Re: [PATCH] memory: tegra: add multi-socket support to the memory interconnect

[Thierry Reding]

On Thu, May 28, 2026 at 02:20:07PM +0200, Krzysztof Kozlowski wrote:
> On 28/05/2026 13:56, Thierry Reding wrote:
> >>>>>
> >>>>> -     mc->debugfs.root = debugfs_create_dir("mc", NULL);
> >>>>> +     if (!mc_debugfs_root)
> >>>> That's a probe path and you created a singletone. Looks like preventing
> >>>> async probing for no real reason.
> >>>>
> >>>> I am very against singletons and debugfs does not look like justified
> >>>> exception.
> >>>
> >>> The singleton was added so multi-socket MC/EMC instances could
> >>> share a "mc"/"emc" parent. I'll drop it in v2.
> >>>
> >>> On single-socket SoCs, the "mc"/"emc" names will be unchanged.
> >>> On multi-socket SoCs, each instance will create a top-level debugfs
> >>> dir named with dev_name(). Same pattern in tegra186-emc.c.
> >>>
> >>>    if (dev_to_node(mc->dev) == NUMA_NO_NODE)
> >>>        mc->debugfs.root = debugfs_create_dir("mc", NULL);
> >>>    else
> >>>        mc->debugfs.root = debugfs_create_dir(dev_name(mc->dev), NULL);
> >>
> >> You assume this is fully synced, so you as well could do a look up and
> >> then use what you found or create new dir. If you think that is racy, so
> >> is this approach... How are other drivers handling per-device debugfs
> >> directories? Do they also create such in the top-level? I think no.
> > 
> > I think we want a top-level directory for a bit more structure in
> > debugfs. But I also think we want to create that top-level directory in
> > the module's init function rather than _probe.
> 
> I was thinking about this as well but that would mean your driver will
> create it on every multi-arch kernel.
> 
> This should be then moved to some core bus (and there are examples of
> that, e.g. USB), except there is no core-MC bus code to do that.

We have a utility function (soc_is_tegra()) that we've used in similar
situations in the past. We haven't used them in a little while, but it
could be useful here. It's not for free, but should be fairly quick to
error out early on multi-arch kernels.

Thierry

Attachment: signature.asc
Description: PGP signature