xref: /linux/Documentation/admin-guide/cgroup-v1/memory.rst (revision 3a2c4d55e32ad65efebdb6de44eef3bfa08bb49d)
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6      This document is hopelessly outdated and it asks for a complete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 - moving (recharging) account at moving a task is selectable.
51 - usage threshold notifier
52 - memory pressure notifier
53 - oom-killer disable knob and oom-notifier
54 - Root cgroup has no limit controls.
55
56 Kernel memory support is a work in progress, and the current version provides
57 basically functionality. (See :ref:`section 2.7
58 <cgroup-v1-memory-kernel-extension>`)
59
60Brief summary of control files.
61
62==================================== ==========================================
63 tasks				     attach a task(thread) and show list of
64				     threads
65 cgroup.procs			     show list of processes
66 cgroup.event_control		     an interface for event_fd()
67				     This knob is not available on CONFIG_PREEMPT_RT systems.
68 memory.usage_in_bytes		     show current usage for memory
69				     (See 5.5 for details)
70 memory.memsw.usage_in_bytes	     show current usage for memory+Swap
71				     (See 5.5 for details)
72 memory.limit_in_bytes		     set/show limit of memory usage
73 memory.memsw.limit_in_bytes	     set/show limit of memory+Swap usage
74 memory.failcnt			     show the number of memory usage hits limits
75 memory.memsw.failcnt		     show the number of memory+Swap hits limits
76 memory.max_usage_in_bytes	     show max memory usage recorded
77 memory.memsw.max_usage_in_bytes     show max memory+Swap usage recorded
78 memory.soft_limit_in_bytes	     This knob is deprecated and has no effect.
79                                     Writes are ignored and reads always
80                                     return the maximum value.
81 memory.stat			     show various statistics
82 memory.use_hierarchy		     set/show hierarchical account enabled
83                                     This knob is deprecated and shouldn't be
84                                     used.
85 memory.force_empty		     trigger forced page reclaim
86 memory.pressure_level		     set memory pressure notifications
87                                     This knob is deprecated and shouldn't be
88                                     used.
89 memory.swappiness		     set/show swappiness parameter of vmscan
90				     (See sysctl's vm.swappiness)
91				     Per memcg knob does not exist in cgroup v2.
92 memory.move_charge_at_immigrate     This knob is deprecated.
93 memory.oom_control		     set/show oom controls.
94                                     This knob is deprecated and shouldn't be
95                                     used.
96 memory.numa_stat		     show the number of memory usage per numa
97				     node
98 memory.kmem.limit_in_bytes          Deprecated knob to set and read the kernel
99                                     memory hard limit. Kernel hard limit is not
100                                     supported since 5.16. Writing any value to
101                                     do file will not have any effect same as if
102                                     nokmem kernel parameter was specified.
103                                     Kernel memory is still charged and reported
104                                     by memory.kmem.usage_in_bytes.
105 memory.kmem.usage_in_bytes          show current kernel memory allocation
106 memory.kmem.failcnt                 show the number of kernel memory usage
107				     hits limits
108 memory.kmem.max_usage_in_bytes      show max kernel memory usage recorded
109
110 memory.kmem.tcp.limit_in_bytes      set/show hard limit for tcp buf memory
111                                     This knob is deprecated and shouldn't be
112                                     used.
113 memory.kmem.tcp.usage_in_bytes      show current tcp buf memory allocation
114                                     This knob is deprecated and shouldn't be
115                                     used.
116 memory.kmem.tcp.failcnt             show the number of tcp buf memory usage
117				     hits limits
118                                     This knob is deprecated and shouldn't be
119                                     used.
120 memory.kmem.tcp.max_usage_in_bytes  show max tcp buf memory usage recorded
121                                     This knob is deprecated and shouldn't be
122                                     used.
123==================================== ==========================================
124
1251. History
126==========
127
128The memory controller has a long history. A request for comments for the memory
129controller was posted by Balbir Singh [1]_. At the time the RFC was posted
130there were several implementations for memory control. The goal of the
131RFC was to build consensus and agreement for the minimal features required
132for memory control. The first RSS controller was posted by Balbir Singh [2]_
133in Feb 2007. Pavel Emelianov [3]_ [4]_ [5]_ has since posted three versions
134of the RSS controller. At OLS, at the resource management BoF, everyone
135suggested that we handle both page cache and RSS together. Another request was
136raised to allow user space handling of OOM. The current memory controller is
137at version 6; it combines both mapped (RSS) and unmapped Page
138Cache Control [11]_.
