On Thu, 28 May 2026, Peter Maydell wrote:
On Thu, 28 May 2026 at 10:23, Sebastian Ott <[email protected]> wrote:
ACPI PCI hotplug of a PCI device using a bar > 2M fails on ARM
with a resource allocation failure in the guest kernel (Linux):
qemu-system-aarch64 [..] -device pcie-root-port,id=rp1,bus=pcie.0 \
-global acpi-ged.acpi-pci-hotplug-with-bridge-support=on \
-pflash /usr/share/edk2/aarch64/QEMU_EFI-silent-pflash.qcow2
(qemu) device_add pci-testdev,bus=rp1,membar=4M
[ 55.106283] pci 0000:01:00.0: [1b36:0005] type 00 class 0x00ff00
conventional PCI endpoint
[ 55.107627] pci 0000:01:00.0: BAR 0 [mem 0x00000000-0x00000fff]
[ 55.107864] pci 0000:01:00.0: BAR 1 [io 0x0000-0x00ff]
[ 55.108015] pci 0000:01:00.0: BAR 2 [mem 0x00000000-0x003fffff 64bit pref]
[ 55.118609] pci 0000:01:00.0: BAR 2 [mem size 0x00400000 64bit pref]: can't
assign; no space
[ 55.118653] pci 0000:01:00.0: BAR 2 [mem size 0x00400000 64bit pref]: failed
to assign
[ 55.118671] pci 0000:01:00.0: BAR 0 [mem 0x10000000-0x10000fff]: assigned
[ 55.122828] pci 0000:01:00.0: BAR 1 [io 0x1000-0x10ff]: assigned
[ 55.127090] pci 0000:01:00.0: BAR 0 [mem 0x10000000-0x10000fff]: releasing
[ 55.127098] pci 0000:01:00.0: BAR 2 [mem size 0x00400000 64bit pref]: can't
assign; no space
[ 55.127102] pci 0000:01:00.0: BAR 2 [mem size 0x00400000 64bit pref]: failed
to assign
[ 55.127106] pci 0000:01:00.0: BAR 0 [mem 0x10000000-0x10000fff]: assigned
This behavior is exactly the same on x86 - except the busses are created
with a default window size of 256M there. If the PCI devices being
hotplugged use a bar > 256M you get the same failure like on ARM.
The workaround on x86 is to specify "pref64-reserve=" for the PCIe
root-port. On ARM this is currently ignored.
From the perspective of somebody who doesn't know much about
ACPI or PCI, this commit message doesn't really tell me why
QEMU needs to change. The above reads like "this is a kernel
problem" to me. Why is the kernel unable to allocate this BAR?
DSM #5 defaults to 1 on ARM which tells the OS that it could
ignore the firmwares PCI configuration at boot time. Linux is
just doing that. The acpi pci hotplug controller can't find a
sufficient resource window of the parent bus and resource
allocation fails. DSM #5 == 0 tells the OS to not ignore the
boot time config from firmware.
The following patch changes ACPI DSM #5 to return 0 (only when
ACPI HP is used) which basically tells the guest kernel to keep
all PCI resource allocations.
If this makes things work why isn't it the default for all ACPI?
Why doesn't the kernel ignore the ACPI table setting and do
the thing that works all the time? Do we break some other use
case by setting it to 0 ?
From what I was able to find in the archives this has been tried
before but led to regressions for legacy devices:
https://lore.kernel.org/all/[email protected]
It was recently selectively enabled for accel SMMU - I guess the
rational was that you don't use accel SMMU with legacy devices:
45c57eaefd hw/arm/virt: Set PCI preserve_config for accel SMMUv3
------->8
From 92cd64b8398f13e51b67ff8baf62f15cac300c94 Mon Sep 17 00:00:00 2001
From: Sebastian Ott <[email protected]>
Date: Wed, 20 May 2026 10:39:26 -0400
Subject: [PATCH] hw/arm/virt: preserve config for ACPI PCI hotplug
Change ACPI DSM #5 to return 0 when ACPI PCI hotplug is used
to tell the guest kernel to keep all PCI resource allocations.
This allows to use the pref64-reserve= option for pcie-root-ports
and thus enables hotplug of PCI devices with a barsize > 2M.
Signed-off-by: Sebastian Ott <[email protected]>
---
hw/arm/virt.c | 1 +
1 file changed, 1 insertion(+)
diff --git a/hw/arm/virt.c b/hw/arm/virt.c
index b090233893..293a0c2163 100644
--- a/hw/arm/virt.c
+++ b/hw/arm/virt.c
@@ -1094,6 +1094,7 @@ static inline DeviceState
*create_acpi_ged(VirtMachineState *vms)
pcihp_region_index = sysbus_mmio_map_name(sbdev,
ACPI_PCIHP_REGION_NAME,
vms->memmap[VIRT_ACPI_PCIHP].base);
assert(pcihp_region_index >= 0);
+ vms->pci_preserve_config = true;
}
If we set this here, do we still need to set it in
virt_machine_device_pre_plug_cb(), or is that now redundant?
I guess you can use accel SMMU without acpi pci hotplug and vice versa.
Thanks,
Sebastian
