linux/Documentation/cgroups/memory.txt
<<
>>
Prefs
   1Memory Resource Controller
   2
   3NOTE: The Memory Resource Controller has generically been referred to as the
   4      memory controller in this document. Do not confuse memory controller
   5      used here with the memory controller that is used in hardware.
   6
   7(For editors)
   8In this document:
   9      When we mention a cgroup (cgroupfs's directory) with memory controller,
  10      we call it "memory cgroup". When you see git-log and source code, you'll
  11      see patch's title and function names tend to use "memcg".
  12      In this document, we avoid using it.
  13
  14Benefits and Purpose of the memory controller
  15
  16The memory controller isolates the memory behaviour of a group of tasks
  17from the rest of the system. The article on LWN [12] mentions some probable
  18uses of the memory controller. The memory controller can be used to
  19
  20a. Isolate an application or a group of applications
  21   Memory-hungry applications can be isolated and limited to a smaller
  22   amount of memory.
  23b. Create a cgroup with a limited amount of memory; this can be used
  24   as a good alternative to booting with mem=XXXX.
  25c. Virtualization solutions can control the amount of memory they want
  26   to assign to a virtual machine instance.
  27d. A CD/DVD burner could control the amount of memory used by the
  28   rest of the system to ensure that burning does not fail due to lack
  29   of available memory.
  30e. There are several other use cases; find one or use the controller just
  31   for fun (to learn and hack on the VM subsystem).
  32
  33Current Status: linux-2.6.34-mmotm(development version of 2010/April)
  34
  35Features:
  36 - accounting anonymous pages, file caches, swap caches usage and limiting them.
  37 - pages are linked to per-memcg LRU exclusively, and there is no global LRU.
  38 - optionally, memory+swap usage can be accounted and limited.
  39 - hierarchical accounting
  40 - soft limit
  41 - moving (recharging) account at moving a task is selectable.
  42 - usage threshold notifier
  43 - memory pressure notifier
  44 - oom-killer disable knob and oom-notifier
  45 - Root cgroup has no limit controls.
  46
  47 Kernel memory support is a work in progress, and the current version provides
  48 basically functionality. (See Section 2.7)
  49
  50Brief summary of control files.
  51
  52 tasks                           # attach a task(thread) and show list of threads
  53 cgroup.procs                    # show list of processes
  54 cgroup.event_control            # an interface for event_fd()
  55 memory.usage_in_bytes           # show current res_counter usage for memory
  56                                 (See 5.5 for details)
  57 memory.memsw.usage_in_bytes     # show current res_counter usage for memory+Swap
  58                                 (See 5.5 for details)
  59 memory.limit_in_bytes           # set/show limit of memory usage
  60 memory.memsw.limit_in_bytes     # set/show limit of memory+Swap usage
  61 memory.failcnt                  # show the number of memory usage hits limits
  62 memory.memsw.failcnt            # show the number of memory+Swap hits limits
  63 memory.max_usage_in_bytes       # show max memory usage recorded
  64 memory.memsw.max_usage_in_bytes # show max memory+Swap usage recorded
  65 memory.soft_limit_in_bytes      # set/show soft limit of memory usage
  66 memory.stat                     # show various statistics
  67 memory.use_hierarchy            # set/show hierarchical account enabled
  68 memory.force_empty              # trigger forced move charge to parent
  69 memory.pressure_level           # set memory pressure notifications
  70 memory.swappiness               # set/show swappiness parameter of vmscan
  71                                 (See sysctl's vm.swappiness)
  72 memory.move_charge_at_immigrate # set/show controls of moving charges
  73 memory.oom_control              # set/show oom controls.
  74 memory.numa_stat                # show the number of memory usage per numa node
  75
  76 memory.kmem.limit_in_bytes      # set/show hard limit for kernel memory
  77 memory.kmem.usage_in_bytes      # show current kernel memory allocation
  78 memory.kmem.failcnt             # show the number of kernel memory usage hits limits
  79 memory.kmem.max_usage_in_bytes  # show max kernel memory usage recorded
  80
  81 memory.kmem.tcp.limit_in_bytes  # set/show hard limit for tcp buf memory
  82 memory.kmem.tcp.usage_in_bytes  # show current tcp buf memory allocation
  83 memory.kmem.tcp.failcnt            # show the number of tcp buf memory usage hits limits
  84 memory.kmem.tcp.max_usage_in_bytes # show max tcp buf memory usage recorded
  85
  861. History
  87
  88The memory controller has a long history. A request for comments for the memory
  89controller was posted by Balbir Singh [1]. At the time the RFC was posted
  90there were several implementations for memory control. The goal of the
  91RFC was to build consensus and agreement for the minimal features required
  92for memory control. The first RSS controller was posted by Balbir Singh[2]
  93in Feb 2007. Pavel Emelianov [3][4][5] has since posted three versions of the
  94RSS controller. At OLS, at the resource management BoF, everyone suggested
  95that we handle both page cache and RSS together. Another request was raised
  96to allow user space handling of OOM. The current memory controller is
  97at version 6; it combines both mapped (RSS) and unmapped Page
  98Cache Control [11].
