On Tue, Sep 9, 2025 at 7:37 AM Jens Axboe <axboe@xxxxxxxxx> wrote:
>
> On 9/9/25 8:35 AM, Eric Dumazet wrote:
> > On Tue, Sep 9, 2025 at 7:04 AM Eric Dumazet <edumazet@xxxxxxxxxx> wrote:
> >>
> >> On Tue, Sep 9, 2025 at 6:32 AM Richard W.M. Jones <rjones@xxxxxxxxxx> wrote:
> >>>
> >>> On Tue, Sep 09, 2025 at 01:22:43PM +0000, Eric Dumazet wrote:
> >>>> Recently, syzbot started to abuse NBD with all kinds of sockets.
> >>>>
> >>>> Commit cf1b2326b734 ("nbd: verify socket is supported during setup")
> >>>> made sure the socket supported a shutdown() method.
> >>>>
> >>>> Explicitely accept TCP and UNIX stream sockets.
> >>>
> >>> I'm not clear what the actual problem is, but I will say that libnbd &
> >>> nbdkit (which are another NBD client & server, interoperable with the
> >>> kernel) we support and use NBD over vsock[1]. And we could support
> >>> NBD over pretty much any stream socket (Infiniband?) [2].
> >>>
> >>> [1] https://libguestfs.org/nbd_aio_connect_vsock.3.html
> >>> https://libguestfs.org/nbdkit-service.1.html#AF_VSOCK
> >>> [2] https://libguestfs.org/nbd_connect_socket.3.html
> >>>
> >>> TCP and Unix domain sockets are by far the most widely used, but I
> >>> don't think it's fair to exclude other socket types.
> >>
> >> If we have known and supported socket types, please send a patch to add them.
> >>
> >> I asked the question last week and got nothing about vsock or other types.
> >>
> >> https://lore.kernel.org/netdev/CANn89iLNFHBMTF2Pb6hHERYpuih9eQZb6A12+ndzBcQs_kZoBA@xxxxxxxxxxxxxx/
> >>
> >> For sure, we do not want datagram sockets, RAW, netlink, and many others.
> >
> > BTW vsock will probably fire lockdep warnings, I see GFP_KERNEL being used
> > in net/vmw_vsock/virtio_transport.c
> >
> > So you will have to fix this.
>
> Rather than play whack-a-mole with this, would it make sense to mark as
> socket as "writeback/reclaim" safe and base the nbd decision on that rather
> than attempt to maintain some allow/deny list of sockets?
Even if a socket type was writeback/reclaim safe, probably NBD would not support
arbitrary socket type, like netlink, af_packet, or af_netrom.
An allow list seems safer to me, with commits with a clear owner.
If future syzbot reports are triggered, the bisection will point to
these commits.
