Re: [PATCH v2 2/6] irqchip/riscv-intc: Create domain using named fwnode

[Jessica Clarke]

On 19 Feb 2022, at 09:32, Marc Zyngier <maz@xxxxxxxxxx> wrote:
> 
> On 2022-02-19 03:38, Anup Patel wrote:
>> On Thu, Feb 17, 2022 at 8:42 PM Marc Zyngier <maz@xxxxxxxxxx> wrote:
>>> On 2022-01-28 05:25, Anup Patel wrote:
>>> > We should create INTC domain using a synthetic fwnode which will allow
>>> > drivers (such as RISC-V SBI IPI driver, RISC-V timer driver, RISC-V
>>> > PMU driver, etc) not having dedicated DT/ACPI node to directly create
>>> > interrupt mapping for standard local interrupt numbers defined by the
>>> > RISC-V privileged specification.
>>> >
>>> > Signed-off-by: Anup Patel <apatel@xxxxxxxxxxxxxxxx>
>>> > ---
>>> >  arch/riscv/include/asm/irq.h      |  2 ++
>>> >  arch/riscv/kernel/irq.c           | 13 +++++++++++++
>>> >  drivers/clocksource/timer-clint.c | 13 +++++++------
>>> >  drivers/clocksource/timer-riscv.c | 11 ++---------
>>> >  drivers/irqchip/irq-riscv-intc.c  | 12 ++++++++++--
>>> >  drivers/irqchip/irq-sifive-plic.c | 19 +++++++++++--------
>>> >  6 files changed, 45 insertions(+), 25 deletions(-)
>>> >
>>> > diff --git a/arch/riscv/include/asm/irq.h
>>> > b/arch/riscv/include/asm/irq.h
>>> > index e4c435509983..f85ebaf07505 100644
>>> > --- a/arch/riscv/include/asm/irq.h
>>> > +++ b/arch/riscv/include/asm/irq.h
>>> > @@ -12,6 +12,8 @@
>>> >
>>> >  #include <asm-generic/irq.h>
>>> >
>>> > +extern struct fwnode_handle *riscv_intc_fwnode(void);
>>> > +
>>> >  extern void __init init_IRQ(void);
>>> >
>>> >  #endif /* _ASM_RISCV_IRQ_H */
>>> > diff --git a/arch/riscv/kernel/irq.c b/arch/riscv/kernel/irq.c
>>> > index 7207fa08d78f..f2fed78ab659 100644
>>> > --- a/arch/riscv/kernel/irq.c
>>> > +++ b/arch/riscv/kernel/irq.c
>>> > @@ -7,9 +7,22 @@
>>> >
>>> >  #include <linux/interrupt.h>
>>> >  #include <linux/irqchip.h>
>>> > +#include <linux/irqdomain.h>
>>> > +#include <linux/module.h>
>>> >  #include <linux/seq_file.h>
>>> >  #include <asm/smp.h>
>>> >
>>> > +static struct fwnode_handle *intc_fwnode;
>>> > +
>>> > +struct fwnode_handle *riscv_intc_fwnode(void)
>>> > +{
>>> > +     if (!intc_fwnode)
>>> > +             intc_fwnode = irq_domain_alloc_named_fwnode("RISCV-INTC");
>>> > +
>>> > +     return intc_fwnode;
>>> > +}
>>> > +EXPORT_SYMBOL_GPL(riscv_intc_fwnode);
>>> Why is this created outside of the root interrupt controller driver?
>>> Furthermore, why do you need to create a new fwnode the first place?
>>> As far as I can tell, the INTC does have a node, and what you don't
>>> have is the firmware linkage between PMU (an others) and the INTC.
>> Fair enough, I will update this patch to not create a synthetic fwnode.
>> The issue is not with INTC driver. We have other drivers and places
>> (such as SBI IPI driver, SBI PMU driver, and KVM RISC-V AIA support)
>> where we don't have a way to locate INTC fwnode.
> 
> And that's exactly what I am talking about: The INTC is OK (sort of),
> but the firmware is too crap for words, and isn't even able to expose
> where the various endpoints route their interrupts to.
> 
> Yes, this is probably fine today because you can describe the topology
> of RISC-V systems on the surface of a post stamp. Once you get to the
> complexity of a server-grade SoC (or worse, a mobile phone style SoC),
> this *implicit topology* stuff doesn't fly, because there is no guarantee
> that all endpoints will always all point to the same controller.
> 
>>> what you should have instead is something like:
>>> static struct fwnode_handle *(*__get_root_intc_node)(void);
>>> struct fwnode_handle *riscv_get_root_intc_hwnode(void)
>>> {
>>>         if (__get_root_intc_node)
>>>                 return __get_root_intc_node();
