- one is about the format of symbol information in the actual ELF binaries which is only an issue if you are not using the standard libc functions for looking up symbols for some strange reason
- one is an issue that impacts targeting a lower version of glibc from a higher one which is a configuration that was never supported (though usually fails more loudly)
- the last one is a security policy change which is legitimately an ABI break, but mostly impacts programs that have their execstack flags set incorrectly
glibc actually goes to a fair bit of effort to be compatible with old binaries unlike most of the rest of the Linux userspace. The binaries I built for my side project back in 2015 (BlastEm 0.3.0) still work fine on modern Linux and they dynamically link against glibc. This is just a hobby project, not a piece of professional software, and a build from before this JangaFX company even existed works fine.
I find it really bizarre when people talk about Linux binary compat and then complain entirely about glibc rather than the sort of problems that the manylinux project has had to deal with. glibc is one of the few parts of userspace you can depend on. Yes, setting up your toolchain to build against an old glibc on a modern distro is a bit annoying. Sure, if you do something sufficiently weird you might find yourself outside what glibc considers part of their stable ABI. But from where I sit, it works pretty well.
I think the problem in a nutshell is that it's not trivial(?) to build an executable on a modern Linux distro that links against an old glibc version number (and if it is trivial then it needs to be better communicated).
It is actually quite trivial when building with the Zig toolchain since you can simply append the requested glibc version to the target-triple (e.g. `-target aarch64-linux-gnu.2.xx`), but I think this doesn't work with regular clang or gcc (makes one wonder why not when Zig can pull it off).
Just to add some more context. Zig cc is a wrapper around clang. It can handle cross compiling to specific glibc versions. See https://andrewkelley.me/post/zig-cc-powerful-drop-in-replace...
I imagine it would help with the glibc problems they are taking about. Glibc tries to provide a backwards compatible abi.
> I think the problem in a nutshell is that it's not trivial(?) to build an executable on a modern Linux distro that links against an old glibc version number (and if it is trivial then it needs to be better communicated).
I wouldn't say it's trivial, but it's not rocket science either. Basically there are two main approaches. One is to just build inside a chroot or container with a sufficiently old distro inside. This is generally the path of least resistance because your build system doesn't really have to have any awareness of what's going on. You just build normally inside the chroot/container. The main downsides with this approach are that it's kind of wasteful (you have a whole distro's filesystem) and if you want to use a newer compiler than what the old distro in question shipped with you generally have to build it yourself inside said chroot/container.
The other main approach is to use a sysroot. gcc and clang both take an optional --sysroot parameter which is an alternative root for header and library lookups. This lets you use a compiler on the normal host, but old headers and libs. You can also bake this parameter in when compiling gcc (and also I assume clang, but less sure there) if you want a dedicated cross-toolchain.
The PEP-600 [0] Rationale section touches on this a bit. The basic problem is that there are things beyond glibc that would be nice to use from the environment for a number of reasons (security updates, avoiding clashes between multiple wheels that depend on the same lib, etc.), but since most libs outside of glibc and libstdc++ don't really have an ABI policy and the distros don't necessarily have a policy on what libraries are guaranteed to be present you sort of have to guess and hope for the best. While the initial list in PEP-513 [1] was a pretty good guess, one of the libraries chosen (libcrypt.so.1) got dropped in Fedora 30 and replaced with an ABI incompatible version. Crypto libraries are an example of something that's actually important to keep up to date so I find this rather unfortunate.
You can ship all of your of the libraries you use with your executable. This isn't possible to do with glibc. It's the exception which is why it's talked about the most.
> It's the exception which is why it's talked about the most.
It's definitely not the only exception. libgl is another obvious example since different GPUs need different userland code. I would be surprised if there there had never been compat regressions in those.
libgl can be dlopened, glibc can't be. That is the problem. If libgl has some incompatibility, I can try to work around that. If glibc has some incompatibility, my executable won't even launch.
> libgl can be dlopened, glibc can't be. That is the problem.
What exactly prevents this for glibc? I assume you'd need a dlopen equivalent from somewhere for bootstrapping, but are there other issues (like TLS or whatnot)?
Yeah, TLS is one reason. I don't remember the details, but last time I looked into it, glibc abd the loader have some private interface that they use to load the program correctly, and there are no stability guarantees for this interface.
Ideally, the loader and libdl will ship with the kernel, and glibc will have to use the public interface they expose.
There is no distinction between system and program libraries in Linux. We used to pretend there was one before usrmigration, but that was never good to take seriously.
The distro as packager model ensures that everything is mixed together in the filesystem and is actively hostile to external packaging. Vendoring dependencies or static linking improves compatibility by choosing known working versions, but decreases incentive and ability for downstream (or users) to upgrade those dependencies.
The libc stuff in this article is mostly glibc-specific, and you'd have fewer issues targeting musl. Mixing static linking and dlopen doesn't make much sense, as said here[1] which is an interesting thread. Even dns resolution on glibc implies dynamic linking due to nsswitch.
Solutions like Snap, Flatpak, and AppImage work to contain the problem by reusing the same abstractions internally rather than introducing anything that directly addresses the issue. We won't have a clean solution until we collectively abandon the FHS for a decentralized filesystem layout where adding an application (not just a program binary) is as easy as extracting a package into a folder and integrates with the rest of the system. I've worked on this off and on for a while, but being so opinionated makes everything an uphill battle while accepting the current reality is easy.
> adding an application (not just a program binary) is as easy as extracting a package into a folder and integrates with the rest of the system
I have fond memories of installed Warlords Battle Cry 3, Warcraft 3, AOE2 etc. directories on flash drives, distributed to 20+ kids in high school (all using the same key). Good days.
Way off topic but you just reminded me of all the time I spent playing Warlords 3 (not Warlords Battlecry 3, the original Warlords games were turn-based). One cool feature it had that I'm surprised I haven't really seen other turn-based games do is a "play by email" option similar to correspondence chess, except you're just emailing save files back and forth and the game makes importing/exporting the save files via email a bit more streamlined.
Civilization V has a "cloud game" option that keeps the save file online. Players can then take their turn when they have a chance (though non-responsive players can break the game), and if you're running the game through Steam you'll even get a Steam notification when it's your turn. You can also define a webhook that gets called when it's somebody's turn. One of these days I'll put together a little tool that takes that webhook message and translates into a Discord PM or Discord channel post @-ing the person whose turn it is.
They specifically say that it's their way of paying tribute to Civ playing by email.
> Even dns resolution on glibc implies dynamic linking due to nsswitch.
Because, as far as I’ve heard, it borrowed that wholesale from Sun, who desperately needed an application to show off their new dynamic linking toy. There’s no reason they couldn’t’ve done a godsdamned daemon (that potentially dynamically loaded plugins) instead, and in fact making some sort of NSS compatibility shim that does work that way (either by linking the daemon with Glibc, or more ambitiously by reimplementing the NSS module APIs on top of a different libc) has been on my potential project list for years. (Long enough that Musl apparently did a different, less-powerful NSS shim in the meantime?)
The same applies to PAM word for word.
> Mixing static linking and dlopen doesn't make much sense, as said [in an oft-cited thread on the musl mailing list].
It’s a meh argument, I think.
It’s true that there’s something of a problem where two copies of a libc can’t coexist in a process, and that entails the problem of pulling in the whole libc that’s mentioned in the thread, but that to me seems more due to a poorly drawn abstraction boundary than anything else. Witness Windows, which has little to no problem with multiple libcs in a process; you may say that’s because most of the difficult-to-share stuff is in KERNEL32 instead, and I’d say that was exactly my point.
The host app would need to pull in a full copy of the dynamic loader? Well duh, but also (again) meh. The dynamic loader is not a trivial program, but it isn’t a huge program, either, especially if we cut down SysV/GNU’s (terrible) dynamic-linking ABI a bit and also only support dlopen()ing ELFs (elves?) that have no DT_NEEDED deps (having presumably been “statically” linked themselves).
So that thread, to me, feels like it has the same fundamental problem as Drepper’s standard rant[1] against static linking in general: it mixes up the problems arising from one libc’s particular implementation with problems inherent to the task of being a libc. (Drepper’s has much more of an attitude problem, of course.)
As for why you’d actually want to dlopen from a static executable, there’s one killer app: exokernels, loading (parts of) system-provided drivers into your process for speed. You might think this an academic fever dream, except that is how talking to the GPU works. Because of that, there’s basically no way to make a statically linked Linux GUI app that makes adequate use of a modern computer’s resources. (Even on a laptop with integrated graphics, using the CPU to shuttle pixels around is patently stupid and wasteful—by which I don’t mean you should never do it, just that there should be an alternative to doing it.)
