1 Memory Resource Controller
3 NOTE: The Memory Resource Controller has been generically been referred
4 to as the memory controller in this document. Do not confuse memory controller
5 used here with the memory controller that is used in hardware.
9 a. Enable control of Anonymous, Page Cache (mapped and unmapped) and
10 Swap Cache memory pages.
11 b. The infrastructure allows easy addition of other types of memory to control
12 c. Provides *zero overhead* for non memory controller users
13 d. Provides a double LRU: global memory pressure causes reclaim from the
14 global LRU; a cgroup on hitting a limit, reclaims from the per
17 Benefits and Purpose of the memory controller
19 The memory controller isolates the memory behaviour of a group of tasks
20 from the rest of the system. The article on LWN [12] mentions some probable
21 uses of the memory controller. The memory controller can be used to
23 a. Isolate an application or a group of applications
24 Memory hungry applications can be isolated and limited to a smaller
26 b. Create a cgroup with limited amount of memory, this can be used
27 as a good alternative to booting with mem=XXXX.
28 c. Virtualization solutions can control the amount of memory they want
29 to assign to a virtual machine instance.
30 d. A CD/DVD burner could control the amount of memory used by the
31 rest of the system to ensure that burning does not fail due to lack
33 e. There are several other use cases, find one or use the controller just
34 for fun (to learn and hack on the VM subsystem).
38 The memory controller has a long history. A request for comments for the memory
39 controller was posted by Balbir Singh [1]. At the time the RFC was posted
40 there were several implementations for memory control. The goal of the
41 RFC was to build consensus and agreement for the minimal features required
42 for memory control. The first RSS controller was posted by Balbir Singh[2]
43 in Feb 2007. Pavel Emelianov [3][4][5] has since posted three versions of the
44 RSS controller. At OLS, at the resource management BoF, everyone suggested
45 that we handle both page cache and RSS together. Another request was raised
46 to allow user space handling of OOM. The current memory controller is
47 at version 6; it combines both mapped (RSS) and unmapped Page
52 Memory is a unique resource in the sense that it is present in a limited
53 amount. If a task requires a lot of CPU processing, the task can spread
54 its processing over a period of hours, days, months or years, but with
55 memory, the same physical memory needs to be reused to accomplish the task.
57 The memory controller implementation has been divided into phases. These
61 2. mlock(2) controller
62 3. Kernel user memory accounting and slab control
63 4. user mappings length controller
65 The memory controller is the first controller developed.
69 The core of the design is a counter called the res_counter. The res_counter
70 tracks the current memory usage and limit of the group of processes associated
71 with the controller. Each cgroup has a memory controller specific data
72 structure (mem_cgroup) associated with it.
76 +--------------------+
79 +--------------------+
82 +---------------+ | +---------------+
83 | mm_struct | |.... | mm_struct |
85 +---------------+ | +---------------+
89 +---------------+ +------+--------+
90 | page +----------> page_cgroup|
92 +---------------+ +---------------+
94 (Figure 1: Hierarchy of Accounting)
97 Figure 1 shows the important aspects of the controller
99 1. Accounting happens per cgroup
100 2. Each mm_struct knows about which cgroup it belongs to
101 3. Each page has a pointer to the page_cgroup, which in turn knows the
104 The accounting is done as follows: mem_cgroup_charge() is invoked to setup
105 the necessary data structures and check if the cgroup that is being charged
106 is over its limit. If it is then reclaim is invoked on the cgroup.
107 More details can be found in the reclaim section of this document.
108 If everything goes well, a page meta-data-structure called page_cgroup is
109 allocated and associated with the page. This routine also adds the page to
112 2.2.1 Accounting details
114 All mapped anon pages (RSS) and cache pages (Page Cache) are accounted.
115 (some pages which never be reclaimable and will not be on global LRU
116 are not accounted. we just accounts pages under usual vm management.)
