RE: New NTB API Issue

From: Allen Hubbe
Date: Fri Jun 23 2017 - 09:18:48 EST

From: Logan Gunthorpe
> On 6/22/2017 4:42 PM, Allen Hubbe wrote:
> > From: Logan Gunthorpe
> >> Any thoughts on changing the semantics of mw_get_align so it must be
> >> called with the link up?
> >
> > The intention of these is that these calls return information from the local port. The calls
> themselves don't reach across the link to the peer, but the information returned from the local port
> needs to be communicated for setting up the translation end-to-end.
> Ok, well if it's from the local port, then splitting up mw_get_range
> into peer_mw_get_addr and mw_get_align is confusing because one has the
> peer designation and the other doesn't. And all the clients apply the
> alignments to the remote bar so they'd technically need to transfer them
> across the link somehow.

By "remote" do you mean the source or destination of a write?

Yes, clients should transfer the address and size information to the peer.

> > I would like to understand why this hardware needs link up. Are there registers on the local port
> that are only valid after link up?
> We only need the link up if we are trying to find the alignment
> requirements (and max_size) of the peer's bar. In theory, switchtec

Maybe this is the confusion. None of these api calls are to reach across to the peer port, as to get the size of the peer's bar. They are to get information from the local port, or to configure the local port.

Some mw configuration is done at the destination, as for Intel and AMD, and some configuration is done on the source side, for IDT. The local configuration of the port on one side could depend on information from the remote port on the other side. For example in IDT, the mw translation configured on the source side needs to know destination address information. Likewise, if there is any difference in the size of the range that can be translated by ports on opposing sides, that needs to be negotiated.

> could have different sizes of bars on both sides of the link and
> different alignment requirements. Though, in practice, this is probably
> unlikely.
> > Can you post snippets of how ntb_mw_get_align and ntb_peer_mw_get_addr might be implemented for the
> Switchtec?
> Hmm, yes, but lets sort out my confusion on the semantics per above first.

Some snippets of code would help me understand your interpretation of the api semantics more exactly.

> Logan