Stretching the definitions a little, the in-proc part of a GPU driver is a very very smart RPC shim, and that’s not the only useful kind: medium-smart RPC shims like KERNEL32 and dumb ones like COM proxy DLLs and the Linux kernel’s VDSO are useful to dynamically load too.
And then there are plugins for stuff that doesn’t really want to pass through a bytestream interface (at all or efficiently), like media format support plugins (avoided by ffmpeg through linking in every media format ever), audio processing plugins, and so on.
Note that all of these intentionally have a very narrow waist[2] of an interface, and when done right they don’t even require both sides to share a malloc implementation. (Not a problem on Windows where there’s malloc at home^W^W^W a shared malloc in KERNEL32; the flip side is the malloc in KERNEL32 sucks ass and they’re stuck with it.) Hell, some of them hardly require wiring together arbitrary symbols and would be OK receiving and returning well-known structs of function pointers in an init function called after dlopen.
> Witness Windows, which has little to no problem with multiple libcs in a process
Only so long as you don't pass data structures from one to the other. The same caveats wrt malloc/free or fopen/fclose across libc boundaries still applies.
Well, not anymore, but only because libc is a system DLL on Windows now with a stable ABI, so for new apps they all share the same copy.
That's one of the reasons that OpenBSD is rather compelling. BSDAuth doesn't open arbitrary libraries to execute code, it forks and execs binaries so it doesn't pollute your program's namespace in unpredictable ways.
> It's true that there's something of a problem where two copies of a libc can't coexist in a process...
That's the meat of this article. It goes beyond complaining about a relatable issue and talks about the work and research they've done to see how it can be mitigated. I think it's a neat exercise to wonder how you could restructure a libc to allow multi-libc compatibility, but question why anyone would even want to statically link to libc in a program that dlopen's other libraries. If you're worried about a stable ABI with your libc, but acknowledge that other libraries you use link to a potentially different and incompatible libc thus making the problem even more complicated, you should probably go the BSDAuth route instead of introducing both additional complexity and incompatibility with existing systems. I think almost everything should be suitable for static linking and that Drepper's clarification is much more interesting than the rant. Polluting the global lib directory with a bunch of your private dependencies should be frowned upon and hides the real scale of applications. Installing an application shouldn't make the rest of your system harder to understand, especially when it doesn't do any special integration. When you have to dynamically link anyway:
> As for why you’d actually want to dlopen from a static executable, there’s one killer app: exokernels, loading (parts of) system-provided drivers into your process for speed.
If you're dealing with system resources like GPU drivers, those should be opaque implementations loaded by intermediaries like libglvnd. [1] This comes to mind as even more reason why dynamic dependencies of even static binaries are terrible. The resolution works, but it would be better if no zlib symbols would leak from mesa at all (using --exclude-libs and linking statically) so a compiled dependency cannot break the program that depends on it. So yes, I agree that dynamic dependencies of static libraries should be static themselves (though enforcing that is questionable), but I don't agree that the libc should be considered part of that problem and statically linked as well. That leads us to:
> ... when done right they don't even require both sides to share a malloc implementation
Better API design for libraries can eliminate a lot of these issues, but enforcing that is much harder problem in the current landscape where both sides are casually expected to share a malloc implementation -- hence the complication described in the article. "How can we force everything that exists into a better paradigm" is a lot less practical of a question than "what are the fewest changes we'd need to ensure this would work with just a recompile". I agree with the idea of a "narrow waist of an interface", but it's not useful in practice until people agree where the boundary should be and you can force everyone to abide by it.
I don't understand why they don't just statically link their binaries. First, they said this:
> Even if you managed to statically link GLIBC—or used an alternative like musl—your application would be unable to load any dynamic libraries at runtime.
But then they immediately said they actually statically link all of their deps aside from libc.
> Instead, we take a different approach: statically linking everything we can.
If they're statically linking everything other than libc, then using musl or statically linking glibc will finish the job. Unless they have some need for loading share libs at runtime which they didn't already have linked into their binary (i.e. manual dlopen), this solves the portability problem on Linux.
What am I missing (assuming I know of the security implications of statically linked binaries -- which they didn't mention as a concern)?
And please, statically linking everything is NOT a solution -- the only reason I can run some games from 20 years ago still on my recent Linux is because they didn't decide to stupidly statically link everything, so I at least _can_ replace the libraries with hooks that make the games work with newer versions.
Neither static nor dynamic linking is looking to solve the 20 year old binaries issue, so both will have different issues.
But I think it's easier for me to find a 20 year old ISO of a Red Hat/Slackware where I can simply run the statically linked binary. Dependency hell for older distros become really difficult when the older packages are not archived anywhere anymore.
I've recently had to do this (to bisect when a change introduced a superficial bug into a 20-year-old program). I think "simply run" is viewing Linux of that era through rose-tinted glasses.
Even for simple 2D "Super VGA" you're needing to choose the correct XFree86 implementation and still tweak your Xorg configuration. The emulated hardware also has bugs, since most of the focus is now on virtio drivers.
(The 20-year-old program was linked against libsdl, which amusingly means on my modern system it supports Wayland with no issues.)
Plus there are other advantages that I always say should never be discounted, such as desktop integration, that you get when you can replace the toolkit (dynamic) libraries rather than just "running everything in a VM". Running everything in a VM is a poor excuse for backwards compatibility.
All since ~mid 2005. Unfortunately a tarball I've been looking for was added before then, but replaced before a release was cut (so it's also missing from those archives).
It's interesting to think how a 20 year old OS plus one program is probably a smaller bundle size than many modern Electron apps ostensibly built "for cross platform compatibility". Maybe microkernels are the way.
Old DOS games are often sold (e.g. on GOG or Steam) with an installer and a full copy of DOSBox, so you get an entire operating system plus hardware emulator to run the game. Looks like this adds less than 10 MB to each game installed in Windows.
rpm-ostree rebase ostree-image-signed:registry:<oci image>
rpm-ostree rebase ostree-image-signed:docker://<oci image>
Native containers run on the host and can host normal containers if a container engine is installed. Compared to an electron runtime, IDK how minimal a native container with systemd and podman, and WASM runtimes, and portable GUI rendering libraries could be.
CoreOS - which was for creating minimal host images that host containers - is now Fedora Atomic is now Fedora Atomic Desktops and rpm-ostree. Silverblue, Kinoite, Sericea; and Bazzite and Secure Blue.
> To handle this correctly, each libc version would need a way to enumerate files across all other libc instances, including dynamically loaded ones, ensuring that every file is visited exactly once without forming cycles. This enumeration must also be thread-safe. Additionally, while enumeration is in progress, another libc could be dynamically loaded (e.g., via dlopen) on a separate thread, or a new file could be opened (e.g., a global constructor in a dynamically loaded library calling fopen).
FWIU, ROP Return-Oriented Programming and Gadgets approaches have implementations of things like dynamic header discovery of static and dynamic libraries at runtime; to compile more at runtime (which isn't safe, though: nothing reverifies what's mutated after loading the PE into process space,
after NX tagging or not, before and after secure enclaves and LD_PRELOAD (which some go binaries don't respect, for example).
> auditwheel is a command line tool to facilitate the creation of Python wheel packages for Linux (containing pre-compiled binary extensions) that are compatible with a wide variety of Linux distributions, consistent with the PEP 600 manylinux_x_y, PEP 513 manylinux1, PEP 571 manylinux2010 and PEP 599 manylinux2014 platform tags.
> auditwheel show: shows external shared libraries that the wheel depends on (beyond the libraries included in the manylinux policies), and checks the extension modules for the use of versioned symbols that exceed the manylinux ABI.
> auditwheel repair: copies these external shared libraries into the wheel itself, and automatically modifies the appropriate RPATH entries such that these libraries will be picked up at runtime. This accomplishes a similar result as if the libraries had been statically linked without requiring changes to the build system. Packagers are advised that bundling, like static linking, may implicate copyright concerns
Should a microkernel implement eBPF and WASM, or, for the same reasons that justify a microkernel should eBPF and most other things be confined or relegated or segregated in userspace; in terms of microkernel goals like separation of concerns and least privilege and then performance?
Linux containers have process isolation features that userspace sandboxes like bubblewrap and runtimes don't.
