Linux kernel mirror (for testing)
git.kernel.org/pub/scm/linux/kernel/git/torvalds/linux.git
kernel
os
linux
1==========================
2Memory Resource Controller
3==========================
4
5.. caution::
6 This document is hopelessly outdated and it asks for a complete
7 rewrite. It still contains a useful information so we are keeping it
8 here but make sure to check the current code if you need a deeper
9 understanding.
10
11.. note::
12 The Memory Resource Controller has generically been referred to as the
13 memory controller in this document. Do not confuse memory controller
14 used here with the memory controller that is used in hardware.
15
16.. hint::
17 When we mention a cgroup (cgroupfs's directory) with memory controller,
18 we call it "memory cgroup". When you see git-log and source code, you'll
19 see patch's title and function names tend to use "memcg".
20 In this document, we avoid using it.
21
22Benefits and Purpose of the memory controller
23=============================================
24
25The memory controller isolates the memory behaviour of a group of tasks
26from the rest of the system. The article on LWN [12]_ mentions some probable
27uses of the memory controller. The memory controller can be used to
28
29a. Isolate an application or a group of applications
30 Memory-hungry applications can be isolated and limited to a smaller
31 amount of memory.
32b. Create a cgroup with a limited amount of memory; this can be used
33 as a good alternative to booting with mem=XXXX.
34c. Virtualization solutions can control the amount of memory they want
35 to assign to a virtual machine instance.
36d. A CD/DVD burner could control the amount of memory used by the
37 rest of the system to ensure that burning does not fail due to lack
38 of available memory.
39e. There are several other use cases; find one or use the controller just
40 for fun (to learn and hack on the VM subsystem).
41
42Current Status: linux-2.6.34-mmotm(development version of 2010/April)
43
44Features:
45
46 - accounting anonymous pages, file caches, swap caches usage and limiting them.
47 - pages are linked to per-memcg LRU exclusively, and there is no global LRU.
48 - optionally, memory+swap usage can be accounted and limited.
49 - hierarchical accounting
50 - soft limit
51 - moving (recharging) account at moving a task is selectable.
52 - usage threshold notifier
53 - memory pressure notifier
54 - oom-killer disable knob and oom-notifier
55 - Root cgroup has no limit controls.
56
57 Kernel memory support is a work in progress, and the current version provides
58 basically functionality. (See :ref:`section 2.7
59 <cgroup-v1-memory-kernel-extension>`)
60
61Brief summary of control files.
62
63==================================== ==========================================
64 tasks attach a task(thread) and show list of
65 threads
66 cgroup.procs show list of processes
67 cgroup.event_control an interface for event_fd()
68 This knob is not available on CONFIG_PREEMPT_RT systems.
69 memory.usage_in_bytes show current usage for memory
70 (See 5.5 for details)
71 memory.memsw.usage_in_bytes show current usage for memory+Swap
72 (See 5.5 for details)
73 memory.limit_in_bytes set/show limit of memory usage
74 memory.memsw.limit_in_bytes set/show limit of memory+Swap usage
75 memory.failcnt show the number of memory usage hits limits
76 memory.memsw.failcnt show the number of memory+Swap hits limits
77 memory.max_usage_in_bytes show max memory usage recorded
78 memory.memsw.max_usage_in_bytes show max memory+Swap usage recorded
79 memory.soft_limit_in_bytes set/show soft limit of memory usage
80 This knob is not available on CONFIG_PREEMPT_RT systems.
81 This knob is deprecated and shouldn't be
82 used.
83 memory.stat show various statistics
84 memory.use_hierarchy set/show hierarchical account enabled
85 This knob is deprecated and shouldn't be
86 used.
87 memory.force_empty trigger forced page reclaim
88 memory.pressure_level set memory pressure notifications
89 This knob is deprecated and shouldn't be
90 used.
91 memory.swappiness set/show swappiness parameter of vmscan
92 (See sysctl's vm.swappiness)
93 Per memcg knob does not exist in cgroup v2.
94 memory.move_charge_at_immigrate This knob is deprecated.
95 memory.oom_control set/show oom controls.
96 This knob is deprecated and shouldn't be
97 used.
98 memory.numa_stat show the number of memory usage per numa
99 node
100 memory.kmem.limit_in_bytes Deprecated knob to set and read the kernel
101 memory hard limit. Kernel hard limit is not
102 supported since 5.16. Writing any value to
103 do file will not have any effect same as if
104 nokmem kernel parameter was specified.
105 Kernel memory is still charged and reported
106 by memory.kmem.usage_in_bytes.
