[ofa-general] Re: [PATCH 1/2] libibumad: fix partition support

[Hal Rosenstock]

On Fri, 2007-06-15 at 17:36, Sean Hefty wrote:
> > Sigh... and opensm (actually libvendor) is the one which uses this
> > incorrectly. I'm worried about existing OpenSM compatibility with the
> > new libibumad when ABI 6 is in effect. I think the long standing ABI 5
> > should be fine, right ?
> 
> ABI 5 should be fine, since the pkey isn't actually passed to the 
> kernel.  ABI 6 would pass down the wrong index.

Right.

> I do print a warning if 
> umad_set_pkey() is called with an index != 0, but we can remove that.

How about if abi_version == 5, setting pkey_index to 0 regardless of
what is set ? Isn't that all that ABI v5 really supports ?

> >> old mode 100644
> >> new mode 100755
> > 
> > Why the mode change ?
> 
> This is just my editor being dumb, and me forgetting to tell git to 
> ignore mode changes.

OK.

> >> +	n = write(fd, data, size);
> >> +	if (n != size) {
> >> +		DEBUG("write returned %d != sizeof mad data %d (%m)", n, size);
> > 
> > Is this really the sizeof the mad data ?
> 
> This is not the size of the data field in a MAD.  It's the size of the 
> write = sizeof ib_user_mad + MAD data length.  I can change the comment 
> to clarify.

Thanks.

-- Hal

> - Sean