1========================== 2Memory Resource Controller 3========================== 4 5.. caution:: 6 This document is hopelessly outdated and it asks for a complete 7 rewrite. It still contains a useful information so we are keeping it 8 here but make sure to check the current code if you need a deeper 9 understanding. 10 11.. note:: 12 The Memory Resource Controller has generically been referred to as the 13 memory controller in this document. Do not confuse memory controller 14 used here with the memory controller that is used in hardware. 15 16.. hint:: 17 When we mention a cgroup (cgroupfs's directory) with memory controller, 18 we call it "memory cgroup". When you see git-log and source code, you'll 19 see patch's title and function names tend to use "memcg". 20 In this document, we avoid using it. 21 22Benefits and Purpose of the memory controller 23============================================= 24 25The memory controller isolates the memory behaviour of a group of tasks 26from the rest of the system. The article on LWN [12]_ mentions some probable 27uses of the memory controller. The memory controller can be used to 28 29a. Isolate an application or a group of applications 30 Memory-hungry applications can be isolated and limited to a smaller 31 amount of memory. 32b. Create a cgroup with a limited amount of memory; this can be used 33 as a good alternative to booting with mem=XXXX. 34c. Virtualization solutions can control the amount of memory they want 35 to assign to a virtual machine instance. 36d. A CD/DVD burner could control the amount of memory used by the 37 rest of the system to ensure that burning does not fail due to lack 38 of available memory. 39e. There are several other use cases; find one or use the controller just 40 for fun (to learn and hack on the VM subsystem). 41 42Current Status: linux-2.6.34-mmotm(development version of 2010/April) 43 44Features: 45 46 - accounting anonymous pages, file caches, swap caches usage and limiting them. 47 - pages are linked to per-memcg LRU exclusively, and there is no global LRU. 48 - optionally, memory+swap usage can be accounted and limited. 49 - hierarchical accounting 50 - soft limit 51 - moving (recharging) account at moving a task is selectable. 52 - usage threshold notifier 53 - memory pressure notifier 54 - oom-killer disable knob and oom-notifier 55 - Root cgroup has no limit controls. 56 57 Kernel memory support is a work in progress, and the current version provides 58 basically functionality. (See :ref:`section 2.7 59 <cgroup-v1-memory-kernel-extension>`) 60 61Brief summary of control files. 62 63==================================== ========================================== 64 tasks attach a task(thread) and show list of 65 threads 66 cgroup.procs show list of processes 67 cgroup.event_control an interface for event_fd() 68 This knob is not available on CONFIG_PREEMPT_RT systems. 69 memory.usage_in_bytes show current usage for memory 70 (See 5.5 for details) 71 memory.memsw.usage_in_bytes show current usage for memory+Swap 72 (See 5.5 for details) 73 memory.limit_in_bytes set/show limit of memory usage 74 memory.memsw.limit_in_bytes set/show limit of memory+Swap usage 75 memory.failcnt show the number of memory usage hits limits 76 memory.memsw.failcnt show the number of memory+Swap hits limits 77 memory.max_usage_in_bytes show max memory usage recorded 78 memory.memsw.max_usage_in_bytes show max memory+Swap usage recorded 79 memory.soft_limit_in_bytes set/show soft limit of memory usage 80 This knob is not available on CONFIG_PREEMPT_RT systems. 81 This knob is deprecated and shouldn't be 82 used. 83 memory.stat show various statistics 84 memory.use_hierarchy set/show hierarchical account enabled 85 This knob is deprecated and shouldn't be 86 used. 87 memory.force_empty trigger forced page reclaim 88 memory.pressure_level set memory pressure notifications 89 This knob is deprecated and shouldn't be 90 used. 91 memory.swappiness set/show swappiness parameter of vmscan 92 (See sysctl's vm.swappiness) 93 memory.move_charge_at_immigrate set/show controls of moving charges 94 This knob is deprecated and shouldn't be 95 used. 96 memory.oom_control set/show oom controls. 97 This knob is deprecated and shouldn't be 98 used. 99 memory.numa_stat show the number of memory usage per numa 100 node 101 memory.kmem.limit_in_bytes Deprecated knob to set and read the kernel 102 memory hard limit. Kernel hard limit is not 103 supported since 5.16. Writing any value to 104 do file will not have any effect same as if 105 nokmem kernel parameter was specified. 106 Kernel memory is still charged and reported 107 by memory.kmem.usage_in_bytes. 