Allowing an unlimited number of clients to any web service is a recipe
for a rudimentary denial of service attack: the client merely needs to
open lots of sockets without closing them, until qemu no longer has
any more fds available to allocate.
For qemu-nbd, we default to allowing only 1 connection unless more are
explicitly asked for (-e or --shared); this was historically picked as
a nice default (without an explicit -t, a non-persistent qemu-nbd goes
away after a client disconnects, without needing any additional
follow-up commands), and we are not going to change that interface now
(besides, someday we want to point people towards qemu-storage-daemon
instead of qemu-nbd).
But for qemu proper, and the newer qemu-storage-daemon, the QMP
nbd-server-start command has historically had a default of unlimited
number of connections, in part because unlike qemu-nbd it is
inherently persistent until nbd-server-stop. Allowing multiple client
sockets is particularly useful for clients that can take advantage of
MULTI_CONN (creating parallel sockets to increase throughput),
although known clients that do so (such as libnbd's nbdcopy) typically
use only 8 or 16 connections (the benefits of scaling diminish once
more sockets are competing for kernel attention). Picking a number
large enough for typical use cases, but not unlimited, makes it
slightly harder for a malicious client to perform a denial of service
merely by opening lots of connections withot progressing through the
handshake.
This change does not eliminate CVE-2024-7409 on its own, but reduces
the chance for fd exhaustion or unlimited memory usage as an attack
surface. On the other hand, by itself, it makes it more obvious that
with a finite limit, we have the problem of an unauthenticated client
holding 100 fds opened as a way to block out a legitimate client from
being able to connect; thus, later patches will further add timeouts
to reject clients that are not making progress.
This is an INTENTIONAL change in behavior, and will break any client
of nbd-server-start that was not passing an explicit max-connections
parameter, yet expects more than 100 simultaneous connections. We are
not aware of any such client (as stated above, most clients aware of
MULTI_CONN get by just fine on 8 or 16 connections, and probably cope
with later connections failing by relying on the earlier connections;
libvirt has not yet been passing max-connections, but generally
creates NBD servers with the intent for a single client for the sake
of live storage migration; meanwhile, the KubeSAN project anticipates
a large cluster sharing multiple clients [up to 8 per node, and up to
100 nodes in a cluster], but it currently uses qemu-nbd with an
explicit --shared=0 rather than qemu-storage-daemon with
nbd-server-start).
We considered using a deprecation period (declare that omitting
max-parameters is deprecated, and make it mandatory in 3 releases -
then we don't need to pick an arbitrary default); that has zero risk
of breaking any apps that accidentally depended on more than 100
connections, and where such breakage might not be noticed under unit
testing but only under the larger loads of production usage. But it
does not close the denial-of-service hole until far into the future,
and requires all apps to change to add the parameter even if 100 was
good enough. It also has a drawback that any app (like libvirt) that
is accidentally relying on an unlimited default should seriously
consider their own CVE now, at which point they are going to change to
pass explicit max-connections sooner than waiting for 3 qemu releases.
Finally, if our changed default breaks an app, that app can always
pass in an explicit max-parameters with a larger value.
It is also intentional that the HMP interface to nbd-server-start is
not changed to expose max-connections (any client needing to fine-tune
things should be using QMP).
Suggested-by: Daniel P. Berrangé <berrange@redhat.com>
Signed-off-by: Eric Blake <eblake@redhat.com>
Message-ID: <20240807174943.771624-12-eblake@redhat.com>
Reviewed-by: Daniel P. Berrangé <berrange@redhat.com>
[ericb: Expand commit message to summarize Dan's argument for why we
break corner-case back-compat behavior without a deprecation period]
Signed-off-by: Eric Blake <eblake@redhat.com>
(cherry picked from commit c8a76dbd90)
Signed-off-by: Michael Tokarev <mjt@tls.msk.ru>
(Mjt: minor fixups in qapi/block-export.json)
dbf946c251