[openib-general] one request

Roland Dreier roland at topspin.com
Tue Aug 24 10:08:51 PDT 2004


    Roland> I don't believe there is a circular dependency (otherwise
    Roland> it would be impossible to load the modules).  The problem
    Roland> is actually that the MAD module does _not_ depend on the
    Roland> Tavor driver, so module dependencies don't automatically
    Roland> load the Tavor driver.

    ron> no, that's not what the code is doing.

    ron> Have you run this stuff? It's really clear if you do.

I wrote most of it :)  I haven't tried the Mellanox gold CD but it
sounds like they may have pulled a broken/old tree.

    ron> tavor load fails on an ib_mad symbol, and if you load mad
    ron> without loading tavor, then the interface list won't get
    ron> initialized correctly. So what the ib_mad stuff does is
    ron> demand-load the tavor driver about halfway through its module
    ron> init.

For a while I did have a request_module() of infiniband_hca in
mad_main.c at the very end of initialization, to help people who
forgot to load the ib_tavor module.  However that broke on 2.6 because
it seems module symbols are not visible until init_module() has
returned, so I took the request_module() out.  I don't think the
request_module was ever halfway through initialization.

You can look at the gen1 tree to see how things stood before we gave
up on that development.

    ron> I'll find the source and quote chapter and verse if that
    ron> helps :-)

That would be interesting....

 - R.



More information about the general mailing list