Re: [RFC PATCH] PCI: remove access to pci_[enable|disable]_msi() for drivers

[Roland Dreier <roland@topspin.com>]

    Greg> Hm, or does it really help for the driver to specify
    Greg> different numbers of msi-x vectors?

    Greg> Roland, any ideas here?  You seem to be the only one who has
    Greg> actually used the msix code so far.

Yes, I think that the driver definitely needs to be in control of how
many MSI-X interrupts it gets.  The current mthca driver knows that it
has three different event queues -- one for firmware command events,
one for async events such as link up/down, and one for actual tx/rx
completions -- and uses a separate MSI-X message for each one.

Pretty soon I'm going to want to split the tx/rx event queue into one
event queue per CPU, so that applications can have their completions
events delivered to the same CPU where they're running, so I'll want
to bump up the number of MSI-X entries I request.

I'm sure other MSI-X capable devices can slice and dice things differently.

 - R.