Based on kernel version 4.16.1. Page generated on 2018-04-09 11:52 EST.
1 Memory Resource Controller 2 3 NOTE: This document is hopelessly outdated and it asks for a complete 4 rewrite. It still contains a useful information so we are keeping it 5 here but make sure to check the current code if you need a deeper 6 understanding. 7 8 NOTE: The Memory Resource Controller has generically been referred to as the 9 memory controller in this document. Do not confuse memory controller 10 used here with the memory controller that is used in hardware. 11 12 (For editors) 13 In this document: 14 When we mention a cgroup (cgroupfs's directory) with memory controller, 15 we call it "memory cgroup". When you see git-log and source code, you'll 16 see patch's title and function names tend to use "memcg". 17 In this document, we avoid using it. 18 19 Benefits and Purpose of the memory controller 20 21 The memory controller isolates the memory behaviour of a group of tasks 22 from the rest of the system. The article on LWN  mentions some probable 23 uses of the memory controller. The memory controller can be used to 24 25 a. Isolate an application or a group of applications 26 Memory-hungry applications can be isolated and limited to a smaller 27 amount of memory. 28 b. Create a cgroup with a limited amount of memory; this can be used 29 as a good alternative to booting with mem=XXXX. 30 c. Virtualization solutions can control the amount of memory they want 31 to assign to a virtual machine instance. 32 d. A CD/DVD burner could control the amount of memory used by the 33 rest of the system to ensure that burning does not fail due to lack 34 of available memory. 35 e. There are several other use cases; find one or use the controller just 36 for fun (to learn and hack on the VM subsystem). 37 38 Current Status: linux-2.6.34-mmotm(development version of 2010/April) 39 40 Features: 41 - accounting anonymous pages, file caches, swap caches usage and limiting them. 42 - pages are linked to per-memcg LRU exclusively, and there is no global LRU. 43 - optionally, memory+swap usage can be accounted and limited. 44 - hierarchical accounting 45 - soft limit 46 - moving (recharging) account at moving a task is selectable. 47 - usage threshold notifier 48 - memory pressure notifier 49 - oom-killer disable knob and oom-notifier 50 - Root cgroup has no limit controls. 51 52 Kernel memory support is a work in progress, and the current version provides 53 basically functionality. (See Section 2.7) 54 55 Brief summary of control files. 56 57 tasks # attach a task(thread) and show list of threads 58 cgroup.procs # show list of processes 59 cgroup.event_control # an interface for event_fd() 60 memory.usage_in_bytes # show current usage for memory 61 (See 5.5 for details) 62 memory.memsw.usage_in_bytes # show current usage for memory+Swap 63 (See 5.5 for details) 64 memory.limit_in_bytes # set/show limit of memory usage 65 memory.memsw.limit_in_bytes # set/show limit of memory+Swap usage 66 memory.failcnt # show the number of memory usage hits limits 67 memory.memsw.failcnt # show the number of memory+Swap hits limits 68 memory.max_usage_in_bytes # show max memory usage recorded 69 memory.memsw.max_usage_in_bytes # show max memory+Swap usage recorded 70 memory.soft_limit_in_bytes # set/show soft limit of memory usage 71 memory.stat # show various statistics 72 memory.use_hierarchy # set/show hierarchical account enabled 73 memory.force_empty # trigger forced move charge to parent 74 memory.pressure_level # set memory pressure notifications 75 memory.swappiness # set/show swappiness parameter of vmscan 76 (See sysctl's vm.swappiness) 77 memory.move_charge_at_immigrate # set/show controls of moving charges 78 memory.oom_control # set/show oom controls. 79 memory.numa_stat # show the number of memory usage per numa node 80 81 memory.kmem.limit_in_bytes # set/show hard limit for kernel memory 82 memory.kmem.usage_in_bytes # show current kernel memory allocation 83 memory.kmem.failcnt # show the number of kernel memory usage hits limits 84 memory.kmem.max_usage_in_bytes # show max kernel memory usage recorded 85 86 memory.kmem.tcp.limit_in_bytes # set/show hard limit for tcp buf memory 87 memory.kmem.tcp.usage_in_bytes # show current tcp buf memory allocation 88 memory.kmem.tcp.failcnt # show the number of tcp buf memory usage hits limits 89 memory.kmem.tcp.max_usage_in_bytes # show max tcp buf memory usage recorded 90 91 1. History 92 93 The memory controller has a long history. A request for comments for the memory 94 controller was posted by Balbir Singh . At the time the RFC was posted 95 there were several implementations for memory control. The goal of the 96 RFC was to build consensus and agreement for the minimal features required 97 for memory control. The first RSS controller was posted by Balbir Singh 98 in Feb 2007. Pavel Emelianov  has since posted three versions of the 99 RSS controller. At OLS, at the resource management BoF, everyone suggested 100 that we handle both page cache and RSS together. Another request was raised 101 to allow user space handling of OOM. The current memory controller is 102 at version 6; it combines both mapped (RSS) and unmapped Page 103 Cache Control . 