Flatpaks bypass selinux and apparmor policies and run unconfined (on DAC but not MAC systems) because the path to the executable in the flatpaks differs from the system policy for */s?bin/* and so wouldn't be relabeled with the necessary extended filesystem attributes even on `restorecon /` (which runs on reboot if /.autorelabel exists).
Thus, e.g. Firefox from a signed package in a container on the host, and Firefox from a package on the host are more process-isolated than Firefox in a Flatpak or from a curl'ed statically-linked binary because one couldn't figure out the build system.
Container-selinux, Kata containers, and GVisor further secure containers without requiring the RAM necessary for full VM virtualization with Xen or Qemu; and that is possible because of container interface standards.
Linux machines run ELF binaries, which could include WASM instructions
> What's new with Wasker is, Wasker generates an OS-independent ELF file where WASI calls from Wasm applications remain unresolved.*
> This unresolved feature allows Wasker's output ELF file to be linked with WASI implementations provided by various operating systems, enabling each OS to execute Wasm applications.
> Wasker empowers your favorite OS to serve as a Wasm runtime!
Why shouldn't we container2wasm everything? Because (rootless) Linux containers better isolate the workload than any current WASM runtime in userspace.
Dynamic linking obviously has benefits or there would not be any incentive to build them or provide the capacity for them.
The problem is they also have problems which motivates people to statically link.
I remember back in the Amiga days when there were multiple libraries that provided file requesters. At one point I saw a unifying file requester library that implemented the interfaces of multiple others so that all requesters had the same look.
It's something that hasn't been done as far as I am aware on Linux. partially because of the problems with Linux dynamic libraries.
I think the answer isn't just static linking.
I think the solution is a commitment.
If you are going to make a dynamic library, commit to backwards compatibility. If you can't provide that, that's ok, but please statically link.
Perhaps making a library at a base level with a forever backwards compatible interface with a static version for breaking changes would help. That might allow for a blend of bug support and adding future features.
All versions of MUSL prior to 1.2.4 (released less than two years ago) would indeed fail to perform DNS lookups in many common cases, and a lot of programs could not run in MUSL as a result. (I'm not aware of what specific deficiencies remain in MUSL, but given the history even when there are explicit standards, I am confident that there are more.) This wasn't related to dynamic linking though.
Glibc's NSS is mostly relevant for LANs. Which is a lot of corporate and home networks.
The configuration of DNS resolution on Linux is quite complicated [1]. Musl just ignores all that. You can build a distro that works with musl, but a static musl binary dropped into an arbitrary Linux system won't necessarily work correctly.
The easy and conforming way to do that would be to call "getent hosts google.com" and use the answer. But this only works for simple use cases where you just need some IPv4/IPv6 address, you can't get other kinds of DNS records like MX or TLSA this way.
You have to bundle your own resolver into your application. But here's the rub, users expect your application to respect nsswitch which requires loading shared libs which execute arbitrary code. How Go handles this is somewhat awkward. They parse /etc/nsswitch and decide if they can cheat and use their own resolver based on what modules they see[1]. Otherwise they farm out to cgo to go through glibc.
[1] They're playing with fire here because you can't really assume to know for sure how the module 'dns' behaves. A user could replace the lib that backs it with their own that resolves everything to zombo.com. It would be one thing if nsswitch described behavior which was well defined and could be emulated but it doesn't, it specifies a specific implementation.
This blows my mind, that in 2025 we still struggle with a simple task such as "read in a string, parse it, and query a cascade of resolvers to discover it's IP". I just can't fathom how that is a difficult problem, or why DNS still is notorious for causing so much pain and suffering. Compared to the advancements in hardware and graphics and so many other areas.
There are resolvers for not just DNS but for users and other lookups. The list of resolvers is dynamic, they are configured in /etc/nsswitch.conf. The /etc/hosts lookup is part of the system.
Where do the resolvers come from? It needs to be possible to install resolvers separately and dynamically load them. Unless you want to have NIS always installed. Better to install LDAP for those who need it.
I bought 3 linux games on DVD between 2006 and 2016.
I stopped buying linux games and instead started again buying windows games. Because there is no easy way to run them. On the other hand I can just run myst1993 and most of windows games without much hustle via wine.
Wine is linux only stable abi
The most annoying part to me is that Linus et all make not breaking compatibility their #1 goal (only rarely superseded by security and performance concerns) but all of the other libraries on top of the kernel don't seem to care about that at all and break things willy-nilly.
The 2006-engine version of Half-Life 2: Episode 1 runs on Windows 10/11 with no configuration [1], outside of getting Steam to download it. I recall installing The Elder Scrolls IV: Oblivion on a Windows 11 machine, which just needed Directx 9c to run.
Valve puts ongoing effort also on the windows games… yet a sibling comment tried a game from valve instead of trying a game from a CD to do a fair comparison -_-'
This is a really great article about binary compatibility!
I disagree with their idea for fixing it by splitting up glibc. I think it's a bad idea because it doesn't actually fix the problems that lead to compat breakage, and it's bad because it's harder than it seems.
They cite these compat bugs as part of their reasoning for why glibc should be split up:
I don't see how a single one of these would be fixed by splitting up glibc. If their proposed libdl or libthread were updated and had one of these regressions, it would cause just as much of a bug as if a monolithic libc updates with one of these regressions.
So, splitting up glibc wouldn't fix the issue.
Also, splitting up glibc would be super nasty because of how the threading, loading, and syscall parts of libc are coupled (some syscalls are implemented with deep threading awareness, like the setxid calls, threads need to know about the loader and vice-versa, and other issues).
I think the problem here is how releases are cut. In an ideal world, glibc devs would have caught all three of those bugs before shipping 2.41. Big corpos like Microsoft manage that by having a binary compatibility team that runs All The Apps on every new version of the OS. I'm guessing that glibc doesn't have (as much of) that kind of process.
Windows having multiple C libraries has its own pain points, in particular it's difficult to ship binary libraries that return allocated memory to their consumer (you either need to have the library consumer allocate the memory, which probably explains why so many Win32 APIs have this behaviour, or allow alloc/free functions to be registered). Not to mention different C libraries having their own file handle, TLS, etc state. Unsurprisingly Microsoft now ships the Universal CRT (UCRT) as part of Windows.
Having consumer allocate the memory where possible has other benefits in that the consumer can use more efficient allocation techniques (stack, arenas etc), and given that Win32 API design goes all the way back to the earliest version of Windows running on very resource-limited PCs, that would be the main reason why.
However, in cases where this wasn't consistently feasible - e.g. COM - Windows instead mandates the use of a common API to manage memory: CoGetMalloc etc.
1. Zig's toolchain statically links with musl libc, producing binaries that depend only on the Linux kernel syscall ABI, not any specific glibc version.
2. This eliminates all the symbol versioning nightmares (`GLIBC_2.xx not found`) that plague distribution across different Linux systems.
3. Vulkan provides a standardized, vendor-neutral GPU API that's supported across all modern graphics hardware, eliminating driver-specific dependencies.
4. The resulting binary is completely self-contained - no external library dependencies, no version mismatches, no containerization needed.
5. You get forward AND backward compatibility - the binary will run on both ancient and cutting-edge distros without modification.
The only real limitation is for NVIDIA CUDA-specific workloads, where you'd still need their proprietary stack.
Furthermore, for those who prefer a higher-level API, Zig CC + WebGPU offers similar benefits with a simpler programming model, though with slightly less mature implementations and a possible small performance overhead.
> More importantly, separating the dynamic linker from the C library itself would allow multiple versions of libc to coexist, eliminating a major source of compatibility issues. This is exactly how Windows handles it, which is one of the reasons Windows maintains such strong binary compatibility. You can still run decades-old Windows software today because Microsoft doesn’t force everything to be tied to a single, ever-changing libc.
One of the questions of multiple versions on the same box is what about security issues of those older versions...
It is, in the sense that the package is bigger, and the package ships "everything it needs".
It isn't in the sense that those parts can be updated independently as long as the DLL interface is backward compatible.
For example, I ship OpenSSL dlls with my app. Which means swapping in a later (compatible) OpenSSL can be done (by the user if necessary.)
If I'm making a small utility I static link it - and I still use utilities daily I compiled 25 years ago. Obviously those dynamically link to KERNEL etc, but Microsoft has insane levels of compatibility there.
And perhaps that's the fundamental root of the issue. Windows has one provider, very committed to the longevity of software. Linux, well, does not.
That's OK. The world has room for different philosophies. And each one will have strengths and weaknesses.
Isn't the compatibility promise what syscalls essentially are?
Historically, they (almost) never break and they are steadily incremented to prevent overlaps in differences of parameters.