  99
 1002. Memory Control
 101
 102Memory is a unique resource in the sense that it is present in a limited
 103amount. If a task requires a lot of CPU processing, the task can spread
 104its processing over a period of hours, days, months or years, but with
 105memory, the same physical memory needs to be reused to accomplish the task.
 106
 107The memory controller implementation has been divided into phases. These
 108are:
 109
 1101. Memory controller
 1112. mlock(2) controller
 1123. Kernel user memory accounting and slab control
 1134. user mappings length controller
 114
 115The memory controller is the first controller developed.
 116
 1172.1. Design
 118
 119The core of the design is a counter called the res_counter. The res_counter
 120tracks the current memory usage and limit of the group of processes associated
 121with the controller. Each cgroup has a memory controller specific data
 122structure (mem_cgroup) associated with it.
 123
 1242.2. Accounting
 125
 126                +--------------------+
 127                |  mem_cgroup     |
 128                |  (res_counter)     |
 129                +--------------------+
 130                 /            ^      \
 131                /             |       \
 132           +---------------+  |        +---------------+
 133           | mm_struct     |  |....    | mm_struct     |
 134           |               |  |        |               |
 135           +---------------+  |        +---------------+
 136                              |
 137                              + --------------+
 138                                              |
 139           +---------------+           +------+--------+
 140           | page          +---------->  page_cgroup|
 141           |               |           |               |
 142           +---------------+           +---------------+
 143
 144             (Figure 1: Hierarchy of Accounting)
 145
 146
 147Figure 1 shows the important aspects of the controller
 148
 1491. Accounting happens per cgroup
 1502. Each mm_struct knows about which cgroup it belongs to
 1513. Each page has a pointer to the page_cgroup, which in turn knows the
 152   cgroup it belongs to
 153
 154The accounting is done as follows: mem_cgroup_charge_common() is invoked to
 155set up the necessary data structures and check if the cgroup that is being
 156charged is over its limit. If it is, then reclaim is invoked on the cgroup.
 157More details can be found in the reclaim section of this document.
 158If everything goes well, a page meta-data-structure called page_cgroup is
 159updated. page_cgroup has its own LRU on cgroup.
 160(*) page_cgroup structure is allocated at boot/memory-hotplug time.
 161
 1622.2.1 Accounting details
 163
 164All mapped anon pages (RSS) and cache pages (Page Cache) are accounted.
 165Some pages which are never reclaimable and will not be on the LRU
 166are not accounted. We just account pages under usual VM management.
 167
 168RSS pages are accounted at page_fault unless they've already been accounted
 169for earlier. A file page will be accounted for as Page Cache when it's
 170inserted into inode (radix-tree). While it's mapped into the page tables of
 171processes, duplicate accounting is carefully avoided.
 172
 173An RSS page is unaccounted when it's fully unmapped. A PageCache page is
 174unaccounted when it's removed from radix-tree. Even if RSS pages are fully
 175unmapped (by kswapd), they may exist as SwapCache in the system until they
 176are really freed. Such SwapCaches are also accounted.
 177A swapped-in page is not accounted until it's mapped.
 178
 179Note: The kernel does swapin-readahead and reads multiple swaps at once.
 180This means swapped-in pages may contain pages for other tasks than a task
 181causing page fault. So, we avoid accounting at swap-in I/O.
 182
 183At page migration, accounting information is kept.
 184
 185Note: we just account pages-on-LRU because our purpose is to control amount
 186of used pages; not-on-LRU pages tend to be out-of-control from VM view.
 187
 1882.3 Shared Page Accounting
 189
 190Shared pages are accounted on the basis of the first touch approach. The
 191cgroup that first touches a page is accounted for the page. The principle
 192behind this approach is that a cgroup that aggressively uses a shared
 193page will eventually get charged for it (once it is uncharged from
 194the cgroup that brought it in -- this will happen on memory pressure).
 195
 196But see section 8.2: when moving a task to another cgroup, its pages may
 197be recharged to the new cgroup, if move_charge_at_immigrate has been chosen.
 198
 199Exception: If CONFIG_MEMCG_SWAP is not used.
 200When you do swapoff and make swapped-out pages of shmem(tmpfs) to
 201be backed into memory in force, charges for pages are accounted against the
 202caller of swapoff rather than the users of shmem.
 203
 2042.4 Swap Extension (CONFIG_MEMCG_SWAP)
 205
 206Swap Extension allows you to record charge for swap. A swapped-in page is
 207charged back to original page allocator if possible.
 208
 209When swap is accounted, following files are added.
 210 - memory.memsw.usage_in_bytes.
 211 - memory.memsw.limit_in_bytes.
