Re: gpio-mt7621 unroutable IRQs to bank0

[Vicente Bergas]

On Tue, May 5, 2026 at 5:05 PM Sergio Paracuellos
<sergio.paracuellos@xxxxxxxxx> wrote:
>
> Hi,
>
> On Tue, May 5, 2026 at 4:21 PM Thomas Gleixner <tglx@xxxxxxxxxx> wrote:
> >
> > On Tue, May 05 2026 at 14:01, Linus Walleij wrote:
> > > On Sat, May 2, 2026 at 11:52 PM Vicente Bergas <vicencb@xxxxxxxxx> wrote:
> > >> As a way to prove that this is indeed the problem,
> > >> the following workaround makes it work.
> > >> It just inverts the sorting order of all matches,
> > >> so it picks Bank0 instead of Bank2.
> > >
> > > That's a tricksy bug, I can't exactly see where the issue
> > > is.
> > >
> > > I think to solve this you might need to allocate an external
> > > irqdomain that deal with the three different gpiochip
> > > instances when translating the irqs.
> >
> > struct gpio_chip has this:
> >
> >         /**
> >          * @of_node_instance_match:
> >          *
> >          * Determine if a chip is the right instance. Must be implemented by
> >          * any driver using more than one gpio_chip per device tree node.
> >          * Returns true if gc is the instance indicated by i (which is the
> >          * first cell in the phandles for GPIO lines and gpio-ranges).
> >          */
> >         bool (*of_node_instance_match)(struct gpio_chip *gc, unsigned int i);
> >
> > That driver falls in the category and lacks that callback, no?
> >
> > Thanks,
> >
> >         tglx
>
> The IP core used inside these SoCs has 3 banks of 32 GPIOs each but
> there is only one device tree node because the registers of all the
> banks are interwoven inside one single IO range. Thus, the driver
> internally sets up a gpio controller instance per bank. The driver
> lacks of_node_instance_match callback but I am not sure if it needs to
> be implemented in this particular case since the driver is using
> of_xlate callback for this.
>
> Thanks,
>     Sergio Paracuellos

Hi all,
just tested with the suggested `of_node_instance_match` by applying
those changes:

```
--- a/drivers/gpio/gpio-mt7621.c
+++ b/drivers/gpio/gpio-mt7621.c
@@ -1,3 +1,5 @@
+#define DEBUG  1
+
 // SPDX-License-Identifier: GPL-2.0
 /*
  * Copyright (C) 2009-2011 Gabor Juhos <juhosg@xxxxxxxxxxx>
@@ -206,6 +208,18 @@
     return gpio % MTK_BANK_WIDTH;
 }

+static bool mediatek_gpio_node_instance_match(struct gpio_chip *chip,
unsigned int i)
+{
+    struct mtk_gc *rg = to_mediatek_gpio(chip);
+
+    struct mtk_gc *rg0 = rg - rg->bank;
+    struct mtk *mtk = container_of(rg0, struct mtk, gc_map[0]);
+    dev_dbg(mtk->dev, "%u <=> %d\n", i, rg->bank);
+    dump_stack();
+
+    return i == rg->bank;
+}
+
 static const struct irq_chip mt7621_irq_chip = {
     .name        = "mt7621-gpio",
     .irq_mask_ack    = mediatek_gpio_irq_mask,
@@ -253,6 +267,7 @@

     rg->chip.gc.of_gpio_n_cells = 2;
     rg->chip.gc.of_xlate = mediatek_gpio_xlate;
+    rg->chip.gc.of_node_instance_match = mediatek_gpio_node_instance_match;
     rg->chip.gc.label = devm_kasprintf(dev, GFP_KERNEL, "%s-bank%d",
                     dev_name(dev), bank);
     if (!rg->chip.gc.label)
```

It turns out that the `mediatek_gpio_node_instance_match` function is
never called.

Regards,
  Vicente.