104 105 2. Memory Control 106 107 Memory is a unique resource in the sense that it is present in a limited 108 amount. If a task requires a lot of CPU processing, the task can spread 109 its processing over a period of hours, days, months or years, but with 110 memory, the same physical memory needs to be reused to accomplish the task. 111 112 The memory controller implementation has been divided into phases. These 113 are: 114 115 1. Memory controller 116 2. mlock(2) controller 117 3. Kernel user memory accounting and slab control 118 4. user mappings length controller 119 120 The memory controller is the first controller developed. 121 122 2.1. Design 123 124 The core of the design is a counter called the page_counter. The 125 page_counter tracks the current memory usage and limit of the group of 126 processes associated with the controller. Each cgroup has a memory controller 127 specific data structure (mem_cgroup) associated with it. 128 129 2.2. Accounting 130 131 +--------------------+ 132 | mem_cgroup | 133 | (page_counter) | 134 +--------------------+ 135 / ^ \ 136 / | \ 137 +---------------+ | +---------------+ 138 | mm_struct | |.... | mm_struct | 139 | | | | | 140 +---------------+ | +---------------+ 141 | 142 + --------------+ 143 | 144 +---------------+ +------+--------+ 145 | page +----------> page_cgroup| 146 | | | | 147 +---------------+ +---------------+ 148 149 (Figure 1: Hierarchy of Accounting) 150 151 152 Figure 1 shows the important aspects of the controller 153 154 1. Accounting happens per cgroup 155 2. Each mm_struct knows about which cgroup it belongs to 156 3. Each page has a pointer to the page_cgroup, which in turn knows the 157 cgroup it belongs to 158 159 The accounting is done as follows: mem_cgroup_charge_common() is invoked to 160 set up the necessary data structures and check if the cgroup that is being 161 charged is over its limit. If it is, then reclaim is invoked on the cgroup. 162 More details can be found in the reclaim section of this document. 163 If everything goes well, a page meta-data-structure called page_cgroup is 164 updated. page_cgroup has its own LRU on cgroup. 165 (*) page_cgroup structure is allocated at boot/memory-hotplug time. 166 167 2.2.1 Accounting details 168 169 All mapped anon pages (RSS) and cache pages (Page Cache) are accounted. 170 Some pages which are never reclaimable and will not be on the LRU 171 are not accounted. We just account pages under usual VM management. 172 173 RSS pages are accounted at page_fault unless they've already been accounted 174 for earlier. A file page will be accounted for as Page Cache when it's 175 inserted into inode (radix-tree). While it's mapped into the page tables of 176 processes, duplicate accounting is carefully avoided. 177 178 An RSS page is unaccounted when it's fully unmapped. A PageCache page is 179 unaccounted when it's removed from radix-tree. Even if RSS pages are fully 180 unmapped (by kswapd), they may exist as SwapCache in the system until they 181 are really freed. Such SwapCaches are also accounted. 182 A swapped-in page is not accounted until it's mapped. 183 184 Note: The kernel does swapin-readahead and reads multiple swaps at once. 185 This means swapped-in pages may contain pages for other tasks than a task 186 causing page fault. So, we avoid accounting at swap-in I/O. 187 188 At page migration, accounting information is kept. 189 190 Note: we just account pages-on-LRU because our purpose is to control amount 191 of used pages; not-on-LRU pages tend to be out-of-control from VM view. 192 193 2.3 Shared Page Accounting 194 195 Shared pages are accounted on the basis of the first touch approach. The 196 cgroup that first touches a page is accounted for the page. The principle 197 behind this approach is that a cgroup that aggressively uses a shared 198 page will eventually get charged for it (once it is uncharged from 199 the cgroup that brought it in -- this will happen on memory pressure). 200 201 But see section 8.2: when moving a task to another cgroup, its pages may 202 be recharged to the new cgroup, if move_charge_at_immigrate has been chosen. 203 204 Exception: If CONFIG_MEMCG_SWAP is not used. 205 When you do swapoff and make swapped-out pages of shmem(tmpfs) to 206 be backed into memory in force, charges for pages are accounted against the 207 caller of swapoff rather than the users of shmem. 208 209 2.4 Swap Extension (CONFIG_MEMCG_SWAP) 210 211 Swap Extension allows you to record charge for swap. A swapped-in page is 212 charged back to original page allocator if possible. 213 214 When swap is accounted, following files are added. 215 - memory.memsw.usage_in_bytes. 216 - memory.memsw.limit_in_bytes. 