108 memory.kmem.usage_in_bytes show current kernel memory allocation 109 memory.kmem.failcnt show the number of kernel memory usage 110 hits limits 111 memory.kmem.max_usage_in_bytes show max kernel memory usage recorded 112 113 memory.kmem.tcp.limit_in_bytes set/show hard limit for tcp buf memory 114 This knob is deprecated and shouldn't be 115 used. 116 memory.kmem.tcp.usage_in_bytes show current tcp buf memory allocation 117 This knob is deprecated and shouldn't be 118 used. 119 memory.kmem.tcp.failcnt show the number of tcp buf memory usage 120 hits limits 121 This knob is deprecated and shouldn't be 122 used. 123 memory.kmem.tcp.max_usage_in_bytes show max tcp buf memory usage recorded 124 This knob is deprecated and shouldn't be 125 used. 126==================================== ========================================== 127 1281. History 129========== 130 131The memory controller has a long history. A request for comments for the memory 132controller was posted by Balbir Singh [1]_. At the time the RFC was posted 133there were several implementations for memory control. The goal of the 134RFC was to build consensus and agreement for the minimal features required 135for memory control. The first RSS controller was posted by Balbir Singh [2]_ 136in Feb 2007. Pavel Emelianov [3]_ [4]_ [5]_ has since posted three versions 137of the RSS controller. At OLS, at the resource management BoF, everyone 138suggested that we handle both page cache and RSS together. Another request was 139raised to allow user space handling of OOM. The current memory controller is 140at version 6; it combines both mapped (RSS) and unmapped Page 141Cache Control [11]_. 142 1432. Memory Control 144================= 145 146Memory is a unique resource in the sense that it is present in a limited 147amount. If a task requires a lot of CPU processing, the task can spread 148its processing over a period of hours, days, months or years, but with 149memory, the same physical memory needs to be reused to accomplish the task. 150 151The memory controller implementation has been divided into phases. These 152are: 153 1541. Memory controller 1552. mlock(2) controller 1563. Kernel user memory accounting and slab control 1574. user mappings length controller 158 159The memory controller is the first controller developed. 160 1612.1. Design 162----------- 163 164The core of the design is a counter called the page_counter. The 165page_counter tracks the current memory usage and limit of the group of 166processes associated with the controller. Each cgroup has a memory controller 167specific data structure (mem_cgroup) associated with it. 168 1692.2. Accounting 170--------------- 171 172.. code-block:: 173 :caption: Figure 1: Hierarchy of Accounting 174 175 +--------------------+ 176 | mem_cgroup | 177 | (page_counter) | 178 +--------------------+ 179 / ^ \ 180 / | \ 181 +---------------+ | +---------------+ 182 | mm_struct | |.... | mm_struct | 183 | | | | | 184 +---------------+ | +---------------+ 185 | 186 + --------------+ 187 | 188 +---------------+ +------+--------+ 189 | page +----------> page_cgroup| 190 | | | | 191 +---------------+ +---------------+ 192 193 194 195Figure 1 shows the important aspects of the controller 196 1971. Accounting happens per cgroup 1982. Each mm_struct knows about which cgroup it belongs to 1993. Each page has a pointer to the page_cgroup, which in turn knows the 200 cgroup it belongs to 201 202The accounting is done as follows: mem_cgroup_charge_common() is invoked to 203set up the necessary data structures and check if the cgroup that is being 204charged is over its limit. If it is, then reclaim is invoked on the cgroup. 205More details can be found in the reclaim section of this document. 206If everything goes well, a page meta-data-structure called page_cgroup is 207updated. page_cgroup has its own LRU on cgroup. 208(*) page_cgroup structure is allocated at boot/memory-hotplug time. 209 2102.2.1 Accounting details 211------------------------ 212 213All mapped anon pages (RSS) and cache pages (Page Cache) are accounted. 214Some pages which are never reclaimable and will not be on the LRU 215are not accounted. We just account pages under usual VM management. 216 217RSS pages are accounted at page_fault unless they've already been accounted 218for earlier. A file page will be accounted for as Page Cache when it's 219inserted into inode (xarray). While it's mapped into the page tables of 220processes, duplicate accounting is carefully avoided. 221 222An RSS page is unaccounted when it's fully unmapped. A PageCache page is 223unaccounted when it's removed from xarray. Even if RSS pages are fully 224unmapped (by kswapd), they may exist as SwapCache in the system until they 225are really freed. Such SwapCaches are also accounted. 226A swapped-in page is accounted after adding into swapcache. 227 228Note: The kernel does swapin-readahead and reads multiple swaps at once. 229Since page's memcg recorded into swap whatever memsw enabled, the page will 230be accounted after swapin. 231 232At page migration, accounting information is kept. 233 234Note: we just account pages-on-LRU because our purpose is to control amount 235of used pages; not-on-LRU pages tend to be out-of-control from VM view. 236 2372.3 Shared Page Accounting 238-------------------------- 239 240Shared pages are accounted on the basis of the first touch approach. The 241cgroup that first touches a page is accounted for the page. The principle 242behind this approach is that a cgroup that aggressively uses a shared 243page will eventually get charged for it (once it is uncharged from 244the cgroup that brought it in -- this will happen on memory pressure). 245 246But see :ref:`section 8.2 <cgroup-v1-memory-movable-charges>` when moving a 247task to another cgroup, its pages may be recharged to the new cgroup, if 248move_charge_at_immigrate has been chosen. 249 2502.4 Swap Extension 251-------------------------------------- 252 253Swap usage is always recorded for each of cgroup. Swap Extension allows you to 254read and limit it. 255 256When CONFIG_SWAP is enabled, following files are added. 