 212
 213memsw means memory+swap. Usage of memory+swap is limited by
 214memsw.limit_in_bytes.
 215
 216Example: Assume a system with 4G of swap. A task which allocates 6G of memory
 217(by mistake) under 2G memory limitation will use all swap.
 218In this case, setting memsw.limit_in_bytes=3G will prevent bad use of swap.
 219By using the memsw limit, you can avoid system OOM which can be caused by swap
 220shortage.
 221
 222* why 'memory+swap' rather than swap.
 223The global LRU(kswapd) can swap out arbitrary pages. Swap-out means
 224to move account from memory to swap...there is no change in usage of
 225memory+swap. In other words, when we want to limit the usage of swap without
 226affecting global LRU, memory+swap limit is better than just limiting swap from
 227an OS point of view.
 228
 229* What happens when a cgroup hits memory.memsw.limit_in_bytes
 230When a cgroup hits memory.memsw.limit_in_bytes, it's useless to do swap-out
 231in this cgroup. Then, swap-out will not be done by cgroup routine and file
 232caches are dropped. But as mentioned above, global LRU can do swapout memory
 233from it for sanity of the system's memory management state. You can't forbid
 234it by cgroup.
 235
 2362.5 Reclaim
 237
 238Each cgroup maintains a per cgroup LRU which has the same structure as
 239global VM. When a cgroup goes over its limit, we first try
 240to reclaim memory from the cgroup so as to make space for the new
 241pages that the cgroup has touched. If the reclaim is unsuccessful,
 242an OOM routine is invoked to select and kill the bulkiest task in the
 243cgroup. (See 10. OOM Control below.)
 244
 245The reclaim algorithm has not been modified for cgroups, except that
 246pages that are selected for reclaiming come from the per-cgroup LRU
 247list.
 248
 249NOTE: Reclaim does not work for the root cgroup, since we cannot set any
 250limits on the root cgroup.
 251
 252Note2: When panic_on_oom is set to "2", the whole system will panic.
 253
 254When oom event notifier is registered, event will be delivered.
 255(See oom_control section)
 256
 2572.6 Locking
 258
 259   lock_page_cgroup()/unlock_page_cgroup() should not be called under
 260   mapping->tree_lock.
 261
 262   Other lock order is following:
 263   PG_locked.
 264   mm->page_table_lock
 265       zone->lru_lock
 266          lock_page_cgroup.
 267  In many cases, just lock_page_cgroup() is called.
 268  per-zone-per-cgroup LRU (cgroup's private LRU) is just guarded by
 269  zone->lru_lock, it has no lock of its own.
 270
 2712.7 Kernel Memory Extension (CONFIG_MEMCG_KMEM)
 272
 273WARNING: Current implementation lacks reclaim support. That means allocation
 274         attempts will fail when close to the limit even if there are plenty of
 275         kmem available for reclaim. That makes this option unusable in real
 276         life so DO NOT SELECT IT unless for development purposes.
 277
 278With the Kernel memory extension, the Memory Controller is able to limit
 279the amount of kernel memory used by the system. Kernel memory is fundamentally
 280different than user memory, since it can't be swapped out, which makes it
 281possible to DoS the system by consuming too much of this precious resource.
 282
 283Kernel memory won't be accounted at all until limit on a group is set. This
 284allows for existing setups to continue working without disruption.  The limit
 285cannot be set if the cgroup have children, or if there are already tasks in the
 286cgroup. Attempting to set the limit under those conditions will return -EBUSY.
 287When use_hierarchy == 1 and a group is accounted, its children will
 288automatically be accounted regardless of their limit value.
 289
 290After a group is first limited, it will be kept being accounted until it
 291is removed. The memory limitation itself, can of course be removed by writing
 292-1 to memory.kmem.limit_in_bytes. In this case, kmem will be accounted, but not
 293limited.
 294
 295Kernel memory limits are not imposed for the root cgroup. Usage for the root
 296cgroup may or may not be accounted. The memory used is accumulated into
 297memory.kmem.usage_in_bytes, or in a separate counter when it makes sense.
 298(currently only for tcp).
 299The main "kmem" counter is fed into the main counter, so kmem charges will
 300also be visible from the user counter.
 301
 302Currently no soft limit is implemented for kernel memory. It is future work
 303to trigger slab reclaim when those limits are reached.
 304
 3052.7.1 Current Kernel Memory resources accounted
 306
 307* stack pages: every process consumes some stack pages. By accounting into
 308kernel memory, we prevent new processes from being created when the kernel
 309memory usage is too high.
 310
 311* slab pages: pages allocated by the SLAB or SLUB allocator are tracked. A copy
 312of each kmem_cache is created every time the cache is touched by the first time
 313from inside the memcg. The creation is done lazily, so some objects can still be
 314skipped while the cache is being created. All objects in a slab page should
 315belong to the same memcg. This only fails to hold when a task is migrated to a
 316different memcg during the page allocation by the cache.