107 memory.kmem.usage_in_bytes show current kernel memory allocation
108 memory.kmem.failcnt show the number of kernel memory usage
109 hits limits
110 memory.kmem.max_usage_in_bytes show max kernel memory usage recorded
111
112 memory.kmem.tcp.limit_in_bytes set/show hard limit for tcp buf memory
113 This knob is deprecated and shouldn't be
114 used.
115 memory.kmem.tcp.usage_in_bytes show current tcp buf memory allocation
116 This knob is deprecated and shouldn't be
117 used.
118 memory.kmem.tcp.failcnt show the number of tcp buf memory usage
119 hits limits
120 This knob is deprecated and shouldn't be
121 used.
122 memory.kmem.tcp.max_usage_in_bytes show max tcp buf memory usage recorded
123 This knob is deprecated and shouldn't be
124 used.
125==================================== ==========================================
126
1271. History
128==========
129
130The memory controller has a long history. A request for comments for the memory
131controller was posted by Balbir Singh [1]_. At the time the RFC was posted
132there were several implementations for memory control. The goal of the
133RFC was to build consensus and agreement for the minimal features required
134for memory control. The first RSS controller was posted by Balbir Singh [2]_
135in Feb 2007. Pavel Emelianov [3]_ [4]_ [5]_ has since posted three versions
136of the RSS controller. At OLS, at the resource management BoF, everyone
137suggested that we handle both page cache and RSS together. Another request was
138raised to allow user space handling of OOM. The current memory controller is
139at version 6; it combines both mapped (RSS) and unmapped Page
140Cache Control [11]_.
141
1422. Memory Control
143=================
144
145Memory is a unique resource in the sense that it is present in a limited
146amount. If a task requires a lot of CPU processing, the task can spread
147its processing over a period of hours, days, months or years, but with
148memory, the same physical memory needs to be reused to accomplish the task.
149
150The memory controller implementation has been divided into phases. These
151are:
152
1531. Memory controller
1542. mlock(2) controller
1553. Kernel user memory accounting and slab control
1564. user mappings length controller
157
158The memory controller is the first controller developed.
159
1602.1. Design
161-----------
162
163The core of the design is a counter called the page_counter. The
164page_counter tracks the current memory usage and limit of the group of
165processes associated with the controller. Each cgroup has a memory controller
166specific data structure (mem_cgroup) associated with it.
167
1682.2. Accounting
169---------------
170
171.. code-block::
172 :caption: Figure 1: Hierarchy of Accounting
173
174 +--------------------+
175 | mem_cgroup |
176 | (page_counter) |
177 +--------------------+
178 / ^ \
179 / | \
180 +---------------+ | +---------------+
181 | mm_struct | |.... | mm_struct |
182 | | | | |
183 +---------------+ | +---------------+
184 |
185 + --------------+
186 |
187 +---------------+ +------+--------+
188 | page +----------> page_cgroup|
189 | | | |
190 +---------------+ +---------------+
191
192
193
194Figure 1 shows the important aspects of the controller
195
1961. Accounting happens per cgroup
1972. Each mm_struct knows about which cgroup it belongs to
1983. Each page has a pointer to the page_cgroup, which in turn knows the
199 cgroup it belongs to
200
201The accounting is done as follows: mem_cgroup_charge_common() is invoked to
202set up the necessary data structures and check if the cgroup that is being
203charged is over its limit. If it is, then reclaim is invoked on the cgroup.
204More details can be found in the reclaim section of this document.
205If everything goes well, a page meta-data-structure called page_cgroup is
206updated. page_cgroup has its own LRU on cgroup.
207(*) page_cgroup structure is allocated at boot/memory-hotplug time.
208
2092.2.1 Accounting details
210------------------------
211
212All mapped anon pages (RSS) and cache pages (Page Cache) are accounted.
213Some pages which are never reclaimable and will not be on the LRU
214are not accounted. We just account pages under usual VM management.
215
216RSS pages are accounted at page_fault unless they've already been accounted
217for earlier. A file page will be accounted for as Page Cache when it's
218inserted into inode (xarray). While it's mapped into the page tables of
219processes, duplicate accounting is carefully avoided.
220
221An RSS page is unaccounted when it's fully unmapped. A PageCache page is
222unaccounted when it's removed from xarray. Even if RSS pages are fully
223unmapped (by kswapd), they may exist as SwapCache in the system until they
224are really freed. Such SwapCaches are also accounted.
225A swapped-in page is accounted after adding into swapcache.
226
227Note: The kernel does swapin-readahead and reads multiple swaps at once.
228Since page's memcg recorded into swap whatever memsw enabled, the page will
229be accounted after swapin.
230
231At page migration, accounting information is kept.
232
233Note: we just account pages-on-LRU because our purpose is to control amount
234of used pages; not-on-LRU pages tend to be out-of-control from VM view.
