Asahi Linux Progresses Linux 7.3 Enablement with Custom Thunderbolt Drivers, M5 Hypervisor Isolation, and Upstream Driver Integration
The Asahi Linux project has released its progress report for the Linux 7.3 cycle, detailing major upstream driver submissions, custom Apple Silicon silicon enablement, and m1n1 hypervisor updates. Key developments include General Availability for M3 SoCs, a multi-VM architecture in m1n1 to virtualize macOS M5 Exclaves and Tightbeam IPC, and custom driver support for Apple's Cortex-M3-based Thunderbolt IP.
Upstream Submissions and Platform Enablement Upstreaming efforts for the Linux 7.3 cycle yielded two major driver submissions alongside low-level SoC block patches: `macaudio` (~3,100 lines of code) and `avd` (~13,000 lines including Devicetree additions). The `macaudio` driver required extensive core ALSA changes and modifications to upstream speaker/headset amplifier drivers to accommodate Apple's audio subsystem architecture. The AVD driver spent under three months in the downstream repository before being submitted upstream.
Initial barebones Devicetrees submitted by contributor Yureka enable TTY boot on M4 Pro, M4 Max, and M5 SoCs. Upstream-first development was also applied to TI USB-PD modifications, SPMI drivers, and initial Thunderbolt integration. M3 platform support has transitioned out of the installer's EXPERT flag into General Availability (GA). Enabling the Display Controller Coprocessor (DCP) on M3 devices delivers system sleep states, improved power efficiency, and rendering support for the display notch region.
Hypervisor Execution: SPTM, Tightbeam, and M5 Exclaves Virtualization capabilities within `m1n1` have been expanded to analyze security architecture changes in newer XNU environments. While M4 reverse engineering relied on virtualizing Secure Page Table Monitor (SPTM) state, M5 SoCs introduce Exclaves—isolated execution tasks running under control of the Secure Kernel.
Under XNU, sensitive features like the hardware indicator for camera usage execute inside Exclaves. Communication between XNU, the Secure Kernel, and these Exclaves relies on SPTM via an IPC protocol named Tightbeam. If XNU cannot establish communication with running Exclaves, the kernel panics. To bypass this, `m1n1` now instantiates three distinct execution contexts at EL1. This configuration runs the Secure Kernel and Exclave userspace inside an isolated guest VM alongside SPTM and macOS.
Reverse Engineering Apple Silicon Hardware Apple's M3 GPU architecture departs significantly from M1 and M2 designs. Following initial reverse engineering detailed at XDC by Alyssa Millburn, Anna Maniscalco, and TellowKrinkle, `m1n1` can initialize the M3 GPU hardware block and render basic primitives, achieving initial triangle rasterization. Upstream Mesa driver integration remains in progress.
Apple's custom Thunderbolt controller implementation diverges from standard Intel hardware. Rather than utilizing standardized off-the-shelf controllers, Apple integrated a per-port controller consisting of a Cortex-M3 core running RTKit coupled with a custom Thunderbolt/USB4 router. This architecture breaks assumptions within the Linux kernel Thunderbolt subsystem regarding hardware topology, shared memory access, and register structures. Furthermore, controller reset states are tied directly to bit manipulation within the Power Management Register (PMGR) MMIO space.
The initial Thunderbolt patchset was submitted upstream-first and pulled into the downstream kernel. Current capabilities cover USB3, XDomain connections, and inter-host Thunderbolt Networking. PCIe and DisplayPort tunneling protocols are not yet supported.
Emulation and Memory Granularity Limits Running x86/amd64 binaries on Asahi Linux relies on FEX for translation, ARM64EC WINE for Windows API translation, and `muvm` for managing a thin hypervisor instance with DRM Native Context. This architecture works around the hardcoded 4 KiB memory page size requirement in x86 binaries on hosts running larger page sizes.
Using `muvm` to present a 4 KiB page size to guest processes incurs a throughput penalty of approximately 20% depending on workload. Operating within the VM boundary also introduces technical trade-offs, including the absence of native Wayland support, input handling complexities, and reduced I/O performance. Unlike Linux, which enforces identical page granularity across host kernel space, XNU allows per-process page granularity, enabling Rosetta to execute 4 KiB processes natively. FEX developers have retired previous downstream 4 KiB hack implementations to preserve translation correctness and system stability.