1 Memory Resource Controller(Memcg) Implementation Memo.
3 Base Kernel Version: based on 2.6.33-rc7-mm(candidate for 34).
5 Because VM is getting complex (one of reasons is memcg...), memcg's behavior
6 is complex. This is a document for memcg's internal behavior.
7 Please note that implementation details can be changed.
9 (*) Topics on API should be in Documentation/cgroup-v1/memory.txt)
11 0. How to record usage ?
14 page_cgroup ....an object per page.
15 Allocated at boot or memory hotplug. Freed at memory hot removal.
17 swap_cgroup ... an entry per swp_entry.
18 Allocated at swapon(). Freed at swapoff().
20 The page_cgroup has USED bit and double count against a page_cgroup never
21 occurs. swap_cgroup is used only when a charged page is swapped-out.
25 a page/swp_entry may be charged (usage += PAGE_SIZE) at
27 mem_cgroup_try_charge()
30 a page/swp_entry may be uncharged (usage -= PAGE_SIZE) by
33 Called when a page's refcount goes down to 0.
35 mem_cgroup_uncharge_swap()
36 Called when swp_entry's refcnt goes down to 0. A charge against swap
39 3. charge-commit-cancel
40 Memcg pages are charged in two steps:
41 mem_cgroup_try_charge()
42 mem_cgroup_commit_charge() or mem_cgroup_cancel_charge()
44 At try_charge(), there are no flags to say "this page is charged".
45 at this point, usage += PAGE_SIZE.
47 At commit(), the page is associated with the memcg.
49 At cancel(), simply usage -= PAGE_SIZE.
51 Under below explanation, we assume CONFIG_MEM_RES_CTRL_SWAP=y.
54 Anonymous page is newly allocated at
55 - page fault into MAP_ANONYMOUS mapping.
59 At swap-in, the page is taken from swap-cache. There are 2 cases.
61 (a) If the SwapCache is newly allocated and read, it has no charges.
62 (b) If the SwapCache has been mapped by processes, it has been
66 At swap-out, typical state transition is below.
68 (a) add to swap cache. (marked as SwapCache)
69 swp_entry's refcnt += 1.
71 swp_entry's refcnt += # of ptes.
72 (c) write back to swap.
73 (d) delete from swap cache. (remove from SwapCache)
74 swp_entry's refcnt -= 1.
77 Finally, at task exit,
78 (e) zap_pte() is called and swp_entry's refcnt -=1 -> 0.
81 Page Cache is charged at
82 - add_to_page_cache_locked().
84 The logic is very clear. (About migration, see below)
85 Note: __remove_from_page_cache() is called by remove_from_page_cache()
86 and __remove_mapping().
88 6. Shmem(tmpfs) Page Cache
89 The best way to understand shmem's page state transition is to read
91 But brief explanation of the behavior of memcg around shmem will be
92 helpful to understand the logic.
94 Shmem's page (just leaf page, not direct/indirect block) can be on
95 - radix-tree of shmem's inode.
97 - Both on radix-tree and SwapCache. This happens at swap-in
101 - A new page is added to shmem's radix-tree.
102 - A swp page is read. (move a charge from swap_cgroup to page_cgroup)
109 Each memcg has its own private LRU. Now, its handling is under global
110 VM's control (means that it's handled under global zone_lru_lock).
111 Almost all routines around memcg's LRU is called by global LRU's
112 list management functions under zone_lru_lock().
114 A special function is mem_cgroup_isolate_pages(). This scans
115 memcg's private LRU and call __isolate_lru_page() to extract a page
117 (By __isolate_lru_page(), the page is removed from both of global and
123 Tests for racy cases.
125 9.1 Small limit to memcg.
126 When you do test to do racy case, it's good test to set memcg's limit
127 to be very small rather than GB. Many races found in the test under
129 (Memory behavior under GB and Memory behavior under MB shows very
130 different situation.)
133 Historically, memcg's shmem handling was poor and we saw some amount
134 of troubles here. This is because shmem is page-cache but can be
135 SwapCache. Test with shmem/tmpfs is always good test.
138 For NUMA, migration is an another special case. To do easy test, cpuset
139 is useful. Following is a sample script to do migration.
