Well current we are again. Linux 7.3 is out and it’s been an eventful few months, so let’s get stuck in to another Asahi Linux advancement report!
Upstreaming continues
As always, upstreaming activity continues to progress. Aside from the customary patches for low-level SoC blocks, two important drivers made it on to the catalog this cycle: macaudio and avd. Getting macaudio prepared for submission has been a challenge. The audio subsystem on Apple Silicon devices is fairly special, and required fairly a figure of center ALSA changes to correctly assistance it. Apple additionally uses particular variants of speaker and headset amplifiers that were already supported upstream, and these drivers additionally needed some pretty important modifications. We have earlier discussed a lot of this activity at length, and we are eventually in a stance anywhere we can present the chief controller for review. The controller and Devicetree changes portray a hefty chunk of our total patchset, about 3100 lines.
AVD is a entire distinct beast. The AVD controller has spent barely 3 months in our downstream tree and is already on the list. This is our shortest turnaround period yet for a major SoC component and bodes fine for a passage to upstream-first betterment several time in the future. This is awesome news, as the AVD controller and connected Devicetree changes are about 13000 lines of code.
Also making an appearance on the catalog this sequence are intial Devicetrees not lone for M4 Pro and Max devices but additionally M5 gratitude to Yureka’s tireless efforts. As customary these initial Devicetrees are barebones and lone adequate to footwear the SoC to a TTY, nevertheless work is progressing on additional enablement. Watch this space!
We have additionally begun shifting to an upstream-first method anywhere practical. Changes to the TI USB-PD and SPMI drivers required for new devices, as fine as first support for Thunderbolt, were all developed upstream-first. We are so thrilled to see our concentration on upstreaming starting to pay off!
M3 goes GA
Last time, we announced first assistance for M3 sequence machines via the EXPERT emblem on the Asahi installer. We are pleased to province that we received nearly no issues from first adopters, so M3 assistance is now mostly accessible without the EXPERT flag!
We have additionally made a figure of improvements to M3 support, most notably enabling DCP. On top of achievement and power division existence improvements, this additionally allows us to enable sleep on M3 sequence devices and create the area about the notch useable on the MacBooks! Keep an eye out for m1n1 and kernel updates in the coming days.
Crowded House
Last time, we teased assistance for virtualising SPTM-enabled versions of macOS in m1n1, which enables reverse engineering on M4 sequence machines. M5 presents a new challenge, as the versions of macOS which assistance it all necessitate the use of Exclaves.
Exclaves are small applications operating in an secluded surroundings controlled by the Secure Kernel. On M5 sequence SoCs and above, macOS uses them for delicate tasks such as example the “camera in use” indicator to the screen. XNU communicates with the Secure Kernel and its Exclaves via SPTM, using yet another IPC protocol called Tightbeam. Much as alongside SPTM, XNU volition panic if it cannot discover the operating Exclaves. This is all starting to appearance extremely akin in idea to Arm’s own TrustZone feature.
While several Exclaves can be bypassed by changing several settings on the command line, we foretell that this volition not continually be the case. As such, Sven started looking at how we could fit all of this in to m1n1. First row of investigation, try putting the Secure Kernel and Exclave userspace in another VM next to SPTM and macOS.
Oh, that worked, and we’re now up to spawning three VMs in EL1 under m1n1! Maybe we can start selling m1n1 to companies fleeing VMWare… With that, we can now virtualise and reverse engineer macOS on the two M4 and M5 sequence machines.
Hello Triangle(s)!
While M3 assistance is now mostly at characteristic parity alongside M1 and M2, the extremely apparent exception to this is the GPU. Apple’s GPU squad seems to prosecute a tick-tock model, and the M3’s GPU was a extremely important departure from the models established on M1 and M2 SoCs. We mentioned quite several period ago that the M3 GPU was being reverse engineered by a new collection of contributors, nevertheless not much has been stated since.
