capacity. 2942 It only exists on the root cgroup. Not all memory can be 2943 allocated by cgroups, as the kernel reserves some for 2944 internal use. 2945 2946 An example for xe follows:: 2947 2948 drm/0000:03:00.0/vram0 8514437120 2949 drm/0000:03:00.0/stolen 67108864 2950 2951 dmem.current 2952 A read-only file that describes current resource usage. 2953 It exists for all the cgroup except root. 2954 2955 An example for xe follows:: 2956 2957 drm/0000:03:00.0/vram0 12550144 2958 drm/0000:03:00.0/stolen 8650752 2959 2960HugeTLB 2961------- 2962 2963The HugeTLB controller allows limiting the HugeTLB usage per control group and 2964enforces the controller limit during page fault. 2965 2966HugeTLB Interface Files 2967~~~~~~~~~~~~~~~~~~~~~~~ 2968 2969 hugetlb.<hugepagesize>.current 2970 Show current usage for "hugepagesize" hugetlb. It exists for all 2971 the cgroup except root. 2972 2973 hugetlb.<hugepagesize>.max 2974 Set/show the hard limit of "hugepagesize" hugetlb usage. 2975 The default value is "max". It exists for all the cgroup except root. 2976 2977 hugetlb.<hugepagesize>.events 2978 A read-only flat-keyed file which exists on non-root cgroups. 2979 2980 max 2981 The number of allocation failure due to HugeTLB limit 2982 2983 hugetlb.<hugepagesize>.events.local 2984 Similar to hugetlb.<hugepagesize>.events but the fields in the file 2985 are local to the cgroup i.e. not hierarchical. The file modified event 2986 generated on this file reflects only the local events. 2987 2988 hugetlb.<hugepagesize>.numa_stat 2989 Similar to memory.numa_stat, it shows the numa information of the 2990 hugetlb pages of <hugepagesize> in this cgroup. Only active in 2991 use hugetlb pages are included. The per-node values are in bytes. 2992 2993Misc 2994---- 2995 2996The Miscellaneous cgroup provides the resource limiting and tracking 2997mechanism for the scalar resources which cannot be abstracted like the other 2998cgroup resources. Controller is enabled by the CONFIG_CGROUP_MISC config 2999option. 3000 3001A resource can be added to the controller via enum misc_res_type{} in the 3002include/linux/misc_cgroup.h file and the corresponding name via misc_res_name[] 3003in the kernel/cgroup/misc.c file. Provider of the resource must set its 3004capacity prior to using the resource by calling misc_cg_set_capacity(). 3005 3006Once a capacity is set then the resource usage can be updated using charge and 3007uncharge APIs. All of the APIs to interact with misc controller are in 3008include/linux/misc_cgroup.h. 3009 3010Misc Interface Files 3011~~~~~~~~~~~~~~~~~~~~ 3012 3013Miscellaneous controller provides the following interface files. If two misc 3014resources (res_a and res_b) are registered then: 3015 3016 misc.capacity 3017 A read-only flat-keyed file shown only in the root cgroup. It shows 3018 miscellaneous scalar resources available on the platform along with 3019 their quantities:: 3020 3021 $ cat misc.capacity 3022 res_a 50 3023 res_b 10 3024 3025 misc.current 3026 A read-only flat-keyed file shown in the all cgroups. It shows 3027 the current usage of the resources in the cgroup and its children.:: 3028 3029 $ cat misc.current 3030 res_a 3 3031 res_b 0 3032 3033 misc.peak 3034 A read-only flat-keyed file shown in all cgroups. It shows the 3035 historical maximum usage of the resources in the cgroup and its 3036 children.:: 3037 3038 $ cat misc.peak 3039 res_a 10 3040 res_b 8 3041 3042 misc.max 3043 A read-write flat-keyed file shown in the non root cgroups. Allowed 3044 maximum usage of the resources in the cgroup and its children.:: 3045 3046 $ cat misc.max 3047 res_a max 3048 res_b 4 3049 3050 Limit can be set by:: 3051 3052 # echo res_a 1 > misc.max 3053 3054 Limit can be set to max by:: 3055 3056 # echo res_a max > misc.max 3057 3058 Limits can be set higher than the capacity value in the misc.capacity 3059 file. 3060 3061 misc.events 3062 A read-only flat-keyed file which exists on non-root cgroups. The 3063 following entries are defined. Unless specified otherwise, a value 3064 change in this file generates a file modified event. All fields in 3065 this file are hierarchical. 3066 3067 max 3068 The number of times the cgroup's resource usage was 3069 about to go over the max boundary. 3070 3071 misc.events.local 3072 Similar to misc.events but the fields in the file are local to the 3073 cgroup i.e. not hierarchical. The file modified event generated on 3074 this file reflects only the local events. 