217 218 memsw means memory+swap. Usage of memory+swap is limited by 219 memsw.limit_in_bytes. 220 221 Example: Assume a system with 4G of swap. A task which allocates 6G of memory 222 (by mistake) under 2G memory limitation will use all swap. 223 In this case, setting memsw.limit_in_bytes=3G will prevent bad use of swap. 224 By using the memsw limit, you can avoid system OOM which can be caused by swap 225 shortage. 226 227 * why 'memory+swap' rather than swap. 228 The global LRU(kswapd) can swap out arbitrary pages. Swap-out means 229 to move account from memory to swap...there is no change in usage of 230 memory+swap. In other words, when we want to limit the usage of swap without 231 affecting global LRU, memory+swap limit is better than just limiting swap from 232 an OS point of view. 233 234 * What happens when a cgroup hits memory.memsw.limit_in_bytes 235 When a cgroup hits memory.memsw.limit_in_bytes, it's useless to do swap-out 236 in this cgroup. Then, swap-out will not be done by cgroup routine and file 237 caches are dropped. But as mentioned above, global LRU can do swapout memory 238 from it for sanity of the system's memory management state. You can't forbid 239 it by cgroup. 240 241 2.5 Reclaim 242 243 Each cgroup maintains a per cgroup LRU which has the same structure as 244 global VM. When a cgroup goes over its limit, we first try 245 to reclaim memory from the cgroup so as to make space for the new 246 pages that the cgroup has touched. If the reclaim is unsuccessful, 247 an OOM routine is invoked to select and kill the bulkiest task in the 248 cgroup. (See 10. OOM Control below.) 249 250 The reclaim algorithm has not been modified for cgroups, except that 251 pages that are selected for reclaiming come from the per-cgroup LRU 252 list. 253 254 NOTE: Reclaim does not work for the root cgroup, since we cannot set any 255 limits on the root cgroup. 256 257 Note2: When panic_on_oom is set to "2", the whole system will panic. 258 259 When oom event notifier is registered, event will be delivered. 260 (See oom_control section) 261 262 2.6 Locking 263 264 lock_page_cgroup()/unlock_page_cgroup() should not be called under 265 mapping->tree_lock. 266 267 Other lock order is following: 268 PG_locked. 269 mm->page_table_lock 270 zone_lru_lock 271 lock_page_cgroup. 272 In many cases, just lock_page_cgroup() is called. 273 per-zone-per-cgroup LRU (cgroup's private LRU) is just guarded by 274 zone_lru_lock, it has no lock of its own. 275 276 2.7 Kernel Memory Extension (CONFIG_MEMCG_KMEM) 277 278 With the Kernel memory extension, the Memory Controller is able to limit 279 the amount of kernel memory used by the system. Kernel memory is fundamentally 280 different than user memory, since it can't be swapped out, which makes it 281 possible to DoS the system by consuming too much of this precious resource. 282 283 Kernel memory accounting is enabled for all memory cgroups by default. But 284 it can be disabled system-wide by passing cgroup.memory=nokmem to the kernel 285 at boot time. In this case, kernel memory will not be accounted at all. 286 287 Kernel memory limits are not imposed for the root cgroup. Usage for the root 288 cgroup may or may not be accounted. The memory used is accumulated into 289 memory.kmem.usage_in_bytes, or in a separate counter when it makes sense. 290 (currently only for tcp). 291 The main "kmem" counter is fed into the main counter, so kmem charges will 292 also be visible from the user counter. 293 294 Currently no soft limit is implemented for kernel memory. It is future work 295 to trigger slab reclaim when those limits are reached. 296 297 2.7.1 Current Kernel Memory resources accounted 298 299 * stack pages: every process consumes some stack pages. By accounting into 300 kernel memory, we prevent new processes from being created when the kernel 301 memory usage is too high. 302 303 * slab pages: pages allocated by the SLAB or SLUB allocator are tracked. A copy 304 of each kmem_cache is created every time the cache is touched by the first time 305 from inside the memcg. The creation is done lazily, so some objects can still be 306 skipped while the cache is being created. All objects in a slab page should 307 belong to the same memcg. This only fails to hold when a task is migrated to a 308 different memcg during the page allocation by the cache. 309 310 * sockets memory pressure: some sockets protocols have memory pressure 311 thresholds. The Memory Controller allows them to be controlled individually 312 per cgroup, instead of globally. 313 314 * tcp memory pressure: sockets memory pressure for the tcp protocol. 315 316 2.7.2 Common use cases 317 318 Because the "kmem" counter is fed to the main user counter, kernel memory can 319 never be limited completely independently of user memory. Say "U" is the user 320 limit, and "K" the kernel limit. There are three possible ways limits can be 321 set: 322 323 U != 0, K = unlimited: 324 This is the standard memcg limitation mechanism already present before kmem 325 accounting. Kernel memory is completely ignored. 