257 258 - memory.memsw.usage_in_bytes. 259 - memory.memsw.limit_in_bytes. 260 261memsw means memory+swap. Usage of memory+swap is limited by 262memsw.limit_in_bytes. 263 264Example: Assume a system with 4G of swap. A task which allocates 6G of memory 265(by mistake) under 2G memory limitation will use all swap. 266In this case, setting memsw.limit_in_bytes=3G will prevent bad use of swap. 267By using the memsw limit, you can avoid system OOM which can be caused by swap 268shortage. 269 2702.4.1 why 'memory+swap' rather than swap 271~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~ 272 273The global LRU(kswapd) can swap out arbitrary pages. Swap-out means 274to move account from memory to swap...there is no change in usage of 275memory+swap. In other words, when we want to limit the usage of swap without 276affecting global LRU, memory+swap limit is better than just limiting swap from 277an OS point of view. 278 2792.4.2. What happens when a cgroup hits memory.memsw.limit_in_bytes 280~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~ 281 282When a cgroup hits memory.memsw.limit_in_bytes, it's useless to do swap-out 283in this cgroup. Then, swap-out will not be done by cgroup routine and file 284caches are dropped. But as mentioned above, global LRU can do swapout memory 285from it for sanity of the system's memory management state. You can't forbid 286it by cgroup. 287 2882.5 Reclaim 289----------- 290 291Each cgroup maintains a per cgroup LRU which has the same structure as 292global VM. When a cgroup goes over its limit, we first try 293to reclaim memory from the cgroup so as to make space for the new 294pages that the cgroup has touched. If the reclaim is unsuccessful, 295an OOM routine is invoked to select and kill the bulkiest task in the 296cgroup. (See :ref:`10. OOM Control <cgroup-v1-memory-oom-control>` below.) 297 298The reclaim algorithm has not been modified for cgroups, except that 299pages that are selected for reclaiming come from the per-cgroup LRU 300list. 301 302.. note:: 303 Reclaim does not work for the root cgroup, since we cannot set any 304 limits on the root cgroup. 305 306.. note:: 307 When panic_on_oom is set to "2", the whole system will panic. 308 309When oom event notifier is registered, event will be delivered. 310(See :ref:`oom_control <cgroup-v1-memory-oom-control>` section) 311 3122.6 Locking 313----------- 314 315Lock order is as follows:: 316 317 folio_lock 318 mm->page_table_lock or split pte_lock 319 folio_memcg_lock (memcg->move_lock) 320 mapping->i_pages lock 321 lruvec->lru_lock. 322 323Per-node-per-memcgroup LRU (cgroup's private LRU) is guarded by 324lruvec->lru_lock; the folio LRU flag is cleared before 325isolating a page from its LRU under lruvec->lru_lock. 326 327.. _cgroup-v1-memory-kernel-extension: 328 3292.7 Kernel Memory Extension 330----------------------------------------------- 331 332With the Kernel memory extension, the Memory Controller is able to limit 333the amount of kernel memory used by the system. Kernel memory is fundamentally 334different than user memory, since it can't be swapped out, which makes it 335possible to DoS the system by consuming too much of this precious resource. 336 337Kernel memory accounting is enabled for all memory cgroups by default. But 338it can be disabled system-wide by passing cgroup.memory=nokmem to the kernel 339at boot time. In this case, kernel memory will not be accounted at all. 340 341Kernel memory limits are not imposed for the root cgroup. Usage for the root 342cgroup may or may not be accounted. The memory used is accumulated into 343memory.kmem.usage_in_bytes, or in a separate counter when it makes sense. 344(currently only for tcp). 345 346The main "kmem" counter is fed into the main counter, so kmem charges will 347also be visible from the user counter. 348 349Currently no soft limit is implemented for kernel memory. It is future work 350to trigger slab reclaim when those limits are reached. 351 3522.7.1 Current Kernel Memory resources accounted 353----------------------------------------------- 354 355stack pages: 356 every process consumes some stack pages. By accounting into 357 kernel memory, we prevent new processes from being created when the kernel 358 memory usage is too high. 359 360slab pages: 361 pages allocated by the SLAB or SLUB allocator are tracked. A copy 362 of each kmem_cache is created every time the cache is touched by the first time 363 from inside the memcg. The creation is done lazily, so some objects can still be 364 skipped while the cache is being created. All objects in a slab page should 365 belong to the same memcg. This only fails to hold when a task is migrated to a 366 different memcg during the page allocation by the cache. 367 368sockets memory pressure: 369 some sockets protocols have memory pressure 370 thresholds. The Memory Controller allows them to be controlled individually 371 per cgroup, instead of globally. 372 373tcp memory pressure: 374 sockets memory pressure for the tcp protocol. 375 3762.7.2 Common use cases 377---------------------- 378 379Because the "kmem" counter is fed to the main user counter, kernel memory can 380never be limited completely independently of user memory. Say "U" is the user 381limit, and "K" the kernel limit. There are three possible ways limits can be 382set: 383 384U != 0, K = unlimited: 385 This is the standard memcg limitation mechanism already present before kmem 386 accounting. Kernel memory is completely ignored. 