On Mon 18-10-21 13:05:35, Vasily Averin wrote:

On 18.10.2021 12:04, Michal Hocko wrote:

quoted

On Mon 18-10-21 11:13:52, Vasily Averin wrote:
[...]

quoted

How could this happen?

User-space task inside the memcg-limited container generated a page fault,
its handler do_user_addr_fault() called handle_mm_fault which could not
allocate the page due to exceeding the memcg limit and returned VM_FAULT_OOM.
Then do_user_addr_fault() called pagefault_out_of_memory() which executed
out_of_memory() without set of memcg.

Partially this problem depends on one of my recent patches, disabled unlimited
memory allocation for dying tasks. However I think the problem can happen
on non-killed tasks too, for example because of kmem limit.

Could you be more specific on how this can happen without your patch? I
have to say I haven't realized this side effect when discussing it.

We can reach obj_cgroup_charge_pages() for example via

do_user_addr_fault
handle_mm_fault
__handle_mm_fault
p4d_alloc
__p4d_alloc
p4d_alloc_one
get_zeroed_page
__get_free_pages
alloc_pages
__alloc_pages
__memcg_kmem_charge_page
obj_cgroup_charge_pages

Here we call try_charge_memcg() that return success and approve the allocation,
however then we hit into kmem limit and fail the allocation.

Just to make sure I understand this would be for the v1 kmem explicit
limit, correct?

If required I can try to search how try_charge_memcg() can reject page allocation 
of non-dying task too.

Yes.

quoted

I will be honest that I am not really happy about pagefault_out_of_memory.
I have tried to remove it in the past. Without much success back then,
unfortunately[1]. 
Maybe we should get rid of it finally. The OOM is always triggered from
inside the allocator where we have much more infromation about the
allocation context. A first step would be to skip pagefault_out_of_memory
for killed or exiting processes.

I like this idea, however it may be not enough, at least in scenario described above.

I original patch has removed the oom killer completely.

-- 
Michal Hocko
SUSE Labs