summaryrefslogtreecommitdiff
path: root/Documentation/cgroups/memcg_test.txt
diff options
context:
space:
mode:
authorLi Zefan <lizf@cn.fujitsu.com>2009-01-15 13:50:59 -0800
committerLinus Torvalds <torvalds@linux-foundation.org>2009-01-15 16:39:37 -0800
commit45ce80fb6b6f9594d1396d44dd7e7c02d596fef8 (patch)
tree2409270f7073c08329ac01c82df0509a264af48c /Documentation/cgroups/memcg_test.txt
parent23964d2d02984d44aeb2d84d7ffb3359e728df43 (diff)
downloadlwn-45ce80fb6b6f9594d1396d44dd7e7c02d596fef8.tar.gz
lwn-45ce80fb6b6f9594d1396d44dd7e7c02d596fef8.zip
cgroups: consolidate cgroup documents
Move Documentation/cpusets.txt and Documentation/controllers/* to Documentation/cgroups/ Signed-off-by: Li Zefan <lizf@cn.fujitsu.com> Acked-by: KAMEZAWA Hiroyuki <kamezawa.hiroyu@jp.fujitsu.com> Acked-by: Balbir Singh <balbir@linux.vnet.ibm.com> Acked-by: Paul Menage <menage@google.com> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
Diffstat (limited to 'Documentation/cgroups/memcg_test.txt')
-rw-r--r--Documentation/cgroups/memcg_test.txt342
1 files changed, 342 insertions, 0 deletions
diff --git a/Documentation/cgroups/memcg_test.txt b/Documentation/cgroups/memcg_test.txt
new file mode 100644
index 000000000000..19533f93b7a2
--- /dev/null
+++ b/Documentation/cgroups/memcg_test.txt
@@ -0,0 +1,342 @@
+Memory Resource Controller(Memcg) Implementation Memo.
+Last Updated: 2008/12/15
+Base Kernel Version: based on 2.6.28-rc8-mm.
+
+Because VM is getting complex (one of reasons is memcg...), memcg's behavior
+is complex. This is a document for memcg's internal behavior.
+Please note that implementation details can be changed.
+
+(*) Topics on API should be in Documentation/cgroups/memory.txt)
+
+0. How to record usage ?
+ 2 objects are used.
+
+ page_cgroup ....an object per page.
+ Allocated at boot or memory hotplug. Freed at memory hot removal.
+
+ swap_cgroup ... an entry per swp_entry.
+ Allocated at swapon(). Freed at swapoff().
+
+ The page_cgroup has USED bit and double count against a page_cgroup never
+ occurs. swap_cgroup is used only when a charged page is swapped-out.
+
+1. Charge
+
+ a page/swp_entry may be charged (usage += PAGE_SIZE) at
+
+ mem_cgroup_newpage_charge()
+ Called at new page fault and Copy-On-Write.
+
+ mem_cgroup_try_charge_swapin()
+ Called at do_swap_page() (page fault on swap entry) and swapoff.
+ Followed by charge-commit-cancel protocol. (With swap accounting)
+ At commit, a charge recorded in swap_cgroup is removed.
+
+ mem_cgroup_cache_charge()
+ Called at add_to_page_cache()
+
+ mem_cgroup_cache_charge_swapin()
+ Called at shmem's swapin.
+
+ mem_cgroup_prepare_migration()
+ Called before migration. "extra" charge is done and followed by
+ charge-commit-cancel protocol.
+ At commit, charge against oldpage or newpage will be committed.
+
+2. Uncharge
+ a page/swp_entry may be uncharged (usage -= PAGE_SIZE) by
+
+ mem_cgroup_uncharge_page()
+ Called when an anonymous page is fully unmapped. I.e., mapcount goes
+ to 0. If the page is SwapCache, uncharge is delayed until
+ mem_cgroup_uncharge_swapcache().