3075 3076Migration and Ownership 3077~~~~~~~~~~~~~~~~~~~~~~~ 3078 3079A miscellaneous scalar resource is charged to the cgroup in which it is used 3080first, and stays charged to that cgroup until that resource is freed. Migrating 3081a process to a different cgroup does not move the charge to the destination 3082cgroup where the process has moved. 3083 3084Others 3085------ 3086 3087perf_event 3088~~~~~~~~~~ 3089 3090perf_event controller, if not mounted on a legacy hierarchy, is 3091automatically enabled on the v2 hierarchy so that perf events can 3092always be filtered by cgroup v2 path. The controller can still be 3093moved to a legacy hierarchy after v2 hierarchy is populated. 3094 3095 3096Non-normative information 3097------------------------- 3098 3099This section contains information that isn't considered to be a part of 3100the stable kernel API and so is subject to change. 3101 3102 3103CPU controller root cgroup process behaviour 3104~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~ 3105 3106When distributing CPU cycles in the root cgroup each thread in this 3107cgroup is treated as if it was hosted in a separate child cgroup of the 3108root cgroup. This child cgroup weight is dependent on its thread nice 3109level. 3110 3111For details of this mapping see sched_prio_to_weight array in 3112kernel/sched/core.c file (values from this array should be scaled 3113appropriately so the neutral - nice 0 - value is 100 instead of 1024). 3114 3115 3116IO controller root cgroup process behaviour 3117~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~ 3118 3119Root cgroup processes are hosted in an implicit leaf child node. 3120When distributing IO resources this implicit child node is taken into 3121account as if it was a normal child cgroup of the root cgroup with a 3122weight value of 200. 3123 3124 3125Namespace 3126========= 3127 3128Basics 3129------ 3130 3131cgroup namespace provides a mechanism to virtualize the view of the 3132"/proc/$PID/cgroup" file and cgroup mounts. The CLONE_NEWCGROUP clone 3133flag can be used with clone(2) and unshare(2) to create a new cgroup 3134namespace. The process running inside the cgroup namespace will have 3135its "/proc/$PID/cgroup" output restricted to cgroupns root. The 3136cgroupns root is the cgroup of the process at the time of creation of 3137the cgroup namespace. 3138 3139Without cgroup namespace, the "/proc/$PID/cgroup" file shows the 3140complete path of the cgroup of a process. In a container setup where 3141a set of cgroups and namespaces are intended to isolate processes the 3142"/proc/$PID/cgroup" file may leak potential system level information 3143to the isolated processes. For example:: 3144 3145 # cat /proc/self/cgroup 3146 0::/batchjobs/container_id1 3147 3148The path '/batchjobs/container_id1' can be considered as system-data 3149and undesirable to expose to the isolated processes. cgroup namespace 3150can be used to restrict visibility of this path. For example, before 3151creating a cgroup namespace, one would see:: 3152 3153 # ls -l /proc/self/ns/cgroup 3154 lrwxrwxrwx 1 root root 0 2014-07-15 10:37 /proc/self/ns/cgroup -> cgroup:[4026531835] 3155 # cat /proc/self/cgroup 3156 0::/batchjobs/container_id1 3157 3158After unsharing a new namespace, the view changes:: 3159 3160 # ls -l /proc/self/ns/cgroup 3161 lrwxrwxrwx 1 root root 0 2014-07-15 10:35 /proc/self/ns/cgroup -> cgroup:[4026532183] 3162 # cat /proc/self/cgroup 3163 0::/ 3164 3165When some thread from a multi-threaded process unshares its cgroup 3166namespace, the new cgroupns gets applied to the entire process (all 3167the threads). This is natural for the v2 hierarchy; however, for the 3168legacy hierarchies, this may be unexpected. 3169 3170A cgroup namespace is alive as long as there are processes inside or 3171mounts pinning it. When the last usage goes away, the cgroup 3172namespace is destroyed. The cgroupns root and the actual cgroups 3173remain. 3174 3175 3176The Root and Views 3177------------------ 3178 3179The 'cgroupns root' for a cgroup namespace is the cgroup in which the 3180process calling unshare(2) is running. For example, if a process in 3181/batchjobs/container_id1 cgroup calls unshare, cgroup 3182/batchjobs/container_id1 becomes the cgroupns root. For the 3183init_cgroup_ns, this is the real root ('/') cgroup. 3184 3185The cgroupns root cgroup does not change even if the namespace creator 3186process later moves to a different cgroup:: 3187 3188 # ~/unshare -c # unshare cgroupns in some cgroup 3189 # cat /proc/self/cgroup 3190 0::/ 3191 # mkdir sub_cgrp_1 3192 # echo 0 > sub_cgrp_1/cgroup.procs 3193 # cat /proc/self/cgroup 3194 0::/sub_cgrp_1 3195 3196Each process gets its namespace-specific view of "/proc/$PID/cgroup" 3197 3198Processes running inside the cgroup namespace will be able to see 3199cgroup paths (in /proc/self/cgroup) only inside their root cgroup. 