On Wed, Feb 1, 2017 at 12:19 PM, Arnd Bergmann [off-list ref] wrote:

On Saturday, January 28, 2017 9:48:36 PM CET Linus Walleij wrote:

quoted

+       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 ?n the slot going and
see if it works the same with just good old vanilla
swizzling.

Yours,
Linus Walleij