[OpenWrt-Devel,1/4] PCI: add DT bindings for Cortina Gemini PCI Host Bridge

This adds device tree bindings for the Cortina Systems Gemini PCI
Host Bridge.

Cc: Janos Laube <janos.dev@gmail.com>
Cc: Paulius Zaleckas <paulius.zaleckas@gmail.com>
Cc: Hans Ulli Kroll <ulli.kroll@googlemail.com>
Cc: Florian Fainelli <f.fainelli@gmail.com>
Cc: devicetree@vger.kernel.org
Signed-off-by: Linus Walleij <linus.walleij@linaro.org>
---
This can be merged to the PCI tree whenever it is considered
fine for inclusion.
---
 .../devicetree/bindings/pci/cortina,gemini-pci.txt | 64 ++++++++++++++++++++++
 1 file changed, 64 insertions(+)
 create mode 100644 Documentation/devicetree/bindings/pci/cortina,gemini-pci.txt

On Tue, Jan 31, 2017 at 1:31 AM, Bjorn Helgaas <helgaas@kernel.org> wrote:

>> +- bus-range: set to <0x00 0x00> (only root bus)
>
> Why is this limited to bus 0?  Is everything completely soldered down
> with no possibility of adding or replacing PCI devices? The interrupt-map
> below suggests slots, though.  If there's a slot, we could plug in a card
> with a bridge, which would mean more than just bus 0.

I thought so because it's a NAS box (the typical usecase) but when I
dismantled it and turned it over, wow, there is indeed a MiniPCI
slot!

OK I will augment this to 0x00 0xff.

Yours,
Linus Walleij

On Wed, Feb 1, 2017 at 5:02 PM, Rob Herring <robh@kernel.org> wrote:
> On Sat, Jan 28, 2017 at 09:48:36PM +0100, Linus Walleij wrote:

>> +Example:
>> +
>> +pci@50000000 {
>> +     compatible = "cortina,gemini-pci";
>> +     reg = <0x50000000 0x100>;
>
> Config space is indirectly accessed?

Yes it is a really annoying construction. The first device, the Faraday
roob hub is at that address, the init code uses that to set up the bus.

As soon as the bus is up, we use bus accessors to talk to the root
hub on slot 0 (this manages interrupts etc).

I guess I could even unmap this reg range at that point ...

I guess I could also get the address directly from the IO range
and start poking around. I don't know if that is any better.

Yours,
Linus Walleij

On Wed, Feb 1, 2017 at 12:09 PM, Arnd Bergmann <arnd@arndb.de> wrote:
> On Saturday, January 28, 2017 9:48:36 PM CET Linus Walleij wrote:
>> +       pci_intc: interrupt-controller {
>> +               interrupt-controller;
>> +               #address-cells = <0>;
>> +               #interrupt-cells = <1>;
>> +       };
>
> I think this one needs at least an interrupt-parent property, otherwise
> you would recursively translate the numbers back through the interrupt-map
> property of its parent.

OK I added this, I think it was the cause of the error Hand Uli Kroll
was seeing. I guess I didn't run into it because I was only testing the
device in slot 9 so everything was defaulting to zero.

Yours,
Linus Walleij

On Wed, Feb 1, 2017 at 12:19 PM, Arnd Bergmann <arnd@arndb.de> wrote:
> On Saturday, January 28, 2017 9:48:36 PM CET Linus Walleij wrote:
>> +       interrupt-map-mask = <0xff00 0 0 7>;
>> +       interrupt-map = <0x4800 0 0 1 &pci_intc 0>, /* Slot 9 */
>> +                       <0x4900 0 0 2 &pci_intc 1>,
>> +                       <0x4a00 0 0 3 &pci_intc 2>,
>> +                       <0x4b00 0 0 4 &pci_intc 3>,
>> +                       <0x5000 0 0 1 &pci_intc 0>, /* Slot 10 */
>> +                       <0x5100 0 0 2 &pci_intc 1>,
>> +                       <0x5200 0 0 3 &pci_intc 2>,
>> +                       <0x5300 0 0 4 &pci_intc 3>,
>> +                       <0x5800 0 0 1 &pci_intc 0>, /* Slot 11 */
>> +                       <0x5900 0 0 2 &pci_intc 1>,
>> +                       <0x5a00 0 0 3 &pci_intc 2>,
>> +                       <0x5b00 0 0 4 &pci_intc 3>,
>> +                       <0x6000 0 0 1 &pci_intc 0>, /* Slot 12 */
>> +                       <0x6100 0 0 2 &pci_intc 1>,
>> +                       <0x6200 0 0 3 &pci_intc 2>,
>> +                       <0x6300 0 0 4 &pci_intc 3>;
>>
>
> The mapping looks wrong here, we normally don't list interrupts per function
> so the mask should be 0xf800.