3200From within an unshared cgroupns:: 3201 3202 # sleep 100000 & 3203 [1] 7353 3204 # echo 7353 > sub_cgrp_1/cgroup.procs 3205 # cat /proc/7353/cgroup 3206 0::/sub_cgrp_1 3207 3208From the initial cgroup namespace, the real cgroup path will be 3209visible:: 3210 3211 $ cat /proc/7353/cgroup 3212 0::/batchjobs/container_id1/sub_cgrp_1 3213 3214From a sibling cgroup namespace (that is, a namespace rooted at a 3215different cgroup), the cgroup path relative to its own cgroup 3216namespace root will be shown. For instance, if PID 7353's cgroup 3217namespace root is at '/batchjobs/container_id2', then it will see:: 3218 3219 # cat /proc/7353/cgroup 3220 0::/../container_id2/sub_cgrp_1 3221 3222Note that the relative path always starts with '/' to indicate that 3223its relative to the cgroup namespace root of the caller. 3224 3225 3226Migration and setns(2) 3227---------------------- 3228 3229Processes inside a cgroup namespace can move into and out of the 3230namespace root if they have proper access to external cgroups. For 3231example, from inside a namespace with cgroupns root at 3232/batchjobs/container_id1, and assuming that the global hierarchy is 3233still accessible inside cgroupns:: 3234 3235 # cat /proc/7353/cgroup 3236 0::/sub_cgrp_1 3237 # echo 7353 > batchjobs/container_id2/cgroup.procs 3238 # cat /proc/7353/cgroup 3239 0::/../container_id2 3240 3241Note that this kind of setup is not encouraged. A task inside cgroup 3242namespace should only be exposed to its own cgroupns hierarchy. 3243 3244setns(2) to another cgroup namespace is allowed when: 3245 3246(a) the process has CAP_SYS_ADMIN against its current user namespace 3247(b) the process has CAP_SYS_ADMIN against the target cgroup 3248 namespace's userns 3249 3250No implicit cgroup changes happen with attaching to another cgroup 3251namespace. It is expected that the someone moves the attaching 3252process under the target cgroup namespace root. 3253 3254 3255Interaction with Other Namespaces 3256--------------------------------- 3257 3258Namespace specific cgroup hierarchy can be mounted by a process 3259running inside a non-init cgroup namespace:: 3260 3261 # mount -t cgroup2 none $MOUNT_POINT 3262 3263This will mount the unified cgroup hierarchy with cgroupns root as the 3264filesystem root. The process needs CAP_SYS_ADMIN against its user and 3265mount namespaces. 3266 3267The virtualization of /proc/self/cgroup file combined with restricting 3268the view of cgroup hierarchy by namespace-private cgroupfs mount 3269provides a properly isolated cgroup view inside the container. 3270 3271 3272Information on Kernel Programming 3273================================= 3274 3275This section contains kernel programming information in the areas 3276where interacting with cgroup is necessary. cgroup core and 3277controllers are not covered. 3278 3279 3280Filesystem Support for Writeback 3281-------------------------------- 3282 3283A filesystem can support cgroup writeback by updating 3284address_space_operations->writepages() to annotate bio's using the 3285following two functions. 3286 3287 wbc_init_bio(@wbc, @bio) 3288 Should be called for each bio carrying writeback data and 3289 associates the bio with the inode's owner cgroup and the 3290 corresponding request queue. This must be called after 3291 a queue (device) has been associated with the bio and 3292 before submission. 3293 3294 wbc_account_cgroup_owner(@wbc, @folio, @bytes) 3295 Should be called for each data segment being written out. 3296 While this function doesn't care exactly when it's called 3297 during the writeback session, it's the easiest and most 3298 natural to call it as data segments are added to a bio. 3299 3300With writeback bio's annotated, cgroup support can be enabled per 3301super_block by setting SB_I_CGROUPWB in ->s_iflags. This allows for 3302selective disabling of cgroup writeback support which is helpful when 3303certain filesystem features, e.g. journaled data mode, are 3304incompatible. 3305 3306wbc_init_bio() binds the specified bio to its cgroup. Depending on 3307the configuration, the bio may be executed at a lower priority and if 3308the writeback session is holding shared resources, e.g. a journal 3309entry, may lead to priority inversion. There is no one easy solution 3310for the problem. Filesystems can try to work around specific problem 3311cases by skipping wbc_init_bio() and using bio_associate_blkg() 3312directly. 