Re: [PATCH BLKTESTS V2 1/3] nvme/002: fix output format

From: Omar Sandoval <osandov@osandov.com>
To: Chaitanya Kulkarni <Chaitanya.Kulkarni@wdc.com>
Cc: "Theodore Y. Ts'o" <tytso@mit.edu>,
	"linux-block@vger.kernel.org" <linux-block@vger.kernel.org>
Subject: Re: [PATCH BLKTESTS V2 1/3] nvme/002: fix output format
Date: Wed, 27 Feb 2019 11:00:43 -0800	[thread overview]
Message-ID: <20190227190043.GB11806@vader> (raw)
In-Reply-To: <SN6PR04MB452790045C42F3DEAD63BE76867F0@SN6PR04MB4527.namprd04.prod.outlook.com>

On Fri, Feb 22, 2019 at 12:00:43AM +0000, Chaitanya Kulkarni wrote:
> On 02/21/2019 07:04 AM, Theodore Y. Ts'o wrote:
> > On Wed, Feb 20, 2019 at 11:35:00AM -0800, Omar Sandoval wrote:
> >>> It should fail on older cli versions, we want latest cli/tools to be
> >>> used when running these testcases so that all the latest tools are
> >>> also been tested for the compliance, else we might be running tests
> >>> with any bugs with older version(s) which are fixed already.
> >>>
> >>> Actually I'd like to add quick nvme-cli version check at the start of
> >>> each test that will certainly help/force user to have a latest version.
> >>
> >> Perfect, I was going to suggest adding a check for the minimum version
> >> needed for each test.
> >
> > It would be great if we could document what is (a) the minimum version
> > for things to work at all,
> Actually the latest version or the HEAD of the git repo should work all
> the time if it is breaking test or the tool or driver we need to add
> a fix respectively.
> 
>   and (b) the ideal version to use so that
> > tests aren't being skipped due to version constraints and so we are
> > testing the kernel as much as possible.
> >
> Agree.
> > Ideally, tests should either be designed primarily to test userspace,
> > or to test the kernel, so that people who want to use blktests aren't
> > having to build the latest versions of userspace.
> >
> At least with the current NVMe infrastructure we need latest tools
> (git HEAD) to be compliant all the time with the tests, since it
> guarantees the tools, driver and kernel compatibility.
> 
> We can a little install script to build and install the
> cli if desired version is not present.
> 
> > (As an aside, Bart has told me that the failures I'm seeing with
> > srp/002 and srp/011 are due to the fact that I'm using a version of
> > multipathd which is too old.  Apparently both the version of
> > multipathd in Debian stable *and* Debian unstable are still too old,
> > since they are still failing.  Next stop, building multipathd from
> > the git repo....   Grrr......)
> This is exact reason why tools needs to be built and check with latest
> HEAD.

I don't think we should be _requiring_ that the tests are run with the
latest tools from git. That just adds another annoying step to someone
trying to get the tests up and running. Like I said, each test should
check for the minimum version it needs, and we should avoid making the
golden output too dependent on version.

However, I do think it's a reasonable expectation that if you want
maximum test coverage, you should build the tools from git.