+
+ mem_cgroup_uncharge_cache_page()
+ Called when a page-cache is deleted from radix-tree. If the page is
+ SwapCache, uncharge is delayed until mem_cgroup_uncharge_swapcache().
+
+ mem_cgroup_uncharge_swapcache()
+ Called when SwapCache is removed from radix-tree. The charge itself
+ is moved to swap_cgroup. (If mem+swap controller is disabled, no
+ charge to swap occurs.)
+
+ mem_cgroup_uncharge_swap()
+ Called when swp_entry's refcnt goes down to 0. A charge against swap
+ disappears.
+
+ mem_cgroup_end_migration(old, new)
+ At success of migration old is uncharged (if necessary), a charge
+ to new page is committed. At failure, charge to old page is committed.
+
+3. charge-commit-cancel
+ In some case, we can't know this "charge" is valid or not at charging
+ (because of races).
+ To handle such case, there are charge-commit-cancel functions.
+ mem_cgroup_try_charge_XXX
+ mem_cgroup_commit_charge_XXX
+ mem_cgroup_cancel_charge_XXX
+ these are used in swap-in and migration.
+
+ At try_charge(), there are no flags to say "this page is charged".
+ at this point, usage += PAGE_SIZE.
+
+ At commit(), the function checks the page should be charged or not
+ and set flags or avoid charging.(usage -= PAGE_SIZE)
+
+ At cancel(), simply usage -= PAGE_SIZE.
+
+Under below explanation, we assume CONFIG_MEM_RES_CTRL_SWAP=y.
+
+4. Anonymous
+ Anonymous page is newly allocated at
+ - page fault into MAP_ANONYMOUS mapping.
+ - Copy-On-Write.
+ It is charged right after it's allocated before doing any page table
+ related operations. Of course, it's uncharged when another page is used
+ for the fault address.
+
+ At freeing anonymous page (by exit() or munmap()), zap_pte() is called
+ and pages for ptes are freed one by one.(see mm/memory.c). Uncharges
+ are done at page_remove_rmap() when page_mapcount() goes down to 0.
+
+ Another page freeing is by page-reclaim (vmscan.c) and anonymous
+ pages are swapped out. In this case, the page is marked as
+ PageSwapCache(). uncharge() routine doesn't uncharge the page marked
+ as SwapCache(). It's delayed until __delete_from_swap_cache().
+
+ 4.1 Swap-in.
+ At swap-in, the page is taken from swap-cache. There are 2 cases.
+
+ (a) If the SwapCache is newly allocated and read, it has no charges.
+ (b) If the SwapCache has been mapped by processes, it has been
+ charged already.
+
+ This swap-in is one of the most complicated work. In do_swap_page(),
+ following events occur when pte is unchanged.
+
+ (1) the page (SwapCache) is looked up.
+ (2) lock_page()
+ (3) try_charge_swapin()
+ (4) reuse_swap_page() (may call delete_swap_cache())
+ (5) commit_charge_swapin()
+ (6) swap_free().
+
+ Considering following situation for example.
+
+ (A) The page has not been charged before (2) and reuse_swap_page()
+ doesn't call delete_from_swap_cache().
+ (B) The page has not been charged before (2) and reuse_swap_page()
+ calls delete_from_swap_cache().
+ (C) The page has been charged before (2) and reuse_swap_page() doesn't
+ call delete_from_swap_cache().
+ (D) The page has been charged before (2) and reuse_swap_page() calls
+ delete_from_swap_cache().
+
+ memory.usage/memsw.usage changes to this page/swp_entry will be
+ Case (A) (B) (C) (D)
+ Event
+ Before (2) 0/ 1 0/ 1 1/ 1 1/ 1
+ ===========================================
+ (3) +1/+1 +1/+1 +1/+1 +1/+1
+ (4) - 0/ 0 - -1/ 0
+ (5) 0/-1 0/ 0 -1/-1 0/ 0
+ (6) - 0/-1 - 0/-1
+ ===========================================
+ Result 1/ 1 1/ 1 1/ 1 1/ 1
+
+ In any cases, charges to this page should be 1/ 1.
