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.
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.
Signed-off-by: Hugh Dickins <[email protected]>
---
mm/ksm.c | 6 +++---
1 file changed, 3 insertions(+), 3 deletions(-)
--- mmotm/mm/ksm.c 2009-11-25 09:28:50.000000000 +0000
+++ linux/mm/ksm.c 2009-11-27 13:49:50.000000000 +0000
@@ -1565,9 +1565,6 @@ again:
struct anon_vma *anon_vma = rmap_item->anon_vma;
struct vm_area_struct *vma;
- if (memcg && !mm_match_cgroup(rmap_item->mm, memcg))
- continue;
-
spin_lock(&anon_vma->lock);
list_for_each_entry(vma, &anon_vma->head, anon_vma_node) {
if (rmap_item->address < vma->vm_start ||
@@ -1582,6 +1579,9 @@ again:
if ((rmap_item->mm == vma->vm_mm) == search_new_forks)
continue;
+ if (memcg && !mm_match_cgroup(vma->vm_mm, memcg))
+ continue;
+
referenced += page_referenced_one(page, vma,
rmap_item->address, &mapcount, vm_flags);
if (!search_new_forks || !mapcount)
* Hugh Dickins <[email protected]> [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