 317
 318* sockets memory pressure: some sockets protocols have memory pressure
 319thresholds. The Memory Controller allows them to be controlled individually
 320per cgroup, instead of globally.
 321
 322* tcp memory pressure: sockets memory pressure for the tcp protocol.
 323
 3242.7.3 Common use cases
 325
 326Because the "kmem" counter is fed to the main user counter, kernel memory can
 327never be limited completely independently of user memory. Say "U" is the user
 328limit, and "K" the kernel limit. There are three possible ways limits can be
 329set:
 330
 331    U != 0, K = unlimited:
 332    This is the standard memcg limitation mechanism already present before kmem
 333    accounting. Kernel memory is completely ignored.
 334
 335    U != 0, K < U:
 336    Kernel memory is a subset of the user memory. This setup is useful in
 337    deployments where the total amount of memory per-cgroup is overcommited.
 338    Overcommiting kernel memory limits is definitely not recommended, since the
 339    box can still run out of non-reclaimable memory.
 340    In this case, the admin could set up K so that the sum of all groups is
 341    never greater than the total memory, and freely set U at the cost of his
 342    QoS.
 343
 344    U != 0, K >= U:
 345    Since kmem charges will also be fed to the user counter and reclaim will be
 346    triggered for the cgroup for both kinds of memory. This setup gives the
 347    admin a unified view of memory, and it is also useful for people who just
 348    want to track kernel memory usage.
 349
 3503. User Interface
 351
 3520. Configuration
 353
 354a. Enable CONFIG_CGROUPS
 355b. Enable CONFIG_RESOURCE_COUNTERS
 356c. Enable CONFIG_MEMCG
 357d. Enable CONFIG_MEMCG_SWAP (to use swap extension)
 358d. Enable CONFIG_MEMCG_KMEM (to use kmem extension)
 359
 3601. Prepare the cgroups (see cgroups.txt, Why are cgroups needed?)
 361# mount -t tmpfs none /sys/fs/cgroup
 362# mkdir /sys/fs/cgroup/memory
 363# mount -t cgroup none /sys/fs/cgroup/memory -o memory
 364
 3652. Make the new group and move bash into it
 366# mkdir /sys/fs/cgroup/memory/0
 367# echo $$ > /sys/fs/cgroup/memory/0/tasks
 368
 369Since now we're in the 0 cgroup, we can alter the memory limit:
 370# echo 4M > /sys/fs/cgroup/memory/0/memory.limit_in_bytes
 371
 372NOTE: We can use a suffix (k, K, m, M, g or G) to indicate values in kilo,
 373mega or gigabytes. (Here, Kilo, Mega, Giga are Kibibytes, Mebibytes, Gibibytes.)
 374
 375NOTE: We can write "-1" to reset the *.limit_in_bytes(unlimited).
 376NOTE: We cannot set limits on the root cgroup any more.
 377
 378# cat /sys/fs/cgroup/memory/0/memory.limit_in_bytes
 3794194304
 380
 381We can check the usage:
 382# cat /sys/fs/cgroup/memory/0/memory.usage_in_bytes
 3831216512
 384
 385A successful write to this file does not guarantee a successful setting of
 386this limit to the value written into the file. This can be due to a
 387number of factors, such as rounding up to page boundaries or the total
 388availability of memory on the system. The user is required to re-read
 389this file after a write to guarantee the value committed by the kernel.
 390
 391# echo 1 > memory.limit_in_bytes
 392# cat memory.limit_in_bytes
 3934096
 394
 395The memory.failcnt field gives the number of times that the cgroup limit was
 396exceeded.
 397
 398The memory.stat file gives accounting information. Now, the number of
 399caches, RSS and Active pages/Inactive pages are shown.
 400
 4014. Testing
 402
 403For testing features and implementation, see memcg_test.txt.
 404
 405Performance test is also important. To see pure memory controller's overhead,
 406testing on tmpfs will give you good numbers of small overheads.
 407Example: do kernel make on tmpfs.
 408
 409Page-fault scalability is also important. At measuring parallel
 410page fault test, multi-process test may be better than multi-thread
 411test because it has noise of shared objects/status.
 412
 413But the above two are testing extreme situations.
 414Trying usual test under memory controller is always helpful.
 415
 4164.1 Troubleshooting
 417
 418Sometimes a user might find that the application under a cgroup is
 419terminated by the OOM killer. There are several causes for this:
 420
 4211. The cgroup limit is too low (just too low to do anything useful)
 4222. The user is using anonymous memory and swap is turned off or too low
 423
 424A sync followed by echo 1 > /proc/sys/vm/drop_caches will help get rid of
 425some of the pages cached in the cgroup (page cache pages).
 426
 427To know what happens, disabling OOM_Kill as per "10. OOM Control" (below) and
 428seeing what happens will be helpful.
