1.. _cgroup-v2: 2 3================ 4Control Group v2 5================ 6 7:Date: October, 2015 8:Author: Tejun Heo <tj@kernel.org> 9 10This is the authoritative documentation on the design, interface and 11conventions of cgroup v2. It describes all userland-visible aspects 12of cgroup including core and specific controller behaviors. All 13future changes must be reflected in this document. Documentation for 14v1 is available under :ref:`Documentation/admin-guide/cgroup-v1/index.rst <cgroup-v1>`. 15 16.. CONTENTS 17 18 1. Introduction 19 1-1. Terminology 20 1-2. What is cgroup? 21 2. Basic Operations 22 2-1. Mounting 23 2-2. Organizing Processes and Threads 24 2-2-1. Processes 25 2-2-2. Threads 26 2-3. [Un]populated Notification 27 2-4. Controlling Controllers 28 2-4-1. Enabling and Disabling 29 2-4-2. Top-down Constraint 30 2-4-3. No Internal Process Constraint 31 2-5. Delegation 32 2-5-1. Model of Delegation 33 2-5-2. Delegation Containment 34 2-6. Guidelines 35 2-6-1. Organize Once and Control 36 2-6-2. Avoid Name Collisions 37 3. Resource Distribution Models 38 3-1. Weights 39 3-2. Limits 40 3-3. Protections 41 3-4. Allocations 42 4. Interface Files 43 4-1. Format 44 4-2. Conventions 45 4-3. Core Interface Files 46 5. Controllers 47 5-1. CPU 48 5-1-1. CPU Interface Files 49 5-2. Memory 50 5-2-1. Memory Interface Files 51 5-2-2. Usage Guidelines 52 5-2-3. Memory Ownership 53 5-3. IO 54 5-3-1. IO Interface Files 55 5-3-2. Writeback 56 5-3-3. IO Latency 57 5-3-3-1. How IO Latency Throttling Works 58 5-3-3-2. IO Latency Interface Files 59 5-3-4. IO Priority 60 5-4. PID 61 5-4-1. PID Interface Files 62 5-5. Cpuset 63 5.5-1. Cpuset Interface Files 64 5-6. Device 65 5-7. RDMA 66 5-7-1. RDMA Interface Files 67 5-8. DMEM 68 5-9. HugeTLB 69 5.9-1. HugeTLB Interface Files 70 5-10. Misc 71 5.10-1 Miscellaneous cgroup Interface Files 72 5.10-2 Migration and Ownership 73 5-11. Others 74 5-11-1. perf_event 75 5-N. Non-normative information 76 5-N-1. CPU controller root cgroup process behaviour 77 5-N-2. IO controller root cgroup process behaviour 78 6. Namespace 79 6-1. Basics 80 6-2. The Root and Views 81 6-3. Migration and setns(2) 82 6-4. Interaction with Other Namespaces 83 P. Information on Kernel Programming 84 P-1. Filesystem Support for Writeback 85 D. Deprecated v1 Core Features 86 R. Issues with v1 and Rationales for v2 87 R-1. Multiple Hierarchies 88 R-2. Thread Granularity 89 R-3. Competition Between Inner Nodes and Threads 90 R-4. Other Interface Issues 91 R-5. Controller Issues and Remedies 92 R-5-1. Memory 93 94 95Introduction 96============ 97 98Terminology 99----------- 100 101"cgroup" stands for "control group" and is never capitalized. The 102singular form is used to designate the whole feature and also as a 103qualifier as in "cgroup controllers". When explicitly referring to 104multiple individual control groups, the plural form "cgroups" is used. 105 106 107What is cgroup? 108--------------- 109 110cgroup is a mechanism to organize processes hierarchically and 111distribute system resources along the hierarchy in a controlled and 112configurable manner. 113 114cgroup is largely composed of two parts - the core and controllers. 115cgroup core is primarily responsible for hierarchically organizing 116processes. A cgroup controller is usually responsible for 117distributing a specific type of system resource along the hierarchy 118although there are utility controllers which serve purposes other than 119resource distribution. 120 121cgroups form a tree structure and every process in the system belongs 122to one and only one cgroup. All threads of a process belong to the 123same cgroup. On creation, all processes are put in the cgroup that 124the parent process belongs to at the time. A process can be migrated 125to another cgroup. Migration of a process doesn't affect already 126existing descendant processes. 127 128Following certain structural constraints, controllers may be enabled or 129disabled selectively on a cgroup. All controller behaviors are 130hierarchical - if a controller is enabled on a cgroup, it affects all 131processes which belong to the cgroups consisting the inclusive 132sub-hierarchy of the cgroup. When a controller is enabled on a nested 133cgroup, it always restricts the resource distribution further. The 134restrictions set closer to the root in the hierarchy can not be 135overridden from further away. 136 137 138Basic Operations 139================ 140 141Mounting 142-------- 143 144Unlike v1, cgroup v2 has only single hierarchy. The cgroup v2 145hierarchy can be mounted with the following mount command:: 146 147 # mount -t cgroup2 none $MOUNT_POINT 148 149cgroup2 filesystem has the magic number 0x63677270 ("cgrp"). All 150controllers which support v2 and are not bound to a v1 hierarchy are 151automatically bound to the v2 hierarchy and show up at the root. 152Controllers which are not in active use in the v2 hierarchy can be 153bound to other hierarchies. This allows mixing v2 hierarchy with the 154legacy v1 multiple hierarchies in a fully backward compatible way. 155 156A controller can be moved across hierarchies only after the controller 157is no longer referenced in its current hierarchy. Because per-cgroup 158controller states are destroyed asynchronously and controllers may 159have lingering references, a controller may not show up immediately on 160the v2 hierarchy after the final umount of the previous hierarchy. 161Similarly, a controller should be fully disabled to be moved out of 162the unified hierarchy and it may take some time for the disabled 163controller to become available for other hierarchies; furthermore, due 164to inter-controller dependencies, other controllers may need to be 165disabled too. 166 167While useful for development and manual configurations, moving 168controllers dynamically between the v2 and other hierarchies is 169strongly discouraged for production use. It is recommended to decide 170the hierarchies and controller associations before starting using the 171controllers after system boot. 172 173During transition to v2, system management software might still 174automount the v1 cgroup filesystem and so hijack all controllers 175during boot, before manual intervention is possible. To make testing 176and experimenting easier, the kernel parameter cgroup_no_v1= allows 177disabling controllers in v1 and make them always available in v2. 178 179cgroup v2 currently supports the following mount options. 180 181 nsdelegate 182 Consider cgroup namespaces as delegation boundaries. This 183 option is system wide and can only be set on mount or modified 184 through remount from the init namespace. The mount option is 185 ignored on non-init namespace mounts. Please refer to the 186 Delegation section for details. 187 188 favordynmods 189 Reduce the latencies of dynamic cgroup modifications such as 190 task migrations and controller on/offs at the cost of making 191 hot path operations such as forks and exits more expensive. 192 The static usage pattern of creating a cgroup, enabling 193 controllers, and then seeding it with CLONE_INTO_CGROUP is 194 not affected by this option. 195 196 memory_localevents 197 Only populate memory.events with data for the current cgroup, 198 and not any subtrees. This is legacy behaviour, the default 199 behaviour without this option is to include subtree counts. 200 This option is system wide and can only be set on mount or 201 modified through remount from the init namespace. The mount 202 option is ignored on non-init namespace mounts. 203 204 memory_recursiveprot 205 Recursively apply memory.min and memory.low protection to 206 entire subtrees, without requiring explicit downward 207 propagation into leaf cgroups. This allows protecting entire 208 subtrees from one another, while retaining free competition 209 within those subtrees. This should have been the default 210 behavior but is a mount-option to avoid regressing setups 211 relying on the original semantics (e.g. specifying bogusly 212 high 'bypass' protection values at higher tree levels). 213 214 memory_hugetlb_accounting 215 Count HugeTLB memory usage towards the cgroup's overall 216 memory usage for the memory controller (for the purpose of 217 statistics reporting and memory protetion). This is a new 218 behavior that could regress existing setups, so it must be 219 explicitly opted in with this mount option. 220 221 A few caveats to keep in mind: 222 223 * There is no HugeTLB pool management involved in the memory 224 controller. The pre-allocated pool does not belong to anyone. 225 Specifically, when a new HugeTLB folio is allocated to 226 the pool, it is not accounted for from the perspective of the 227 memory controller. It is only charged to a cgroup when it is 228 actually used (for e.g at page fault time). Host memory 229 overcommit management has to consider this when configuring 230 hard limits. In general, HugeTLB pool management should be 231 done via other mechanisms (such as the HugeTLB controller). 232 * Failure to charge a HugeTLB folio to the memory controller 233 results in SIGBUS. This could happen even if the HugeTLB pool 234 still has pages available (but the cgroup limit is hit and 235 reclaim attempt fails). 236 * Charging HugeTLB memory towards the memory controller affects 237 memory protection and reclaim dynamics. Any userspace tuning 238 (of low, min limits for e.g) needs to take this into account. 239 * HugeTLB pages utilized while this option is not selected 240 will not be tracked by the memory controller (even if cgroup 241 v2 is remounted later on). 242 243 pids_localevents 244 The option restores v1-like behavior of pids.events:max, that is only 245 local (inside cgroup proper) fork failures are counted. Without this 246 option pids.events.max represents any pids.max enforcemnt across 247 cgroup's subtree. 248 249 250 251Organizing Processes and Threads 252-------------------------------- 253 254Processes 255~~~~~~~~~ 256 257Initially, only the root cgroup exists to which all processes belong. 258A child cgroup can be created by creating a sub-directory:: 259 260 # mkdir $CGROUP_NAME 261 262A given cgroup may have multiple child cgroups forming a tree 263structure. Each cgroup has a read-writable interface file 264"cgroup.procs". When read, it lists the PIDs of all processes which 265belong to the cgroup one-per-line. The PIDs are not ordered and the 266same PID may show up more than once if the process got moved to 267another cgroup and then back or the PID got recycled while reading. 268 269A process can be migrated into a cgroup by writing its PID to the 270target cgroup's "cgroup.procs" file. Only one process can be migrated 271on a single write(2) call. If a process is composed of multiple 272threads, writing the PID of any thread migrates all threads of the 273process. 274 275When a process forks a child process, the new process is born into the 276cgroup that the forking process belongs to at the time of the 277operation. After exit, a process stays associated with the cgroup 278that it belonged to at the time of exit until it's reaped; however, a 279zombie process does not appear in "cgroup.procs" and thus can't be 280moved to another cgroup. 281 282A cgroup which doesn't have any children or live processes can be 283destroyed by removing the directory. Note that a cgroup which doesn't 284have any children and is associated only with zombie processes is 285considered empty and can be removed:: 286 287 # rmdir $CGROUP_NAME 288 289"/proc/$PID/cgroup" lists a process's cgroup membership. If legacy 290cgroup is in use in the system, this file may contain multiple lines, 291one for each hierarchy. The entry for cgroup v2 is always in the 292format "0::$PATH":: 293 294 # cat /proc/842/cgroup 295 ... 296 0::/test-cgroup/test-cgroup-nested 297 298If the process becomes a zombie and the cgroup it was associated with 299is removed subsequently, " (deleted)" is appended to the path:: 300 301 # cat /proc/842/cgroup 302 ... 303 0::/test-cgroup/test-cgroup-nested (deleted) 304 305 306Threads 307~~~~~~~ 308 309cgroup v2 supports thread granularity for a subset of controllers to 310support use cases requiring hierarchical resource distribution across 311the threads of a group of processes. By default, all threads of a 312process belong to the same cgroup, which also serves as the resource 313domain to host resource consumptions which are not specific to a 314process or thread. The thread mode allows threads to be spread across 315a subtree while still maintaining the common resource domain for them. 316 317Controllers which support thread mode are called threaded controllers. 318The ones which don't are called domain controllers. 319 320Marking a cgroup threaded makes it join the resource domain of its 321parent as a threaded cgroup. The parent may be another threaded 322cgroup whose resource domain is further up in the hierarchy. The root 323of a threaded subtree, that is, the nearest ancestor which is not 324threaded, is called threaded domain or thread root interchangeably and 325serves as the resource domain for the entire subtree. 326 327Inside a threaded subtree, threads of a process can be put in 328different cgroups and are not subject to the no internal process 329constraint - threaded controllers can be enabled on non-leaf cgroups 330whether they have threads in them or not. 331 332As the threaded domain cgroup hosts all the domain resource 333consumptions of the subtree, it is considered to have internal 334resource consumptions whether there are processes in it or not and 335can't have populated child cgroups which aren't threaded. Because the 336root cgroup is not subject to no internal process constraint, it can 337serve both as a threaded domain and a parent to domain cgroups. 338 339The current operation mode or type of the cgroup is shown in the 340"cgroup.type" file which indicates whether the cgroup is a normal 341domain, a domain which is serving as the domain of a threaded subtree, 342or a threaded cgroup. 343 344On creation, a cgroup is always a domain cgroup and can be made 345threaded by writing "threaded" to the "cgroup.type" file. The 346operation is single direction:: 347 348 # echo threaded > cgroup.type 349 350Once threaded, the cgroup can't be made a domain again. To enable the 351thread mode, the following conditions must be met. 352 353- As the cgroup will join the parent's resource domain. The parent 354 must either be a valid (threaded) domain or a threaded cgroup. 355 356- When the parent is an unthreaded domain, it must not have any domain 357 controllers enabled or populated domain children. The root is 358 exempt from this requirement. 359 360Topology-wise, a cgroup can be in an invalid state. Please consider 361the following topology:: 362 363 A (threaded domain) - B (threaded) - C (domain, just created) 364 365C is created as a domain but isn't connected to a parent which can 366host child domains. C can't be used until it is turned into a 367threaded cgroup. "cgroup.type" file will report "domain (invalid)" in 368these cases. Operations which fail due to invalid topology use 369EOPNOTSUPP as the errno. 370 371A domain cgroup is turned into a threaded domain when one of its child 372cgroup becomes threaded or threaded controllers are enabled in the 373"cgroup.subtree_control" file while there are processes in the cgroup. 374A threaded domain reverts to a normal domain when the conditions 375clear. 376 377When read, "cgroup.threads" contains the list of the thread IDs of all 378threads in the cgroup. Except that the operations are per-thread 379instead of per-process, "cgroup.threads" has the same format and 380behaves the same way as "cgroup.procs". While "cgroup.threads" can be 381written to in any cgroup, as it can only move threads inside the same 382threaded domain, its operations are confined inside each threaded 383subtree. 384 385The threaded domain cgroup serves as the resource domain for the whole 386subtree, and, while the threads can be scattered across the subtree, 387all the processes are considered to be in the threaded domain cgroup. 388"cgroup.procs" in a threaded domain cgroup contains the PIDs of all 389processes in the subtree and is not readable in the subtree proper. 390However, "cgroup.procs" can be written to from anywhere in the subtree 391to migrate all threads of the matching process to the cgroup. 392 393Only threaded controllers can be enabled in a threaded subtree. When 394a threaded controller is enabled inside a threaded subtree, it only 395accounts for and controls resource consumptions associated with the 396threads in the cgroup and its descendants. All consumptions which 397aren't tied to a specific thread belong to the threaded domain cgroup. 398 399Because a threaded subtree is exempt from no internal process 400constraint, a threaded controller must be able to handle competition 401between threads in a non-leaf cgroup and its child cgroups. Each 402threaded controller defines how such competitions are handled. 