326 327 U != 0, K < U: 328 Kernel memory is a subset of the user memory. This setup is useful in 329 deployments where the total amount of memory per-cgroup is overcommited. 330 Overcommiting kernel memory limits is definitely not recommended, since the 331 box can still run out of non-reclaimable memory. 332 In this case, the admin could set up K so that the sum of all groups is 333 never greater than the total memory, and freely set U at the cost of his 334 QoS. 335 WARNING: In the current implementation, memory reclaim will NOT be 336 triggered for a cgroup when it hits K while staying below U, which makes 337 this setup impractical. 338 339 U != 0, K >= U: 340 Since kmem charges will also be fed to the user counter and reclaim will be 341 triggered for the cgroup for both kinds of memory. This setup gives the 342 admin a unified view of memory, and it is also useful for people who just 343 want to track kernel memory usage. 344 345 3. User Interface 346 347 3.0. Configuration 348 349 a. Enable CONFIG_CGROUPS 350 b. Enable CONFIG_MEMCG 351 c. Enable CONFIG_MEMCG_SWAP (to use swap extension) 352 d. Enable CONFIG_MEMCG_KMEM (to use kmem extension) 353 354 3.1. Prepare the cgroups (see cgroups.txt, Why are cgroups needed?) 355 # mount -t tmpfs none /sys/fs/cgroup 356 # mkdir /sys/fs/cgroup/memory 357 # mount -t cgroup none /sys/fs/cgroup/memory -o memory 358 359 3.2. Make the new group and move bash into it 360 # mkdir /sys/fs/cgroup/memory/0 361 # echo $$ > /sys/fs/cgroup/memory/0/tasks 362 363 Since now we're in the 0 cgroup, we can alter the memory limit: 364 # echo 4M > /sys/fs/cgroup/memory/0/memory.limit_in_bytes 365 366 NOTE: We can use a suffix (k, K, m, M, g or G) to indicate values in kilo, 367 mega or gigabytes. (Here, Kilo, Mega, Giga are Kibibytes, Mebibytes, Gibibytes.) 368 369 NOTE: We can write "-1" to reset the *.limit_in_bytes(unlimited). 370 NOTE: We cannot set limits on the root cgroup any more. 371 372 # cat /sys/fs/cgroup/memory/0/memory.limit_in_bytes 373 4194304 374 375 We can check the usage: 376 # cat /sys/fs/cgroup/memory/0/memory.usage_in_bytes 377 1216512 378 379 A successful write to this file does not guarantee a successful setting of 380 this limit to the value written into the file. This can be due to a 381 number of factors, such as rounding up to page boundaries or the total 382 availability of memory on the system. The user is required to re-read 383 this file after a write to guarantee the value committed by the kernel. 384 385 # echo 1 > memory.limit_in_bytes 386 # cat memory.limit_in_bytes 387 4096 388 389 The memory.failcnt field gives the number of times that the cgroup limit was 390 exceeded. 391 392 The memory.stat file gives accounting information. Now, the number of 393 caches, RSS and Active pages/Inactive pages are shown. 394 395 4. Testing 396 397 For testing features and implementation, see memcg_test.txt. 398 399 Performance test is also important. To see pure memory controller's overhead, 400 testing on tmpfs will give you good numbers of small overheads. 401 Example: do kernel make on tmpfs. 402 403 Page-fault scalability is also important. At measuring parallel 404 page fault test, multi-process test may be better than multi-thread 405 test because it has noise of shared objects/status. 406 407 But the above two are testing extreme situations. 408 Trying usual test under memory controller is always helpful. 409 410 4.1 Troubleshooting 411 412 Sometimes a user might find that the application under a cgroup is 413 terminated by the OOM killer. There are several causes for this: 414 415 1. The cgroup limit is too low (just too low to do anything useful) 416 2. The user is using anonymous memory and swap is turned off or too low 417 418 A sync followed by echo 1 > /proc/sys/vm/drop_caches will help get rid of 419 some of the pages cached in the cgroup (page cache pages). 420 421 To know what happens, disabling OOM_Kill as per "10. OOM Control" (below) and 422 seeing what happens will be helpful. 423 424 4.2 Task migration 425 426 When a task migrates from one cgroup to another, its charge is not 427 carried forward by default. The pages allocated from the original cgroup still 428 remain charged to it, the charge is dropped when the page is freed or 429 reclaimed. 430 431 You can move charges of a task along with task migration. 432 See 8. "Move charges at task migration" 433 434 4.3 Removing a cgroup 435 436 A cgroup can be removed by rmdir, but as discussed in sections 4.1 and 4.2, a 437 cgroup might have some charge associated with it, even though all 438 tasks have migrated away from it. (because we charge against pages, not 439 against tasks.) 440 441 We move the stats to root (if use_hierarchy==0) or parent (if 442 use_hierarchy==1), and no change on the charge except uncharging 443 from the child. 444 445 Charges recorded in swap information is not updated at removal of cgroup. 