>>>         return NULL;
>>> }
>>> and the corresponding registration interface.
>> Thanks, I will follow this suggestion. This is a much better approach
>> and it will avoid touching existing drivers.
>>> But either way, something breaks: the INTC has one node per CPU, and
>>> expect one irqdomain per CPU. Having a single fwnode completely breaks
>>> the INTC driver (and probably the irqdomain list, as we don't check for
>>> duplicate entries).
>>> > diff --git a/drivers/irqchip/irq-riscv-intc.c
>>> > b/drivers/irqchip/irq-riscv-intc.c
>>> > index b65bd8878d4f..26ed62c11768 100644
>>> > --- a/drivers/irqchip/irq-riscv-intc.c
>>> > +++ b/drivers/irqchip/irq-riscv-intc.c
>>> > @@ -112,8 +112,16 @@ static int __init riscv_intc_init(struct
>>> > device_node *node,
>>> >       if (riscv_hartid_to_cpuid(hartid) != smp_processor_id())
>>> >               return 0;
>>> >
>>> > -     intc_domain = irq_domain_add_linear(node, BITS_PER_LONG,
>>> > -                                         &riscv_intc_domain_ops, NULL);
>>> > +     /*
>>> > +      * Create INTC domain using a synthetic fwnode which will allow
>>> > +      * drivers (such as RISC-V SBI IPI driver, RISC-V timer driver,
>>> > +      * RISC-V PMU driver, etc) not having dedicated DT/ACPI node to
>>> > +      * directly create interrupt mapping for standard local interrupt
>>> > +      * numbers defined by the RISC-V privileged specification.
>>> > +      */
>>> > +     intc_domain = irq_domain_create_linear(riscv_intc_fwnode(),
>>> > +                                            BITS_PER_LONG,
>>> > +                                            &riscv_intc_domain_ops, NULL);
>>> This is what I'm talking about. It is simply broken. So either you don't
>>> need a per-CPU node (and the DT was bad the first place), or you
>>> absolutely need
>>> one (and the whole 'well-known/default domain' doesn't work at all).
>>> Either way, this patch is plain wrong.
>> Okay, I will update this patch with the new approach which you suggested.
> 
> But how do you plan to work around the fact that everything is currently
> build around having a node (and an irqdomain) per CPU? The PLIC, for example,
> clearly has one parent per CPU, not one global parent.
> 
> I'm sure there was a good reason for this, and I suspect merging the domains
> will simply end up breaking things.

On the contrary, the drivers rely on the controller being the same
across all harts, with riscv_intc_init skipping initialisation for all
but the boot hart’s controller. The bindings are a complete pain to
deal with as a result, what you *want* is like you have in the Arm
world where there is just one interrupt controller in the device tree
with some of the interrupts per-processor, but instead we have this
overengineered nuisance. The only reason there are per-hart interrupt
controllers is because that’s how the contexts for the CLINT/PLIC are
specified, but that really should have been done another way rather
than abusing the interrupts-extended property for that. In the FreeBSD
world we’ve been totally ignoring the device tree nodes for the local
interrupt controllers but for my AIA and ACLINT branch I started a few
months ago (though ACLINT's now been completely screwed up by RVI
politics, things have been renamed and split up differently in the past
few days and software interrupts de-prioritised with no current path to
ratification, so that was a waste of my time) I just hang the driver
off the boot hart’s node and leave all the others as totally ignored
and a waste of space other than to figure out the contexts for the PLIC
etc.

TL;DR yes the bindings are awful, no there’s no issue with merging the
domains.

Jess