3313 3314 3315Deprecated v1 Core Features 3316=========================== 3317 3318- Multiple hierarchies including named ones are not supported. 3319 3320- All v1 mount options are not supported. 3321 3322- The "tasks" file is removed and "cgroup.procs" is not sorted. 3323 3324- "cgroup.clone_children" is removed. 3325 3326- /proc/cgroups is meaningless for v2. Use "cgroup.controllers" or 3327 "cgroup.stat" files at the root instead. 3328 3329 3330Issues with v1 and Rationales for v2 3331==================================== 3332 3333Multiple Hierarchies 3334-------------------- 3335 3336cgroup v1 allowed an arbitrary number of hierarchies and each 3337hierarchy could host any number of controllers. While this seemed to 3338provide a high level of flexibility, it wasn't useful in practice. 3339 3340For example, as there is only one instance of each controller, utility 3341type controllers such as freezer which can be useful in all 3342hierarchies could only be used in one. The issue is exacerbated by 3343the fact that controllers couldn't be moved to another hierarchy once 3344hierarchies were populated. Another issue was that all controllers 3345bound to a hierarchy were forced to have exactly the same view of the 3346hierarchy. It wasn't possible to vary the granularity depending on 3347the specific controller. 3348 3349In practice, these issues heavily limited which controllers could be 3350put on the same hierarchy and most configurations resorted to putting 3351each controller on its own hierarchy. Only closely related ones, such 3352as the cpu and cpuacct controllers, made sense to be put on the same 3353hierarchy. This often meant that userland ended up managing multiple 3354similar hierarchies repeating the same steps on each hierarchy 3355whenever a hierarchy management operation was necessary. 3356 3357Furthermore, support for multiple hierarchies came at a steep cost. 3358It greatly complicated cgroup core implementation but more importantly 3359the support for multiple hierarchies restricted how cgroup could be 3360used in general and what controllers was able to do. 3361 3362There was no limit on how many hierarchies there might be, which meant 3363that a thread's cgroup membership couldn't be described in finite 3364length. The key might contain any number of entries and was unlimited 3365in length, which made it highly awkward to manipulate and led to 3366addition of controllers which existed only to identify membership, 3367which in turn exacerbated the original problem of proliferating number 3368of hierarchies. 3369 3370Also, as a controller couldn't have any expectation regarding the 3371topologies of hierarchies other controllers might be on, each 3372controller had to assume that all other controllers were attached to 3373completely orthogonal hierarchies. This made it impossible, or at 3374least very cumbersome, for controllers to cooperate with each other. 3375 3376In most use cases, putting controllers on hierarchies which are 3377completely orthogonal to each other isn't necessary. What usually is 3378called for is the ability to have differing levels of granularity 3379depending on the specific controller. In other words, hierarchy may 3380be collapsed from leaf towards root when viewed from specific 3381controllers. For example, a given configuration might not care about 3382how memory is distributed beyond a certain level while still wanting 3383to control how CPU cycles are distributed. 3384 3385 3386Thread Granularity 3387------------------ 3388 3389cgroup v1 allowed threads of a process to belong to different cgroups. 3390This didn't make sense for some controllers and those controllers 3391ended up implementing different ways to ignore such situations but 3392much more importantly it blurred the line between API exposed to 3393individual applications and system management interface. 3394 3395Generally, in-process knowledge is available only to the process 3396itself; thus, unlike service-level organization of processes, 3397categorizing threads of a process requires active participation from 3398the application which owns the target process. 3399 3400cgroup v1 had an ambiguously defined delegation model which got abused 3401in combination with thread granularity. cgroups were delegated to 3402individual applications so that they can create and manage their own 3403sub-hierarchies and control resource distributions along them. This 3404effectively raised cgroup to the status of a syscall-like API exposed 3405to lay programs. 