Yup it works that way too, and indeed the USB hub on function
2 was requesting the right pin and everything.

> Note that the interrupt map is board specific, so this should probably
> go in the board.dts file rather than platform.dtsi.

OK moving it down there.

> For this particular board, the interrupt lines appear to have been badly
> configured so all slots use the same interrupt 0 for IntA. IIRC This also
> means you can probably use <0 0 0 7> as the mask and just specify each of
> the four interrupts once. A properly wired board would swizzle the
> interrupts so that each slot has a different IRQ for its IntA line.

They are swizzled, I just have very bad hardware docs. (Only
code.) It turns out when I'm browsing through old board support
code that there is a comment followed by a piece of code like this:

/*
 *      No swizzle on SL2312
 */
static u8 __init sl2312_pci_swizzle(struct pci_dev *dev, u8 *pinp)
{
        return PCI_SLOT(dev->devfn);
}

/*
 * map the specified device/slot/pin to an IRQ.  This works out such
 * that slot 9 pin 1 is INT0, pin 2 is INT1, and slot 10 pin 1 is INT1.
 */
static int __init sl2312_pci_map_irq(struct pci_dev *dev, u8 slot, u8 pin)
{
        int intnr = ((slot  + (pin - 1)) & 3) + 4;  /* the IRQ number
of PCI bridge */

        printk("%s : slot = %d  pin = %d \n",__func__,slot,pin);
    switch (slot)
    {
        case 12:
                if (pin==1)
                {
                        intnr = 3;
                    }
                    else
                    {
                        intnr = 0;
                    }
#ifdef CONFIG_DUAL_PCI
                    return IRQ_PCI_INTD;
#endif
            break;
        case 11:
                    intnr = (2 + (pin - 1)) & 3;
#ifdef CONFIG_DUAL_PCI
                    return IRQ_PCI_INTC;
#endif
            break;
        case 10:
                    intnr = (1 + (pin - 1)) & 3;
#ifdef CONFIG_DUAL_PCI
                    return IRQ_PCI_INTB;
#endif
            break;
        case  9:
                    intnr = (pin - 1) & 3;
            break;
    }
//      if (slot == 10)
//              intnr = (1 + (pin - 1)) & 3;
//      else if (slot == 9)
//              intnr = (pin - 1) & 3;
        return (IRQ_PCI_INTA + intnr);
}

If I understand correctly they say that the IRQs are not swizzled
on the PCI bridge side but on the IRQ handler side.

So if I put it in the device tree like so:

+                         interrupt-map =
+                               <0x4800 0 0 1 &pci_intc 0>, /* Slot 9 */
+                               <0x4800 0 0 2 &pci_intc 1>,
+                               <0x4800 0 0 3 &pci_intc 2>,
+                               <0x4800 0 0 4 &pci_intc 3>,
+                               <0x5000 0 0 1 &pci_intc 1>, /* Slot 10 */
+                               <0x5000 0 0 2 &pci_intc 2>,
+                               <0x5000 0 0 3 &pci_intc 3>,
+                               <0x5000 0 0 4 &pci_intc 0>,
+                               <0x5800 0 0 1 &pci_intc 2>, /* Slot 11 */
+                               <0x5800 0 0 2 &pci_intc 3>,
+                               <0x5800 0 0 3 &pci_intc 0>,
+                               <0x5800 0 0 4 &pci_intc 1>,
+                               <0x6000 0 0 1 &pci_intc 3>, /* Slot 12 */
+                               <0x6000 0 0 2 &pci_intc 0>,
+                               <0x6000 0 0 3 &pci_intc 1>,
+                               <0x6000 0 0 4 &pci_intc 2>;

So the IRQs on the right side are swizzled instead
of the pins being swizzled.

...but that looks a bit insane.

Isn't that exactly the same thing just exposed in some
inverse way?

I'll try to get the RALink MiniPCI I have ín the slot going and
see if it works the same with just good old vanilla
swizzling.

Yours,
Linus Walleij