> Mach’s virtual memory (VM) system was influential beyond the project – it was adopted by 4.4BSD and later FreeBSD as their memory management subsystem.

…and NetBSD[0], OpenBSD[1], but apparently not DragonFly BSD[2].

[0] https://netbsd.org/docs/kernel/uvm.html

[1] https://man.openbsd.org/OpenBSD-3.0/uvm.9

[2] https://www.dragonflybsd.org/mailarchive/kernel/2011-04/msg0...

Darwin is interesting because of the pace of radical changes to its core components. From dropping syscall backwards compatibility to mandatory code signing to dyld_shared_cache eliminating individual system library files to speed up dynamic executable loading. It's a very results-oriented design approach with no nostalgia and no sacred cows. I think only a big hardware vendor like Apple could pull it off.

The article states that "pager daemons" that manage swap files runs in user space, and the kernel memory can also get swapped out, but never explained how a user space daemon swaps out kernel memory. Do they have hard-coded exceptions for special daemons, or use special system calls? Where can I find out more details about the user space memory management specifically?

Whenever I see the Darwin kernel brought into the discussion I can't help but wonder how different things could have been if Apple had just forked Linux and ran their OS services on top of that.

Especially when I think about how committed they are to Darwin it really paints a poor image in my mind. The loss that open source suffers from that, and the time and money Apple has to dedicate to this with a disproportionate return.

Lots of love and work went into this article, as someone around for most of this history, ported code from NeXTSTEP into Windows, dived into the GNUStep attempts to clone the experience, remembers YellowBox and OpenStep, read the internals books, regular consumer of WWDC content, I would say the article matches pretty much my recollection on most of the systems have evolved.

Jobs tried to hire Torvalds to work on Mac OS X and Linus declined: https://www.macrumors.com/2012/03/22/steve-jobs-tried-to-hir...

I’m not sure if I/O kit was written in this c++ subset just for speed. There was this controversial at the time. Apple announced MacOS X and said that it won’t be compatible with current software. All partners would need to rewrite the software in Objective-C. This didn’t go over well. Apple back paddelt and introduced “carbon”. An API layer for cpp applications as well as “Core Foundation” an underpinning to the objective-c base Framework “Foundation”. Also the reason why we have Obj-c++. The interesting part is that they managed to get the memory management toll free. Means an object allocated in the c/cpp world can be passed to obj-c without extra overhead.

They should have fostered a better FOSS community around XNU, now that they moved to ARM there should have been a runnable distribution for x64

I’ve been wanting to understand Darwin at this depth for a long time. Great read!

That's a good history, but it skips over a lot of the nice security work that really distinguishes Apple's operating systems from Linux or Windows. There's a lack of appreciation out there for just how far ahead Apple now is when it comes to security. I sometimes wonder if one day awareness of this will grow and people working in sensitive contexts will be required to use a Mac by their CISO.

The keystone is the code signing system. It's what allows apps to be granted permissions, or to be sandboxed, and for that to actually stick. Apple doesn't use ELF like most UNIXs do, they use a format called Mach-O. The differences between ELF and Mach-O aren't important except for one: Mach-O supports an extra section containing a signed code directory. The code directory contains a series of hashes over code pages. The kernel has some understanding of this data structure and dyld can associate it with the binary or library as it gets loaded. XNU checks the signature over the code directory and the VMM subsystem then hashes code pages as they are loaded on demand, verifying the hashes match the signed hash in the directory. The hash of the code directory therefore can act as a unique identifier for any program in the Apple ecosystem. There's a bug here: the association hangs off the Mach vnode structure so if you overwrite a signed binary and then run it the kernel gets upset and kills the process, even if the new file has a valid signature. You have to actually replace the file as a whole for it to recognize the new situation.

On top of this foundation Apple adds code requirements. These are programs written in a small expression language that specifies constraints over aspects of a code signature. You can write a requirement like, "this binary must be signed by Apple" or "this binary can be of any version signed by an entity whose identity is X according to certificate authority Y" or "this binary must have a cdhash of Z" (i.e. be that exact binary). Binaries can also expose a designated requirement, which is the requirement by which they'd like to be known by other parties. This system initially looks like overkill but enables programs to evolve whilst retaining a stable and unforgeable identity.