446 Recorded information is discarded and a cgroup which uses swap (swapcache) 447 will be charged as a new owner of it. 448 449 About use_hierarchy, see Section 6. 450 451 5. Misc. interfaces. 452 453 5.1 force_empty 454 memory.force_empty interface is provided to make cgroup's memory usage empty. 455 When writing anything to this 456 457 # echo 0 > memory.force_empty 458 459 the cgroup will be reclaimed and as many pages reclaimed as possible. 460 461 The typical use case for this interface is before calling rmdir(). 462 Because rmdir() moves all pages to parent, some out-of-use page caches can be 463 moved to the parent. If you want to avoid that, force_empty will be useful. 464 465 Also, note that when memory.kmem.limit_in_bytes is set the charges due to 466 kernel pages will still be seen. This is not considered a failure and the 467 write will still return success. In this case, it is expected that 468 memory.kmem.usage_in_bytes == memory.usage_in_bytes. 469 470 About use_hierarchy, see Section 6. 471 472 5.2 stat file 473 474 memory.stat file includes following statistics 475 476 # per-memory cgroup local status 477 cache - # of bytes of page cache memory. 478 rss - # of bytes of anonymous and swap cache memory (includes 479 transparent hugepages). 480 rss_huge - # of bytes of anonymous transparent hugepages. 481 mapped_file - # of bytes of mapped file (includes tmpfs/shmem) 482 pgpgin - # of charging events to the memory cgroup. The charging 483 event happens each time a page is accounted as either mapped 484 anon page(RSS) or cache page(Page Cache) to the cgroup. 485 pgpgout - # of uncharging events to the memory cgroup. The uncharging 486 event happens each time a page is unaccounted from the cgroup. 487 swap - # of bytes of swap usage 488 dirty - # of bytes that are waiting to get written back to the disk. 489 writeback - # of bytes of file/anon cache that are queued for syncing to 490 disk. 491 inactive_anon - # of bytes of anonymous and swap cache memory on inactive 492 LRU list. 493 active_anon - # of bytes of anonymous and swap cache memory on active 494 LRU list. 495 inactive_file - # of bytes of file-backed memory on inactive LRU list. 496 active_file - # of bytes of file-backed memory on active LRU list. 497 unevictable - # of bytes of memory that cannot be reclaimed (mlocked etc). 498 499 # status considering hierarchy (see memory.use_hierarchy settings) 500 501 hierarchical_memory_limit - # of bytes of memory limit with regard to hierarchy 502 under which the memory cgroup is 503 hierarchical_memsw_limit - # of bytes of memory+swap limit with regard to 504 hierarchy under which memory cgroup is. 505 506 total_<counter> - # hierarchical version of <counter>, which in 507 addition to the cgroup's own value includes the 508 sum of all hierarchical children's values of 509 <counter>, i.e. total_cache 510 511 # The following additional stats are dependent on CONFIG_DEBUG_VM. 512 513 recent_rotated_anon - VM internal parameter. (see mm/vmscan.c) 514 recent_rotated_file - VM internal parameter. (see mm/vmscan.c) 515 recent_scanned_anon - VM internal parameter. (see mm/vmscan.c) 516 recent_scanned_file - VM internal parameter. (see mm/vmscan.c) 517 518 Memo: 519 recent_rotated means recent frequency of LRU rotation. 520 recent_scanned means recent # of scans to LRU. 521 showing for better debug please see the code for meanings. 522 523 Note: 524 Only anonymous and swap cache memory is listed as part of 'rss' stat. 525 This should not be confused with the true 'resident set size' or the 526 amount of physical memory used by the cgroup. 527 'rss + mapped_file" will give you resident set size of cgroup. 528 (Note: file and shmem may be shared among other cgroups. In that case, 529 mapped_file is accounted only when the memory cgroup is owner of page 530 cache.) 531 532 5.3 swappiness 533 534 Overrides /proc/sys/vm/swappiness for the particular group. The tunable 535 in the root cgroup corresponds to the global swappiness setting. 536 537 Please note that unlike during the global reclaim, limit reclaim 538 enforces that 0 swappiness really prevents from any swapping even if 539 there is a swap storage available. This might lead to memcg OOM killer 540 if there are no file pages to reclaim. 541 542 5.4 failcnt 543 544 A memory cgroup provides memory.failcnt and memory.memsw.failcnt files. 545 This failcnt(== failure count) shows the number of times that a usage counter 546 hit its limit. When a memory cgroup hits a limit, failcnt increases and 547 memory under it will be reclaimed. 548 549 You can reset failcnt by writing 0 to failcnt file. 550 # echo 0 > .../memory.failcnt 551 552 5.5 usage_in_bytes 553 554 For efficiency, as other kernel components, memory cgroup uses some optimization 555 to avoid unnecessary cacheline false sharing. usage_in_bytes is affected by the 556 method and doesn't show 'exact' value of memory (and swap) usage, it's a fuzz 557 value for efficient access. (Of course, when necessary, it's synchronized.) 