On Thu, 27 Sep 2018, Christoph Paasch wrote: > Hello, > > I am trying to rework our handling of the skb-building in tcp_sendmsg > (avoiding custom function-pointers for select_size,...) to reduce our impact > on regular TCP. > > So, I started looking into what it takes to use do_tcp_sendpages when > pushing from the MPTCP-level to the subflow-level. > > A problem is that do_tcp_sendpages() only works on subflows that have > SG-support (see the check in tcp_sendpage_locked()). > > So, that would be kind of problematic, no? > > I'm also not 100% sure how that is handled for the in-kernel TLS :-/ Hi Christoph - I'm not quite sure why tcp_sendpage_locked() has that check. It may even force extra copying by GSO. If the net device does not have NETIF_F_SG when an skb is to be sent to the device, the skb will get linearized just before getting pushed to the device (see __dev_queue_xmit() and validate_xmit_skb()) What's more, NETIF_F_SG is not necessarily fixed for a given device over time (see commit message for 113f99c3358564a0647d444c2ae34e8b1abfd5b9), so the stack must be able to handle scatter/gather skbs from do_tcp_sendpages even if the device turns off that feature while those skbs are queued above the device layer. GSO being always enabled now may also be a factor in not requiring linearized skbs. This is what I've pieced together this afternoon, anyway - if anyone has better information, please share!