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