Linux Begins Seeing Workarounds For Fujitsu MONAKA CPU Bugs

Last month Fujitsu formally announced their MONAKA CPU as a 144 core Arm-based server processor geared for AI infrastructure. This comes after two years of seeing various Monaka patches to the GCC compiler and other open-source components. Now that MONAKA is announced, we are seeing more patches surface for the Linux kernel with this week seeing several workarounds for MONAKA CPU bugs.

With being an ARM64 CPU design, the Fujitsu MONAKA is largely in good shape already for the Linux support -- especially with being standards-compliant in the server space where not having to mess around with system Device Trees and the like for MONAKA servers. The CPU type detection in the Linux kernel for MONAKA is introduced via this week's patches in order to accommodate workarounds for different bugs.

The first workaround for MONAKA is range-based TLBI operations targeting a VA or IPA may fail to invalidate TLB entries in certain situations. So the Linux kernel needs to issue additional TLBI instructions to ensure the proper TLB invalidation occurs. This is only present if using a 64K kernel for MONAKA, but that is certainly recommended anyhow for performance reasons.

There is also a workaround for another bug where a PMU overflow may be incorrectly detected. In turn this could cause an interrupt storm on MONAKA.

The third workaround this week for MONAKA on Linux is dealing with a possible CPU core hang. It's around an SGI being generated and possibly lost if the operation races with a CPU processing a higher-priority interrupt, an update to a pending interrupt, or a transition of the PE to the sleep state. In this condition the system register write does not complete and could lead to a hang on the issuing CPU core.

Those interested can learn more via this MONAKA patch series.