139
1402. Memory Control
141=================
142
143Memory is a unique resource in the sense that it is present in a limited
144amount. If a task requires a lot of CPU processing, the task can spread
145its processing over a period of hours, days, months or years, but with
146memory, the same physical memory needs to be reused to accomplish the task.
147
148The memory controller implementation has been divided into phases. These
149are:
150
1511. Memory controller
1522. mlock(2) controller
1533. Kernel user memory accounting and slab control
1544. user mappings length controller
155
156The memory controller is the first controller developed.
157
1582.1. Design
159-----------
160
161The core of the design is a counter called the page_counter. The
162page_counter tracks the current memory usage and limit of the group of
163processes associated with the controller. Each cgroup has a memory controller
164specific data structure (mem_cgroup) associated with it.
165
1662.2. Accounting
167---------------
168
169.. code-block::
170   :caption: Figure 1: Hierarchy of Accounting
171
172		+--------------------+
173		|  mem_cgroup        |
174		|  (page_counter)    |
175		+--------------------+
176		 /            ^      \
177		/             |       \
178           +---------------+  |        +---------------+
179           | mm_struct     |  |....    | mm_struct     |
180           |               |  |        |               |
181           +---------------+  |        +---------------+
182                              |
183                              + --------------+
184                                              |
185           +---------------+           +------+--------+
186           | page          +---------->  page_cgroup|
187           |               |           |               |
188           +---------------+           +---------------+
189
190
191
192Figure 1 shows the important aspects of the controller
193
1941. Accounting happens per cgroup
1952. Each mm_struct knows about which cgroup it belongs to
1963. Each page has a pointer to the page_cgroup, which in turn knows the
197   cgroup it belongs to
198
199The accounting is done as follows: mem_cgroup_charge_common() is invoked to
200set up the necessary data structures and check if the cgroup that is being
201charged is over its limit. If it is, then reclaim is invoked on the cgroup.
202More details can be found in the reclaim section of this document.
203If everything goes well, a page meta-data-structure called page_cgroup is
204updated. page_cgroup has its own LRU on cgroup.
205(*) page_cgroup structure is allocated at boot/memory-hotplug time.
206
2072.2.1 Accounting details
208------------------------
209
210All mapped anon pages (RSS) and cache pages (Page Cache) are accounted.
211Some pages which are never reclaimable and will not be on the LRU
212are not accounted. We just account pages under usual VM management.
213
214RSS pages are accounted at page_fault unless they've already been accounted
215for earlier. A file page will be accounted for as Page Cache when it's
216inserted into inode (xarray). While it's mapped into the page tables of
217processes, duplicate accounting is carefully avoided.
218
219An RSS page is unaccounted when it's fully unmapped. A PageCache page is
220unaccounted when it's removed from xarray. Even if RSS pages are fully
221unmapped (by kswapd), they may exist as SwapCache in the system until they
222are really freed. Such SwapCaches are also accounted.
223A swapped-in page is accounted after adding into swapcache.
224
225Note: The kernel does swapin-readahead and reads multiple swaps at once.
226Since page's memcg recorded into swap whatever memsw enabled, the page will
227be accounted after swapin.
228
229At page migration, accounting information is kept.
230
231Note: we just account pages-on-LRU because our purpose is to control amount
232of used pages; not-on-LRU pages tend to be out-of-control from VM view.
233
2342.3 Shared Page Accounting
235--------------------------
236
237Shared pages are accounted on the basis of the first touch approach. The
238cgroup that first touches a page is accounted for the page. The principle
239behind this approach is that a cgroup that aggressively uses a shared
240page will eventually get charged for it (once it is uncharged from
241the cgroup that brought it in -- this will happen on memory pressure).
242
2432.4 Swap Extension
244--------------------------------------
245
246Swap usage is always recorded for each of cgroup. Swap Extension allows you to
247read and limit it.
248
249When CONFIG_SWAP is enabled, following files are added.
250
251 - memory.memsw.usage_in_bytes.
252 - memory.memsw.limit_in_bytes.
253
254memsw means memory+swap. Usage of memory+swap is limited by
255memsw.limit_in_bytes.
256
257Example: Assume a system with 4G of swap. A task which allocates 6G of memory
258(by mistake) under 2G memory limitation will use all swap.
