|
[Date Prev][Date Next][Thread Prev][Thread Next][Date Index][Thread Index] Re: [PATCH] xen/arm: domain_build: Ignore device nodes with invalid addresses
On Tue, 2 Feb 2021, Julien Grall wrote:
> On 02/02/2021 18:12, Julien Grall wrote:
> > On 02/02/2021 17:47, Elliott Mitchell wrote:
> > > The handle_device() function has been returning failure upon
> > > encountering a device address which was invalid. A device tree which
> > > had such an entry has now been seen in the wild. As it causes no
> > > failures to simply ignore the entries, ignore them. >
> > > Signed-off-by: Elliott Mitchell <ehem+xenn@xxxxxxx>
> > >
> > > ---
> > > I'm starting to suspect there are an awful lot of places in the various
> > > domain_build.c files which should simply ignore errors. This is now the
> > > second place I've encountered in 2 months where ignoring errors was the
> > > correct action.
> >
> > Right, as a counterpoint, we run Xen on Arm HW for several years now and
> > this is the first time I heard about issue parsing the DT. So while I
> > appreciate that you are eager to run Xen on the RPI...
> >
> > > I know failing in case of error is an engineer's
> > > favorite approach, but there seem an awful lot of harmless failures
> > > causing panics.
> > >
> > > This started as the thread "[RFC PATCH] xen/arm: domain_build: Ignore
> > > empty memory bank". Now it seems clear the correct approach is to simply
> > > ignore these entries.
> >
> > ... we first need to fully understand the issues. Here a few questions:
> > 1) Can you provide more information why you believe the address is
> > invalid?
> > 2) How does Linux use the node?
> > 3) Is it happening with all the RPI DT? If not, what are the
> > differences?
>
> So I had another look at the device-tree you provided earlier on. The node is
> the following (copied directly from the DTS):
>
> &pcie0 {
> pci@1,0 {
> #address-cells = <3>;
> #size-cells = <2>;
> ranges;
>
> reg = <0 0 0 0 0>;
>
> usb@1,0 {
> reg = <0x10000 0 0 0 0>;
> resets = <&reset RASPBERRYPI_FIRMWARE_RESET_ID_USB>;
> };
> };
> };
>
> pcie0: pcie@7d500000 {
> compatible = "brcm,bcm2711-pcie";
> reg = <0x0 0x7d500000 0x0 0x9310>;
> device_type = "pci";
> #address-cells = <3>;
> #interrupt-cells = <1>;
> #size-cells = <2>;
> interrupts = <GIC_SPI 148 IRQ_TYPE_LEVEL_HIGH>,
> <GIC_SPI 148 IRQ_TYPE_LEVEL_HIGH>;
> interrupt-names = "pcie", "msi";
> interrupt-map-mask = <0x0 0x0 0x0 0x7>;
> interrupt-map = <0 0 0 1 &gicv2 GIC_SPI 143
> IRQ_TYPE_LEVEL_HIGH>;
> msi-controller;
> msi-parent = <&pcie0>;
>
> ranges = <0x02000000 0x0 0xc0000000 0x6 0x00000000
> 0x0 0x40000000>;
> /*
> * The wrapper around the PCIe block has a bug
> * preventing it from accessing beyond the first 3GB of
> * memory.
> */
> dma-ranges = <0x02000000 0x0 0x00000000 0x0 0x00000000
> 0x0 0xc0000000>;
> brcm,enable-ssc;
> };
>
> The interpretation of "reg" depends on the context. In this case, we are
> trying to interpret as a memory address from the CPU PoV when it has a
> different meaning (I am not exactly sure what).
>
> In fact, you are lucky that Xen doesn't manage to interpret it. Xen should
> really stop trying to look region to map when it discover a PCI bus. I wrote a
> quick hack patch that should ignore it:
Yes, I think you are right. There are a few instances where "reg" is not
a address ready to be remapped. It is not just PCI, although that's the
most common. Maybe we need a list, like skip_matches in handle_node.
> diff --git a/xen/arch/arm/domain_build.c b/xen/arch/arm/domain_build.c
> index 374bf655ee34..937fd1e387b7 100644
> --- a/xen/arch/arm/domain_build.c
> +++ b/xen/arch/arm/domain_build.c
> @@ -1426,7 +1426,7 @@ static int __init handle_device(struct domain *d, struct
> dt_device_node *dev,
>
> static int __init handle_node(struct domain *d, struct kernel_info *kinfo,
> struct dt_device_node *node,
> - p2m_type_t p2mt)
> + p2m_type_t p2mt, bool pci_bus)
> {
> static const struct dt_device_match skip_matches[] __initconst =
> {
> @@ -1532,9 +1532,14 @@ static int __init handle_node(struct domain *d, struct
> kernel_info *kinfo,
> "WARNING: Path %s is reserved, skip the node as we may re-use
> the path.\n",
> path);
>
> - res = handle_device(d, node, p2mt);
> - if ( res)
> - return res;
> + if ( !pci_bus )
> + {
> + res = handle_device(d, node, p2mt);
> + if ( res)
> + return res;
> +
> + pci_bus = dt_device_type_is_equal(node, "pci");
> + }
>
> /*
> * The property "name" is used to have a different name on older FDT
> @@ -1554,7 +1559,7 @@ static int __init handle_node(struct domain *d, struct
> kernel_info *kinfo,
>
> for ( child = node->child; child != NULL; child = child->sibling )
> {
> - res = handle_node(d, kinfo, child, p2mt);
> + res = handle_node(d, kinfo, child, p2mt, pci_bus);
> if ( res )
> return res;
> }
> @@ -2192,7 +2197,7 @@ static int __init prepare_dtb_hwdom(struct domain *d,
> struct kernel_info *kinfo)
>
> fdt_finish_reservemap(kinfo->fdt);
>
> - ret = handle_node(d, kinfo, dt_host, default_p2mt);
> + ret = handle_node(d, kinfo, dt_host, default_p2mt, false);
> if ( ret )
> goto err;
>
> A less hackish possibility would be to modify dt_number_of_address() and
> return 0 when the device is a child of a PCI below.
>
> Stefano, do you have any opinions?
Would PCIe even work today? Because if it doesn't, we could just add it
to skip_matches until we get PCI passthrough properly supported.
But aside from PCIe, let's say that we know of a few nodes for which
"reg" needs a special treatment. I am not sure it makes sense to proceed
with parsing those nodes without knowing how to deal with that. So maybe
we should add those nodes to skip_matches until we know what to do with
them. At that point, I would imagine we would introduce a special
handle_device function that knows what to do. In the case of PCIe,
something like "handle_device_pcie".
|
![]() |
Lists.xenproject.org is hosted with RackSpace, monitoring our |