Re: Comments on addr-select-sol

[Rémi Denis-Courmont <rdenis@simphalempin.com>]

On Wednesday 25 July 2007 07:34:43 you wrote:
> IMO, trial-and-error approach can work with RFC3493.
> You may suppose that the kernel changes the source address of the
> existing connections, but it doesn't. The kernel just stores cache
> for successful/unworkable address selection and utilize it for the
> next address selection.
>
> Rather, I am suspicious about "question-and-answer" approach about
> this point. How can we implement that in getaddrinfo() without
> changing the APIs ?

Indeed, if an application does:

getaddrinfo("2001:db8::1", "53", { 0, SOCK_DGRAM, 0, AI_NUMERICHOST }, ...)
socket()
connect()

none of the function calls are allowed to block (UDP connect() cannot block, 
numeric getaddrinfo() cannot block), so you cannot ask anything from the 
network. So it is indeed doubtful whether the Q&A approach works. In fact, I 
suspect only the RFC3484-ish approach can work at all.

As such, I'd much rather go for the approach put forward by Brian Carpenter, 
which if I understand correctly, consists of extending RFC3484 so you can 
deliver the policy tables at network configuration time.


I might be wrong. But I think any approach will need to be prototyped before 
it's selected.

-- 
Rémi Denis-Courmont