On Wed, Jun 5, 2019 at 1:38 PM Alexey Kardashevskiy <a...@ozlabs.ru> wrote:
>
> When the firmware does PCI BAR resource allocation, it passes the assigned
> addresses and flags (prefetch/64bit/...) via the "reg" property of
> a PCI device device tree node so the kernel does not need to do
> resource allocation.
>
> The flags are stored in resource::flags - the lower byte stores
> PCI_BASE_ADDRESS_SPACE/etc bits and the other bytes are IORESOURCE_IO/etc.
> Some flags from PCI_BASE_ADDRESS_xxx and IORESOURCE_xxx are duplicated,
> such as PCI_BASE_ADDRESS_MEM_PREFETCH/PCI_BASE_ADDRESS_MEM_TYPE_64/etc.
> When parsing the "reg" property, we copy the prefetch flag but we skip
> on PCI_BASE_ADDRESS_MEM_TYPE_64 which leaves the flags out of sync.
>
> The missing IORESOURCE_MEM_64 flag comes into play under 2 conditions:
> 1. we remove PCI_PROBE_ONLY for pseries (by hacking pSeries_setup_arch()
> or by passing "/chosen/linux,pci-probe-only");
> 2. we request resource alignment (by passing pci=resource_alignment=
> via the kernel cmd line to request PAGE_SIZE alignment or defining
> ppc_md.pcibios_default_alignment which returns anything but 0). Note that
> the alignment requests are ignored if PCI_PROBE_ONLY is enabled.
>
> With 1) and 2), the generic PCI code in the kernel unconditionally
> decides to:
> - reassign the BARs in pci_specified_resource_alignment() (works fine)
> - write new BARs to the device - this fails for 64bit BARs as the generic
> code looks at IORESOURCE_MEM_64 (not set) and writes only lower 32bits
> of the BAR and leaves the upper 32bit unmodified which breaks BAR mapping
> in the hypervisor.
>
> This fixes the issue by copying the flag. This is useful if we want to
> enforce certain BAR alignment per platform as handling subpage sized BARs
> is proven to cause problems with hotplug (SLOF already aligns BARs to 64k).
>
> Signed-off-by: Alexey Kardashevskiy <a...@ozlabs.ru>
> ---
>
> This code is there for ages (from 200x) hence no "Fixes:".
>
> Ideally I want to enforce /chosen/linux,pci-probe-only in QEMU as
> at the moment:
> - pci=resource_alignment= alone does not do anything;
> - /chosen/linux,pci-probe-only alone does not cause the kernel to
> reassign resources;
> - pci=resource_alignment= with /chosen/linux,pci-probe-only is broken
> anyway.
>
> ---
> arch/powerpc/kernel/pci_of_scan.c | 2 ++
> 1 file changed, 2 insertions(+)
>
> diff --git a/arch/powerpc/kernel/pci_of_scan.c
> b/arch/powerpc/kernel/pci_of_scan.c
> index 24191ea2d9a7..64ad92016b63 100644
> --- a/arch/powerpc/kernel/pci_of_scan.c
> +++ b/arch/powerpc/kernel/pci_of_scan.c
> @@ -45,6 +45,8 @@ unsigned int pci_parse_of_flags(u32 addr0, int bridge)
> if (addr0 & 0x02000000) {
> flags = IORESOURCE_MEM | PCI_BASE_ADDRESS_SPACE_MEMORY;
> flags |= (addr0 >> 22) & PCI_BASE_ADDRESS_MEM_TYPE_64;
> + if (flags & PCI_BASE_ADDRESS_MEM_TYPE_64)
> + flags |= IORESOURCE_MEM_64;
> flags |= (addr0 >> 28) & PCI_BASE_ADDRESS_MEM_TYPE_1M;
> if (addr0 & 0x40000000)
> flags |= IORESOURCE_PREFETCH
> --
> 2.17.1
Seems like an oversight that PROBE_ONLY has been papering over for years.
Reviewed-by: Oliver O'Halloran <ooh...@gmail.com>
