Re: [PATCH v5] mm/oom_kill.c: futex: Close a race between do_exit and the oom_reaper

[Nico Pache]

On 3/21/22 02:55, Michal Hocko wrote:
> On Thu 17-03-22 21:36:21, Nico Pache wrote:
>> The pthread struct is allocated on PRIVATE|ANONYMOUS memory [1] which can
>> be targeted by the oom reaper. This mapping is used to store the futex
>> robust list; the kernel does not keep a copy of the robust list and instead
>> references a userspace address to maintain the robustness during a process
>> death. A race can occur between exit_mm and the oom reaper that allows
>> the oom reaper to free the memory of the futex robust list before the
>> exit path has handled the futex death:
>>
>>     CPU1                               CPU2
>> ------------------------------------------------------------------------
>>     page_fault
>>     out_of_memory
>>     do_exit "signal"
>>     wake_oom_reaper
>> 					oom_reaper
>>                         		oom_reap_task_mm (invalidates mm)
>>     exit_mm
>>     exit_mm_release
>>     futex_exit_release
>>     futex_cleanup
>>     exit_robust_list
>>     get_user (EFAULT- can't access memory)
> 
> I still think it is useful to be explicit about the consequences of the
> EFAULT here. Did you want to mention that a failing get_user in this
> path would result in a hang because nobody is woken up when the current
> holder of the lock terminates.

Sounds good! You make a good point-- We had that in all the other versions, but
I forgot to include it in this commit log.
> 
>> While in the oom reaper thread, we must handle the futex cleanup without
>> sleeping. To achieve this, add the boolean `try` to futex_exit_begin().
>> This will control weather or not we use a trylock. Also introduce
>> try_futex_exit_release() which will utilize the trylock version of the
>> futex_cleanup_begin(). Also call kthread_use_mm in this context to assure
>> the get_user call in futex_cleanup() does not return with EFAULT.
> 
> This alone is not sufficient. get_user can sleep in the #PF handler path
> (e.g. by waiting for swap in). Or is there any guarantee that the page
> is never swapped out? If we cannot rule out #PF then this is not a
> viable way to address the problem I am afraid.>
> Please also note that this all is done after mmap_lock has been already
> taken so a page fault could deadlock on the mmap_lock.
>
I don't think we can guarantee that page is not swapped out. Good catch, I was
concerned when I saw the 'might_fault' in get_user, but I wasn't fully sure of
its consequences. I'm still learning the labyrinth that is the MM space, so
thanks for the context!

> The more I am thinking about this the more I am getting convinced that
> we should rather approach this differently and skip over vmas which can
> be holding the list. Have you considered this option?

We've discussed it and it seems very doable, but we haven't attempted to
implement it yet. I'll give it a shot and see what I can come up with!

Thank you for your feedback and reviews :)
-- Nico