The kernel exposes the signing identity of tasks to other tasks via ports. Requirements can then be imposed on those ports using a userspace library that interprets the constraint language. For example, if a program stores a key in the system keychain (which is implemented in user space) the keychain daemon examines the designated requirement of the program sending the RPC and ensures it matches future requests to use the key.

This system is abstracted by entitlements. These are key=value pairs that express permissions. Entitlements are an open system and apps can define their own. However, most entitlements are defined by Apple. Some are purely opt-in: you obtain the permission merely by asking for it and the OS grants it automatically and silently. These seem useless at first, but allow the App Store to explain what an app will do up front, and more generally enable a least-privilege stance where apps don't have access to things unless they need them. Some require additional evidence like a provisioning profile: this is a signed CMS data structure provided by Apple that basically says "apps with designated requirement X are allowed to use restricted entitlement Y", and so you must get Apple's permission to use them. And some are basically abused as a generic signed flags system; they aren't security related at all.

The system is then extended further, again through cooperation of userspace and XNU. Binaries being signable is a start but many programs have data files too. At this point the Apple security system becomes a bit hacky IMHO: the kernel isn't involved in checking the integrity of data files. Instead a plist is included at a special place in the slightly ad-hoc bundle directory layout format, the plist contains hashes of every data file in the bundle (at file not page granularity), the hash of the plist is placed in the code signature, and finally the whole thing is checked by Gatekeeper on first run. Gatekeeper is asked by the kernel if it's willing to let a program run and it decides based on the presence of extended attributes that are placed on files and then propagated by GUI tools like web browsers and decompression utilities. The userspace OS code like Finder invokes Gatekeeper to check out a program when it's been first downloaded, and Gatekeeper hashes every file in the bundle to ensure it matches what's signed in the binaries. This is why macOS has this slow "Verifying app" dialog that pops up on first run. Presumably it's done this way to avoid causing apps to stall when they open large data files without using mmap, but it's a pity because on fast networks the unoptimized Gatekeeper verification can actually be slower than the download itself. Apple doesn't care because they view out-of-store distribution as legacy tech.

Finally there is Seatbelt, a Lisp-based programming language for expressing sandbox rules. These files are compiled in userspace to some sort of bytecode that's evaluated by the kernel. The language is quite sophisticated and lets you express arbitrary rules for how different system components interact and what they can do, all based on the code signing identities.

The above scheme has an obvious loophole that was only closed in recent releases: data files might contain code and they're only checked once. In fact for any Electron or JVM app this is true because the code is in a portable format. So, one app could potentially inject code into another by editing data files and thus subvert code signing. To block this in modern macOS Seatbelt actually sandboxes every single app running. AFAIK there is no unsandboxed code in a modern macOS. One of the policies the sandbox imposes is that apps aren't allowed to modify the data files of other apps unless they've been granted that permission. The policy is quite sophisticated: apps can modify other apps if they're signed by the same legal entity as verified by Apple, apps can allow others matching code requirements to modify them, and users can grant permission on demand. To see this in action go into Settings -> Privacy & Security -> App Management, then turn it off for Terminal.app and (re)start it. Run something like "vim /Applications/Google Chrome.app/Contents/Info.plist" and observe that although the file has rw permissions vim thinks it's read-only.

Now, I'll admit that my understanding of how this works ends here because I don't work for Apple. AFAIK the kernel doesn't understand app bundles, and I'm not sure how it decides whether an open() syscall should be converted to read only or not. My guess is that the default Seatbelt policy tells the kernel to do an upcall to a security daemon which understands the bundle format and how to read the SQLite permission database. It then compares the designated requirement of the opener against the policies expressed by the bundle and the sandbox to make the decision.

Couldn't Apple have used ZFS instead of inventing APFS? Maybe modifying it to use less physical memory?

It seems like the XNU kernel is architecturally super close to the Mach kernel, and XNU drivers architecturally work like Mach drivers, but just that they are compiled into the kernel instead of running in userspace as a separate process. And it seems like the only reason for doing so is performance.

That makes me wonder: how hard would it be to run the XNU kernel in something like a “Mach mode”, where you take the same kernel and drivers but run them separately as the Mach microkernel was intended?