259In this case, setting memsw.limit_in_bytes=3G will prevent bad use of swap.
260By using the memsw limit, you can avoid system OOM which can be caused by swap
261shortage.
262
2632.4.1 why 'memory+swap' rather than swap
264~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
265
266The global LRU(kswapd) can swap out arbitrary pages. Swap-out means
267to move account from memory to swap...there is no change in usage of
268memory+swap. In other words, when we want to limit the usage of swap without
269affecting global LRU, memory+swap limit is better than just limiting swap from
270an OS point of view.
271
2722.4.2. What happens when a cgroup hits memory.memsw.limit_in_bytes
273~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
274
275When a cgroup hits memory.memsw.limit_in_bytes, it's useless to do swap-out
276in this cgroup. Then, swap-out will not be done by cgroup routine and file
277caches are dropped. But as mentioned above, global LRU can do swapout memory
278from it for sanity of the system's memory management state. You can't forbid
279it by cgroup.
280
2812.5 Reclaim
282-----------
283
284Each cgroup maintains a per cgroup LRU which has the same structure as
285global VM. When a cgroup goes over its limit, we first try
286to reclaim memory from the cgroup so as to make space for the new
287pages that the cgroup has touched. If the reclaim is unsuccessful,
288an OOM routine is invoked to select and kill the bulkiest task in the
289cgroup. (See :ref:`10. OOM Control <cgroup-v1-memory-oom-control>` below.)
290
291The reclaim algorithm has not been modified for cgroups, except that
292pages that are selected for reclaiming come from the per-cgroup LRU
293list.
294
295.. note::
296   Reclaim does not work for the root cgroup, since we cannot set any
297   limits on the root cgroup.
298
299.. note::
300   When panic_on_oom is set to "2", the whole system will panic.
301
302When oom event notifier is registered, event will be delivered.
303(See :ref:`oom_control <cgroup-v1-memory-oom-control>` section)
304
3052.6 Locking
306-----------
307
308Lock order is as follows::
309
310  folio_lock
311    mm->page_table_lock or split pte_lock
312      mapping->i_pages lock
313        lruvec->lru_lock.
314
315Per-node-per-memcgroup LRU (cgroup's private LRU) is guarded by
316lruvec->lru_lock; the folio LRU flag is cleared before
317isolating a page from its LRU under lruvec->lru_lock.
318
319.. _cgroup-v1-memory-kernel-extension:
320
3212.7 Kernel Memory Extension
322-----------------------------------------------
323
324With the Kernel memory extension, the Memory Controller is able to limit
325the amount of kernel memory used by the system. Kernel memory is fundamentally
326different than user memory, since it can't be swapped out, which makes it
327possible to DoS the system by consuming too much of this precious resource.
328
329Kernel memory accounting is enabled for all memory cgroups by default. But
330it can be disabled system-wide by passing cgroup.memory=nokmem to the kernel
331at boot time. In this case, kernel memory will not be accounted at all.
332
333Kernel memory limits are not imposed for the root cgroup. Usage for the root
334cgroup may or may not be accounted. The memory used is accumulated into
335memory.kmem.usage_in_bytes, or in a separate counter when it makes sense.
336(currently only for tcp).
337
338The main "kmem" counter is fed into the main counter, so kmem charges will
339also be visible from the user counter.
340
3412.7.1 Current Kernel Memory resources accounted
342-----------------------------------------------
343
344stack pages:
345  every process consumes some stack pages. By accounting into
346  kernel memory, we prevent new processes from being created when the kernel
347  memory usage is too high.
348
349slab pages:
350  pages allocated by the SLAB or SLUB allocator are tracked. A copy
351  of each kmem_cache is created every time the cache is touched by the first time
352  from inside the memcg. The creation is done lazily, so some objects can still be
353  skipped while the cache is being created. All objects in a slab page should
354  belong to the same memcg. This only fails to hold when a task is migrated to a
355  different memcg during the page allocation by the cache.
356
357sockets memory pressure:
358  some sockets protocols have memory pressure
359  thresholds. The Memory Controller allows them to be controlled individually
360  per cgroup, instead of globally.
361
362tcp memory pressure:
363  sockets memory pressure for the tcp protocol.
