Bug#1040343: linux-image-5.10.0-9-amd64: Kernel silenty de-supported nfsv3 UDP mounts
On 7/5/23 10:20, Bastian Blank wrote:
On Tue, Jul 04, 2023 at 06:35:44PM +0200, Hauke Fath wrote:
/misc /etc/auto.misc -nfsvers=3,proto=udp,resvport,retrans=5,rsize=16384,wsize=16384,rw,hard
And if you set it to TCP (the default) or better directly switch to
NFSv4?
nfsv3 over tcp works, but is subobtimal , as described - when the router
goes down, the tcp mounts will hang, and the machine will have to be
rebooted.
We do not use nfsv4 here.
This upstream decision
<https://www.spinics.net/lists/linux-nfs/msg74889.html> is more than
debatable -- we have been running nfsv3 over UDP for ~20 years here
without ever seeing the data corruption that was claimed as
motivation.
That you have to take up with upstream, not us.
Note we are not talking about code changes here. This is about a kernel
configuration option, which is very much at the discretion of a
distribution (as the Arch example has shown).
Also this talks about
problem with fragment reassembly, not data corruption itself. But
usually I would trust upstream to know more about it then yourself.
At this point, we have 20 years of experience with running nfsv3 over
udp for ~40 clients that mount user homes over nfs.
On 7/5/23 14:34, Ben Hutchings wrote:
This was an upstream change in Linux 5.6 that we won't override. NFS-
over-TCP has been well supported on Linux, and better performing, for a
long time.
This request is not about defaulting, or even preferring, udp over tcp.
It is simply about having the option, for interoperability as well as
for situations (and they do exist, despite the blanket statement), where
nfsv3 over udp provides more robust service -- without having to deploy
a self-compiled kernel.
Please re-consider, and re-open the ticket.
Cheerio,
Hauke
--
The ASCII Ribbon Campaign Hauke Fath
() No HTML/RTF in email Institut für Nachrichtentechnik
/\ No Word docs in email TU Darmstadt
Respect for open standards Ruf +49-6151-16-21344
Reply to: