Apple’s Darwin OS and XNU Kernel Deep Dive

> 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.