As WASI is also implementing syscalls for WASM, I'd argue that the binary format doesn't really matter as long as it's using the same syscalls in the end.
I understand this topic is mostly focussing on glibc/muslc problems, but if you want to develop stable software, use CGo free Go binaries. They likely will run in 10 years the same way they do today.
C ABI compatibility is a mess on linux mostly because upstream maintainers don't give a damn about semantic versioning. Just take a look at the SO file headers, and how they differ from upstream "semantic" versions of the library. As long as shared objects differ in versions due to breaking changes, and as long as the C ecosystem doesn't enforce correct versioning, this won't change.
There’s no reason to believe that widespread Linux adoption would not irrevocably damage the experience. It would end up looking something like Android. The same thing happened to the internet. It was ruined by mass adoption. You can call this gatekeeping, but nothing positive has come from getting most of the world onto social media.
I'm inclined to agree with you here. As much as I'd love to see native apps from some big vendors (Adobe, in particular) if what it takes to get there is mass adoption of Linux on the desktop, I'm not sure I want it.
In a way, we already have it in the form of ChromeOS, and I certainly don't want ChromeOS or Android to be the default experience, or for the community to start to cater to that model.
All of the things Linux would need to become to get Windows level marketshare would strip away what makes it appealing in the first place, in particular, full user control. I don't want a nanny desktop, but that's exactly what it would become.
Linux can stay niche for those who appreciate it as it is.
> GLIBC is an example of a "system library" that cannot be bundled with your application because it includes the dynamic linker itself. This linker is responsible for loading other libraries, some of which may also depend on GLIBC—but not always.
Running WordPerfect on modern Linux is done by shipping both of those components:
Wow. This is huge. This patch doesn’t risk weird runtime errors due to missing polyfills or anything? Does it risk polyfilling irrelevant functions that aren’t referenced and bloating?
golang seems to manage being 100% statically linked (unless you go out of your way, by using plugins). It just forgoes the stuff that the article mentions glibc doing dynamic linking for, and does it the simple and direct way.
What do you use golang for? If your answer is only web-dev and if you don't touch anything GUI, system libraries, PAM or VPNs, you simply haven't compiled something complex enough.
This is simply too complex. There are quite a bit dependencies on Glibc and other GNU-based behavior.
What you're actually asking from them is creating a new Android or Chrome OS i.e. recreating the entire userspace. "Just split up musl" will not work. How the system headers work needs to be redefined. This is basically a complete redefinition of entire API of all Linux programs and libraries. Which library will supply unistd.h or any of the sys/whatever.h? Any interdependencies? I think even POSIX itself can be hostile against this. So the fix will have to deviate a lot from that (that's probably why Windows cannot be made POSIX easily either).
That's a surprisingly hard nut to crack when containers won't work for your use case. We found https://github.com/silitics/rugix to work well in that situation.
Really the title should be something like “the difficulty of releasing binaries on Linux and how to work around it.” It isn’t really an atrocious state, the goal of a distro should be to get useful software into the repos. Software not distributed in an open source format doesn’t really help there.
I think the problem is that people treat Linux as OS instead of just a kernel.
You should assume every Linux distro is a different OS so when you are shipping your app for Linux you're actually shipping your app for Debian, Fedora, Ubuntu etc
This is definitely the root of it. Linux isn't the OS, your distro is.
Somewhere along the way we seem to have collectively forgotten that the distro, via it's package manager, is responsible for delivering software, ensuring compatibility, and integrating it all - at least for open source software. That was one of Debian's biggest selling points "back in the day" - that they had a huge repository and big team of maintainers packaging everything under the sun so it works on Debian.
Now we have app image, flatpak, snap, etc. All means of getting developers to ship binaries directly to users, bypassing the package managers, instead of shipping code to distros, but we shouldn't expect devs to ship to the 100s of different OSes in this ecosystem. They ship the code, distros pick it up (or don't) and make it work on their distro.
The shift to direct-to-user binaries doesn't make sense for OSS. It should be packaged by the distro, not the dev (of course, the devs can help with packaging if they are so inclined).
Obviously this doesn't apply to proprietary software, but like you said, you can just target on of the LTS enterprise distros and get a well-defined, stable base (for 10 years with Ubuntu and Red hat).
That's the sane way to tackle it. If you're the vendor, just target the top N (whatever value of N you can cope with).
I don't mean disrespect towards people running Alpine (hi), Arch, or Gentoo, but you wouldn't be running these distros if you aren't ready to handle their quirks.
TFA mostly talks about binary compat. Even if you can get away with statically linking everything, you still have to cope with the mess that is userspace fragmentation: <https://tailscale.com/blog/sisyphean-dns-client-linux>
So yeah, supporting the top N gets you approximately sqrt(N/(N+1))% of the way. (Assuming desktop Linux market share is about 1%.)
>I don't mean disrespect towards people running Alpine (hi), Arch, or Gentoo, but you wouldn't be running these distros if you aren't ready to handle their quirks.
You'd think so, but you'd be surprised how many gamers switching from Windows to Linux get recommended Arch and Manjaro
This article missed a critical point which is "the right way" to select a glibc ABI version: see binutils ld documentation, second part of the page related to VERSION support.
This must include glibc internal symbols.
This will allow to craft ELF binaries on a modern distro which will run on "older" distros. This is critical for games and game engines. There is an significant upfront only-once work in order to select an "old" glibc ABI.
The quick and dirty alternative being having a toolchain configured to link with an "old" glibc on the side.
This article missed the -static-libstdc++ critical option for c++ applications (the c++ ABI is hell on earth), but did not miss the -static-libgcc and the dynamic loading of system interface shared libs.
Is there a reason glibc can't just do a better job at keeping some legacy symbols around? It's not like it's big stuff. They're things like legacy string functions. We're talking a few kilobytes of code in most cases.
The Linux kernel goes to a lot of effort to not break user space, at least for non-exotic core features and syscalls. It seems like a lot of user-space in Linux-land does not make the same effort.
It's particularly bad when it's the C library doing this, since that's at the center of the dependency graph for almost everything.
The problem is the opposite: they are trying to run executables built using a newer glibc in a system that has an older glibc. glibc keeps all the old function definitions since practically forever.
Frankly, I do not understand who would think glibc symbols themselves would be the challenge in this case. Even if you statically link glibc there's zero guarantee the syscalls will be present in the older Linux (cue .ABI-tag failures). Or even damn ELF format changes (e.g. gnu-style hashes). The simple solution is to build in the older Linux (&glibc).
In my long experience with ancient binaries, glibc has almost never been the problem, and its ability to _run_ ancient binaries is all but excellent; even Linux is more of a problem than glibc is (for starters paths to everywhere in /proc, /sys change every other half-decade).
Ehm... you _can_ use the older toolchain even in the newer "Linux", in the same way you have to use a older Visual Studio to target Windows 8 from Windows 10.
So maybe the difference here is between can and actually doing it?
I tend to stay on the oldest supported version of Windows until they drop support and haven't ever seen an application that wouldn't run because it's built on a newer version of Windows.
I am saying that compiler toolchains on Linux should never ever under any circumstances ever rely on anything on the system for compiling. Compiling based on the system global version of glibc is stupid, bad, wrong, and Linus should be ashamed for letting it happen.
It should be trivial for Windows to cross-compile for Linux for any distro and for any ancient version of glibc.
> a lot of user-space in Linux-land does not make the same effort
I believe that, what the article misses is that glibc is maintained and extended with an entirely different community and development model. Windows remains compatible over decades because Microsoft (a) is the sole distributor, and (b) puts an immense effort towards backwards compat. In Linux userspace, it's simply a non-goal across distributions. If you want to ship a binary for a particular distro, you need to build the binary on / for that distro; even within a distro, a major release bump (or especially a major release downgrade) may break a binary.
Ultimately, it's a consequence of Conway’s Law. Microsoft is the sole distributor of Windows, so they can enforce compatibilty with an iron fist, and there are people working for Microsoft whose pay depends on said compatibility. With "Linux" in general, there is no common authority to appeal to, and (again) most vendors don't even promise a seamless userspace upgrade path from one major release to another.
This is unfixable; it will never change -- as long as independent parties are permitted to distribute different operating systems yet call them all "Linux".
Ship multiple binaries, or distribute the source code (and let users build it).
EDIT: you'll notice that "ship multiple binaries" is what distros (especially commercial distros) do. They maintain separate branches, backport fixes to old branches, and employ software maintenance engineers to focus on this kind of work. If you want to target multiple major releases, this is what you have to do, too.