403 404Currently, the following controllers are threaded and can be enabled 405in a threaded cgroup:: 406 407- cpu 408- cpuset 409- perf_event 410- pids 411 412[Un]populated Notification 413-------------------------- 414 415Each non-root cgroup has a "cgroup.events" file which contains 416"populated" field indicating whether the cgroup's sub-hierarchy has 417live processes in it. Its value is 0 if there is no live process in 418the cgroup and its descendants; otherwise, 1. poll and [id]notify 419events are triggered when the value changes. This can be used, for 420example, to start a clean-up operation after all processes of a given 421sub-hierarchy have exited. The populated state updates and 422notifications are recursive. Consider the following sub-hierarchy 423where the numbers in the parentheses represent the numbers of processes 424in each cgroup:: 425 426 A(4) - B(0) - C(1) 427 \ D(0) 428 429A, B and C's "populated" fields would be 1 while D's 0. After the one 430process in C exits, B and C's "populated" fields would flip to "0" and 431file modified events will be generated on the "cgroup.events" files of 432both cgroups. 433 434 435Controlling Controllers 436----------------------- 437 438Enabling and Disabling 439~~~~~~~~~~~~~~~~~~~~~~ 440 441Each cgroup has a "cgroup.controllers" file which lists all 442controllers available for the cgroup to enable:: 443 444 # cat cgroup.controllers 445 cpu io memory 446 447No controller is enabled by default. Controllers can be enabled and 448disabled by writing to the "cgroup.subtree_control" file:: 449 450 # echo "+cpu +memory -io" > cgroup.subtree_control 451 452Only controllers which are listed in "cgroup.controllers" can be 453enabled. When multiple operations are specified as above, either they 454all succeed or fail. If multiple operations on the same controller 455are specified, the last one is effective. 456 457Enabling a controller in a cgroup indicates that the distribution of 458the target resource across its immediate children will be controlled. 459Consider the following sub-hierarchy. The enabled controllers are 460listed in parentheses:: 461 462 A(cpu,memory) - B(memory) - C() 463 \ D() 464 465As A has "cpu" and "memory" enabled, A will control the distribution 466of CPU cycles and memory to its children, in this case, B. As B has 467"memory" enabled but not "CPU", C and D will compete freely on CPU 468cycles but their division of memory available to B will be controlled. 469 470As a controller regulates the distribution of the target resource to 471the cgroup's children, enabling it creates the controller's interface 472files in the child cgroups. In the above example, enabling "cpu" on B 473would create the "cpu." prefixed controller interface files in C and 474D. Likewise, disabling "memory" from B would remove the "memory." 475prefixed controller interface files from C and D. This means that the 476controller interface files - anything which doesn't start with 477"cgroup." are owned by the parent rather than the cgroup itself. 478 479 480Top-down Constraint 481~~~~~~~~~~~~~~~~~~~ 482 483Resources are distributed top-down and a cgroup can further distribute 484a resource only if the resource has been distributed to it from the 485parent. This means that all non-root "cgroup.subtree_control" files 486can only contain controllers which are enabled in the parent's 487"cgroup.subtree_control" file. A controller can be enabled only if 488the parent has the controller enabled and a controller can't be 489disabled if one or more children have it enabled. 490 491 492No Internal Process Constraint 493~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~ 494 495Non-root cgroups can distribute domain resources to their children 496only when they don't have any processes of their own. In other words, 497only domain cgroups which don't contain any processes can have domain 498controllers enabled in their "cgroup.subtree_control" files. 499 500This guarantees that, when a domain controller is looking at the part 501of the hierarchy which has it enabled, processes are always only on 502the leaves. This rules out situations where child cgroups compete 503against internal processes of the parent. 504 505The root cgroup is exempt from this restriction. Root contains 506processes and anonymous resource consumption which can't be associated 507with any other cgroups and requires special treatment from most 508controllers. How resource consumption in the root cgroup is governed 509is up to each controller (for more information on this topic please 510refer to the Non-normative information section in the Controllers 511chapter). 512 513Note that the restriction doesn't get in the way if there is no 514enabled controller in the cgroup's "cgroup.subtree_control". This is 515important as otherwise it wouldn't be possible to create children of a 516populated cgroup. To control resource distribution of a cgroup, the 517cgroup must create children and transfer all its processes to the 518children before enabling controllers in its "cgroup.subtree_control" 519file. 520 521 522Delegation 523---------- 524 525Model of Delegation 526~~~~~~~~~~~~~~~~~~~ 527 528A cgroup can be delegated in two ways. First, to a less privileged 529user by granting write access of the directory and its "cgroup.procs", 530"cgroup.threads" and "cgroup.subtree_control" files to the user. 531Second, if the "nsdelegate" mount option is set, automatically to a 532cgroup namespace on namespace creation. 533 534Because the resource control interface files in a given directory 535control the distribution of the parent's resources, the delegatee 536shouldn't be allowed to write to them. For the first method, this is 537achieved by not granting access to these files. For the second, files 538outside the namespace should be hidden from the delegatee by the means 539of at least mount namespacing, and the kernel rejects writes to all 540files on a namespace root from inside the cgroup namespace, except for 541those files listed in "/sys/kernel/cgroup/delegate" (including 542"cgroup.procs", "cgroup.threads", "cgroup.subtree_control", etc.). 543 544The end results are equivalent for both delegation types. Once 545delegated, the user can build sub-hierarchy under the directory, 546organize processes inside it as it sees fit and further distribute the 547resources it received from the parent. The limits and other settings 548of all resource controllers are hierarchical and regardless of what 549happens in the delegated sub-hierarchy, nothing can escape the 550resource restrictions imposed by the parent. 551 552Currently, cgroup doesn't impose any restrictions on the number of 553cgroups in or nesting depth of a delegated sub-hierarchy; however, 554this may be limited explicitly in the future. 555 556 557Delegation Containment 558~~~~~~~~~~~~~~~~~~~~~~ 559 560A delegated sub-hierarchy is contained in the sense that processes 561can't be moved into or out of the sub-hierarchy by the delegatee. 562 563For delegations to a less privileged user, this is achieved by 564requiring the following conditions for a process with a non-root euid 565to migrate a target process into a cgroup by writing its PID to the 566"cgroup.procs" file. 567 568- The writer must have write access to the "cgroup.procs" file. 569 570- The writer must have write access to the "cgroup.procs" file of the 571 common ancestor of the source and destination cgroups. 572 573The above two constraints ensure that while a delegatee may migrate 574processes around freely in the delegated sub-hierarchy it can't pull 575in from or push out to outside the sub-hierarchy. 576 577For an example, let's assume cgroups C0 and C1 have been delegated to 578user U0 who created C00, C01 under C0 and C10 under C1 as follows and 579all processes under C0 and C1 belong to U0:: 580 581 ~~~~~~~~~~~~~ - C0 - C00 582 ~ cgroup ~ \ C01 583 ~ hierarchy ~ 584 ~~~~~~~~~~~~~ - C1 - C10 585 586Let's also say U0 wants to write the PID of a process which is 587currently in C10 into "C00/cgroup.procs". U0 has write access to the 588file; however, the common ancestor of the source cgroup C10 and the 589destination cgroup C00 is above the points of delegation and U0 would 590not have write access to its "cgroup.procs" files and thus the write 591will be denied with -EACCES. 592 593For delegations to namespaces, containment is achieved by requiring 594that both the source and destination cgroups are reachable from the 595namespace of the process which is attempting the migration. If either 596is not reachable, the migration is rejected with -ENOENT. 597 598 599Guidelines 600---------- 601 602Organize Once and Control 603~~~~~~~~~~~~~~~~~~~~~~~~~ 604 605Migrating a process across cgroups is a relatively expensive operation 606and stateful resources such as memory are not moved together with the 607process. This is an explicit design decision as there often exist 608inherent trade-offs between migration and various hot paths in terms 609of synchronization cost. 610 611As such, migrating processes across cgroups frequently as a means to 612apply different resource restrictions is discouraged. A workload 613should be assigned to a cgroup according to the system's logical and 614resource structure once on start-up. Dynamic adjustments to resource 615distribution can be made by changing controller configuration through 616the interface files. 617 618 619Avoid Name Collisions 620~~~~~~~~~~~~~~~~~~~~~ 621 622Interface files for a cgroup and its children cgroups occupy the same 623directory and it is possible to create children cgroups which collide 624with interface files. 625 626All cgroup core interface files are prefixed with "cgroup." and each 627controller's interface files are prefixed with the controller name and 628a dot. A controller's name is composed of lower case alphabets and 629'_'s but never begins with an '_' so it can be used as the prefix 630character for collision avoidance. Also, interface file names won't 631start or end with terms which are often used in categorizing workloads 632such as job, service, slice, unit or workload. 633 634cgroup doesn't do anything to prevent name collisions and it's the 635user's responsibility to avoid them. 636 637 638Resource Distribution Models 639============================ 640 641cgroup controllers implement several resource distribution schemes 642depending on the resource type and expected use cases. This section 643describes major schemes in use along with their expected behaviors. 644 645 646Weights 647------- 648 649A parent's resource is distributed by adding up the weights of all 650active children and giving each the fraction matching the ratio of its 651weight against the sum. As only children which can make use of the 652resource at the moment participate in the distribution, this is 653work-conserving. Due to the dynamic nature, this model is usually 654used for stateless resources. 655 656All weights are in the range [1, 10000] with the default at 100. This 657allows symmetric multiplicative biases in both directions at fine 658enough granularity while staying in the intuitive range. 659 660As long as the weight is in range, all configuration combinations are 661valid and there is no reason to reject configuration changes or 662process migrations. 663 664"cpu.weight" proportionally distributes CPU cycles to active children 665and is an example of this type. 666 667 668.. _cgroupv2-limits-distributor: 669 670Limits 671------ 672 673A child can only consume up to the configured amount of the resource. 674Limits can be over-committed - the sum of the limits of children can 675exceed the amount of resource available to the parent. 676 677Limits are in the range [0, max] and defaults to "max", which is noop. 678 679As limits can be over-committed, all configuration combinations are 680valid and there is no reason to reject configuration changes or 681process migrations. 682 683"io.max" limits the maximum BPS and/or IOPS that a cgroup can consume 684on an IO device and is an example of this type. 685 686.. _cgroupv2-protections-distributor: 687 688Protections 689----------- 690 691A cgroup is protected up to the configured amount of the resource 692as long as the usages of all its ancestors are under their 693protected levels. Protections can be hard guarantees or best effort 694soft boundaries. Protections can also be over-committed in which case 695only up to the amount available to the parent is protected among 696children. 697 698Protections are in the range [0, max] and defaults to 0, which is 699noop. 700 701As protections can be over-committed, all configuration combinations 702are valid and there is no reason to reject configuration changes or 703process migrations. 704 705"memory.low" implements best-effort memory protection and is an 706example of this type. 707 708 709Allocations 710----------- 711 712A cgroup is exclusively allocated a certain amount of a finite 713resource. Allocations can't be over-committed - the sum of the 714allocations of children can not exceed the amount of resource 715available to the parent. 716 717Allocations are in the range [0, max] and defaults to 0, which is no 718resource. 719 720As allocations can't be over-committed, some configuration 721combinations are invalid and should be rejected. Also, if the 722resource is mandatory for execution of processes, process migrations 723may be rejected. 724 725"cpu.rt.max" hard-allocates realtime slices and is an example of this 726type. 727 728 729Interface Files 730=============== 731 732Format 733------ 734 735All interface files should be in one of the following formats whenever 736possible:: 737 738 New-line separated values 739 (when only one value can be written at once) 740 741 VAL0\n 742 VAL1\n 743 ... 744 745 Space separated values 746 (when read-only or multiple values can be written at once) 747 748 VAL0 VAL1 ...\n 749 750 Flat keyed 751 752 KEY0 VAL0\n 753 KEY1 VAL1\n 754 ... 755 756 Nested keyed 757 758 KEY0 SUB_KEY0=VAL00 SUB_KEY1=VAL01... 759 KEY1 SUB_KEY0=VAL10 SUB_KEY1=VAL11... 760 ... 761 762For a writable file, the format for writing should generally match 763reading; however, controllers may allow omitting later fields or 764implement restricted shortcuts for most common use cases. 765 766For both flat and nested keyed files, only the values for a single key 767can be written at a time. For nested keyed files, the sub key pairs 768may be specified in any order and not all pairs have to be specified. 769 770 771Conventions 772----------- 773 774- Settings for a single feature should be contained in a single file. 775 776- The root cgroup should be exempt from resource control and thus 777 shouldn't have resource control interface files. 778 779- The default time unit is microseconds. If a different unit is ever 780 used, an explicit unit suffix must be present. 781 782- A parts-per quantity should use a percentage decimal with at least 783 two digit fractional part - e.g. 13.40. 784 785- If a controller implements weight based resource distribution, its 786 interface file should be named "weight" and have the range [1, 787 10000] with 100 as the default. The values are chosen to allow 788 enough and symmetric bias in both directions while keeping it 789 intuitive (the default is 100%). 790 791- If a controller implements an absolute resource guarantee and/or 792 limit, the interface files should be named "min" and "max" 793 respectively. If a controller implements best effort resource 794 guarantee and/or limit, the interface files should be named "low" 795 and "high" respectively. 796 797 In the above four control files, the special token "max" should be 798 used to represent upward infinity for both reading and writing. 799 800- If a setting has a configurable default value and keyed specific 801 overrides, the default entry should be keyed with "default" and 802 appear as the first entry in the file. 803 804 The default value can be updated by writing either "default $VAL" or 805 "$VAL". 806 807 When writing to update a specific override, "default" can be used as 808 the value to indicate removal of the override. Override entries 809 with "default" as the value must not appear when read. 810 811 For example, a setting which is keyed by major:minor device numbers 812 with integer values may look like the following:: 813 814 # cat cgroup-example-interface-file 815 default 150 816 8:0 300 817 818 The default value can be updated by:: 819 820 # echo 125 > cgroup-example-interface-file 821 822 or:: 823 824 # echo "default 125" > cgroup-example-interface-file 825 826 An override can be set by:: 827 828 # echo "8:16 170" > cgroup-example-interface-file 829 830 and cleared by:: 831 832 # echo "8:0 default" > cgroup-example-interface-file 833 # cat cgroup-example-interface-file 834 default 125 835 8:16 170 836 837- For events which are not very high frequency, an interface file 838 "events" should be created which lists event key value pairs. 839 Whenever a notifiable event happens, file modified event should be 840 generated on the file. 841 842 843Core Interface Files 844-------------------- 845 846All cgroup core files are prefixed with "cgroup." 847 848 cgroup.type 849 A read-write single value file which exists on non-root 850 cgroups. 