#osfc2024 — Public Fediverse posts
Live and recent posts from across the Fediverse tagged #osfc2024, aggregated by home.social.
-
Three members of Criteo's Hardware Team attended Open Source Firmware Conference 2024 to share their vision and meet members of the community. This article highlights the presentations that resonated the most with us.
-
Prem'Day: feedback on the first infra on-prem conference and creation of a user group by Erwan Velu
Prem'Day was a one-day conference in May from infrastructure users.
Many companies participated: server operators giving talks (Scaleway, i3d, Qarnot, Moji, Criteo), and server vendors (Dell, AMD, Intel HPE, Gigabyte, Supermicro, etc.) listening to all this feedback.
(with yours truly giving the opening talk)
Videos are available at https://www.youtube.com/playlist?list=PLjPeZjanhgqW0P-F40DJYS3Pg0nGJlXUg
Every presentation talked about the firmware ecosystem and how it can be a pain when operating servers.
Having so many server buyers on stage giving feedback about the same issues had a bigger impact towards server vendors, because they can understand that it's a pervasive issue.
The conference was just the starting point: a User Group for Infra owners is the main goal, in order to have a point of contact to ease communication between vendors and their user communities. This user group will own and support the conference, in addition to working on various topics.
The group is currently organized on the Open Source Firmware slack channel #premday, and the next edition of the conference should be in 2025.
-
"Operating system provided device-trees" by Heinrich Schuchardt
While it was not the initial goal, in practice, when booting Linux, device trees need to be coupled with a given kernel version.
This brings its own sets of challenge, because the bootloader needs to patch the device tree to provide information to the OS.
When using EFI, there is now the EFI_DT_FIXUP_PROTOCOL.
Outside of EFI, the flash-kernel tool helps picking the correct device tree depending on the board model.
The way forward would be to use Unified Kernel Images, which would contain all device trees for supported platform, with the efistub used to select the correct one.
-
Next is "OpenBMC - The state of multi-host platform support" by Oliver Brewka
Multi-host in this context means having a single BMC managing multiple host nodes.
First multi-host system Yosemite appeared in 2015; the platform support was maintained in Meta-OpenBMC, and for the latest version it migrated to LF-OpenBMC.
One of the challenges was to go from a static to a dynamic design: the 1<->1 relation is broken, and the number of hosts might change (empty node slot). Changing the design in OpenBMC without breaking the many single-host platforms was the hard part.
As of lately, multi-host has been getting more attention in OpenBMC; Aspeed announced new multi-host capabilities for their next BMC SoC, for example.
Oliver says that multi-host in OpenBMC will improve the design, getting rid of much hard-coding, and getting closer the Redfish specification.
-
We start the last OSFC day with two OCP-related lightning talks: "Open Compute Project Europe and Open Source Firmware Foundation: Intro and collaboration opportunities" by Martin L Roth, Paul Grimes and Raul Alvarez
-
Tom takes over to talk about the firmware update side of the AMC with OpenBMC.
At NVIDIA, they have strong KPIs for firmware update performance and reliability, while still using standard, PLDM type 5 firmware update with Redfish.
-
"Adapting OpenBMC for NVIDIA Platforms: Challenges and Solutions" by Deepak Kodihalli and Tom Joseph
An 8-GPU Tray has an Accelerator Management Controller (AMC), which might generate 3000 metrics per seconds, which might not scale with Redfish while communicating with the BMC.
-
In 2021, the initial protocol support for MCTP was merged into the Linux Kernel.
The goal is to make it standard socket calls to use it with socket(AF_MCTP)/sendmsg/recvmsg, etc.
In the future, more transports will be merged. serial, i2c and i3c are already merged; USB, PCC and PCIe VDM are upcoming.
-
RHF (Rust Hypervisor Firmware) is highly portable. It can run in QEMU, as a coreboot payload, or under Cloud Hypervisor.
Recently, it was integrated with m1n1, Asahi Linux's bootloader as a demo (not ready for upstream support yet).
-
Akira Moroo will now present "Firmware in Rust: More Than Just 'Rewrite It In Rust'"
-
"x86 coreboot with long mode: where are we at and where are we going" by Arthur Heymans and
Patrick Rudolphx86 CPUs do no start in protected mode, it needs to be enabled from 16 bit mode (!). Today's x86 processors are all 16-bit and 32-bit mode compatible.
And coreboot still uses 32-bit mode by default, which means payload must have a 32-bit entrypoint.Long-mode is the (optional) 64-bit version, also called amd64 or x86_64.
-
FPGA-based CI is also run on every repository change request to the OpenTitan repository.
It takes a lot of maintenance to keep the stack of FPGAs working.
This is expensive hardware, the latest ones cost up to $10k for each board.
-
James Wainwright is now presenting "Open source all the way down: developing firmware in parallel with open silicon".
In the context of the OpenTitan open source root-of-trust project, the hardware, its tests and firmware are all developed in lockstep in a monorepo.
-
"An integer overflow caused $370M to vaporize in 40 seconds."
When a rocket exploded in flight because of two simultaneous integer overflow in independently written systems.
Luckily the industry has learned from this, right? Not really, with some Boeing planes having to be rebooted every 51 days to prevent catastrophically wrong information being shown to pilots.
This is Samir's talk on Tock OS, or how to secure firmware at compile time.