1==========================
2Memory Resource Controller
3==========================
4
5.. caution::
6      This document is hopelessly outdated and it asks for a complete
7      rewrite. It still contains a useful information so we are keeping it
8      here but make sure to check the current code if you need a deeper
9      understanding.
10
11.. note::
12      The Memory Resource Controller has generically been referred to as the
13      memory controller in this document. Do not confuse memory controller
14      used here with the memory controller that is used in hardware.
15
16.. hint::
17      When we mention a cgroup (cgroupfs's directory) with memory controller,
18      we call it "memory cgroup". When you see git-log and source code, you'll
19      see patch's title and function names tend to use "memcg".
20      In this document, we avoid using it.
21
22Benefits and Purpose of the memory controller
23=============================================
24
25The memory controller isolates the memory behaviour of a group of tasks
26from the rest of the system. The article on LWN [12]_ mentions some probable
27uses of the memory controller. The memory controller can be used to
28
29a. Isolate an application or a group of applications
30   Memory-hungry applications can be isolated and limited to a smaller
31   amount of memory.
32b. Create a cgroup with a limited amount of memory; this can be used
33   as a good alternative to booting with mem=XXXX.
34c. Virtualization solutions can control the amount of memory they want
35   to assign to a virtual machine instance.
36d. A CD/DVD burner could control the amount of memory used by the
37   rest of the system to ensure that burning does not fail due to lack
38   of available memory.
39e. There are several other use cases; find one or use the controller just
40   for fun (to learn and hack on the VM subsystem).
41
42Current Status: linux-2.6.34-mmotm(development version of 2010/April)
43
44Features:
45
46 - accounting anonymous pages, file caches, swap caches usage and limiting them.
47 - pages are linked to per-memcg LRU exclusively, and there is no global LRU.
48 - optionally, memory+swap usage can be accounted and limited.
49 - hierarchical accounting
50 - soft limit
51 - moving (recharging) account at moving a task is selectable.
52 - usage threshold notifier
53 - memory pressure notifier
54 - oom-killer disable knob and oom-notifier
55 - Root cgroup has no limit controls.
56
57 Kernel memory support is a work in progress, and the current version provides
58 basically functionality. (See :ref:`section 2.7
59 <cgroup-v1-memory-kernel-extension>`)
60
61Brief summary of control files.
62
63==================================== ==========================================
64 tasks				     attach a task(thread) and show list of
65				     threads
66 cgroup.procs			     show list of processes
67 cgroup.event_control		     an interface for event_fd()
68				     This knob is not available on CONFIG_PREEMPT_RT systems.
69 memory.usage_in_bytes		     show current usage for memory
70				     (See 5.5 for details)
71 memory.memsw.usage_in_bytes	     show current usage for memory+Swap
72				     (See 5.5 for details)
73 memory.limit_in_bytes		     set/show limit of memory usage
74 memory.memsw.limit_in_bytes	     set/show limit of memory+Swap usage
75 memory.failcnt			     show the number of memory usage hits limits
76 memory.memsw.failcnt		     show the number of memory+Swap hits limits
77 memory.max_usage_in_bytes	     show max memory usage recorded
78 memory.memsw.max_usage_in_bytes     show max memory+Swap usage recorded
79 memory.soft_limit_in_bytes	     set/show soft limit of memory usage
80				     This knob is not available on CONFIG_PREEMPT_RT systems.
81 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 memory.swappiness		     set/show swappiness parameter of vmscan
88				     (See sysctl's vm.swappiness)
89 memory.move_charge_at_immigrate     set/show controls of moving charges
90                                     This knob is deprecated and shouldn't be
91                                     used.
92 memory.oom_control		     set/show oom controls.
93 memory.numa_stat		     show the number of memory usage per numa
94				     node
95 memory.kmem.limit_in_bytes          This knob is deprecated and writing to
96                                     it will return -ENOTSUPP.
97 memory.kmem.usage_in_bytes          show current kernel memory allocation
98 memory.kmem.failcnt                 show the number of kernel memory usage
99				     hits limits
100 memory.kmem.max_usage_in_bytes      show max kernel memory usage recorded
101
102 memory.kmem.tcp.limit_in_bytes      set/show hard limit for tcp buf memory
103 memory.kmem.tcp.usage_in_bytes      show current tcp buf memory allocation
104 memory.kmem.tcp.failcnt             show the number of tcp buf memory usage
105				     hits limits
106 memory.kmem.tcp.max_usage_in_bytes  show max tcp buf memory usage recorded
107==================================== ==========================================
108
1091. History
110==========
111
112The memory controller has a long history. A request for comments for the memory
113controller was posted by Balbir Singh [1]_. At the time the RFC was posted
114there were several implementations for memory control. The goal of the
115RFC was to build consensus and agreement for the minimal features required
116for memory control. The first RSS controller was posted by Balbir Singh [2]_
117in Feb 2007. Pavel Emelianov [3]_ [4]_ [5]_ has since posted three versions
118of the RSS controller. At OLS, at the resource management BoF, everyone
119suggested that we handle both page cache and RSS together. Another request was
120raised to allow user space handling of OOM. The current memory controller is
121at version 6; it combines both mapped (RSS) and unmapped Page
122Cache Control [11]_.