851 852 When read, it indicates the current type of the cgroup, which 853 can be one of the following values. 854 855 - "domain" : A normal valid domain cgroup. 856 857 - "domain threaded" : A threaded domain cgroup which is 858 serving as the root of a threaded subtree. 859 860 - "domain invalid" : A cgroup which is in an invalid state. 861 It can't be populated or have controllers enabled. It may 862 be allowed to become a threaded cgroup. 863 864 - "threaded" : A threaded cgroup which is a member of a 865 threaded subtree. 866 867 A cgroup can be turned into a threaded cgroup by writing 868 "threaded" to this file. 869 870 cgroup.procs 871 A read-write new-line separated values file which exists on 872 all cgroups. 873 874 When read, it lists the PIDs of all processes which belong to 875 the cgroup one-per-line. The PIDs are not ordered and the 876 same PID may show up more than once if the process got moved 877 to another cgroup and then back or the PID got recycled while 878 reading. 879 880 A PID can be written to migrate the process associated with 881 the PID to the cgroup. The writer should match all of the 882 following conditions. 883 884 - It must have write access to the "cgroup.procs" file. 885 886 - It must have write access to the "cgroup.procs" file of the 887 common ancestor of the source and destination cgroups. 888 889 When delegating a sub-hierarchy, write access to this file 890 should be granted along with the containing directory. 891 892 In a threaded cgroup, reading this file fails with EOPNOTSUPP 893 as all the processes belong to the thread root. Writing is 894 supported and moves every thread of the process to the cgroup. 895 896 cgroup.threads 897 A read-write new-line separated values file which exists on 898 all cgroups. 899 900 When read, it lists the TIDs of all threads which belong to 901 the cgroup one-per-line. The TIDs are not ordered and the 902 same TID may show up more than once if the thread got moved to 903 another cgroup and then back or the TID got recycled while 904 reading. 905 906 A TID can be written to migrate the thread associated with the 907 TID to the cgroup. The writer should match all of the 908 following conditions. 909 910 - It must have write access to the "cgroup.threads" file. 911 912 - The cgroup that the thread is currently in must be in the 913 same resource domain as the destination cgroup. 914 915 - It must have write access to the "cgroup.procs" file of the 916 common ancestor of the source and destination cgroups. 917 918 When delegating a sub-hierarchy, write access to this file 919 should be granted along with the containing directory. 920 921 cgroup.controllers 922 A read-only space separated values file which exists on all 923 cgroups. 924 925 It shows space separated list of all controllers available to 926 the cgroup. The controllers are not ordered. 927 928 cgroup.subtree_control 929 A read-write space separated values file which exists on all 930 cgroups. Starts out empty. 931 932 When read, it shows space separated list of the controllers 933 which are enabled to control resource distribution from the 934 cgroup to its children. 935 936 Space separated list of controllers prefixed with '+' or '-' 937 can be written to enable or disable controllers. A controller 938 name prefixed with '+' enables the controller and '-' 939 disables. If a controller appears more than once on the list, 940 the last one is effective. When multiple enable and disable 941 operations are specified, either all succeed or all fail. 942 943 cgroup.events 944 A read-only flat-keyed file which exists on non-root cgroups. 945 The following entries are defined. Unless specified 946 otherwise, a value change in this file generates a file 947 modified event. 948 949 populated 950 1 if the cgroup or its descendants contains any live 951 processes; otherwise, 0. 952 frozen 953 1 if the cgroup is frozen; otherwise, 0. 954 955 cgroup.max.descendants 956 A read-write single value files. The default is "max". 957 958 Maximum allowed number of descent cgroups. 959 If the actual number of descendants is equal or larger, 960 an attempt to create a new cgroup in the hierarchy will fail. 961 962 cgroup.max.depth 963 A read-write single value files. The default is "max". 964 965 Maximum allowed descent depth below the current cgroup. 966 If the actual descent depth is equal or larger, 967 an attempt to create a new child cgroup will fail. 968 969 cgroup.stat 970 A read-only flat-keyed file with the following entries: 971 972 nr_descendants 973 Total number of visible descendant cgroups. 974 975 nr_dying_descendants 976 Total number of dying descendant cgroups. A cgroup becomes 977 dying after being deleted by a user. The cgroup will remain 978 in dying state for some time undefined time (which can depend 979 on system load) before being completely destroyed. 980 981 A process can't enter a dying cgroup under any circumstances, 982 a dying cgroup can't revive. 983 984 A dying cgroup can consume system resources not exceeding 985 limits, which were active at the moment of cgroup deletion. 986 987 nr_subsys_<cgroup_subsys> 988 Total number of live cgroup subsystems (e.g memory 989 cgroup) at and beneath the current cgroup. 990 991 nr_dying_subsys_<cgroup_subsys> 992 Total number of dying cgroup subsystems (e.g. memory 993 cgroup) at and beneath the current cgroup. 994 995 cgroup.freeze 996 A read-write single value file which exists on non-root cgroups. 997 Allowed values are "0" and "1". The default is "0". 998 999 Writing "1" to the file causes freezing of the cgroup and all 1000 descendant cgroups. This means that all belonging processes will 1001 be stopped and will not run until the cgroup will be explicitly 1002 unfrozen. Freezing of the cgroup may take some time; when this action 1003 is completed, the "frozen" value in the cgroup.events control file 1004 will be updated to "1" and the corresponding notification will be 1005 issued. 1006 1007 A cgroup can be frozen either by its own settings, or by settings 1008 of any ancestor cgroups. If any of ancestor cgroups is frozen, the 1009 cgroup will remain frozen. 1010 1011 Processes in the frozen cgroup can be killed by a fatal signal. 1012 They also can enter and leave a frozen cgroup: either by an explicit 1013 move by a user, or if freezing of the cgroup races with fork(). 1014 If a process is moved to a frozen cgroup, it stops. If a process is 1015 moved out of a frozen cgroup, it becomes running. 1016 1017 Frozen status of a cgroup doesn't affect any cgroup tree operations: 1018 it's possible to delete a frozen (and empty) cgroup, as well as 1019 create new sub-cgroups. 1020 1021 cgroup.kill 1022 A write-only single value file which exists in non-root cgroups. 1023 The only allowed value is "1". 1024 1025 Writing "1" to the file causes the cgroup and all descendant cgroups to 1026 be killed. This means that all processes located in the affected cgroup 1027 tree will be killed via SIGKILL. 1028 1029 Killing a cgroup tree will deal with concurrent forks appropriately and 1030 is protected against migrations. 1031 1032 In a threaded cgroup, writing this file fails with EOPNOTSUPP as 1033 killing cgroups is a process directed operation, i.e. it affects 1034 the whole thread-group. 1035 1036 cgroup.pressure 1037 A read-write single value file that allowed values are "0" and "1". 1038 The default is "1". 1039 1040 Writing "0" to the file will disable the cgroup PSI accounting. 1041 Writing "1" to the file will re-enable the cgroup PSI accounting. 1042 1043 This control attribute is not hierarchical, so disable or enable PSI 1044 accounting in a cgroup does not affect PSI accounting in descendants 1045 and doesn't need pass enablement via ancestors from root. 1046 1047 The reason this control attribute exists is that PSI accounts stalls for 1048 each cgroup separately and aggregates it at each level of the hierarchy. 1049 This may cause non-negligible overhead for some workloads when under 1050 deep level of the hierarchy, in which case this control attribute can 1051 be used to disable PSI accounting in the non-leaf cgroups. 1052 1053 irq.pressure 1054 A read-write nested-keyed file. 1055 1056 Shows pressure stall information for IRQ/SOFTIRQ. See 1057 :ref:`Documentation/accounting/psi.rst <psi>` for details. 1058 1059Controllers 1060=========== 1061 1062.. _cgroup-v2-cpu: 1063 1064CPU 1065--- 1066 1067The "cpu" controllers regulates distribution of CPU cycles. This 1068controller implements weight and absolute bandwidth limit models for 1069normal scheduling policy and absolute bandwidth allocation model for 1070realtime scheduling policy. 1071 1072In all the above models, cycles distribution is defined only on a temporal 1073base and it does not account for the frequency at which tasks are executed. 1074The (optional) utilization clamping support allows to hint the schedutil 1075cpufreq governor about the minimum desired frequency which should always be 1076provided by a CPU, as well as the maximum desired frequency, which should not 1077be exceeded by a CPU. 1078 1079WARNING: cgroup2 cpu controller doesn't yet support the (bandwidth) control of 1080realtime processes. For a kernel built with the CONFIG_RT_GROUP_SCHED option 1081enabled for group scheduling of realtime processes, the cpu controller can only 1082be enabled when all RT processes are in the root cgroup. Be aware that system 1083management software may already have placed RT processes into non-root cgroups 1084during the system boot process, and these processes may need to be moved to the 1085root cgroup before the cpu controller can be enabled with a 1086CONFIG_RT_GROUP_SCHED enabled kernel. 1087 1088With CONFIG_RT_GROUP_SCHED disabled, this limitation does not apply and some of 1089the interface files either affect realtime processes or account for them. See 1090the following section for details. Only the cpu controller is affected by 1091CONFIG_RT_GROUP_SCHED. Other controllers can be used for the resource control of 1092realtime processes irrespective of CONFIG_RT_GROUP_SCHED. 1093 1094 1095CPU Interface Files 1096~~~~~~~~~~~~~~~~~~~ 1097 1098The interaction of a process with the cpu controller depends on its scheduling 1099policy and the underlying scheduler. From the point of view of the cpu controller, 1100processes can be categorized as follows: 1101 1102* Processes under the fair-class scheduler 1103* Processes under a BPF scheduler with the ``cgroup_set_weight`` callback 1104* Everything else: ``SCHED_{FIFO,RR,DEADLINE}`` and processes under a BPF scheduler 1105 without the ``cgroup_set_weight`` callback 1106 1107For details on when a process is under the fair-class scheduler or a BPF scheduler, 1108check out :ref:`Documentation/scheduler/sched-ext.rst <sched-ext>`. 1109 1110For each of the following interface files, the above categories 1111will be referred to. All time durations are in microseconds. 1112 1113 cpu.stat 1114 A read-only flat-keyed file. 1115 This file exists whether the controller is enabled or not. 1116 1117 It always reports the following three stats, which account for all the 1118 processes in the cgroup: 1119 1120 - usage_usec 1121 - user_usec 1122 - system_usec 1123 1124 and the following five when the controller is enabled, which account for 1125 only the processes under the fair-class scheduler: 1126 1127 - nr_periods 1128 - nr_throttled 1129 - throttled_usec 1130 - nr_bursts 1131 - burst_usec 1132 1133 cpu.weight 1134 A read-write single value file which exists on non-root 1135 cgroups. The default is "100". 1136 1137 For non idle groups (cpu.idle = 0), the weight is in the 1138 range [1, 10000]. 1139 1140 If the cgroup has been configured to be SCHED_IDLE (cpu.idle = 1), 1141 then the weight will show as a 0. 1142 1143 This file affects only processes under the fair-class scheduler and a BPF 1144 scheduler with the ``cgroup_set_weight`` callback depending on what the 1145 callback actually does. 1146 1147 cpu.weight.nice 1148 A read-write single value file which exists on non-root 1149 cgroups. The default is "0". 1150 1151 The nice value is in the range [-20, 19]. 1152 1153 This interface file is an alternative interface for 1154 "cpu.weight" and allows reading and setting weight using the 1155 same values used by nice(2). Because the range is smaller and 1156 granularity is coarser for the nice values, the read value is 1157 the closest approximation of the current weight. 1158 1159 This file affects only processes under the fair-class scheduler and a BPF 1160 scheduler with the ``cgroup_set_weight`` callback depending on what the 1161 callback actually does. 1162 1163 cpu.max 1164 A read-write two value file which exists on non-root cgroups. 1165 The default is "max 100000". 1166 1167 The maximum bandwidth limit. It's in the following format:: 1168 1169 $MAX $PERIOD 1170 1171 which indicates that the group may consume up to $MAX in each 1172 $PERIOD duration. "max" for $MAX indicates no limit. If only 1173 one number is written, $MAX is updated. 1174 1175 This file affects only processes under the fair-class scheduler. 1176 1177 cpu.max.burst 1178 A read-write single value file which exists on non-root 1179 cgroups. The default is "0". 1180 1181 The burst in the range [0, $MAX]. 1182 1183 This file affects only processes under the fair-class scheduler. 1184 1185 cpu.pressure 1186 A read-write nested-keyed file. 1187 1188 Shows pressure stall information for CPU. See 1189 :ref:`Documentation/accounting/psi.rst <psi>` for details. 1190 1191 This file accounts for all the processes in the cgroup. 1192 1193 cpu.uclamp.min 1194 A read-write single value file which exists on non-root cgroups. 1195 The default is "0", i.e. no utilization boosting. 1196 1197 The requested minimum utilization (protection) as a percentage 1198 rational number, e.g. 12.34 for 12.34%. 1199 1200 This interface allows reading and setting minimum utilization clamp 1201 values similar to the sched_setattr(2). This minimum utilization 1202 value is used to clamp the task specific minimum utilization clamp, 1203 including those of realtime processes. 1204 1205 The requested minimum utilization (protection) is always capped by 1206 the current value for the maximum utilization (limit), i.e. 1207 `cpu.uclamp.max`. 1208 1209 This file affects all the processes in the cgroup. 1210 1211 cpu.uclamp.max 1212 A read-write single value file which exists on non-root cgroups. 1213 The default is "max". i.e. no utilization capping 1214 1215 The requested maximum utilization (limit) as a percentage rational 1216 number, e.g. 98.76 for 98.76%. 1217 1218 This interface allows reading and setting maximum utilization clamp 1219 values similar to the sched_setattr(2). This maximum utilization 1220 value is used to clamp the task specific maximum utilization clamp, 1221 including those of realtime processes. 1222 1223 This file affects all the processes in the cgroup. 1224 1225 cpu.idle 1226 A read-write single value file which exists on non-root cgroups. 1227 The default is 0. 1228 1229 This is the cgroup analog of the per-task SCHED_IDLE sched policy. 1230 Setting this value to a 1 will make the scheduling policy of the 1231 cgroup SCHED_IDLE. The threads inside the cgroup will retain their 1232 own relative priorities, but the cgroup itself will be treated as 1233 very low priority relative to its peers. 1234 1235 This file affects only processes under the fair-class scheduler. 1236 1237Memory 1238------ 1239 1240The "memory" controller regulates distribution of memory. Memory is 1241stateful and implements both limit and protection models. Due to the 1242intertwining between memory usage and reclaim pressure and the 1243stateful nature of memory, the distribution model is relatively 1244complex. 1245 1246While not completely water-tight, all major memory usages by a given 1247cgroup are tracked so that the total memory consumption can be 1248accounted and controlled to a reasonable extent. Currently, the 1249following types of memory usages are tracked. 1250 1251- Userland memory - page cache and anonymous memory. 1252 1253- Kernel data structures such as dentries and inodes. 1254 1255- TCP socket buffers. 1256 1257The above list may expand in the future for better coverage. 1258 1259 1260Memory Interface Files 1261~~~~~~~~~~~~~~~~~~~~~~ 1262 1263All memory amounts are in bytes. If a value which is not aligned to 1264PAGE_SIZE is written, the value may be rounded up to the closest 1265PAGE_SIZE multiple when read back. 1266 1267 memory.current 1268 A read-only single value file which exists on non-root 1269 cgroups. 1270 1271 The total amount of memory currently being used by the cgroup 1272 and its descendants. 1273 1274 memory.min 1275 A read-write single value file which exists on non-root 1276 cgroups. The default is "0". 1277 1278 Hard memory protection. If the memory usage of a cgroup 1279 is within its effective min boundary, the cgroup's memory 1280 won't be reclaimed under any conditions. If there is no 1281 unprotected reclaimable memory available, OOM killer 1282 is invoked. Above the effective min boundary (or 1283 effective low boundary if it is higher), pages are reclaimed 1284 proportionally to the overage, reducing reclaim pressure for 1285 smaller overages. 