[PATCH blktests v2 0/3] Test different queue counts

[Daniel Wagner]

On Thu, Mar 23, 2023 at 11:06:53AM +0000, Shinichiro Kawasaki wrote:
> On Mar 22, 2023 / 11:16, Daniel Wagner wrote:
> > Setup different queues, e.g. read and poll queues.
> > 
> > There is still the problem that _require_nvme_trtype_is_fabrics also includes
> > the loop transport which has no support for different queue types.
> > 
> > See also https://lore.kernel.org/linux-nvme/20230322002350.4038048-1-kbusch@meta.com/
> 
> Hi Daniel, thanks for the patches. The new test case catches some bugs. Looks
> valuable.
> 
> I ran the test case using various nvme_trtype on kernel v6.2 and v6.3-rc3, and
> observed hangs. I applied the 3rd patch in the link above on top of v6.3-rc3 and
> confirmed the hang disappears. I would like to wait for the kernel fix patch
> delivered to upstream, before adding this test case to blktests master.

Okay makes sense.

> When I ran the test case without setting nvme_trtype, kernel reported messages
> below:
> 
> [  199.621431][ T1001] nvme_fabrics: invalid parameter 'nr_write_queues=%d'
> [  201.271200][ T1030] nvme_fabrics: invalid parameter 'nr_write_queues=%d'
> [  201.272155][ T1030] nvme_fabrics: invalid parameter 'nr_poll_queues=%d'

BTW, I've added a '|| echo FAIL' to catch those.

> Is it useful to run the test case with default nvme_trtype=loop?

No, we should run this test only for those transport which actually support the
different queue types. Christoph suggest to figure out before running the test
if it is actually supported. So my first idea was to check what options are
supported by reading /dev/nvme-fabrics. But this will return all options we are
parsed by fabrics.c but not the subset which each transport might only support.

So to figure this out we would need to do a full setup just to figure out if it
is supported. I think the currently best approach would just to limit this test
to tcp and rdma. Maybe we could add something like

rc:
_require_nvme_trtype() {
	local trtype
	for trtype in "$@"; do
		if [[ "${nvme_trtype}" == "$trtype" ]]; then
			return 0
		fi
	done
	SKIP_REASONS+=("nvme_trtype=${nvme_trtype} is not supported in this test")
	return 1
}

047:
requires() {
	_nvme_requires
	_have_xfs
	_have_fio
	_require_nvme_trtype tcp rdma
	_have_kver 4 21
}

What do you think?

Thanks,
Daniel