123
1242. Memory Control
125=================
126
127Memory is a unique resource in the sense that it is present in a limited
128amount. If a task requires a lot of CPU processing, the task can spread
129its processing over a period of hours, days, months or years, but with
130memory, the same physical memory needs to be reused to accomplish the task.
131
132The memory controller implementation has been divided into phases. These
133are:
134
1351. Memory controller
1362. mlock(2) controller
1373. Kernel user memory accounting and slab control
1384. user mappings length controller
139
140The memory controller is the first controller developed.
141
1422.1. Design
143-----------
144
145The core of the design is a counter called the page_counter. The
146page_counter tracks the current memory usage and limit of the group of
147processes associated with the controller. Each cgroup has a memory controller
148specific data structure (mem_cgroup) associated with it.
149
1502.2. Accounting
151---------------
152
153.. code-block::
154   :caption: Figure 1: Hierarchy of Accounting
155
156		+--------------------+
157		|  mem_cgroup        |
158		|  (page_counter)    |
159		+--------------------+
160		 /            ^      \
161		/             |       \
162           +---------------+  |        +---------------+
163           | mm_struct     |  |....    | mm_struct     |
164           |               |  |        |               |
165           +---------------+  |        +---------------+
166                              |
167                              + --------------+
168                                              |
169           +---------------+           +------+--------+
170           | page          +---------->  page_cgroup|
171           |               |           |               |
172           +---------------+           +---------------+
173
174
175
176Figure 1 shows the important aspects of the controller
177
1781. Accounting happens per cgroup
1792. Each mm_struct knows about which cgroup it belongs to
1803. Each page has a pointer to the page_cgroup, which in turn knows the
181   cgroup it belongs to
182
183The accounting is done as follows: mem_cgroup_charge_common() is invoked to
184set up the necessary data structures and check if the cgroup that is being
185charged is over its limit. If it is, then reclaim is invoked on the cgroup.
186More details can be found in the reclaim section of this document.
187If everything goes well, a page meta-data-structure called page_cgroup is
188updated. page_cgroup has its own LRU on cgroup.
189(*) page_cgroup structure is allocated at boot/memory-hotplug time.
190
1912.2.1 Accounting details
192------------------------
193
194All mapped anon pages (RSS) and cache pages (Page Cache) are accounted.
195Some pages which are never reclaimable and will not be on the LRU
196are not accounted. We just account pages under usual VM management.
197
198RSS pages are accounted at page_fault unless they've already been accounted
199for earlier. A file page will be accounted for as Page Cache when it's
200inserted into inode (radix-tree). While it's mapped into the page tables of
201processes, duplicate accounting is carefully avoided.
202
203An RSS page is unaccounted when it's fully unmapped. A PageCache page is
204unaccounted when it's removed from radix-tree. Even if RSS pages are fully
205unmapped (by kswapd), they may exist as SwapCache in the system until they
206are really freed. Such SwapCaches are also accounted.
207A swapped-in page is accounted after adding into swapcache.
208
209Note: The kernel does swapin-readahead and reads multiple swaps at once.
210Since page's memcg recorded into swap whatever memsw enabled, the page will
211be accounted after swapin.
212
213At page migration, accounting information is kept.
214
215Note: we just account pages-on-LRU because our purpose is to control amount
216of used pages; not-on-LRU pages tend to be out-of-control from VM view.
217
2182.3 Shared Page Accounting
219--------------------------
220
221Shared pages are accounted on the basis of the first touch approach. The
222cgroup that first touches a page is accounted for the page. The principle
223behind this approach is that a cgroup that aggressively uses a shared
224page will eventually get charged for it (once it is uncharged from
225the cgroup that brought it in -- this will happen on memory pressure).
226
227But see :ref:`section 8.2 <cgroup-v1-memory-movable-charges>` when moving a
228task to another cgroup, its pages may be recharged to the new cgroup, if
229move_charge_at_immigrate has been chosen.
230
2312.4 Swap Extension
232--------------------------------------
233
234Swap usage is always recorded for each of cgroup. Swap Extension allows you to
235read and limit it.
236
237When CONFIG_SWAP is enabled, following files are added.
238
239 - memory.memsw.usage_in_bytes.
240 - memory.memsw.limit_in_bytes.
241
242memsw means memory+swap. Usage of memory+swap is limited by
243memsw.limit_in_bytes.
244
245Example: Assume a system with 4G of swap. A task which allocates 6G of memory
246(by mistake) under 2G memory limitation will use all swap.
247In this case, setting memsw.limit_in_bytes=3G will prevent bad use of swap.
248By using the memsw limit, you can avoid system OOM which can be caused by swap
249shortage.