235
2362.3 Shared Page Accounting
237--------------------------
238
239Shared pages are accounted on the basis of the first touch approach. The
240cgroup that first touches a page is accounted for the page. The principle
241behind this approach is that a cgroup that aggressively uses a shared
242page will eventually get charged for it (once it is uncharged from
243the cgroup that brought it in -- this will happen on memory pressure).
244
2452.4 Swap Extension
246--------------------------------------
247
248Swap usage is always recorded for each of cgroup. Swap Extension allows you to
249read and limit it.
250
251When CONFIG_SWAP is enabled, following files are added.
252
253 - memory.memsw.usage_in_bytes.
254 - memory.memsw.limit_in_bytes.
255
256memsw means memory+swap. Usage of memory+swap is limited by
257memsw.limit_in_bytes.
258
259Example: Assume a system with 4G of swap. A task which allocates 6G of memory
260(by mistake) under 2G memory limitation will use all swap.
261In this case, setting memsw.limit_in_bytes=3G will prevent bad use of swap.
262By using the memsw limit, you can avoid system OOM which can be caused by swap
263shortage.
264
2652.4.1 why 'memory+swap' rather than swap
266~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
267
268The global LRU(kswapd) can swap out arbitrary pages. Swap-out means
269to move account from memory to swap...there is no change in usage of
270memory+swap. In other words, when we want to limit the usage of swap without
271affecting global LRU, memory+swap limit is better than just limiting swap from
272an OS point of view.
273
2742.4.2. What happens when a cgroup hits memory.memsw.limit_in_bytes
275~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
276
277When a cgroup hits memory.memsw.limit_in_bytes, it's useless to do swap-out
278in this cgroup. Then, swap-out will not be done by cgroup routine and file
279caches are dropped. But as mentioned above, global LRU can do swapout memory
280from it for sanity of the system's memory management state. You can't forbid
281it by cgroup.
282
2832.5 Reclaim
284-----------
285
286Each cgroup maintains a per cgroup LRU which has the same structure as
287global VM. When a cgroup goes over its limit, we first try
288to reclaim memory from the cgroup so as to make space for the new
289pages that the cgroup has touched. If the reclaim is unsuccessful,
290an OOM routine is invoked to select and kill the bulkiest task in the
291cgroup. (See :ref:`10. OOM Control <cgroup-v1-memory-oom-control>` below.)
292
293The reclaim algorithm has not been modified for cgroups, except that
294pages that are selected for reclaiming come from the per-cgroup LRU
295list.
296
297.. note::
298 Reclaim does not work for the root cgroup, since we cannot set any
299 limits on the root cgroup.
300
301.. note::
302 When panic_on_oom is set to "2", the whole system will panic.
303
304When oom event notifier is registered, event will be delivered.
305(See :ref:`oom_control <cgroup-v1-memory-oom-control>` section)
306
3072.6 Locking
308-----------
309
310Lock order is as follows::
311
312 folio_lock
313 mm->page_table_lock or split pte_lock
314 mapping->i_pages lock
315 lruvec->lru_lock.
316
317Per-node-per-memcgroup LRU (cgroup's private LRU) is guarded by
318lruvec->lru_lock; the folio LRU flag is cleared before
319isolating a page from its LRU under lruvec->lru_lock.
320
321.. _cgroup-v1-memory-kernel-extension:
322
3232.7 Kernel Memory Extension
324-----------------------------------------------
325
326With the Kernel memory extension, the Memory Controller is able to limit
327the amount of kernel memory used by the system. Kernel memory is fundamentally
328different than user memory, since it can't be swapped out, which makes it
329possible to DoS the system by consuming too much of this precious resource.
330
331Kernel memory accounting is enabled for all memory cgroups by default. But
332it can be disabled system-wide by passing cgroup.memory=nokmem to the kernel
333at boot time. In this case, kernel memory will not be accounted at all.
334
335Kernel memory limits are not imposed for the root cgroup. Usage for the root
336cgroup may or may not be accounted. The memory used is accumulated into
337memory.kmem.usage_in_bytes, or in a separate counter when it makes sense.
338(currently only for tcp).
339
340The main "kmem" counter is fed into the main counter, so kmem charges will
341also be visible from the user counter.
342
343Currently no soft limit is implemented for kernel memory. It is future work
344to trigger slab reclaim when those limits are reached.
345
3462.7.1 Current Kernel Memory resources accounted
347-----------------------------------------------
348
349stack pages:
350 every process consumes some stack pages. By accounting into
351 kernel memory, we prevent new processes from being created when the kernel
352 memory usage is too high.
