Message-ID: <91b50d9a362b3f108df55cd766739b79637e4797.camel@svanheule.net>
Date:   Thu, 23 Dec 2021 20:17:28 +0100
From:   Sander Vanheule <sander@...nheule.net>
To:     Marc Zyngier <maz@...nel.org>
Cc:     Thomas Gleixner <tglx@...utronix.de>, devicetree@...r.kernel.org,
        Rob Herring <robh+dt@...nel.org>,
        Birger Koblitz <mail@...ger-koblitz.de>,
        Bert Vermeulen <bert@...t.com>,
        John Crispin <john@...ozen.org>, linux-kernel@...r.kernel.org
Subject: Re: [RFC PATCH v1 2/4] irqchip: realtek-rtl: use per-parent irq
 handling

Hi Mark,

Thanks for your feedback.

On Thu, 2021-12-23 at 17:57 +0000, Marc Zyngier wrote:
> Hi Sander,
> 
> nit: please check the way the irqchip patches have their title
> formatted, and make sure you follow these rules. In this case, it
> should read:
> 
> irqchip/realtek-rtl: Use per-parent...

I'll update the titles.

> On Thu, 23 Dec 2021 12:08:32 +0000,
> Sander Vanheule <sander@...nheule.net> wrote:
> > 
> > The interrupt router controller is used to route 32 SoC interrupts to up
> > to 6 MIPS CPU interrupts. This means that the SoC interrupts inherit the
> > priority of to the target CPU interrupt.
> > 
> > Currently the driver handles all SoC interrupts equally, independent of
> > which CPU interrupt it is routed to. The use of __ffs actually gives
> > higher priority to lower IRQ lines, effectively bypassing the CPU
> > interrupt priority.
> > 
> > Additionally, this indiscriminate handling of SoC interrupts masked
> > another issue. There is an actually an offset between routing values
> > (1..6) and CPU interrupts (2..7), but the current mapping makes no
> > distinction between these two values. This issue was also hidden during
> > testing, because an interrupt mapping was used where for each required
> > interrupt another (unused) routing was configured, with an offset of +1.
> > 
> > Rework the driver to use a separate handler for each used CPU interrupt,
> > and use the correct routing values. Instead of assuming that the parent
> > interrupt controller is the MIPS CPU interrupt controller
> > ("mti,cpu-interrupt-controller"), this is now checked explicitly to
> > correctly handle the timer interrupt.
> > 
> > Fixes: 9f3a0f34b84a ("irqchip: Add support for Realtek RTL838x/RTL839x interrupt
> > controller")
> > Signed-off-by: Sander Vanheule <sander@...nheule.net>
> > ---
> > 
> > This patch makes a few changes at the same time, and introduces the
> > *_irr functions, which aren't strictly required. This allows the last
> > patch to be a bit smaller, and seeks to add some clarity to the code.
> > 
> > Please let me know if this should be split into separate patches with
> > more incremental changes (in addition to other likely comments).
> > ---
> >  drivers/irqchip/irq-realtek-rtl.c | 153 +++++++++++++++++++++---------
> >  1 file changed, 108 insertions(+), 45 deletions(-)
> > 
> > diff --git a/drivers/irqchip/irq-realtek-rtl.c b/drivers/irqchip/irq-realtek-rtl.c
> > index d6788dd93c7b..71366f1cf721 100644
> > --- a/drivers/irqchip/irq-realtek-rtl.c
> > +++ b/drivers/irqchip/irq-realtek-rtl.c
> > @@ -7,6 +7,7 @@
> >  
> >  #include <linux/of_irq.h>
> >  #include <linux/irqchip.h>
> > +#include <linux/interrupt.h>
> >  #include <linux/spinlock.h>
> >  #include <linux/of_address.h>
> >  #include <linux/irqchip/chained_irq.h>
> > @@ -21,10 +22,43 @@
> >  #define RTL_ICTL_IRR2          0x10
> >  #define RTL_ICTL_IRR3          0x14
> >  
> > -#define REG(x)         (realtek_ictl_base + x)
> > +#define RTL_ICTL_NUM_PRIO      6
> > +
> > +#define REG(x)                 (realtek_ictl_base + x)
> 
> Spurious change?

The indentation didn't match with the other defines, but I can leave out this change here.


