Oleg Nesterov <o...@redhat.com> writes:
> On 05/03, Eric W. Biederman wrote:
>>
>> Oleg Nesterov <o...@redhat.com> writes:
>>
>> > On 05/02, Eric W. Biederman wrote:
>> >>
>> >> +static void mem_cgroup_fork(struct task_struct *tsk)
>> >> +{
>> >> + struct cgroup_subsys_state *css;
>> >> +
>> >> + rcu_read_lock();
>> >> + css = task_css(tsk, memory_cgrp_id);
>> >> + if (css && css_tryget(css))
>> >> + task_update_memcg(tsk, mem_cgroup_from_css(css));
>> >> + rcu_read_unlock();
>> >> +}
>> >
>> > Why do we need it?
>> >
>> > The child's mm->memcg was already initialized by mm_init_memcg() and we
>> > can't
>> > race with migrate until cgroup_threadgroup_change_end() ?
>>
>> I admit I missed the cgroup_threadgroup_change_begin
>> cgroup_threadgroup_change_end pair in fs fork. In this case it doesn't
>> matter because mm_init_memcg is called from:
>>
>> copy_mm
>> dup_mm
>> mm_init
>>
>> And copy_mm is called before we call cgroup_threadgroup_change_begin.
>> So the race remains.
>
> Ah yes, you are right.
>
>> We could move move cgroup_threadgroup_change_begin earlier, to remove
>> the need for mem_cgroup_fork. But I have not analyzed that.
>
> No, cgroup_threadgroup_change_begin() was called early and this was wrong, see
> 568ac888215c7fb2fabe8ea739b00ec3c1f5d440. Actually there were more problems,
> say
> copy_net() could deadlock because cleanup_net() does do_wait() with net_mutex
> held.
>
>
> OK, what about exec() ? mm_init_memcg() initializes bprm->mm->memcg early in
> bprm_mm_init(). What if the execing task migrates before exec_mmap() ?
We need the the cgroup when the mm is initialized. That way we have the
cgroup information when initializing the mm.
I don't know if a lock preventing changing the cgroup in exec or just
a little bit of code in exec_mmap to ensure mm->memcg is properly set
is the better approach. I have not analyzed that code path.
This does look like a very good place for an incremental patch to close
that race.
Eric
