svn commit: r328129 - head/sys/fs/nfsserver

[Rick Macklem]

Emmanuel Vadot wrote:
[stuff snipped]
> So should we warn once or maybe return EBUSY on unloading if there is
>still lock structures ?
My intent was that a module unload would clear out all data structures,
so I would so no. I would also say that I envisioned an unload of nfsd.ko as
a last resort, to free up the data structures.
It should refuse to unload if the nfsd daemon is running.

Isolan used to unload/load the nfsd.ko a lot, because they preferred that
for testing and software updates.
I, personally, have never done so (except once in a while to see if it works),
so I'll admit I have trouble understanding why you would do so?

Rebooting an NFSv4 server when clients are mounted forces the clients
to do state recovery during a grace period (2 minutes+) right after boot.
This isn't something that I would recommend as a routine exercise, but
at least it should recover the opens/locks if it goes smoothly.
(I can't recall for sure, but I don't think an unload/reload of the nfsd.ko
 followed by a startup of the nfsd daemon does this? If I am wrong and
 it does do this, then this is exactly like a reboot w.r.t. nfs service.
 If you look at a packet trace in wireshark just after the server restart,
 you would see errors like NFS4ERR_STALE_CLIENTID, NFS4ERR_STALE_STATEID,
 NFS4ERR_BAD_SESSION that indicates to the client that recovery of state is
 needed.
--> If this state recovery isn't happening then your mounts will probably be
      broken, at least w.r.t. byte range locking.
 If the unload/reload doesn't do this, then a reboot is preferable.

Because of this rather complex state recovery exercise, any reboot or
unload/reload/restart of the nfsd should be done only when absolutely
necessary, imho.
If you can't run a server for an extended period with an unload of the
nfsd.ko or a reboot, using NFSv3 mounts would be a safer bet.

I wrote:
> For this case, normally all lock structures should go away when clients
> unmount and unloading the nfsd module while there are active mounts
> is not a safe practice. (Again NFSv4 isn't like NFSv3, there is server state
> for NFSv4.)
>
> rick
rick