[PATCH 4/4] nvme: use per device timeout waits over depending on global default value

[David Jeffery]

On Fri, Sep 13, 2024 at 11:17 AM Keith Busch <kbusch at kernel.org> wrote:
>
> On Fri, Sep 13, 2024 at 09:44:30AM -0400, David Jeffery wrote:
> > Instead of passing NVME_IO_TIMEOUT as a parameter with every call to
> > nvme_wait_freeze_timeout, use the largest timeout value from all the
> > namespaces for the nvme controller. This will match the wait to the
> > configuration of the devices instead of assuming a single value can
> > fulfill the needs of all devices on the system.
>
> The module parameter is obviously too coarse, but does IO timeout really
> need to be per-namespace? I'd think setting the preferred timeout at the
> controller level is sufficient.
>

The expansion of nvme to be a remote storage interface means different
namespaces provided by the same controller can have very different
characteristics. Take someone making an nvme-tcp target where one
namespace is backed by local, high speed SSDs while another namespace
is storage backed by a slower or more finicky data solution like nfs
or various clustered storage products. I would expect such cases to be
rare, but they are still cases where independent timeout values may be
needed to better handle different namespaces through the same
controller.