353
354slab pages:
355 pages allocated by the SLAB or SLUB allocator are tracked. A copy
356 of each kmem_cache is created every time the cache is touched by the first time
357 from inside the memcg. The creation is done lazily, so some objects can still be
358 skipped while the cache is being created. All objects in a slab page should
359 belong to the same memcg. This only fails to hold when a task is migrated to a
360 different memcg during the page allocation by the cache.
361
362sockets memory pressure:
363 some sockets protocols have memory pressure
364 thresholds. The Memory Controller allows them to be controlled individually
365 per cgroup, instead of globally.
366
367tcp memory pressure:
368 sockets memory pressure for the tcp protocol.
369
3702.7.2 Common use cases
371----------------------
372
373Because the "kmem" counter is fed to the main user counter, kernel memory can
374never be limited completely independently of user memory. Say "U" is the user
375limit, and "K" the kernel limit. There are three possible ways limits can be
376set:
377
378U != 0, K = unlimited:
379 This is the standard memcg limitation mechanism already present before kmem
380 accounting. Kernel memory is completely ignored.
381
382U != 0, K < U:
383 Kernel memory is a subset of the user memory. This setup is useful in
384 deployments where the total amount of memory per-cgroup is overcommitted.
385 Overcommitting kernel memory limits is definitely not recommended, since the
386 box can still run out of non-reclaimable memory.
387 In this case, the admin could set up K so that the sum of all groups is
388 never greater than the total memory, and freely set U at the cost of his
389 QoS.
390
391 .. warning::
392 In the current implementation, memory reclaim will NOT be triggered for
393 a cgroup when it hits K while staying below U, which makes this setup
394 impractical.
395
396U != 0, K >= U:
397 Since kmem charges will also be fed to the user counter and reclaim will be
398 triggered for the cgroup for both kinds of memory. This setup gives the
399 admin a unified view of memory, and it is also useful for people who just
400 want to track kernel memory usage.
401
4023. User Interface
403=================
404
405To use the user interface:
406
4071. Enable CONFIG_CGROUPS and CONFIG_MEMCG options
4082. Prepare the cgroups (see :ref:`Why are cgroups needed?
409 <cgroups-why-needed>` for the background information)::
410
411 # mount -t tmpfs none /sys/fs/cgroup
412 # mkdir /sys/fs/cgroup/memory
413 # mount -t cgroup none /sys/fs/cgroup/memory -o memory
414
4153. Make the new group and move bash into it::
416
417 # mkdir /sys/fs/cgroup/memory/0
418 # echo $$ > /sys/fs/cgroup/memory/0/tasks
419
4204. Since now we're in the 0 cgroup, we can alter the memory limit::
421
422 # echo 4M > /sys/fs/cgroup/memory/0/memory.limit_in_bytes
423
424 The limit can now be queried::
425
426 # cat /sys/fs/cgroup/memory/0/memory.limit_in_bytes
427 4194304
428
429.. note::
430 We can use a suffix (k, K, m, M, g or G) to indicate values in kilo,
431 mega or gigabytes. (Here, Kilo, Mega, Giga are Kibibytes, Mebibytes,
432 Gibibytes.)
433
434.. note::
435 We can write "-1" to reset the ``*.limit_in_bytes(unlimited)``.
436
437.. note::
438 We cannot set limits on the root cgroup any more.
439
440
441We can check the usage::
442
443 # cat /sys/fs/cgroup/memory/0/memory.usage_in_bytes
444 1216512
445
446A successful write to this file does not guarantee a successful setting of
447this limit to the value written into the file. This can be due to a
448number of factors, such as rounding up to page boundaries or the total
449availability of memory on the system. The user is required to re-read
450this file after a write to guarantee the value committed by the kernel::
451
452 # echo 1 > memory.limit_in_bytes
453 # cat memory.limit_in_bytes
454 4096
455
456The memory.failcnt field gives the number of times that the cgroup limit was
457exceeded.
458
459The memory.stat file gives accounting information. Now, the number of
460caches, RSS and Active pages/Inactive pages are shown.
461
4624. Testing
463==========
464
465For testing features and implementation, see memcg_test.txt.
466
467Performance test is also important. To see pure memory controller's overhead,
468testing on tmpfs will give you good numbers of small overheads.
469Example: do kernel make on tmpfs.
470
471Page-fault scalability is also important. At measuring parallel
472page fault test, multi-process test may be better than multi-thread
473test because it has noise of shared objects/status.
474
475But the above two are testing extreme situations.
476Trying usual test under memory controller is always helpful.
477
478.. _cgroup-v1-memory-test-troubleshoot:
479
4804.1 Troubleshooting
481-------------------
482
483Sometimes a user might find that the application under a cgroup is
484terminated by the OOM killer. There are several causes for this:
485
4861. The cgroup limit is too low (just too low to do anything useful)
4872. The user is using anonymous memory and swap is turned off or too low
488
489A sync followed by echo 1 > /proc/sys/vm/drop_caches will help get rid of
490some of the pages cached in the cgroup (page cache pages).