364
3652.7.2 Common use cases
366----------------------
367
368Because the "kmem" counter is fed to the main user counter, kernel memory can
369never be limited completely independently of user memory. Say "U" is the user
370limit, and "K" the kernel limit. There are three possible ways limits can be
371set:
372
373U != 0, K = unlimited:
374    This is the standard memcg limitation mechanism already present before kmem
375    accounting. Kernel memory is completely ignored.
376
377U != 0, K < U:
378    Kernel memory is a subset of the user memory. This setup is useful in
379    deployments where the total amount of memory per-cgroup is overcommitted.
380    Overcommitting kernel memory limits is definitely not recommended, since the
381    box can still run out of non-reclaimable memory.
382    In this case, the admin could set up K so that the sum of all groups is
383    never greater than the total memory, and freely set U at the cost of his
384    QoS.
385
386    .. warning::
387       In the current implementation, memory reclaim will NOT be triggered for
388       a cgroup when it hits K while staying below U, which makes this setup
389       impractical.
390
391U != 0, K >= U:
392    Since kmem charges will also be fed to the user counter and reclaim will be
393    triggered for the cgroup for both kinds of memory. This setup gives the
394    admin a unified view of memory, and it is also useful for people who just
395    want to track kernel memory usage.
396
3973. User Interface
398=================
399
400To use the user interface:
401
4021. Enable CONFIG_CGROUPS and CONFIG_MEMCG options
4032. Prepare the cgroups (see :ref:`Why are cgroups needed?
404   <cgroups-why-needed>` for the background information)::
405
406	# mount -t tmpfs none /sys/fs/cgroup
407	# mkdir /sys/fs/cgroup/memory
408	# mount -t cgroup none /sys/fs/cgroup/memory -o memory
409
4103. Make the new group and move bash into it::
411
412	# mkdir /sys/fs/cgroup/memory/0
413	# echo $$ > /sys/fs/cgroup/memory/0/tasks
414
4154. Since now we're in the 0 cgroup, we can alter the memory limit::
416
417	# echo 4M > /sys/fs/cgroup/memory/0/memory.limit_in_bytes
418
419   The limit can now be queried::
420
421	# cat /sys/fs/cgroup/memory/0/memory.limit_in_bytes
422	4194304
423
424.. note::
425   We can use a suffix (k, K, m, M, g or G) to indicate values in kilo,
426   mega or gigabytes. (Here, Kilo, Mega, Giga are Kibibytes, Mebibytes,
427   Gibibytes.)
428
429.. note::
430   We can write "-1" to reset the ``*.limit_in_bytes(unlimited)``.
431
432.. note::
433   We cannot set limits on the root cgroup any more.
434
435
436We can check the usage::
437
438  # cat /sys/fs/cgroup/memory/0/memory.usage_in_bytes
439  1216512
440
441A successful write to this file does not guarantee a successful setting of
442this limit to the value written into the file. This can be due to a
443number of factors, such as rounding up to page boundaries or the total
444availability of memory on the system. The user is required to re-read
445this file after a write to guarantee the value committed by the kernel::
446
447  # echo 1 > memory.limit_in_bytes
448  # cat memory.limit_in_bytes
449  4096
450
451The memory.failcnt field gives the number of times that the cgroup limit was
452exceeded.
453
454The memory.stat file gives accounting information. Now, the number of
455caches, RSS and Active pages/Inactive pages are shown.
456
4574. Testing
458==========
459
460For testing features and implementation, see memcg_test.txt.
461
462Performance test is also important. To see pure memory controller's overhead,
463testing on tmpfs will give you good numbers of small overheads.
464Example: do kernel make on tmpfs.
465
466Page-fault scalability is also important. At measuring parallel
467page fault test, multi-process test may be better than multi-thread
468test because it has noise of shared objects/status.
469
470But the above two are testing extreme situations.
471Trying usual test under memory controller is always helpful.
472
473.. _cgroup-v1-memory-test-troubleshoot:
474
4754.1 Troubleshooting
476-------------------
477
478Sometimes a user might find that the application under a cgroup is
479terminated by the OOM killer. There are several causes for this:
480
4811. The cgroup limit is too low (just too low to do anything useful)
4822. The user is using anonymous memory and swap is turned off or too low
483
484A sync followed by echo 1 > /proc/sys/vm/drop_caches will help get rid of
485some of the pages cached in the cgroup (page cache pages).
486
487To know what happens, disabling OOM_Kill as per :ref:`"10. OOM Control"
488<cgroup-v1-memory-oom-control>` (below) and seeing what happens will be
489helpful.
