Re: [Qemu-devel] [PATCH V4 for 3.1 1/4] net: drop too large packet early

[Jason Wang]

On 2018/12/4 上午12:18, Eric Blake wrote:
> On 12/3/18 4:06 AM, Jason Wang wrote:
> > We try to detect and drop too large packet (>INT_MAX) in 1592a9947036
> > ("net: ignore packet size greater than INT_MAX") during packet
> > delivering. Unfortunately, this is not sufficient as we may hit
> > another integer overflow when trying to queue such large packet in
> > qemu_net_queue_append_iov():
> >
> > - size of the allocation may overflow on 32bit
> > - packet->size is integer which may overflow even on 64bit
> >
> > Fixing this by move the check to qemu_sendv_packet_async() which is
>
> s/move/moving/


Ok.


> > the entrance of all networking codes and reduce the limit to
> > NET_BUFSIZE to be more conservative.
>
> Please mention commit 1592a994 in the commit message (since you are 
> effectively reverting that with this as its replacement),


I think I did it?


> and if this is (as I suspect) an additional patch required for the 
> complete fix to CVE-2018-10839, also mention that.


Ok.


> > Cc: address@hidden
> > Cc: Li Qiang <address@hidden>
> > Reported-by: Li Qiang <address@hidden>
> > Reviewed-by: Li Qiang <address@hidden>
> > Signed-off-by: Jason Wang <address@hidden>
> > ---
> >   net/net.c | 13 +++++++------
> >   1 file changed, 7 insertions(+), 6 deletions(-)
> >
> > diff --git a/net/net.c b/net/net.c
> > index 07c194a8f6..affe1877cf 100644
> > --- a/net/net.c
> > +++ b/net/net.c
> > @@ -712,15 +712,11 @@ ssize_t qemu_deliver_packet_iov(NetClientState 
> > *sender,
> >                                   void *opaque)
> >   {
> >       NetClientState *nc = opaque;
> > -    size_t size = iov_size(iov, iovcnt);
> >       int ret;
> >   -    if (size > INT_MAX) {
> > -        return size;
> > -    }
> >         if (nc->link_down) {
> > -        return size;
> > +        return iov_size(iov, iovcnt);
>
> Reverts 1592a994, and...
>
> >       }
> >         if (nc->receive_disabled) {
> > @@ -745,10 +741,15 @@ ssize_t qemu_sendv_packet_async(NetClientState 
> > *sender,
> >                                   NetPacketSent *sent_cb)
> >   {
> >       NetQueue *queue;
> > +    size_t size = iov_size(iov, iovcnt);
> >       int ret;
> >   +    if (size > NET_BUFSIZE) {
> > +        return size;
> > +    }
>
> ...returns early for packets larger than 68k (a much smaller limit 
> than INT_MAX, which makes analysis for int overflow a lot easier) at a 
> saner point in the code.  Returning a large value is weird, 


Might be, but we did this for years, see the following return value when 
link is down.


> but auditing all callers:
>
> hw/net/virtio-net.c:        ret = 
> qemu_sendv_packet_async(qemu_get_subqueue(n->nic, queue_index),
>  - only checks if ret is 0 (returns -EBUSY) or not (assumes packet was 
> sent)
> net/netmap.c:        iovsize = qemu_sendv_packet_async(&s->nc, s->iov, 
> iovcnt,
>  - only checks if ret is 0 (starts polling) or not (assumes packet was 
> sent)
> net/net.c:    return qemu_sendv_packet_async(nc, iov, iovcnt, NULL);
>  - implementation of qemu_sendv_packet() - so we have to audit those 
> callers as well:
>
> hw/net/net_tx_pkt.c:        qemu_sendv_packet(nc, iov, iov_cnt);
> hw/net/rocker/rocker_fp.c:        qemu_sendv_packet(nc, iov, iovcnt);
> hw/net/rtl8139.c: qemu_sendv_packet(qemu_get_queue(s->nic), iov, 3);
> net/hub.c:        qemu_sendv_packet(&port->nc, iov, iovcnt);
> - all four of these do not check the return status
>
> So, it looks like none of the callers cares if the return value is 
> overlarge (no further math on the values), just that it is non-zero 
> (where the callers then presumably assume the packet was sent). 


Yes, the caller may only care if it returns zero.


> However, I am not familiar enough with the code to know if skipping 
> the packet by returning a non-zero value is going to have knock-on 
> effects - that is, my audit shows what the callers do, but not whether 
> it was sane.

The difference between qemu_sendv_packet() and qemu_sendv_packet_async() 
is that the latter can trigger a callback (sent_cb) when peer can accept 
more packets. This could be used by some high speed networking 
implementation to prevent the source from producing more packets and 
wasting cpu cycles in dropping packets. After peer can accept more, 
sent_cb usually enable the source to produce packets. Those who use 
qemu_sendv_packet() will just waste some cpu in dropping the packets.

Consider we are emulating ethernet and packet will be copied if queued, 
it's safe to assume that the packet was sent.


> > +
> >       if (sender->link_down || !sender->peer) {
> > -        return iov_size(iov, iovcnt);
> > +        return size;
> >       }
>
> If this is indeed CVE fixing, then we want it in -rc4, but I don't 
> know if the patch is correctly secure yet without answers to my 
> questions. Especially on a CVE fix for -rc4, you want to make the 
> reviewer's life as easy as possible by providing a commit message that 
> includes enough details to make analysis easy.


Hope my answer help. If it is, I will add them to the commit log.

Thanks for the reviewing.