 429
 4304.2 Task migration
 431
 432When a task migrates from one cgroup to another, its charge is not
 433carried forward by default. The pages allocated from the original cgroup still
 434remain charged to it, the charge is dropped when the page is freed or
 435reclaimed.
 436
 437You can move charges of a task along with task migration.
 438See 8. "Move charges at task migration"
 439
 4404.3 Removing a cgroup
 441
 442A cgroup can be removed by rmdir, but as discussed in sections 4.1 and 4.2, a
 443cgroup might have some charge associated with it, even though all
 444tasks have migrated away from it. (because we charge against pages, not
 445against tasks.)
 446
 447We move the stats to root (if use_hierarchy==0) or parent (if
 448use_hierarchy==1), and no change on the charge except uncharging
 449from the child.
 450
 451Charges recorded in swap information is not updated at removal of cgroup.
 452Recorded information is discarded and a cgroup which uses swap (swapcache)
 453will be charged as a new owner of it.
 454
 455About use_hierarchy, see Section 6.
 456
 4575. Misc. interfaces.
 458
 4595.1 force_empty
 460  memory.force_empty interface is provided to make cgroup's memory usage empty.
 461  When writing anything to this
 462
 463  # echo 0 > memory.force_empty
 464
 465  the cgroup will be reclaimed and as many pages reclaimed as possible.
 466
 467  The typical use case for this interface is before calling rmdir().
 468  Because rmdir() moves all pages to parent, some out-of-use page caches can be
 469  moved to the parent. If you want to avoid that, force_empty will be useful.
 470
 471  Also, note that when memory.kmem.limit_in_bytes is set the charges due to
 472  kernel pages will still be seen. This is not considered a failure and the
 473  write will still return success. In this case, it is expected that
 474  memory.kmem.usage_in_bytes == memory.usage_in_bytes.
 475
 476  About use_hierarchy, see Section 6.
 477
 4785.2 stat file
 479
 480memory.stat file includes following statistics
 481
 482# per-memory cgroup local status
 483cache           - # of bytes of page cache memory.
 484rss             - # of bytes of anonymous and swap cache memory (includes
 485                transparent hugepages).
 486rss_huge        - # of bytes of anonymous transparent hugepages.
 487mapped_file     - # of bytes of mapped file (includes tmpfs/shmem)
 488pgpgin          - # of charging events to the memory cgroup. The charging
 489                event happens each time a page is accounted as either mapped
 490                anon page(RSS) or cache page(Page Cache) to the cgroup.
 491pgpgout         - # of uncharging events to the memory cgroup. The uncharging
 492                event happens each time a page is unaccounted from the cgroup.
 493swap            - # of bytes of swap usage
 494writeback       - # of bytes of file/anon cache that are queued for syncing to
 495                disk.
 496inactive_anon   - # of bytes of anonymous and swap cache memory on inactive
 497                LRU list.
 498active_anon     - # of bytes of anonymous and swap cache memory on active
 499                LRU list.
 500inactive_file   - # of bytes of file-backed memory on inactive LRU list.
 501active_file     - # of bytes of file-backed memory on active LRU list.
 502unevictable     - # of bytes of memory that cannot be reclaimed (mlocked etc).
 503
 504# status considering hierarchy (see memory.use_hierarchy settings)
 505
 506hierarchical_memory_limit - # of bytes of memory limit with regard to hierarchy
 507                        under which the memory cgroup is
 508hierarchical_memsw_limit - # of bytes of memory+swap limit with regard to
 509                        hierarchy under which memory cgroup is.
 510
 511total_<counter>         - # hierarchical version of <counter>, which in
 512                        addition to the cgroup's own value includes the
 513                        sum of all hierarchical children's values of
 514                        <counter>, i.e. total_cache
 515
 516# The following additional stats are dependent on CONFIG_DEBUG_VM.
 517
 518recent_rotated_anon     - VM internal parameter. (see mm/vmscan.c)
 519recent_rotated_file     - VM internal parameter. (see mm/vmscan.c)
 520recent_scanned_anon     - VM internal parameter. (see mm/vmscan.c)
 521recent_scanned_file     - VM internal parameter. (see mm/vmscan.c)
 522
 523Memo:
 524        recent_rotated means recent frequency of LRU rotation.
 525        recent_scanned means recent # of scans to LRU.
 526        showing for better debug please see the code for meanings.
 527
 528Note:
 529        Only anonymous and swap cache memory is listed as part of 'rss' stat.
 530        This should not be confused with the true 'resident set size' or the
 531        amount of physical memory used by the cgroup.
 532        'rss + file_mapped" will give you resident set size of cgroup.
 533        (Note: file and shmem may be shared among other cgroups. In that case,
 534         file_mapped is accounted only when the memory cgroup is owner of page
 535         cache.)
