On Fri, 2025-03-21 at 09:46 +0100, Xose Vazquez Perez wrote:
> A keep alive timer is implemented in the NVMe-oF protocol by default,
> which acts like a path checker.
> Therefore, change path_checker to default, "none" for NVMe arrays.
I'm not sure about this. NVMe keep-alive will check the connection to
the NVMe controller, while our path checker is designed to check the
availability of a given NVMe namespace, which is only almost the same.
With the "none" checker, multipathd will simply trust the sysfs state.
We have seen issues in the past (e.g. with the LIO target) where even
use of the TUR checker resulted in paths being reinstated by multipathd
even though they weren't actually usable, causing them to toggle back
and forth between UP and DOWN state. I can imagine this happening for
NVMe, too.
Does it do any harm to keep the directio checker enabled for NVMe as an
extra safety guard?
Most people will be using native kernel multipath for NMVe these days.
I consider dm-multipath on NVMe as a feature in maintenance mode, and
I'd prefer applying only necessary bug fixes in the area.
Thanks,
Martin
>
> Cc: Martin Wilck <mwilck@xxxxxxxx>
> Cc: Benjamin Marzinski <bmarzins@xxxxxxxxxx>
> Cc: Christophe Varoqui <christophe.varoqui@xxxxxxxxxxx>
> Cc: DM-DEVEL ML <dm-devel@xxxxxxxxxxxxxxx>
> Signed-off-by: Xose Vazquez Perez <xose.vazquez@xxxxxxxxx>
> ---
> libmultipath/hwtable.c | 1 -
> 1 file changed, 1 deletion(-)
>
> diff --git a/libmultipath/hwtable.c b/libmultipath/hwtable.c
> index cf39a50d..862bfdaa 100644
> --- a/libmultipath/hwtable.c
> +++ b/libmultipath/hwtable.c
> @@ -1100,7 +1100,6 @@ static struct hwentry default_hw[] = {
> /* OceanStor NVMe */
> .vendor = "NVM[eE]",
> .product = "Huawei-XSG1",
> - .checker_name = DIRECTIO,
> .no_path_retry = 12,
> },
> /*