490
491.. _cgroup-v1-memory-test-task-migration:
492
4934.2 Task migration
494------------------
495
496When a task migrates from one cgroup to another, its charge is not
497carried forward by default. The pages allocated from the original cgroup still
498remain charged to it, the charge is dropped when the page is freed or
499reclaimed.
500
501You can move charges of a task along with task migration.
502See :ref:`8. "Move charges at task migration" <cgroup-v1-memory-move-charges>`
503
5044.3 Removing a cgroup
505---------------------
506
507A cgroup can be removed by rmdir, but as discussed in :ref:`sections 4.1
508<cgroup-v1-memory-test-troubleshoot>` and :ref:`4.2
509<cgroup-v1-memory-test-task-migration>`, a cgroup might have some charge
510associated with it, even though all tasks have migrated away from it. (because
511we charge against pages, not against tasks.)
512
513We move the stats to parent, and no change on the charge except uncharging
514from the child.
515
516Charges recorded in swap information is not updated at removal of cgroup.
517Recorded information is discarded and a cgroup which uses swap (swapcache)
518will be charged as a new owner of it.
519
5205. Misc. interfaces
521===================
522
5235.1 force_empty
524---------------
525  memory.force_empty interface is provided to make cgroup's memory usage empty.
526  When writing anything to this::
527
528    # echo 0 > memory.force_empty
529
530  the cgroup will be reclaimed and as many pages reclaimed as possible.
531
532  The typical use case for this interface is before calling rmdir().
533  Though rmdir() offlines memcg, but the memcg may still stay there due to
534  charged file caches. Some out-of-use page caches may keep charged until
535  memory pressure happens. If you want to avoid that, force_empty will be useful.
536
5375.2 stat file
538-------------
539
540memory.stat file includes following statistics:
541
542  * per-memory cgroup local status
543
544    =============== ===============================================================
545    cache           # of bytes of page cache memory.
546    rss             # of bytes of anonymous and swap cache memory (includes
547                    transparent hugepages).
548    rss_huge        # of bytes of anonymous transparent hugepages.
549    mapped_file     # of bytes of mapped file (includes tmpfs/shmem)
550    pgpgin          # of charging events to the memory cgroup. The charging
551                    event happens each time a page is accounted as either mapped
552                    anon page(RSS) or cache page(Page Cache) to the cgroup.
553    pgpgout         # of uncharging events to the memory cgroup. The uncharging
554                    event happens each time a page is unaccounted from the
555                    cgroup.
556    swap            # of bytes of swap usage
557    swapcached      # of bytes of swap cached in memory
558    dirty           # of bytes that are waiting to get written back to the disk.
559    writeback       # of bytes of file/anon cache that are queued for syncing to
560                    disk.
561    inactive_anon   # of bytes of anonymous and swap cache memory on inactive
562                    LRU list.
563    active_anon     # of bytes of anonymous and swap cache memory on active
564                    LRU list.
565    inactive_file   # of bytes of file-backed memory and MADV_FREE anonymous
566                    memory (LazyFree pages) on inactive LRU list.
567    active_file     # of bytes of file-backed memory on active LRU list.
568    unevictable     # of bytes of memory that cannot be reclaimed (mlocked etc).
569    =============== ===============================================================
570
571  * status considering hierarchy (see memory.use_hierarchy settings):
572
573    ========================= ===================================================
574    hierarchical_memory_limit # of bytes of memory limit with regard to
575                              hierarchy
576                              under which the memory cgroup is
577    hierarchical_memsw_limit  # of bytes of memory+swap limit with regard to
578                              hierarchy under which memory cgroup is.
579
580    total_<counter>           # hierarchical version of <counter>, which in
581                              addition to the cgroup's own value includes the
582                              sum of all hierarchical children's values of
583                              <counter>, i.e. total_cache
584    ========================= ===================================================
585
586  * additional vm parameters (depends on CONFIG_DEBUG_VM):
587
588    ========================= ========================================
589    recent_rotated_anon       VM internal parameter. (see mm/vmscan.c)
590    recent_rotated_file       VM internal parameter. (see mm/vmscan.c)
591    recent_scanned_anon       VM internal parameter. (see mm/vmscan.c)
592    recent_scanned_file       VM internal parameter. (see mm/vmscan.c)
593    ========================= ========================================
594
595.. hint::
596	recent_rotated means recent frequency of LRU rotation.