250
2512.4.1 why 'memory+swap' rather than swap
252~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
253
254The global LRU(kswapd) can swap out arbitrary pages. Swap-out means
255to move account from memory to swap...there is no change in usage of
256memory+swap. In other words, when we want to limit the usage of swap without
257affecting global LRU, memory+swap limit is better than just limiting swap from
258an OS point of view.
259
2602.4.2. What happens when a cgroup hits memory.memsw.limit_in_bytes
261~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
262
263When a cgroup hits memory.memsw.limit_in_bytes, it's useless to do swap-out
264in this cgroup. Then, swap-out will not be done by cgroup routine and file
265caches are dropped. But as mentioned above, global LRU can do swapout memory
266from it for sanity of the system's memory management state. You can't forbid
267it by cgroup.
268
2692.5 Reclaim
270-----------
271
272Each cgroup maintains a per cgroup LRU which has the same structure as
273global VM. When a cgroup goes over its limit, we first try
274to reclaim memory from the cgroup so as to make space for the new
275pages that the cgroup has touched. If the reclaim is unsuccessful,
276an OOM routine is invoked to select and kill the bulkiest task in the
277cgroup. (See :ref:`10. OOM Control <cgroup-v1-memory-oom-control>` below.)
278
279The reclaim algorithm has not been modified for cgroups, except that
280pages that are selected for reclaiming come from the per-cgroup LRU
281list.
282
283.. note::
284   Reclaim does not work for the root cgroup, since we cannot set any
285   limits on the root cgroup.
286
287.. note::
288   When panic_on_oom is set to "2", the whole system will panic.
289
290When oom event notifier is registered, event will be delivered.
291(See :ref:`oom_control <cgroup-v1-memory-oom-control>` section)
292
2932.6 Locking
294-----------
295
296Lock order is as follows::
297
298  Page lock (PG_locked bit of page->flags)
299    mm->page_table_lock or split pte_lock
300      folio_memcg_lock (memcg->move_lock)
301        mapping->i_pages lock
302          lruvec->lru_lock.
303
304Per-node-per-memcgroup LRU (cgroup's private LRU) is guarded by
305lruvec->lru_lock; PG_lru bit of page->flags is cleared before
306isolating a page from its LRU under lruvec->lru_lock.
307
308.. _cgroup-v1-memory-kernel-extension:
309
3102.7 Kernel Memory Extension
311-----------------------------------------------
312
313With the Kernel memory extension, the Memory Controller is able to limit
314the amount of kernel memory used by the system. Kernel memory is fundamentally
315different than user memory, since it can't be swapped out, which makes it
316possible to DoS the system by consuming too much of this precious resource.
317
318Kernel memory accounting is enabled for all memory cgroups by default. But
319it can be disabled system-wide by passing cgroup.memory=nokmem to the kernel
320at boot time. In this case, kernel memory will not be accounted at all.
321
322Kernel memory limits are not imposed for the root cgroup. Usage for the root
323cgroup may or may not be accounted. The memory used is accumulated into
324memory.kmem.usage_in_bytes, or in a separate counter when it makes sense.
325(currently only for tcp).
326
327The main "kmem" counter is fed into the main counter, so kmem charges will
328also be visible from the user counter.
329
330Currently no soft limit is implemented for kernel memory. It is future work
331to trigger slab reclaim when those limits are reached.
332
3332.7.1 Current Kernel Memory resources accounted
334-----------------------------------------------
335
336stack pages:
337  every process consumes some stack pages. By accounting into
338  kernel memory, we prevent new processes from being created when the kernel
339  memory usage is too high.
340
341slab pages:
342  pages allocated by the SLAB or SLUB allocator are tracked. A copy
343  of each kmem_cache is created every time the cache is touched by the first time
344  from inside the memcg. The creation is done lazily, so some objects can still be
345  skipped while the cache is being created. All objects in a slab page should
346  belong to the same memcg. This only fails to hold when a task is migrated to a
347  different memcg during the page allocation by the cache.
348
349sockets memory pressure:
350  some sockets protocols have memory pressure
351  thresholds. The Memory Controller allows them to be controlled individually
352  per cgroup, instead of globally.
353
354tcp memory pressure:
355  sockets memory pressure for the tcp protocol.
356
3572.7.2 Common use cases
358----------------------
359
360Because the "kmem" counter is fed to the main user counter, kernel memory can
361never be limited completely independently of user memory. Say "U" is the user
362limit, and "K" the kernel limit. There are three possible ways limits can be
363set:
364
365U != 0, K = unlimited:
366    This is the standard memcg limitation mechanism already present before kmem
367    accounting. Kernel memory is completely ignored.
368
369U != 0, K < U:
370    Kernel memory is a subset of the user memory. This setup is useful in
371    deployments where the total amount of memory per-cgroup is overcommitted.
372    Overcommitting kernel memory limits is definitely not recommended, since the
373    box can still run out of non-reclaimable memory.