I feel like from a security standpoint, a lot of situations would gladly call for giving up a little bit of performance for the process isolation security benefits that come from running a microkernel.

Is anybody here familiar enough with XNU to opine on this?

Can someone speak to the below statement from the article? I thought Objective-C did not have a runtime like memory managed languages like C#.

> avoid the runtime overhead of Objective-C in the kernel

From Apple docs[0]:

You typically don’t need to use the Objective-C runtime library directly when programming in Objective-C. This API is useful primarily for developing bridge layers between Objective-C and other languages, or for low-level debugging.

0. https://developer.apple.com/documentation/objectivec/objecti...

At the risk of nitpicking, there are a bunch of things that are not quite right. Nonexhaustive list:

- Discussion of paging mixes together some concepts as I described in [1].

- Mach port "rights" are not directly related to entitlements. Port rights are port of the original Mach design; entitlements are part of a very different, Apple-specific security system grafted on much later. They are connected in the sense that Mach IPC lets the receiver get an "audit token" describing the process that sent them, which it can then use to look up entitlements.

- All IOKit calls go through Mach IPC, not just asynchronous events.

- "kmem" (assuming this refers to the kmem_* functions) is not really a “general-purpose kernel malloc”; that would be kalloc. The kmem_* functions are sometimes used for allocations, but they’re closer to a “kernel mmap” in the sense that they always allocate new whole pages.

- It’s true that xnu can map the same physical pages into multiple tasks read-only, but that’s nothing special. Every OS does that if you use mmap or similar APIs. What does make the shared cache special is that it can also share physical page tables between tasks.

- The discussion about “shared address space” is mixing things up.

The current 64-bit behavior is the same as the traditional 32-bit behavior: the lower half of the address space is reserved for the current user process, and the upper half is reserved for the kernel. This is typically called a shared address space, in the sense that the kernel page tables are always loaded, and only page permissions prevent userland from accessing kernel memory. Though you could also think of it as a 'separate' address space in the sense that userland and kernel stick to separate addresses. Anyway, this approach is more efficient (because you don't have to swap page tables for every syscall) and it's the standard thing kernels do.

What was tricky and unusual was the intermediate 32-bit behavior where the kernel and user page tables actually were completely independent (so the same address would mean one thing in user mode and another thing in kernel mode). This allowed 32-bit user processes to use more memory (4GB rather than 2GB), but at the cost of making syscalls more expensive.

Even weirder, in the same era, xnu could even run 64-bit processes while itself being 32-bit! [2]

- The part about Secure Enclave / Exclaves does not explain the main difference between them: the Secure Enclave is its own CPU, while Exclaves are running on the main CPU, just in a more-trusted context.

- Probably shouldn't describe dispatch queues as a "new technique". They're more than 15 years old, and now they're sort of being phased out, at least as a programming model you interact with directly, in favor of Swift Concurrency. To be fair, Swift Concurrency uses libdispatch as a backend.

[1] https://news.ycombinator.com/item?id=43599230

[2] https://superuser.com/questions/23214/why-does-my-mac-os-x-1...

> This is likely the origin of boosts and throttles that iOS uses to ensure the foreground app gets more CPU than background apps.

The purpose of these is for daemons to inherit the priority of their client apps when doing work for them.

I don't remember what thread groups/workloads are for.

I suppose with unified memory there's no real difference between the kernel and userspace; it's just different security zones.

The MMU era used separate memory spaces to enforce security, but it's probably safer in the log run to actually have secure areas instead of 'accidentslly secure areas" that aren't that secure.

Question for the author, who is here in the comments: for clarification, to what extent is the article a deep dive into the OS itself (e.g., reverse engineering) vs. a deep dive into the extant literature on the OS?

Where is Scott Forstall mark in all of this evolution I wonder. He was responsible for adapting macOS, hence XNU, to iPhone. The most successful Apple product of all times. And where is he now?

Does this give us any indication on when Apple will be ditching x86 support?

I'm a bit disappointed at the lack of interest Apple seems to have on Rust, given their focus on performance, UX and security.

oh interesting

What impresses me most about technical documentation like this is how it structures knowledge into comprehensible layers. This article manages to explain an extremely complex system by establishing clear relationships between components.