118 RSS pages are accounted at page_fault unless they've already been accounted
119 for earlier. A file page will be accounted for as Page Cache when it's
120 inserted into inode (radix-tree). While it's mapped into the page tables of
121 processes, duplicate accounting is carefully avoided.
123 A RSS page is unaccounted when it's fully unmapped. A PageCache page is
124 unaccounted when it's removed from radix-tree.
126 At page migration, accounting information is kept.
128 Note: we just account pages-on-lru because our purpose is to control amount
129 of used pages. not-on-lru pages are tend to be out-of-control from vm view.
131 2.3 Shared Page Accounting
133 Shared pages are accounted on the basis of the first touch approach. The
134 cgroup that first touches a page is accounted for the page. The principle
135 behind this approach is that a cgroup that aggressively uses a shared
136 page will eventually get charged for it (once it is uncharged from
137 the cgroup that brought it in -- this will happen on memory pressure).
139 Exception: If CONFIG_CGROUP_CGROUP_MEM_RES_CTLR_SWAP is not used..
140 When you do swapoff and make swapped-out pages of shmem(tmpfs) to
141 be backed into memory in force, charges for pages are accounted against the
142 caller of swapoff rather than the users of shmem.
145 2.4 Swap Extension (CONFIG_CGROUP_MEM_RES_CTLR_SWAP)
146 Swap Extension allows you to record charge for swap. A swapped-in page is
147 charged back to original page allocator if possible.
149 When swap is accounted, following files are added.
150 - memory.memsw.usage_in_bytes.
151 - memory.memsw.limit_in_bytes.
153 usage of mem+swap is limited by memsw.limit_in_bytes.
155 * why 'mem+swap' rather than swap.
156 The global LRU(kswapd) can swap out arbitrary pages. Swap-out means
157 to move account from memory to swap...there is no change in usage of
158 mem+swap. In other words, when we want to limit the usage of swap without
159 affecting global LRU, mem+swap limit is better than just limiting swap from
162 * What happens when a cgroup hits memory.memsw.limit_in_bytes
163 When a cgroup his memory.memsw.limit_in_bytes, it's useless to do swap-out
164 in this cgroup. Then, swap-out will not be done by cgroup routine and file
165 caches are dropped. But as mentioned above, global LRU can do swapout memory
166 from it for sanity of the system's memory management state. You can't forbid
171 Each cgroup maintains a per cgroup LRU that consists of an active
172 and inactive list. When a cgroup goes over its limit, we first try
173 to reclaim memory from the cgroup so as to make space for the new
174 pages that the cgroup has touched. If the reclaim is unsuccessful,
175 an OOM routine is invoked to select and kill the bulkiest task in the
178 The reclaim algorithm has not been modified for cgroups, except that
179 pages that are selected for reclaiming come from the per cgroup LRU
182 NOTE: Reclaim does not work for the root cgroup, since we cannot set any
183 limits on the root cgroup.
185 Note2: When panic_on_oom is set to "2", the whole system will panic.
187 When oom event notifier is registered, event will be delivered.
188 (See oom_control section)
192 The memory controller uses the following hierarchy
194 1. zone->lru_lock is used for selecting pages to be isolated
195 2. mem->per_zone->lru_lock protects the per cgroup LRU (per zone)
196 3. lock_page_cgroup() is used to protect page->page_cgroup
202 a. Enable CONFIG_CGROUPS
203 b. Enable CONFIG_RESOURCE_COUNTERS
204 c. Enable CONFIG_CGROUP_MEM_RES_CTLR
206 1. Prepare the cgroups
208 # mount -t cgroup none /cgroups -o memory
210 2. Make the new group and move bash into it
212 # echo $$ > /cgroups/0/tasks
214 Since now we're in the 0 cgroup,
215 We can alter the memory limit:
216 # echo 4M > /cgroups/0/memory.limit_in_bytes
218 NOTE: We can use a suffix (k, K, m, M, g or G) to indicate values in kilo,
220 NOTE: We can write "-1" to reset the *.limit_in_bytes(unlimited).
221 NOTE: We cannot set limits on the root cgroup any more.
223 # cat /cgroups/0/memory.limit_in_bytes
226 NOTE: The interface has now changed to display the usage in bytes
229 We can check the usage:
230 # cat /cgroups/0/memory.usage_in_bytes
233 A successful write to this file does not guarantee a successful set of
234 this limit to the value written into the file. This can be due to a
235 number of factors, such as rounding up to page boundaries or the total
236 availability of memory on the system. The user is required to re-read
237 this file after a write to guarantee the value committed by the kernel.