374    In this case, the admin could set up K so that the sum of all groups is
375    never greater than the total memory, and freely set U at the cost of his
376    QoS.
377
378    .. warning::
379       In the current implementation, memory reclaim will NOT be triggered for
380       a cgroup when it hits K while staying below U, which makes this setup
381       impractical.
382
383U != 0, K >= U:
384    Since kmem charges will also be fed to the user counter and reclaim will be
385    triggered for the cgroup for both kinds of memory. This setup gives the
386    admin a unified view of memory, and it is also useful for people who just
387    want to track kernel memory usage.
388
3893. User Interface
390=================
391
392To use the user interface:
393
3941. Enable CONFIG_CGROUPS and CONFIG_MEMCG options
3952. Prepare the cgroups (see :ref:`Why are cgroups needed?
396   <cgroups-why-needed>` for the background information)::
397
398	# mount -t tmpfs none /sys/fs/cgroup
399	# mkdir /sys/fs/cgroup/memory
400	# mount -t cgroup none /sys/fs/cgroup/memory -o memory
401
4023. Make the new group and move bash into it::
403
404	# mkdir /sys/fs/cgroup/memory/0
405	# echo $$ > /sys/fs/cgroup/memory/0/tasks
406
4074. Since now we're in the 0 cgroup, we can alter the memory limit::
408
409	# echo 4M > /sys/fs/cgroup/memory/0/memory.limit_in_bytes
410
411   The limit can now be queried::
412
413	# cat /sys/fs/cgroup/memory/0/memory.limit_in_bytes
414	4194304
415
416.. note::
417   We can use a suffix (k, K, m, M, g or G) to indicate values in kilo,
418   mega or gigabytes. (Here, Kilo, Mega, Giga are Kibibytes, Mebibytes,
419   Gibibytes.)
420
421.. note::
422   We can write "-1" to reset the ``*.limit_in_bytes(unlimited)``.
423
424.. note::
425   We cannot set limits on the root cgroup any more.
426
427
428We can check the usage::
429
430  # cat /sys/fs/cgroup/memory/0/memory.usage_in_bytes
431  1216512
432
433A successful write to this file does not guarantee a successful setting of
434this limit to the value written into the file. This can be due to a
435number of factors, such as rounding up to page boundaries or the total
436availability of memory on the system. The user is required to re-read
437this file after a write to guarantee the value committed by the kernel::
438
439  # echo 1 > memory.limit_in_bytes
440  # cat memory.limit_in_bytes
441  4096
442
443The memory.failcnt field gives the number of times that the cgroup limit was
444exceeded.
445
446The memory.stat file gives accounting information. Now, the number of
447caches, RSS and Active pages/Inactive pages are shown.
448
4494. Testing
450==========
451
452For testing features and implementation, see memcg_test.txt.
453
454Performance test is also important. To see pure memory controller's overhead,
455testing on tmpfs will give you good numbers of small overheads.
456Example: do kernel make on tmpfs.
457
458Page-fault scalability is also important. At measuring parallel
459page fault test, multi-process test may be better than multi-thread
460test because it has noise of shared objects/status.
461
462But the above two are testing extreme situations.
463Trying usual test under memory controller is always helpful.
464
465.. _cgroup-v1-memory-test-troubleshoot:
466
4674.1 Troubleshooting
468-------------------
469
470Sometimes a user might find that the application under a cgroup is
471terminated by the OOM killer. There are several causes for this:
472
4731. The cgroup limit is too low (just too low to do anything useful)
4742. The user is using anonymous memory and swap is turned off or too low
475
476A sync followed by echo 1 > /proc/sys/vm/drop_caches will help get rid of
477some of the pages cached in the cgroup (page cache pages).
478
479To know what happens, disabling OOM_Kill as per :ref:`"10. OOM Control"
480<cgroup-v1-memory-oom-control>` (below) and seeing what happens will be
481helpful.
482
483.. _cgroup-v1-memory-test-task-migration:
484
4854.2 Task migration
486------------------
487
488When a task migrates from one cgroup to another, its charge is not
489carried forward by default. The pages allocated from the original cgroup still
490remain charged to it, the charge is dropped when the page is freed or
491reclaimed.
492
493You can move charges of a task along with task migration.
494See :ref:`8. "Move charges at task migration" <cgroup-v1-memory-move-charges>`
495
4964.3 Removing a cgroup
497---------------------
498
499A cgroup can be removed by rmdir, but as discussed in :ref:`sections 4.1
500<cgroup-v1-memory-test-troubleshoot>` and :ref:`4.2
501<cgroup-v1-memory-test-task-migration>`, a cgroup might have some charge
502associated with it, even though all tasks have migrated away from it. (because
503we charge against pages, not against tasks.)
504
505We move the stats to parent, and no change on the charge except uncharging
506from the child.
507
508Charges recorded in swap information is not updated at removal of cgroup.
509Recorded information is discarded and a cgroup which uses swap (swapcache)
510will be charged as a new owner of it.
