Re: PATCH: Support binding to a local IPv4 address when mounting a server.

[Ben Greear <greearb@candelatech.com>]

Chuck Lever wrote:
> Hi Ben-
>
> On Feb 21, 2009, at Feb 21, 2009, 2:43 AM, Ben Greear wrote:
>> I re-worked the kernel nfs local-address-binding logic against 
>> 2.6.29-rc5.  In quick testing,
>> this works with IPv4 and NFSv3 at least.  I changed the attribute to 
>> be called 'bindaddr'
>> as previously suggested.
>>
>> I didn't actually make any further changes to the mount.nfs tool and 
>> it took the bindaddr=a.b.c.d
>> just fine, so maybe there are no changes at all needed in user-space.
>
> You probably want the code in support/nfs/getport.c to send requests 
> from your bindaddr, if the mount command has to contact the server to 
> renegotiate mount options.

Any clue as to which part of this code to modify and how to test if it 
actually
does anything useful?

>>
>>     /* Don't talk to strangers */
>> -    clp = nfs_find_client(svc_addr(rqstp), 4);
>> +    clp = nfs_find_client(svc_daddr(rqstp), svc_addr(rqstp), 4);
>
> It's not clear to me why the callback server needs to be aware of the 
> mount point's bind address.  Can you explain this a little more?  I 
> would think the bind address would be pertinent for sending callback 
> service replies, but I don't see code here to do that.
I want to have a unique mount per local IP, so if I mount the same 
server from 1.1.1.1 and from 1.1.1.2, I
want two unique mounts.  I believe the way to do this is to 
differentiate based on the IP addr,
which is what this code is supposed to be doing.
> Would lockd also need to have this information too (passed in via 
> nlmclnt_inet)?
>
> Would we also want kernel rpcbind requests to be sensitive to the 
> passed-in bind address?
As much as possible, I want each mount to appear to be a unique machine 
with a particular
IP..so likely the answer is yes to both the above questions.  If you 
have an idea of which
code should be modified, please point me at it...I'll try searching on 
my own in the meantime.

>> +        case Opt_bindaddr:
>> +            string = match_strdup(args);
>> +            if (string == NULL)
>> +                goto out_nomem;
>> +            nfs_parse_ip_address(string, strlen(string),
>> +                         (struct sockaddr *)&mnt->bindaddr.address,
>> +                         &mnt->bindaddr.addrlen);
>> +            kfree(mnt->bindaddr.hostname);
>> +            mnt->bindaddr.hostname = string;
>
> You don't appear to use bindaddr.hostname anywhere, and I don't think 
> it is adequate to free it only here.
>
> Maybe you don't need it at all?

Yeah, I'll get rid of that.

Thanks,
Ben

-- 
Ben Greear <greearb@candelatech.com> 
Candela Technologies Inc  http://www.candelatech.com