239 # echo 1 > memory.limit_in_bytes
240 # cat memory.limit_in_bytes
243 The memory.failcnt field gives the number of times that the cgroup limit was
246 The memory.stat file gives accounting information. Now, the number of
247 caches, RSS and Active pages/Inactive pages are shown.
251 Balbir posted lmbench, AIM9, LTP and vmmstress results [10] and [11].
252 Apart from that v6 has been tested with several applications and regular
253 daily use. The controller has also been tested on the PPC64, x86_64 and
258 Sometimes a user might find that the application under a cgroup is
259 terminated. There are several causes for this:
261 1. The cgroup limit is too low (just too low to do anything useful)
262 2. The user is using anonymous memory and swap is turned off or too low
264 A sync followed by echo 1 > /proc/sys/vm/drop_caches will help get rid of
265 some of the pages cached in the cgroup (page cache pages).
269 When a task migrates from one cgroup to another, its charge is not
270 carried forward by default. The pages allocated from the original cgroup still
271 remain charged to it, the charge is dropped when the page is freed or
274 Note: You can move charges of a task along with task migration. See 8.
276 4.3 Removing a cgroup
278 A cgroup can be removed by rmdir, but as discussed in sections 4.1 and 4.2, a
279 cgroup might have some charge associated with it, even though all
280 tasks have migrated away from it.
281 Such charges are freed(at default) or moved to its parent. When moved,
282 both of RSS and CACHES are moved to parent.
283 If both of them are busy, rmdir() returns -EBUSY. See 5.1 Also.
285 Charges recorded in swap information is not updated at removal of cgroup.
286 Recorded information is discarded and a cgroup which uses swap (swapcache)
287 will be charged as a new owner of it.
293 memory.force_empty interface is provided to make cgroup's memory usage empty.
294 You can use this interface only when the cgroup has no tasks.
295 When writing anything to this
297 # echo 0 > memory.force_empty
299 Almost all pages tracked by this memcg will be unmapped and freed. Some of
300 pages cannot be freed because it's locked or in-use. Such pages are moved
301 to parent and this cgroup will be empty. But this may return -EBUSY in
304 Typical use case of this interface is that calling this before rmdir().
305 Because rmdir() moves all pages to parent, some out-of-use page caches can be
306 moved to the parent. If you want to avoid that, force_empty will be useful.
310 memory.stat file includes following statistics
312 cache - # of bytes of page cache memory.
313 rss - # of bytes of anonymous and swap cache memory.
314 pgpgin - # of pages paged in (equivalent to # of charging events).
315 pgpgout - # of pages paged out (equivalent to # of uncharging events).
316 active_anon - # of bytes of anonymous and swap cache memory on active
318 inactive_anon - # of bytes of anonymous memory and swap cache memory on
320 active_file - # of bytes of file-backed memory on active lru list.
321 inactive_file - # of bytes of file-backed memory on inactive lru list.
322 unevictable - # of bytes of memory that cannot be reclaimed (mlocked etc).
324 The following additional stats are dependent on CONFIG_DEBUG_VM.
326 inactive_ratio - VM internal parameter. (see mm/page_alloc.c)
327 recent_rotated_anon - VM internal parameter. (see mm/vmscan.c)
328 recent_rotated_file - VM internal parameter. (see mm/vmscan.c)
329 recent_scanned_anon - VM internal parameter. (see mm/vmscan.c)
330 recent_scanned_file - VM internal parameter. (see mm/vmscan.c)
333 recent_rotated means recent frequency of lru rotation.