511
5125. Misc. interfaces
513===================
514
5155.1 force_empty
516---------------
517  memory.force_empty interface is provided to make cgroup's memory usage empty.
518  When writing anything to this::
519
520    # echo 0 > memory.force_empty
521
522  the cgroup will be reclaimed and as many pages reclaimed as possible.
523
524  The typical use case for this interface is before calling rmdir().
525  Though rmdir() offlines memcg, but the memcg may still stay there due to
526  charged file caches. Some out-of-use page caches may keep charged until
527  memory pressure happens. If you want to avoid that, force_empty will be useful.
528
5295.2 stat file
530-------------
531
532memory.stat file includes following statistics:
533
534  * per-memory cgroup local status
535
536    =============== ===============================================================
537    cache           # of bytes of page cache memory.
538    rss             # of bytes of anonymous and swap cache memory (includes
539                    transparent hugepages).
540    rss_huge        # of bytes of anonymous transparent hugepages.
541    mapped_file     # of bytes of mapped file (includes tmpfs/shmem)
542    pgpgin          # of charging events to the memory cgroup. The charging
543                    event happens each time a page is accounted as either mapped
544                    anon page(RSS) or cache page(Page Cache) to the cgroup.
545    pgpgout         # of uncharging events to the memory cgroup. The uncharging
546                    event happens each time a page is unaccounted from the
547                    cgroup.
548    swap            # of bytes of swap usage
549    dirty           # of bytes that are waiting to get written back to the disk.
550    writeback       # of bytes of file/anon cache that are queued for syncing to
551                    disk.
552    inactive_anon   # of bytes of anonymous and swap cache memory on inactive
553                    LRU list.
554    active_anon     # of bytes of anonymous and swap cache memory on active
555                    LRU list.
556    inactive_file   # of bytes of file-backed memory and MADV_FREE anonymous
557                    memory (LazyFree pages) on inactive LRU list.
558    active_file     # of bytes of file-backed memory on active LRU list.
559    unevictable     # of bytes of memory that cannot be reclaimed (mlocked etc).
560    =============== ===============================================================
561
562  * status considering hierarchy (see memory.use_hierarchy settings):
563
564    ========================= ===================================================
565    hierarchical_memory_limit # of bytes of memory limit with regard to
566                              hierarchy
567                              under which the memory cgroup is
568    hierarchical_memsw_limit  # of bytes of memory+swap limit with regard to
569                              hierarchy under which memory cgroup is.
570
571    total_<counter>           # hierarchical version of <counter>, which in
572                              addition to the cgroup's own value includes the
573                              sum of all hierarchical children's values of
574                              <counter>, i.e. total_cache
575    ========================= ===================================================
576
577  * additional vm parameters (depends on CONFIG_DEBUG_VM):
578
579    ========================= ========================================
580    recent_rotated_anon       VM internal parameter. (see mm/vmscan.c)
581    recent_rotated_file       VM internal parameter. (see mm/vmscan.c)
582    recent_scanned_anon       VM internal parameter. (see mm/vmscan.c)
583    recent_scanned_file       VM internal parameter. (see mm/vmscan.c)
584    ========================= ========================================
585
586.. hint::
587	recent_rotated means recent frequency of LRU rotation.
588	recent_scanned means recent # of scans to LRU.
589	showing for better debug please see the code for meanings.
590
591.. note::
592	Only anonymous and swap cache memory is listed as part of 'rss' stat.
593	This should not be confused with the true 'resident set size' or the
594	amount of physical memory used by the cgroup.
595
596	'rss + mapped_file" will give you resident set size of cgroup.
597
598	(Note: file and shmem may be shared among other cgroups. In that case,
599	mapped_file is accounted only when the memory cgroup is owner of page
600	cache.)
601
6025.3 swappiness
603--------------
604
605Overrides /proc/sys/vm/swappiness for the particular group. The tunable
606in the root cgroup corresponds to the global swappiness setting.
607
608Please note that unlike during the global reclaim, limit reclaim
609enforces that 0 swappiness really prevents from any swapping even if
610there is a swap storage available. This might lead to memcg OOM killer
611if there are no file pages to reclaim.
612
6135.4 failcnt
614-----------
615
616A memory cgroup provides memory.failcnt and memory.memsw.failcnt files.
617This failcnt(== failure count) shows the number of times that a usage counter
618hit its limit. When a memory cgroup hits a limit, failcnt increases and
619memory under it will be reclaimed.
620
621You can reset failcnt by writing 0 to failcnt file::
622
623	# echo 0 > .../memory.failcnt
624
6255.5 usage_in_bytes
626------------------
627
628For efficiency, as other kernel components, memory cgroup uses some optimization
629to avoid unnecessary cacheline false sharing. usage_in_bytes is affected by the
630method and doesn't show 'exact' value of memory (and swap) usage, it's a fuzz
631value for efficient access. (Of course, when necessary, it's synchronized.)
632If you want to know more exact memory usage, you should use RSS+CACHE(+SWAP)
633value in memory.stat(see 5.2).
