On Thu, Jan 10, 2013 at 12:24:17PM -0700, Jason Gunthorpe wrote: > On Thu, Jan 10, 2013 at 08:03:27PM +0100, Thierry Reding wrote: > > > > > You'd piece a mapping together, each bus requires 16 64k mappings, a > > > > simple 2d array of busnr*16 of pointers would do the trick. A more > > > > clever solution would be to allocate contiguous virtual memory and > > > > split that up.. > > > > Oh, I see. I'm not very familiar with the internals of remapping, so > > > I'll need to do some more reading. Thanks for the hints. > > > > I forgot to ask. What's the advantage of having a contiguous virtual > > memory area and splitting it up versus remapping each chunk separately? > > Not alot, really, but it saves you from the pointer array and > associated overhead. IIRC it is fairly easy to do in the kernel. I've been investigating this a bit, and one problem is that it will prevent the driver from ever building as a module because the necessary functions aren't exported and I'm not sure exporting them would be acceptable. Currently PCI host controller drivers with MSI support can't be built as modules because the MSI infrastructure requires it, but I briefly discussed this with Bjorn at some point and it should be easy to remove that requirement. > Arnd's version is good too, but you would be restricted to aligned > powers of two for the bus number range in the DT, which is probably > not that big a deal either? Stephen suggested on IRC that we could try to keep a bit of dynamicity in the allocation scheme if we create the bus mapping when the first device on the bus is probed and discard the mapping if no devices are found. Sounds like a good plan to me. Does anybody see any potential pitfalls? Thierry
Attachment:
pgpxWDe8FXNc0.pgp
Description: PGP signature