+
+ 4.2 Swap-out.
+ At swap-out, typical state transition is below.
+
+ (a) add to swap cache. (marked as SwapCache)
+ swp_entry's refcnt += 1.
+ (b) fully unmapped.
+ swp_entry's refcnt += # of ptes.
+ (c) write back to swap.
+ (d) delete from swap cache. (remove from SwapCache)
+ swp_entry's refcnt -= 1.
+
+
+ At (b), the page is marked as SwapCache and not uncharged.
+ At (d), the page is removed from SwapCache and a charge in page_cgroup
+ is moved to swap_cgroup.
+
+ Finally, at task exit,
+ (e) zap_pte() is called and swp_entry's refcnt -=1 -> 0.
+ Here, a charge in swap_cgroup disappears.
+
+5. Page Cache
+ Page Cache is charged at
+ - add_to_page_cache_locked().
+
+ uncharged at
+ - __remove_from_page_cache().
+
+ The logic is very clear. (About migration, see below)
+ Note: __remove_from_page_cache() is called by remove_from_page_cache()
+ and __remove_mapping().
+
+6. Shmem(tmpfs) Page Cache
+ Memcg's charge/uncharge have special handlers of shmem. The best way
+ to understand shmem's page state transition is to read mm/shmem.c.
+ But brief explanation of the behavior of memcg around shmem will be
+ helpful to understand the logic.
+
+ Shmem's page (just leaf page, not direct/indirect block) can be on
+ - radix-tree of shmem's inode.
+ - SwapCache.
+ - Both on radix-tree and SwapCache. This happens at swap-in
+ and swap-out,
+
+ It's charged when...
+ - A new page is added to shmem's radix-tree.
+ - A swp page is read. (move a charge from swap_cgroup to page_cgroup)
+ It's uncharged when
+ - A page is removed from radix-tree and not SwapCache.
+ - When SwapCache is removed, a charge is moved to swap_cgroup.
+ - When swp_entry's refcnt goes down to 0, a charge in swap_cgroup
+ disappears.
+
+7. Page Migration
+ One of the most complicated functions is page-migration-handler.
+ Memcg has 2 routines. Assume that we are migrating a page's contents
+ from OLDPAGE to NEWPAGE.
+
+ Usual migration logic is..
+ (a) remove the page from LRU.
+ (b) allocate NEWPAGE (migration target)
+ (c) lock by lock_page().
+ (d) unmap all mappings.
+ (e-1) If necessary, replace entry in radix-tree.
+ (e-2) move contents of a page.
+ (f) map all mappings again.
+ (g) pushback the page to LRU.
+ (-) OLDPAGE will be freed.
+
+ Before (g), memcg should complete all necessary charge/uncharge to
+ NEWPAGE/OLDPAGE.
+
+ The point is....
+ - If OLDPAGE is anonymous, all charges will be dropped at (d) because
+ try_to_unmap() drops all mapcount and the page will not be
+ SwapCache.
+
+ - If OLDPAGE is SwapCache, charges will be kept at (g) because
+ __delete_from_swap_cache() isn't called at (e-1)
+
+ - If OLDPAGE is page-cache, charges will be kept at (g) because
+ remove_from_swap_cache() isn't called at (e-1)
+
+ memcg provides following hooks.
+
+ - mem_cgroup_prepare_migration(OLDPAGE)
+ Called after (b) to account a charge (usage += PAGE_SIZE) against
+ memcg which OLDPAGE belongs to.
+
+ - mem_cgroup_end_migration(OLDPAGE, NEWPAGE)
+ Called after (f) before (g).
+ If OLDPAGE is used, commit OLDPAGE again. If OLDPAGE is already
+ charged, a charge by prepare_migration() is automatically canceled.