597	recent_scanned means recent # of scans to LRU.
598	showing for better debug please see the code for meanings.
599
600.. note::
601	Only anonymous and swap cache memory is listed as part of 'rss' stat.
602	This should not be confused with the true 'resident set size' or the
603	amount of physical memory used by the cgroup.
604
605	'rss + mapped_file" will give you resident set size of cgroup.
606
607	Note that some kernel configurations might account complete larger
608	allocations (e.g., THP) towards 'rss' and 'mapped_file', even if
609	only some, but not all that memory is mapped.
610
611	(Note: file and shmem may be shared among other cgroups. In that case,
612	mapped_file is accounted only when the memory cgroup is owner of page
613	cache.)
614
6155.3 swappiness
616--------------
617
618Overrides /proc/sys/vm/swappiness for the particular group. The tunable
619in the root cgroup corresponds to the global swappiness setting.
620
621Please note that unlike during the global reclaim, limit reclaim
622enforces that 0 swappiness really prevents from any swapping even if
623there is a swap storage available. This might lead to memcg OOM killer
624if there are no file pages to reclaim.
625
6265.4 failcnt
627-----------
628
629A memory cgroup provides memory.failcnt and memory.memsw.failcnt files.
630This failcnt(== failure count) shows the number of times that a usage counter
631hit its limit. When a memory cgroup hits a limit, failcnt increases and
632memory under it will be reclaimed.
633
634You can reset failcnt by writing 0 to failcnt file::
635
636	# echo 0 > .../memory.failcnt
637
6385.5 usage_in_bytes
639------------------
640
641For efficiency, as other kernel components, memory cgroup uses some optimization
642to avoid unnecessary cacheline false sharing. usage_in_bytes is affected by the
643method and doesn't show 'exact' value of memory (and swap) usage, it's a fuzz
644value for efficient access. (Of course, when necessary, it's synchronized.)
645If you want to know more exact memory usage, you should use RSS+CACHE(+SWAP)
646value in memory.stat(see 5.2).
647
6485.6 numa_stat
649-------------
650
651This is similar to numa_maps but operates on a per-memcg basis.  This is
652useful for providing visibility into the numa locality information within
653an memcg since the pages are allowed to be allocated from any physical
654node.  One of the use cases is evaluating application performance by
655combining this information with the application's CPU allocation.
656
657Each memcg's numa_stat file includes "total", "file", "anon" and "unevictable"
658per-node page counts including "hierarchical_<counter>" which sums up all
659hierarchical children's values in addition to the memcg's own value.
660
661The output format of memory.numa_stat is::
662
663  total=<total pages> N0=<node 0 pages> N1=<node 1 pages> ...
664  file=<total file pages> N0=<node 0 pages> N1=<node 1 pages> ...
665  anon=<total anon pages> N0=<node 0 pages> N1=<node 1 pages> ...
666  unevictable=<total anon pages> N0=<node 0 pages> N1=<node 1 pages> ...
667  hierarchical_<counter>=<counter pages> N0=<node 0 pages> N1=<node 1 pages> ...
668
669The "total" count is sum of file + anon + unevictable.
670
6716. Hierarchy support
672====================
673
674The memory controller supports a deep hierarchy and hierarchical accounting.
675The hierarchy is created by creating the appropriate cgroups in the
676cgroup filesystem. Consider for example, the following cgroup filesystem
677hierarchy::
678
679	       root
680	     /  |   \
681            /	|    \
682	   a	b     c
683		      | \
684		      |  \
685		      d   e
686
687In the diagram above, with hierarchical accounting enabled, all memory
688usage of e, is accounted to its ancestors up until the root (i.e, c and root).
689If one of the ancestors goes over its limit, the reclaim algorithm reclaims
690from the tasks in the ancestor and the children of the ancestor.
691
6926.1 Hierarchical accounting and reclaim
693---------------------------------------
694
695Hierarchical accounting is enabled by default. Disabling the hierarchical
696accounting is deprecated. An attempt to do it will result in a failure
697and a warning printed to dmesg.
698
699For compatibility reasons writing 1 to memory.use_hierarchy will always pass::
700
701	# echo 1 > memory.use_hierarchy
702
7037. Soft limits (DEPRECATED)
704===========================
705
706THIS IS DEPRECATED!
707
708Writing to memory.soft_limit_in_bytes has no effect and reading it will
709always return the maximum value.
