Date: Fri, 29 Mar 2002 10:16:31 -0800 From: Maksim Yevmenkin <myevmenk@digisle.net> To: freebsd-current@freebsd.org, freebsd-net@freebsd.org Subject: Flow control in -current Netgraph (long) Message-ID: <3CA4AF7F.4B0A1C93@digisle.net>
next in thread | raw e-mail | index | archive | help
Hackers, i'm writing Bluetooth L2CAP sockets layer in Netgraph. the basic idea is very similar to Netgraph sockets except L2CAP is a reliable protocol. before L2CAP channel is open, each peer must negotiate incoming and outgoing MTU for the channel. so the idea is that each peer sends and receives datagram of fixed MTU. note that send and receive MTU might be different. L2CAP protocol also provides protocol/service multiplexing (PSM) so upper layer protocols can use L2CAP as a transport (for example SDP, RFCOMM etc.) new L2CAP node is created for every Bluetooth unit. L2CAP downstream hook is connected to the HCI layer. also L2CAP node presents four upstream hooks that represents protocols (i.e. sdp, rfcomm, tcp and orphan - for other protocols). for now every L2CAP upstream hook of every L2CAP node is connected to the same L2CAP sockets layer. so the sockets layer knows which PSM/unit downstream hook is connected to. the socket address contains PSM and BDADDR. (BDADDR is a Bluetooth unit address similar to MAC address in Ethernet). when socket requesting new connection (via connect() call) new Netgraph control message is sent to appropriate downstream hook (based on which address socket was bound via bind() and remote peer's address from connect()). the L2CAP node responds to control message and eventually socket goes into CONNECTED state. now the data transfer is a bit of a problem. currently i'm using Netgraph NG_SEND_DATA macros to send data from socket to the L2CAP node. when i do stress testing and sending data as fast as i can the system runs out of mbufs. packets gets dropped and everything goes into unpredictable state. bottom line: in order to guarantee reliable L2CAP channel i must acknowledge every data packet. the upper layer must not send any more until is gets an acknowledgment. note that single L2CAP packet must be fragmented before it can be send to the HCI layer. it is not allowed to reorder/mix fragments on the same physical link. so, as of now i have three options 1) put data in queue on every layer and try not to overflow them. i do not know the easy and fast way to implement this. note that upstream hooks may be connected to the different nodes and L2CAP node must somehow synchronize them and make sure than everyone gets bandwidth. 2) prepend every data mbuf with some soft of envelope and as soon as packet was sent acknowledge it by sending Netgraph control message back to sender (using information from envelope). i can live with that, but it might be confusing. the sender sends data, but gets a message back. 3) convert every send request into Netgraph message and as soon as packet was sent acknowledge it by sending response message back. i like this too, but there is another problem. i can not pass mbuf in Netgraph control message. here is an example: socket layer pru_send function accepts mbuf. it converts it into Netgraph message (m_copydata) and sends it to downstream hook. lower layer would have to convert it back to mbuf (m_copyback) because it wants to fragment it, add headers, etc. yuck! i would go with option 2 for now unless someone can suggest a better way. thanks max To Unsubscribe: send mail to majordomo@FreeBSD.org with "unsubscribe freebsd-current" in the body of the message
Want to link to this message? Use this URL: <https://mail-archive.FreeBSD.org/cgi/mid.cgi?3CA4AF7F.4B0A1C93>