Message-ID: <4B58B8FE.80800@oracle.com>
Date: Thu, 21 Jan 2010 15:28:46 -0500
From: Chuck Lever <chuck.lever@oracle.com>
To: "J. Bruce Fields" <bfields@fieldses.org>
CC: Jeff Layton <jlayton@redhat.com>, linux-nfs@vger.kernel.org,
        steved@redhat.com, trond.myklebust@fys.uio.no
Subject: Re: [PATCH] mount.nfs: prefer IPv4 addresses over IPv6 (try #2)
References: <1263907662-19107-1-git-send-email-jlayton@redhat.com> <1667647A-2BB1-478D-8881-CE8EA2191F97@oracle.com> <20100120082905.1825f806@tlielax.poochiereds.net> <8E556CE2-569B-4E6D-BD02-7EF5CA84900D@oracle.com> <20100121191515.GA22021@fieldses.org> <4B58AD16.8040309@oracle.com> <20100121195746.GC22021@fieldses.org>
In-Reply-To: <20100121195746.GC22021@fieldses.org>
Content-Type: text/plain; charset=us-ascii; format=flowed
Sender: linux-nfs-owner@vger.kernel.org
MIME-Version: 1.0

On 01/21/2010 02:57 PM, J. Bruce Fields wrote:
> On Thu, Jan 21, 2010 at 02:37:58PM -0500, Chuck Lever wrote:
>> On 01/21/2010 02:15 PM, J. Bruce Fields wrote:
>>> On Wed, Jan 20, 2010 at 10:36:36AM -0500, Chuck Lever wrote:
>>>> For the record, we looked at Solaris behavior yesterday.  With bi-family
>>>> servers, its mount command tries IPv6 first, but appears smart enough to
>>>> fall back to IPv4.  One thing we haven't tried is to see how difficult it
>>>> would be to fix the real problem by adding proper protocol family
>>>> negotiation to our own mount command.
>>>
>>> Sorry, I probably just haven't been following: what's "proper protocol
>>> family negotiation"?  I thought the only ways to negotiate were either
>>> rpcbind (v2, v3) or trial and error (v4)?
>>
>> In TI-RPC parlance, a "protocol" is the transport protocol (UDP, for
>> example), and a "protocol family" is the address family ("inet6", for
>> example).  A netid represents a particular combination of the two:  the
>> netid "udp6" represents UDP over "inet6".
>>
>> The "protocol family" is really the value that is passed to socket(2).
>> This call generally takes PF_INET or something like that as its first
>> argument.  All of the PF_FOO thingies have the same integer value as
>> their AF_FOO counterparts.  For TI-RPC, we have "inet" and "inet6",
>> which are strings that match up with the AF_FOO and PF_FOO names.
>>
>> rpcb_getaddr(3t) is designed to use the rpcbind protocol to determine
>> the address and transport to use when contacting a remote service.  Our
>> mount command has its own negotiation mechanism that is a superset of
>> rpcbind calls, in addition to having a faster timeout than
>> rpcb_getaddr(3t).
>
> What does "is a superset of rpcbind calls" mean?

rpcb_getaddr(3t) performs a single specific rpcbind query with a long 
fixed timeout.  mount.nfs uses several rpcbind queries, in a particular 
order, to identify which NFS-related services are available.  mount.nfs 
uses individual queries rather than a single DUMPALL in order to enable 
firewalls to detect which ports should be opened.

> I still don't
> understand what the proper protocol family negotiation is: what actually
> happens on the wire?

If a particular RPC service (including rpcbind) cannot be contacted via 
"inet6," and the server has an "inet" address listed in DNS, then 
mount.nfs should be smart enough to try the mount request via the "inet" 
address too.  This is in addition to support for rpcbind queries that 
can return a netid, which would include information about which protocol 
family to use).

Currently our mount.nfs command fails if the target server has at least 
one IPv6 address listed in DNS in addition to an IPv4 address, but does 
not support NFS/IPv6.

For NFSv4, a server that has an IPv6 address but does not support 
NFS/IPv6 will refuse connection to port 2049 over IPv6.  In that case, 
mount.nfs should tell the kernel to retry the mount with the server's 
IPv4 address, if it has one.

For NFSv3, a server that has an IPv6 address, but does not support 
NFS/IPv6, will not register any inet6 netids in its rpcbind database. 
Thus the mount.nfs command has to be smart enough to retry 
PROGNOTREGISTERED results with the server's IPv4 address, if it has one.

If the server has an IPv6 address, but is running portmap instead of 
rpcbind, the initial rpcbind query connection will be refused (portmap 
does not set up an IPv6 listener).  In that case, the mount request 
should be retried with the server's IPv4 address, if it has one.

Note that in any of these cases, if an NFS server does not have any IPv6 
addresses listed in DNS, then behavior should be the same as before.

-- 
chuck[dot]lever[at]oracle[dot]com