Re: [PATCH] nvme: make controller 'state' sysfs attribute pollable

[Sagi Grimberg]

On 25/09/17 08:59, Hannes Reinecke wrote:
> On 09/25/2017 07:37 AM, Sagi Grimberg wrote:
> > > So why exposing it then in the first time? I know you don't want
> > > dm-mpath in
> > > NVMe (neither do I) but we have to have something until your patchset
> > > and ANA
> > > is merged. And with this patch it's trivial to build a path checker
> > > that just
> > > looks at the state attribute in sysfs.
> >
> > Can't we just not use path-checkers for nvme (we already have one in
> > nvme)?
>
> Really? For NVMe?
> How would you do that, then?

Quick and dirty is to have a path-checker that returns path-up always,
when the path go down, nvme will detect it and fast-fail the io.

> Anyway: the entire point is that you don't _need_ a path checker for NVMe.
> The primary reason for path checkers is to check with the transport
> layer if the remote endpoint is reachable.
> (I know, that's not quite what they're doing now, but that's beside the
> point).
> For NVMf we do have KATO, so the NVMe subsystem knows exactly if the
> connection is live or not. So it should be perfectly sufficient to check
> the connection state instead of running a path checker of sorts.
> But for doing so we need something in sysfs which we could check.
>
> Mind you, I wouldn't be adverse to have some common sysfs attribute,
> with some common values (eg path up, path down, path blocked), and have
> NVMf translating the internal state into that.

We could have such an interface I assume. But it would suck to maintain
yet another state (we are already having enough trouble to have a
coherent controller state machine).