On Fri, Apr 20, 2001 at 02:15:23PM -0300, Fernando Barreto wrote: > Hi... > > I'm working to do a new network stack... I can now receive my > packets throught the wire without any problems, but: How can I transfer > one packet that arrived to one process? And how can I block the process > until the data arrive to the device driver, or until the data become > already transmited? > Is there anywhere in the kernel that can I find this > implementation? net/core/dev.c net/ipv4/ -- lots -- Follow the call of netif_rx() at the ethernet card RX path INSIDE THE INTERRUPT CONTEXT. See how it puts things into some queue, and triggers a BH instance, and returns away. Triggered BH is actually net_rx_action(), which continues with processing of the packet -- doing diversions, fastroute, bridging, and finally looking up the Ethernet content TYPE field (first 2 bytes of the frame after the MAC addresses). When a matching one is found, associated protocol processing receiver function is called. See at first how IPv4 does init itself by registering different levels of protocols. Then observe the processing path of for example UDP frames. I wonder what "data already transmitted" means ? Shall the caller stop until the remote end acks having received the data ? Or is there some flow-control and kernel side buffering ? (Ever heard of TCP of the IP suite ?) Whatever, the answer is most likely used deep inside the TCP, and UDP codes of IPv4 protocol family. > I'm thinking to mask one sys_call of the system to do that... Is > it the better way? Definitely. Register your protocol as AF_XYZ with numeric value that is not presently in use. Make it "just another socket protocol family." > Thanks for any help... > :-) /Matti Aarnio - : send the line "unsubscribe linux-net" in the body of a message to majordomo@vger.kernel.org