+ If NEWPAGE is used, commit NEWPAGE and uncharge OLDPAGE.
+
+ But zap_pte() (by exit or munmap) can be called while migration,
+ we have to check if OLDPAGE/NEWPAGE is a valid page after commit().
+
+8. LRU
+ Each memcg has its own private LRU. Now, it's handling is under global
+ VM's control (means that it's handled under global zone->lru_lock).
+ Almost all routines around memcg's LRU is called by global LRU's
+ list management functions under zone->lru_lock().
+
+ A special function is mem_cgroup_isolate_pages(). This scans
+ memcg's private LRU and call __isolate_lru_page() to extract a page
+ from LRU.
+ (By __isolate_lru_page(), the page is removed from both of global and
+ private LRU.)
+
+
+9. Typical Tests.
+
+ Tests for racy cases.
+
+ 9.1 Small limit to memcg.
+ When you do test to do racy case, it's good test to set memcg's limit
+ to be very small rather than GB. Many races found in the test under
+ xKB or xxMB limits.
+ (Memory behavior under GB and Memory behavior under MB shows very
+ different situation.)
+
+ 9.2 Shmem
+ Historically, memcg's shmem handling was poor and we saw some amount
+ of troubles here. This is because shmem is page-cache but can be
+ SwapCache. Test with shmem/tmpfs is always good test.
+
+ 9.3 Migration
+ For NUMA, migration is an another special case. To do easy test, cpuset
+ is useful. Following is a sample script to do migration.
+
+ mount -t cgroup -o cpuset none /opt/cpuset
+
+ mkdir /opt/cpuset/01
+ echo 1 > /opt/cpuset/01/cpuset.cpus
+ echo 0 > /opt/cpuset/01/cpuset.mems
+ echo 1 > /opt/cpuset/01/cpuset.memory_migrate
+ mkdir /opt/cpuset/02
+ echo 1 > /opt/cpuset/02/cpuset.cpus
+ echo 1 > /opt/cpuset/02/cpuset.mems
+ echo 1 > /opt/cpuset/02/cpuset.memory_migrate
+
+ In above set, when you moves a task from 01 to 02, page migration to
+ node 0 to node 1 will occur. Following is a script to migrate all
+ under cpuset.
+ --
+ move_task()
+ {
+ for pid in $1
+ do
+ /bin/echo $pid >$2/tasks 2>/dev/null
+ echo -n $pid
+ echo -n " "
+ done
+ echo END
+ }
+
+ G1_TASK=`cat ${G1}/tasks`
+ G2_TASK=`cat ${G2}/tasks`
+ move_task "${G1_TASK}" ${G2} &
+ --
+ 9.4 Memory hotplug.
+ memory hotplug test is one of good test.
+ to offline memory, do following.
+ # echo offline > /sys/devices/system/memory/memoryXXX/state
+ (XXX is the place of memory)
+ This is an easy way to test page migration, too.
+
+ 9.5 mkdir/rmdir
+ When using hierarchy, mkdir/rmdir test should be done.
+ Use tests like the following.
+
+ echo 1 >/opt/cgroup/01/memory/use_hierarchy
+ mkdir /opt/cgroup/01/child_a
+ mkdir /opt/cgroup/01/child_b
+
+ set limit to 01.
+ add limit to 01/child_b
+ run jobs under child_a and child_b
+
+ create/delete following groups at random while jobs are running.
+ /opt/cgroup/01/child_a/child_aa
+ /opt/cgroup/01/child_b/child_bb
+ /opt/cgroup/01/child_c
+
+ running new jobs in new group is also good.
+
+ 9.6 Mount with other subsystems.
+ Mounting with other subsystems is a good test because there is a
+ race and lock dependency with other cgroup subsystems.
+
+ example)
+ # mount -t cgroup none /cgroup -t cpuset,memory,cpu,devices
+
+ and do task move, mkdir, rmdir etc...under this.