You're better off asking this on netdev or kernel-newbies. regards, dan carpenter On Wed, Jan 15, 2025 at 08:34:55PM +0100, Alexandre Ferrieux wrote: > Hi, > > Recently digging into a driver-specific issue with sk_buff->protocol, I > discovered (with surprise) that the computation of this field (which is *always* > necessary for protocol handlers to kick in) was somehow "delegated" to > individual network interface drivers. This can be seen by looking for callers of > eth_type_trans(), e.g: > > - tg3: called from tg3_poll_work > - ixgbe: called from ixgbe_clean_rx_irq > - veth: called from veth_xmit (though indirectly via __dev_forward_skb) > > This is a surprise as one would naively expect this ubiquitous behavior to be > triggered from generic code, depending only on the L2 header structure (but not > on the specific NIC hardware at hand). Another surprise was *not* to find any > mention of this "contract" in Documentation/*. > > So, is it an unspoken tradition for NIC driver developers, to > "just know" that prior to emitting an skb from the rx path, they must fill > skb->protocol (along with who knows how many other metadata items) ? > > Or, is there *somewhere* some reference documentation listing (exhaustively) the > required metadata computations that fall within a NIC driver's responsibility ? > > Thanks in advance, > > -Alex
