Re: [PATCH v8] oom_kill.c: futex: Don't OOM reap the VMA containing the robust_list_head

[Nico Pache]

On 4/12/22 12:20, Thomas Gleixner wrote:
> On Mon, Apr 11 2022 at 19:51, Nico Pache wrote:
>> On 4/8/22 09:54, Thomas Gleixner wrote:
>>> The below reproduces the problem nicely, i.e. the lock() in the parent
>>> times out. So why would the OOM killer fail to cause the same problem
>>> when it reaps the private anon mapping where the private futex sits?
>>>
>>> If you revert the lock order in the child the robust muck works.
>>
>> Thanks for the reproducer Thomas :)
>>
>> I think I need to re-up my knowledge around COW and how it effects
>> that stack. There are increased oddities when you add the pthread
>> library that I cant fully wrap my head around at the moment.
> 
> The pthread library functions are just conveniance so I did not have to
> hand code the futex and robust list handling.
> 
>> My confusion lies in how the parent/child share a robust list here, but they
>> obviously do. In my mind the mut_s would be different in the child/parent after
>> the fork and pthread_mutex_init (and friends) are done in the child.
> 
> They don't share a robust list, each thread has it's own.
> 
> The shared mutex mut_s is initialized in the parent before fork and it's
> the same address in the child and it's not COWed because the mapping is
> MAP_SHARED.
> 
> The child allocates private memory and initializes the private mutex in
> that private mapping.
> 
> So now child does:
> 
>    pthread_mutex_lock(mut_s);
> 
> That's the mutex in the memory shared with the parent. After that the
> childs robusts list head points to mut_s::robust_list.
> 
> Now it does:
> 
>    pthread_mutex_lock(mut_p);
> 
> after that the childs robust list head points to mut_p::robust_list and
> mut_p::robust_list points to mut_s::robust_list.
> 
> So now the child unmaps the private memory and exists.
> 
> The kernel tries to walk the robust list pointer and faults when trying
> to access mut_p. End of walk and mut_s stays locked.
> 
> So now think about the OOM case. The killed process has a shared mapping
> with some other unrelated process (file, shmem) where mut_p sits.
> 
> It gets killed after:
> 		pthread_mutex_lock(mut_s);
> 		pthread_mutex_lock(mut_p);
> 
> So the OOM reaper rips the VMA which contains mut_p and therefore breaks
> the chain which is necessary to reach mut_p.
> 
> See?
Yes, thank you for the detailed explanation, the missing piece just clicked in
my head :)

Cheers,
-- Nico
> 
> Thanks,
> 
>         tglx
> 
> 
>