491
492To know what happens, disabling OOM_Kill as per :ref:`"10. OOM Control"
493<cgroup-v1-memory-oom-control>` (below) and seeing what happens will be
494helpful.
495
496.. _cgroup-v1-memory-test-task-migration:
497
4984.2 Task migration
499------------------
500
501When a task migrates from one cgroup to another, its charge is not
502carried forward by default. The pages allocated from the original cgroup still
503remain charged to it, the charge is dropped when the page is freed or
504reclaimed.
505
506You can move charges of a task along with task migration.
507See :ref:`8. "Move charges at task migration" <cgroup-v1-memory-move-charges>`
508
5094.3 Removing a cgroup
510---------------------
511
512A cgroup can be removed by rmdir, but as discussed in :ref:`sections 4.1
513<cgroup-v1-memory-test-troubleshoot>` and :ref:`4.2
514<cgroup-v1-memory-test-task-migration>`, a cgroup might have some charge
515associated with it, even though all tasks have migrated away from it. (because
516we charge against pages, not against tasks.)
517
518We move the stats to parent, and no change on the charge except uncharging
519from the child.
520
521Charges recorded in swap information is not updated at removal of cgroup.
522Recorded information is discarded and a cgroup which uses swap (swapcache)
523will be charged as a new owner of it.
524
5255. Misc. interfaces
526===================
527
5285.1 force_empty
529---------------
530 memory.force_empty interface is provided to make cgroup's memory usage empty.
531 When writing anything to this::
532
533 # echo 0 > memory.force_empty
534
535 the cgroup will be reclaimed and as many pages reclaimed as possible.
536
537 The typical use case for this interface is before calling rmdir().
538 Though rmdir() offlines memcg, but the memcg may still stay there due to
539 charged file caches. Some out-of-use page caches may keep charged until
540 memory pressure happens. If you want to avoid that, force_empty will be useful.
541
5425.2 stat file
543-------------
544
545memory.stat file includes following statistics:
546
547 * per-memory cgroup local status
548
549 =============== ===============================================================
550 cache # of bytes of page cache memory.
551 rss # of bytes of anonymous and swap cache memory (includes
552 transparent hugepages).
553 rss_huge # of bytes of anonymous transparent hugepages.
554 mapped_file # of bytes of mapped file (includes tmpfs/shmem)
555 pgpgin # of charging events to the memory cgroup. The charging
556 event happens each time a page is accounted as either mapped
557 anon page(RSS) or cache page(Page Cache) to the cgroup.
558 pgpgout # of uncharging events to the memory cgroup. The uncharging
559 event happens each time a page is unaccounted from the
560 cgroup.
561 swap # of bytes of swap usage
562 swapcached # of bytes of swap cached in memory
563 dirty # of bytes that are waiting to get written back to the disk.
564 writeback # of bytes of file/anon cache that are queued for syncing to
565 disk.
566 inactive_anon # of bytes of anonymous and swap cache memory on inactive
567 LRU list.
568 active_anon # of bytes of anonymous and swap cache memory on active
569 LRU list.
570 inactive_file # of bytes of file-backed memory and MADV_FREE anonymous
571 memory (LazyFree pages) on inactive LRU list.
572 active_file # of bytes of file-backed memory on active LRU list.
573 unevictable # of bytes of memory that cannot be reclaimed (mlocked etc).
574 =============== ===============================================================
575
576 * status considering hierarchy (see memory.use_hierarchy settings):
577
578 ========================= ===================================================
579 hierarchical_memory_limit # of bytes of memory limit with regard to
580 hierarchy
581 under which the memory cgroup is
582 hierarchical_memsw_limit # of bytes of memory+swap limit with regard to
583 hierarchy under which memory cgroup is.
584
585 total_<counter> # hierarchical version of <counter>, which in
586 addition to the cgroup's own value includes the
587 sum of all hierarchical children's values of
588 <counter>, i.e. total_cache
589 ========================= ===================================================
590
591 * additional vm parameters (depends on CONFIG_DEBUG_VM):
592
593 ========================= ========================================
594 recent_rotated_anon VM internal parameter. (see mm/vmscan.c)
595 recent_rotated_file VM internal parameter. (see mm/vmscan.c)
596 recent_scanned_anon VM internal parameter. (see mm/vmscan.c)
597 recent_scanned_file VM internal parameter. (see mm/vmscan.c)
598 ========================= ========================================
599
600.. hint::
601 recent_rotated means recent frequency of LRU rotation.
602 recent_scanned means recent # of scans to LRU.