387 388U != 0, K < U: 389 Kernel memory is a subset of the user memory. This setup is useful in 390 deployments where the total amount of memory per-cgroup is overcommitted. 391 Overcommitting kernel memory limits is definitely not recommended, since the 392 box can still run out of non-reclaimable memory. 393 In this case, the admin could set up K so that the sum of all groups is 394 never greater than the total memory, and freely set U at the cost of his 395 QoS. 396 397 .. warning:: 398 In the current implementation, memory reclaim will NOT be triggered for 399 a cgroup when it hits K while staying below U, which makes this setup 400 impractical. 401 402U != 0, K >= U: 403 Since kmem charges will also be fed to the user counter and reclaim will be 404 triggered for the cgroup for both kinds of memory. This setup gives the 405 admin a unified view of memory, and it is also useful for people who just 406 want to track kernel memory usage. 407 4083. User Interface 409================= 410 411To use the user interface: 412 4131. Enable CONFIG_CGROUPS and CONFIG_MEMCG options 4142. Prepare the cgroups (see :ref:`Why are cgroups needed? 415 <cgroups-why-needed>` for the background information):: 416 417 # mount -t tmpfs none /sys/fs/cgroup 418 # mkdir /sys/fs/cgroup/memory 419 # mount -t cgroup none /sys/fs/cgroup/memory -o memory 420 4213. Make the new group and move bash into it:: 422 423 # mkdir /sys/fs/cgroup/memory/0 424 # echo $$ > /sys/fs/cgroup/memory/0/tasks 425 4264. Since now we're in the 0 cgroup, we can alter the memory limit:: 427 428 # echo 4M > /sys/fs/cgroup/memory/0/memory.limit_in_bytes 429 430 The limit can now be queried:: 431 432 # cat /sys/fs/cgroup/memory/0/memory.limit_in_bytes 433 4194304 434 435.. note:: 436 We can use a suffix (k, K, m, M, g or G) to indicate values in kilo, 437 mega or gigabytes. (Here, Kilo, Mega, Giga are Kibibytes, Mebibytes, 438 Gibibytes.) 439 440.. note:: 441 We can write "-1" to reset the ``*.limit_in_bytes(unlimited)``. 442 443.. note:: 444 We cannot set limits on the root cgroup any more. 445 446 447We can check the usage:: 448 449 # cat /sys/fs/cgroup/memory/0/memory.usage_in_bytes 450 1216512 451 452A successful write to this file does not guarantee a successful setting of 453this limit to the value written into the file. This can be due to a 454number of factors, such as rounding up to page boundaries or the total 455availability of memory on the system. The user is required to re-read 456this file after a write to guarantee the value committed by the kernel:: 457 458 # echo 1 > memory.limit_in_bytes 459 # cat memory.limit_in_bytes 460 4096 461 462The memory.failcnt field gives the number of times that the cgroup limit was 463exceeded. 464 465The memory.stat file gives accounting information. Now, the number of 466caches, RSS and Active pages/Inactive pages are shown. 467 4684. Testing 469========== 470 471For testing features and implementation, see memcg_test.txt. 472 473Performance test is also important. To see pure memory controller's overhead, 474testing on tmpfs will give you good numbers of small overheads. 475Example: do kernel make on tmpfs. 476 477Page-fault scalability is also important. At measuring parallel 478page fault test, multi-process test may be better than multi-thread 479test because it has noise of shared objects/status. 480 481But the above two are testing extreme situations. 482Trying usual test under memory controller is always helpful. 483 484.. _cgroup-v1-memory-test-troubleshoot: 485 4864.1 Troubleshooting 487------------------- 488 489Sometimes a user might find that the application under a cgroup is 490terminated by the OOM killer. There are several causes for this: 491 4921. The cgroup limit is too low (just too low to do anything useful) 4932. The user is using anonymous memory and swap is turned off or too low 494 495A sync followed by echo 1 > /proc/sys/vm/drop_caches will help get rid of 496some of the pages cached in the cgroup (page cache pages). 497 498To know what happens, disabling OOM_Kill as per :ref:`"10. OOM Control" 499<cgroup-v1-memory-oom-control>` (below) and seeing what happens will be 500helpful. 501 502.. _cgroup-v1-memory-test-task-migration: 503 5044.2 Task migration 505------------------ 506 507When a task migrates from one cgroup to another, its charge is not 508carried forward by default. The pages allocated from the original cgroup still 509remain charged to it, the charge is dropped when the page is freed or 510reclaimed. 511 512You can move charges of a task along with task migration. 513See :ref:`8. "Move charges at task migration" <cgroup-v1-memory-move-charges>` 514 5154.3 Removing a cgroup 516--------------------- 517 518A cgroup can be removed by rmdir, but as discussed in :ref:`sections 4.1 519<cgroup-v1-memory-test-troubleshoot>` and :ref:`4.2 520<cgroup-v1-memory-test-task-migration>`, a cgroup might have some charge 521associated with it, even though all tasks have migrated away from it. (because 522we charge against pages, not against tasks.) 523 524We move the stats to parent, and no change on the charge except uncharging 525from the child. 526 527Charges recorded in swap information is not updated at removal of cgroup. 