Re: [External] Re: [RFC PATCH 0/9] Support for Virtio-fs daemon crash reconnection

[Stefan Hajnoczi]

On Tue, Mar 23, 2021 at 01:54:46PM +0100, Christian Schoenebeck wrote:
> On Montag, 22. März 2021 11:54:56 CET Stefan Hajnoczi wrote:
> > > > Thanks, Christian. I am still trying to figure out the details of the
> > > > ROP
> > > > attacks.
> > > > 
> > > > However, QEMU's vhost-user reconnection is based on chardev socket
> > > > reconnection. The socket reconnection can be enabled by the "--chardev
> > > > socket,...,reconnect=N" in QEMU command options, in which N means QEMU
> > > > will
> > > > try to connect the disconnected socket every N seconds. We can increase
> > > > N
> > > > to increase the reconnect delay. If we want to change the reconnect
> > > > delay
> > > > dynamically, I think we should change the chardev socket reconnection
> > > > code.
> > > > It is a more generic mechanism than vhost-user-fs and vhost-user
> > > > backend.
> > > > 
> > > > By the way, I also considered the socket reconnection delay time in the
> > > > performance aspect. As the reconnection delay increase, if an
> > > > application
> > > > in the guest is doing I/Os, it will suffer larger tail latency. And for
> > > > now, the smallest delay is 1 second, which is rather large for
> > > > high-performance virtual I/O devices today. I think maybe a more
> > > > performant
> > > > and safer reconnect delay adjustment mechanism should be considered in
> > > > the
> > > > future. What are your thoughts?
> > > 
> > > So with N=1 an attacker could e.g. bypass a 16-bit PAC by brute-force in
> > > ~18 hours (e.g. on Arm if PAC + MTE was enabled). With 24-bit PAC (no
> > > MTE) it would be ~194 days. Independent of what architecture and defend
> > > mechanism is used, there is always the possibility though that some kind
> > > of side channel attack exists that might require a much lower amount of
> > > attempts. So in an untrusted environment I would personally limit the
> > > amount of automatic reconnects and rather accept a down time for further
> > > investigation if a suspicious high amount of crashes happened.
> > > 
> > > And yes, if a dynamic delay scheme was deployed in future then starting
> > > with a value smaller than 1 second would make sense.
> > 
> > If we're talking about repeatedly crashing the process to find out its
> > memory map, shouldn't each process have a different randomized memory
> > layout?
> > 
> > Stefan
> 
> Yes, ASLR is enabled on Linux and other OSes by default for more than 10 
> years. But ASLR does not prevent ROP attacks which are commonly using 
> relative 
> offsets, tweaking the stack, indirect jumps, as well as heap spraying. Plus 
> side channels exist to gain access to direct addresses.
> 
> The situation might improve significantly when shadow stacks (e.g. Intel CET) 
> become widely used in future. But in the meantime I would be cautious if 
> something is crashing too often in a certain time frame.

It's a good point for performance as well. A broken service should not
hog CPU by constantly restarting itself. If it's broken badly it may
never come back up and should be throttled.

Stefan

signature.asc
Description: PGP signature