3406 3407First of all, cgroup has a fundamentally inadequate interface to be 3408exposed this way. For a process to access its own knobs, it has to 3409extract the path on the target hierarchy from /proc/self/cgroup, 3410construct the path by appending the name of the knob to the path, open 3411and then read and/or write to it. This is not only extremely clunky 3412and unusual but also inherently racy. There is no conventional way to 3413define transaction across the required steps and nothing can guarantee 3414that the process would actually be operating on its own sub-hierarchy. 3415 3416cgroup controllers implemented a number of knobs which would never be 3417accepted as public APIs because they were just adding control knobs to 3418system-management pseudo filesystem. cgroup ended up with interface 3419knobs which were not properly abstracted or refined and directly 3420revealed kernel internal details. These knobs got exposed to 3421individual applications through the ill-defined delegation mechanism 3422effectively abusing cgroup as a shortcut to implementing public APIs 3423without going through the required scrutiny. 3424 3425This was painful for both userland and kernel. Userland ended up with 3426misbehaving and poorly abstracted interfaces and kernel exposing and 3427locked into constructs inadvertently. 3428 3429 3430Competition Between Inner Nodes and Threads 3431------------------------------------------- 3432 3433cgroup v1 allowed threads to be in any cgroups which created an 3434interesting problem where threads belonging to a parent cgroup and its 3435children cgroups competed for resources. This was nasty as two 3436different types of entities competed and there was no obvious way to 3437settle it. Different controllers did different things. 3438 3439The cpu controller considered threads and cgroups as equivalents and 3440mapped nice levels to cgroup weights. This worked for some cases but 3441fell flat when children wanted to be allocated specific ratios of CPU 3442cycles and the number of internal threads fluctuated - the ratios 3443constantly changed as the number of competing entities fluctuated. 3444There also were other issues. The mapping from nice level to weight 3445wasn't obvious or universal, and there were various other knobs which 3446simply weren't available for threads. 3447 3448The io controller implicitly created a hidden leaf node for each 3449cgroup to host the threads. The hidden leaf had its own copies of all 3450the knobs with ``leaf_`` prefixed. While this allowed equivalent 3451control over internal threads, it was with serious drawbacks. It 3452always added an extra layer of nesting which wouldn't be necessary 3453otherwise, made the interface messy and significantly complicated the 3454implementation. 3455 3456The memory controller didn't have a way to control what happened 3457between internal tasks and child cgroups and the behavior was not 3458clearly defined. There were attempts to add ad-hoc behaviors and 3459knobs to tailor the behavior to specific workloads which would have 3460led to problems extremely difficult to resolve in the long term. 3461 3462Multiple controllers struggled with internal tasks and came up with 3463different ways to deal with it; unfortunately, all the approaches were 3464severely flawed and, furthermore, the widely different behaviors 3465made cgroup as a whole highly inconsistent. 3466 3467This clearly is a problem which needs to be addressed from cgroup core 3468in a uniform way. 3469 3470 3471Other Interface Issues 3472---------------------- 3473 3474cgroup v1 grew without oversight and developed a large number of 3475idiosyncrasies and inconsistencies. One issue on the cgroup core side 3476was how an empty cgroup was notified - a userland helper binary was 3477forked and executed for each event. The event delivery wasn't 3478recursive or delegatable. The limitations of the mechanism also led 3479to in-kernel event delivery filtering mechanism further complicating 3480the interface. 3481 3482Controller interfaces were problematic too. An extreme example is 3483controllers completely ignoring hierarchical organization and treating 3484all cgroups as if they were all located directly under the root 3485cgroup. Some controllers exposed a large amount of inconsistent 3486implementation details to userland. 3487 3488There also was no consistency across controllers. When a new cgroup 3489was created, some controllers defaulted to not imposing extra 3490restrictions while others disallowed any resource usage until 3491explicitly configured. Configuration knobs for the same type of 3492control used widely differing naming schemes and formats. Statistics 3493and information knobs were named arbitrarily and used different 3494formats and units even in the same controller. 