334 recent_scanned means recent # of scans to lru.
335 showing for better debug please see the code for meanings.
338 Only anonymous and swap cache memory is listed as part of 'rss' stat.
339 This should not be confused with the true 'resident set size' or the
340 amount of physical memory used by the cgroup. Per-cgroup rss
341 accounting is not done yet.
344 Similar to /proc/sys/vm/swappiness, but affecting a hierarchy of groups only.
346 Following cgroups' swappiness can't be changed.
347 - root cgroup (uses /proc/sys/vm/swappiness).
348 - a cgroup which uses hierarchy and it has child cgroup.
349 - a cgroup which uses hierarchy and not the root of hierarchy.
354 The memory controller supports a deep hierarchy and hierarchical accounting.
355 The hierarchy is created by creating the appropriate cgroups in the
356 cgroup filesystem. Consider for example, the following cgroup filesystem
367 In the diagram above, with hierarchical accounting enabled, all memory
368 usage of e, is accounted to its ancestors up until the root (i.e, c and root),
369 that has memory.use_hierarchy enabled. If one of the ancestors goes over its
370 limit, the reclaim algorithm reclaims from the tasks in the ancestor and the
371 children of the ancestor.
373 6.1 Enabling hierarchical accounting and reclaim
375 The memory controller by default disables the hierarchy feature. Support
376 can be enabled by writing 1 to memory.use_hierarchy file of the root cgroup
378 # echo 1 > memory.use_hierarchy
380 The feature can be disabled by
382 # echo 0 > memory.use_hierarchy
384 NOTE1: Enabling/disabling will fail if the cgroup already has other
385 cgroups created below it.
387 NOTE2: When panic_on_oom is set to "2", the whole system will panic in
388 case of an oom event in any cgroup.
392 Soft limits allow for greater sharing of memory. The idea behind soft limits
393 is to allow control groups to use as much of the memory as needed, provided
395 a. There is no memory contention
396 b. They do not exceed their hard limit
398 When the system detects memory contention or low memory control groups
399 are pushed back to their soft limits. If the soft limit of each control
400 group is very high, they are pushed back as much as possible to make
401 sure that one control group does not starve the others of memory.
403 Please note that soft limits is a best effort feature, it comes with
404 no guarantees, but it does its best to make sure that when memory is
405 heavily contended for, memory is allocated based on the soft limit
406 hints/setup. Currently soft limit based reclaim is setup such that
407 it gets invoked from balance_pgdat (kswapd).
411 Soft limits can be setup by using the following commands (in this example we
412 assume a soft limit of 256 megabytes)
414 # echo 256M > memory.soft_limit_in_bytes
416 If we want to change this to 1G, we can at any time use
418 # echo 1G > memory.soft_limit_in_bytes
420 NOTE1: Soft limits take effect over a long period of time, since they involve
421 reclaiming memory for balancing between memory cgroups
422 NOTE2: It is recommended to set the soft limit always below the hard limit,
423 otherwise the hard limit will take precedence.
425 8. Move charges at task migration
427 Users can move charges associated with a task along with task migration, that
428 is, uncharge task's pages from the old cgroup and charge them to the new cgroup.
429 This feature is not supported in !CONFIG_MMU environments because of lack of
434 This feature is disabled by default. It can be enabled(and disabled again) by
435 writing to memory.move_charge_at_immigrate of the destination cgroup.
437 If you want to enable it:
439 # echo (some positive value) > memory.move_charge_at_immigrate
441 Note: Each bits of move_charge_at_immigrate has its own meaning about what type
442 of charges should be moved. See 8.2 for details.
443 Note: Charges are moved only when you move mm->owner, IOW, a leader of a thread
445 Note: If we cannot find enough space for the task in the destination cgroup, we
446 try to make space by reclaiming memory. Task migration may fail if we
447 cannot make enough space.
