Instead of allocating PCI devices I/O port bus addresses from the
000xxxxx I/O port range as intended, due to a bus versus physical
address mixup, the Kirkwood PCIe handling code inadvertently
allocated I/O port bus addresses from the f20xxxxx address range
(which is the physical address range of the PCIe I/O mapping window),
but then direct all I/O port accesses to bus addresses 000xxxxx,
which would then not be decoded at all.
Fix this by setting the base address of the PCIe I/O space struct
resource to KIRKWOOD_PCIE_IO_BUS_BASE instead of the incorrect
KIRKWOOD_PCIE_IO_PHYS_BASE, and fix up __io() to expect addresses
offsetted by the former instead of the latter.
(The suggested fix of directing I/O port accesses from the host to
bus addresses f20xxxxx instead has the problem that assigning full
32bit I/O port bus addresses (f20xxxxx) doesn't work on all PCI
devices, as not all PCI devices implement full 32 bit BAR registers
for I/O ports. We should really try to allocate I/O port bus
addresses that fit in 16 bits.)
Signed-off-by: Lennert Buytenhek <buytenh@marvell.com>
Signed-off-by: Nicolas Pitre <nico@marvell.com>
static inline void __iomem *__io(unsigned long addr)
{
- return (void __iomem *)((addr - KIRKWOOD_PCIE_IO_PHYS_BASE)
+ return (void __iomem *)((addr - KIRKWOOD_PCIE_IO_BUS_BASE)
+ KIRKWOOD_PCIE_IO_VIRT_BASE);
}
*/
res[0].name = "PCIe I/O Space";
res[0].flags = IORESOURCE_IO;
- res[0].start = KIRKWOOD_PCIE_IO_PHYS_BASE;
+ res[0].start = KIRKWOOD_PCIE_IO_BUS_BASE;
res[0].end = res[0].start + KIRKWOOD_PCIE_IO_SIZE - 1;
if (request_resource(&ioport_resource, &res[0]))
panic("Request PCIe IO resource failed\n");
git://ftp.safe.ca / safe / jmp / linux-2.6 / commitdiff