On Thu, Jul 26, 2018 at 1:57 PM, Christoph Hellwig [off-list ref] wrote:

On Thu, Jul 26, 2018 at 09:08:00AM +0530, Anup Patel wrote:

quoted

Actually, RISCV HLIC and PLIC are very similar to RPi2 and RPi3 SOCs.

On RPi2 and RPi3, we have per-CPU BCM2836 local intc and the global
interrupts are managed using BCM2835 intc. You should certainly have
a look a this drivers because these very simple compared to GICv2 and
GICv3 drivers.

Yes, using that model makes writing the per-cpu irq controller driver
trivial.  But retrofitting it into the device tree, where the existing
bootloader (bbl) assumes the old DT layout is a giant pain in the neck.

This can also be taken care in HLIC driver probe function with something
like below:

if (parent)
return 0;

if (of_is_compatible(parent, "riscv,cpu")) {
/*
* Legacy DT binding so we have HLIC DT node
* under each CPU DT node. To provide backwared
* compatiblity we go forward for only one HLIC
* DT node
*/
if (atomic_inc_return(&hlic_lottery) > 1)
return 0;
}

In PLIC driver probe, register nested IRQ handler for only first two
entries of interrupts-extended because it is HLIC IRQs are per-CPU. We
can happily ignore other entries in interrupts-extended of Legacy DTS.

At the same time I'm still not conveninced RISC-V really needs a full
irqchip driver for the per-cpu interrupt 'controller' really is nothing
but 1 and a half architectural control registers:

  - the scause register that contains the reason for an exception
    (any exception including syscalls and page faults) for the entry
    into supervisor mode.  This includes a bit to indicate interrupts,
    and then logical interrupt reason, out of which only three are
    interesting for supervisor mode (timer, software, external)
  - the sie register that allows to to enable/disable each of the above
    causes individually

Biggest plus is the ability show stats for per-CPU interrupts via
"cat /proc/interrupts" (just like other architectures).

Currently, we have only three per-CPU interrupts (timer, software,
and external) but in-future people might come-up with interesting
devices which might have per-CPU interrupts.

So after burning out on DT hacking (never mind retrofitting that into
actual shipping SOCs vs just qemu) I'm going to try a version that
doesn't add an irqchip for this but just handles it hardcoded in
RISC-V do_IRQ.  I'll still keep the irqchip for the PLIC, which while
specificed in the RISC-V spec isn't architectural but an actual
periphal.

I believe it possible to have RICV HLIC driver which maintains
backward compatibility with legacy DTS. I haven't tried above
approach myself on QEMU so I will let you decide.

Regards,
Anup