ARP response with link local IP, why not broadcast
Subsystems (2)
Hello!
According to RFC3927 every ARP packet (reply and request) should be sent as link layer broadcast as long as the sender IP is a link local address. (see chapter 2.5). That functionality would help me a lot with a use case I have with our application.
But it is not implemented in the kernel that way.
Does anyone know why?
Regards,
Sebastian
Quoted 5 lines · Sebastian Fett
Hello!
According to RFC3927 every ARP packet (reply and request) should be sent as link layer broadcast as long as the sender IP is a link local address. (see chapter 2.5).
Because broadcast replies are noisy and should be avoided.
if possible- it creates a broadcast flood that would wake up all receivers, and is especially undesirable in today's world, where bcast would wake up sleepy devices, or require other inefficient processes in a cloud env. See also https://www.ietf.org/id/draft-nordmark-6man-dad-approaches-01.txt
Quoted 2 lines
That functionality would help me a lot with a use case I have with our application.
what is your use case?
Quoted 3 lines
But it is not implemented in the kernel that way.
Does anyone know why?
--Sowmini
Quoted 13 lines · Sebastian Fett
Quoted 5 lines
Hello!
According to RFC3927 every ARP packet (reply and request) should be sent as link layer broadcast as long as the sender IP is a link local address. (see chapter 2.5).
Because broadcast replies are noisy and should be avoided.
if possible- it creates a broadcast flood that would wake up all receivers, and is especially undesirable in today's world, where bcast would wake up sleepy devices, or require other inefficient processes in a cloud env. See also https://www.ietf.org/id/draft-nordmark-6man-dad-approaches-01.txt
Quoted 5 lines
Quoted 2 lines
That functionality would help me a lot with a use case I have with our application.
what is your use case?
My problem ist a local network of audio devices. It is a valid possibility that two halfs of the setup are set up individually (Stage left and stage right). Both local networks will auto configure themselves via link local and will be stable. But there always can be two devices with the same IP in both networks. At one point those two networks will be connected. With the current behaviour the conflicting devices will never know of each other and the address conflict.
The devices are controlled by a central PC using avahi/bonjour. It will know of all conflicting devices, but will only be able to talk to the one that happens to be in it's ARP cache. And renewing that cache will not change anything, because it will happen with unicast messages.
I looked at a Dante Controller (an audio data streaming device). And here all ARP messages are answered with broadcasts.
I think that behaviour is acceptable because it only happens in local networks. Waking up sleeping devices will not be a concern there.
That brings me to another question. When I react to an ARP packet in a userspace program, can I keep that packet from reaching the kernel as well? I would like to avoid to completely handle ARP in userspace.
Quoted 3 lines
Quoted 3 lines
But it is not implemented in the kernel that way.
Does anyone know why?
Sebastian
Quoted 11 lines · Sebastian Fett
Quoted 2 lines
what is your use case?
My problem ist a local network of audio devices. It is a valid possibility that two halfs of the setup are set up individually (Stage left and stage right). Both local networks will auto configure themselves via link local and will be stable. But there always can be two devices with the same IP in both networks.
At one point those two networks will be connected. With the current behaviour the conflicting devices will never know of each other and the address conflict.
Ah yes, this is a valid problem (Partition-Join tolerance) and one that is being discussed in the Ipv6 context on 6man:
http://www.ietf.org/mail-archive/web/ipv6/current/msg22712.html
FWIW, when Solaris implemented ACD (rfc 5227) the compromise that was made between bcasting *every* ARP response whle solving the type or issue that you describe was to use a periodic ARP "announce", advertising the IP address (a Grat ARP) with exponential backoff. If a duplicate address is triggered (as would happen in the scenario that you describe) the system would fall into the aggressive defend mode.
ARP announcemnts were bcast, but the noise is mitigated by tunable exponential backoff.
Of course, all of this only helps to *detect* the duplicate- eventually some other entity has to jump in and arbitrate on which one should "own" the address.
Quoted 10 lines
The devices are controlled by a central PC using avahi/bonjour. It will know of all conflicting devices, but will only be able to talk to the one that happens to be in it's ARP cache. And renewing that cache will not change anything, because it will happen with unicast messages.
I looked at a Dante Controller (an audio data streaming device). And here all ARP messages are answered with broadcasts.
I think that behaviour is acceptable because it only happens in local networks. Waking up sleeping devices will not be a concern there.
I dont know if a short term solution that makes sense here is to have a tunable for this.
But even the "always bcast arp response" will fail if you have a silent rejoin of the partitioned network- there is a reliance on the owner of an address bcasting their ARP resp at some point right?
(there's also a DoS vector here- I can create a lot of bcast traffic by arping for an address..)
Quoted 5 lines
That brings me to another question. When I react to an ARP packet in a userspace program, can I keep that packet from reaching the kernel as well? I would like to avoid to completely handle ARP in userspace.
David Miller
maintainerFrom: Sebastian Fett <[email protected]>
Date: Wed, 22 Jul 2015 09:49:49 +0200
Quoted 2 lines
I think that behaviour is acceptable because it only happens in local networks. Waking up sleeping devices will not be a concern there.
No, it is not acceptable.
If your laptop or cell phone's wireless interface is sleeping, they are on a "local" network and shouldn't be woken up.