603 showing for better debug please see the code for meanings.
604
605.. note::
606 Only anonymous and swap cache memory is listed as part of 'rss' stat.
607 This should not be confused with the true 'resident set size' or the
608 amount of physical memory used by the cgroup.
609
610 'rss + mapped_file" will give you resident set size of cgroup.
611
612 Note that some kernel configurations might account complete larger
613 allocations (e.g., THP) towards 'rss' and 'mapped_file', even if
614 only some, but not all that memory is mapped.
615
616 (Note: file and shmem may be shared among other cgroups. In that case,
617 mapped_file is accounted only when the memory cgroup is owner of page
618 cache.)
619
6205.3 swappiness
621--------------
622
623Overrides /proc/sys/vm/swappiness for the particular group. The tunable
624in the root cgroup corresponds to the global swappiness setting.
625
626Please note that unlike during the global reclaim, limit reclaim
627enforces that 0 swappiness really prevents from any swapping even if
628there is a swap storage available. This might lead to memcg OOM killer
629if there are no file pages to reclaim.
630
6315.4 failcnt
632-----------
633
634A memory cgroup provides memory.failcnt and memory.memsw.failcnt files.
635This failcnt(== failure count) shows the number of times that a usage counter
636hit its limit. When a memory cgroup hits a limit, failcnt increases and
637memory under it will be reclaimed.
638
639You can reset failcnt by writing 0 to failcnt file::
640
641 # echo 0 > .../memory.failcnt
642
6435.5 usage_in_bytes
644------------------
645
646For efficiency, as other kernel components, memory cgroup uses some optimization
647to avoid unnecessary cacheline false sharing. usage_in_bytes is affected by the
648method and doesn't show 'exact' value of memory (and swap) usage, it's a fuzz
649value for efficient access. (Of course, when necessary, it's synchronized.)
650If you want to know more exact memory usage, you should use RSS+CACHE(+SWAP)
651value in memory.stat(see 5.2).
652
6535.6 numa_stat
654-------------
655
656This is similar to numa_maps but operates on a per-memcg basis. This is
657useful for providing visibility into the numa locality information within
658an memcg since the pages are allowed to be allocated from any physical
659node. One of the use cases is evaluating application performance by
660combining this information with the application's CPU allocation.
661
662Each memcg's numa_stat file includes "total", "file", "anon" and "unevictable"
663per-node page counts including "hierarchical_<counter>" which sums up all
664hierarchical children's values in addition to the memcg's own value.
665
666The output format of memory.numa_stat is::
667
668 total=<total pages> N0=<node 0 pages> N1=<node 1 pages> ...
669 file=<total file pages> N0=<node 0 pages> N1=<node 1 pages> ...
670 anon=<total anon pages> N0=<node 0 pages> N1=<node 1 pages> ...
671 unevictable=<total anon pages> N0=<node 0 pages> N1=<node 1 pages> ...
672 hierarchical_<counter>=<counter pages> N0=<node 0 pages> N1=<node 1 pages> ...
673
674The "total" count is sum of file + anon + unevictable.
675
6766. Hierarchy support
677====================
678
679The memory controller supports a deep hierarchy and hierarchical accounting.
680The hierarchy is created by creating the appropriate cgroups in the
681cgroup filesystem. Consider for example, the following cgroup filesystem
682hierarchy::
683
684 root
685 / | \
686 / | \
687 a b c
688 | \
689 | \
690 d e
691
692In the diagram above, with hierarchical accounting enabled, all memory
693usage of e, is accounted to its ancestors up until the root (i.e, c and root).
694If one of the ancestors goes over its limit, the reclaim algorithm reclaims
695from the tasks in the ancestor and the children of the ancestor.
696
6976.1 Hierarchical accounting and reclaim
698---------------------------------------
699
700Hierarchical accounting is enabled by default. Disabling the hierarchical
701accounting is deprecated. An attempt to do it will result in a failure
702and a warning printed to dmesg.
703
704For compatibility reasons writing 1 to memory.use_hierarchy will always pass::
705
706 # echo 1 > memory.use_hierarchy
707
7087. Soft limits (DEPRECATED)
709===========================
710
711THIS IS DEPRECATED!
712
713Soft limits allow for greater sharing of memory. The idea behind soft limits
714is to allow control groups to use as much of the memory as needed, provided
715
716a. There is no memory contention
717b. They do not exceed their hard limit
718
719When the system detects memory contention or low memory, control groups
720are pushed back to their soft limits. If the soft limit of each control
721group is very high, they are pushed back as much as possible to make
722sure that one control group does not starve the others of memory.