1286 1287 Effective min boundary is limited by memory.min values of 1288 all ancestor cgroups. If there is memory.min overcommitment 1289 (child cgroup or cgroups are requiring more protected memory 1290 than parent will allow), then each child cgroup will get 1291 the part of parent's protection proportional to its 1292 actual memory usage below memory.min. 1293 1294 Putting more memory than generally available under this 1295 protection is discouraged and may lead to constant OOMs. 1296 1297 If a memory cgroup is not populated with processes, 1298 its memory.min is ignored. 1299 1300 memory.low 1301 A read-write single value file which exists on non-root 1302 cgroups. The default is "0". 1303 1304 Best-effort memory protection. If the memory usage of a 1305 cgroup is within its effective low boundary, the cgroup's 1306 memory won't be reclaimed unless there is no reclaimable 1307 memory available in unprotected cgroups. 1308 Above the effective low boundary (or 1309 effective min boundary if it is higher), pages are reclaimed 1310 proportionally to the overage, reducing reclaim pressure for 1311 smaller overages. 1312 1313 Effective low boundary is limited by memory.low values of 1314 all ancestor cgroups. If there is memory.low overcommitment 1315 (child cgroup or cgroups are requiring more protected memory 1316 than parent will allow), then each child cgroup will get 1317 the part of parent's protection proportional to its 1318 actual memory usage below memory.low. 1319 1320 Putting more memory than generally available under this 1321 protection is discouraged. 1322 1323 memory.high 1324 A read-write single value file which exists on non-root 1325 cgroups. The default is "max". 1326 1327 Memory usage throttle limit. If a cgroup's usage goes 1328 over the high boundary, the processes of the cgroup are 1329 throttled and put under heavy reclaim pressure. 1330 1331 Going over the high limit never invokes the OOM killer and 1332 under extreme conditions the limit may be breached. The high 1333 limit should be used in scenarios where an external process 1334 monitors the limited cgroup to alleviate heavy reclaim 1335 pressure. 1336 1337 If memory.high is opened with O_NONBLOCK then the synchronous 1338 reclaim is bypassed. This is useful for admin processes that 1339 need to dynamically adjust the job's memory limits without 1340 expending their own CPU resources on memory reclamation. The 1341 job will trigger the reclaim and/or get throttled on its 1342 next charge request. 1343 1344 Please note that with O_NONBLOCK, there is a chance that the 1345 target memory cgroup may take indefinite amount of time to 1346 reduce usage below the limit due to delayed charge request or 1347 busy-hitting its memory to slow down reclaim. 1348 1349 memory.max 1350 A read-write single value file which exists on non-root 1351 cgroups. The default is "max". 1352 1353 Memory usage hard limit. This is the main mechanism to limit 1354 memory usage of a cgroup. If a cgroup's memory usage reaches 1355 this limit and can't be reduced, the OOM killer is invoked in 1356 the cgroup. Under certain circumstances, the usage may go 1357 over the limit temporarily. 1358 1359 In default configuration regular 0-order allocations always 1360 succeed unless OOM killer chooses current task as a victim. 1361 1362 Some kinds of allocations don't invoke the OOM killer. 1363 Caller could retry them differently, return into userspace 1364 as -ENOMEM or silently ignore in cases like disk readahead. 1365 1366 If memory.max is opened with O_NONBLOCK, then the synchronous 1367 reclaim and oom-kill are bypassed. This is useful for admin 1368 processes that need to dynamically adjust the job's memory limits 1369 without expending their own CPU resources on memory reclamation. 1370 The job will trigger the reclaim and/or oom-kill on its next 1371 charge request. 1372 1373 Please note that with O_NONBLOCK, there is a chance that the 1374 target memory cgroup may take indefinite amount of time to 1375 reduce usage below the limit due to delayed charge request or 1376 busy-hitting its memory to slow down reclaim. 1377 1378 memory.reclaim 1379 A write-only nested-keyed file which exists for all cgroups. 1380 1381 This is a simple interface to trigger memory reclaim in the 1382 target cgroup. 1383 1384 Example:: 1385 1386 echo "1G" > memory.reclaim 1387 1388 Please note that the kernel can over or under reclaim from 1389 the target cgroup. If less bytes are reclaimed than the 1390 specified amount, -EAGAIN is returned. 1391 1392 Please note that the proactive reclaim (triggered by this 1393 interface) is not meant to indicate memory pressure on the 1394 memory cgroup. Therefore socket memory balancing triggered by 1395 the memory reclaim normally is not exercised in this case. 1396 This means that the networking layer will not adapt based on 1397 reclaim induced by memory.reclaim. 1398 1399The following nested keys are defined. 1400 1401 ========== ================================ 1402 swappiness Swappiness value to reclaim with 1403 ========== ================================ 1404 1405 Specifying a swappiness value instructs the kernel to perform 1406 the reclaim with that swappiness value. Note that this has the 1407 same semantics as vm.swappiness applied to memcg reclaim with 1408 all the existing limitations and potential future extensions. 1409 1410 The valid range for swappiness is [0-200, max], setting 1411 swappiness=max exclusively reclaims anonymous memory. 1412 1413 memory.peak 1414 A read-write single value file which exists on non-root cgroups. 1415 1416 The max memory usage recorded for the cgroup and its descendants since 1417 either the creation of the cgroup or the most recent reset for that FD. 1418 1419 A write of any non-empty string to this file resets it to the 1420 current memory usage for subsequent reads through the same 1421 file descriptor. 1422 1423 memory.oom.group 1424 A read-write single value file which exists on non-root 1425 cgroups. The default value is "0". 1426 1427 Determines whether the cgroup should be treated as 1428 an indivisible workload by the OOM killer. If set, 1429 all tasks belonging to the cgroup or to its descendants 1430 (if the memory cgroup is not a leaf cgroup) are killed 1431 together or not at all. This can be used to avoid 1432 partial kills to guarantee workload integrity. 1433 1434 Tasks with the OOM protection (oom_score_adj set to -1000) 1435 are treated as an exception and are never killed. 1436 1437 If the OOM killer is invoked in a cgroup, it's not going 1438 to kill any tasks outside of this cgroup, regardless 1439 memory.oom.group values of ancestor cgroups. 1440 1441 memory.events 1442 A read-only flat-keyed file which exists on non-root cgroups. 1443 The following entries are defined. Unless specified 1444 otherwise, a value change in this file generates a file 1445 modified event. 1446 1447 Note that all fields in this file are hierarchical and the 1448 file modified event can be generated due to an event down the 1449 hierarchy. For the local events at the cgroup level see 1450 memory.events.local. 1451 1452 low 1453 The number of times the cgroup is reclaimed due to 1454 high memory pressure even though its usage is under 1455 the low boundary. This usually indicates that the low 1456 boundary is over-committed. 1457 1458 high 1459 The number of times processes of the cgroup are 1460 throttled and routed to perform direct memory reclaim 1461 because the high memory boundary was exceeded. For a 1462 cgroup whose memory usage is capped by the high limit 1463 rather than global memory pressure, this event's 1464 occurrences are expected. 1465 1466 max 1467 The number of times the cgroup's memory usage was 1468 about to go over the max boundary. If direct reclaim 1469 fails to bring it down, the cgroup goes to OOM state. 1470 1471 oom 1472 The number of time the cgroup's memory usage was 1473 reached the limit and allocation was about to fail. 1474 1475 This event is not raised if the OOM killer is not 1476 considered as an option, e.g. for failed high-order 1477 allocations or if caller asked to not retry attempts. 1478 1479 oom_kill 1480 The number of processes belonging to this cgroup 1481 killed by any kind of OOM killer. 1482 1483 oom_group_kill 1484 The number of times a group OOM has occurred. 1485 1486 memory.events.local 1487 Similar to memory.events but the fields in the file are local 1488 to the cgroup i.e. not hierarchical. The file modified event 1489 generated on this file reflects only the local events. 1490 1491 memory.stat 1492 A read-only flat-keyed file which exists on non-root cgroups. 1493 1494 This breaks down the cgroup's memory footprint into different 1495 types of memory, type-specific details, and other information 1496 on the state and past events of the memory management system. 1497 1498 All memory amounts are in bytes. 1499 1500 The entries are ordered to be human readable, and new entries 1501 can show up in the middle. Don't rely on items remaining in a 1502 fixed position; use the keys to look up specific values! 1503 1504 If the entry has no per-node counter (or not show in the 1505 memory.numa_stat). We use 'npn' (non-per-node) as the tag 1506 to indicate that it will not show in the memory.numa_stat. 1507 1508 anon 1509 Amount of memory used in anonymous mappings such as 1510 brk(), sbrk(), and mmap(MAP_ANONYMOUS). Note that 1511 some kernel configurations might account complete larger 1512 allocations (e.g., THP) if only some, but not all the 1513 memory of such an allocation is mapped anymore. 1514 1515 file 1516 Amount of memory used to cache filesystem data, 1517 including tmpfs and shared memory. 1518 1519 kernel (npn) 1520 Amount of total kernel memory, including 1521 (kernel_stack, pagetables, percpu, vmalloc, slab) in 1522 addition to other kernel memory use cases. 1523 1524 kernel_stack 1525 Amount of memory allocated to kernel stacks. 1526 1527 pagetables 1528 Amount of memory allocated for page tables. 1529 1530 sec_pagetables 1531 Amount of memory allocated for secondary page tables, 1532 this currently includes KVM mmu allocations on x86 1533 and arm64 and IOMMU page tables. 1534 1535 percpu (npn) 1536 Amount of memory used for storing per-cpu kernel 1537 data structures. 1538 1539 sock (npn) 1540 Amount of memory used in network transmission buffers 1541 1542 vmalloc (npn) 1543 Amount of memory used for vmap backed memory. 1544 1545 shmem 1546 Amount of cached filesystem data that is swap-backed, 1547 such as tmpfs, shm segments, shared anonymous mmap()s 1548 1549 zswap 1550 Amount of memory consumed by the zswap compression backend. 1551 1552 zswapped 1553 Amount of application memory swapped out to zswap. 1554 1555 file_mapped 1556 Amount of cached filesystem data mapped with mmap(). Note 1557 that some kernel configurations might account complete 1558 larger allocations (e.g., THP) if only some, but not 1559 not all the memory of such an allocation is mapped. 1560 1561 file_dirty 1562 Amount of cached filesystem data that was modified but 1563 not yet written back to disk 1564 1565 file_writeback 1566 Amount of cached filesystem data that was modified and 1567 is currently being written back to disk 1568 1569 swapcached 1570 Amount of swap cached in memory. The swapcache is accounted 1571 against both memory and swap usage. 1572 1573 anon_thp 1574 Amount of memory used in anonymous mappings backed by 1575 transparent hugepages 1576 1577 file_thp 1578 Amount of cached filesystem data backed by transparent 1579 hugepages 1580 1581 shmem_thp 1582 Amount of shm, tmpfs, shared anonymous mmap()s backed by 1583 transparent hugepages 1584 1585 inactive_anon, active_anon, inactive_file, active_file, unevictable 1586 Amount of memory, swap-backed and filesystem-backed, 1587 on the internal memory management lists used by the 1588 page reclaim algorithm. 1589 1590 As these represent internal list state (eg. shmem pages are on anon 1591 memory management lists), inactive_foo + active_foo may not be equal to 1592 the value for the foo counter, since the foo counter is type-based, not 1593 list-based. 1594 1595 slab_reclaimable 1596 Part of "slab" that might be reclaimed, such as 1597 dentries and inodes. 1598 1599 slab_unreclaimable 1600 Part of "slab" that cannot be reclaimed on memory 1601 pressure. 1602 1603 slab (npn) 1604 Amount of memory used for storing in-kernel data 1605 structures. 1606 1607 workingset_refault_anon 1608 Number of refaults of previously evicted anonymous pages. 1609 1610 workingset_refault_file 1611 Number of refaults of previously evicted file pages. 1612 1613 workingset_activate_anon 1614 Number of refaulted anonymous pages that were immediately 1615 activated. 1616 1617 workingset_activate_file 1618 Number of refaulted file pages that were immediately activated. 1619 1620 workingset_restore_anon 1621 Number of restored anonymous pages which have been detected as 1622 an active workingset before they got reclaimed. 1623 1624 workingset_restore_file 1625 Number of restored file pages which have been detected as an 1626 active workingset before they got reclaimed. 1627 1628 workingset_nodereclaim 1629 Number of times a shadow node has been reclaimed 1630 1631 pswpin (npn) 1632 Number of pages swapped into memory 1633 1634 pswpout (npn) 1635 Number of pages swapped out of memory 1636 1637 pgscan (npn) 1638 Amount of scanned pages (in an inactive LRU list) 1639 1640 pgsteal (npn) 1641 Amount of reclaimed pages 1642 1643 pgscan_kswapd (npn) 1644 Amount of scanned pages by kswapd (in an inactive LRU list) 1645 1646 pgscan_direct (npn) 1647 Amount of scanned pages directly (in an inactive LRU list) 1648 1649 pgscan_khugepaged (npn) 1650 Amount of scanned pages by khugepaged (in an inactive LRU list) 1651 1652 pgscan_proactive (npn) 1653 Amount of scanned pages proactively (in an inactive LRU list) 1654 1655 pgsteal_kswapd (npn) 1656 Amount of reclaimed pages by kswapd 1657 1658 pgsteal_direct (npn) 1659 Amount of reclaimed pages directly 1660 1661 pgsteal_khugepaged (npn) 1662 Amount of reclaimed pages by khugepaged 1663 1664 pgsteal_proactive (npn) 1665 Amount of reclaimed pages proactively 1666 1667 pgfault (npn) 1668 Total number of page faults incurred 1669 1670 pgmajfault (npn) 1671 Number of major page faults incurred 1672 1673 pgrefill (npn) 1674 Amount of scanned pages (in an active LRU list) 1675 1676 pgactivate (npn) 1677 Amount of pages moved to the active LRU list 1678 1679 pgdeactivate (npn) 1680 Amount of pages moved to the inactive LRU list 1681 1682 pglazyfree (npn) 1683 Amount of pages postponed to be freed under memory pressure 1684 1685 pglazyfreed (npn) 1686 Amount of reclaimed lazyfree pages 1687 1688 swpin_zero 1689 Number of pages swapped into memory and filled with zero, where I/O 1690 was optimized out because the page content was detected to be zero 1691 during swapout. 1692 1693 swpout_zero 1694 Number of zero-filled pages swapped out with I/O skipped due to the 1695 content being detected as zero. 1696 1697 zswpin 1698 Number of pages moved in to memory from zswap. 1699 1700 zswpout 1701 Number of pages moved out of memory to zswap. 1702 1703 zswpwb 1704 Number of pages written from zswap to swap. 1705 1706 thp_fault_alloc (npn) 1707 Number of transparent hugepages which were allocated to satisfy 1708 a page fault. This counter is not present when CONFIG_TRANSPARENT_HUGEPAGE 1709 is not set. 1710 1711 thp_collapse_alloc (npn) 1712 Number of transparent hugepages which were allocated to allow 1713 collapsing an existing range of pages. This counter is not 1714 present when CONFIG_TRANSPARENT_HUGEPAGE is not set. 1715 1716 thp_swpout (npn) 1717 Number of transparent hugepages which are swapout in one piece 1718 without splitting. 1719 1720 thp_swpout_fallback (npn) 1721 Number of transparent hugepages which were split before swapout. 1722 Usually because failed to allocate some continuous swap space 1723 for the huge page. 1724 1725 numa_pages_migrated (npn) 1726 Number of pages migrated by NUMA balancing. 1727 1728 numa_pte_updates (npn) 1729 Number of pages whose page table entries are modified by 1730 NUMA balancing to produce NUMA hinting faults on access. 1731 1732 numa_hint_faults (npn) 1733 Number of NUMA hinting faults. 1734 1735 numa_task_migrated (npn) 1736 Number of task migration by NUMA balancing. 1737 1738 numa_task_swapped (npn) 1739 Number of task swap by NUMA balancing. 1740 1741 pgdemote_kswapd 1742 Number of pages demoted by kswapd. 1743 1744 pgdemote_direct 1745 Number of pages demoted directly. 1746 1747 pgdemote_khugepaged 1748 Number of pages demoted by khugepaged. 1749 1750 pgdemote_proactive 1751 Number of pages demoted by proactively. 1752 1753 hugetlb 1754 Amount of memory used by hugetlb pages. This metric only shows 1755 up if hugetlb usage is accounted for in memory.current (i.e. 1756 cgroup is mounted with the memory_hugetlb_accounting option). 1757 1758 memory.numa_stat 1759 A read-only nested-keyed file which exists on non-root cgroups. 