Re: Mapping PCI memory to user-space

[Roland Dreier <rdreier@cisco.com>]

 > I am writing a driver to map a PCI board memory space (pcibar2) into a
 > user-space vma via 'mmap'.  What is the relationship between the address
 > returned from ioremap and the type of address needed in the
 > 'io_remap_page_range' or 'remap_pfn_range' functions?  How about the
 > following? (I am developing under RHEL4 and a 2.6.9 kernel)

There is no relationship between the address returned from ioremap and
what you pass into io_remap_page_range().  ioremap gives you a kernel
virtual address for the PCI address you remap.  io_remap_page_range()
creates a userspace mapping in the same way, and you should pass in
the PCI address exactly the same way you pass in the PCI address into
ioremap.  io_remap_pfn_range() takes a PFN ("page frame number"),
which is basically the PCI address you want to map divided by
PAGE_SIZE.  The main reason for using PFNs is that they allow you to
map addresses above 4G even if sizeof long is only 4.

In your code:

 >      dev.pcibar2 = ioremap_nocache(resource,size);
 >      dev.region_start = dev.pcibar2 + offset;     // RAM is at some offset from base

This gives you a kernel mapping that you can use with readl(),
writel() etc to access the PCI memory from the kernel.

To map to userspace, this:

 >   if (io_remap_page_range(vma, phyaddr, vma->vm_start, vsize, vma->vm_page_prot))

should use phyaddr as you have it here:

 >   // phyaddr   = physical address of PCI memory area

This is just wrong:

 >   unsigned long phy            = __pa(dev->region_start + off);

__pa() doesn't work on addresses returned from ioremap.  Just use the
same resource address you passed into ioremap.

 - R.