On 02/02/2017 07:21 AM, Eggert, Lars wrote: > On 2017-2-2, at 10:54, Fernando Gont <fgont@xxxxxxxxxxxxxxx> wrote: >> On 02/02/2017 06:37 AM, Eggert, Lars wrote: >>> Also, even if ICMP delivery is assured, there are additional >>> complications for UDP, which has been seeing a lot of interest >>> both as a tunneling encapsulation and for applications (e.g., >>> QUIC). Many platforms do not provide UDP-sending applications any >>> information about arriving ICMP messages that were triggered by >>> their transmissions. So even if the path delivers ICMP, the OS >>> makes ICMP-based PMTUD for UDP often impossible. Another reason >>> to recommend 4821? >> >> Agreed... although in this case this would be more of an app-layer >> implementation than one at the transport layer? > > There are two dimensions here, one is in kernel vs. in userspace, the > other which "layer" something is at. It used to be that "transport > layer" (or "network layer" always implied "in kernel", but those days > are past. I was refering to "layer" as in a reference model. i.e., PMTU assumes that you can repacketize your data. For TCP, that can be done at the transport layer, since it's byte-stream oriented. OTOH, UDP is essentially record-based... so any "re-packetization" must be done at the upper (app) layer. Thanks, -- Fernando Gont SI6 Networks e-mail: fgont@xxxxxxxxxxxxxxx PGP Fingerprint: 6666 31C6 D484 63B2 8FB1 E3C4 AE25 0D55 1D4E 7492