Only last IP frag sent if ARP entry absent

[Mike Karels]

Another $.02 (inline):

On 17 Aug 2017, at 18:39, Gopakumar Pillai wrote:

> Thank You Bjoern. Could you please point me to the RFC?

I don’t know if there is anything more recent than RFC1122 on this.  
IIRC, it requires queuing at least one packet.  Queing one packet is 
what BSD has done essentially since ARP was implemented.

> If this is not a MUST behavior in RFC, would my fix be good? I agree 
> that this would affect only ICMP/UDP traffic.

People have been asking for queuing of multiple packets for years.  That 
is a more general change.  Consider another dumb application that starts 
out by sending multiple UDP packets back-to-back.  However, 
well-designed application protocols don’t experience problems like 
this.  I’ll quickly note that ping isn’t an application, but a 
network measuring tool.  If you ask the question “what happens if I 
start off a session with a single large packet and I don’t support 
retransmission”, ping answers that question correctly.

If badly-designed protocols get bad performance, that doesn’t seem 
like a bug to me, but a feature.

> On 8/17/17, 2:40 PM, "Bjoern A. Zeeb" <bzeeb-lists at lists.zabbadoz.net> 
> wrote:
>
>     On 17 Aug 2017, at 21:16, Gopakumar Pillai wrote:
>
>     > Hi FreeBSD Networking Gurus,
>     > I came across an issue with an old version of FreeBSD and 
> looking at
>     > the latest FreeBSD code, seems it exists even now. I am assuming 
> that
>     > this issue is not reported.
>     >
>     > Observation:
>     > When a ping was performed with larger payload than MTU, the 
> first ping
>     > failed when the ARP entry was absent for that IP.
>
>     That is because ping/ICMP has no retransmit.
>
>
>     > Noticed on the wire that the last IP fragment was sent for the 
> first
>     > request and then the subsequent requests were fine.
>     >
>     > Root Cause:
>     >   * ip_output fragments the packets and loops through the 
> fragments to
>     > send them to ether_output.
>     >   * ether_output does an arpresolve and if there is no existing 
> ARP
>     > entry it'll return EWOULDBLOCK after sending ARP Request.
>     >   * ether_output ignores the error and propagates success to 
> ip_output
>     > and it continues to send the remaining fragments.
>     >   * llentry keeps only one mbuf and the last fragment is 
> retained when
>     > the ARP Reply comes and the fragment is sent.
>
>     Yes, according to the spec (RFC) we are supposed to throw the 
> packet
>     away entirely and simply report that to the next upper layer.  
> However
>     over the years people realised that this sucks for a TCP SYN 
> packet with
>     a retransmit timer and hence we store one of them.
>
>     A large UDP packet would btw see the same behaviour to your ping.
>     There’s no guarantee any of these packets will not be dropped 
> anywhere
>     on the network, so we can as well.
>
>     Just my 2ct
>
>     /bz

		Mike