Re: [PATCH V2 3/5] PCI: xilinx: Modifying AXI PCIe Host Bridge driver to work on both Zynq and Microblaze

[Arnd Bergmann <arnd@arndb.de>]

On Tuesday 26 January 2016 10:59:12 Michal Simek wrote:
> >> diff --git a/drivers/pci/host/pcie-xilinx.c b/drivers/pci/host/pcie-xilinx.c
> >> index 3e3757f..1981948 100644
> >> --- a/drivers/pci/host/pcie-xilinx.c
> >> +++ b/drivers/pci/host/pcie-xilinx.c
> >> @@ -92,7 +92,12 @@
> >>  #define ECAM_DEV_NUM_SHIFT          12
> >>  
> >>  /* Number of MSI IRQs */
> >> -#define XILINX_NUM_MSI_IRQS         128
> >> +#define XILINX_NUM_MSI_IRQS 128
> >> +#ifdef CONFIG_ARM
> >> +#define TOT_NR_IRQS         XILINX_NUM_MSI_IRQS
> >> +#else
> >> +#define TOT_NR_IRQS         (NR_IRQS + XILINX_NUM_MSI_IRQS)
> >> +#endif
> > 
> > Something looks wrong here in the microblaze variant. What does NR_IRQS
> > have to do with it?
> 
> Arnd: What was the story regarding NR_IRQS?
> I remember some discussion about it but just forget.
> 
> Default value in include/asm-generic/irq.h is 64.
> Current value is 32 because microblaze primary interrupt controller is
> axi_intc core which has up to 32 input lines.

The value in asm-generic is completely arbitrary, it's just something
that happens to work for a number of the simpler architectures.

Traditionally, there is a a fixed NR_IRQS which defines the maximum
number of interrupts that can be used, and each irqchip has a fixed
start offset below that number.

On modern systems, you have CONFIG_SPARSE_IRQ, which lets an irqchip
allocate its own interrupts, without an upper limit. This is more
flexible and avoids preallocating space for all irq_desc instances,
so it saves memory.

This code however doesn't do either of the two on microblaze:

+       irq = pos;
+#ifdef CONFIG_MICROBLAZE
+#define TOT_NR_IRQS            (NR_IRQS + XILINX_NUM_MSI_IRQS)
+       irq = XILINX_NUM_MSI_IRQS + pos;
+#endif
+       if (irq < TOT_NR_IRQS)
                set_bit(pos, msi_irq_in_use);

So you define XILINX_NUM_MSI_IRQS to mean the number of interrupts
that the xilinx_pcie_port can handle itself, but then pick a number
outside of this range by making the hwirq number something between
XILINX_NUM_MSI_IRQS and (2*XILINX_NUM_MSI_IRQS - 1), and in the
end compare it to (NR_IRQS + XILINX_NUM_MSI_IRQS), which is the
sum of two things that are not related: the total number of interrupts
including the MSIs and the number of MSI.

	Arnd