On Wed, Sep 3, 2025 at 7:44 AM Jan Palus <jpalus@xxxxxxxxxxxx> wrote:
>
> On 02.09.2025 17:13, Klaus Kudielka wrote:
> > The blamed commit simplifies code, by using the for_each_of_range()
> > iterator. But it results in no pci devices being detected anymore on
> > Turris Omnia (and probably other mvebu targets).
> >
> > Analysis:
> >
> > To determine range.flags, of_pci_range_parser_one() uses bus->get_flags(),
> > which resolves to of_bus_pci_get_flags(). That function already returns an
> > IORESOURCE bit field, and NOT the original flags from the "ranges"
> > resource.
> >
> > Then mvebu_get_tgt_attr() attempts the very same conversion again.
> > But this is a misinterpretation of range.flags.
> >
> > Remove the misinterpretation of range.flags in mvebu_get_tgt_addr(),
> > to restore the intended behavior.
> >
> > Signed-off-by: Klaus Kudielka <klaus.kudielka@xxxxxxxxx>
> > Fixes: 5da3d94a23c6 ("PCI: mvebu: Use for_each_of_range() iterator for parsing "ranges"")
> > Reported-by: Bjorn Helgaas <helgaas@xxxxxxxxxx>
> > Closes: https://lore.kernel.org/r/20250820184603.GA633069@bhelgaas/
> > Reported-by: Jan Palus <jpalus@xxxxxxxxxxxx>
> > Closes: https://bugzilla.kernel.org/show_bug.cgi?id=220479
> > ---
> > drivers/pci/controller/pci-mvebu.c | 14 ++------------
> > 1 file changed, 2 insertions(+), 12 deletions(-)
> >
> > diff --git a/drivers/pci/controller/pci-mvebu.c b/drivers/pci/controller/pci-mvebu.c
> > index 755651f338..4e2e1fa197 100644
> > --- a/drivers/pci/controller/pci-mvebu.c
> > +++ b/drivers/pci/controller/pci-mvebu.c
> > @@ -1168,9 +1168,6 @@ static void __iomem *mvebu_pcie_map_registers(struct platform_device *pdev,
> > return devm_ioremap_resource(&pdev->dev, &port->regs);
> > }
> >
> > -#define DT_FLAGS_TO_TYPE(flags) (((flags) >> 24) & 0x03)
> > -#define DT_TYPE_IO 0x1
> > -#define DT_TYPE_MEM32 0x2
> > #define DT_CPUADDR_TO_TARGET(cpuaddr) (((cpuaddr) >> 56) & 0xFF)
> > #define DT_CPUADDR_TO_ATTR(cpuaddr) (((cpuaddr) >> 48) & 0xFF)
> >
> > @@ -1189,17 +1186,10 @@ static int mvebu_get_tgt_attr(struct device_node *np, int devfn,
> > return -EINVAL;
> >
> > for_each_of_range(&parser, &range) {
> > - unsigned long rtype;
> > u32 slot = upper_32_bits(range.bus_addr);
> >
> > - if (DT_FLAGS_TO_TYPE(range.flags) == DT_TYPE_IO)
> > - rtype = IORESOURCE_IO;
> > - else if (DT_FLAGS_TO_TYPE(range.flags) == DT_TYPE_MEM32)
> > - rtype = IORESOURCE_MEM;
> > - else
> > - continue;
> > -
> > - if (slot == PCI_SLOT(devfn) && type == rtype) {
> > + if (slot == PCI_SLOT(devfn) &&
> > + type == (range.flags & IORESOURCE_TYPE_BITS)) {
> > *tgt = DT_CPUADDR_TO_TARGET(range.cpu_addr);
> > *attr = DT_CPUADDR_TO_ATTR(range.cpu_addr);
>
> Thanks for the patch Klaus! While it does improve situation we're not
> quite there yet. It appears that what used to be stored in `cpuaddr` var
> is also very different from `range.cpu_addr` value so the results
> in both `*tgt` and `*attr` are both wrong.
>
> Previously `cpuaddr` had a value like ie 0x8e8000000000000 or
> 0x4d0000000000000. Now `range.cpu_addr` is always 0xffffffffffffffff.
> Luckily what used to be stored in `cpuaddr`:
~0 is OF_BAD_ADDR which means we couldn't translate the address. Seems
it is not needed here, but it should work. Can you define DEBUG in
drivers/of/address.c and post the log?
Rob