634
6355.6 numa_stat
636-------------
637
638This is similar to numa_maps but operates on a per-memcg basis.  This is
639useful for providing visibility into the numa locality information within
640an memcg since the pages are allowed to be allocated from any physical
641node.  One of the use cases is evaluating application performance by
642combining this information with the application's CPU allocation.
643
644Each memcg's numa_stat file includes "total", "file", "anon" and "unevictable"
645per-node page counts including "hierarchical_<counter>" which sums up all
646hierarchical children's values in addition to the memcg's own value.
647
648The output format of memory.numa_stat is::
649
650  total=<total pages> N0=<node 0 pages> N1=<node 1 pages> ...
651  file=<total file pages> N0=<node 0 pages> N1=<node 1 pages> ...
652  anon=<total anon pages> N0=<node 0 pages> N1=<node 1 pages> ...
653  unevictable=<total anon pages> N0=<node 0 pages> N1=<node 1 pages> ...
654  hierarchical_<counter>=<counter pages> N0=<node 0 pages> N1=<node 1 pages> ...
655
656The "total" count is sum of file + anon + unevictable.
657
6586. Hierarchy support
659====================
660
661The memory controller supports a deep hierarchy and hierarchical accounting.
662The hierarchy is created by creating the appropriate cgroups in the
663cgroup filesystem. Consider for example, the following cgroup filesystem
664hierarchy::
665
666	       root
667	     /  |   \
668            /	|    \
669	   a	b     c
670		      | \
671		      |  \
672		      d   e
673
674In the diagram above, with hierarchical accounting enabled, all memory
675usage of e, is accounted to its ancestors up until the root (i.e, c and root).
676If one of the ancestors goes over its limit, the reclaim algorithm reclaims
677from the tasks in the ancestor and the children of the ancestor.
678
6796.1 Hierarchical accounting and reclaim
680---------------------------------------
681
682Hierarchical accounting is enabled by default. Disabling the hierarchical
683accounting is deprecated. An attempt to do it will result in a failure
684and a warning printed to dmesg.
685
686For compatibility reasons writing 1 to memory.use_hierarchy will always pass::
687
688	# echo 1 > memory.use_hierarchy
689
6907. Soft limits
691==============
692
693Soft limits allow for greater sharing of memory. The idea behind soft limits
694is to allow control groups to use as much of the memory as needed, provided
695
696a. There is no memory contention
697b. They do not exceed their hard limit
698
699When the system detects memory contention or low memory, control groups
700are pushed back to their soft limits. If the soft limit of each control
701group is very high, they are pushed back as much as possible to make
702sure that one control group does not starve the others of memory.
703
704Please note that soft limits is a best-effort feature; it comes with
705no guarantees, but it does its best to make sure that when memory is
706heavily contended for, memory is allocated based on the soft limit
707hints/setup. Currently soft limit based reclaim is set up such that
708it gets invoked from balance_pgdat (kswapd).
709
7107.1 Interface
711-------------
712
713Soft limits can be setup by using the following commands (in this example we
714assume a soft limit of 256 MiB)::
715
716	# echo 256M > memory.soft_limit_in_bytes
717
718If we want to change this to 1G, we can at any time use::
719
720	# echo 1G > memory.soft_limit_in_bytes
721
722.. note::
723       Soft limits take effect over a long period of time, since they involve
724       reclaiming memory for balancing between memory cgroups
725
726.. note::
727       It is recommended to set the soft limit always below the hard limit,
728       otherwise the hard limit will take precedence.
729
730.. _cgroup-v1-memory-move-charges:
731
7328. Move charges at task migration (DEPRECATED!)
733===============================================
734
735THIS IS DEPRECATED!
736
737It's expensive and unreliable! It's better practice to launch workload
738tasks directly from inside their target cgroup. Use dedicated workload
739cgroups to allow fine-grained policy adjustments without having to
740move physical pages between control domains.
741
742Users can move charges associated with a task along with task migration, that
743is, uncharge task's pages from the old cgroup and charge them to the new cgroup.
744This feature is not supported in !CONFIG_MMU environments because of lack of
745page tables.
746
7478.1 Interface
748-------------
749
750This feature is disabled by default. It can be enabled (and disabled again) by
751writing to memory.move_charge_at_immigrate of the destination cgroup.
752
753If you want to enable it::
754
755	# echo (some positive value) > memory.move_charge_at_immigrate
756
757.. note::
758      Each bits of move_charge_at_immigrate has its own meaning about what type
759      of charges should be moved. See :ref:`section 8.2
760      <cgroup-v1-memory-movable-charges>` for details.
761
762.. note::
763      Charges are moved only when you move mm->owner, in other words,
764      a leader of a thread group.
765
766.. note::
767      If we cannot find enough space for the task in the destination cgroup, we
768      try to make space by reclaiming memory. Task migration may fail if we
769      cannot make enough space.
