Trond Myklebust wrote:
> On Thu, 2007-11-15 at 12:23 -0500, Chuck Lever wrote:
>> Trond Myklebust wrote:
>>> On Thu, 2007-11-15 at 08:55 -0500, Chuck Lever wrote:
>>>
>>>> What do you think of changing the rq_daddr field to be a sockaddr_storage?
>>> Why? You've already got rq_addr.
>> OK, here's what I've ended up with.  I think this is what you and Neil 
>> have been suggesting.  Comments?
>>
>> /*
>>   * Find an NLM client handle in the cache. If there is none, create it.
>>   *
>>   * Manufacture a specific source address in case we're using multiple
>>   * IP addresses on a single NIC.  NB: the family of the address in
>>   * rq_daddr is guaranteed to be the same as the family of rq_addr.
>>   */
>> struct nlm_host *
>> nlmsvc_lookup_host(struct svc_rqst *rqstp,
>> 			const char *hostname, unsigned int hostname_len)
>> {
>> 	struct sockaddr_in6 source = {
>> 		.sin6_family	= AF_INET6,
>> 	};
>>
>> 	switch(svc_addr(rqstp)->sa_family) {
>> 	case AF_INET: {
>> 		struct sockaddr_in *sin = (struct sockaddr_in *)&source;
>> 		sin->sin_family = AF_INET;
>> 		sin->sin_addr.s_addr = rqstp->rq_daddr.addr.s_addr;
>> 		break;
>> 	}
>> 	case AF_INET6:
>> 		ipv6_addr_copy(&source.sin6_addr,
>>                                 &rqstp->rq_daddr.addr6);
>> 		break;
>> 	}
>>
>> 	return nlm_lookup_host(1, svc_addr_in(rqstp),
>> 			       rqstp->rq_prot, rqstp->rq_vers,
>> 			       hostname, hostname_len,
>> 			       (struct sockaddr *)&source);
>> }
> 
> That works for me. Technically, I suppose we should really pass a
> source_len to nlm_lookup_host() if only to ensure that the comparisons
> don't overrun memory, however in practice I doubt this is a problem. We
> always know that the kernel allocates enough space.

My IPv6-aware version of nlm_cmp_addr takes a pair of sockaddr *'s; no 
lengths are needed for the comparison of either AF_INET or AF_INET6 
addresses.

In fact, in most cases, it appears that the address length is unneeded 
for us.  Both INET families use fixed-size addresses.