1760 1761 This breaks down the cgroup's memory footprint into different 1762 types of memory, type-specific details, and other information 1763 per node on the state of the memory management system. 1764 1765 This is useful for providing visibility into the NUMA locality 1766 information within an memcg since the pages are allowed to be 1767 allocated from any physical node. One of the use case is evaluating 1768 application performance by combining this information with the 1769 application's CPU allocation. 1770 1771 All memory amounts are in bytes. 1772 1773 The output format of memory.numa_stat is:: 1774 1775 type N0=<bytes in node 0> N1=<bytes in node 1> ... 1776 1777 The entries are ordered to be human readable, and new entries 1778 can show up in the middle. Don't rely on items remaining in a 1779 fixed position; use the keys to look up specific values! 1780 1781 The entries can refer to the memory.stat. 1782 1783 memory.swap.current 1784 A read-only single value file which exists on non-root 1785 cgroups. 1786 1787 The total amount of swap currently being used by the cgroup 1788 and its descendants. 1789 1790 memory.swap.high 1791 A read-write single value file which exists on non-root 1792 cgroups. The default is "max". 1793 1794 Swap usage throttle limit. If a cgroup's swap usage exceeds 1795 this limit, all its further allocations will be throttled to 1796 allow userspace to implement custom out-of-memory procedures. 1797 1798 This limit marks a point of no return for the cgroup. It is NOT 1799 designed to manage the amount of swapping a workload does 1800 during regular operation. Compare to memory.swap.max, which 1801 prohibits swapping past a set amount, but lets the cgroup 1802 continue unimpeded as long as other memory can be reclaimed. 1803 1804 Healthy workloads are not expected to reach this limit. 1805 1806 memory.swap.peak 1807 A read-write single value file which exists on non-root cgroups. 1808 1809 The max swap usage recorded for the cgroup and its descendants since 1810 the creation of the cgroup or the most recent reset for that FD. 1811 1812 A write of any non-empty string to this file resets it to the 1813 current memory usage for subsequent reads through the same 1814 file descriptor. 1815 1816 memory.swap.max 1817 A read-write single value file which exists on non-root 1818 cgroups. The default is "max". 1819 1820 Swap usage hard limit. If a cgroup's swap usage reaches this 1821 limit, anonymous memory of the cgroup will not be swapped out. 1822 1823 memory.swap.events 1824 A read-only flat-keyed file which exists on non-root cgroups. 1825 The following entries are defined. Unless specified 1826 otherwise, a value change in this file generates a file 1827 modified event. 1828 1829 high 1830 The number of times the cgroup's swap usage was over 1831 the high threshold. 1832 1833 max 1834 The number of times the cgroup's swap usage was about 1835 to go over the max boundary and swap allocation 1836 failed. 1837 1838 fail 1839 The number of times swap allocation failed either 1840 because of running out of swap system-wide or max 1841 limit. 1842 1843 When reduced under the current usage, the existing swap 1844 entries are reclaimed gradually and the swap usage may stay 1845 higher than the limit for an extended period of time. This 1846 reduces the impact on the workload and memory management. 1847 1848 memory.zswap.current 1849 A read-only single value file which exists on non-root 1850 cgroups. 1851 1852 The total amount of memory consumed by the zswap compression 1853 backend. 1854 1855 memory.zswap.max 1856 A read-write single value file which exists on non-root 1857 cgroups. The default is "max". 1858 1859 Zswap usage hard limit. If a cgroup's zswap pool reaches this 1860 limit, it will refuse to take any more stores before existing 1861 entries fault back in or are written out to disk. 1862 1863 memory.zswap.writeback 1864 A read-write single value file. The default value is "1". 1865 Note that this setting is hierarchical, i.e. the writeback would be 1866 implicitly disabled for child cgroups if the upper hierarchy 1867 does so. 1868 1869 When this is set to 0, all swapping attempts to swapping devices 1870 are disabled. This included both zswap writebacks, and swapping due 1871 to zswap store failures. If the zswap store failures are recurring 1872 (for e.g if the pages are incompressible), users can observe 1873 reclaim inefficiency after disabling writeback (because the same 1874 pages might be rejected again and again). 1875 1876 Note that this is subtly different from setting memory.swap.max to 1877 0, as it still allows for pages to be written to the zswap pool. 1878 This setting has no effect if zswap is disabled, and swapping 1879 is allowed unless memory.swap.max is set to 0. 1880 1881 memory.pressure 1882 A read-only nested-keyed file. 1883 1884 Shows pressure stall information for memory. See 1885 :ref:`Documentation/accounting/psi.rst <psi>` for details. 1886 1887 1888Usage Guidelines 1889~~~~~~~~~~~~~~~~ 1890 1891"memory.high" is the main mechanism to control memory usage. 1892Over-committing on high limit (sum of high limits > available memory) 1893and letting global memory pressure to distribute memory according to 1894usage is a viable strategy. 1895 1896Because breach of the high limit doesn't trigger the OOM killer but 1897throttles the offending cgroup, a management agent has ample 1898opportunities to monitor and take appropriate actions such as granting 1899more memory or terminating the workload. 1900 1901Determining whether a cgroup has enough memory is not trivial as 1902memory usage doesn't indicate whether the workload can benefit from 1903more memory. For example, a workload which writes data received from 1904network to a file can use all available memory but can also operate as 1905performant with a small amount of memory. A measure of memory 1906pressure - how much the workload is being impacted due to lack of 1907memory - is necessary to determine whether a workload needs more 1908memory; unfortunately, memory pressure monitoring mechanism isn't 1909implemented yet. 1910 1911 1912Memory Ownership 1913~~~~~~~~~~~~~~~~ 1914 1915A memory area is charged to the cgroup which instantiated it and stays 1916charged to the cgroup until the area is released. Migrating a process 1917to a different cgroup doesn't move the memory usages that it 1918instantiated while in the previous cgroup to the new cgroup. 1919 1920A memory area may be used by processes belonging to different cgroups. 1921To which cgroup the area will be charged is in-deterministic; however, 1922over time, the memory area is likely to end up in a cgroup which has 1923enough memory allowance to avoid high reclaim pressure. 1924 1925If a cgroup sweeps a considerable amount of memory which is expected 1926to be accessed repeatedly by other cgroups, it may make sense to use 1927POSIX_FADV_DONTNEED to relinquish the ownership of memory areas 1928belonging to the affected files to ensure correct memory ownership. 1929 1930 1931IO 1932-- 1933 1934The "io" controller regulates the distribution of IO resources. This 1935controller implements both weight based and absolute bandwidth or IOPS 1936limit distribution; however, weight based distribution is available 1937only if cfq-iosched is in use and neither scheme is available for 1938blk-mq devices. 1939 1940 1941IO Interface Files 1942~~~~~~~~~~~~~~~~~~ 1943 1944 io.stat 1945 A read-only nested-keyed file. 1946 1947 Lines are keyed by $MAJ:$MIN device numbers and not ordered. 1948 The following nested keys are defined. 1949 1950 ====== ===================== 1951 rbytes Bytes read 1952 wbytes Bytes written 1953 rios Number of read IOs 1954 wios Number of write IOs 1955 dbytes Bytes discarded 1956 dios Number of discard IOs 1957 ====== ===================== 1958 1959 An example read output follows:: 1960 1961 8:16 rbytes=1459200 wbytes=314773504 rios=192 wios=353 dbytes=0 dios=0 1962 8:0 rbytes=90430464 wbytes=299008000 rios=8950 wios=1252 dbytes=50331648 dios=3021 1963 1964 io.cost.qos 1965 A read-write nested-keyed file which exists only on the root 1966 cgroup. 1967 1968 This file configures the Quality of Service of the IO cost 1969 model based controller (CONFIG_BLK_CGROUP_IOCOST) which 1970 currently implements "io.weight" proportional control. Lines 1971 are keyed by $MAJ:$MIN device numbers and not ordered. The 1972 line for a given device is populated on the first write for 1973 the device on "io.cost.qos" or "io.cost.model". The following 1974 nested keys are defined. 1975 1976 ====== ===================================== 1977 enable Weight-based control enable 1978 ctrl "auto" or "user" 1979 rpct Read latency percentile [0, 100] 1980 rlat Read latency threshold 1981 wpct Write latency percentile [0, 100] 1982 wlat Write latency threshold 1983 min Minimum scaling percentage [1, 10000] 1984 max Maximum scaling percentage [1, 10000] 1985 ====== ===================================== 1986 1987 The controller is disabled by default and can be enabled by 1988 setting "enable" to 1. "rpct" and "wpct" parameters default 1989 to zero and the controller uses internal device saturation 1990 state to adjust the overall IO rate between "min" and "max". 1991 1992 When a better control quality is needed, latency QoS 1993 parameters can be configured. For example:: 1994 1995 8:16 enable=1 ctrl=auto rpct=95.00 rlat=75000 wpct=95.00 wlat=150000 min=50.00 max=150.0 1996 1997 shows that on sdb, the controller is enabled, will consider 1998 the device saturated if the 95th percentile of read completion 1999 latencies is above 75ms or write 150ms, and adjust the overall 2000 IO issue rate between 50% and 150% accordingly. 2001 2002 The lower the saturation point, the better the latency QoS at 2003 the cost of aggregate bandwidth. The narrower the allowed 2004 adjustment range between "min" and "max", the more conformant 2005 to the cost model the IO behavior. Note that the IO issue 2006 base rate may be far off from 100% and setting "min" and "max" 2007 blindly can lead to a significant loss of device capacity or 2008 control quality. "min" and "max" are useful for regulating 2009 devices which show wide temporary behavior changes - e.g. a 2010 ssd which accepts writes at the line speed for a while and 2011 then completely stalls for multiple seconds. 2012 2013 When "ctrl" is "auto", the parameters are controlled by the 2014 kernel and may change automatically. Setting "ctrl" to "user" 2015 or setting any of the percentile and latency parameters puts 2016 it into "user" mode and disables the automatic changes. The 2017 automatic mode can be restored by setting "ctrl" to "auto". 2018 2019 io.cost.model 2020 A read-write nested-keyed file which exists only on the root 2021 cgroup. 2022 2023 This file configures the cost model of the IO cost model based 2024 controller (CONFIG_BLK_CGROUP_IOCOST) which currently 2025 implements "io.weight" proportional control. Lines are keyed 2026 by $MAJ:$MIN device numbers and not ordered. The line for a 2027 given device is populated on the first write for the device on 2028 "io.cost.qos" or "io.cost.model". The following nested keys 2029 are defined. 2030 2031 ===== ================================ 2032 ctrl "auto" or "user" 2033 model The cost model in use - "linear" 2034 ===== ================================ 2035 2036 When "ctrl" is "auto", the kernel may change all parameters 2037 dynamically. When "ctrl" is set to "user" or any other 2038 parameters are written to, "ctrl" become "user" and the 2039 automatic changes are disabled. 2040 2041 When "model" is "linear", the following model parameters are 2042 defined. 2043 2044 ============= ======================================== 2045 [r|w]bps The maximum sequential IO throughput 2046 [r|w]seqiops The maximum 4k sequential IOs per second 2047 [r|w]randiops The maximum 4k random IOs per second 2048 ============= ======================================== 2049 2050 From the above, the builtin linear model determines the base 2051 costs of a sequential and random IO and the cost coefficient 2052 for the IO size. While simple, this model can cover most 2053 common device classes acceptably. 2054 2055 The IO cost model isn't expected to be accurate in absolute 2056 sense and is scaled to the device behavior dynamically. 2057 2058 If needed, tools/cgroup/iocost_coef_gen.py can be used to 2059 generate device-specific coefficients. 2060 2061 io.weight 2062 A read-write flat-keyed file which exists on non-root cgroups. 2063 The default is "default 100". 2064 2065 The first line is the default weight applied to devices 2066 without specific override. The rest are overrides keyed by 2067 $MAJ:$MIN device numbers and not ordered. The weights are in 2068 the range [1, 10000] and specifies the relative amount IO time 2069 the cgroup can use in relation to its siblings. 2070 2071 The default weight can be updated by writing either "default 2072 $WEIGHT" or simply "$WEIGHT". Overrides can be set by writing 2073 "$MAJ:$MIN $WEIGHT" and unset by writing "$MAJ:$MIN default". 2074 2075 An example read output follows:: 2076 2077 default 100 2078 8:16 200 2079 8:0 50 2080 2081 io.max 2082 A read-write nested-keyed file which exists on non-root 2083 cgroups. 2084 2085 BPS and IOPS based IO limit. Lines are keyed by $MAJ:$MIN 2086 device numbers and not ordered. The following nested keys are 2087 defined. 2088 2089 ===== ================================== 2090 rbps Max read bytes per second 2091 wbps Max write bytes per second 2092 riops Max read IO operations per second 2093 wiops Max write IO operations per second 2094 ===== ================================== 2095 2096 When writing, any number of nested key-value pairs can be 2097 specified in any order. "max" can be specified as the value 2098 to remove a specific limit. If the same key is specified 2099 multiple times, the outcome is undefined. 2100 2101 BPS and IOPS are measured in each IO direction and IOs are 2102 delayed if limit is reached. Temporary bursts are allowed. 2103 2104 Setting read limit at 2M BPS and write at 120 IOPS for 8:16:: 2105 2106 echo "8:16 rbps=2097152 wiops=120" > io.max 2107 2108 Reading returns the following:: 2109 2110 8:16 rbps=2097152 wbps=max riops=max wiops=120 2111 2112 Write IOPS limit can be removed by writing the following:: 2113 2114 echo "8:16 wiops=max" > io.max 2115 2116 Reading now returns the following:: 2117 2118 8:16 rbps=2097152 wbps=max riops=max wiops=max 2119 2120 io.pressure 2121 A read-only nested-keyed file. 2122 2123 Shows pressure stall information for IO. See 2124 :ref:`Documentation/accounting/psi.rst <psi>` for details. 2125 2126 2127Writeback 2128~~~~~~~~~ 2129 2130Page cache is dirtied through buffered writes and shared mmaps and 2131written asynchronously to the backing filesystem by the writeback 2132mechanism. Writeback sits between the memory and IO domains and 2133regulates the proportion of dirty memory by balancing dirtying and 2134write IOs. 2135 2136The io controller, in conjunction with the memory controller, 2137implements control of page cache writeback IOs. The memory controller 2138defines the memory domain that dirty memory ratio is calculated and 2139maintained for and the io controller defines the io domain which 2140writes out dirty pages for the memory domain. Both system-wide and 2141per-cgroup dirty memory states are examined and the more restrictive 2142of the two is enforced. 2143 2144cgroup writeback requires explicit support from the underlying 2145filesystem. Currently, cgroup writeback is implemented on ext2, ext4, 2146btrfs, f2fs, and xfs. On other filesystems, all writeback IOs are 2147attributed to the root cgroup. 2148 2149There are inherent differences in memory and writeback management 2150which affects how cgroup ownership is tracked. Memory is tracked per 2151page while writeback per inode. For the purpose of writeback, an 2152inode is assigned to a cgroup and all IO requests to write dirty pages 2153from the inode are attributed to that cgroup. 2154 2155As cgroup ownership for memory is tracked per page, there can be pages 2156which are associated with different cgroups than the one the inode is 2157associated with. These are called foreign pages. The writeback 2158constantly keeps track of foreign pages and, if a particular foreign 2159cgroup becomes the majority over a certain period of time, switches 2160the ownership of the inode to that cgroup. 2161 2162While this model is enough for most use cases where a given inode is 2163mostly dirtied by a single cgroup even when the main writing cgroup 2164changes over time, use cases where multiple cgroups write to a single 2165inode simultaneously are not supported well. In such circumstances, a 2166significant portion of IOs are likely to be attributed incorrectly. 2167As memory controller assigns page ownership on the first use and 2168doesn't update it until the page is released, even if writeback 2169strictly follows page ownership, multiple cgroups dirtying overlapping 2170areas wouldn't work as expected. It's recommended to avoid such usage 2171patterns. 2172 2173The sysctl knobs which affect writeback behavior are applied to cgroup 2174writeback as follows. 