723
724Please note that soft limits is a best-effort feature; it comes with
725no guarantees, but it does its best to make sure that when memory is
726heavily contended for, memory is allocated based on the soft limit
727hints/setup. Currently soft limit based reclaim is set up such that
728it gets invoked from balance_pgdat (kswapd).
729
7307.1 Interface
731-------------
732
733Soft limits can be setup by using the following commands (in this example we
734assume a soft limit of 256 MiB)::
735
736 # echo 256M > memory.soft_limit_in_bytes
737
738If we want to change this to 1G, we can at any time use::
739
740 # echo 1G > memory.soft_limit_in_bytes
741
742.. note::
743 Soft limits take effect over a long period of time, since they involve
744 reclaiming memory for balancing between memory cgroups
745
746.. note::
747 It is recommended to set the soft limit always below the hard limit,
748 otherwise the hard limit will take precedence.
749
750.. _cgroup-v1-memory-move-charges:
751
7528. Move charges at task migration (DEPRECATED!)
753===============================================
754
755THIS IS DEPRECATED!
756
757Reading memory.move_charge_at_immigrate will always return 0 and writing
758to it will always return -EINVAL.
759
7609. Memory thresholds
761====================
762
763Memory cgroup implements memory thresholds using the cgroups notification
764API (see cgroups.txt). It allows to register multiple memory and memsw
765thresholds and gets notifications when it crosses.
766
767To register a threshold, an application must:
768
769- create an eventfd using eventfd(2);
770- open memory.usage_in_bytes or memory.memsw.usage_in_bytes;
771- write string like "<event_fd> <fd of memory.usage_in_bytes> <threshold>" to
772 cgroup.event_control.
773
774Application will be notified through eventfd when memory usage crosses
775threshold in any direction.
776
777It's applicable for root and non-root cgroup.
778
779.. _cgroup-v1-memory-oom-control:
780
78110. OOM Control (DEPRECATED)
782============================
783
784THIS IS DEPRECATED!
785
786memory.oom_control file is for OOM notification and other controls.
787
788Memory cgroup implements OOM notifier using the cgroup notification
789API (See cgroups.txt). It allows to register multiple OOM notification
790delivery and gets notification when OOM happens.
791
792To register a notifier, an application must:
793
794 - create an eventfd using eventfd(2)
795 - open memory.oom_control file
796 - write string like "<event_fd> <fd of memory.oom_control>" to
797 cgroup.event_control
798
799The application will be notified through eventfd when OOM happens.
800OOM notification doesn't work for the root cgroup.
801
802You can disable the OOM-killer by writing "1" to memory.oom_control file, as:
803
804 #echo 1 > memory.oom_control
805
806If OOM-killer is disabled, tasks under cgroup will hang/sleep
807in memory cgroup's OOM-waitqueue when they request accountable memory.
808
809For running them, you have to relax the memory cgroup's OOM status by
810
811 * enlarge limit or reduce usage.
812
813To reduce usage,
814
815 * kill some tasks.
816 * move some tasks to other group with account migration.
817 * remove some files (on tmpfs?)
818
819Then, stopped tasks will work again.
820
821At reading, current status of OOM is shown.
822
823 - oom_kill_disable 0 or 1
824 (if 1, oom-killer is disabled)
825 - under_oom 0 or 1
826 (if 1, the memory cgroup is under OOM, tasks may be stopped.)
827 - oom_kill integer counter
828 The number of processes belonging to this cgroup killed by any
829 kind of OOM killer.
830
83111. Memory Pressure (DEPRECATED)
832================================
833
834THIS IS DEPRECATED!
835
836The pressure level notifications can be used to monitor the memory
837allocation cost; based on the pressure, applications can implement
838different strategies of managing their memory resources. The pressure
839levels are defined as following:
840
841The "low" level means that the system is reclaiming memory for new
842allocations. Monitoring this reclaiming activity might be useful for
843maintaining cache level. Upon notification, the program (typically
844"Activity Manager") might analyze vmstat and act in advance (i.e.
845prematurely shutdown unimportant services).
846
847The "medium" level means that the system is experiencing medium memory
848pressure, the system might be making swap, paging out active file caches,
849etc. Upon this event applications may decide to further analyze
850vmstat/zoneinfo/memcg or internal memory usage statistics and free any
851resources that can be easily reconstructed or re-read from a disk.
852
853The "critical" level means that the system is actively thrashing, it is
854about to out of memory (OOM) or even the in-kernel OOM killer is on its
855way to trigger. Applications should do whatever they can to help the
856system. It might be too late to consult with vmstat or any other
857statistics, so it's advisable to take an immediate action.