528Recorded information is discarded and a cgroup which uses swap (swapcache) 529will be charged as a new owner of it. 530 5315. Misc. interfaces 532=================== 533 5345.1 force_empty 535--------------- 536 memory.force_empty interface is provided to make cgroup's memory usage empty. 537 When writing anything to this:: 538 539 # echo 0 > memory.force_empty 540 541 the cgroup will be reclaimed and as many pages reclaimed as possible. 542 543 The typical use case for this interface is before calling rmdir(). 544 Though rmdir() offlines memcg, but the memcg may still stay there due to 545 charged file caches. Some out-of-use page caches may keep charged until 546 memory pressure happens. If you want to avoid that, force_empty will be useful. 547 5485.2 stat file 549------------- 550 551memory.stat file includes following statistics: 552 553 * per-memory cgroup local status 554 555 =============== =============================================================== 556 cache # of bytes of page cache memory. 557 rss # of bytes of anonymous and swap cache memory (includes 558 transparent hugepages). 559 rss_huge # of bytes of anonymous transparent hugepages. 560 mapped_file # of bytes of mapped file (includes tmpfs/shmem) 561 pgpgin # of charging events to the memory cgroup. The charging 562 event happens each time a page is accounted as either mapped 563 anon page(RSS) or cache page(Page Cache) to the cgroup. 564 pgpgout # of uncharging events to the memory cgroup. The uncharging 565 event happens each time a page is unaccounted from the 566 cgroup. 567 swap # of bytes of swap usage 568 swapcached # of bytes of swap cached in memory 569 dirty # of bytes that are waiting to get written back to the disk. 570 writeback # of bytes of file/anon cache that are queued for syncing to 571 disk. 572 inactive_anon # of bytes of anonymous and swap cache memory on inactive 573 LRU list. 574 active_anon # of bytes of anonymous and swap cache memory on active 575 LRU list. 576 inactive_file # of bytes of file-backed memory and MADV_FREE anonymous 577 memory (LazyFree pages) on inactive LRU list. 578 active_file # of bytes of file-backed memory on active LRU list. 579 unevictable # of bytes of memory that cannot be reclaimed (mlocked etc). 580 =============== =============================================================== 581 582 * status considering hierarchy (see memory.use_hierarchy settings): 583 584 ========================= =================================================== 585 hierarchical_memory_limit # of bytes of memory limit with regard to 586 hierarchy 587 under which the memory cgroup is 588 hierarchical_memsw_limit # of bytes of memory+swap limit with regard to 589 hierarchy under which memory cgroup is. 590 591 total_<counter> # hierarchical version of <counter>, which in 592 addition to the cgroup's own value includes the 593 sum of all hierarchical children's values of 594 <counter>, i.e. total_cache 595 ========================= =================================================== 596 597 * additional vm parameters (depends on CONFIG_DEBUG_VM): 598 599 ========================= ======================================== 600 recent_rotated_anon VM internal parameter. (see mm/vmscan.c) 601 recent_rotated_file VM internal parameter. (see mm/vmscan.c) 602 recent_scanned_anon VM internal parameter. (see mm/vmscan.c) 603 recent_scanned_file VM internal parameter. (see mm/vmscan.c) 604 ========================= ======================================== 605 606.. hint:: 607 recent_rotated means recent frequency of LRU rotation. 608 recent_scanned means recent # of scans to LRU. 609 showing for better debug please see the code for meanings. 610 611.. note:: 612 Only anonymous and swap cache memory is listed as part of 'rss' stat. 613 This should not be confused with the true 'resident set size' or the 614 amount of physical memory used by the cgroup. 615 616 'rss + mapped_file" will give you resident set size of cgroup. 617 618 (Note: file and shmem may be shared among other cgroups. In that case, 619 mapped_file is accounted only when the memory cgroup is owner of page 620 cache.) 621 6225.3 swappiness 623-------------- 624 625Overrides /proc/sys/vm/swappiness for the particular group. The tunable 626in the root cgroup corresponds to the global swappiness setting. 627 628Please note that unlike during the global reclaim, limit reclaim 629enforces that 0 swappiness really prevents from any swapping even if 630there is a swap storage available. This might lead to memcg OOM killer 631if there are no file pages to reclaim. 632 6335.4 failcnt 634----------- 635 636A memory cgroup provides memory.failcnt and memory.memsw.failcnt files. 637This failcnt(== failure count) shows the number of times that a usage counter 638hit its limit. When a memory cgroup hits a limit, failcnt increases and 639memory under it will be reclaimed. 640 641You can reset failcnt by writing 0 to failcnt file:: 642 643 # echo 0 > .../memory.failcnt 644 6455.5 usage_in_bytes 646------------------ 647 648For efficiency, as other kernel components, memory cgroup uses some optimization 649to avoid unnecessary cacheline false sharing. usage_in_bytes is affected by the 650method and doesn't show 'exact' value of memory (and swap) usage, it's a fuzz 651value for efficient access. (Of course, when necessary, it's synchronized.) 