770
771.. note::
772      It can take several seconds if you move charges much.
773
774And if you want disable it again::
775
776	# echo 0 > memory.move_charge_at_immigrate
777
778.. _cgroup-v1-memory-movable-charges:
779
7808.2 Type of charges which can be moved
781--------------------------------------
782
783Each bit in move_charge_at_immigrate has its own meaning about what type of
784charges should be moved. But in any case, it must be noted that an account of
785a page or a swap can be moved only when it is charged to the task's current
786(old) memory cgroup.
787
788+---+--------------------------------------------------------------------------+
789|bit| what type of charges would be moved ?                                    |
790+===+==========================================================================+
791| 0 | A charge of an anonymous page (or swap of it) used by the target task.   |
792|   | You must enable Swap Extension (see 2.4) to enable move of swap charges. |
793+---+--------------------------------------------------------------------------+
794| 1 | A charge of file pages (normal file, tmpfs file (e.g. ipc shared memory) |
795|   | and swaps of tmpfs file) mmapped by the target task. Unlike the case of  |
796|   | anonymous pages, file pages (and swaps) in the range mmapped by the task |
797|   | will be moved even if the task hasn't done page fault, i.e. they might   |
798|   | not be the task's "RSS", but other task's "RSS" that maps the same file. |
799|   | And mapcount of the page is ignored (the page can be moved even if       |
800|   | page_mapcount(page) > 1). You must enable Swap Extension (see 2.4) to    |
801|   | enable move of swap charges.                                             |
802+---+--------------------------------------------------------------------------+
803
8048.3 TODO
805--------
806
807- All of moving charge operations are done under cgroup_mutex. It's not good
808  behavior to hold the mutex too long, so we may need some trick.
809
8109. Memory thresholds
811====================
812
813Memory cgroup implements memory thresholds using the cgroups notification
814API (see cgroups.txt). It allows to register multiple memory and memsw
815thresholds and gets notifications when it crosses.
816
817To register a threshold, an application must:
818
819- create an eventfd using eventfd(2);
820- open memory.usage_in_bytes or memory.memsw.usage_in_bytes;
821- write string like "<event_fd> <fd of memory.usage_in_bytes> <threshold>" to
822  cgroup.event_control.
823
824Application will be notified through eventfd when memory usage crosses
825threshold in any direction.
826
827It's applicable for root and non-root cgroup.
828
829.. _cgroup-v1-memory-oom-control:
830
83110. OOM Control
832===============
833
834memory.oom_control file is for OOM notification and other controls.
835
836Memory cgroup implements OOM notifier using the cgroup notification
837API (See cgroups.txt). It allows to register multiple OOM notification
838delivery and gets notification when OOM happens.
839
840To register a notifier, an application must:
841
842 - create an eventfd using eventfd(2)
843 - open memory.oom_control file
844 - write string like "<event_fd> <fd of memory.oom_control>" to
845   cgroup.event_control
846
847The application will be notified through eventfd when OOM happens.
848OOM notification doesn't work for the root cgroup.
849
850You can disable the OOM-killer by writing "1" to memory.oom_control file, as:
851
852	#echo 1 > memory.oom_control
853
854If OOM-killer is disabled, tasks under cgroup will hang/sleep
855in memory cgroup's OOM-waitqueue when they request accountable memory.
856
857For running them, you have to relax the memory cgroup's OOM status by
858
859	* enlarge limit or reduce usage.
860
861To reduce usage,
862
863	* kill some tasks.
864	* move some tasks to other group with account migration.
865	* remove some files (on tmpfs?)
866
867Then, stopped tasks will work again.
868
869At reading, current status of OOM is shown.
870
871	- oom_kill_disable 0 or 1
872	  (if 1, oom-killer is disabled)
873	- under_oom	   0 or 1
874	  (if 1, the memory cgroup is under OOM, tasks may be stopped.)
875        - oom_kill         integer counter
876          The number of processes belonging to this cgroup killed by any
877          kind of OOM killer.
878
87911. Memory Pressure
880===================
881
882The pressure level notifications can be used to monitor the memory
883allocation cost; based on the pressure, applications can implement
884different strategies of managing their memory resources. The pressure
885levels are defined as following:
886
887The "low" level means that the system is reclaiming memory for new
888allocations. Monitoring this reclaiming activity might be useful for
889maintaining cache level. Upon notification, the program (typically
890"Activity Manager") might analyze vmstat and act in advance (i.e.
891prematurely shutdown unimportant services).
892
893The "medium" level means that the system is experiencing medium memory
894pressure, the system might be making swap, paging out active file caches,
895etc. Upon this event applications may decide to further analyze
896vmstat/zoneinfo/memcg or internal memory usage statistics and free any
897resources that can be easily reconstructed or re-read from a disk.
898
899The "critical" level means that the system is actively thrashing, it is
900about to out of memory (OOM) or even the in-kernel OOM killer is on its
901way to trigger. Applications should do whatever they can to help the
902system. It might be too late to consult with vmstat or any other
903statistics, so it's advisable to take an immediate action.
