Skip to contentSkip to conversation

Strange PCI mmap problem on MPC8555

DISCUSSION willy jacobs 1 participant 1 message Jul 13, 2007

lore
Share & tools
lore
#0 lore raw

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?

Signature · 2 lines
willy
Unclassified