On Tue, Nov 08, 2022 at 09:54:18AM +0100, Stefano Brivio wrote:
> I got all paranoid after triggering a divide-by-zero general
> protection fault in passt with a qemu version without the virtio_net
> TX hang fix, while flooding UDP. I start thinking this was actually
> coming from some random changes I was playing with, but before
> reaching this conclusion I reviewed once more the relatively short
> path in tap_handler_passt() before we start using packet_*()
> functions, and found this.
> 
> Never observed in practice, but artificially reproduced with changes
> in qemu's socket interface: if we don't receive from qemu a complete
> length descriptor in one recv() call, or if we receive a partial one
> at the end of one call, we currently disregard the rest, which would
> make the stream inconsistent.
> 
> Nothing really bad happens, except that from that point on we would
> disregard all the packets we get until, if ever, we get the stream
> back in sync by chance.
> 
> Force reading a complete packet length descriptor with a blocking
> recv(), if needed -- not just a complete packet later.
> 
> Signed-off-by: Stefano Brivio <sbrivio@redhat.com>

This seems an ok short term fix, but I think we want another approach
in the slightly longer term.  Read on..

> ---
>  tap.c | 18 +++++++++++++++---
>  1 file changed, 15 insertions(+), 3 deletions(-)
> 
> diff --git a/tap.c b/tap.c
> index f8314ef..11ac732 100644
> --- a/tap.c
> +++ b/tap.c
> @@ -747,14 +747,26 @@ redo:
>  		return -ECONNRESET;
>  	}
>  
> -	while (n > (ssize_t)sizeof(uint32_t)) {
> -		ssize_t len = ntohl(*(uint32_t *)p);
> +	while (n > 0) {
> +		ssize_t len;
> +
> +		/* Force receiving at least a complete length descriptor to
> +		 * avoid an inconsistent stream.
> +		 */

Is it actually enough for this to be blocking?  AFAICT, recv() on a
stream socket, like read(), can return less data than you requested.

> +		if (n < (ssize_t)sizeof(uint32_t)) {
> +			rem = recv(c->fd_tap, p + n,
> +				   (ssize_t)sizeof(uint32_t) - n, 0);
> +			if ((n += rem) != (ssize_t)sizeof(uint32_t))
> +				return 0;
> +		}
> +
> +		len = ntohl(*(uint32_t *)p);
>  
>  		p += sizeof(uint32_t);
>  		n -= sizeof(uint32_t);
>  
>  		/* At most one packet might not fit in a single read, and this
> -		 * needs to be blocking.
> +		 * also needs to be blocking.

Same issue here (obviously not introduced by this patch, though).

>  		 */
>  		if (len > n) {
>  			rem = recv(c->fd_tap, p + n, len - n, 0);

Can we handle both these cases more neatly (and without blocking
recv()) calls, if we maintain two pointers into pkt_buf.  The first
one tracks how much we've read from the qemu socket, the second tracks
how much has been parsed into packets.  When we get an epoll
notification on the qemu socket, we recv() and advance the first
pointer.  Then we discern as many full packets as we can, advancing
the second pointer.

-- 
David Gibson			| I'll have my music baroque, and my code
david AT gibson.dropbear.id.au	| minimalist, thank you.  NOT _the_ _other_
				| _way_ _around_!
http://www.ozlabs.org/~dgibson