448 Note: It can take several seconds if you move charges in giga bytes order.
450 And if you want disable it again:
452 # echo 0 > memory.move_charge_at_immigrate
454 8.2 Type of charges which can be move
456 Each bits of move_charge_at_immigrate has its own meaning about what type of
457 charges should be moved.
459 bit | what type of charges would be moved ?
460 -----+------------------------------------------------------------------------
461 0 | A charge of an anonymous page(or swap of it) used by the target task.
462 | Those pages and swaps must be used only by the target task. You must
463 | enable Swap Extension(see 2.4) to enable move of swap charges.
465 Note: Those pages and swaps must be charged to the old cgroup.
466 Note: More type of pages(e.g. file cache, shmem,) will be supported by other
471 - Add support for other types of pages(e.g. file cache, shmem, etc.).
472 - Implement madvise(2) to let users decide the vma to be moved or not to be
474 - All of moving charge operations are done under cgroup_mutex. It's not good
475 behavior to hold the mutex too long, so we may need some trick.
479 Memory controler implements memory thresholds using cgroups notification
480 API (see cgroups.txt). It allows to register multiple memory and memsw
481 thresholds and gets notifications when it crosses.
483 To register a threshold application need:
484 - create an eventfd using eventfd(2);
485 - open memory.usage_in_bytes or memory.memsw.usage_in_bytes;
486 - write string like "<event_fd> <memory.usage_in_bytes> <threshold>" to
487 cgroup.event_control.
489 Application will be notified through eventfd when memory usage crosses
490 threshold in any direction.
492 It's applicable for root and non-root cgroup.
496 Memory controler implements oom notifier using cgroup notification
497 API (See cgroups.txt). It allows to register multiple oom notification
498 delivery and gets notification when oom happens.
500 To register a notifier, application need:
501 - create an eventfd using eventfd(2)
502 - open memory.oom_control file
503 - write string like "<event_fd> <memory.oom_control>" to cgroup.event_control
505 Application will be notifier through eventfd when oom happens.
506 OOM notification doesn't work for root cgroup.
511 1. Add support for accounting huge pages (as a separate controller)
512 2. Make per-cgroup scanner reclaim not-shared pages first
513 3. Teach controller to account for shared-pages
514 4. Start reclamation in the background when the limit is
515 not yet hit but the usage is getting closer
519 Overall, the memory controller has been a stable controller and has been
520 commented and discussed quite extensively in the community.
524 1. Singh, Balbir. RFC: Memory Controller, http://lwn.net/Articles/206697/
525 2. Singh, Balbir. Memory Controller (RSS Control),
526 http://lwn.net/Articles/222762/
527 3. Emelianov, Pavel. Resource controllers based on process cgroups
528 http://lkml.org/lkml/2007/3/6/198
529 4. Emelianov, Pavel. RSS controller based on process cgroups (v2)
530 http://lkml.org/lkml/2007/4/9/78
531 5. Emelianov, Pavel. RSS controller based on process cgroups (v3)
532 http://lkml.org/lkml/2007/5/30/244
533 6. Menage, Paul. Control Groups v10, http://lwn.net/Articles/236032/
534 7. Vaidyanathan, Srinivasan, Control Groups: Pagecache accounting and control
535 subsystem (v3), http://lwn.net/Articles/235534/
536 8. Singh, Balbir. RSS controller v2 test results (lmbench),
537 http://lkml.org/lkml/2007/5/17/232
538 9. Singh, Balbir. RSS controller v2 AIM9 results
539 http://lkml.org/lkml/2007/5/18/1
540 10. Singh, Balbir. Memory controller v6 test results,
541 http://lkml.org/lkml/2007/8/19/36
542 11. Singh, Balbir. Memory controller introduction (v6),
543 http://lkml.org/lkml/2007/8/17/69
544 12. Corbet, Jonathan, Controlling memory use in cgroups,
545 http://lwn.net/Articles/243795/