558 If you want to know more exact memory usage, you should use RSS+CACHE(+SWAP) 559 value in memory.stat(see 5.2). 560 561 5.6 numa_stat 562 563 This is similar to numa_maps but operates on a per-memcg basis. This is 564 useful for providing visibility into the numa locality information within 565 an memcg since the pages are allowed to be allocated from any physical 566 node. One of the use cases is evaluating application performance by 567 combining this information with the application's CPU allocation. 568 569 Each memcg's numa_stat file includes "total", "file", "anon" and "unevictable" 570 per-node page counts including "hierarchical_<counter>" which sums up all 571 hierarchical children's values in addition to the memcg's own value. 572 573 The output format of memory.numa_stat is: 574 575 total=<total pages> N0=<node 0 pages> N1=<node 1 pages> ... 576 file=<total file pages> N0=<node 0 pages> N1=<node 1 pages> ... 577 anon=<total anon pages> N0=<node 0 pages> N1=<node 1 pages> ... 578 unevictable=<total anon pages> N0=<node 0 pages> N1=<node 1 pages> ... 579 hierarchical_<counter>=<counter pages> N0=<node 0 pages> N1=<node 1 pages> ... 580 581 The "total" count is sum of file + anon + unevictable. 582 583 6. Hierarchy support 584 585 The memory controller supports a deep hierarchy and hierarchical accounting. 586 The hierarchy is created by creating the appropriate cgroups in the 587 cgroup filesystem. Consider for example, the following cgroup filesystem 588 hierarchy 589 590 root 591 / | \ 592 / | \ 593 a b c 594 | \ 595 | \ 596 d e 597 598 In the diagram above, with hierarchical accounting enabled, all memory 599 usage of e, is accounted to its ancestors up until the root (i.e, c and root), 600 that has memory.use_hierarchy enabled. If one of the ancestors goes over its 601 limit, the reclaim algorithm reclaims from the tasks in the ancestor and the 602 children of the ancestor. 603 604 6.1 Enabling hierarchical accounting and reclaim 605 606 A memory cgroup by default disables the hierarchy feature. Support 607 can be enabled by writing 1 to memory.use_hierarchy file of the root cgroup 608 609 # echo 1 > memory.use_hierarchy 610 611 The feature can be disabled by 612 613 # echo 0 > memory.use_hierarchy 614 615 NOTE1: Enabling/disabling will fail if either the cgroup already has other 616 cgroups created below it, or if the parent cgroup has use_hierarchy 617 enabled. 618 619 NOTE2: When panic_on_oom is set to "2", the whole system will panic in 620 case of an OOM event in any cgroup. 621 622 7. Soft limits 623 624 Soft limits allow for greater sharing of memory. The idea behind soft limits 625 is to allow control groups to use as much of the memory as needed, provided 626 627 a. There is no memory contention 628 b. They do not exceed their hard limit 629 630 When the system detects memory contention or low memory, control groups 631 are pushed back to their soft limits. If the soft limit of each control 632 group is very high, they are pushed back as much as possible to make 633 sure that one control group does not starve the others of memory. 634 635 Please note that soft limits is a best-effort feature; it comes with 636 no guarantees, but it does its best to make sure that when memory is 637 heavily contended for, memory is allocated based on the soft limit 638 hints/setup. Currently soft limit based reclaim is set up such that 639 it gets invoked from balance_pgdat (kswapd). 640 641 7.1 Interface 642 643 Soft limits can be setup by using the following commands (in this example we 644 assume a soft limit of 256 MiB) 645 646 # echo 256M > memory.soft_limit_in_bytes 647 648 If we want to change this to 1G, we can at any time use 649 650 # echo 1G > memory.soft_limit_in_bytes 651 652 NOTE1: Soft limits take effect over a long period of time, since they involve 653 reclaiming memory for balancing between memory cgroups 654 NOTE2: It is recommended to set the soft limit always below the hard limit, 655 otherwise the hard limit will take precedence. 656 657 8. Move charges at task migration 658 659 Users can move charges associated with a task along with task migration, that 660 is, uncharge task's pages from the old cgroup and charge them to the new cgroup. 661 This feature is not supported in !CONFIG_MMU environments because of lack of 662 page tables. 663 664 8.1 Interface 665 666 This feature is disabled by default. It can be enabled (and disabled again) by 667 writing to memory.move_charge_at_immigrate of the destination cgroup. 668 669 If you want to enable it: 670 671 # echo (some positive value) > memory.move_charge_at_immigrate 672 673 Note: Each bits of move_charge_at_immigrate has its own meaning about what type 674 of charges should be moved. See 8.2 for details. 675 Note: Charges are moved only when you move mm->owner, in other words, 676 a leader of a thread group. 