2175 2176 vm.dirty_background_ratio, vm.dirty_ratio 2177 These ratios apply the same to cgroup writeback with the 2178 amount of available memory capped by limits imposed by the 2179 memory controller and system-wide clean memory. 2180 2181 vm.dirty_background_bytes, vm.dirty_bytes 2182 For cgroup writeback, this is calculated into ratio against 2183 total available memory and applied the same way as 2184 vm.dirty[_background]_ratio. 2185 2186 2187IO Latency 2188~~~~~~~~~~ 2189 2190This is a cgroup v2 controller for IO workload protection. You provide a group 2191with a latency target, and if the average latency exceeds that target the 2192controller will throttle any peers that have a lower latency target than the 2193protected workload. 2194 2195The limits are only applied at the peer level in the hierarchy. This means that 2196in the diagram below, only groups A, B, and C will influence each other, and 2197groups D and F will influence each other. Group G will influence nobody:: 2198 2199 [root] 2200 / | \ 2201 A B C 2202 / \ | 2203 D F G 2204 2205 2206So the ideal way to configure this is to set io.latency in groups A, B, and C. 2207Generally you do not want to set a value lower than the latency your device 2208supports. Experiment to find the value that works best for your workload. 2209Start at higher than the expected latency for your device and watch the 2210avg_lat value in io.stat for your workload group to get an idea of the 2211latency you see during normal operation. Use the avg_lat value as a basis for 2212your real setting, setting at 10-15% higher than the value in io.stat. 2213 2214How IO Latency Throttling Works 2215~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~ 2216 2217io.latency is work conserving; so as long as everybody is meeting their latency 2218target the controller doesn't do anything. Once a group starts missing its 2219target it begins throttling any peer group that has a higher target than itself. 2220This throttling takes 2 forms: 2221 2222- Queue depth throttling. This is the number of outstanding IO's a group is 2223 allowed to have. We will clamp down relatively quickly, starting at no limit 2224 and going all the way down to 1 IO at a time. 2225 2226- Artificial delay induction. There are certain types of IO that cannot be 2227 throttled without possibly adversely affecting higher priority groups. This 2228 includes swapping and metadata IO. These types of IO are allowed to occur 2229 normally, however they are "charged" to the originating group. If the 2230 originating group is being throttled you will see the use_delay and delay 2231 fields in io.stat increase. The delay value is how many microseconds that are 2232 being added to any process that runs in this group. Because this number can 2233 grow quite large if there is a lot of swapping or metadata IO occurring we 2234 limit the individual delay events to 1 second at a time. 2235 2236Once the victimized group starts meeting its latency target again it will start 2237unthrottling any peer groups that were throttled previously. If the victimized 2238group simply stops doing IO the global counter will unthrottle appropriately. 2239 2240IO Latency Interface Files 2241~~~~~~~~~~~~~~~~~~~~~~~~~~ 2242 2243 io.latency 2244 This takes a similar format as the other controllers. 2245 2246 "MAJOR:MINOR target=<target time in microseconds>" 2247 2248 io.stat 2249 If the controller is enabled you will see extra stats in io.stat in 2250 addition to the normal ones. 2251 2252 depth 2253 This is the current queue depth for the group. 2254 2255 avg_lat 2256 This is an exponential moving average with a decay rate of 1/exp 2257 bound by the sampling interval. The decay rate interval can be 2258 calculated by multiplying the win value in io.stat by the 2259 corresponding number of samples based on the win value. 2260 2261 win 2262 The sampling window size in milliseconds. This is the minimum 2263 duration of time between evaluation events. Windows only elapse 2264 with IO activity. Idle periods extend the most recent window. 2265 2266IO Priority 2267~~~~~~~~~~~ 2268 2269A single attribute controls the behavior of the I/O priority cgroup policy, 2270namely the io.prio.class attribute. The following values are accepted for 2271that attribute: 2272 2273 no-change 2274 Do not modify the I/O priority class. 2275 2276 promote-to-rt 2277 For requests that have a non-RT I/O priority class, change it into RT. 2278 Also change the priority level of these requests to 4. Do not modify 2279 the I/O priority of requests that have priority class RT. 2280 2281 restrict-to-be 2282 For requests that do not have an I/O priority class or that have I/O 2283 priority class RT, change it into BE. Also change the priority level 2284 of these requests to 0. Do not modify the I/O priority class of 2285 requests that have priority class IDLE. 2286 2287 idle 2288 Change the I/O priority class of all requests into IDLE, the lowest 2289 I/O priority class. 2290 2291 none-to-rt 2292 Deprecated. Just an alias for promote-to-rt. 2293 2294The following numerical values are associated with the I/O priority policies: 2295 2296+----------------+---+ 2297| no-change | 0 | 2298+----------------+---+ 2299| promote-to-rt | 1 | 2300+----------------+---+ 2301| restrict-to-be | 2 | 2302+----------------+---+ 2303| idle | 3 | 2304+----------------+---+ 2305 2306The numerical value that corresponds to each I/O priority class is as follows: 2307 2308+-------------------------------+---+ 2309| IOPRIO_CLASS_NONE | 0 | 2310+-------------------------------+---+ 2311| IOPRIO_CLASS_RT (real-time) | 1 | 2312+-------------------------------+---+ 2313| IOPRIO_CLASS_BE (best effort) | 2 | 2314+-------------------------------+---+ 2315| IOPRIO_CLASS_IDLE | 3 | 2316+-------------------------------+---+ 2317 2318The algorithm to set the I/O priority class for a request is as follows: 2319 2320- If I/O priority class policy is promote-to-rt, change the request I/O 2321 priority class to IOPRIO_CLASS_RT and change the request I/O priority 2322 level to 4. 2323- If I/O priority class policy is not promote-to-rt, translate the I/O priority 2324 class policy into a number, then change the request I/O priority class 2325 into the maximum of the I/O priority class policy number and the numerical 2326 I/O priority class. 2327 2328PID 2329--- 2330 2331The process number controller is used to allow a cgroup to stop any 2332new tasks from being fork()'d or clone()'d after a specified limit is 2333reached. 2334 2335The number of tasks in a cgroup can be exhausted in ways which other 2336controllers cannot prevent, thus warranting its own controller. For 2337example, a fork bomb is likely to exhaust the number of tasks before 2338hitting memory restrictions. 2339 2340Note that PIDs used in this controller refer to TIDs, process IDs as 2341used by the kernel. 2342 2343 2344PID Interface Files 2345~~~~~~~~~~~~~~~~~~~ 2346 2347 pids.max 2348 A read-write single value file which exists on non-root 2349 cgroups. The default is "max". 2350 2351 Hard limit of number of processes. 2352 2353 pids.current 2354 A read-only single value file which exists on non-root cgroups. 2355 2356 The number of processes currently in the cgroup and its 2357 descendants. 2358 2359 pids.peak 2360 A read-only single value file which exists on non-root cgroups. 2361 2362 The maximum value that the number of processes in the cgroup and its 2363 descendants has ever reached. 2364 2365 pids.events 2366 A read-only flat-keyed file which exists on non-root cgroups. Unless 2367 specified otherwise, a value change in this file generates a file 2368 modified event. The following entries are defined. 2369 2370 max 2371 The number of times the cgroup's total number of processes hit the pids.max 2372 limit (see also pids_localevents). 2373 2374 pids.events.local 2375 Similar to pids.events but the fields in the file are local 2376 to the cgroup i.e. not hierarchical. The file modified event 2377 generated on this file reflects only the local events. 2378 2379Organisational operations are not blocked by cgroup policies, so it is 2380possible to have pids.current > pids.max. This can be done by either 2381setting the limit to be smaller than pids.current, or attaching enough 2382processes to the cgroup such that pids.current is larger than 2383pids.max. However, it is not possible to violate a cgroup PID policy 2384through fork() or clone(). These will return -EAGAIN if the creation 2385of a new process would cause a cgroup policy to be violated. 2386 2387 2388Cpuset 2389------ 2390 2391The "cpuset" controller provides a mechanism for constraining 2392the CPU and memory node placement of tasks to only the resources 2393specified in the cpuset interface files in a task's current cgroup. 2394This is especially valuable on large NUMA systems where placing jobs 2395on properly sized subsets of the systems with careful processor and 2396memory placement to reduce cross-node memory access and contention 2397can improve overall system performance. 2398 2399The "cpuset" controller is hierarchical. That means the controller 2400cannot use CPUs or memory nodes not allowed in its parent. 2401 2402 2403Cpuset Interface Files 2404~~~~~~~~~~~~~~~~~~~~~~ 2405 2406 cpuset.cpus 2407 A read-write multiple values file which exists on non-root 2408 cpuset-enabled cgroups. 2409 2410 It lists the requested CPUs to be used by tasks within this 2411 cgroup. The actual list of CPUs to be granted, however, is 2412 subjected to constraints imposed by its parent and can differ 2413 from the requested CPUs. 2414 2415 The CPU numbers are comma-separated numbers or ranges. 2416 For example:: 2417 2418 # cat cpuset.cpus 2419 0-4,6,8-10 2420 2421 An empty value indicates that the cgroup is using the same 2422 setting as the nearest cgroup ancestor with a non-empty 2423 "cpuset.cpus" or all the available CPUs if none is found. 2424 2425 The value of "cpuset.cpus" stays constant until the next update 2426 and won't be affected by any CPU hotplug events. 2427 2428 cpuset.cpus.effective 2429 A read-only multiple values file which exists on all 2430 cpuset-enabled cgroups. 2431 2432 It lists the onlined CPUs that are actually granted to this 2433 cgroup by its parent. These CPUs are allowed to be used by 2434 tasks within the current cgroup. 2435 2436 If "cpuset.cpus" is empty, the "cpuset.cpus.effective" file shows 2437 all the CPUs from the parent cgroup that can be available to 2438 be used by this cgroup. Otherwise, it should be a subset of 2439 "cpuset.cpus" unless none of the CPUs listed in "cpuset.cpus" 2440 can be granted. In this case, it will be treated just like an 2441 empty "cpuset.cpus". 2442 2443 Its value will be affected by CPU hotplug events. 2444 2445 cpuset.mems 2446 A read-write multiple values file which exists on non-root 2447 cpuset-enabled cgroups. 2448 2449 It lists the requested memory nodes to be used by tasks within 2450 this cgroup. The actual list of memory nodes granted, however, 2451 is subjected to constraints imposed by its parent and can differ 2452 from the requested memory nodes. 2453 2454 The memory node numbers are comma-separated numbers or ranges. 2455 For example:: 2456 2457 # cat cpuset.mems 2458 0-1,3 2459 2460 An empty value indicates that the cgroup is using the same 2461 setting as the nearest cgroup ancestor with a non-empty 2462 "cpuset.mems" or all the available memory nodes if none 2463 is found. 2464 2465 The value of "cpuset.mems" stays constant until the next update 2466 and won't be affected by any memory nodes hotplug events. 2467 2468 Setting a non-empty value to "cpuset.mems" causes memory of 2469 tasks within the cgroup to be migrated to the designated nodes if 2470 they are currently using memory outside of the designated nodes. 2471 2472 There is a cost for this memory migration. The migration 2473 may not be complete and some memory pages may be left behind. 2474 So it is recommended that "cpuset.mems" should be set properly 2475 before spawning new tasks into the cpuset. Even if there is 2476 a need to change "cpuset.mems" with active tasks, it shouldn't 2477 be done frequently. 2478 2479 cpuset.mems.effective 2480 A read-only multiple values file which exists on all 2481 cpuset-enabled cgroups. 2482 2483 It lists the onlined memory nodes that are actually granted to 2484 this cgroup by its parent. These memory nodes are allowed to 2485 be used by tasks within the current cgroup. 2486 2487 If "cpuset.mems" is empty, it shows all the memory nodes from the 2488 parent cgroup that will be available to be used by this cgroup. 2489 Otherwise, it should be a subset of "cpuset.mems" unless none of 2490 the memory nodes listed in "cpuset.mems" can be granted. In this 2491 case, it will be treated just like an empty "cpuset.mems". 2492 2493 Its value will be affected by memory nodes hotplug events. 2494 2495 cpuset.cpus.exclusive 2496 A read-write multiple values file which exists on non-root 2497 cpuset-enabled cgroups. 2498 2499 It lists all the exclusive CPUs that are allowed to be used 2500 to create a new cpuset partition. Its value is not used 2501 unless the cgroup becomes a valid partition root. See the 2502 "cpuset.cpus.partition" section below for a description of what 2503 a cpuset partition is. 2504 2505 When the cgroup becomes a partition root, the actual exclusive 2506 CPUs that are allocated to that partition are listed in 2507 "cpuset.cpus.exclusive.effective" which may be different 2508 from "cpuset.cpus.exclusive". If "cpuset.cpus.exclusive" 2509 has previously been set, "cpuset.cpus.exclusive.effective" 2510 is always a subset of it. 2511 2512 Users can manually set it to a value that is different from 2513 "cpuset.cpus". One constraint in setting it is that the list of 2514 CPUs must be exclusive with respect to "cpuset.cpus.exclusive" 2515 of its sibling. If "cpuset.cpus.exclusive" of a sibling cgroup 2516 isn't set, its "cpuset.cpus" value, if set, cannot be a subset 2517 of it to leave at least one CPU available when the exclusive 2518 CPUs are taken away. 2519 2520 For a parent cgroup, any one of its exclusive CPUs can only 2521 be distributed to at most one of its child cgroups. Having an 2522 exclusive CPU appearing in two or more of its child cgroups is 2523 not allowed (the exclusivity rule). A value that violates the 2524 exclusivity rule will be rejected with a write error. 2525 2526 The root cgroup is a partition root and all its available CPUs 2527 are in its exclusive CPU set. 2528 2529 cpuset.cpus.exclusive.effective 2530 A read-only multiple values file which exists on all non-root 2531 cpuset-enabled cgroups. 2532 2533 This file shows the effective set of exclusive CPUs that 2534 can be used to create a partition root. The content 2535 of this file will always be a subset of its parent's 2536 "cpuset.cpus.exclusive.effective" if its parent is not the root 2537 cgroup. It will also be a subset of "cpuset.cpus.exclusive" 2538 if it is set. If "cpuset.cpus.exclusive" is not set, it is 2539 treated to have an implicit value of "cpuset.cpus" in the 2540 formation of local partition. 2541 2542 cpuset.cpus.isolated 2543 A read-only and root cgroup only multiple values file. 2544 2545 This file shows the set of all isolated CPUs used in existing 2546 isolated partitions. It will be empty if no isolated partition 2547 is created. 2548 2549 cpuset.cpus.partition 2550 A read-write single value file which exists on non-root 2551 cpuset-enabled cgroups. This flag is owned by the parent cgroup 2552 and is not delegatable. 2553 2554 It accepts only the following input values when written to. 2555 2556 ========== ===================================== 2557 "member" Non-root member of a partition 2558 "root" Partition root 2559 "isolated" Partition root without load balancing 2560 ========== ===================================== 2561 2562 A cpuset partition is a collection of cpuset-enabled cgroups with 2563 a partition root at the top of the hierarchy and its descendants 2564 except those that are separate partition roots themselves and 2565 their descendants. A partition has exclusive access to the 2566 set of exclusive CPUs allocated to it. Other cgroups outside 2567 of that partition cannot use any CPUs in that set. 2568 2569 There are two types of partitions - local and remote. A local 2570 partition is one whose parent cgroup is also a valid partition 2571 root. A remote partition is one whose parent cgroup is not a 2572 valid partition root itself. Writing to "cpuset.cpus.exclusive" 2573 is optional for the creation of a local partition as its 2574 "cpuset.cpus.exclusive" file will assume an implicit value that 2575 is the same as "cpuset.cpus" if it is not set. Writing the 2576 proper "cpuset.cpus.exclusive" values down the cgroup hierarchy 2577 before the target partition root is mandatory for the creation 2578 of a remote partition. 2579 2580 Currently, a remote partition cannot be created under a local 2581 partition. All the ancestors of a remote partition root except 2582 the root cgroup cannot be a partition root. 2583 2584 The root cgroup is always a partition root and its state cannot 2585 be changed. All other non-root cgroups start out as "member". 