If you (as a commercial ISV) target a commercial distro with long-term support, and can convince your users to use / license the same distro, you'll have a good, stable development experience. You only need to port like once every decade, when you jump major releases.
The Linux user base / the Linux market is fragmented; that's the whole goal. The technical proliferation / inconsistency is just a consequence. Unless you take away the freedom of users to run their own flavors of "Linux", there won't be a uniform Linux target.
In a way, it's funny to even expect otherwise. Why do you expect to ship the same binaries when the foundations are diverse, with no standardization authority that all Linux distributors recognize as such? And even POSIX is an API spec, not an ABI spec.
And, any authority that controls binary aspects will immediately accrue political capital. This is exactly what shouldn't happen in Linux. The fact that anyone can fork (or start) a distro, and contribute to the chaos, is good for freedom.
> If you (as a commercial ISV) target a commercial distro with long-term support, and can convince your users to use / license the same distro, you'll have a good, stable development experience. You only need to port like once every decade, when you jump major releases.
If things go well, it's even better than that: If you target ex. RHEL 8, there's a very good chance that your binaries will work on RHEL 9 and a decent shot at RHEL 10 with zero changes (though of course you should test all versions you want to work). And the same for Ubuntu 20.04/22.04/24.04/... and Debian/SUSE/whatever. Backwards incompatibilities can happen, but within a single stable distro they're not super common so the lazy ISV can probably only really port forward after more than a decade if they really want.
(Incidentally, this isn't a hypothetical: I once had the joy of working on software that targeted RHEL 5, and those binaries ran on RHEL/CentOS 7 without any problems.)
> Ship multiple binaries, or distribute the source code (and let users build it).
And that's why we have package managers and distro maintainers/packagers. You'll get no help from the community if your stuff is proprietary, just the way it is. Ship the code, distros will pick it up and do the packaging for you to make it available in their distro. It's part of the free software culture that surrounds the Linux ecosystem.
If you absolutely must ship proprietary software, then target an enterprise distro. Ship it for RHEL or Ubuntu LTS and you get, at least, 10 years of a stable base.
This theory would check out if a proprietary vendor could easily get away with shipping a single binary package for all supported versions of, say, Ubuntu.
Having to build and maintain a binary packege separately for each version of the same distro probably isn't that appealing to them.
Rust toolchain has already a bug since 2015: it cannot link statically libgcc, namely it does not have "-static-libgcc" option since 2015.
I got the bug with "TinyGlade" video game (extremely good BTW), which is written in rust, and with the dev we hit that bug. Namely... better have a libgcc with the right ABI... and I can tell you, this has a been a HUGE issue since valve started to distribute games more than a decade ago.
Here's a thought: just distribute source code. ABI issues should be mostly fixed. Most computers can compile source code fast enough for the user not to notice and cache the results so that it's never a problem again. If you want optimised code, you can do a source to source optimisation then zip and minify the file. You could compile such a library to approximately native speeds without much user-end lag using modern JIT methods, and maybe even run LTO in a background thread so that the exectuables outdo dynamically linked ones.
I hate compiling. 9/10 something goes wrong. Sometimes I can fix it, other times I just abandon the effort and use something else. These days I just use packages or docker images and if that doesn't work out I'm moving on, ain't nobody got time for this. You really can't expect people who just want to use their computers and don't even know what a compiler is to get involved in a process like that.
More than 20 years, with a single-core Pentium 4, it could take indeed something like 3 days of continuous compilation to compile an entire Gentoo distribution, in order to have a personal computer with every application that one might want.
However, already after the appearance of the first dual-core AMD Athlon64, 20 years ago, that time could be reduced to not much more than a half of day, while nowadays, with a decent desktop CPU from 5 years ago, most Gentoo packages can be compiled and installed in less than a minute.
There are only a few packages whose compilation and installation can take a noticeable time, of up to tens of minutes, depending on the chosen options and on the number of cores of the CPU, e.g. Firefox, LibreOffice, LLVM.
There is only a single package whose compilation may take ages unless you have an expensive CPU and enough memory per core: Google Chromium (including its derivatives that use the same code base).
Not Gentoo. The system I'm suggesting is one where you replace the linker with one which accepts a platform agnostic assembly code with similar semantics to C, and perhaps some additional tools for stack traversal to allow the implementation of garbage collected languages and exceptions. Compiling would be fast because the code you distribute would have already had optimisations applied to it in a pre-compilation step. It would also be automatic in the same way the present linker system is automatic.
The idea is that software would only be shipped if it actually works, and part of this would be ensuring it compiles. I'm not suggesting we use Make files or some other foul approach like that. Code should ideally be distributed as a single zip file with an executable header that triggers it to be unpacked in some cache directory by a program analogous to the linker, and compiled by the simple process of compiling all source files. Additionally this could be sped up drastically by having the compiler be a single executable with uniform flags across files. Header files could be parsed once and shared between build threads in parallel, which alone would safe much work on the part of the compiler. If this still proves insufficient, a language with superior semantics to C could be chosen which can be compiled faster. Most code these days shipped is JS which is compiled on the user's computer. This would be an order of magnitude improvement on that from a performance perspective.
The compilation of firefox could take a few hours on some laptop dual-core Skylake CPU from 10 years ago.
Nowadays, on any decent dektop CPU with many cores the compilation of Firefox should take significantly less than an hour, though it remains one of the handful of open-source applications with a really long and non-negligible compilation time.
The Linux kernel is normally compiled much faster than Firefox, except when one would enable the compilation of all existing kernel modules, for all the hardware that could be supported by Linux, even if almost all of that is not present and it would never be present on the target computer system.
Just because you are used to slow compilers doesn't mean fast ones are impossible. As I said in the original post, code optimisation can be done before compilation on the developers machine, so all that need be done on the target is a simple debug build. An example is Jonathon Blow's JAI compiler which compiles around 250,000 lines of code per second. Even on very slow hardware this is reduced to perhaps 80,000LoC/s, two minutes for 10 million lines of code. Most users would be tolerant of an "install" box with a progress bar that takes at most two minutes for only the largest application on the computer. Or perhaps the browser people would opt not to distribute their software in this way. Advanced software may have advanced needs, beyond that of the average application. Most code is already written in Javascript these days, meaning it begins life being interpreted and is JIT compiled on the user's computer. Few seem to notice the compilation but many complain about the general sluggishness.
If you offered users a deal: spend 1 minute installing the software and it is snappy and responsive when you use it, I suspect all would accept.
I thought you were advocating "just distribute source code" – JAI is a closed-source language that, in its decade of development, has never been used for a significant project.
I fail to see how that relates to my using it as an example of a fast compiler. For that matter, my suggested approach doesn't need to to release source code, just some kind of programming language code. You could minify it, pre-optimise it, or do any number of things that would mean it isn't source code but instead becomes derived code (as I said in my original comment). Distributing textual code is not the same as being open source. I'm really just saying you should distribute something that isn't a platform specific blob with a fixed binary interface. You could, for instance, use LLVM IR code to achieve this, which is little more than a kind of portable assembly. LLVM can apply heavy optimisations to this code on the developer's machine, then take only the last step of code generation on the user's machine.
Absolutely ridiculous that scrollbars are disabled for this page. The kids are too cool for such pedestrian web parts! Grow up and stop forcing your shitty UX on people.
Edit: Found it - it's black on black - even worse!
So of the 3 glibc issues they link
- one is about the format of symbol information in the actual ELF binaries which is only an issue if you are not using the standard libc functions for looking up symbols for some strange reason
- one is an issue that impacts targeting a lower version of glibc from a higher one which is a configuration that was never supported (though usually fails more loudly)
- the last one is a security policy change which is legitimately an ABI break, but mostly impacts programs that have their execstack flags set incorrectly
glibc actually goes to a fair bit of effort to be compatible with old binaries unlike most of the rest of the Linux userspace. The binaries I built for my side project back in 2015 (BlastEm 0.3.0) still work fine on modern Linux and they dynamically link against glibc. This is just a hobby project, not a piece of professional software, and a build from before this JangaFX company even existed works fine.
I find it really bizarre when people talk about Linux binary compat and then complain entirely about glibc rather than the sort of problems that the manylinux project has had to deal with. glibc is one of the few parts of userspace you can depend on. Yes, setting up your toolchain to build against an old glibc on a modern distro is a bit annoying. Sure, if you do something sufficiently weird you might find yourself outside what glibc considers part of their stable ABI. But from where I sit, it works pretty well.