> > 
> >  static DEFINE_RAW_SPINLOCK(irq_lock);
> >  static void __iomem *realtek_ictl_base;
> > +static struct irq_domain *realtek_ictl_domain;
> > +
> > +struct realtek_ictl_priority {
> > +       unsigned int routing_value;
> > +       u32 child_mask;
> > +};
> > +
> > +static struct realtek_ictl_priority priorities[RTL_ICTL_NUM_PRIO];
> > +
> > +/*
> > + * IRR0-IRR3 store 4 bits per interrupt, but Realtek uses inverted
> > + * numbering, placing IRQ 31 in the first four bits.
> > + */
> > +#define IRR_OFFSET(idx)                (4 * (3 - (idx * 4) / 32))
> > +#define IRR_SHIFT(idx)         ((idx * 4) % 32)
> > +
> > +static inline u32 read_irr(void __iomem *irr0, int idx)
> > +{
> > +       return (readl(irr0 + IRR_OFFSET(idx)) >> IRR_SHIFT(idx)) & 0xf;
> > +}
> > +
> > +static inline void write_irr(void __iomem *irr0, int idx, u32 value)
> > +{
> > +       unsigned int offset = IRR_OFFSET(idx);
> > +       unsigned int shift = IRR_SHIFT(idx);
> > +       u32 irr;
> > +
> > +       irr = readl(irr0 + offset) & ~(0xf << shift);
> > +       irr |= (value & 0xf) << shift;
> > +       writel(irr, irr0 + offset);
> 
> Are you always in a situation where this doesn't need any locking?

These are currently only used on initialisation, so that's before any interrupts should be
active.


> > +}
> >  
> >  static void realtek_ictl_unmask_irq(struct irq_data *i)
> >  {
> > @@ -72,43 +106,67 @@ static const struct irq_domain_ops irq_domain_ops = {
> >         .xlate = irq_domain_xlate_onecell,
> >  };
> >  
> > -static void realtek_irq_dispatch(struct irq_desc *desc)
> > +static irqreturn_t realtek_irq_dispatch(int irq, void *devid)
> 
> No, that's definitely not on. Interrupt handlers are not chained
> handlers. They have different guarantees, and they aren't
> interchangeable. It is only in limited circumstances that you can do
> this change (threaded interrupts).
> 
> >  {
> > -       struct irq_chip *chip = irq_desc_get_chip(desc);
> > -       struct irq_domain *domain;
> > -       unsigned int pending;
> > -
> > -       chained_irq_enter(chip, desc);
> > -       pending = readl(REG(RTL_ICTL_GIMR)) & readl(REG(RTL_ICTL_GISR));
> > -       if (unlikely(!pending)) {
> > -               spurious_interrupt();
> > -               goto out;
> > +       struct realtek_ictl_priority *priority = devid;
> 
> So this is *why* you made this change. We have per-interrupt data that
> you can use, and get rid of this horrible hack.

Aha, I'll need to have another look then. I'll make sure to switch back to chained
handlers for the next version.


> > +       unsigned long pending;
> > +       int soc_irq;
> > +       int ret = 0;
> > +
> > +       pending = readl(REG(RTL_ICTL_GIMR)) & readl(REG(RTL_ICTL_GISR))
> > +               & priority->child_mask;
> > +
> > +       for_each_set_bit(soc_irq, &pending, BITS_PER_LONG) {
> > +               generic_handle_domain_irq(realtek_ictl_domain, soc_irq);
> > +               ret = 1;
> >         }
> > -       domain = irq_desc_get_handler_data(desc);
> > -       generic_handle_domain_irq(domain, __ffs(pending));
> >  
> > -out:
> > -       chained_irq_exit(chip, desc);
> > +       return IRQ_RETVAL(ret);
> >  }
> >  
> > -/*
> > - * SoC interrupts are cascaded to MIPS CPU interrupts according to the
> > - * interrupt-map in the device tree. Each SoC interrupt gets 4 bits for
> > - * the CPU interrupt in an Interrupt Routing Register. Max 32 SoC interrupts
> > - * thus go into 4 IRRs.
> > - */
> > -static int __init map_interrupts(struct device_node *node, struct irq_domain *domain)
> > +static void __init set_routing(struct realtek_ictl_priority *priority, unsigned int
> > soc_int)
> >  {
> > +       unsigned int priority_old;
> > +
> > +       priority_old = read_irr(REG(RTL_ICTL_IRR0), soc_int);
> > +       if (priority_old) {
> > +               pr_warn("int %d already routed to %d, not updating\n", soc_int,
> > priority_old);
> > +               return;
> > +       }
> > +
> > +       priority->child_mask |= BIT(soc_int);
> > +       write_irr(REG(RTL_ICTL_IRR0), soc_int, priority->routing_value);
> > +}
> > +
> > +static int __init setup_parent_interrupt(struct realtek_ictl_priority *prio_ctl, int
> > parent)
> > +{
> > +       struct device_node *parent_node;
> > +       struct irq_data *irqd;
> > +       unsigned int flags;
> > +       int parent_hwirq;
> > +
> > +       irqd = irq_get_irq_data(parent);
> > +       if (!irqd)
> > +               return -ENOENT;
> > +
> > +       parent_node = to_of_node(irqd->domain->fwnode);
> > +       parent_hwirq = irqd_to_hwirq(irqd);
> > +
> > +       flags = IRQF_PERCPU | IRQF_SHARED;
> > +       if (of_device_is_compatible(parent_node, "mti,cpu-interrupt-controller")
> > +               && parent_hwirq == 7)
> > +               flags |= IRQF_TIMER;
> > +
> > +       return request_irq(parent, realtek_irq_dispatch, flags, "rtl-intc", prio_ctl);
> 
> This really is mixing two different things. Why aren't the flags on
> the actual endpoint interrupt? This really is the business of the
> driver requesting the interrupt, and not the irqchip.

