Messages in this thread |  | | | Date | Mon, 30 Nov 2009 09:46:17 +0530 | | From | Balbir Singh <> | | Subject | Re: [PATCH] ksm: hold anon_vma in rmap_item fix |
| |
* Hugh Dickins <hugh.dickins@tiscali.co.uk> [2009-11-29 15:50:32]:
> KSM mem_cgroup testing oopsed on NULL pointer in mem_cgroup_from_task(), > called from the mm_match_cgroup() in page_referenced_ksm(). > > Right, it is inappropriate to use mm_match_cgroup() on rmap_item->mm > there: that mm could be waiting for ksmd's final mmdrop(), with its > mm->owner task long gone. > > Move the mm_match_cgroup() test down into the anon_vma loop, which is > where it now should be to match page_referenced_anon(). The anon_vma > guarantees its vmas are valid, which guarantee their mms are valid. > > However... although this moves the oops from easy-to-reproduce to > never-seen, I think we shall want to do more later: so far as I can > see, with or without KSM, the use of mm->owner from page_referenced() > is unsafe. No problem when NULL, but it may have been left pointing > to a task_struct freed by now, with nonsense in mm->owner->cgroups. >
Ideally we should not be left pointing to a stale task struct, unless our assumption about mm_users is incorrect (discussed below).
> But let's put this patch in while we discuss that separately: perhaps > mm_need_new_owner() should not short-circuit when mm_users <= 1, or > perhaps it should then set mm->owner to NULL, or perhaps we abandon > mm->owner as more trouble than it's worth, or... perhaps I'm wrong. >
We short circuit, since the task is exiting and mm_users <= 1 and we are shorting going to do a mmput(). I suspect what you are seeing is mm_count >= 1 and mm_users == 0. With users == 0, we should set owner to NULL
We could look for the above condition in mmput() and clear the owner when users become 0.
-- Balbir
|  |