I think the problem in a nutshell is that it's not trivial(?) to build an executable on a modern Linux distro that links against an old glibc version number (and if it is trivial then it needs to be better communicated).
It is actually quite trivial when building with the Zig toolchain since you can simply append the requested glibc version to the target-triple (e.g. `-target aarch64-linux-gnu.2.xx`), but I think this doesn't work with regular clang or gcc (makes one wonder why not when Zig can pull it off).
Just to add some more context. Zig cc is a wrapper around clang. It can handle cross compiling to specific glibc versions. See https://andrewkelley.me/post/zig-cc-powerful-drop-in-replace... I imagine it would help with the glibc problems they are taking about. Glibc tries to provide a backwards compatible abi.
> I think the problem in a nutshell is that it's not trivial(?) to build an executable on a modern Linux distro that links against an old glibc version number (and if it is trivial then it needs to be better communicated).
I wouldn't say it's trivial, but it's not rocket science either. Basically there are two main approaches. One is to just build inside a chroot or container with a sufficiently old distro inside. This is generally the path of least resistance because your build system doesn't really have to have any awareness of what's going on. You just build normally inside the chroot/container. The main downsides with this approach are that it's kind of wasteful (you have a whole distro's filesystem) and if you want to use a newer compiler than what the old distro in question shipped with you generally have to build it yourself inside said chroot/container.
The other main approach is to use a sysroot. gcc and clang both take an optional --sysroot parameter which is an alternative root for header and library lookups. This lets you use a compiler on the normal host, but old headers and libs. You can also bake this parameter in when compiling gcc (and also I assume clang, but less sure there) if you want a dedicated cross-toolchain.
Is there a good summary of the problems manylinux had to deal with?
The PEP-600 [0] Rationale section touches on this a bit. The basic problem is that there are things beyond glibc that would be nice to use from the environment for a number of reasons (security updates, avoiding clashes between multiple wheels that depend on the same lib, etc.), but since most libs outside of glibc and libstdc++ don't really have an ABI policy and the distros don't necessarily have a policy on what libraries are guaranteed to be present you sort of have to guess and hope for the best. While the initial list in PEP-513 [1] was a pretty good guess, one of the libraries chosen (libcrypt.so.1) got dropped in Fedora 30 and replaced with an ABI incompatible version. Crypto libraries are an example of something that's actually important to keep up to date so I find this rather unfortunate.
[0] https://peps.python.org/pep-0600/
[1] https://peps.python.org/pep-0513/
>then complain entirely about glibc
You can ship all of your of the libraries you use with your executable. This isn't possible to do with glibc. It's the exception which is why it's talked about the most.
> It's the exception which is why it's talked about the most.
It's definitely not the only exception. libgl is another obvious example since different GPUs need different userland code. I would be surprised if there there had never been compat regressions in those.
libgl can be dlopened, glibc can't be. That is the problem. If libgl has some incompatibility, I can try to work around that. If glibc has some incompatibility, my executable won't even launch.
I agree with this article completely.
> libgl can be dlopened, glibc can't be. That is the problem.
What exactly prevents this for glibc? I assume you'd need a dlopen equivalent from somewhere for bootstrapping, but are there other issues (like TLS or whatnot)?
Yeah, TLS is one reason. I don't remember the details, but last time I looked into it, glibc abd the loader have some private interface that they use to load the program correctly, and there are no stability guarantees for this interface. Ideally, the loader and libdl will ship with the kernel, and glibc will have to use the public interface they expose.
If that's the case then it feels like it proves they just don't care enough about compatibility...
> then complain entirely about glibc
That’s because when you’re trying to ship a product glibc is one of the primary sources of pain.
If you have 3 hours or less of development experience on linux.
There is no distinction between system and program libraries in Linux. We used to pretend there was one before usrmigration, but that was never good to take seriously.
The distro as packager model ensures that everything is mixed together in the filesystem and is actively hostile to external packaging. Vendoring dependencies or static linking improves compatibility by choosing known working versions, but decreases incentive and ability for downstream (or users) to upgrade those dependencies.
The libc stuff in this article is mostly glibc-specific, and you'd have fewer issues targeting musl. Mixing static linking and dlopen doesn't make much sense, as said here[1] which is an interesting thread. Even dns resolution on glibc implies dynamic linking due to nsswitch.
Solutions like Snap, Flatpak, and AppImage work to contain the problem by reusing the same abstractions internally rather than introducing anything that directly addresses the issue. We won't have a clean solution until we collectively abandon the FHS for a decentralized filesystem layout where adding an application (not just a program binary) is as easy as extracting a package into a folder and integrates with the rest of the system. I've worked on this off and on for a while, but being so opinionated makes everything an uphill battle while accepting the current reality is easy.
[1] https://musl.openwall.narkive.com/lW4KCyXd/static-linking-an...
> adding an application (not just a program binary) is as easy as extracting a package into a folder and integrates with the rest of the system
I have fond memories of installed Warlords Battle Cry 3, Warcraft 3, AOE2 etc. directories on flash drives, distributed to 20+ kids in high school (all using the same key). Good days.
Way off topic but you just reminded me of all the time I spent playing Warlords 3 (not Warlords Battlecry 3, the original Warlords games were turn-based). One cool feature it had that I'm surprised I haven't really seen other turn-based games do is a "play by email" option similar to correspondence chess, except you're just emailing save files back and forth and the game makes importing/exporting the save files via email a bit more streamlined.
PBEM was not uncommon in games from that era. The other title I can think of that had the same feature built-in was Age of Wonders.
Civilization V has a "cloud game" option that keeps the save file online. Players can then take their turn when they have a chance (though non-responsive players can break the game), and if you're running the game through Steam you'll even get a Steam notification when it's your turn. You can also define a webhook that gets called when it's somebody's turn. One of these days I'll put together a little tool that takes that webhook message and translates into a Discord PM or Discord channel post @-ing the person whose turn it is.
They specifically say that it's their way of paying tribute to Civ playing by email.
> Even dns resolution on glibc implies dynamic linking due to nsswitch.
Because, as far as I’ve heard, it borrowed that wholesale from Sun, who desperately needed an application to show off their new dynamic linking toy. There’s no reason they couldn’t’ve done a godsdamned daemon (that potentially dynamically loaded plugins) instead, and in fact making some sort of NSS compatibility shim that does work that way (either by linking the daemon with Glibc, or more ambitiously by reimplementing the NSS module APIs on top of a different libc) has been on my potential project list for years. (Long enough that Musl apparently did a different, less-powerful NSS shim in the meantime?)
The same applies to PAM word for word.
> Mixing static linking and dlopen doesn't make much sense, as said [in an oft-cited thread on the musl mailing list].
It’s a meh argument, I think.
It’s true that there’s something of a problem where two copies of a libc can’t coexist in a process, and that entails the problem of pulling in the whole libc that’s mentioned in the thread, but that to me seems more due to a poorly drawn abstraction boundary than anything else. Witness Windows, which has little to no problem with multiple libcs in a process; you may say that’s because most of the difficult-to-share stuff is in KERNEL32 instead, and I’d say that was exactly my point.
The host app would need to pull in a full copy of the dynamic loader? Well duh, but also (again) meh. The dynamic loader is not a trivial program, but it isn’t a huge program, either, especially if we cut down SysV/GNU’s (terrible) dynamic-linking ABI a bit and also only support dlopen()ing ELFs (elves?) that have no DT_NEEDED deps (having presumably been “statically” linked themselves).
So that thread, to me, feels like it has the same fundamental problem as Drepper’s standard rant[1] against static linking in general: it mixes up the problems arising from one libc’s particular implementation with problems inherent to the task of being a libc. (Drepper’s has much more of an attitude problem, of course.)
As for why you’d actually want to dlopen from a static executable, there’s one killer app: exokernels, loading (parts of) system-provided drivers into your process for speed. You might think this an academic fever dream, except that is how talking to the GPU works. Because of that, there’s basically no way to make a statically linked Linux GUI app that makes adequate use of a modern computer’s resources. (Even on a laptop with integrated graphics, using the CPU to shuttle pixels around is patently stupid and wasteful—by which I don’t mean you should never do it, just that there should be an alternative to doing it.)
Stretching the definitions a little, the in-proc part of a GPU driver is a very very smart RPC shim, and that’s not the only useful kind: medium-smart RPC shims like KERNEL32 and dumb ones like COM proxy DLLs and the Linux kernel’s VDSO are useful to dynamically load too.