At XDC this year, Alyssa Millburn, Anna Maniscalco and TellowKrinkle gave a talk on their progress reverse engineering the M3 GPU. Work has progressed far adequate for m1n1 to successfully initialise the GPU and for the squad to diagram a triangle! Fans of the “Think Different” sections of these advancement reports volition discover the conversation edifying; Apple have formerly again made a figure of… engaging choices which volition necessitate important engineering effort to merge into upstream Mesa. It is fine value a observe if you’re interested in additional detail.
A huge gratitude goes out to everyone who has made this activity possible. More progress is on the way, so remain tuned.
Kept you waiting, huh?
All the way rear in January of 2024, we made the mistake of promising Thunderbolt assistance inside the year. It’s a error we won’t quickly repeat. We’ve stated nearly nothing concerning Thunderbolt in the two years since, which begs the question: anywhere is it?
Thunderbolt was initially developed by Apple and Intel, alongside Intel handling the design and fabrication of the hardware required. Since Intel made these controllers available to anyone who wanted them at a (mostly) sensible price, no one else bothered to scheme their own. It is lone extremely lately — mostly coinciding with the transfer of the norm to the USB-IF — that companies akin AMD have decided to rotate their own controllers. Apple is of way no exception.
As part of their shove for insourcing most silicon IP, Apple made the choice to design their own Thunderbolt controller before most others. Surprising no one, the controller is of way a Cortex-M3 operating RTKit coupled alongside a bespoke Thunderbolt/USB4 router. There is one of these for all Thunderbolt-enabled USB-C port on all Apple Silicon Macs.
Since Intel supplies all another extant Thunderbolt controller on the market, the Linux kernel’s Thunderbolt subsystem is efficiently designed about them. It makes a figure of assumptions concerning the hardware topology of the operating machine, including hardware capabilities and admission patterns for certain registers and regions of shared memory. Apple’s controller breaks a lot of these assumptions. Great.
The Thunderbolt controller is additionally oddly coupled to PMGR. The reset row for each controller is asserted or deasserted by environment and clearing a set of bits in the middle of PMGR’s MMIO range. No another hardware obstacle plant fairly akin this, and it took several period to realise that these apparently random writes to PMGR were not fair noise.
Luckily for us, Sven was up to the project of dealing alongside all this nonsense. It just took slightly longer than expected! While the patchset to allow Thunderbolt was sent upstream archetypal (another enormous win for us!), we have additionally pulled it into our downstream kernel and are now shipping it to users!
But there is a catch. Currently, lone USB3, XDomain and Thunderbolt Networking are supported. You can plug in USB devices or nexus two Thunderbolt computers together and create them ping all other, but that’s concerning it. PCIe and DisplayPort tunnelling — required for most Thunderbolt docks — are not supported. Yet. Watch this space.
An enormous amount of gratitude is due to Sven, who has stuck alongside this thankless chore for nearly 5 years in between all of his another amazing work.
“The Holy Grail, Doctor Jones”
Playing retailing PC games on Asahi Linux is a bit of a ache point, to say the least. The application stack required to allow this is incredibly complex. We need to translate x86/amd64 binaries to arm64, and do so accelerated adequate for low-latency gaming. In most cases, we additionally need to agreement alongside the Windows API. On top of that, nearly all games bake in an assumed 4 KiB recollection leaf size to allow certain achievement optimisation. Over the former few years, most of these problems have been solved:
- FEX has improved leaps and bounds. The achievement and reliability of binary translation is barely a deliberation in most cases now;
- WINE is additionally awesome now. Support for ARM64EC is mature and reliable, meaning that FEX now lone has to translate the equivalent binaries fairly than WINE’s complete Windows API implementation; and
- muvm allows us to rotation up a lean VM alongside DRM Native Context to present a 4 KiB recollection leaf size to x86/amd64 applications.
When muvm was archetypal developed, it was alongside the assumption that DRM Native Context would allow near-native GPU achievement from inner the VM. Unfortunately, this assumption did not hold. The muvm achievement penalty is fairly important for a figure of reasons, concerning 20% depending on the workload. Engineering around the restrictions inherent to using a VM additionally introduces issues; muvm cannot use Wayland, has dilatory I/O performance, and struggles alongside certain types of input devices.
