Re: [PATCH] cache: sifive_ccache: skip disabled device tree nodes

From: Conor Dooley

Date: Mon Nov 16 2026 - 12:20:55 EST


On Sat, Nov 14, 2026 at 07:40:42PM +0200, Avi wrote:
> On Thu, Nov 5, 2026 at 8:24 PM Conor Dooley <conor@xxxxxxxxxx> wrote:
> >
> > On Wed, Oct 28, 2026 at 03:19:40PM +0200, Avi Weiss wrote:
> > > sifive_ccache_init() uses of_find_matching_node_and_match(), which also
> > > returns nodes marked as disabled. The initializer can therefore map and
> > > access a cache controller that the device tree marks unavailable.
> > >
> > > Select the first available matching node before accessing its registers.
> > > This also handles a disabled node preceding an enabled one.
> > >
> > > Fixes: a967a289f169 ("RISC-V: sifive_l2_cache: Add L2 cache controller driver for SiFive SoCs")
> > > Signed-off-by: Avi Weiss <thnkslprpt@xxxxxxxxx>
> > > ---
> > > drivers/cache/sifive_ccache.c | 4 +++-
> > > 1 file changed, 3 insertions(+), 1 deletion(-)
> > >
> > > diff --git a/drivers/cache/sifive_ccache.c b/drivers/cache/sifive_ccache.c
> > > index 71d3f94987fa..02878d54381d 100644
> > > --- a/drivers/cache/sifive_ccache.c
> > > +++ b/drivers/cache/sifive_ccache.c
> > > @@ -297,7 +297,9 @@ static int __init sifive_ccache_init(void)
> > > unsigned long quirks __maybe_unused;
> > > int rc;
> > >
> > > - np = of_find_matching_node_and_match(NULL, sifive_ccache_ids, &match);
> > > + for_each_matching_node_and_match(np, sifive_ccache_ids, &match)
> > > + if (of_device_is_available(np))
> >
> > Are there any devices where this is actually possible? IOW, where there
> > are 2 ccache instances.

> It's more of a defensive thing. The 2 ccache instance was an example -
> the issue is also relevant for one disabled ccache node no?

That is true.
And on the two node front, the driver doesn't support that anyway either
before or after the patch. Consider what the new code does if there are
two nodes, both enabled.

But I don't think the ccache or llc-cache successor are intended for
that kind of situation anyway.

> Since of_find_matching_node_and_match() doesn't check whether a node
> is enabled, it can return a disabled one.
>
> sifive_ccache_init() then goes ahead and maps and accesses its
> registers, even though the device tree marks it as disabled and normal
> OF platform-device creation would skip it.

Could we instead drop the loop and just check for availability, while
retaining of_find_matching_node_and_match()? I kinda feel like the loop
gives the impression that the driver supports multiple nodes, when it
realistically does not.

Cheers,
Conor.

>
> >
> > > + break;
> > > if (!np)
> > > return -ENODEV;
> > >
> > >
> > > base-commit: 65ba8f34ec886ec72b2c07ddc8cf3d815c32869d
> > > --
> > > 2.43.0
> > >

Attachment: signature.asc
Description: PGP signature