Re: [PATCH v3 1/4] PCI: of: Clear 64-bit flag for non-prefetchable memory below 4GB

[Bjorn Helgaas <helgaas@kernel.org>]

[+cc Leonardo]

On Mon, Jun 07, 2021 at 08:28:53PM +0900, Punit Agrawal wrote:
> Some host bridges advertise non-prefetchable memory windows that are
> entirely located below 4GB but are marked as 64-bit address memory.
> 
> Since commit 9d57e61bf723 ("of/pci: Add IORESOURCE_MEM_64 to resource
> flags for 64-bit memory addresses"), the OF PCI range parser takes a
> stricter view and treats 64-bit address ranges as advertised while
> before such ranges were treated as 32-bit.
> 
> A PCI root port modelled as a PCI-to-PCI bridge cannot forward 64-bit
> non-prefetchable memory ranges. As a result, the change in behaviour
> due to the commit causes failure to allocate 32-bit BAR from a 64-bit
> non-prefetchable window.
> 
> In order to not break platforms where non-prefetchable memory ranges
> lie entirely below 4GB, clear the 64-bit flag.

I don't think we should care about the address width DT supplies for a
host bridge window.  Prior to 9d57e61bf723, I don't think we *did*
care because of_bus_pci_get_flags() threw away that information.

My proposal for a commit log, including information about the problem
report and a "Fixes:" tag:

  Alexandru and Qu reported this resource allocation failure on
  ROCKPro64 v2 and ROCK Pi 4B, both based on the RK3399:

    pci_bus 0000:00: root bus resource [mem 0xfa000000-0xfbdfffff 64bit]
    pci 0000:00:00.0: PCI bridge to [bus 01]
    pci 0000:00:00.0: BAR 14: no space for [mem size 0x00100000]
    pci 0000:01:00.0: reg 0x10: [mem 0x00000000-0x00003fff 64bit]

  "BAR 14" is the PCI bridge's 32-bit non-prefetchable window, and our
  PCI allocation code isn't smart enough to allocate it in a host
  bridge window marked as 64-bit, even though this should work fine.

  A DT host bridge description includes the windows from the CPU
  address space to the PCI bus space.  On a few architectures
  (microblaze, powerpc, sparc), the DT may also describe PCI devices
  themselves, including their BARs.

  Before 9d57e61bf723 ("of/pci: Add IORESOURCE_MEM_64 to resource
  flags for 64-bit memory addresses"), of_bus_pci_get_flags() ignored
  the fact that some DT addresses described 64-bit windows and BARs.
  That was a problem because the virtio virtual NIC has a 32-bit BAR
  and a 64-bit BAR, and the driver couldn't distinguish them.

  9d57e61bf723 set IORESOURCE_MEM_64 for those 64-bit DT ranges, which
  fixed the virtio driver.  But it also set IORESOURCE_MEM_64 for host
  bridge windows, which exposed the fact that the PCI allocator isn't
  smart enough to put 32-bit resources in those 64-bit windows.

  Clear IORESOURCE_MEM_64 from host bridge windows since we don't need
  that information.

  Fixes: 9d57e61bf723 ("of/pci: Add IORESOURCE_MEM_64 to resource flags for 64-bit memory addresses")
  Reported-at: https://lore.kernel.org/lkml/7a1e2ebc-f7d8-8431-d844-41a9c36a8911@arm.com/
  Reported-by: Alexandru Elisei <alexandru.elisei@arm.com>
  Reported-by: Qu Wenruo <wqu@suse.com>

> Suggested-by: Ard Biesheuvel <ardb@kernel.org>
> Link: https://lore.kernel.org/r/7a1e2ebc-f7d8-8431-d844-41a9c36a8911@arm.com
> Signed-off-by: Punit Agrawal <punitagrawal@gmail.com>
> Tested-by: Alexandru Elisei <alexandru.elisei@arm.com>
> Cc: Bjorn Helgaas <bhelgaas@google.com>
> Cc: Rob Herring <robh+dt@kernel.org>
> ---
>  drivers/pci/of.c | 8 ++++++++
>  1 file changed, 8 insertions(+)
> 
> diff --git a/drivers/pci/of.c b/drivers/pci/of.c
> index 85dcb7097da4..1e45186a5715 100644
> --- a/drivers/pci/of.c
> +++ b/drivers/pci/of.c
> @@ -353,6 +353,14 @@ static int devm_of_pci_get_host_bridge_resources(struct device *dev,
>  				dev_warn(dev, "More than one I/O resource converted for %pOF. CPU base address for old range lost!\n",
>  					 dev_node);
>  			*io_base = range.cpu_addr;
> +		} else if (resource_type(res) == IORESOURCE_MEM) {
> +			if (!(res->flags & IORESOURCE_PREFETCH)) {
> +				if (res->flags & IORESOURCE_MEM_64)
> +					if (!upper_32_bits(range.pci_addr + range.size - 1)) {
> +						dev_warn(dev, "Clearing 64-bit flag for non-prefetchable memory below 4GB\n");
> +						res->flags &= ~IORESOURCE_MEM_64;
> +					}
> +			}

Why do we need to check IORESOURCE_PREFETCH, IORESOURCE_MEM_64, and
upper_32_bits()?  If I understand this correctly, prior to
9d57e61bf723, IORESOURCE_MEM_64 was *never* set here.  Isn't something
like this sufficient?

  } else if (resource_type(res) == IORESOURCE_MEM) {
    res->flags &= ~IORESOURCE_MEM_64;
  }

I'm not sure we need a warning either.  We didn't warn before
9d57e61bf723, and there's nothing the user needs to do anyway.

>  		}
>  
>  		pci_add_resource_offset(resources, res,	res->start - range.pci_addr);
> -- 
> 2.30.2
>