Messages in this thread | | | Date | Sat, 05 Jul 2008 12:19:11 +0530 | From | Balbir Singh <> | Subject | Re: [PATCH] memcg: handle shmem's swap cache (Was 2.6.26-rc8-mm1 |
| |
KAMEZAWA Hiroyuki wrote: > On Sat, 05 Jul 2008 11:11:10 +0530 > Balbir Singh <balbir@linux.vnet.ibm.com> wrote: > >> KAMEZAWA Hiroyuki wrote: >>> My swapcache accounting under memcg patch failed to catch tmpfs(shmem)'s one. >>> Can I test this under -mm tree ? >>> (If -mm is busy, I'm not in hurry.) >>> This patch works well in my box. >>> = >>> SwapCache handling fix. >>> >>> shmem's swapcache behavior is a little different from anonymous's one and >>> memcg failed to handle it. This patch tries to fix it. >>> >>> After this: >>> >>> Any page marked as SwapCache is not uncharged. (delelte_from_swap_cache() >>> delete the SwapCache flag.) >>> >>> To check a shmem-page-cache is alive or not we use >>> page->mapping && !PageAnon(page) instead of >>> pc->flags & PAGE_CGROUP_FLAG_CACHE. >>> >>> Signed-off-by: KAMEZAWA Hiroyuki <kamezawa.hiroyu@jp.fujitsu.com> >> Though I am not opposed to this, I do sit up and think if keeping the reference >> count around could avoid this complexity and from my point, the maintenance >> overhead of this logic/code (I fear there might be more special cases :( ) > > yes, to me. but we have to fix.. > > But I don't like old code's refcnt handling which does > - increment > - does this increment was really neccesary ? > No? ok, decrement it again. > > This was much more complex to me than current code. >
That can be redone -- the moment a page is used by a path, refcnt (increment) it. Undo the same when the page is no longer in use.
I expect
rmap path to increment/decrement it on mapping radix-tree (cache's) to do the same
Using a kref we should be able to get this logic right - no?
> And old ones will needs the check at treating swap-cache. (it couldn't but if we want) > >> The trade-off is complexity versus the overhead of reference counting. >> > refcnt was also very complex ;)
I think that is easier to simply, instead of adding the complex checks we have right now. refcnt is easier to prove as working correct than the checks.
-- Warm Regards, Balbir Singh Linux Technology Center IBM, ISTL
| |