This is one area anywhere macOS has us entirely licked. While the achievement gap between FEX and Rosetta has narrowed, XNU allows userspace processes to map themselves at any leaf granularity, so all Rosetta needs to do is inform XNU to map 4 KiB pages for its kid processes. Linux does not have this flexibility. All pages must be mapped at the identical granularity as the kernel. While we did tease Windows games operating exterior of muvm final year, this relied on ugly hacks inside FEX which have since been removed in the interests of maintainability and translation correctness. Bummer. Even so, compatibility was hit-and-miss, alongside a lot of (mostly older) software motionless refusing to work. Double bummer.
Android is currently going through a passage from 4 KiB pages to 16 KiB pages for a assortment of reasons. Google Play now requires apps to be built for 16 KiB systems, and any use updates which do not test successfully on specified systems volition not be approved for publish on the Play Store. What concerning applications which are no longer updated though? What concerning sideloaded endeavor slop apps?
Google engineers Kalesh Singh and Frederick Mayle gave a extremely engaging talk at Linux Plumbers Conference this twelvemonth concerning exactly this. Their resolution is simple: just fix Linux so that userspace can map 4 KiB pages!
Their supplement to the Android kernel is called P3S, or Per Process Page Size. On ARM64 systems, pagetable granularity is controlled by bits set in the Translation Control Registers (TCR). P3S plant by tracking the granularity of all process’s pagetables, and writing the suitable bits to the TCR whenever a procedure alongside a distinct pagetable size is switched to. Because this is a sole enroll write, and since this is a no-op for “native” leaf size processes, the achievement penalty is effectively zero!
While this attempt is chiefly targeted at Android, the patchset mostly rebases cleanly on top of mainline Linux. What happens if we rebase it on top of our Asahi tree (and fix several bugs)? As it turns out:
Okay okay, Battlefront is a extremely uncomplicated Direct3D 9.0 equivalent from 22 years ago. Not exactly an notable demo, right? Well what about:
Wolfenstein: The New Order is among the most fine optimised native Vulkan titles always released, and being 64-bit uses WINE’s ARM64EC pathways to massively decrease the amount of activity FEX has to do. Again, not exactly an impressive demo if we’re judging according to specialized difficulty.
Let’s discover item else. Something… worse. Something 32-bit so that FEX and WINE have to activity harder. Something that volition exercise DXVK too. Something that is so comically buggy and poorly optimised that it doesn’t even activity correctly on genuine Windows machines, and relies on community patches and command row switches to commencement at all on contemporary computers:
That’s additional akin it! But we can do even worse. We can equal run… Bethesda games!
Performance is of way motionless patchy. Both Battlefront and Wolfenstein run basically flawlessly, alongside the second hitting 144+ FPS on an M2 Max Mac Studio. GTA IV on the another hand is… not fantastic. Performance is motionless massively improved complete using muvm, and eliminating the use of the VM in broad should additionally eliminate a few classes of bugs. Debugging the GPU controller should additionally be much easier alongside this approach.
As continually there is a catch, and unfortunately it’s a big one. The current resolution is hard-coded for 16 KiB hosts and 4 KiB processes on those hosts. This evidently satisfies Google’s business requirements, but adding pagesize-specific kludges to the kernel’s center recollection administration code is not suitable upstream. We are unsure if Google aims to generalise this method and upstream it, or if this is going to remain an Android-only patchset. As such, we do not desire to container this in our downstream tree and have folks become reliant on it whenever there is this flat of uncertainty.
It additionally requires userspace integration. P3S plant by checking the binary metadata for its sections’ leaf alignments. This plant for Google’s use case, as P3S is intended as a fallback scheme for applications aligned to 4 KiB pages. Modern arm64 binaries are aligned to 64 KiB however, as this guarantees compatibility alongside systems of all native pagesizes. Thus, a shim is required to power P3S on. The shim triggers P3S, and any child processes volition inherit its 4 KiB pagesize. It is not value spending the attempt on implementing this until we have several certainty about what Google aims to do alongside this patchset.
We have reached out to Google alongside several questions, and we anticipation to have more to say soon.
