Re: [PATCH V4] binder: ipc namespace support for android binder

[Todd Kjos]

On Tue, Nov 13, 2018 at 12:12 AM chouryzhou(åå) <chouryzhou@xxxxxxxxxxx> wrote:
>
> > I have not received an answer to my questions in the last version of this patch
> > set. Also it would be good if I could be Cc'ed by default. I can't hunt down all
> > patches.
> > I do not know of any kernel entity, specifically devices, that change namespaces
> > on open().
> > This seems like an invitation for all kinds of security bugs.
> > A device node belongs to one namespace only which is attached to the
> > underlying kobject. Opening the device should never change that.
> > Please look at how mqueue or shm are doing this. They don't change
> > namespaces on open either.
> > I have to say that is one of the main reasons why I disagree with that design.
> >
> >
>
> If we must return the same context when every open in proc, we can only isolate
> binder with mnt namespace instead of ipc namespace, what do you think, Todd?

I don't have strong feelings on this -- it seems like a bizarre
use-case to send the fd through a backchannel as christian describes,
but I do agree it is strange behavior (though it seems safe to me
since it prevents communication between unrelated entities). I don't
know how mqueue and shm work, its worth a look since this patch is
modelling their behavior...

We'll talk about it here at LPC and then on this thread.