Strange PCI mmap problem on MPC8555

willy jacobs willy.jacobs at nl.thalesgroup.com
Fri Jul 13 21:21:40 EST 2007


Setup: ELDK 4.1, U-Boot 1.2.0 and Linux 2.6.20.2 (arch/ppc)

The PPC can communicate (memory mapped) with a FGPA through the PCI bus.
A user program should be able to access (read/write) the FPGA memory areas
directly via mmap() through a small Linux driver.

In the driver initialisation bar0 (1 KByte memory) the PCI physical 
address
is remapped to virtual address space via ioremap().
A memory dump (with expected values) in the kernel driver shows:
             0  1  2  3  4  5  6  7  8  9  A  B  C  D  E  F
0xe1056000 17 31 04 00 02 00 00 00 07 20 07 03 00 00 00 00

The mmap call in driver looks like this:

static int
rvc_mmap (struct file *filp, struct vm_area_struct *vma)
{
...
	if (remap_pfn_range(vma,
			    vma->vm_start,
			    phys_bar0 >> PAGE_SHIFT,
			    vma->vm_end - vma->vm_start,
			    vma->vm_page_prot)) {
	    printk(KERN_CRIT "rvc_mmap: remap_page bar0 failed\n");
	    return -EAGAIN;
...
}

The user space test program does a mmap() to bar0 and the memory dump
(erroneous values) now shows:

            0  1  2  3  4  5  6  7  8  9  A  B  C  D  E  F
0x30019000 17 31 04 00 17 31 04 00 07 20 07 03 07 20 07 03
                        ^^^^^^^^^^^             ^^^^^^^^^^^
So the first 32-bit word is now exactly the same as the second 32-bit 
word,
the 4th 32-bit word is the same as the 3rd 32-bit word, etc.
Also the same behaviour for bar1.
It looks like a 32-/64 bit problem, but the kernel I use is 32-bits.
Any idea what causes this strange behaviour?
-- 
willy
Unclassified



More information about the Linuxppc-embedded mailing list