I've been experimenting with similar approaches for documentation in open source projects, using knowledge graphs to link concepts and architectural decisions. The biggest challenge is always keeping documentation synchronized with evolving code.

Has anyone found effective tools for maintaining this synchronization between documented architecture and implemented code? Large projects like Darwin must have established processes for this.

The claim is inaccurate and mixes together multiple different things:

- The Mach microkernel originally supported true userland paging, like mmap but with an arbitrary daemon in place of the filesystem. You can see the interface here:

https://web.mit.edu/darwin/src/modules/xnu/osfmk/man/memory_...

But I'm not sure if Darwin ever used this functionality; it certainly hasn't used it for the last ~20 years.

- dynamic_pager never used this interface. It used a different, much more limited Mach interface where xnu could alert it when it was low on swap; dynamic_pager would create swap files, and pass them back into the kernel using macx_swapon and macx_swapoff syscalls. But the actual swapping was done by the kernel. Here is what dynamic_pager used to look like:

https://github.com/apple-oss-distributions/system_cmds/blob/...

But that functionality has since moved into the kernel, so now dynamic_pager does basically nothing:

https://github.com/apple-oss-distributions/system_cmds/blob/...

- The vast majority of kernel memory is wired and cannot be paged out. But the kernel can explicitly ask for pageable memory (e.g. with IOMallocPageable), and yes, that memory can be swapped to disk. It's just rarely used.

Still, any code that does this needs to be careful to avoid deadlocks. Even though userland is no longer involved in "paging" per se, it's still possible and in fact common for userland to get involved one or two layers down. You can have userland filesystems with FSKit (or third-party FUSE). You can have filesystems mounted on disk images which rely on userland to convert reads and writes to the virtual block device into reads and writes to the underlying dmg file (see `man hdiutil`). You can have NFS or SMB connections going through userland networking extensions. There are probably other cases I'm not thinking of.

EDIT: Actually, I may be wrong about that last bit. You can definitely have filesystems that block on userspace, but it may not be supported to put swap on those filesystems.

There was never a right time for Apple to make such a switch. NeXTSTEP predates Linux, and when it was adapted into Mac OS X, Apple couldn't afford a wholesale kernel replacement project on top of everything else, and Linux in the late 1990s was far from being an obviously superior choice. Once they were a few versions in to OS X and solidly established as the most successful UNIX-like OS for consumer PCs, switching to a Linux base would have been an expensive risk with very little short-term upside.

Maybe if Apple had been able to keep classic MacOS going five years longer, or Linux had matured five years earlier, the OS X transition could have been very different. But throwing out XNU in favor of a pre-2.6 Linux kernel wouldn't have made much sense.

Back in the days when Apple acquired NeXT, Linux was undergoing lots of development and wasn't well established. Linux being a monolithic kernel didn't offer the levels of compartmentalization that Mach did.

As things now stand, FreeBSD represents many of the benefits of Darwin and the open source nature of Linux. If you seek a more secure environment without Apple's increasing levels of lock-in, then FreeBSD (and the other BSDs) merit consideration for deployment.

Interestingly enough, Apple did contribute to porting Linux to PowerPC Macs in the mid-1990s under the MkLinux project, which started in 1996 before Apple’s purchase of NeXT later that year:

https://en.m.wikipedia.org/wiki/MkLinux

I don’t think there was any work done on bringing the Macintosh GUI and application ecosystem to Linux. However, until the purchase of NeXT, Apple already had the Macintosh environment running on top of Unix via A/UX (for 68k Macs) and later the Macintosh Application Environment for Solaris and HP-UX; the latter ran Mac OS as a Unix process. If I remember correctly, the work Apple did for creating the Macintosh Application Environment laid the groundwork for Rhapsody’s Blue Box, which later became Mac OS X’s Classic environment. It is definitely possible to imagine the Macintosh Application Environment being ported to MkLinux. The modern FOSS BSDs were also available in 1996, since this was after the settlement of the lawsuit affecting the BSDs.

