[ofa-general] ***SPAM*** ibdm network topology format

[Hal Rosenstock]

Sasha,

On Thu, Oct 2, 2008 at 6:22 PM, Hal Rosenstock <hal.rosenstock at gmail.com> wrote:
> Sasha,
>
> On Thu, Oct 2, 2008 at 1:00 PM, Sasha Khapyorsky <sashak at voltaire.com> wrote:
>> Hi Hal,
>>
>> On 10:18 Thu 02 Oct     , Hal Rosenstock wrote:
>>> >
>>> > 2. ibis doesn't register class 0x81 - SM direct routed, only SM lid
>>> > routed (0x1). In comment in ibutils/ibis/src/ibsm.c line 118 is stated:
>>> >
>>> >  /* no need to bind the Directed Route class as it will automatically
>>> >     be handled by the osm_vendor_bind if asked for LID route */
>>> >
>>> > As far as I can see in osm_vendor_bind() it is not (but it is in
>>> > opposite order - when class 0x81 is registered class 0x1 will be
>>> > registered too).
>>>
>>> Yes that is what osm_vendor_ibumad.c:osm_vendor_bind does.
>>>
>>> So either ibdiagnet needs to register 0x81 r.t.1 or
>>> osm_vendor_ibumad.c:osm_vendor_bind needs to be "symmetric" in terms
>>> of registering the other SM class when only one is requested. This is
>>> a minor change in the underlying semantics. [Popping up a level in
>>> terms of this, (other than applications taking advantage of this
>>> "feature",) I'm not sure why the vendor layer should assume that just
>>> because one SM class is requested, the other should be too].  I just
>>> looked and the latter appears to be consistent with the other vendor
>>> layers. I think either solution will work. Your solution below also
>>> looks like it would work but don't that should be done in a sim layer.
>>
>> I'm not like this "solution" too, but the fact that ibis works with real
>> stack without registering 0x81 class is unclear for me.
>
> Me too. See below.
>
>>> > Somehow it works without ibsim - so I suspect user_mad handles it.
>>> >
>>> > (Hal, could you clarify?)
>>>
>>> The kernel (user_mad/mad) does not change the requested registrations
>>> but I'm not sure I understand the question you are asking to be
>>> clarified. Is that what you're asking ?
>>
>> ibis works somehow with real stack. It registers 0x1 class only and
>> uses direct routing SMPs. Do you have any idea about why
>> (osm_vendor_idumad and/or libibumad don't help)?
>
> libibumad umad_register does not do anything that would affect this
> either. I can only conclude there must be something in ibutils that
> fixes this if it does work with the real stack. It shouldn't be too
> hard to track down where that registration for class 0x81 comes from.

Are you sure this is the only registration and not DR class too ?
That's the first thing to confirm or maybe you've already confirmed
this and it wasn't clear to me in what you wrote. If so, I have a
theory about what could be occuring. It may be the case that it is an
effect of the kernel MAD layer in that a MAD agent can send any class
and when using request/response it matches on transaction ID which
contains the MAD agent. Unsolicited messages on that other class
wouldn't get through though. I just ran a simple test of this and that
appears to be the case.

-- Hal

>
> -- Hal
>
>> Sasha
>>
>