858
859By default, events are propagated upward until the event is handled, i.e. the
860events are not pass-through. For example, you have three cgroups: A->B->C. Now
861you set up an event listener on cgroups A, B and C, and suppose group C
862experiences some pressure. In this situation, only group C will receive the
863notification, i.e. groups A and B will not receive it. This is done to avoid
864excessive "broadcasting" of messages, which disturbs the system and which is
865especially bad if we are low on memory or thrashing. Group B, will receive
866notification only if there are no event listeners for group C.
867
868There are three optional modes that specify different propagation behavior:
869
870 - "default": this is the default behavior specified above. This mode is the
871 same as omitting the optional mode parameter, preserved by backwards
872 compatibility.
873
874 - "hierarchy": events always propagate up to the root, similar to the default
875 behavior, except that propagation continues regardless of whether there are
876 event listeners at each level, with the "hierarchy" mode. In the above
877 example, groups A, B, and C will receive notification of memory pressure.
878
879 - "local": events are pass-through, i.e. they only receive notifications when
880 memory pressure is experienced in the memcg for which the notification is
881 registered. In the above example, group C will receive notification if
882 registered for "local" notification and the group experiences memory
883 pressure. However, group B will never receive notification, regardless if
884 there is an event listener for group C or not, if group B is registered for
885 local notification.
886
887The level and event notification mode ("hierarchy" or "local", if necessary) are
888specified by a comma-delimited string, i.e. "low,hierarchy" specifies
889hierarchical, pass-through, notification for all ancestor memcgs. Notification
890that is the default, non pass-through behavior, does not specify a mode.
891"medium,local" specifies pass-through notification for the medium level.
892
893The file memory.pressure_level is only used to setup an eventfd. To
894register a notification, an application must:
895
896- create an eventfd using eventfd(2);
897- open memory.pressure_level;
898- write string as "<event_fd> <fd of memory.pressure_level> <level[,mode]>"
899 to cgroup.event_control.
900
901Application will be notified through eventfd when memory pressure is at
902the specific level (or higher). Read/write operations to
903memory.pressure_level are no implemented.
904
905Test:
906
907 Here is a small script example that makes a new cgroup, sets up a
908 memory limit, sets up a notification in the cgroup and then makes child
909 cgroup experience a critical pressure::
910
911 # cd /sys/fs/cgroup/memory/
912 # mkdir foo
913 # cd foo
914 # cgroup_event_listener memory.pressure_level low,hierarchy &
915 # echo 8000000 > memory.limit_in_bytes
916 # echo 8000000 > memory.memsw.limit_in_bytes
917 # echo $$ > tasks
918 # dd if=/dev/zero | read x
919
920 (Expect a bunch of notifications, and eventually, the oom-killer will
921 trigger.)
922
92312. TODO
924========
925
9261. Make per-cgroup scanner reclaim not-shared pages first
9272. Teach controller to account for shared-pages
9283. Start reclamation in the background when the limit is
929 not yet hit but the usage is getting closer
930
931Summary
932=======
933
934Overall, the memory controller has been a stable controller and has been
935commented and discussed quite extensively in the community.
936
937References
938==========
939
940.. [1] Singh, Balbir. RFC: Memory Controller, http://lwn.net/Articles/206697/
941.. [2] Singh, Balbir. Memory Controller (RSS Control),
942 http://lwn.net/Articles/222762/
943.. [3] Emelianov, Pavel. Resource controllers based on process cgroups
944 https://lore.kernel.org/r/45ED7DEC.7010403@sw.ru
945.. [4] Emelianov, Pavel. RSS controller based on process cgroups (v2)
946 https://lore.kernel.org/r/461A3010.90403@sw.ru
947.. [5] Emelianov, Pavel. RSS controller based on process cgroups (v3)
948 https://lore.kernel.org/r/465D9739.8070209@openvz.org
949
9506. Menage, Paul. Control Groups v10, http://lwn.net/Articles/236032/
9517. Vaidyanathan, Srinivasan, Control Groups: Pagecache accounting and control
952 subsystem (v3), http://lwn.net/Articles/235534/
9538. Singh, Balbir. RSS controller v2 test results (lmbench),
954 https://lore.kernel.org/r/464C95D4.7070806@linux.vnet.ibm.com
9559. Singh, Balbir. RSS controller v2 AIM9 results
956 https://lore.kernel.org/r/464D267A.50107@linux.vnet.ibm.com
95710. Singh, Balbir. Memory controller v6 test results,
958 https://lore.kernel.org/r/20070819094658.654.84837.sendpatchset@balbir-laptop
959
960.. [11] Singh, Balbir. Memory controller introduction (v6),
961 https://lore.kernel.org/r/20070817084228.26003.12568.sendpatchset@balbir-laptop
962.. [12] Corbet, Jonathan, Controlling memory use in cgroups,
963 http://lwn.net/Articles/243795/