Re: [Qemu-devel] [PATCH 3/3] s390x/css: generate channel path initialized CRW for channel path hotplug

[Cornelia Huck <cohuck@redhat.com>]

On Fri, 28 Jul 2017 16:29:14 +0200
Halil Pasic <pasic@linux.vnet.ibm.com> wrote:

> On 07/28/2017 02:58 PM, Cornelia Huck wrote:
> > On Fri, 28 Jul 2017 14:32:11 +0200
> > Halil Pasic <pasic@linux.vnet.ibm.com> wrote:
> >   
> >> On 07/28/2017 12:11 PM, Cornelia Huck wrote:  
> >>> On Thu, 27 Jul 2017 18:15:07 +0200
> >>> Halil Pasic <pasic@linux.vnet.ibm.com> wrote:  
> >   
> >>>> So my intention was to ask: What benefits do we expect from these 'real'
> >>>> virtual channel paths?     
> >>>
> >>> Path grouping and friends come to mind. This depends on whether you
> >>> want to pass-through channel paths to the guest, of course, but you
> >>> really need management to deal with things like reserve/release on ECKD
> >>> correctly.    
> >>
> >> Pass-through means dedicated in this case (that is the passed trough paths
> >> are not used by the host -- correct me if my understanding is wrong).  
> > 
> > There's nothing that speaks against path sharing, I think.  
> 
> That is a nice to hear. I could not form an opinion on this
> myself yet. Theoretically we speak about shared physical resources here,
> and in such situations I'm wary of interference. A quick look into
> the AR documents was not conclusive.

I'm afraid that much of it will be either underdocumented, confusingly
worded, or even druidic knowledge. Experimenting a bit might be helpful.

> 
> I'm still trying to figure out this whole channel path handling,
> and frankly you are a big help right now.

Thanks.

> 
> > Especially as e.g. SetPGID is "first one gets to set it".  
> 
> Hm, I don't understand this. (I've found a description of SETPGID
> in "IBM 3880 Storage Control Models ... " but could not get your
> point based no that.)

The first OS that does SetPGID after a reset (or removal of a PGID),
sets it. Subsequent SetPGIDs are rejected if the PGID does not match.
(See the SensePGID/SetPGID handling in the Linux common I/O layer -
this is needed e.g. for running under z/VM.)


> >>  
> >>> Also failover etc. Preferred channel paths are not relevant
> >>> on modern hardware anymore, fortunately (AFAIK).
> >>>    
> >>
> >> If I understand you correctly it ain't possible to handle these
> >> in the host (and let the guest a simple 'non-real' virtual
> >> channel path whose reliability depends on what the host does),
> >> or?  
> > 
> > It is possible. Mapping to a virtual channel path or not is basically a
> > design decision (IIRC, z/VM supports both).
> > 
> > Mapping everything to a virtual chpid basically concentrates all
> > path-related handling in the hypervisor. This allows for a dumb guest
> > OS, but can make errors really hard to debug from the guest side.
> >   
> 
> IMHO the same is true for virtio for example (the abstraction
> hides the backend and the backing: if there is a problem there it's
> hard to debug from the guest side).

In a way, yes. But it is way more on the virtual side of things :)

> 
> Because of my lack of understanding, this option appeared simpler to
> me: clear ownership, and probably also less places where things can
> go wrong.

My gut feeling is that exposing channel paths is the easier way in the
long run.

> 
> > Exposing real channel paths to the guest means that the guest OS needs
> > to be able to deal with path-related things, but OTOH it has more
> > control. As I don't think we'll ever want to support a guest OS that
> > does not also run under LPAR, I'd prefer that way.  
> 
> Nod. And this makes a full circle, namely the question of benefit of
> having more control. But since we did one full circle I'm much smarter
> now than at the beginning.
> 
> Thank you very much for all the background information and for your
> patience.

Well, I hope I'm not confusing everyone too much :)