2586 2587 When set to "root", the current cgroup is the root of a new 2588 partition or scheduling domain. The set of exclusive CPUs is 2589 determined by the value of its "cpuset.cpus.exclusive.effective". 2590 2591 When set to "isolated", the CPUs in that partition will be in 2592 an isolated state without any load balancing from the scheduler 2593 and excluded from the unbound workqueues. Tasks placed in such 2594 a partition with multiple CPUs should be carefully distributed 2595 and bound to each of the individual CPUs for optimal performance. 2596 2597 A partition root ("root" or "isolated") can be in one of the 2598 two possible states - valid or invalid. An invalid partition 2599 root is in a degraded state where some state information may 2600 be retained, but behaves more like a "member". 2601 2602 All possible state transitions among "member", "root" and 2603 "isolated" are allowed. 2604 2605 On read, the "cpuset.cpus.partition" file can show the following 2606 values. 2607 2608 ============================= ===================================== 2609 "member" Non-root member of a partition 2610 "root" Partition root 2611 "isolated" Partition root without load balancing 2612 "root invalid (<reason>)" Invalid partition root 2613 "isolated invalid (<reason>)" Invalid isolated partition root 2614 ============================= ===================================== 2615 2616 In the case of an invalid partition root, a descriptive string on 2617 why the partition is invalid is included within parentheses. 2618 2619 For a local partition root to be valid, the following conditions 2620 must be met. 2621 2622 1) The parent cgroup is a valid partition root. 2623 2) The "cpuset.cpus.exclusive.effective" file cannot be empty, 2624 though it may contain offline CPUs. 2625 3) The "cpuset.cpus.effective" cannot be empty unless there is 2626 no task associated with this partition. 2627 2628 For a remote partition root to be valid, all the above conditions 2629 except the first one must be met. 2630 2631 External events like hotplug or changes to "cpuset.cpus" or 2632 "cpuset.cpus.exclusive" can cause a valid partition root to 2633 become invalid and vice versa. Note that a task cannot be 2634 moved to a cgroup with empty "cpuset.cpus.effective". 2635 2636 A valid non-root parent partition may distribute out all its CPUs 2637 to its child local partitions when there is no task associated 2638 with it. 2639 2640 Care must be taken to change a valid partition root to "member" 2641 as all its child local partitions, if present, will become 2642 invalid causing disruption to tasks running in those child 2643 partitions. These inactivated partitions could be recovered if 2644 their parent is switched back to a partition root with a proper 2645 value in "cpuset.cpus" or "cpuset.cpus.exclusive". 2646 2647 Poll and inotify events are triggered whenever the state of 2648 "cpuset.cpus.partition" changes. That includes changes caused 2649 by write to "cpuset.cpus.partition", cpu hotplug or other 2650 changes that modify the validity status of the partition. 2651 This will allow user space agents to monitor unexpected changes 2652 to "cpuset.cpus.partition" without the need to do continuous 2653 polling. 2654 2655 A user can pre-configure certain CPUs to an isolated state 2656 with load balancing disabled at boot time with the "isolcpus" 2657 kernel boot command line option. If those CPUs are to be put 2658 into a partition, they have to be used in an isolated partition. 2659 2660 2661Device controller 2662----------------- 2663 2664Device controller manages access to device files. It includes both 2665creation of new device files (using mknod), and access to the 2666existing device files. 2667 2668Cgroup v2 device controller has no interface files and is implemented 2669on top of cgroup BPF. To control access to device files, a user may 2670create bpf programs of type BPF_PROG_TYPE_CGROUP_DEVICE and attach 2671them to cgroups with BPF_CGROUP_DEVICE flag. On an attempt to access a 2672device file, corresponding BPF programs will be executed, and depending 2673on the return value the attempt will succeed or fail with -EPERM. 2674 2675A BPF_PROG_TYPE_CGROUP_DEVICE program takes a pointer to the 2676bpf_cgroup_dev_ctx structure, which describes the device access attempt: 2677access type (mknod/read/write) and device (type, major and minor numbers). 2678If the program returns 0, the attempt fails with -EPERM, otherwise it 2679succeeds. 2680 2681An example of BPF_PROG_TYPE_CGROUP_DEVICE program may be found in 2682tools/testing/selftests/bpf/progs/dev_cgroup.c in the kernel source tree. 2683 2684 2685RDMA 2686---- 2687 2688The "rdma" controller regulates the distribution and accounting of 2689RDMA resources. 2690 2691RDMA Interface Files 2692~~~~~~~~~~~~~~~~~~~~ 2693 2694 rdma.max 2695 A readwrite nested-keyed file that exists for all the cgroups 2696 except root that describes current configured resource limit 2697 for a RDMA/IB device. 2698 2699 Lines are keyed by device name and are not ordered. 2700 Each line contains space separated resource name and its configured 2701 limit that can be distributed. 2702 2703 The following nested keys are defined. 2704 2705 ========== ============================= 2706 hca_handle Maximum number of HCA Handles 2707 hca_object Maximum number of HCA Objects 2708 ========== ============================= 2709 2710 An example for mlx4 and ocrdma device follows:: 2711 2712 mlx4_0 hca_handle=2 hca_object=2000 2713 ocrdma1 hca_handle=3 hca_object=max 2714 2715 rdma.current 2716 A read-only file that describes current resource usage. 2717 It exists for all the cgroup except root. 2718 2719 An example for mlx4 and ocrdma device follows:: 2720 2721 mlx4_0 hca_handle=1 hca_object=20 2722 ocrdma1 hca_handle=1 hca_object=23 2723 2724DMEM 2725---- 2726 2727The "dmem" controller regulates the distribution and accounting of 2728device memory regions. Because each memory region may have its own page size, 2729which does not have to be equal to the system page size, the units are always bytes. 2730 2731DMEM Interface Files 2732~~~~~~~~~~~~~~~~~~~~ 2733 2734 dmem.max, dmem.min, dmem.low 2735 A readwrite nested-keyed file that exists for all the cgroups 2736 except root that describes current configured resource limit 2737 for a region. 2738 2739 An example for xe follows:: 2740 2741 drm/0000:03:00.0/vram0 1073741824 2742 drm/0000:03:00.0/stolen max 2743 2744 The semantics are the same as for the memory cgroup controller, and are 2745 calculated in the same way. 2746 2747 dmem.capacity 2748 A read-only file that describes maximum region capacity. 2749 It only exists on the root cgroup. Not all memory can be 2750 allocated by cgroups, as the kernel reserves some for 2751 internal use. 2752 2753 An example for xe follows:: 2754 2755 drm/0000:03:00.0/vram0 8514437120 2756 drm/0000:03:00.0/stolen 67108864 2757 2758 dmem.current 2759 A read-only file that describes current resource usage. 2760 It exists for all the cgroup except root. 2761 2762 An example for xe follows:: 2763 2764 drm/0000:03:00.0/vram0 12550144 2765 drm/0000:03:00.0/stolen 8650752 2766 2767HugeTLB 2768------- 2769 2770The HugeTLB controller allows to limit the HugeTLB usage per control group and 2771enforces the controller limit during page fault. 2772 2773HugeTLB Interface Files 2774~~~~~~~~~~~~~~~~~~~~~~~ 2775 2776 hugetlb.<hugepagesize>.current 2777 Show current usage for "hugepagesize" hugetlb. It exists for all 2778 the cgroup except root. 2779 2780 hugetlb.<hugepagesize>.max 2781 Set/show the hard limit of "hugepagesize" hugetlb usage. 2782 The default value is "max". It exists for all the cgroup except root. 2783 2784 hugetlb.<hugepagesize>.events 2785 A read-only flat-keyed file which exists on non-root cgroups. 2786 2787 max 2788 The number of allocation failure due to HugeTLB limit 2789 2790 hugetlb.<hugepagesize>.events.local 2791 Similar to hugetlb.<hugepagesize>.events but the fields in the file 2792 are local to the cgroup i.e. not hierarchical. The file modified event 2793 generated on this file reflects only the local events. 2794 2795 hugetlb.<hugepagesize>.numa_stat 2796 Similar to memory.numa_stat, it shows the numa information of the 2797 hugetlb pages of <hugepagesize> in this cgroup. Only active in 2798 use hugetlb pages are included. The per-node values are in bytes. 2799 2800Misc 2801---- 2802 2803The Miscellaneous cgroup provides the resource limiting and tracking 2804mechanism for the scalar resources which cannot be abstracted like the other 2805cgroup resources. Controller is enabled by the CONFIG_CGROUP_MISC config 2806option. 2807 2808A resource can be added to the controller via enum misc_res_type{} in the 2809include/linux/misc_cgroup.h file and the corresponding name via misc_res_name[] 2810in the kernel/cgroup/misc.c file. Provider of the resource must set its 2811capacity prior to using the resource by calling misc_cg_set_capacity(). 2812 2813Once a capacity is set then the resource usage can be updated using charge and 2814uncharge APIs. All of the APIs to interact with misc controller are in 2815include/linux/misc_cgroup.h. 2816 2817Misc Interface Files 2818~~~~~~~~~~~~~~~~~~~~ 2819 2820Miscellaneous controller provides 3 interface files. If two misc resources (res_a and res_b) are registered then: 2821 2822 misc.capacity 2823 A read-only flat-keyed file shown only in the root cgroup. It shows 2824 miscellaneous scalar resources available on the platform along with 2825 their quantities:: 2826 2827 $ cat misc.capacity 2828 res_a 50 2829 res_b 10 2830 2831 misc.current 2832 A read-only flat-keyed file shown in the all cgroups. It shows 2833 the current usage of the resources in the cgroup and its children.:: 2834 2835 $ cat misc.current 2836 res_a 3 2837 res_b 0 2838 2839 misc.peak 2840 A read-only flat-keyed file shown in all cgroups. It shows the 2841 historical maximum usage of the resources in the cgroup and its 2842 children.:: 2843 2844 $ cat misc.peak 2845 res_a 10 2846 res_b 8 2847 2848 misc.max 2849 A read-write flat-keyed file shown in the non root cgroups. Allowed 2850 maximum usage of the resources in the cgroup and its children.:: 2851 2852 $ cat misc.max 2853 res_a max 2854 res_b 4 2855 2856 Limit can be set by:: 2857 2858 # echo res_a 1 > misc.max 2859 2860 Limit can be set to max by:: 2861 2862 # echo res_a max > misc.max 2863 2864 Limits can be set higher than the capacity value in the misc.capacity 2865 file. 2866 2867 misc.events 2868 A read-only flat-keyed file which exists on non-root cgroups. The 2869 following entries are defined. Unless specified otherwise, a value 2870 change in this file generates a file modified event. All fields in 2871 this file are hierarchical. 2872 2873 max 2874 The number of times the cgroup's resource usage was 2875 about to go over the max boundary. 2876 2877 misc.events.local 2878 Similar to misc.events but the fields in the file are local to the 2879 cgroup i.e. not hierarchical. The file modified event generated on 2880 this file reflects only the local events. 2881 2882Migration and Ownership 2883~~~~~~~~~~~~~~~~~~~~~~~ 2884 2885A miscellaneous scalar resource is charged to the cgroup in which it is used 2886first, and stays charged to that cgroup until that resource is freed. Migrating 2887a process to a different cgroup does not move the charge to the destination 2888cgroup where the process has moved. 2889 2890Others 2891------ 2892 2893perf_event 2894~~~~~~~~~~ 2895 2896perf_event controller, if not mounted on a legacy hierarchy, is 2897automatically enabled on the v2 hierarchy so that perf events can 2898always be filtered by cgroup v2 path. The controller can still be 2899moved to a legacy hierarchy after v2 hierarchy is populated. 2900 2901 2902Non-normative information 2903------------------------- 2904 2905This section contains information that isn't considered to be a part of 2906the stable kernel API and so is subject to change. 2907 2908 2909CPU controller root cgroup process behaviour 2910~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~ 2911 2912When distributing CPU cycles in the root cgroup each thread in this 2913cgroup is treated as if it was hosted in a separate child cgroup of the 2914root cgroup. This child cgroup weight is dependent on its thread nice 2915level. 2916 2917For details of this mapping see sched_prio_to_weight array in 2918kernel/sched/core.c file (values from this array should be scaled 2919appropriately so the neutral - nice 0 - value is 100 instead of 1024). 2920 2921 2922IO controller root cgroup process behaviour 2923~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~ 2924 2925Root cgroup processes are hosted in an implicit leaf child node. 2926When distributing IO resources this implicit child node is taken into 2927account as if it was a normal child cgroup of the root cgroup with a 2928weight value of 200. 2929 2930 2931Namespace 2932========= 2933 2934Basics 2935------ 2936 2937cgroup namespace provides a mechanism to virtualize the view of the 2938"/proc/$PID/cgroup" file and cgroup mounts. The CLONE_NEWCGROUP clone 2939flag can be used with clone(2) and unshare(2) to create a new cgroup 2940namespace. The process running inside the cgroup namespace will have 2941its "/proc/$PID/cgroup" output restricted to cgroupns root. The 2942cgroupns root is the cgroup of the process at the time of creation of 2943the cgroup namespace. 2944 2945Without cgroup namespace, the "/proc/$PID/cgroup" file shows the 2946complete path of the cgroup of a process. In a container setup where 2947a set of cgroups and namespaces are intended to isolate processes the 2948"/proc/$PID/cgroup" file may leak potential system level information 2949to the isolated processes. For example:: 2950 2951 # cat /proc/self/cgroup 2952 0::/batchjobs/container_id1 2953 2954The path '/batchjobs/container_id1' can be considered as system-data 2955and undesirable to expose to the isolated processes. cgroup namespace 2956can be used to restrict visibility of this path. For example, before 2957creating a cgroup namespace, one would see:: 2958 2959 # ls -l /proc/self/ns/cgroup 2960 lrwxrwxrwx 1 root root 0 2014-07-15 10:37 /proc/self/ns/cgroup -> cgroup:[4026531835] 2961 # cat /proc/self/cgroup 2962 0::/batchjobs/container_id1 2963 2964After unsharing a new namespace, the view changes:: 2965 2966 # ls -l /proc/self/ns/cgroup 2967 lrwxrwxrwx 1 root root 0 2014-07-15 10:35 /proc/self/ns/cgroup -> cgroup:[4026532183] 2968 # cat /proc/self/cgroup 2969 0::/ 2970 2971When some thread from a multi-threaded process unshares its cgroup 2972namespace, the new cgroupns gets applied to the entire process (all 2973the threads). This is natural for the v2 hierarchy; however, for the 2974legacy hierarchies, this may be unexpected. 2975 2976A cgroup namespace is alive as long as there are processes inside or 2977mounts pinning it. When the last usage goes away, the cgroup 2978namespace is destroyed. The cgroupns root and the actual cgroups 2979remain. 2980 2981 2982The Root and Views 2983------------------ 2984 2985The 'cgroupns root' for a cgroup namespace is the cgroup in which the 2986process calling unshare(2) is running. For example, if a process in 2987/batchjobs/container_id1 cgroup calls unshare, cgroup 2988/batchjobs/container_id1 becomes the cgroupns root. For the 2989init_cgroup_ns, this is the real root ('/') cgroup. 2990 2991The cgroupns root cgroup does not change even if the namespace creator 2992process later moves to a different cgroup:: 2993 2994 # ~/unshare -c # unshare cgroupns in some cgroup 2995 # cat /proc/self/cgroup 2996 0::/ 2997 # mkdir sub_cgrp_1 2998 # echo 0 > sub_cgrp_1/cgroup.procs 2999 # cat /proc/self/cgroup 3000 0::/sub_cgrp_1 3001 3002Each process gets its namespace-specific view of "/proc/$PID/cgroup" 3003 3004Processes running inside the cgroup namespace will be able to see 3005cgroup paths (in /proc/self/cgroup) only inside their root cgroup. 3006From within an unshared cgroupns:: 3007 3008 # sleep 100000 & 3009 [1] 7353 3010 # echo 7353 > sub_cgrp_1/cgroup.procs 3011 # cat /proc/7353/cgroup 3012 0::/sub_cgrp_1 3013 3014From the initial cgroup namespace, the real cgroup path will be 3015visible:: 3016 3017 $ cat /proc/7353/cgroup 3018 0::/batchjobs/container_id1/sub_cgrp_1 3019 3020From a sibling cgroup namespace (that is, a namespace rooted at a 3021different cgroup), the cgroup path relative to its own cgroup 3022namespace root will be shown. For instance, if PID 7353's cgroup 3023namespace root is at '/batchjobs/container_id2', then it will see:: 3024 3025 # cat /proc/7353/cgroup 3026 0::/../container_id2/sub_cgrp_1 3027 3028Note that the relative path always starts with '/' to indicate that 3029its relative to the cgroup namespace root of the caller. 