Re: [PATCH 2/2] mkfs: remove notion of config "type"

From: Eric Sandeen <sandeen@sandeen.net>
To: Dave Chinner <david@fromorbit.com>
Cc: "Luis R. Rodriguez" <mcgrof@gmail.com>,
	"Darrick J. Wong" <darrick.wong@oracle.com>,
	linux-xfs <linux-xfs@vger.kernel.org>
Subject: Re: [PATCH 2/2] mkfs: remove notion of config "type"
Date: Fri, 15 Jun 2018 19:55:02 -0500	[thread overview]
Message-ID: <558124ff-bc3f-f969-072c-3db8c9493bfd@sandeen.net> (raw)
In-Reply-To: <20180616001743.GT10363@dastard>

t

On 6/15/18 7:17 PM, Dave Chinner wrote:
> On Thu, Jun 14, 2018 at 09:45:49PM -0500, Eric Sandeen wrote:
>>
>>
>> On 6/14/18 9:33 PM, Dave Chinner wrote:
>>> On Thu, Jun 14, 2018 at 07:10:09PM -0700, Luis R. Rodriguez wrote:
>>>> Silly mobile gmail interface not letting me bottom-post... What if we treat
>>>> no version being present as version 0?
>>>
>>> We haven't released anything yet so we should put it in there from
>>> the start rather than having to work around the lack of a version
>>> field later.
>>
>> So, pretend I'm dumb ('cause I often am) and spell it out for me, what would
>> we do with a version?
>>
>> If a config file contains a section or token that some version of mkfs doesn't
>> understand, it'll fail.
>>
>> If we try to read a config file with a too-new version, we'd ... fail?
> 
> Fail with a useful error message, rather than do something
> unexpected or incorrect.

like "section 'foo' uknown" or "token 'bar' unknown?"
We'd do that today...

> Let's face it - the config file is a persistent, on-disk structure
> that we have to handle in both forwards and backwards compatible
> manners for many, many years. It's no different to the on-disk
> format in that respect. Why wouldn't we apply the same guards for
> format changes we apply to syscalls, ioctls, on-disk formats, etc
> that all have the same long term compatibility requirements?

Indeed, with on-disk formats we have compat, incompat, ro-compat ....

Are you proposing forward and backward, compat/imcompat ... tokens?
Pobably not.

Disk formats have that because we may corrupt the things around us if
we don't understand the existence of a new metadata type; others we can
safely ignore.

With ioctls we have a set size; the only time we have versions is if we
have padding that should now be parsed in a newer version.

I'm trying to find a concrete example of how an incompatible config
file version would fail more gracefully than an unknown section/token already
fails today.  Maybe I lack imagination, but help me out?

In the abstract it sounds good.  In practice, I'm not yet seeing it.

-Eric