[SLOF] [PATCH 1/2] ping: use gateway address for routing

[Thomas Huth]

On 22.04.2016 07:23, Nikunj A Dadhania wrote:
> Thomas Huth <thuth at redhat.com> writes:
> 
>> On 13.04.2016 09:03, Nikunj A Dadhania wrote:
>>> ping was failing for machine across the subnet with
>>> statically assinged IP address. The parsed gateway address
>>> was ignored in the stack because the router variable was not
>>> set.
>>>
>>> Currently, ping does not take a netmask parameter, set netmask parameter
>>> with the common top order bits of client IP and gateway IP.
>>>
>>> Signed-off-by: Nikunj A Dadhania <nikunj at linux.vnet.ibm.com>
>>> ---
>>>  clients/net-snk/app/netapps/ping.c | 6 ++++++
>>>  1 file changed, 6 insertions(+)
>>>
>>> diff --git a/clients/net-snk/app/netapps/ping.c b/clients/net-snk/app/netapps/ping.c
>>> index 2c7dadb..1522792 100644
>>> --- a/clients/net-snk/app/netapps/ping.c
>>> +++ b/clients/net-snk/app/netapps/ping.c
>>> @@ -163,6 +163,12 @@ ping(int argc, char *argv[])
>>>  
>>>  	} else {
>>>  		memcpy(&fn_ip.own_ip, &ping_args.client_ip.integer, 4);
>>> +		if (ping_args.gateway_ip.integer) {
>>> +			uint32_t netmask;
>>> +			netmask = ping_args.gateway_ip.integer & ping_args.client_ip.integer;
>>> +			set_ipv4_netmask(netmask & 0xFFFFFF00);
>>
>> That netmask calculation looks somewhat strange to me ... what's the
>> rationale here?
> 
> Ping currently only supports IPv4. Case here was that subnet mask is not
> provided and we are trying to guess the subnet mask. Dont think this
> will be accurate always, but will work.

I think this will be very fragile. For example, let's assume client
address is 192.168.1.10, router address is 192.168.1.1 and TFTP server
address is 192.168.3.2 (i.e. TFTP server is in another subnet).

With your calculation, you get netmask = 192.168.1.0. When the
send_ipv4() code then tries to find out whether the TFTP server is in
the same network as the client (to find out whether a router MAC has to
be used), it does:

   if((subnet_mask & own_ip) == (subnet_mask & ip->ip_dst)) ...

and thus in above example, it will think that the TFTP server is in the
same subnet as the client and fail to use the router address.

>> If you need automatic netmask calculation, shouldn't
>> that rather be done by the type of IP address that has been specified?
> 
> Here its only IPv4. We still dont have ipv6 support here. I have that in
> my list though.

I didn't mean IPv6, I was talking about IPv4 address classes (sorry,
should have used that term). For Class A, B and C, the default netmask
should be applied as long as no CIDR prefix length has been specified
(see the first table on
https://en.wikipedia.org/wiki/IPv4_subnetting_reference)

>> (or maybe via an "/xx" suffix of the specified IP address string?)
> 
> I am not sure about this, can you please elaborate?

Sorry again for not using the right terms ... I meant the specifying the
prefix length (and thus the netmask) with a CIDR prefix notation, see
https://en.wikipedia.org/wiki/Classless_Inter-Domain_Routing#CIDR_notation

That means, it might maybe be useful to support specifying the client
address with a prefix length like this:

 load net:192.168.3.2,filename,192.168.1.10/24

What do you think?

 Thomas