 536
 5375.3 swappiness
 538
 539Overrides /proc/sys/vm/swappiness for the particular group. The tunable
 540in the root cgroup corresponds to the global swappiness setting.
 541
 542Please note that unlike during the global reclaim, limit reclaim
 543enforces that 0 swappiness really prevents from any swapping even if
 544there is a swap storage available. This might lead to memcg OOM killer
 545if there are no file pages to reclaim.
 546
 5475.4 failcnt
 548
 549A memory cgroup provides memory.failcnt and memory.memsw.failcnt files.
 550This failcnt(== failure count) shows the number of times that a usage counter
 551hit its limit. When a memory cgroup hits a limit, failcnt increases and
 552memory under it will be reclaimed.
 553
 554You can reset failcnt by writing 0 to failcnt file.
 555# echo 0 > .../memory.failcnt
 556
 5575.5 usage_in_bytes
 558
 559For efficiency, as other kernel components, memory cgroup uses some optimization
 560to avoid unnecessary cacheline false sharing. usage_in_bytes is affected by the
 561method and doesn't show 'exact' value of memory (and swap) usage, it's a fuzz
 562value for efficient access. (Of course, when necessary, it's synchronized.)
 563If you want to know more exact memory usage, you should use RSS+CACHE(+SWAP)
 564value in memory.stat(see 5.2).
 565
 5665.6 numa_stat
 567
 568This is similar to numa_maps but operates on a per-memcg basis.  This is
 569useful for providing visibility into the numa locality information within
 570an memcg since the pages are allowed to be allocated from any physical
 571node.  One of the use cases is evaluating application performance by
 572combining this information with the application's CPU allocation.
 573
 574Each memcg's numa_stat file includes "total", "file", "anon" and "unevictable"
 575per-node page counts including "hierarchical_<counter>" which sums up all
 576hierarchical children's values in addition to the memcg's own value.
 577
 578The output format of memory.numa_stat is:
 579
 580total=<total pages> N0=<node 0 pages> N1=<node 1 pages> ...
 581file=<total file pages> N0=<node 0 pages> N1=<node 1 pages> ...
 582anon=<total anon pages> N0=<node 0 pages> N1=<node 1 pages> ...
 583unevictable=<total anon pages> N0=<node 0 pages> N1=<node 1 pages> ...
 584hierarchical_<counter>=<counter pages> N0=<node 0 pages> N1=<node 1 pages> ...
 585
 586The "total" count is sum of file + anon + unevictable.
 587
 5886. Hierarchy support
 589
 590The memory controller supports a deep hierarchy and hierarchical accounting.
 591The hierarchy is created by creating the appropriate cgroups in the
 592cgroup filesystem. Consider for example, the following cgroup filesystem
 593hierarchy
 594
 595               root
 596             /  |   \
 597            /   |    \
 598           a    b     c
 599                      | \
 600                      |  \
 601                      d   e
 602
 603In the diagram above, with hierarchical accounting enabled, all memory
 604usage of e, is accounted to its ancestors up until the root (i.e, c and root),
 605that has memory.use_hierarchy enabled. If one of the ancestors goes over its
 606limit, the reclaim algorithm reclaims from the tasks in the ancestor and the
 607children of the ancestor.
 608
 6096.1 Enabling hierarchical accounting and reclaim
 610
 611A memory cgroup by default disables the hierarchy feature. Support
 612can be enabled by writing 1 to memory.use_hierarchy file of the root cgroup
 613
 614# echo 1 > memory.use_hierarchy
 615
 616The feature can be disabled by
 617
 618# echo 0 > memory.use_hierarchy
 619
 620NOTE1: Enabling/disabling will fail if either the cgroup already has other
 621       cgroups created below it, or if the parent cgroup has use_hierarchy
 622       enabled.
 623
 624NOTE2: When panic_on_oom is set to "2", the whole system will panic in
 625       case of an OOM event in any cgroup.
 626
 6277. Soft limits
 628
 629Soft limits allow for greater sharing of memory. The idea behind soft limits
 630is to allow control groups to use as much of the memory as needed, provided
 631
 632a. There is no memory contention
 633b. They do not exceed their hard limit
 634
 635When the system detects memory contention or low memory, control groups
 636are pushed back to their soft limits. If the soft limit of each control
 637group is very high, they are pushed back as much as possible to make
 638sure that one control group does not starve the others of memory.
 639
 640Please note that soft limits is a best-effort feature; it comes with
 641no guarantees, but it does its best to make sure that when memory is
 642heavily contended for, memory is allocated based on the soft limit
 643hints/setup. Currently soft limit based reclaim is set up such that
 644it gets invoked from balance_pgdat (kswapd).