Of course, running the classic Mac OS as a process on top of Linux, FreeBSD, BeOS, Windows NT, or some other contemporary OS was not a viable consumer desktop OS strategy in the mid 1990s, since this required workstation-level resources at a time when Apple was still supporting 68k Macs (Mac OS 8 ran on some 68030 and 68040 machines). This idea would’ve been more viable in the G3/G4 era, and by the 2000s it would have be feasible to give each classic Macintosh program its own Mac OS process running on top of a modern OS, but I don’t think Apple would have made it past 1998 without Jobs’ return, not to mention that the NeXT purchase brought other important components to the Mac such as Cocoa, IOKit, Quartz (the successor to Display PostScript) and other now-fundamental technologies.

Why would we want more of a monoculture? We've put so many eggs in one basket already. I hope we see more diversity in kernels, not further consolidation.

Taken a different way, it feels similar to suggesting Apple should rebase safari on chromium.

> Whenever I see the Darwin kernel brought into the discussion I can't help but wonder how different things could have been if Apple had just forked Linux

XNU is only partially open sourced – the core is open sourced, but significant chunks are missing, e.g. APFS filesystem.

Forking Linux might have legally compelled them to make all kernel modules open source–which while that would likely be a positive for humanity, isn't what Apple wants to do

1) We are talking about the late 90s, well before Ubuntu, where Desktop Linux was pretty poor in terms of features and polish.

2) Apple had no money or time to invest in rewriting NeXTStep for a completely new kernel they had no experience in. Especially when so many of the dev team was involved in sorting out Apple's engineering and tech strategy as well as all the features needed to make it more Mac like.

3) Apple was still using PowerPC at the time which NeXTStep supported but Linux did not. It took IBM a couple of years to get Linux running.

Diverse systems are more resilient. It's probably a good thing for IT in a general sense, even if it's not the most efficient

Keep in mind they were also looking at BeOS which is more real time and notably not unix/Linux. I wish I lived in the timeline that they went with it as I'm a huge Be fan.

Control is important. Apple has never had to fight with Torvalds or IBM or Microsoft over getting something added to the kernel. Just look at the fiasco when Microsoft wanted to add a driver for their virtualization system to the kernel.

Also, one thing you'll notice about big companies - they know that not only is time valuable, worst-case time is important too. If someone in an open-source ecosystem CAN delay your project, that's almost as bad as if they regularly DO delay your project. This is why big companies like Google tend to invent everything themselves, I.E. Google may have "invented Kubernetes" (really, an engineer at Google uninvolved with the progenitor of K8s - Borg - invented it based on Borg), but they still use Borg, which everyone Xoogler here likes to say is "not as good as k8s". Yet they still use it. Because it gives them full control, and no possibility of outsiders slowing them down.

>Whenever I see the Darwin kernel brought into the discussion I can't help but wonder how different things could have been if Apple had just forked Linux and ran their OS services on top of that.

They have a long history with XNU and BSD. And Linux has s GPL license which might not suit Apple.

>Especially when I think about how committed they are to Darwin it really paints a poor image in my mind. The loss that open source suffers from that, and the time and money Apple has to dedicate to this with a disproportionate return.

They share a lot of code with FreeBSD, NetBSD and OpenBSD. Which are open source. And Darwin is open source, too. So there's no loss that open source suffers.

The world is better with multiple flavors instead of one bloated one.

One must consider the loss of control moving to Linux would bring. Even Google is reconsidering with fuchsia inline to replace Linux on Android.

It would never have worked. So many of the things that owning xnu has made possible would never have happened on top of Linux. The things you can do when you know each and every customer of the stack, and you all belong to the same business with common objectives and leadership direction just can’t be done in the open-source context.

Based on how often they pull in updated bits from FreeBSD (pretty much never), an Apple fork of Linux would be more or less Linux 2.4 today.

I don't know what the loss that open source suffers is in this context?

I don't think Apple would need to spend less time or money on their kernel grafted ontop of Linux 2.4 vs their kernel grafted on top of FreeBSD 4.4

Seen differently, I think it's great that there is yet another kernel being maintained out there.

Imagine if Apple decided to open source Darwin: wouldn't that be a big win for open source?

Mac OS X Internals by Singh is one of my favorite books, such a great in depth examination of Mac OS X circa 10.4. I really wish there was an updated version.

Edit: I see it's even cited at the end of this article. Truly a source for the (macOS) ages.