3030 3031 3032Migration and setns(2) 3033---------------------- 3034 3035Processes inside a cgroup namespace can move into and out of the 3036namespace root if they have proper access to external cgroups. For 3037example, from inside a namespace with cgroupns root at 3038/batchjobs/container_id1, and assuming that the global hierarchy is 3039still accessible inside cgroupns:: 3040 3041 # cat /proc/7353/cgroup 3042 0::/sub_cgrp_1 3043 # echo 7353 > batchjobs/container_id2/cgroup.procs 3044 # cat /proc/7353/cgroup 3045 0::/../container_id2 3046 3047Note that this kind of setup is not encouraged. A task inside cgroup 3048namespace should only be exposed to its own cgroupns hierarchy. 3049 3050setns(2) to another cgroup namespace is allowed when: 3051 3052(a) the process has CAP_SYS_ADMIN against its current user namespace 3053(b) the process has CAP_SYS_ADMIN against the target cgroup 3054 namespace's userns 3055 3056No implicit cgroup changes happen with attaching to another cgroup 3057namespace. It is expected that the someone moves the attaching 3058process under the target cgroup namespace root. 3059 3060 3061Interaction with Other Namespaces 3062--------------------------------- 3063 3064Namespace specific cgroup hierarchy can be mounted by a process 3065running inside a non-init cgroup namespace:: 3066 3067 # mount -t cgroup2 none $MOUNT_POINT 3068 3069This will mount the unified cgroup hierarchy with cgroupns root as the 3070filesystem root. The process needs CAP_SYS_ADMIN against its user and 3071mount namespaces. 3072 3073The virtualization of /proc/self/cgroup file combined with restricting 3074the view of cgroup hierarchy by namespace-private cgroupfs mount 3075provides a properly isolated cgroup view inside the container. 3076 3077 3078Information on Kernel Programming 3079================================= 3080 3081This section contains kernel programming information in the areas 3082where interacting with cgroup is necessary. cgroup core and 3083controllers are not covered. 3084 3085 3086Filesystem Support for Writeback 3087-------------------------------- 3088 3089A filesystem can support cgroup writeback by updating 3090address_space_operations->writepages() to annotate bio's using the 3091following two functions. 3092 3093 wbc_init_bio(@wbc, @bio) 3094 Should be called for each bio carrying writeback data and 3095 associates the bio with the inode's owner cgroup and the 3096 corresponding request queue. This must be called after 3097 a queue (device) has been associated with the bio and 3098 before submission. 3099 3100 wbc_account_cgroup_owner(@wbc, @folio, @bytes) 3101 Should be called for each data segment being written out. 3102 While this function doesn't care exactly when it's called 3103 during the writeback session, it's the easiest and most 3104 natural to call it as data segments are added to a bio. 3105 3106With writeback bio's annotated, cgroup support can be enabled per 3107super_block by setting SB_I_CGROUPWB in ->s_iflags. This allows for 3108selective disabling of cgroup writeback support which is helpful when 3109certain filesystem features, e.g. journaled data mode, are 3110incompatible. 3111 3112wbc_init_bio() binds the specified bio to its cgroup. Depending on 3113the configuration, the bio may be executed at a lower priority and if 3114the writeback session is holding shared resources, e.g. a journal 3115entry, may lead to priority inversion. There is no one easy solution 3116for the problem. Filesystems can try to work around specific problem 3117cases by skipping wbc_init_bio() and using bio_associate_blkg() 3118directly. 3119 3120 3121Deprecated v1 Core Features 3122=========================== 3123 3124- Multiple hierarchies including named ones are not supported. 3125 3126- All v1 mount options are not supported. 3127 3128- The "tasks" file is removed and "cgroup.procs" is not sorted. 3129 3130- "cgroup.clone_children" is removed. 3131 3132- /proc/cgroups is meaningless for v2. Use "cgroup.controllers" or 3133 "cgroup.stat" files at the root instead. 3134 3135 3136Issues with v1 and Rationales for v2 3137==================================== 3138 3139Multiple Hierarchies 3140-------------------- 3141 3142cgroup v1 allowed an arbitrary number of hierarchies and each 3143hierarchy could host any number of controllers. While this seemed to 3144provide a high level of flexibility, it wasn't useful in practice. 3145 3146For example, as there is only one instance of each controller, utility 3147type controllers such as freezer which can be useful in all 3148hierarchies could only be used in one. The issue is exacerbated by 3149the fact that controllers couldn't be moved to another hierarchy once 3150hierarchies were populated. Another issue was that all controllers 3151bound to a hierarchy were forced to have exactly the same view of the 3152hierarchy. It wasn't possible to vary the granularity depending on 3153the specific controller. 3154 3155In practice, these issues heavily limited which controllers could be 3156put on the same hierarchy and most configurations resorted to putting 3157each controller on its own hierarchy. Only closely related ones, such 3158as the cpu and cpuacct controllers, made sense to be put on the same 3159hierarchy. This often meant that userland ended up managing multiple 3160similar hierarchies repeating the same steps on each hierarchy 3161whenever a hierarchy management operation was necessary. 3162 3163Furthermore, support for multiple hierarchies came at a steep cost. 3164It greatly complicated cgroup core implementation but more importantly 3165the support for multiple hierarchies restricted how cgroup could be 3166used in general and what controllers was able to do. 3167 3168There was no limit on how many hierarchies there might be, which meant 3169that a thread's cgroup membership couldn't be described in finite 3170length. The key might contain any number of entries and was unlimited 3171in length, which made it highly awkward to manipulate and led to 3172addition of controllers which existed only to identify membership, 3173which in turn exacerbated the original problem of proliferating number 3174of hierarchies. 3175 3176Also, as a controller couldn't have any expectation regarding the 3177topologies of hierarchies other controllers might be on, each 3178controller had to assume that all other controllers were attached to 3179completely orthogonal hierarchies. This made it impossible, or at 3180least very cumbersome, for controllers to cooperate with each other. 3181 3182In most use cases, putting controllers on hierarchies which are 3183completely orthogonal to each other isn't necessary. What usually is 3184called for is the ability to have differing levels of granularity 3185depending on the specific controller. In other words, hierarchy may 3186be collapsed from leaf towards root when viewed from specific 3187controllers. For example, a given configuration might not care about 3188how memory is distributed beyond a certain level while still wanting 3189to control how CPU cycles are distributed. 3190 3191 3192Thread Granularity 3193------------------ 3194 3195cgroup v1 allowed threads of a process to belong to different cgroups. 3196This didn't make sense for some controllers and those controllers 3197ended up implementing different ways to ignore such situations but 3198much more importantly it blurred the line between API exposed to 3199individual applications and system management interface. 3200 3201Generally, in-process knowledge is available only to the process 3202itself; thus, unlike service-level organization of processes, 3203categorizing threads of a process requires active participation from 3204the application which owns the target process. 3205 3206cgroup v1 had an ambiguously defined delegation model which got abused 3207in combination with thread granularity. cgroups were delegated to 3208individual applications so that they can create and manage their own 3209sub-hierarchies and control resource distributions along them. This 3210effectively raised cgroup to the status of a syscall-like API exposed 3211to lay programs. 3212 3213First of all, cgroup has a fundamentally inadequate interface to be 3214exposed this way. For a process to access its own knobs, it has to 3215extract the path on the target hierarchy from /proc/self/cgroup, 3216construct the path by appending the name of the knob to the path, open 3217and then read and/or write to it. This is not only extremely clunky 3218and unusual but also inherently racy. There is no conventional way to 3219define transaction across the required steps and nothing can guarantee 3220that the process would actually be operating on its own sub-hierarchy. 3221 3222cgroup controllers implemented a number of knobs which would never be 3223accepted as public APIs because they were just adding control knobs to 3224system-management pseudo filesystem. cgroup ended up with interface 3225knobs which were not properly abstracted or refined and directly 3226revealed kernel internal details. These knobs got exposed to 3227individual applications through the ill-defined delegation mechanism 3228effectively abusing cgroup as a shortcut to implementing public APIs 3229without going through the required scrutiny. 3230 3231This was painful for both userland and kernel. Userland ended up with 3232misbehaving and poorly abstracted interfaces and kernel exposing and 3233locked into constructs inadvertently. 3234 3235 3236Competition Between Inner Nodes and Threads 3237------------------------------------------- 3238 3239cgroup v1 allowed threads to be in any cgroups which created an 3240interesting problem where threads belonging to a parent cgroup and its 3241children cgroups competed for resources. This was nasty as two 3242different types of entities competed and there was no obvious way to 3243settle it. Different controllers did different things. 3244 3245The cpu controller considered threads and cgroups as equivalents and 3246mapped nice levels to cgroup weights. This worked for some cases but 3247fell flat when children wanted to be allocated specific ratios of CPU 3248cycles and the number of internal threads fluctuated - the ratios 3249constantly changed as the number of competing entities fluctuated. 3250There also were other issues. The mapping from nice level to weight 3251wasn't obvious or universal, and there were various other knobs which 3252simply weren't available for threads. 3253 3254The io controller implicitly created a hidden leaf node for each 3255cgroup to host the threads. The hidden leaf had its own copies of all 3256the knobs with ``leaf_`` prefixed. While this allowed equivalent 3257control over internal threads, it was with serious drawbacks. It 3258always added an extra layer of nesting which wouldn't be necessary 3259otherwise, made the interface messy and significantly complicated the 3260implementation. 3261 3262The memory controller didn't have a way to control what happened 3263between internal tasks and child cgroups and the behavior was not 3264clearly defined. There were attempts to add ad-hoc behaviors and 3265knobs to tailor the behavior to specific workloads which would have 3266led to problems extremely difficult to resolve in the long term. 3267 3268Multiple controllers struggled with internal tasks and came up with 3269different ways to deal with it; unfortunately, all the approaches were 3270severely flawed and, furthermore, the widely different behaviors 3271made cgroup as a whole highly inconsistent. 3272 3273This clearly is a problem which needs to be addressed from cgroup core 3274in a uniform way. 3275 3276 3277Other Interface Issues 3278---------------------- 3279 3280cgroup v1 grew without oversight and developed a large number of 3281idiosyncrasies and inconsistencies. One issue on the cgroup core side 3282was how an empty cgroup was notified - a userland helper binary was 3283forked and executed for each event. The event delivery wasn't 3284recursive or delegatable. The limitations of the mechanism also led 3285to in-kernel event delivery filtering mechanism further complicating 3286the interface. 3287 3288Controller interfaces were problematic too. An extreme example is 3289controllers completely ignoring hierarchical organization and treating 3290all cgroups as if they were all located directly under the root 3291cgroup. Some controllers exposed a large amount of inconsistent 3292implementation details to userland. 3293 3294There also was no consistency across controllers. When a new cgroup 3295was created, some controllers defaulted to not imposing extra 3296restrictions while others disallowed any resource usage until 3297explicitly configured. Configuration knobs for the same type of 3298control used widely differing naming schemes and formats. Statistics 3299and information knobs were named arbitrarily and used different 3300formats and units even in the same controller. 3301 3302cgroup v2 establishes common conventions where appropriate and updates 3303controllers so that they expose minimal and consistent interfaces. 3304 3305 3306Controller Issues and Remedies 3307------------------------------ 3308 3309Memory 3310~~~~~~ 3311 3312The original lower boundary, the soft limit, is defined as a limit 3313that is per default unset. As a result, the set of cgroups that 3314global reclaim prefers is opt-in, rather than opt-out. The costs for 3315optimizing these mostly negative lookups are so high that the 3316implementation, despite its enormous size, does not even provide the 3317basic desirable behavior. First off, the soft limit has no 3318hierarchical meaning. All configured groups are organized in a global 3319rbtree and treated like equal peers, regardless where they are located 3320in the hierarchy. This makes subtree delegation impossible. Second, 3321the soft limit reclaim pass is so aggressive that it not just 3322introduces high allocation latencies into the system, but also impacts 3323system performance due to overreclaim, to the point where the feature 3324becomes self-defeating. 3325 3326The memory.low boundary on the other hand is a top-down allocated 3327reserve. A cgroup enjoys reclaim protection when it's within its 3328effective low, which makes delegation of subtrees possible. It also 3329enjoys having reclaim pressure proportional to its overage when 3330above its effective low. 3331 3332The original high boundary, the hard limit, is defined as a strict 3333limit that can not budge, even if the OOM killer has to be called. 3334But this generally goes against the goal of making the most out of the 3335available memory. The memory consumption of workloads varies during 3336runtime, and that requires users to overcommit. But doing that with a 3337strict upper limit requires either a fairly accurate prediction of the 3338working set size or adding slack to the limit. Since working set size 3339estimation is hard and error prone, and getting it wrong results in 3340OOM kills, most users tend to err on the side of a looser limit and 3341end up wasting precious resources. 3342 3343The memory.high boundary on the other hand can be set much more 3344conservatively. When hit, it throttles allocations by forcing them 3345into direct reclaim to work off the excess, but it never invokes the 3346OOM killer. As a result, a high boundary that is chosen too 3347aggressively will not terminate the processes, but instead it will 3348lead to gradual performance degradation. The user can monitor this 3349and make corrections until the minimal memory footprint that still 3350gives acceptable performance is found. 3351 3352In extreme cases, with many concurrent allocations and a complete 3353breakdown of reclaim progress within the group, the high boundary can 3354be exceeded. But even then it's mostly better to satisfy the 3355allocation from the slack available in other groups or the rest of the 3356system than killing the group. Otherwise, memory.max is there to 3357limit this type of spillover and ultimately contain buggy or even 3358malicious applications. 3359 3360Setting the original memory.limit_in_bytes below the current usage was 3361subject to a race condition, where concurrent charges could cause the 3362limit setting to fail. memory.max on the other hand will first set the 3363limit to prevent new charges, and then reclaim and OOM kill until the 3364new limit is met - or the task writing to memory.max is killed. 3365 3366The combined memory+swap accounting and limiting is replaced by real 3367control over swap space. 3368 3369The main argument for a combined memory+swap facility in the original 3370cgroup design was that global or parental pressure would always be 3371able to swap all anonymous memory of a child group, regardless of the 3372child's own (possibly untrusted) configuration. However, untrusted 3373groups can sabotage swapping by other means - such as referencing its 3374anonymous memory in a tight loop - and an admin can not assume full 3375swappability when overcommitting untrusted jobs. 3376 3377For trusted jobs, on the other hand, a combined counter is not an 3378intuitive userspace interface, and it flies in the face of the idea 3379that cgroup controllers should account and limit specific physical 3380resources. Swap space is a resource like all others in the system, 3381and that's why unified hierarchy allows distributing it separately. 3382