Skip to main content

Siemens Innexis ANA: Arm Zena CSS

This page records a bringup of the Avocado OS arm64-uefi-v9 target on the Arm Zena Compute Subsystem (CSS) virtual platform of a Siemens Innexis Architecture Native Acceleration (ANA) evaluation chamber. The target is in progress for the 2026 release. This is not a supported platform, and the page only states what was observed on the chamber. A command that was not run for this page is marked "not re-run".

Platform​

The project was created from the chamber's Arm-Zena-CSS template, which carries the EWAOL software stack.

ItemValue
TemplateArm-Zena-CSS (EWAOL software stack)
Simulation CPUs4, which the guest sees as 16 CPUs
HostA Graviton 2 server (Armv8.2), as stated by the vendor and not observed
Boot chainRSE, TF-A BL2, OP-TEE, U-Boot as UEFI firmware, systemd-boot, Linux

What booted​

Avocado OS booted from a QCOW2 made from the provisioned arm64-uefi-v9 image and reached a root shell. The boot went through RSE, TF-A BL2, OP-TEE, U-Boot as UEFI firmware, systemd-boot, Linux 6.18.52 and systemd 259.5.

Boot record of the second run:

  • The project reported running 15 s after the power-on call. The record does not say which event this was read from, so it is not comparable with the overlay run's 287 s below.
  • The login prompt appeared 20 s later.
  • 7.2 s passed from the start of the kernel to the prompt.

Guest facts from a root login:

ItemValue
CPUCortex-A720AE (CPU implementer 0x41, CPU part 0xd89), 16 CPUs
Featuresfp asimd evtstrm aes pmull sha1 crc32 atomics fphp asimdhp cpuid asimdrdm lrcpc dcpop asimddp
Memory46 GiB of RAM
TEEOP-TEE 4.5 with /dev/tee0, and no /dev/tpm*
Networketh0 on DHCP
Diskvda with the partitions boot-a, boot-b, recovery, rootfs-a, rootfs-b and var

systemctl is-system-running reported degraded. The only failed unit was the systemd-growfs instance for /var; its cause was not investigated.

Obstacle 1: memory​

The template's vsoc.json sets memory > populated-ram to 128 (GB). The chamber offers 112 GiB to a virtual SoC, so Start failed with:

Requested RAM: 128GiB, but available RAM: 112GiB

The value is editable in the Web IDE editor (find and replace, then save). The vendor's Zena guide keeps it at 48 GB or less for GPU passthrough.

The firmware must be rebuilt with the same size, because TF-A and OP-TEE carry it: EWAOL_ADDONS_DRAM_SIZE_GB feeds DRAM_SIZE_GB and CFG_DRAM_SIZE_GB. DRAM0 is 2 GB at 0x80000000 and DRAM1 is the rest, at 0x200_00000000.

The value used was 48. U-Boot printed DRAM: 2 GiB (effective 48 GiB) and the guest had 46 GiB.

Obstacle 2: firmware for a Graviton 2 host​

The prebuilt Zena stack is compiled for -march=armv9.2-a, so it was rebuilt for the Graviton 2 host.

The source is the ewaol-addons.tar.gz staged in every Zena project (Web IDE, right-click, Download). It is a kas tree over Arm's sw-ref-stack (commit 6bac7e67). The menu offers the option "Cortex A720AE on Graviton2 (compile for Armv8.2)", which applies a neoverse-n1 tune patch:

kas menu ewaol-addons/Kconfig

The choices were:

  • the Arm model EULA accepted
  • ANA-RD-Zena
  • the EWAOL demo
  • Graviton2
  • DRAM 48
  • GPU, virtual camera and virtio sound off
  • RAS off

Of the firmware outputs only fip_with_bl2.bin changed size (2039905 to 2039817 bytes). The RSE images, the provisioning message and the OTP image kept their sizes; their contents were not compared. Only the FIP was replaced, in firmware/secure-world/. The platform does not keep the original once it is replaced, so a local copy was saved first.

Build notes:

  • The build ran in an Ubuntu 22.04 docker image with kas 4.3.2.
  • The image needed iproute2.
  • Ubuntu's security-updated tar 1.34+dfsg-1ubuntu0.1.22.04.6 broke Arm's pinned pseudo (got *at() syscall for unknown directory). Pinning tar=1.34+dfsg-1build3 and holding the package fixed it.
  • The build ran 6802 tasks in about 26 minutes, with 3797 tasks reused from earlier attempts. The build directory was about 112 GB.