I've also wanted to understand Windows NT at this depth for a while. Skip the Win32 stuff, and discuss what's underneath it. As I understand Win32 is just one personality; there was also Windows Services for UNIX in the Windows XP days and Subsystem for UNIX-based Applications in Windows Vista. The underlying NT kernel is flexible enough to allow POSIX compliance. That would be an interesting read.

It’s my first time condensing my research notes into a blog post like this, glad you liked it!

I do not think that "security" is the appropriate name for such features.

In my opinion "security" should always refer to the security of the computer owners or users.

These Apple features may be used for enhancing security, but the main purpose for which they have been designed is to provide enhanced control of the computer vendor on how the computer that they have sold, and which is supposed to no longer belong to them, is used by its theoretical owner, i.e. by allowing Apple to decide which programs are run by the end user.

I remember reading back in 2007-2008 that Apple was interested in bringing ZFS support to Mac OS X, but discussions ended once Oracle purchased Sun. This was a bummer; I would’ve loved ZFS on a Mac.

After a cursory Google search, I found this article:

https://www.zdnet.com/article/zfs-on-snow-leopard-forget-abo...

Supporting ZFS in a UNIX kernel requires excessively extensive modifications to the design and implementation of the VMM, namely:

  1. Integration of the kernel's VM with ZFS's adaptive replacement cache which runs in user space – memory pressure cooperation, page accounting and unified memory management. It also requires extensive VM modifications to support ZFS's controlled page eviction, fine-grained dirty page tracking, plus other stuff.

  2. VMM alignment with the ZFS transactional semantics and intent logs – delayed write optimisations, efficient page syncing.

  3. Support for large memory pages and proper memory page alignment – support for superpages (to reduce the TLB pressure and to efficiently map large ZFS blocks efficiently) and I/O alignment awareness (to ensure proper alignment of memory pages to avoid unnecessary copies).

  4. Memory-mapped I/O: different implementation of mmap and support for lazy checksumming for mmap pages.

  5. Integration with kernel thread management and scheduling, co-opertation with VMM memmory allocators.

  6. … and the list goes on and on.

ZFS is just not the right answer for consumer facing and mobile/portable devices due being a heavyweight server design with vastly different design provisions and due to being the answer to a entirely different question.

IIRC that was something they had been working on, but it got axed when ZFS changed hands and licensing became potentially thorny. My memory may be failing me though.

Around the time of Snow Leopard, it was rumored. I assume the Oracle buyout of Sun around the same time had a big part in killing that particular idea.

Also,

> As of iOS 15, Apple even allows virtualization on iOS (to run Linux inside an iPad app, for example, which some developers have demoed), indicating the XNU hypervisor is capable on mobile as well, though subject to entitlement.

Apple definitely does not allow this; in fact the hypervisor code has been removed from the kernel as of late.

These are very helpful, thank you! A lot of good corrections and suggestions were provided in the comments, I’ll verify and correct the portions in the post for future readers.

The AT&T lawsuits are a moot point, as they were all settled in the early 1990s. They are the sole reason why FreeBSD and NetBSD even came into existence – by forking the 4.4BSD-Lite codebase after the disputed code had been eliminated or replaced with non-encumbered reimplementations. Otherwise, we would all be running on descendants of 4.4BSD-Lite today.

Linux has been running uninterruptedly on s/390 since October 1999 (31-bit support, Linux v2.2.13) and since January 2001 for 64-bit (Linux v2.4.0). Linux mainlined PPC64 support in August 2002 (Linux v2.4.19), and it has been running on ppc64 happily ever since, whereas FreeBSD dropped ppc64 support around 2008–2010. Both s/390 and ppc64 (as well as many others) are hardly hobbyist platforms, and both remain in active use today. Yes, IBM was behind each port, although the Linux community has been a net $0 beneficiary of the porting efforts.

I am also of the opinion that licensing is a red herring, as BSD/MIT licences are best suited for proprietary, closed-source development. However, the real issue with proprietary development is its siloed nature, and the fact that closed-source design and development very quickly start diverging from the mainline and become prohibitively expensive to maintain in-house long-term. So the big wigs quickly figured out that they could make a sacrifice and embrace the GPL to reduce ongoing costs. Now, with the *BSD core team-led development, new contributors (including commercial entities) would be promptly shown the door, whereas the Linux community would give them the warmest welcome. That was the second major reason for the downfall of all things BSD.