 645
 6467.1 Interface
 647
 648Soft limits can be setup by using the following commands (in this example we
 649assume a soft limit of 256 MiB)
 650
 651# echo 256M > memory.soft_limit_in_bytes
 652
 653If we want to change this to 1G, we can at any time use
 654
 655# echo 1G > memory.soft_limit_in_bytes
 656
 657NOTE1: Soft limits take effect over a long period of time, since they involve
 658       reclaiming memory for balancing between memory cgroups
 659NOTE2: It is recommended to set the soft limit always below the hard limit,
 660       otherwise the hard limit will take precedence.
 661
 6628. Move charges at task migration
 663
 664Users can move charges associated with a task along with task migration, that
 665is, uncharge task's pages from the old cgroup and charge them to the new cgroup.
 666This feature is not supported in !CONFIG_MMU environments because of lack of
 667page tables.
 668
 6698.1 Interface
 670
 671This feature is disabled by default. It can be enabled (and disabled again) by
 672writing to memory.move_charge_at_immigrate of the destination cgroup.
 673
 674If you want to enable it:
 675
 676# echo (some positive value) > memory.move_charge_at_immigrate
 677
 678Note: Each bits of move_charge_at_immigrate has its own meaning about what type
 679      of charges should be moved. See 8.2 for details.
 680Note: Charges are moved only when you move mm->owner, in other words,
 681      a leader of a thread group.
 682Note: If we cannot find enough space for the task in the destination cgroup, we
 683      try to make space by reclaiming memory. Task migration may fail if we
 684      cannot make enough space.
 685Note: It can take several seconds if you move charges much.
 686
 687And if you want disable it again:
 688
 689# echo 0 > memory.move_charge_at_immigrate
 690
 6918.2 Type of charges which can be moved
 692
 693Each bit in move_charge_at_immigrate has its own meaning about what type of
 694charges should be moved. But in any case, it must be noted that an account of
 695a page or a swap can be moved only when it is charged to the task's current
 696(old) memory cgroup.
 697
 698  bit | what type of charges would be moved ?
 699 -----+------------------------------------------------------------------------
 700   0  | A charge of an anonymous page (or swap of it) used by the target task.
 701      | You must enable Swap Extension (see 2.4) to enable move of swap charges.
 702 -----+------------------------------------------------------------------------
 703   1  | A charge of file pages (normal file, tmpfs file (e.g. ipc shared memory)
 704      | and swaps of tmpfs file) mmapped by the target task. Unlike the case of
 705      | anonymous pages, file pages (and swaps) in the range mmapped by the task
 706      | will be moved even if the task hasn't done page fault, i.e. they might
 707      | not be the task's "RSS", but other task's "RSS" that maps the same file.
 708      | And mapcount of the page is ignored (the page can be moved even if
 709      | page_mapcount(page) > 1). You must enable Swap Extension (see 2.4) to
 710      | enable move of swap charges.
 711
 7128.3 TODO
 713
 714- All of moving charge operations are done under cgroup_mutex. It's not good
 715  behavior to hold the mutex too long, so we may need some trick.
 716
 7179. Memory thresholds
 718
 719Memory cgroup implements memory thresholds using the cgroups notification
 720API (see cgroups.txt). It allows to register multiple memory and memsw
 721thresholds and gets notifications when it crosses.
 722
 723To register a threshold, an application must:
 724- create an eventfd using eventfd(2);
 725- open memory.usage_in_bytes or memory.memsw.usage_in_bytes;
 726- write string like "<event_fd> <fd of memory.usage_in_bytes> <threshold>" to
 727  cgroup.event_control.
 728
 729Application will be notified through eventfd when memory usage crosses
 730threshold in any direction.
 731
 732It's applicable for root and non-root cgroup.
 733
 73410. OOM Control
 735
 736memory.oom_control file is for OOM notification and other controls.
 737
 738Memory cgroup implements OOM notifier using the cgroup notification
 739API (See cgroups.txt). It allows to register multiple OOM notification
 740delivery and gets notification when OOM happens.
 741
 742To register a notifier, an application must:
 743 - create an eventfd using eventfd(2)
 744 - open memory.oom_control file
 745 - write string like "<event_fd> <fd of memory.oom_control>" to
 746   cgroup.event_control
 747
 748The application will be notified through eventfd when OOM happens.
 749OOM notification doesn't work for the root cgroup.
 750
 751You can disable the OOM-killer by writing "1" to memory.oom_control file, as:
 752
 753        #echo 1 > memory.oom_control
 754
 755If OOM-killer is disabled, tasks under cgroup will hang/sleep
 756in memory cgroup's OOM-waitqueue when they request accountable memory.
 757
 758For running them, you have to relax the memory cgroup's OOM status by
 759        * enlarge limit or reduce usage.
 760To reduce usage,
 761        * kill some tasks.
 762        * move some tasks to other group with account migration.
 763        * remove some files (on tmpfs?)
 764
 765Then, stopped tasks will work again.
 766
 767At reading, current status of OOM is shown.
 768        oom_kill_disable 0 or 1 (if 1, oom-killer is disabled)
 769        under_oom        0 or 1 (if 1, the memory cgroup is under OOM, tasks may
 770                                 be stopped.)
