> On Mar 5, 2021, at 9:36 AM, J. Bruce Fields <bfields@xxxxxxxxxxxx> wrote: > > On Thu, Mar 04, 2021 at 11:31:53AM -0500, Steve Dickson wrote: >> >> >> On 3/4/21 9:06 AM, J. Bruce Fields wrote: >>> On Thu, Mar 04, 2021 at 08:57:28AM -0500, Steve Dickson wrote: >>>> Personally I see this is the first step away from V3... >>>> >>>> So what we don't need is all that RPC code, all the different mounting >>>> versions... no RPC code at all, which also means no need for libtirpc... >>>> That is a lot of code that goes away, which I think is a good thing. >>> >>> libtirpc is a shared library, it'll still be loaded as long as anyone >>> needs it, and I'm not convinced we'll be able to get rid of all users. >>> >>>> I never thought it was a good idea to have mountd process >>>> the v4 upcalls... I always thought it should be a different >>>> deamon... and now we have one. >>>> >>>> A simple daemon that only processes v4 upcalls. >>> >>> I really do get the appeal, I've always liked the idea too. >>> >>> I'm not sure it's bringing us a real practical advantage at this point, >>> compared to rpc.mountd, which can act either as a daemon that only >>> processes v4 upcalls or can do both, depending on how you start it. >> Right with some configuration changes... But I do think there is >> value with have a package that will work right out of the box! I didn't understand this claim. Don't we already have that? If not, why not? >> Boom! Install the package and you have a working v4 server >> with no configure changes... I do think there is value there. > > Installing rpms and enabling systemd units is also a form of > configuration. Don't understand this argument either. What does "with no configuration changes" mean exactly? Installing and enabling is, as Bruce says, about the same administrative hassle as editing /etc/nfs.conf. What's being avoided here? > So maybe it comes down to whether we'd rather configure a v4-only server > with: > > dnf install nfsv4-only-server > systemctl enable nfsv4-server > > vs. > > edit some stuff in /etc/nfs.conf > > My preference is for the second, but it's just a feeling, I don't really > have an objective argument either way. Anyone else? Now, I'm making these comments in good faith. I don't want Steve to roll his eyes and think "what a pain in the ass". Please read and digest. Maybe have a Scotch and get a good night's sleep before replying. I really really don't want an argument, especially not right before a weekend. (And note please that I've given some alternatives here, so I'm not just trying to knock down the ideas). I feel strongly that the first approach is wrong, but it's Steve's ballgame so I'm not going to go as far as a NAK. We've spent a decade trying to fix the mistake of having an "nfs" and "nfs4" filesystem type on the client. Now we want to do the same thing on the server, and in the end, if and when NFSv3 goes away, we will be stuck with nfsv4-only-server. All I'm saying is we should proceed very carefully so we don't paint ourselves into a corner in the long run. This game is ultimately about lowering long-term maintenance cost. First, if only NFSv4 is configured in /etc/nfs.conf, then just don't start rpcbind. We already do this for statd when NFSv2 or NFSv3 is mounted on an NFS client. Let's automate and hide as much complexity as possible. Second, "dnf install nfsv4-only-server" by itself makes sense if you want to remove rpcbind and a few other components that are not ever going to be used, and you want to prevent dnf from ever installing them again during an update. But instead of nfsv4-only-server I might consider splitting the RPM generated by nfs-utils.src: one called nfs4-utils that installs the nfsv4-only stuff, and nfs-utils, which depends on nfs4-utils and installs the non-v4 stuff too. No changes to systemd units, please! And "dnf install nfs-utils" as before still provides the whole user space suite. Folks who just want an NFSv4 server can uninstall nfs-utils and install nfs4-utils (or if nothing is on the system, just install nfs4-utils). The configuration steps, and this is critical, are otherwise exactly the same in both cases. Third, as for exportd... if we're not changing the upcall mechanism, then just having mountd not start network listeners seems entirely adequate. I don't see the sense in replacing it. If you want smaller binaries, then let's focus on putting the whole nfs-utils package on a diet. Everyone wins that way. -- Chuck Lever