710
711Use memory.low and memory.min in cgroup v2 instead.
712
713.. _cgroup-v1-memory-move-charges:
714
7158. Move charges at task migration (DEPRECATED!)
716===============================================
717
718THIS IS DEPRECATED!
719
720Reading memory.move_charge_at_immigrate will always return 0 and writing
721to it will always return -EINVAL.
722
7239. Memory thresholds
724====================
725
726Memory cgroup implements memory thresholds using the cgroups notification
727API (see cgroups.txt). It allows to register multiple memory and memsw
728thresholds and gets notifications when it crosses.
729
730To register a threshold, an application must:
731
732- create an eventfd using eventfd(2);
733- open memory.usage_in_bytes or memory.memsw.usage_in_bytes;
734- write string like "<event_fd> <fd of memory.usage_in_bytes> <threshold>" to
735  cgroup.event_control.
736
737Application will be notified through eventfd when memory usage crosses
738threshold in any direction.
739
740It's applicable for root and non-root cgroup.
741
742.. _cgroup-v1-memory-oom-control:
743
74410. OOM Control (DEPRECATED)
745============================
746
747THIS IS DEPRECATED!
748
749memory.oom_control file is for OOM notification and other controls.
750
751Memory cgroup implements OOM notifier using the cgroup notification
752API (See cgroups.txt). It allows to register multiple OOM notification
753delivery and gets notification when OOM happens.
754
755To register a notifier, an application must:
756
757 - create an eventfd using eventfd(2)
758 - open memory.oom_control file
759 - write string like "<event_fd> <fd of memory.oom_control>" to
760   cgroup.event_control
761
762The application will be notified through eventfd when OOM happens.
763OOM notification doesn't work for the root cgroup.
764
765You can disable the OOM-killer by writing "1" to memory.oom_control file, as:
766
767	#echo 1 > memory.oom_control
768
769If OOM-killer is disabled, tasks under cgroup will hang/sleep
770in memory cgroup's OOM-waitqueue when they request accountable memory.
771
772For running them, you have to relax the memory cgroup's OOM status by
773
774	* enlarge limit or reduce usage.
775
776To reduce usage,
777
778	* kill some tasks.
779	* move some tasks to other group with account migration.
780	* remove some files (on tmpfs?)
781
782Then, stopped tasks will work again.
783
784At reading, current status of OOM is shown.
785
786	- oom_kill_disable 0 or 1
787	  (if 1, oom-killer is disabled)
788	- under_oom	   0 or 1
789	  (if 1, the memory cgroup is under OOM, tasks may be stopped.)
790        - oom_kill         integer counter
791          The number of processes belonging to this cgroup killed by any
792          kind of OOM killer.
793
79411. Memory Pressure (DEPRECATED)
795================================
796
797THIS IS DEPRECATED!
798
799The pressure level notifications can be used to monitor the memory
800allocation cost; based on the pressure, applications can implement
801different strategies of managing their memory resources. The pressure
802levels are defined as following:
803
804The "low" level means that the system is reclaiming memory for new
805allocations. Monitoring this reclaiming activity might be useful for
806maintaining cache level. Upon notification, the program (typically
807"Activity Manager") might analyze vmstat and act in advance (i.e.
808prematurely shutdown unimportant services).
809
810The "medium" level means that the system is experiencing medium memory
811pressure, the system might be making swap, paging out active file caches,
812etc. Upon this event applications may decide to further analyze
813vmstat/zoneinfo/memcg or internal memory usage statistics and free any
814resources that can be easily reconstructed or re-read from a disk.
815
816The "critical" level means that the system is actively thrashing, it is
817about to out of memory (OOM) or even the in-kernel OOM killer is on its
818way to trigger. Applications should do whatever they can to help the
819system. It might be too late to consult with vmstat or any other
820statistics, so it's advisable to take an immediate action.
821
822By default, events are propagated upward until the event is handled, i.e. the
823events are not pass-through. For example, you have three cgroups: A->B->C. Now
824you set up an event listener on cgroups A, B and C, and suppose group C
825experiences some pressure. In this situation, only group C will receive the
826notification, i.e. groups A and B will not receive it. This is done to avoid
827excessive "broadcasting" of messages, which disturbs the system and which is
828especially bad if we are low on memory or thrashing. Group B, will receive
829notification only if there are no event listeners for group C.