 771
 77211. Memory Pressure
 773
 774The pressure level notifications can be used to monitor the memory
 775allocation cost; based on the pressure, applications can implement
 776different strategies of managing their memory resources. The pressure
 777levels are defined as following:
 778
 779The "low" level means that the system is reclaiming memory for new
 780allocations. Monitoring this reclaiming activity might be useful for
 781maintaining cache level. Upon notification, the program (typically
 782"Activity Manager") might analyze vmstat and act in advance (i.e.
 783prematurely shutdown unimportant services).
 784
 785The "medium" level means that the system is experiencing medium memory
 786pressure, the system might be making swap, paging out active file caches,
 787etc. Upon this event applications may decide to further analyze
 788vmstat/zoneinfo/memcg or internal memory usage statistics and free any
 789resources that can be easily reconstructed or re-read from a disk.
 790
 791The "critical" level means that the system is actively thrashing, it is
 792about to out of memory (OOM) or even the in-kernel OOM killer is on its
 793way to trigger. Applications should do whatever they can to help the
 794system. It might be too late to consult with vmstat or any other
 795statistics, so it's advisable to take an immediate action.
 796
 797The events are propagated upward until the event is handled, i.e. the
 798events are not pass-through. Here is what this means: for example you have
 799three cgroups: A->B->C. Now you set up an event listener on cgroups A, B
 800and C, and suppose group C experiences some pressure. In this situation,
 801only group C will receive the notification, i.e. groups A and B will not
 802receive it. This is done to avoid excessive "broadcasting" of messages,
 803which disturbs the system and which is especially bad if we are low on
 804memory or thrashing. So, organize the cgroups wisely, or propagate the
 805events manually (or, ask us to implement the pass-through events,
 806explaining why would you need them.)
 807
 808The file memory.pressure_level is only used to setup an eventfd. To
 809register a notification, an application must:
 810
 811- create an eventfd using eventfd(2);
 812- open memory.pressure_level;
 813- write string like "<event_fd> <fd of memory.pressure_level> <level>"
 814  to cgroup.event_control.
 815
 816Application will be notified through eventfd when memory pressure is at
 817the specific level (or higher). Read/write operations to
 818memory.pressure_level are no implemented.
 819
 820Test:
 821
 822   Here is a small script example that makes a new cgroup, sets up a
 823   memory limit, sets up a notification in the cgroup and then makes child
 824   cgroup experience a critical pressure:
 825
 826   # cd /sys/fs/cgroup/memory/
 827   # mkdir foo
 828   # cd foo
 829   # cgroup_event_listener memory.pressure_level low &
 830   # echo 8000000 > memory.limit_in_bytes
 831   # echo 8000000 > memory.memsw.limit_in_bytes
 832   # echo $$ > tasks
 833   # dd if=/dev/zero | read x
 834
 835   (Expect a bunch of notifications, and eventually, the oom-killer will
 836   trigger.)
 837
 83812. TODO
 839
 8401. Make per-cgroup scanner reclaim not-shared pages first
 8412. Teach controller to account for shared-pages
 8423. Start reclamation in the background when the limit is
 843   not yet hit but the usage is getting closer
 844
 845Summary
 846
 847Overall, the memory controller has been a stable controller and has been
 848commented and discussed quite extensively in the community.
 849
 850References
 851
 8521. Singh, Balbir. RFC: Memory Controller, http://lwn.net/Articles/206697/
 8532. Singh, Balbir. Memory Controller (RSS Control),
 854   http://lwn.net/Articles/222762/
 8553. Emelianov, Pavel. Resource controllers based on process cgroups
 856   http://lkml.org/lkml/2007/3/6/198
 8574. Emelianov, Pavel. RSS controller based on process cgroups (v2)
 858   http://lkml.org/lkml/2007/4/9/78
 8595. Emelianov, Pavel. RSS controller based on process cgroups (v3)
 860   http://lkml.org/lkml/2007/5/30/244
 8616. Menage, Paul. Control Groups v10, http://lwn.net/Articles/236032/
 8627. Vaidyanathan, Srinivasan, Control Groups: Pagecache accounting and control
 863   subsystem (v3), http://lwn.net/Articles/235534/
 8648. Singh, Balbir. RSS controller v2 test results (lmbench),
 865   http://lkml.org/lkml/2007/5/17/232
 8669. Singh, Balbir. RSS controller v2 AIM9 results
 867   http://lkml.org/lkml/2007/5/18/1
 86810. Singh, Balbir. Memory controller v6 test results,
 869    http://lkml.org/lkml/2007/8/19/36
 87011. Singh, Balbir. Memory controller introduction (v6),
 871    http://lkml.org/lkml/2007/8/17/69
 87212. Corbet, Jonathan, Controlling memory use in cgroups,
 873    http://lwn.net/Articles/243795/
 874