And then there are plugins for stuff that doesn’t really want to pass through a bytestream interface (at all or efficiently), like media format support plugins (avoided by ffmpeg through linking in every media format ever), audio processing plugins, and so on.
Note that all of these intentionally have a very narrow waist[2] of an interface, and when done right they don’t even require both sides to share a malloc implementation. (Not a problem on Windows where there’s malloc at home^W^W^W a shared malloc in KERNEL32; the flip side is the malloc in KERNEL32 sucks ass and they’re stuck with it.) Hell, some of them hardly require wiring together arbitrary symbols and would be OK receiving and returning well-known structs of function pointers in an init function called after dlopen.
[1] https://www.akkadia.org/drepper/no_static_linking.html
[2] https://www.oilshell.org/blog/2022/02/diagrams.html
> Witness Windows, which has little to no problem with multiple libcs in a process
Only so long as you don't pass data structures from one to the other. The same caveats wrt malloc/free or fopen/fclose across libc boundaries still applies.
Well, not anymore, but only because libc is a system DLL on Windows now with a stable ABI, so for new apps they all share the same copy.
> The same applies to PAM word for word.
That's one of the reasons that OpenBSD is rather compelling. BSDAuth doesn't open arbitrary libraries to execute code, it forks and execs binaries so it doesn't pollute your program's namespace in unpredictable ways.
> It's true that there's something of a problem where two copies of a libc can't coexist in a process...
That's the meat of this article. It goes beyond complaining about a relatable issue and talks about the work and research they've done to see how it can be mitigated. I think it's a neat exercise to wonder how you could restructure a libc to allow multi-libc compatibility, but question why anyone would even want to statically link to libc in a program that dlopen's other libraries. If you're worried about a stable ABI with your libc, but acknowledge that other libraries you use link to a potentially different and incompatible libc thus making the problem even more complicated, you should probably go the BSDAuth route instead of introducing both additional complexity and incompatibility with existing systems. I think almost everything should be suitable for static linking and that Drepper's clarification is much more interesting than the rant. Polluting the global lib directory with a bunch of your private dependencies should be frowned upon and hides the real scale of applications. Installing an application shouldn't make the rest of your system harder to understand, especially when it doesn't do any special integration. When you have to dynamically link anyway:
> As for why you’d actually want to dlopen from a static executable, there’s one killer app: exokernels, loading (parts of) system-provided drivers into your process for speed.
If you're dealing with system resources like GPU drivers, those should be opaque implementations loaded by intermediaries like libglvnd. [1] This comes to mind as even more reason why dynamic dependencies of even static binaries are terrible. The resolution works, but it would be better if no zlib symbols would leak from mesa at all (using --exclude-libs and linking statically) so a compiled dependency cannot break the program that depends on it. So yes, I agree that dynamic dependencies of static libraries should be static themselves (though enforcing that is questionable), but I don't agree that the libc should be considered part of that problem and statically linked as well. That leads us to:
> ... when done right they don't even require both sides to share a malloc implementation
Better API design for libraries can eliminate a lot of these issues, but enforcing that is much harder problem in the current landscape where both sides are casually expected to share a malloc implementation -- hence the complication described in the article. "How can we force everything that exists into a better paradigm" is a lot less practical of a question than "what are the fewest changes we'd need to ensure this would work with just a recompile". I agree with the idea of a "narrow waist of an interface", but it's not useful in practice until people agree where the boundary should be and you can force everyone to abide by it.
[1] https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/28...
I don't understand why they don't just statically link their binaries. First, they said this:
> Even if you managed to statically link GLIBC—or used an alternative like musl—your application would be unable to load any dynamic libraries at runtime.
But then they immediately said they actually statically link all of their deps aside from libc.
> Instead, we take a different approach: statically linking everything we can.
If they're statically linking everything other than libc, then using musl or statically linking glibc will finish the job. Unless they have some need for loading share libs at runtime which they didn't already have linked into their binary (i.e. manual dlopen), this solves the portability problem on Linux.
What am I missing (assuming I know of the security implications of statically linked binaries -- which they didn't mention as a concern)?
And please, statically linking everything is NOT a solution -- the only reason I can run some games from 20 years ago still on my recent Linux is because they didn't decide to stupidly statically link everything, so I at least _can_ replace the libraries with hooks that make the games work with newer versions.
As long as the library is available.
Neither static nor dynamic linking is looking to solve the 20 year old binaries issue, so both will have different issues.
But I think it's easier for me to find a 20 year old ISO of a Red Hat/Slackware where I can simply run the statically linked binary. Dependency hell for older distros become really difficult when the older packages are not archived anywhere anymore.
I've recently had to do this (to bisect when a change introduced a superficial bug into a 20-year-old program). I think "simply run" is viewing Linux of that era through rose-tinted glasses.
Even for simple 2D "Super VGA" you're needing to choose the correct XFree86 implementation and still tweak your Xorg configuration. The emulated hardware also has bugs, since most of the focus is now on virtio drivers.
(The 20-year-old program was linked against libsdl, which amusingly means on my modern system it supports Wayland with no issues.)
Plus there are other advantages that I always say should never be discounted, such as desktop integration, that you get when you can replace the toolkit (dynamic) libraries rather than just "running everything in a VM". Running everything in a VM is a poor excuse for backwards compatibility.
Debian archives all of our binaries (and source) here:
https://snapshot.debian.org/
Some things built on top of that:
https://manpages.debian.org/man/debsnap https://manpages.debian.org/man/debbisect https://wiki.debian.org/BisectDebian https://metasnap.debian.net/ https://reproduce.debian.net/
All since ~mid 2005. Unfortunately a tarball I've been looking for was added before then, but replaced before a release was cut (so it's also missing from those archives).
Which one?
The OpenAL source tarball that was imported into Debian as version "0.0.200101102349-1".
It's interesting to think how a 20 year old OS plus one program is probably a smaller bundle size than many modern Electron apps ostensibly built "for cross platform compatibility". Maybe microkernels are the way.
Old DOS games are often sold (e.g. on GOG or Steam) with an installer and a full copy of DOSBox, so you get an entire operating system plus hardware emulator to run the game. Looks like this adds less than 10 MB to each game installed in Windows.
How should a microkernel run (WASI) WASM runtimes?
Docker can run WASM runtimes, but I don't think podman or nerdctl can yet.
From https://news.ycombinator.com/item?id=38779803 :
From https://news.ycombinator.com/item?id=41306658 :> ostree native containers are bootable host images that can also be built and signed with a SLSA provenance attestation; https://coreos.github.io/rpm-ostree/container/ :
Native containers run on the host and can host normal containers if a container engine is installed. Compared to an electron runtime, IDK how minimal a native container with systemd and podman, and WASM runtimes, and portable GUI rendering libraries could be.CoreOS - which was for creating minimal host images that host containers - is now Fedora Atomic is now Fedora Atomic Desktops and rpm-ostree. Silverblue, Kinoite, Sericea; and Bazzite and Secure Blue.
Secureblue has a hardened_malloc implementation.
From https://jangafx.com/insights/linux-binary-compatibility :
> To handle this correctly, each libc version would need a way to enumerate files across all other libc instances, including dynamically loaded ones, ensuring that every file is visited exactly once without forming cycles. This enumeration must also be thread-safe. Additionally, while enumeration is in progress, another libc could be dynamically loaded (e.g., via dlopen) on a separate thread, or a new file could be opened (e.g., a global constructor in a dynamically loaded library calling fopen).
FWIU, ROP Return-Oriented Programming and Gadgets approaches have implementations of things like dynamic header discovery of static and dynamic libraries at runtime; to compile more at runtime (which isn't safe, though: nothing reverifies what's mutated after loading the PE into process space, after NX tagging or not, before and after secure enclaves and LD_PRELOAD (which some go binaries don't respect, for example).
Can a microkernel do eBPF?
What about a RISC machine for WASM and WASI?
"Customasm – An assembler for custom, user-defined instruction sets" (2024) https://news.ycombinator.com/item?id=42717357
Maybe that would shrink some of these flatpaks which ship their own Electron runtimes instead of like the Gnome and KDE shared runtimes.
Python's manylinux project specifies a number of libc versions that manylinux packages portably target.
Manylinux requires a tool called auditwheel for Linux, delicate for MacOS, and delvewheel for windows;
Auditwheel > Overview: https://github.com/pypa/auditwheel#overview :
> auditwheel is a command line tool to facilitate the creation of Python wheel packages for Linux (containing pre-compiled binary extensions) that are compatible with a wide variety of Linux distributions, consistent with the PEP 600 manylinux_x_y, PEP 513 manylinux1, PEP 571 manylinux2010 and PEP 599 manylinux2014 platform tags.
