Loopback mtu 16436 ?

[SkyHiRider]

Why does the loopback interface has such a high mtu? I get it that it needs to work out even large packets, which is logical but why this specific number?
I can set the MTU even higher with the ip link command, so it's not a limitation of the variable used.

[stevea]

From linux-2.6-stable/drivers/net/loopback.c

/*
 * The loopback device is special. There is only one instance
 * per network namespace.
 */
static void loopback_setup(struct net_device *dev)
{
        dev->mtu                = (16 * 1024) + 20 + 20 + 12;
        dev->hard_header_len    = ETH_HLEN;     /* 14   */
        dev->addr_len           = ETH_ALEN;     /* 6    */
        dev->tx_queue_len       = 0;
        dev->type               = ARPHRD_LOOPBACK;      /* 0x0001*/
        dev->flags              = IFF_LOOPBACK;
        dev->priv_flags        &= ~IFF_XMIT_DST_RELEASE;
        dev->features           = NETIF_F_SG | NETIF_F_FRAGLIST
                | NETIF_F_TSO
                | NETIF_F_NO_CSUM
                | NETIF_F_HIGHDMA
                | NETIF_F_LLTX
                | NETIF_F_NETNS_LOCAL;
        dev->ethtool_ops        = &loopback_ethtool_ops;
        dev->header_ops         = &eth_header_ops;
        dev->netdev_ops         = &loopback_ops;
        dev->destructor         = loopback_dev_free;

16436 = = (16 * 1024) + 20 + 20 + 12

Remember that the lo device is not an 802 media - so there is no MAC address and encapsulation. The flags above indicate there is no checksum. I suppose the extra 52 bytes are sufficient to create the IP header with a 16KiB paymoad.

[DBelton]

ummm.. 16 * 1024 would be 16384 then if you add the 52 byte TCP/IP header information, it comes out to the 16436 number.

(52 bytes for the header is.. 20 bytes for the IP header, 20 bytes for the TCP header, then it pads it to the next 32 bit boundary. )


you can send a 16kb packet along with the header information without fragmenting the packet. That's why that magic number

Edit: I forgot that it also adds in each option as well, before it pads to the 32 bit boundary

[SkyHiRider]

Thanks, that number was really bugging me

[stevea]

Yeah - it's just an efficiency thing. Nice to be able to drop a 16k data unit in one "packet".

Still there is something wrong with your logic. 20bytes for IP + 20 for TCP is already at a 32bit boundary.

The 802 wrapper is 14B at the top and 4B at the tail, but lo: is NOT an 802 media so these don't exist.

TCP can take several options described in RFC1323 & 793. It seems the timestamp option is the longest
at length 10B, but it must be padded to 12B with.'NOP' option bytes. The other PAWS option is only on
syn or syn/ack packets that are typically very short. I am not aware of any other options.

Technically it's possible a TCP header would contain multiple options pushing the total header beyond
52 bytes, but that would be rare enough to ignore, and may require the invention of new options.