Obstacle 3: serial console​

On a device-tree platform the kernel uses the /chosen/stdout-path UART as a console only when the command line names no console=. The Avocado arm64 boot entry carries console=tty0, which turns that off. The kernel printed until it disabled the early console at 0.6 s and then went silent on all four serial consoles while it kept running. On an ACPI machine the SPCR table adds the UART whatever console= says.

This cause comes from reading the kernel source: a console= argument sets console_set_on_cmdline, and the serial core's of_console_check then returns without adding the stdout-path console. The tree read was a recent mainline one, not the guest's 6.18.52 build.

A QEMU boot of the same 6.18.52 kernel in device-tree mode (direct kernel boot on the virt machine, no disk) reproduces it. With the shipped command line the serial output ends at the line that disables the early console, 0.011 s in, and ttyAMA0 is never enabled. Without console=tty0 the kernel enables ttyAMA0 at 0.19 s and the log continues. With the shipped line plus console=ttyAMA0,115200n8 it does the same. No boot without console=tty0 was run on the chamber.

The fix that was observed on the chamber the first time was adding console=ttyAMA0,115200n8 after console=tty0 in loader/entries/avocado.conf on the first ESP of a copy of the image. The edit used mtools on the raw image, then qemu-img convert -c wrote the result back to QCOW2. The file was uploaded under a new name and linux-disk was pointed at it. loader.conf has editor no, so the arguments cannot be edited at boot.

The kas/feature/pl011-console.yml overlay in meta-avocado is the repeatable form of that edit. The image that booted the first time was built before the overlay existed. The overlay image was booted under QEMU and then on the chamber, as the next section records.

Overlay image on the chamber​

Booted on the chamber with the console overlay: the serial console showed the login prompt and /proc/cmdline held console=ttyAMA0,115200n8.

The power-on job completed with running 287 s after the call (12:44:33Z to 12:49:20Z), and the login prompt appeared 21 s after the console connected. The last kernel timestamp printed before the prompt was 7.195 s. The kernel printed past 0.6 s on the serial console, where the first boot had gone silent. The kernel line read:

Kernel command line: initrd=\initramfs.cpio.zst root=PARTLABEL=rootfs-a ro console=tty0 earlycon pci=pcie_bus_safe console=ttyAMA0,115200n8

The guest kernel was 6.18.52. systemctl is-system-running reported degraded, with only the systemd-growfs instance for /var failed, as in the first boot. The image has no systemd-analyze, so no analyze figure was taken.

The image was built with kas/feature/armv8-userland.yml and kas/feature/pl011-console.yml. No hand edit of the ESP was needed. The hand edit described under the serial console obstacle is the first boot's method, not this one.

Under QEMU both images reached the login prompt, only the console image's kernel command line held the argument, and the same unit failed in both.

Build the image​

The build below is the kas equivalent of the build that was run through a wrapper. It was not re-run with kas directly (not re-run).

kas build --target avocado-stone \
kas/machine/arm64-uefi-v9.yml:kas/target/bringup.yml:kas/target/qemu-provision.yml:kas/feature/armv8-userland.yml:kas/feature/pl011-console.yml

kas/target/bringup.yml builds a development image with an empty root password and openssh installed. The root login on the console needed no password. Keep such an image off a network that others can reach.

The guest reported an Armv8.2-class feature set (no FEAT_LRCPC2), which is why the chain includes kas/feature/armv8-userland.yml. The default Armv9 tune was not tried on this model. It dies with an illegal instruction on the other virtual SoC of the same chamber, see Siemens Innexis ANA virtual platform.

Loading the image​

Convert the provisioned image to a compressed QCOW2 (not re-run in this exact form). The build writes the image as avocado-os-avocado-arm64-uefi-v9.img under tmp/deploy/stone/_build/ in the build directory:

qemu-img convert -c -O qcow2 avocado-os-avocado-arm64-uefi-v9.img avocado.qcow2

Upload it through the Web IDE. Select the target folder first, because uploads go into the selected node and not the node that was right-clicked. The Zena project's software is the subsystem form, with a linux-disk that points at the uploaded file.

  • A 145 MiB compressed QCOW2 uploaded without error.
  • The rebuilt vendor image of 889 MB was built but never uploaded, so no upload size limit is known between 145 MiB (worked) and a 1.9 GiB raw image (failed after at least 746 MB, on the other virtual SoC).

Limits seen​

  • systemctl is-system-running reports degraded, caused only by the systemd-growfs instance for /var (not investigated).
  • The SSH tunnel's Authorize step failed on this project as on the Innexis ANA one, so SSH and GDB access through the tunnel was not available.
  • The platform shuts a project down after 60 minutes by default.