Re: client caching and locks

From: "NeilBrown" <neilb@suse.de>
To: "'bfields@fieldses.org'" <bfields@fieldses.org>
Cc: "inoguchi.yuki@fujitsu.com" <inoguchi.yuki@fujitsu.com>,
	"'Trond Myklebust'" <trondmy@hammerspace.com>,
	"'linux-nfs@vger.kernel.org'" <linux-nfs@vger.kernel.org>,
	"'mbenjami@redhat.com'" <mbenjami@redhat.com>
Subject: Re: client caching and locks
Date: Mon, 10 Jan 2022 09:16:21 +1100	[thread overview]
Message-ID: <164176658189.25899.1767614988647740830@noble.neil.brown.name> (raw)
In-Reply-To: <20220106141628.GA7105@fieldses.org>

On Fri, 07 Jan 2022, 'bfields@fieldses.org' wrote:
> On Thu, Jan 06, 2022 at 07:23:16AM +0000, inoguchi.yuki@fujitsu.com wrote:
> > > How about this?  I also updated the lock/nolock description and deleted
> > > the end of this subsection since it's redundant with that. And removed
> > > the bit about using nolock to mount /var with v2/v3 as that seems like a
> > > bit of a niche case at this point.  If we still want to document that, I
> > > think it belongs elsewhere.
> > 
> > Thank you so much for creating the patch!
> > 
> > For the most part I agree with you, but I feel unsafe to remove the part 
> > "using nolock to mount /var with v2/v3" even if it seems niche case. 
> > I'm also not sure if there is another suitable document to migrate it. 
> > 
> > Therefore, at the end of "lock/nolock" subsection ...
> 
> I could live with that.
> 
> Though the other reason I cut it was because I think it needs updates
> too and I wasn't sure exactly how to handle them.
> 
> The v4 case is more important and should probably be dealt with first.
> I think the answer there is just "don't mount /var over NFSv4", period.

Why not? An NFSv4 client has no business accessing anything in
/var/lib/nfs, so how can it cause a problem?

> 
> And maybe we should be more specific: the problem is with /var/lib/nfs,
> not all of /var.

According to "3.2. CLIENT STARTUP" section "C/" in the README that comes
with nfs-utils,
      statd should be run before any NFSv2 or NFSv3 filesystem is
      mounted with remote locking (i.e. without -o nolock).

so it is only fair to say "the problem is with /var/lib/nfs" if that is
a separate filesystem from /var.

Note that "-o nolock" should also be used for the root filesystem for
root-over-NFSv3, as that must be mounted before statd can be running.

Maybe that it how it should be explained in the man page:

  -o nolock should be used to mount any NFSv3 filesystems that are
   mounted before rpc.statd can be started, which can only be started
   after /var/lib/nfs is available.

NeilBrown