677 Note: If we cannot find enough space for the task in the destination cgroup, we 678 try to make space by reclaiming memory. Task migration may fail if we 679 cannot make enough space. 680 Note: It can take several seconds if you move charges much. 681 682 And if you want disable it again: 683 684 # echo 0 > memory.move_charge_at_immigrate 685 686 8.2 Type of charges which can be moved 687 688 Each bit in move_charge_at_immigrate has its own meaning about what type of 689 charges should be moved. But in any case, it must be noted that an account of 690 a page or a swap can be moved only when it is charged to the task's current 691 (old) memory cgroup. 692 693 bit | what type of charges would be moved ? 694 -----+------------------------------------------------------------------------ 695 0 | A charge of an anonymous page (or swap of it) used by the target task. 696 | You must enable Swap Extension (see 2.4) to enable move of swap charges. 697 -----+------------------------------------------------------------------------ 698 1 | A charge of file pages (normal file, tmpfs file (e.g. ipc shared memory) 699 | and swaps of tmpfs file) mmapped by the target task. Unlike the case of 700 | anonymous pages, file pages (and swaps) in the range mmapped by the task 701 | will be moved even if the task hasn't done page fault, i.e. they might 702 | not be the task's "RSS", but other task's "RSS" that maps the same file. 703 | And mapcount of the page is ignored (the page can be moved even if 704 | page_mapcount(page) > 1). You must enable Swap Extension (see 2.4) to 705 | enable move of swap charges. 706 707 8.3 TODO 708 709 - All of moving charge operations are done under cgroup_mutex. It's not good 710 behavior to hold the mutex too long, so we may need some trick. 711 712 9. Memory thresholds 713 714 Memory cgroup implements memory thresholds using the cgroups notification 715 API (see cgroups.txt). It allows to register multiple memory and memsw 716 thresholds and gets notifications when it crosses. 717 718 To register a threshold, an application must: 719 - create an eventfd using eventfd(2); 720 - open memory.usage_in_bytes or memory.memsw.usage_in_bytes; 721 - write string like "<event_fd> <fd of memory.usage_in_bytes> <threshold>" to 722 cgroup.event_control. 723 724 Application will be notified through eventfd when memory usage crosses 725 threshold in any direction. 726 727 It's applicable for root and non-root cgroup. 728 729 10. OOM Control 730 731 memory.oom_control file is for OOM notification and other controls. 732 733 Memory cgroup implements OOM notifier using the cgroup notification 734 API (See cgroups.txt). It allows to register multiple OOM notification 735 delivery and gets notification when OOM happens. 736 737 To register a notifier, an application must: 738 - create an eventfd using eventfd(2) 739 - open memory.oom_control file 740 - write string like "<event_fd> <fd of memory.oom_control>" to 741 cgroup.event_control 742 743 The application will be notified through eventfd when OOM happens. 744 OOM notification doesn't work for the root cgroup. 745 746 You can disable the OOM-killer by writing "1" to memory.oom_control file, as: 747 748 #echo 1 > memory.oom_control 749 750 If OOM-killer is disabled, tasks under cgroup will hang/sleep 751 in memory cgroup's OOM-waitqueue when they request accountable memory. 752 753 For running them, you have to relax the memory cgroup's OOM status by 754 * enlarge limit or reduce usage. 755 To reduce usage, 756 * kill some tasks. 757 * move some tasks to other group with account migration. 758 * remove some files (on tmpfs?) 759 760 Then, stopped tasks will work again. 761 762 At reading, current status of OOM is shown. 763 oom_kill_disable 0 or 1 (if 1, oom-killer is disabled) 764 under_oom 0 or 1 (if 1, the memory cgroup is under OOM, tasks may 765 be stopped.) 766 767 11. Memory Pressure 768 769 The pressure level notifications can be used to monitor the memory 770 allocation cost; based on the pressure, applications can implement 771 different strategies of managing their memory resources. The pressure 772 levels are defined as following: 773 774 The "low" level means that the system is reclaiming memory for new 775 allocations. Monitoring this reclaiming activity might be useful for 776 maintaining cache level. Upon notification, the program (typically 777 "Activity Manager") might analyze vmstat and act in advance (i.e. 778 prematurely shutdown unimportant services). 779 780 The "medium" level means that the system is experiencing medium memory 781 pressure, the system might be making swap, paging out active file caches, 782 etc. Upon this event applications may decide to further analyze 783 vmstat/zoneinfo/memcg or internal memory usage statistics and free any 784 resources that can be easily reconstructed or re-read from a disk. 785 786 The "critical" level means that the system is actively thrashing, it is 787 about to out of memory (OOM) or even the in-kernel OOM killer is on its 788 way to trigger. Applications should do whatever they can to help the 789 system. It might be too late to consult with vmstat or any other 790 statistics, so it's advisable to take an immediate action. 