830
831There are three optional modes that specify different propagation behavior:
832
833 - "default": this is the default behavior specified above. This mode is the
834   same as omitting the optional mode parameter, preserved by backwards
835   compatibility.
836
837 - "hierarchy": events always propagate up to the root, similar to the default
838   behavior, except that propagation continues regardless of whether there are
839   event listeners at each level, with the "hierarchy" mode. In the above
840   example, groups A, B, and C will receive notification of memory pressure.
841
842 - "local": events are pass-through, i.e. they only receive notifications when
843   memory pressure is experienced in the memcg for which the notification is
844   registered. In the above example, group C will receive notification if
845   registered for "local" notification and the group experiences memory
846   pressure. However, group B will never receive notification, regardless if
847   there is an event listener for group C or not, if group B is registered for
848   local notification.
849
850The level and event notification mode ("hierarchy" or "local", if necessary) are
851specified by a comma-delimited string, i.e. "low,hierarchy" specifies
852hierarchical, pass-through, notification for all ancestor memcgs. Notification
853that is the default, non pass-through behavior, does not specify a mode.
854"medium,local" specifies pass-through notification for the medium level.
855
856The file memory.pressure_level is only used to setup an eventfd. To
857register a notification, an application must:
858
859- create an eventfd using eventfd(2);
860- open memory.pressure_level;
861- write string as "<event_fd> <fd of memory.pressure_level> <level[,mode]>"
862  to cgroup.event_control.
863
864Application will be notified through eventfd when memory pressure is at
865the specific level (or higher). Read/write operations to
866memory.pressure_level are no implemented.
867
868Test:
869
870   Here is a small script example that makes a new cgroup, sets up a
871   memory limit, sets up a notification in the cgroup and then makes child
872   cgroup experience a critical pressure::
873
874	# cd /sys/fs/cgroup/memory/
875	# mkdir foo
876	# cd foo
877	# cgroup_event_listener memory.pressure_level low,hierarchy &
878	# echo 8000000 > memory.limit_in_bytes
879	# echo 8000000 > memory.memsw.limit_in_bytes
880	# echo $$ > tasks
881	# dd if=/dev/zero | read x
882
883   (Expect a bunch of notifications, and eventually, the oom-killer will
884   trigger.)
885
88612. TODO
887========
888
8891. Make per-cgroup scanner reclaim not-shared pages first
8902. Teach controller to account for shared-pages
8913. Start reclamation in the background when the limit is
892   not yet hit but the usage is getting closer
893
894Summary
895=======
896
897Overall, the memory controller has been a stable controller and has been
898commented and discussed quite extensively in the community.
899
900References
901==========
902
903.. [1] Singh, Balbir. RFC: Memory Controller, http://lwn.net/Articles/206697/
904.. [2] Singh, Balbir. Memory Controller (RSS Control),
905   http://lwn.net/Articles/222762/
906.. [3] Emelianov, Pavel. Resource controllers based on process cgroups
907   https://lore.kernel.org/r/45ED7DEC.7010403@sw.ru
908.. [4] Emelianov, Pavel. RSS controller based on process cgroups (v2)
909   https://lore.kernel.org/r/461A3010.90403@sw.ru
910.. [5] Emelianov, Pavel. RSS controller based on process cgroups (v3)
911   https://lore.kernel.org/r/465D9739.8070209@openvz.org
912
9136. Menage, Paul. Control Groups v10, http://lwn.net/Articles/236032/
9147. Vaidyanathan, Srinivasan, Control Groups: Pagecache accounting and control
915   subsystem (v3), http://lwn.net/Articles/235534/
9168. Singh, Balbir. RSS controller v2 test results (lmbench),
917   https://lore.kernel.org/r/464C95D4.7070806@linux.vnet.ibm.com
9189. Singh, Balbir. RSS controller v2 AIM9 results
919   https://lore.kernel.org/r/464D267A.50107@linux.vnet.ibm.com
92010. Singh, Balbir. Memory controller v6 test results,
921    https://lore.kernel.org/r/20070819094658.654.84837.sendpatchset@balbir-laptop
922
923.. [11] Singh, Balbir. Memory controller introduction (v6),
924   https://lore.kernel.org/r/20070817084228.26003.12568.sendpatchset@balbir-laptop
925.. [12] Corbet, Jonathan, Controlling memory use in cgroups,
926   http://lwn.net/Articles/243795/
927