On 10.04.2025 22:55, Jason Andryuk wrote: > On 2025-04-04 06:28, Roger Pau Monné wrote: >> On Thu, Apr 03, 2025 at 06:01:42PM -0700, Stefano Stabellini wrote: >>> On one Sapphire AMD x86 board, I see this: >>> >>> >>> (XEN) [0000003943ca6ff2] [00000000f0000000, 00000000f7ffffff] (reserved) >>> (XEN) [00000039460886d9] [00000000fd000000, 00000000ffffffff] (reserved) >>> [...] >>> (XEN) [ 4.612235] 0000:02:00.0: not mapping BAR [fea00, fea03] invalid >>> position >>> >>> >>> Linux boots fine on this platform but Linux as Dom0 on Xen does not. >>> This is because the pci_check_bar->is_memory_hole check fails due to the >>> MMIO region overlapping with the EFI reserved region. >> >> That's weird. (Partially) the reason to not attempt to map such BAR >> is that it should already be mapped, because at dom0 creation time all >> reserved regions are added to the p2m (see arch_iommu_hwdom_init()). >> If that's not the case we should figure out why this reserved region >> is not added to dom0 p2m as part of arch_iommu_hwdom_init(). > > Victor discovered these regions are type 11 EfiMemoryMappedIO, but they > get converted to e820 RESERVED. The BAR points into it. > > 00000f0000000-00000f7ffffff type=11 attr=800000000000100d > 00000fd000000-00000fedfffff type=11 attr=800000000000100d > 00000fee00000-00000fee00fff type=11 attr=8000000000000001 > 00000fee01000-00000ffffffff type=11 attr=800000000000100d > > Xenia discovered Linux keeps small regions like this reserved, but lets > larger ones (>= 256kb) become holes. See the comment in Linux > arch/x86/platform/efi/efi.c:efi_remove_e820_mmio() around line 301.
What a hack. And the mentioning of MMCFG space isn't even correct there, I think. That space may legitimately be reserved (and at least older Linux actually checked for that), unlike any ranges where BARs may live. As to host bridge windows - I dare to question that they always need to be "large". Similarly nothing guarantees the some non-window space included in _CRS may not point at a "large" region. Jan
