sounds good.. can you send us a patch to look at?
On Wed, 6 Mar 2002, Terry Lambert wrote: > There are redundant calls to the in_pcblookup_hash() in the > ip_fw_chk() function called via (*ip_fw_chk_ptr)() in the > ip_input path. > > Would it be useful to modify the (*pr_input) function pointer > in the struct ipprotosw to take a fourth argument (perhaps it > should be cast to a "void *" to keep it generalized?) to pass > the pre-looked-up struct inpcb * to TCP, if the lookup has > already been done? > > Profiling indicates that this is one of the most expensive > calls in the code path, particularly when there are a lot > of sockets open. Increasing the hash table size only works > so far; at "a lot", the number of connections makes the > lookup expensive anyway (it's a linear traversal of the > collision chain for the bucket). > > Since there are reasons other than firewalling to do the > lookup early, it seems that it would be useful to pass a > pointer the a pointer that was non-NULL, if the lookup had > already taken place. For example, moving the ipflow to > use an overlay structure for the inpcb would mean that a > single lookup was used for fast forwarding, firewalling, > and inpcb identification for tcpcb retrieval for TCP. > > Note that I'm only talking about the packet input path here, > at this time, so the firewall code isn't really generalizable > (the inpcb is already known on output, except to the ip_fw > code; it probably doesn't make sense to push knwledge of it > into the ip_output path, at least without more thought). > > Right now, I'm just talking about a way ip_input could pass > the already looked up input inpcb to tcp_input, udp_input, > or udp_ctlinput -- all of which repeat the lookup operation. > > -- Terry > > To Unsubscribe: send mail to [EMAIL PROTECTED] > with "unsubscribe freebsd-net" in the body of the message > To Unsubscribe: send mail to [EMAIL PROTECTED] with "unsubscribe freebsd-hackers" in the body of the message