652If you want to know more exact memory usage, you should use RSS+CACHE(+SWAP) 653value in memory.stat(see 5.2). 654 6555.6 numa_stat 656------------- 657 658This is similar to numa_maps but operates on a per-memcg basis. This is 659useful for providing visibility into the numa locality information within 660an memcg since the pages are allowed to be allocated from any physical 661node. One of the use cases is evaluating application performance by 662combining this information with the application's CPU allocation. 663 664Each memcg's numa_stat file includes "total", "file", "anon" and "unevictable" 665per-node page counts including "hierarchical_<counter>" which sums up all 666hierarchical children's values in addition to the memcg's own value. 667 668The output format of memory.numa_stat is:: 669 670 total=<total pages> N0=<node 0 pages> N1=<node 1 pages> ... 671 file=<total file pages> N0=<node 0 pages> N1=<node 1 pages> ... 672 anon=<total anon pages> N0=<node 0 pages> N1=<node 1 pages> ... 673 unevictable=<total anon pages> N0=<node 0 pages> N1=<node 1 pages> ... 674 hierarchical_<counter>=<counter pages> N0=<node 0 pages> N1=<node 1 pages> ... 675 676The "total" count is sum of file + anon + unevictable. 677 6786. Hierarchy support 679==================== 680 681The memory controller supports a deep hierarchy and hierarchical accounting. 682The hierarchy is created by creating the appropriate cgroups in the 683cgroup filesystem. Consider for example, the following cgroup filesystem 684hierarchy:: 685 686 root 687 / | \ 688 / | \ 689 a b c 690 | \ 691 | \ 692 d e 693 694In the diagram above, with hierarchical accounting enabled, all memory 695usage of e, is accounted to its ancestors up until the root (i.e, c and root). 696If one of the ancestors goes over its limit, the reclaim algorithm reclaims 697from the tasks in the ancestor and the children of the ancestor. 698 6996.1 Hierarchical accounting and reclaim 700--------------------------------------- 701 702Hierarchical accounting is enabled by default. Disabling the hierarchical 703accounting is deprecated. An attempt to do it will result in a failure 704and a warning printed to dmesg. 705 706For compatibility reasons writing 1 to memory.use_hierarchy will always pass:: 707 708 # echo 1 > memory.use_hierarchy 709 7107. Soft limits (DEPRECATED) 711=========================== 712 713THIS IS DEPRECATED! 714 715Soft limits allow for greater sharing of memory. The idea behind soft limits 716is to allow control groups to use as much of the memory as needed, provided 717 718a. There is no memory contention 719b. They do not exceed their hard limit 720 721When the system detects memory contention or low memory, control groups 722are pushed back to their soft limits. If the soft limit of each control 723group is very high, they are pushed back as much as possible to make 724sure that one control group does not starve the others of memory. 725 726Please note that soft limits is a best-effort feature; it comes with 727no guarantees, but it does its best to make sure that when memory is 728heavily contended for, memory is allocated based on the soft limit 729hints/setup. Currently soft limit based reclaim is set up such that 730it gets invoked from balance_pgdat (kswapd). 731 7327.1 Interface 733------------- 734 735Soft limits can be setup by using the following commands (in this example we 736assume a soft limit of 256 MiB):: 737 738 # echo 256M > memory.soft_limit_in_bytes 739 740If we want to change this to 1G, we can at any time use:: 741 742 # echo 1G > memory.soft_limit_in_bytes 743 744.. note:: 745 Soft limits take effect over a long period of time, since they involve 746 reclaiming memory for balancing between memory cgroups 747 748.. note:: 749 It is recommended to set the soft limit always below the hard limit, 750 otherwise the hard limit will take precedence. 751 752.. _cgroup-v1-memory-move-charges: 753 7548. Move charges at task migration (DEPRECATED!) 755=============================================== 756 757THIS IS DEPRECATED! 758 759It's expensive and unreliable! It's better practice to launch workload 760tasks directly from inside their target cgroup. Use dedicated workload 761cgroups to allow fine-grained policy adjustments without having to 762move physical pages between control domains. 763 764Users can move charges associated with a task along with task migration, that 765is, uncharge task's pages from the old cgroup and charge them to the new cgroup. 766This feature is not supported in !CONFIG_MMU environments because of lack of 767page tables. 768 7698.1 Interface 770------------- 771 772This feature is disabled by default. It can be enabled (and disabled again) by 773writing to memory.move_charge_at_immigrate of the destination cgroup. 774 775If you want to enable it:: 776 777 # echo (some positive value) > memory.move_charge_at_immigrate 778 779.. note:: 780 Each bits of move_charge_at_immigrate has its own meaning about what type 781 of charges should be moved. See :ref:`section 8.2 782 <cgroup-v1-memory-movable-charges>` for details. 783 784.. note:: 785 Charges are moved only when you move mm->owner, in other words, 786 a leader of a thread group. 787 788.. note:: 789 If we cannot find enough space for the task in the destination cgroup, we 790 try to make space by reclaiming memory. Task migration may fail if we 791 cannot make enough space. 