Mics, but faster
Our tradition beamformer for the microphone gathering established in the MacBooks is written in Rust. Rust’s specialized computing ecosystem is not as mature as C++’s, which is a bit of a issue as beamforming requires realtime FFT and extremely dense matrix operations to be done on buffers filled alongside audio samples. Mature scientific computing libraries have hand-optimised code, frequently using intrinsics tied to the target ISA. We were not doing this however, and relied on the compiler autovectorising code anywhere it could to accomplish our achievement targets.
This cut distinct whenever we eliminated the FFT-based analytic transform, replacing it with an implementation according to infinite encourage reply (IIR) filters. This should be faster, nevertheless the compiler was failing to autovectorise the new implementation. Linkmauve kindly stepped in and added SIMD-optimised code for the two arm64 and amd64. This, coupled alongside other performance optimisations courtesy of Frédéric Bour has caused Triforce becoming nearly twice as fast, all during using small CPU time!
Keeping up
Fedora Linux 45 was released on 3 November, and we shipped Fedora Asahi Remix on the identical day. This is the archetypal period we have been capable to get a new type of Fedora Asahi Remix out on the identical day as upstream!
Learning from history
We obtain awesome pride in our work, and have continually done all of our betterment in the open. We communicate alongside our users, our followers and our financial sponsors as transparently as is humanly possible. Our finances are in the open, our betterment branches are in the open, and we study our advancement regularly and candidly in the open. Our IRC channels and Matrix rooms are harmless spaces, open to anyone alongside equal a passing involvement in Asahi Linux or hacking on Apple Silicon.
We do all of this since we firmly accept in Free and Open Source Software, not fair as a means to an end but as the end in and of itself. The contemporary digital earth is lone possible because of the Free Software movement, and it is additional crucial now than it always has been that we do all we can to defend the Free Software ecosystem and uphold the ideals upon which it is built.
It may astonishment several to study that WINE was initially made accessible under the MIT licence. The 90s and 2000s were a extremely distinct period in the Free Software ecosystem, and permissively licenced projects were viewed by tech companies as liberated genuine estate to pilfer and exploit. WINE was not immune to this, and many monetised proprietary forks of WINE were marketed and sold by assorted bad actors. The WINE betterment team grew increasingly concerned that these proprietary forks, during fine inside their rights to exist, would eventually logic so much fragmentation in the ecosystem that WINE itself would rotate into unviable. In 2002, WINE was relicenced under the LGPL to forestall further proprietary forks from being created, and to advance straightforward upstream contribution.
When Asahi Linux was archetypal started, it was alongside the intent that we create a Free Software ecosystem for the Apple Silicon platform. To facilitate and advance the particiation of our allies at the BSDs and another Free Software functioning systems, we endeavoured to licence as much of our code as imaginable under the permissive MIT licence.
Unfortunately, this is no longer viable. The malicious practices of the 90s and 2000s have reared their ugly heads formerly again, this period driven in ample part by the two the insatiable appetite of frontier AI labs for high-quality training data, and militant animosity toward the Free Software community.
Some period ago, we were made conscious of multiple attempts to utilize our permissive licensing to hijack our work, alongside a straightforward perspective to replacing the project and aggressively monetising Linux on Apple Silicon. In several extreme cases, in-development characteristic code was being lifted direct out of our individual forks and published without attribution or lazily rewritten alongside LLM aid to obfuscate its origin. The MIT licence explicitly allows this, and has no carveouts to defend against bad belief actors.
As such, we made the difficult decision to commencement the procedure of transitioning all of our repositories to the GNU General Public License type 2 or later. Going forward all Asahi Linux code volition be licenced under powerful copyleft licences only. Those who do not study from former are doomed to reiterate it.
Until next time!
As always, we would akin to sincerely appreciate our supporters on OpenCollective and GitHub Sponsors. It is lone through your generous assistance that we are capable to dedicate the period and resources to continuing our activity reverse engineering the phase and bringing to it a world-class, polished, and battle-tested Linux desktop experience. You average the earth to us.
James Calligeros · 2026-11-07
