Re: [PATCH V2 7/9] vhost: do not use RCU to synchronize MMU notifier with worker

[Jason Wang]

On 2019/8/5 äå2:28, Michael S. Tsirkin wrote:
> On Mon, Aug 05, 2019 at 12:33:45PM +0800, Jason Wang wrote:
> > On 2019/8/2 äå10:03, Michael S. Tsirkin wrote:
> > > On Fri, Aug 02, 2019 at 05:40:07PM +0800, Jason Wang wrote:
> > > > Btw, I come up another idea, that is to disable preemption when vhost thread
> > > > need to access the memory. Then register preempt notifier and if vhost
> > > > thread is preempted, we're sure no one will access the memory and can do the
> > > > cleanup.
> > > Great, more notifiers :(
> > >
> > > Maybe can live with
> > > 1- disable preemption while using the cached pointer
> > > 2- teach vhost to recover from memory access failures,
> > >      by switching to regular from/to user path
> >
> > I don't get this, I believe we want to recover from regular from/to user
> > path, isn't it?
> That (disable copy to/from user completely) would be a nice to have
> since it would reduce the attack surface of the driver, but e.g. your
> code already doesn't do that.

Yes since it requires a lot of changes.


> > > So if you want to try that, fine since it's a step in
> > > the right direction.
> > >
> > > But I think fundamentally it's not what we want to do long term.
> >
> > Yes.
> >
> >
> > > It's always been a fundamental problem with this patch series that only
> > > metadata is accessed through a direct pointer.
> > >
> > > The difference in ways you handle metadata and data is what is
> > > now coming and messing everything up.
> >
> > I do propose soemthing like this in the past:
> > https://www.spinics.net/lists/linux-virtualization/msg36824.html. But looks
> > like you have some concern about its locality.
> Right and it doesn't go away. You'll need to come up
> with a test that messes it up and triggers a worst-case
> scenario, so we can measure how bad is that worst-case.




> > But the problem still there, GUP can do page fault, so still need to
> > synchronize it with MMU notifiers.
> I think the idea was, if GUP would need a pagefault, don't
> do a GUP and do to/from user instead.


But this still need to be synchronized with MMU notifiers (or using 
dedicated work for GUP).


>   Hopefully that
> will fault the page in and the next access will go through.
>
> > The solution might be something like
> > moving GUP to a dedicated kind of vhost work.
> Right, generally GUP.
>
> > > So if continuing the direct map approach,
> > > what is needed is a cache of mapped VM memory, then on a cache miss
> > > we'd queue work along the lines of 1-2 above.
> > >
> > > That's one direction to take. Another one is to give up on that and
> > > write our own version of uaccess macros.  Add a "high security" flag to
> > > the vhost module and if not active use these for userspace memory
> > > access.
> >
> > Or using SET_BACKEND_FEATURES?
> No, I don't think it's considered best practice to allow unpriveledged
> userspace control over whether kernel enables security features.


Get this.


> > But do you mean permanent GUP as I did in
> > original RFC https://lkml.org/lkml/2018/12/13/218?
> >
> > Thanks
> Permanent GUP breaks THP and NUMA.


Yes.

Thanks


> > >