792 793.. note:: 794 It can take several seconds if you move charges much. 795 796And if you want disable it again:: 797 798 # echo 0 > memory.move_charge_at_immigrate 799 800.. _cgroup-v1-memory-movable-charges: 801 8028.2 Type of charges which can be moved 803-------------------------------------- 804 805Each bit in move_charge_at_immigrate has its own meaning about what type of 806charges should be moved. But in any case, it must be noted that an account of 807a page or a swap can be moved only when it is charged to the task's current 808(old) memory cgroup. 809 810+---+--------------------------------------------------------------------------+ 811|bit| what type of charges would be moved ? | 812+===+==========================================================================+ 813| 0 | A charge of an anonymous page (or swap of it) used by the target task. | 814| | You must enable Swap Extension (see 2.4) to enable move of swap charges. | 815+---+--------------------------------------------------------------------------+ 816| 1 | A charge of file pages (normal file, tmpfs file (e.g. ipc shared memory) | 817| | and swaps of tmpfs file) mmapped by the target task. Unlike the case of | 818| | anonymous pages, file pages (and swaps) in the range mmapped by the task | 819| | will be moved even if the task hasn't done page fault, i.e. they might | 820| | not be the task's "RSS", but other task's "RSS" that maps the same file. | 821| | The mapcount of the page is ignored (the page can be moved independent | 822| | of the mapcount). You must enable Swap Extension (see 2.4) to | 823| | enable move of swap charges. | 824+---+--------------------------------------------------------------------------+ 825 8268.3 TODO 827-------- 828 829- All of moving charge operations are done under cgroup_mutex. It's not good 830 behavior to hold the mutex too long, so we may need some trick. 831 8329. Memory thresholds 833==================== 834 835Memory cgroup implements memory thresholds using the cgroups notification 836API (see cgroups.txt). It allows to register multiple memory and memsw 837thresholds and gets notifications when it crosses. 838 839To register a threshold, an application must: 840 841- create an eventfd using eventfd(2); 842- open memory.usage_in_bytes or memory.memsw.usage_in_bytes; 843- write string like "<event_fd> <fd of memory.usage_in_bytes> <threshold>" to 844 cgroup.event_control. 845 846Application will be notified through eventfd when memory usage crosses 847threshold in any direction. 848 849It's applicable for root and non-root cgroup. 850 851.. _cgroup-v1-memory-oom-control: 852 85310. OOM Control (DEPRECATED) 854============================ 855 856THIS IS DEPRECATED! 857 858memory.oom_control file is for OOM notification and other controls. 859 860Memory cgroup implements OOM notifier using the cgroup notification 861API (See cgroups.txt). It allows to register multiple OOM notification 862delivery and gets notification when OOM happens. 863 864To register a notifier, an application must: 865 866 - create an eventfd using eventfd(2) 867 - open memory.oom_control file 868 - write string like "<event_fd> <fd of memory.oom_control>" to 869 cgroup.event_control 870 871The application will be notified through eventfd when OOM happens. 872OOM notification doesn't work for the root cgroup. 873 874You can disable the OOM-killer by writing "1" to memory.oom_control file, as: 875 876 #echo 1 > memory.oom_control 877 878If OOM-killer is disabled, tasks under cgroup will hang/sleep 879in memory cgroup's OOM-waitqueue when they request accountable memory. 880 881For running them, you have to relax the memory cgroup's OOM status by 882 883 * enlarge limit or reduce usage. 884 885To reduce usage, 886 887 * kill some tasks. 888 * move some tasks to other group with account migration. 889 * remove some files (on tmpfs?) 890 891Then, stopped tasks will work again. 892 893At reading, current status of OOM is shown. 894 895 - oom_kill_disable 0 or 1 896 (if 1, oom-killer is disabled) 897 - under_oom 0 or 1 898 (if 1, the memory cgroup is under OOM, tasks may be stopped.) 899 - oom_kill integer counter 900 The number of processes belonging to this cgroup killed by any 901 kind of OOM killer. 902 90311. Memory Pressure (DEPRECATED) 904================================ 905 906THIS IS DEPRECATED! 907 908The pressure level notifications can be used to monitor the memory 909allocation cost; based on the pressure, applications can implement 910different strategies of managing their memory resources. The pressure 911levels are defined as following: 912 913The "low" level means that the system is reclaiming memory for new 914allocations. Monitoring this reclaiming activity might be useful for 915maintaining cache level. Upon notification, the program (typically 916"Activity Manager") might analyze vmstat and act in advance (i.e. 917prematurely shutdown unimportant services). 918 919The "medium" level means that the system is experiencing medium memory 920pressure, the system might be making swap, paging out active file caches, 921etc. Upon this event applications may decide to further analyze 922vmstat/zoneinfo/memcg or internal memory usage statistics and free any 923resources that can be easily reconstructed or re-read from a disk. 924 925The "critical" level means that the system is actively thrashing, it is 926about to out of memory (OOM) or even the in-kernel OOM killer is on its 927way to trigger. Applications should do whatever they can to help the 928system. It might be too late to consult with vmstat or any other 929statistics, so it's advisable to take an immediate action. 