Re: [PATCH] arm64: PCI: Validate the node before setting node id for root bus

[Baolin Wang <baolin.wang@linux.alibaba.com>]

On Mon, Sep 28, 2020 at 03:00:55PM +0100, Will Deacon wrote:
> [+ Lorenzo]
> 
> On Tue, Sep 22, 2020 at 06:33:24PM +0800, Baolin Wang wrote:
> > If the BIOS disabled the NUMA configuration, but did not change the
> > proximity domain description in the SRAT table, so the PCI root bus
> > device may get a incorrect node id by acpi_get_node().
> 
> How "incorrect" are we talking here? What actually goes wrong? At some
> point, we have to trust what the firmware is telling us.

What I mean is, if we disable the NUMA from BIOS, but we did not change
the PXM for the PCI devices, so the PCI devices can still get a numa
node id from acpi_get_node(). For example, we can still get the numa
node id = 1 in this case from acpi_get_node(), but the numa_nodes_parsed
is empty, which means the node id 1 is invalid. We should add a
validation for the node id when setting the root bus node id.

> 
> > Thus better to add a numa node validation before setting numa node
> > for the PCI root bus, like pci_acpi_root_get_node() does for X86
> > architecture.
> > 
> > Signed-off-by: Baolin Wang <baolin.wang@linux.alibaba.com>
> > ---
> >  arch/arm64/kernel/pci.c | 6 +++++-
> >  1 file changed, 5 insertions(+), 1 deletion(-)
> > 
> > diff --git a/arch/arm64/kernel/pci.c b/arch/arm64/kernel/pci.c
> > index 1006ed2..24fe2bd 100644
> > --- a/arch/arm64/kernel/pci.c
> > +++ b/arch/arm64/kernel/pci.c
> > @@ -86,9 +86,13 @@ int pcibios_root_bridge_prepare(struct pci_host_bridge *bridge)
> >  		struct pci_config_window *cfg = bridge->bus->sysdata;
> >  		struct acpi_device *adev = to_acpi_device(cfg->parent);
> >  		struct device *bus_dev = &bridge->bus->dev;
> > +		int node = acpi_get_node(acpi_device_handle(adev));
> > +
> > +		if (node != NUMA_NO_NODE && !node_online(node))
> > +			node = NUMA_NO_NODE;
> 
> Hmm. afaict, acpi_get_node() tries quite hard to return a valid node when
> it gets back NUMA_NO_NODE in acpi_map_pxm_to_node(). Seems like we're
> undoing all of that here, which worries me because NUMA_NO_NODE is a bit
> of a loaded gun if you interpret it as a valid node.

I did not treate NUMA_NO_NODE as a valid node, I just add a validation
to validate if it is a valid node before setting. See my previous comments,
hopes I make things clear. Thanks.

> 
> Anyway, I defer to Lorenzo on this.
> 
> Will