[openib-general] svn patches.

Wed Apr 28 15:12:28 PDT 2004

    Greg> The proper pci_register_driver() interface is what you need
    Greg> to do.  That works just fine for 2.4 and 2.6, and will
    Greg> handle all of the hotplug issues properly.

Right, but the HCA HW driver has to have an interface to give the
device pointer to upper-level protocols like SRP.

    Roland> In fact that's one of the things that I'm trying to figure
    Roland> out how to do in a way portable between 2.4 and 2.6.  For
    Roland> 2.6 we clearly want to just use a struct device * (since
    Roland> there are devices we care about on things like a PPC4xx
    Roland> EPB rather than PCI), but the device model and generic DMA
    Roland> mapping API doesn't exist on 2.4.

    Greg> And it will not ever get backported either (for good reason,
    Greg> 2.4 development is over.)  So, I was just lucky in finding
    Greg> the main reason why 2.4 support is going to be almost
    Greg> impossible?  :)

You picked a good one :) Now that I think about it though, for 2.4 I
think we can get away with just giving a PCI device pointer but
calling it a struct device *.  Then we just have to add trivial DMA
API wrappers around PCI DMA API functions in our ib_kcompat module.

    Roland> Hotplug is an even bigger can of worms.  I don't know what
    Roland> the heck we do for stuff like userspace processes that
    Roland> have registered memory with an HCA that is no longer
    Roland> attached to the system.

    Greg> If you use pci_register_driver() everything will work just
    Greg> fine.  The driver has to tell userspace that it's resources
    Greg> are now gone just before the card goes away, just like it
    Greg> happens if you unload the module.

But the card can now go away with a nonzero reference count...

With typical hotplug PCI hardware, what happens to a userspace process
that writes to a PCI devices mmio region after the device is no longer
there?  Is the write silently discarded or does the process get some
sort of fault?

    Greg> Also, why is the memory model backwards from the traditional
    Greg> Unix mmap model?  (kernel gives out the memory to the user,
    Greg> the user never gives a memory chunk to the kernel.)  Is
    Greg> there some IB reason for this that I'm missing?

Unfortunately this is deeply embedded in the IB spec.  A userspace
consumer can call the register memory verb on any virtual address
space it owns.  It does make sense too: otherwise the IB driver ends
up reinventing interfaces to give user processes hugetlb pages, etc.

 - Roland

-- 
To unsubscribe send an email with subject unsubscribe to openib-general at openib.org.
Please contact moderator at openib.org for questions.