On Fri, Aug 16, 2024 at 04:14:52PM GMT, Anup Patel wrote:
> On Fri, Aug 16, 2024 at 3:25 PM Andrew Jones <ajones@xxxxxxxxxxxxxxxx> wrote:
> >
> > An 'msi-parent' property with a single entry and no accompanying
> > '#msi-cells' property is considered the legacy definition as opposed
> > to its definition after being expanded with commit 126b16e2ad98
> > ("Docs: dt: add generic MSI bindings"). However, the legacy
> > definition is completely compatible with the current definition and,
> > since of_phandle_iterator_next() tolerates missing and present-but-
> > zero *cells properties since commit e42ee61017f5 ("of: Let
> > of_for_each_phandle fallback to non-negative cell_count"), there's no
> > need anymore to special case the legacy definition in
> > of_msi_get_domain().
> >
> > Indeed, special casing has turned out to be harmful, because, as of
> > commit 7c025238b47a ("dt-bindings: irqchip: Describe the IMX MU block
> > as a MSI controller"), MSI controller DT bindings have started
> > specifying '#msi-cells' as a required property (even when the value
> > must be zero) as an effort to make the bindings more explicit. But,
> > since the special casing of 'msi-parent' only uses the existence of
> > '#msi-cells' for its heuristic, and not whether or not it's also
> > nonzero, the legacy path is not taken. Furthermore, the path to
> > support the new, broader definition isn't taken either since that
> > path has been restricted to the platform-msi bus.
> >
> > But, neither the definition of 'msi-parent' nor the definition of
> > '#msi-cells' is platform-msi-specific (the platform-msi bus was just
> > the first bus that needed '#msi-cells'), so remove both the special
> > casing and the restriction. This not only simplifies the code but
> > also resolves an issue with PCI devices finding their MSI controllers
> > on riscv, as the riscv,imsics binding requires '#msi-cells=<0>'.
> >
> > Signed-off-by: Andrew Jones <ajones@xxxxxxxxxxxxxxxx>
> > ---
> > drivers/of/irq.c | 37 +++++++++++--------------------------
> > 1 file changed, 11 insertions(+), 26 deletions(-)
> >
> > diff --git a/drivers/of/irq.c b/drivers/of/irq.c
> > index c94203ce65bb..026b52c8ee63 100644
> > --- a/drivers/of/irq.c
> > +++ b/drivers/of/irq.c
> > @@ -709,8 +709,7 @@ struct irq_domain *of_msi_map_get_device_domain(struct device *dev, u32 id,
> > * @np: device node for @dev
> > * @token: bus type for this domain
> > *
> > - * Parse the msi-parent property (both the simple and the complex
> > - * versions), and returns the corresponding MSI domain.
> > + * Parse the msi-parent property and returns the corresponding MSI domain.
> > *
> > * Returns: the MSI domain for this device (or NULL on failure).
> > */
> > @@ -718,33 +717,19 @@ struct irq_domain *of_msi_get_domain(struct device *dev,
> > struct device_node *np,
> > enum irq_domain_bus_token token)
> > {
> > - struct device_node *msi_np;
> > + struct of_phandle_args args;
> > struct irq_domain *d;
> > + int index = 0;
> >
> > - /* Check for a single msi-parent property */
> > - msi_np = of_parse_phandle(np, "msi-parent", 0);
> > - if (msi_np && !of_property_read_bool(msi_np, "#msi-cells")) {
> > - d = irq_find_matching_host(msi_np, token);
> > - if (!d)
> > - of_node_put(msi_np);
> > - return d;
> > - }
>
> Dropping this special case of single msi-parent property breaks
> RISC-V KVM guest created using KVMTOOl with AIA available
> on the host.
Ouch. This is a paperbag moment as I can easily reproduce that.
>
> Let's not drop this special case.
I'll work out a [much better tested] v2 now.
Thanks,
drew