141 mount -t cgroup -o cpuset none /opt/cpuset
144 echo 1 > /opt/cpuset/01/cpuset.cpus
145 echo 0 > /opt/cpuset/01/cpuset.mems
146 echo 1 > /opt/cpuset/01/cpuset.memory_migrate
148 echo 1 > /opt/cpuset/02/cpuset.cpus
149 echo 1 > /opt/cpuset/02/cpuset.mems
150 echo 1 > /opt/cpuset/02/cpuset.memory_migrate
152 In above set, when you moves a task from 01 to 02, page migration to
153 node 0 to node 1 will occur. Following is a script to migrate all
160 /bin/echo $pid >$2/tasks 2>/dev/null
167 G1_TASK=`cat ${G1}/tasks`
168 G2_TASK=`cat ${G2}/tasks`
169 move_task "${G1_TASK}" ${G2} &
172 memory hotplug test is one of good test.
173 to offline memory, do following.
174 # echo offline > /sys/devices/system/memory/memoryXXX/state
175 (XXX is the place of memory)
176 This is an easy way to test page migration, too.
179 When using hierarchy, mkdir/rmdir test should be done.
180 Use tests like the following.
182 echo 1 >/opt/cgroup/01/memory/use_hierarchy
183 mkdir /opt/cgroup/01/child_a
184 mkdir /opt/cgroup/01/child_b
187 add limit to 01/child_b
188 run jobs under child_a and child_b
190 create/delete following groups at random while jobs are running.
191 /opt/cgroup/01/child_a/child_aa
192 /opt/cgroup/01/child_b/child_bb
193 /opt/cgroup/01/child_c
195 running new jobs in new group is also good.
197 9.6 Mount with other subsystems.
198 Mounting with other subsystems is a good test because there is a
199 race and lock dependency with other cgroup subsystems.
202 # mount -t cgroup none /cgroup -o cpuset,memory,cpu,devices
204 and do task move, mkdir, rmdir etc...under this.
207 Besides management of swap is one of complicated parts of memcg,
208 call path of swap-in at swapoff is not same as usual swap-in path..
209 It's worth to be tested explicitly.
211 For example, test like following is good.
213 # mount -t cgroup none /cgroup -o memory
215 # echo 40M > /cgroup/test/memory.limit_in_bytes
216 # echo 0 > /cgroup/test/tasks
217 Run malloc(100M) program under this. You'll see 60M of swaps.
219 # move all tasks in /cgroup/test to /cgroup
224 Of course, tmpfs v.s. swapoff test should be tested, too.
227 Out-of-memory caused by memcg's limit will kill tasks under
228 the memcg. When hierarchy is used, a task under hierarchy
229 will be killed by the kernel.
230 In this case, panic_on_oom shouldn't be invoked and tasks
231 in other groups shouldn't be killed.
233 It's not difficult to cause OOM under memcg as following.
234 Case A) when you can swapoff
236 #echo 50M > /memory.limit_in_bytes
239 Case B) when you use mem+swap limitation.
240 #echo 50M > memory.limit_in_bytes
241 #echo 50M > memory.memsw.limit_in_bytes
244 9.9 Move charges at task migration
245 Charges associated with a task can be moved along with task migration.
249 #echo $$ >/cgroup/A/tasks
250 run some programs which uses some amount of memory in /cgroup/A.
254 #echo 1 >/cgroup/B/memory.move_charge_at_immigrate
255 #echo "pid of the program running in group A" >/cgroup/B/tasks
257 You can see charges have been moved by reading *.usage_in_bytes or
258 memory.stat of both A and B.
259 See 8.2 of Documentation/cgroup-v1/memory.txt to see what value should be
260 written to move_charge_at_immigrate.
262 9.10 Memory thresholds
263 Memory controller implements memory thresholds using cgroups notification
264 API. You can use tools/cgroup/cgroup_event_listener.c to test it.
266 (Shell-A) Create cgroup and run event listener
268 # ./cgroup_event_listener /cgroup/A/memory.usage_in_bytes 5M
270 (Shell-B) Add task to cgroup and try to allocate and free memory
271 # echo $$ >/cgroup/A/tasks
272 # a="$(dd if=/dev/zero bs=1M count=10)"
275 You will see message from cgroup_event_listener every time you cross
278 Use /cgroup/A/memory.memsw.usage_in_bytes to test memsw thresholds.
280 It's good idea to test root cgroup as well.