791 792 By default, events are propagated upward until the event is handled, i.e. the 793 events are not pass-through. For example, you have three cgroups: A->B->C. Now 794 you set up an event listener on cgroups A, B and C, and suppose group C 795 experiences some pressure. In this situation, only group C will receive the 796 notification, i.e. groups A and B will not receive it. This is done to avoid 797 excessive "broadcasting" of messages, which disturbs the system and which is 798 especially bad if we are low on memory or thrashing. Group B, will receive 799 notification only if there are no event listers for group C. 800 801 There are three optional modes that specify different propagation behavior: 802 803 - "default": this is the default behavior specified above. This mode is the 804 same as omitting the optional mode parameter, preserved by backwards 805 compatibility. 806 807 - "hierarchy": events always propagate up to the root, similar to the default 808 behavior, except that propagation continues regardless of whether there are 809 event listeners at each level, with the "hierarchy" mode. In the above 810 example, groups A, B, and C will receive notification of memory pressure. 811 812 - "local": events are pass-through, i.e. they only receive notifications when 813 memory pressure is experienced in the memcg for which the notification is 814 registered. In the above example, group C will receive notification if 815 registered for "local" notification and the group experiences memory 816 pressure. However, group B will never receive notification, regardless if 817 there is an event listener for group C or not, if group B is registered for 818 local notification. 819 820 The level and event notification mode ("hierarchy" or "local", if necessary) are 821 specified by a comma-delimited string, i.e. "low,hierarchy" specifies 822 hierarchical, pass-through, notification for all ancestor memcgs. Notification 823 that is the default, non pass-through behavior, does not specify a mode. 824 "medium,local" specifies pass-through notification for the medium level. 825 826 The file memory.pressure_level is only used to setup an eventfd. To 827 register a notification, an application must: 828 829 - create an eventfd using eventfd(2); 830 - open memory.pressure_level; 831 - write string as "<event_fd> <fd of memory.pressure_level> <level[,mode]>" 832 to cgroup.event_control. 833 834 Application will be notified through eventfd when memory pressure is at 835 the specific level (or higher). Read/write operations to 836 memory.pressure_level are no implemented. 837 838 Test: 839 840 Here is a small script example that makes a new cgroup, sets up a 841 memory limit, sets up a notification in the cgroup and then makes child 842 cgroup experience a critical pressure: 843 844 # cd /sys/fs/cgroup/memory/ 845 # mkdir foo 846 # cd foo 847 # cgroup_event_listener memory.pressure_level low,hierarchy & 848 # echo 8000000 > memory.limit_in_bytes 849 # echo 8000000 > memory.memsw.limit_in_bytes 850 # echo $$ > tasks 851 # dd if=/dev/zero | read x 852 853 (Expect a bunch of notifications, and eventually, the oom-killer will 854 trigger.) 855 856 12. TODO 857 858 1. Make per-cgroup scanner reclaim not-shared pages first 859 2. Teach controller to account for shared-pages 860 3. Start reclamation in the background when the limit is 861 not yet hit but the usage is getting closer 862 863 Summary 864 865 Overall, the memory controller has been a stable controller and has been 866 commented and discussed quite extensively in the community. 867 868 References 869 870 1. Singh, Balbir. RFC: Memory Controller, http://lwn.net/Articles/206697/ 871 2. Singh, Balbir. Memory Controller (RSS Control), 872 http://lwn.net/Articles/222762/ 873 3. Emelianov, Pavel. Resource controllers based on process cgroups 874 http://lkml.org/lkml/2007/3/6/198 875 4. Emelianov, Pavel. RSS controller based on process cgroups (v2) 876 http://lkml.org/lkml/2007/4/9/78 877 5. Emelianov, Pavel. RSS controller based on process cgroups (v3) 878 http://lkml.org/lkml/2007/5/30/244 879 6. Menage, Paul. Control Groups v10, http://lwn.net/Articles/236032/ 880 7. Vaidyanathan, Srinivasan, Control Groups: Pagecache accounting and control 881 subsystem (v3), http://lwn.net/Articles/235534/ 882 8. Singh, Balbir. RSS controller v2 test results (lmbench), 883 http://lkml.org/lkml/2007/5/17/232 884 9. Singh, Balbir. RSS controller v2 AIM9 results 885 http://lkml.org/lkml/2007/5/18/1 886 10. Singh, Balbir. Memory controller v6 test results, 887 http://lkml.org/lkml/2007/8/19/36 888 11. Singh, Balbir. Memory controller introduction (v6), 889 http://lkml.org/lkml/2007/8/17/69 890 12. Corbet, Jonathan, Controlling memory use in cgroups, 891 http://lwn.net/Articles/243795/