Re: [PATCH 00/10] exposing knfsd opens to userspace

From: NeilBrown <neilb@suse.com>
To: "J. Bruce Fields" <bfields@fieldses.org>
Cc: Andreas Dilger <adilger@dilger.ca>,
	"J. Bruce Fields" <bfields@redhat.com>,
	linux-nfs <linux-nfs@vger.kernel.org>,
	linux-fsdevel <linux-fsdevel@vger.kernel.org>,
	abe@purdue.edu, lsof-l@lists.purdue.edu,
	util-linux@vger.kernel.org, Jeff Layton <jlayton@redhat.com>,
	James Simmons <uja.ornl@gmail.com>
Subject: Re: [PATCH 00/10] exposing knfsd opens to userspace
Date: Mon, 29 Apr 2019 08:57:37 +1000	[thread overview]
Message-ID: <874l6h68m6.fsf@notabene.neil.brown.name> (raw)
In-Reply-To: <20190427190038.GB9568@fieldses.org>

[-- Attachment #1: Type: text/plain, Size: 2422 bytes --]

On Sat, Apr 27 2019, J. Bruce Fields wrote:

> On Sat, Apr 27, 2019 at 09:55:23AM +1000, NeilBrown wrote:
>> On Fri, Apr 26 2019, J. Bruce Fields wrote:
>> > But it's true that from the start nfsd didn't really fit the model
>> > of a single (possibly writeable) attribute per file.
>> 
>> Depends on what you mean by that.  Original files where write-only and
>> where slightly complex attributes.
>
> Yes I thought it was just those too, but then I looked at the original
> commit it also included at least the "exports" file.

Maybe it depends on how one chooses to interpret history - never an
exact science.

The "exports" file pre-existed the nfsd filesystem - it was (and still
is) a file in procfs: /proc/fs/nfs/exports.  So the nfsd filesystem was
not created to provide that file.  It was created to replace a
systemcall.
As I said, it subsequently had a variety of things added to it.  exports
was just the first of these.  At least, that is how I choose to see it.

>
>> Writing performed an action, like
>> adding an entry to the export table (first you add a client, then add a
>> client+filesystem to export it).
>> 
>> This idea for a file performing an action, rather than presenting an
>> attribute, is much the same as the "bind" and "unbind" files you can
>> find in sysfs.
>> 
>> (see also https://lwn.net/Articles/378884/ for examples of sysfs files
>> that are not one-attribute-per-file)
>
> I'll give that a re-read, thanks.
>
> I did spend maybe a few minutes looking into basing nfsd code on kernfs
> and didn't think it was worth it.  I could take a more serious look.

I think that in your use-case it make lots of sense to have a structured
file for the "opens" (similar to /proc/locks and /proc/mounts).
The "info" could reasonably be several attribute files (clientid,
address, name, minor_version), but I don't think it benefits anyone for
'opens' to be a directory full of directories each with a file for each
of the fields.

So using kernfs would mean pushing for allowing structured files in
kernfs.  I'd be happy to support that, but I think you would need to go
into it convinced that you really wanted to persist.

I think using kernfs is mostly about embedding a kobject in everything,
then setting an appropriate ktype with appropriate attributes.  Not
particularly complex, but certainly a bit of work.

Thanks,
NeilBrown

[-- Attachment #2: signature.asc --]
[-- Type: application/pgp-signature, Size: 832 bytes --]