I needed to set the flags, because otherwise the CEVT-R4K timer couldn't use the timer
interrupt any more. But I assume that's a side effect of using an interrupt handler
instead of a chained handler, and this will disappear when I use the correct handler type.


> > +}
> > +
> > +static int __init map_interrupts(struct device_node *node)
> > +{
> > +       struct realtek_ictl_priority *prio_ctl;
> >         struct device_node *cpu_ictl;
> >         const __be32 *imap;
> > -       u32 imaplen, soc_int, cpu_int, tmp, regs[4];
> > -       int ret, i, irr_regs[] = {
> > -               RTL_ICTL_IRR3,
> > -               RTL_ICTL_IRR2,
> > -               RTL_ICTL_IRR1,
> > -               RTL_ICTL_IRR0,
> > -       };
> > -       u8 mips_irqs_set;
> > +       u32 imaplen, soc_int, priority, tmp;
> > +       int ret, i;
> >  
> >         ret = of_property_read_u32(node, "#address-cells", &tmp);
> >         if (ret || tmp)
> > @@ -118,8 +176,6 @@ static int __init map_interrupts(struct device_node *node, struct
> > irq_domain *do
> >         if (!imap || imaplen % 3)
> >                 return -EINVAL;
> >  
> > -       mips_irqs_set = 0;
> > -       memset(regs, 0, sizeof(regs));
> >         for (i = 0; i < imaplen; i += 3 * sizeof(u32)) {
> >                 soc_int = be32_to_cpup(imap);
> >                 if (soc_int > 31)
> > @@ -133,42 +189,49 @@ static int __init map_interrupts(struct device_node *node,
> > struct irq_domain *do
> >                         return -EINVAL;
> >                 of_node_put(cpu_ictl);
> >  
> > -               cpu_int = be32_to_cpup(imap + 2);
> > -               if (cpu_int > 7)
> > +               /* Map priority (1..6) to MIPS CPU interrupt (2..7) */
> > +               priority = be32_to_cpup(imap + 2);
> 
> I don't understand this. As far as the binding describes it, this is
> the target interrupt, and not a priority. Either the binding is wrong,
> and it needs fixing, or this is wrong. What gives?

The "interrupt-map" values that my (and OpenWrt's) DTS-s have could also be amended to
provide correct <soc_int &parent_node parent_hwirq> tuples, instead of the current (wrong)
<soc_int &parent_node routing_value> tuples. That would probably make it possible to limit
this patch only to the per-parent chained handlers.

However, that would require me to keep the assumption that output lines (1..6) are mapped
to CPU interrupts (2..7). This is what I wanted to get rid of by using "interrupt-parent"
+ "realtek,interrupt-routing". Maybe this will be clearer if I can isolate these two
issues in a v2.


> If the binding is really describing a priority, how is the interrupt
> priority actually mapped to the CPU interrupt? Why can't you just
> ignore the DT-provided priority and simply have a flat priority
> scheme, allocating mapping input lines to output lines as they get
> allocated?

I updated the bindings quite late while cleaning up my changes, the priority naming is
wrong. Because the SoC interrupts are always routed to MIPS CPU interrupts (in my
hardware), they "inherit" a priority, and this is what I had in mind when I started on
these patches.

While updating the binding, I realised that there doesn't need to be a priority to the
output interrupts of this controller. We don't have any real documentation on the
hardware, so for all we know they could maybe even be routed to different interrupt
controllers. I'll update the naming to something else throughout the patch.


Best,
Sander

Powered by blists - more mailing lists