Re: [RFT] Change E820 memory map handling
From: Nikolas Joshua Britton
Date: Tue Nov 17 2026 - 09:18:00 EST
On Tue, 10 Nov 2026, Mario Limonciello (AMD) wrote:
> Here is the flow I would like to see:
> 1. Test 7.4-rc2 (vanilla - no changes please)
> 2. Test this branch (superm1/genoa-mess)
> 3. If this branch fails, please revert this commit from the branch and test again.
> commit 35c9c60ea8c1c ("Revert "x86/PCI: Add E820 quirks for Clevo NL4XLU and Lenovo Yoga C940"")
>
> In all cases please check whether your original issue exists and comment if it does.
I tested on the Mac Pro 7,1 from my bridge window report [1]: two Radeon
Pro Vega II Duo cards, four Vega 20 dies behind two root ports, BAR0
resized to 32G on every die. Both kernels were built from the same
.config.
1. v7.4-rc2: the original issue exists. Both root port windows stay at
32770M, the second die's branch (08:10.0 and 16:10.0) gets no
prefetchable window, and 0e:00.0 and 1c:00.0 are left without BARs:
pci 0000:0e:00.0: BAR 0 [mem size 0x800000000 64bit pref]: can't assign; no space
pci 0000:0e:00.0: BAR 2 [mem size 0x00200000 64bit pref]: can't assign; no space
2. superm1/genoa-mess (35c9c60ea8c1): the original issue is gone. Every
BAR is assigned, both root port windows are 96G, and all four dies
get their 32G BAR0. The layout is identical to the one I reported
for v5 on v7.2.15 [2]. The first root port:
pci 0000:06:00.0: bridge window [mem 0x9e800000000-0x9ffffffffff 64bit pref]: assigned
pci 0000:07:00.0: bridge window [mem 0x9e800000000-0x9ffffffffff 64bit pref]: assigned
pci 0000:08:08.0: bridge window [mem 0x9f7ffe00000-0x9ffffffffff 64bit pref]: assigned
pci 0000:08:10.0: bridge window [mem 0x9e800000000-0x9f0001fffff 64bit pref]: assigned
pci 0000:0b:00.0: BAR 0 [mem 0x9f800000000-0x9ffffffffff 64bit pref]: assigned
pci 0000:0b:00.0: BAR 2 [mem 0x9f7ffe00000-0x9f7ffffffff 64bit pref]: assigned
pci 0000:0e:00.0: BAR 0 [mem 0x9e800000000-0x9efffffffff 64bit pref]: assigned
pci 0000:0e:00.0: BAR 2 [mem 0x9f000000000-0x9f0001fffff 64bit pref]: assigned
amdgpu binds all four dies, and compute and peer-to-peer transfers
across them work.
3. Not needed, since the branch works.
As for E820 itself, 07eab0901ede removes a single descriptor on this
machine:
efi: Remove mem126: MMIO range=[0x60000000-0x600fffff] (1MB) from e820 map
That is the first 1MB of ECAM (0x60000000-0x6fffffff), below all host
bridge windows. On the branch the line is gone and /proc/iomem keeps
5e000000-600fffff as Reserved. The firmware is dated 2026, so both
kernels print "PCI: Ignoring E820 reservations for host bridge windows"
and the restored entry could not clip a window here anyway. Apart from
the efi/e820 lines and what follows from the BAR assignment (amdgpu
binding two more dies), the kernel logs of the two boots do not differ.
Also, the boot failure on v7.3 and later that I mentioned in the v5
thread does not reproduce with v7.4-rc2, so both kernels here are
plain upstream. It does not reproduce with Ubuntu 26.10's 7.3.0-10
kernel either, and I have since tested v5 on that kernel as well.
Unpatched, it shows the original issue (two of the four dies without
BARs). With the 14 patches applied (they apply without fuzz and need no
prerequisites on 7.3), all four dies get their BARs with the same
layout as above, and a multi-hour inference benchmark across the four
dies showed no regression against our reference runs.
Tested-by: Nikolas Joshua Britton <nbritton@xxxxxxxxx>
[1] https://lore.kernel.org/linux-pci/20260903063124.9316-1-nbritton@xxxxxxxxx/
[2] https://lore.kernel.org/linux-pci/20261106152642.23118-1-nbritton@xxxxxxxxx/