Re: [RFC PATCH 1/2] Fix: sched/membarrier: p->mm->membarrier_state racy load

[Mathieu Desnoyers <mathieu.desnoyers@efficios.com>]

----- On Sep 4, 2019, at 7:11 AM, Oleg Nesterov oleg@redhat.com wrote:

> with or without these changes...
> 
> Why do membarrier_register_*_expedited() check get_nr_threads() == 1?
> This makes no sense to me, atomic_read(mm_users) == 1 should be enough.

Indeed, if every thread within a process hold a mm_users refcount, then
the get_nr_threads() == 1 check becomes redundant.

AFAIR, this check started out as "get_nr_threads() == 1", and then I changed
the code to also cover the multi-process CLONE_VM use-case by adding the
additional check.

> And I am not sure I understand membarrier_mm_sync_core_before_usermode().
> OK, membarrier_private_expedited() can race with user -> kernel -> user
> transition, but we do not care unless both user's above have the same mm?
> Shouldn't membarrier_mm_sync_core_before_usermode() do
> 
>	if (current->mm != mm)
>		return;
> 
> at the start to make it more clear and avoid sync_core_before_usermode()
> if possible?

Indeed, if we have taskA -> kernel -> taskB, it implies that we go through
switch_mm() when scheduling taskB, which provides the required core serializing
guarantees.

Moreover, if we look closely at the call to membarrier_mm_sync_core_before_usermode(),
the mm it receives as parameter is the rq->prev_mm. So using the prev_mm membarrier
state to decide whether we need to issue a sync_core before returning to a
different next mm is not really relevant unless the next mm == rq->prev_mm.

Nothing there seem to be actively buggy, but those are indeed nice cleanups.

Thanks,

Mathieu

-- 
Mathieu Desnoyers
EfficiOS Inc.
http://www.efficios.com