904
905By default, events are propagated upward until the event is handled, i.e. the
906events are not pass-through. For example, you have three cgroups: A->B->C. Now
907you set up an event listener on cgroups A, B and C, and suppose group C
908experiences some pressure. In this situation, only group C will receive the
909notification, i.e. groups A and B will not receive it. This is done to avoid
910excessive "broadcasting" of messages, which disturbs the system and which is
911especially bad if we are low on memory or thrashing. Group B, will receive
912notification only if there are no event listers for group C.
913
914There are three optional modes that specify different propagation behavior:
915
916 - "default": this is the default behavior specified above. This mode is the
917   same as omitting the optional mode parameter, preserved by backwards
918   compatibility.
919
920 - "hierarchy": events always propagate up to the root, similar to the default
921   behavior, except that propagation continues regardless of whether there are
922   event listeners at each level, with the "hierarchy" mode. In the above
923   example, groups A, B, and C will receive notification of memory pressure.
924
925 - "local": events are pass-through, i.e. they only receive notifications when
926   memory pressure is experienced in the memcg for which the notification is
927   registered. In the above example, group C will receive notification if
928   registered for "local" notification and the group experiences memory
929   pressure. However, group B will never receive notification, regardless if
930   there is an event listener for group C or not, if group B is registered for
931   local notification.
932
933The level and event notification mode ("hierarchy" or "local", if necessary) are
934specified by a comma-delimited string, i.e. "low,hierarchy" specifies
935hierarchical, pass-through, notification for all ancestor memcgs. Notification
936that is the default, non pass-through behavior, does not specify a mode.
937"medium,local" specifies pass-through notification for the medium level.
938
939The file memory.pressure_level is only used to setup an eventfd. To
940register a notification, an application must:
941
942- create an eventfd using eventfd(2);
943- open memory.pressure_level;
944- write string as "<event_fd> <fd of memory.pressure_level> <level[,mode]>"
945  to cgroup.event_control.
946
947Application will be notified through eventfd when memory pressure is at
948the specific level (or higher). Read/write operations to
949memory.pressure_level are no implemented.
950
951Test:
952
953   Here is a small script example that makes a new cgroup, sets up a
954   memory limit, sets up a notification in the cgroup and then makes child
955   cgroup experience a critical pressure::
956
957	# cd /sys/fs/cgroup/memory/
958	# mkdir foo
959	# cd foo
960	# cgroup_event_listener memory.pressure_level low,hierarchy &
961	# echo 8000000 > memory.limit_in_bytes
962	# echo 8000000 > memory.memsw.limit_in_bytes
963	# echo $$ > tasks
964	# dd if=/dev/zero | read x
965
966   (Expect a bunch of notifications, and eventually, the oom-killer will
967   trigger.)
968
96912. TODO
970========
971
9721. Make per-cgroup scanner reclaim not-shared pages first
9732. Teach controller to account for shared-pages
9743. Start reclamation in the background when the limit is
975   not yet hit but the usage is getting closer
976
977Summary
978=======
979
980Overall, the memory controller has been a stable controller and has been
981commented and discussed quite extensively in the community.
982
983References
984==========
985
986.. [1] Singh, Balbir. RFC: Memory Controller, http://lwn.net/Articles/206697/
987.. [2] Singh, Balbir. Memory Controller (RSS Control),
988   http://lwn.net/Articles/222762/
989.. [3] Emelianov, Pavel. Resource controllers based on process cgroups
990   https://lore.kernel.org/r/45ED7DEC.7010403@sw.ru
991.. [4] Emelianov, Pavel. RSS controller based on process cgroups (v2)
992   https://lore.kernel.org/r/461A3010.90403@sw.ru
993.. [5] Emelianov, Pavel. RSS controller based on process cgroups (v3)
994   https://lore.kernel.org/r/465D9739.8070209@openvz.org
995
9966. Menage, Paul. Control Groups v10, http://lwn.net/Articles/236032/
9977. Vaidyanathan, Srinivasan, Control Groups: Pagecache accounting and control
998   subsystem (v3), http://lwn.net/Articles/235534/
9998. Singh, Balbir. RSS controller v2 test results (lmbench),
1000   https://lore.kernel.org/r/464C95D4.7070806@linux.vnet.ibm.com
10019. Singh, Balbir. RSS controller v2 AIM9 results
1002   https://lore.kernel.org/r/464D267A.50107@linux.vnet.ibm.com
100310. Singh, Balbir. Memory controller v6 test results,
1004    https://lore.kernel.org/r/20070819094658.654.84837.sendpatchset@balbir-laptop
1005
1006.. [11] Singh, Balbir. Memory controller introduction (v6),
1007   https://lore.kernel.org/r/20070817084228.26003.12568.sendpatchset@balbir-laptop
1008.. [12] Corbet, Jonathan, Controlling memory use in cgroups,
1009   http://lwn.net/Articles/243795/
1010