[Ltib] Re: 8544ds pcie mmap fails

[Todd Blackmon]

I never got a response, but found the answer by sheer dumb luck and persistence.  So I'm sending it for documentation purposes.  My problem was apparently that I needed to add the following line before the call to remap_pfn_range.

vma->vm_page_prot = pgprot_noncached(vma->vm_page_prot);

Apparently it was doing something odd due to the cache operations.  The x86 port didn't have this issue, though.

Todd

On Wed, Jul 15, 2009 at 1:38 PM, Todd Blackmon <address@hidden> wrote:

I think I was completely confused on this before.  I understand it better now, but I still have an issue.
I have to communicate with a PCIe endpoint on an FPGA card using the mpc8544ds.  The example code that I have uses the mmap system call to map the BAR0 addresses into the user space.  The mmap succeeds, but when the user code attempts to access the space (any size and it is page/byte/word aligned), the program maxes out the CPU, and can be stopped via ctrl-c.

The driver mmap handler function uses remap_pfn_range, which returns 0 so it seems to work.  The call is:
    sysErr = remap_pfn_range(vma,
                  (unsigned long)vma->vm_start,
                  (phys_start >> PAGE_SHIFT),      // phys_start = BAR0 address + (vma->vm_pgoff << PAGE_SHIFT)
                  mapSize,                                         // mapSize = vma->vm_end - vma->vm_start
                  vma->vm_page_prot);

This code works on ubuntu on an x86 architecture, but not on this powerpc arch.  Is there a fundamental reason (architectural difference) that this would happen?  I haven't been able to find another PCI/PCIe driver that uses this mmaping method.  Is that not the *right* way to do it?  Are there examples of the *right* way?

Thanks for all your help,
Todd Blackmon