930 931By default, events are propagated upward until the event is handled, i.e. the 932events are not pass-through. For example, you have three cgroups: A->B->C. Now 933you set up an event listener on cgroups A, B and C, and suppose group C 934experiences some pressure. In this situation, only group C will receive the 935notification, i.e. groups A and B will not receive it. This is done to avoid 936excessive "broadcasting" of messages, which disturbs the system and which is 937especially bad if we are low on memory or thrashing. Group B, will receive 938notification only if there are no event listeners for group C. 939 940There are three optional modes that specify different propagation behavior: 941 942 - "default": this is the default behavior specified above. This mode is the 943 same as omitting the optional mode parameter, preserved by backwards 944 compatibility. 945 946 - "hierarchy": events always propagate up to the root, similar to the default 947 behavior, except that propagation continues regardless of whether there are 948 event listeners at each level, with the "hierarchy" mode. In the above 949 example, groups A, B, and C will receive notification of memory pressure. 950 951 - "local": events are pass-through, i.e. they only receive notifications when 952 memory pressure is experienced in the memcg for which the notification is 953 registered. In the above example, group C will receive notification if 954 registered for "local" notification and the group experiences memory 955 pressure. However, group B will never receive notification, regardless if 956 there is an event listener for group C or not, if group B is registered for 957 local notification. 958 959The level and event notification mode ("hierarchy" or "local", if necessary) are 960specified by a comma-delimited string, i.e. "low,hierarchy" specifies 961hierarchical, pass-through, notification for all ancestor memcgs. Notification 962that is the default, non pass-through behavior, does not specify a mode. 963"medium,local" specifies pass-through notification for the medium level. 964 965The file memory.pressure_level is only used to setup an eventfd. To 966register a notification, an application must: 967 968- create an eventfd using eventfd(2); 969- open memory.pressure_level; 970- write string as "<event_fd> <fd of memory.pressure_level> <level[,mode]>" 971 to cgroup.event_control. 972 973Application will be notified through eventfd when memory pressure is at 974the specific level (or higher). Read/write operations to 975memory.pressure_level are no implemented. 976 977Test: 978 979 Here is a small script example that makes a new cgroup, sets up a 980 memory limit, sets up a notification in the cgroup and then makes child 981 cgroup experience a critical pressure:: 982 983 # cd /sys/fs/cgroup/memory/ 984 # mkdir foo 985 # cd foo 986 # cgroup_event_listener memory.pressure_level low,hierarchy & 987 # echo 8000000 > memory.limit_in_bytes 988 # echo 8000000 > memory.memsw.limit_in_bytes 989 # echo $$ > tasks 990 # dd if=/dev/zero | read x 991 992 (Expect a bunch of notifications, and eventually, the oom-killer will 993 trigger.) 994 99512. TODO 996======== 997 9981. Make per-cgroup scanner reclaim not-shared pages first 9992. Teach controller to account for shared-pages 10003. Start reclamation in the background when the limit is 1001 not yet hit but the usage is getting closer 1002 1003Summary 1004======= 1005 1006Overall, the memory controller has been a stable controller and has been 1007commented and discussed quite extensively in the community. 1008 1009References 1010========== 1011 1012.. [1] Singh, Balbir. RFC: Memory Controller, http://lwn.net/Articles/206697/ 1013.. [2] Singh, Balbir. Memory Controller (RSS Control), 1014 http://lwn.net/Articles/222762/ 1015.. [3] Emelianov, Pavel. Resource controllers based on process cgroups 1016 https://lore.kernel.org/r/45ED7DEC.7010403@sw.ru 1017.. [4] Emelianov, Pavel. RSS controller based on process cgroups (v2) 1018 https://lore.kernel.org/r/461A3010.90403@sw.ru 1019.. [5] Emelianov, Pavel. RSS controller based on process cgroups (v3) 1020 https://lore.kernel.org/r/465D9739.8070209@openvz.org 1021 10226. Menage, Paul. Control Groups v10, http://lwn.net/Articles/236032/ 10237. Vaidyanathan, Srinivasan, Control Groups: Pagecache accounting and control 1024 subsystem (v3), http://lwn.net/Articles/235534/ 10258. Singh, Balbir. RSS controller v2 test results (lmbench), 1026 https://lore.kernel.org/r/464C95D4.7070806@linux.vnet.ibm.com 10279. Singh, Balbir. RSS controller v2 AIM9 results 1028 https://lore.kernel.org/r/464D267A.50107@linux.vnet.ibm.com 102910. Singh, Balbir. Memory controller v6 test results, 1030 https://lore.kernel.org/r/20070819094658.654.84837.sendpatchset@balbir-laptop 1031 1032.. [11] Singh, Balbir. Memory controller introduction (v6), 1033 https://lore.kernel.org/r/20070817084228.26003.12568.sendpatchset@balbir-laptop 1034.. [12] Corbet, Jonathan, Controlling memory use in cgroups, 1035 http://lwn.net/Articles/243795/ 1036