3495 3496cgroup v2 establishes common conventions where appropriate and updates 3497controllers so that they expose minimal and consistent interfaces. 3498 3499 3500Controller Issues and Remedies 3501------------------------------ 3502 3503Memory 3504~~~~~~ 3505 3506The original lower boundary, the soft limit, is defined as a limit 3507that is per default unset. As a result, the set of cgroups that 3508global reclaim prefers is opt-in, rather than opt-out. The costs for 3509optimizing these mostly negative lookups are so high that the 3510implementation, despite its enormous size, does not even provide the 3511basic desirable behavior. First off, the soft limit has no 3512hierarchical meaning. All configured groups are organized in a global 3513rbtree and treated like equal peers, regardless where they are located 3514in the hierarchy. This makes subtree delegation impossible. Second, 3515the soft limit reclaim pass is so aggressive that it not just 3516introduces high allocation latencies into the system, but also impacts 3517system performance due to overreclaim, to the point where the feature 3518becomes self-defeating. 3519 3520The memory.low boundary on the other hand is a top-down allocated 3521reserve. A cgroup enjoys reclaim protection when it's within its 3522effective low, which makes delegation of subtrees possible. It also 3523enjoys having reclaim pressure proportional to its overage when 3524above its effective low. 3525 3526The original high boundary, the hard limit, is defined as a strict 3527limit that can not budge, even if the OOM killer has to be called. 3528But this generally goes against the goal of making the most out of the 3529available memory. The memory consumption of workloads varies during 3530runtime, and that requires users to overcommit. But doing that with a 3531strict upper limit requires either a fairly accurate prediction of the 3532working set size or adding slack to the limit. Since working set size 3533estimation is hard and error prone, and getting it wrong results in 3534OOM kills, most users tend to err on the side of a looser limit and 3535end up wasting precious resources. 3536 3537The memory.high boundary on the other hand can be set much more 3538conservatively. When hit, it throttles allocations by forcing them 3539into direct reclaim to work off the excess, but it never invokes the 3540OOM killer. As a result, a high boundary that is chosen too 3541aggressively will not terminate the processes, but instead it will 3542lead to gradual performance degradation. The user can monitor this 3543and make corrections until the minimal memory footprint that still 3544gives acceptable performance is found. 3545 3546In extreme cases, with many concurrent allocations and a complete 3547breakdown of reclaim progress within the group, the high boundary can 3548be exceeded. But even then it's mostly better to satisfy the 3549allocation from the slack available in other groups or the rest of the 3550system than killing the group. Otherwise, memory.max is there to 3551limit this type of spillover and ultimately contain buggy or even 3552malicious applications. 3553 3554Setting the original memory.limit_in_bytes below the current usage was 3555subject to a race condition, where concurrent charges could cause the 3556limit setting to fail. memory.max on the other hand will first set the 3557limit to prevent new charges, and then reclaim and OOM kill until the 3558new limit is met - or the task writing to memory.max is killed. 3559 3560The combined memory+swap accounting and limiting is replaced by real 3561control over swap space. 3562 3563The main argument for a combined memory+swap facility in the original 3564cgroup design was that global or parental pressure would always be 3565able to swap all anonymous memory of a child group, regardless of the 3566child's own (possibly untrusted) configuration. However, untrusted 3567groups can sabotage swapping by other means - such as referencing its 3568anonymous memory in a tight loop - and an admin can not assume full 3569swappability when overcommitting untrusted jobs. 3570 3571For trusted jobs, on the other hand, a combined counter is not an 3572intuitive userspace interface, and it flies in the face of the idea 3573that cgroup controllers should account and limit specific physical 3574resources. Swap space is a resource like all others in the system, 3575and that's why unified hierarchy allows distributing it separately. 3576