Jeremy Harris wrote:
Erik Nordmark wrote:
Yes, but for UDP performance we want to handle the case when a socket
is repeatedly sending to the same destination differently, for
instance by caching the IRE and the source address that was selected.
If/when we fully do this, we can also latch the policy lookup.
Is this a common case, or are application writers sensible enough
to use a connected UDP socket? What's the cost/benefit analysis?
I don't know how common connected UDP sockets vs. repeated
sendto/sendmsg to the same address are.
But I don't think it matters that much for this discussion.
The issue is the complexity of ip_wput/ip_wput_ire.
If we can cleanly move most or all of the part of ip_wput* that require
access to the conn_t up to a sub-layer above ip_wput, then we can do
this complexity reduction.
A side effect of this is that handling caching of the IRE, source
address selection, and IPsec policy lookup for unconnected UDP sockets
more or less come for free.
Erik
_______________________________________________
networking-discuss mailing list
[email protected]