> auditwheel show: shows external shared libraries that the wheel depends on (beyond the libraries included in the manylinux policies), and checks the extension modules for the use of versioned symbols that exceed the manylinux ABI.
> auditwheel repair: copies these external shared libraries into the wheel itself, and automatically modifies the appropriate RPATH entries such that these libraries will be picked up at runtime. This accomplishes a similar result as if the libraries had been statically linked without requiring changes to the build system. Packagers are advised that bundling, like static linking, may implicate copyright concerns
github/choosealicense.com: https://github.com/github/choosealicense.com
From https://news.ycombinator.com/item?id=42347468 :
> A manylinux_x_y wheel requires glibc>=x.y. A musllinux_x_y wheel requires musl libc>=x.y; per PEP 600
> How should a microkernel run (WASI) WASM runtimes?
Same as any other kernel—the runtime is just a userspace program.
> Can a microkernel do eBPF?
If it implements it, why not?
Should a microkernel implement eBPF and WASM, or, for the same reasons that justify a microkernel should eBPF and most other things be confined or relegated or segregated in userspace; in terms of microkernel goals like separation of concerns and least privilege and then performance?
Linux containers have process isolation features that userspace sandboxes like bubblewrap and runtimes don't.
Flatpaks bypass selinux and apparmor policies and run unconfined (on DAC but not MAC systems) because the path to the executable in the flatpaks differs from the system policy for */s?bin/* and so wouldn't be relabeled with the necessary extended filesystem attributes even on `restorecon /` (which runs on reboot if /.autorelabel exists).
Thus, e.g. Firefox from a signed package in a container on the host, and Firefox from a package on the host are more process-isolated than Firefox in a Flatpak or from a curl'ed statically-linked binary because one couldn't figure out the build system.
Container-selinux, Kata containers, and GVisor further secure containers without requiring the RAM necessary for full VM virtualization with Xen or Qemu; and that is possible because of container interface standards.
Linux machines run ELF binaries, which could include WASM instructions
/? ELF binary WASM : https://www.google.com/search?q=elf+binary+wasm :
mewz-project/wasker https://github.com/mewz-project/wasker :
> What's new with Wasker is, Wasker generates an OS-independent ELF file where WASI calls from Wasm applications remain unresolved.*
> This unresolved feature allows Wasker's output ELF file to be linked with WASI implementations provided by various operating systems, enabling each OS to execute Wasm applications.
> Wasker empowers your favorite OS to serve as a Wasm runtime!
Why shouldn't we container2wasm everything? Because (rootless) Linux containers better isolate the workload than any current WASM runtime in userspace.
Return oriented programming: https://en.wikipedia.org/wiki/Return-oriented_programming
/? awesome return oriented programming sire:github.com https://www.google.com/search?q=awesome+return+oriented+prog...
This can probably find multiple versions of libc at runtime, too: https://github.com/0vercl0k/rp :
> rp++ is a fast C++ ROP gadget finder for PE/ELF/Mach-O x86/x64/ARM/ARM64 binaries.
Software running for 20 years is not always a reasonable requirement.
But sometimes it is. And even if it's not a requirement, it might be nice to have.
How do you troubleshoot and figure that out?
Dynamic linking obviously has benefits or there would not be any incentive to build them or provide the capacity for them.
The problem is they also have problems which motivates people to statically link.
I remember back in the Amiga days when there were multiple libraries that provided file requesters. At one point I saw a unifying file requester library that implemented the interfaces of multiple others so that all requesters had the same look.
It's something that hasn't been done as far as I am aware on Linux. partially because of the problems with Linux dynamic libraries.
I think the answer isn't just static linking.
I think the solution is a commitment.
If you are going to make a dynamic library, commit to backwards compatibility. If you can't provide that, that's ok, but please statically link.
Perhaps making a library at a base level with a forever backwards compatible interface with a static version for breaking changes would help. That might allow for a blend of bug support and adding future features.
At least for some apps, perhaps it’s Wine and the Win32 API which is the answer.
OpenGL and Vulkan are provided by the GPU vendor you can't statically link them.
Various things including name (DNS) resolution rely on dynamic linking.
Are you saying that a statically linked binary cannot make an HTTP request to `google.com` because it would be unable to resolve the domain name?
There are entire distros, like alpine, built on musl. I find this very hard to believe.
All versions of MUSL prior to 1.2.4 (released less than two years ago) would indeed fail to perform DNS lookups in many common cases, and a lot of programs could not run in MUSL as a result. (I'm not aware of what specific deficiencies remain in MUSL, but given the history even when there are explicit standards, I am confident that there are more.) This wasn't related to dynamic linking though.
Glibc's NSS is mostly relevant for LANs. Which is a lot of corporate and home networks.
The configuration of DNS resolution on Linux is quite complicated [1]. Musl just ignores all that. You can build a distro that works with musl, but a static musl binary dropped into an arbitrary Linux system won't necessarily work correctly.
[1]: https://news.ycombinator.com/item?id=43451861
The easy and conforming way to do that would be to call "getent hosts google.com" and use the answer. But this only works for simple use cases where you just need some IPv4/IPv6 address, you can't get other kinds of DNS records like MX or TLSA this way.
You have to bundle your own resolver into your application. But here's the rub, users expect your application to respect nsswitch which requires loading shared libs which execute arbitrary code. How Go handles this is somewhat awkward. They parse /etc/nsswitch and decide if they can cheat and use their own resolver based on what modules they see[1]. Otherwise they farm out to cgo to go through glibc.
[1] They're playing with fire here because you can't really assume to know for sure how the module 'dns' behaves. A user could replace the lib that backs it with their own that resolves everything to zombo.com. It would be one thing if nsswitch described behavior which was well defined and could be emulated but it doesn't, it specifies a specific implementation.
Fascinating. Thanks for breaking this down more. I think the article could've explained this point further.
Static linking makes it impossible for a library to evolve external to applications. That’s not a great outcome for a lot of reasons.
musl and glibc static links are their own Pandora’s box of pain and suffering. They don’t “just work” like you’d hope and dream.
This blows my mind, that in 2025 we still struggle with a simple task such as "read in a string, parse it, and query a cascade of resolvers to discover it's IP". I just can't fathom how that is a difficult problem, or why DNS still is notorious for causing so much pain and suffering. Compared to the advancements in hardware and graphics and so many other areas.
There are resolvers for not just DNS but for users and other lookups. The list of resolvers is dynamic, they are configured in /etc/nsswitch.conf. The /etc/hosts lookup is part of the system.
Where do the resolvers come from? It needs to be possible to install resolvers separately and dynamically load them. Unless you want to have NIS always installed. Better to install LDAP for those who need it.
It could be handled by a system daemon instead. We don't really need this in-process all the time.
I bought 3 linux games on DVD between 2006 and 2016. I stopped buying linux games and instead started again buying windows games. Because there is no easy way to run them. On the other hand I can just run myst1993 and most of windows games without much hustle via wine. Wine is linux only stable abi
> Wine is linux only stable abi
https://blog.hiler.eu/win32-the-only-stable-abi/
The most annoying part to me is that Linus et all make not breaking compatibility their #1 goal (only rarely superseded by security and performance concerns) but all of the other libraries on top of the kernel don't seem to care about that at all and break things willy-nilly.
As if a windows game from 2006 is going to run on windows 11 :)
I was able to run Age of Empires 2 on Windows 11. It didn't like running on an ultrawide, but it worked well enough at smaller resolutions.
a 3d game…
I've just installed and run Star Wars Empire at war (2005) from steam, it seems to run just fine in WQHD. The UI isn't even blurry
On steam? Games that get updated are updated? Do you have any more groundbreaking revelations?
instead of being toxic you could point any particular one to check. If I don't have it purchased I'll grab the OG scene rel.
The 2006-engine version of Half-Life 2: Episode 1 runs on Windows 10/11 with no configuration [1], outside of getting Steam to download it. I recall installing The Elder Scrolls IV: Oblivion on a Windows 11 machine, which just needed Directx 9c to run.
[1] https://steamcommunity.com/sharedfiles/filedetails/?id=28643...
all of the half life work fine natively on linux as well on steam. Do you have something more challenging?
Because valve puts in the ongoing effort. Tell loki software they need to update the binaries of their games. Oh wait they don't exist anymore.
Valve puts ongoing effort also on the windows games… yet a sibling comment tried a game from valve instead of trying a game from a CD to do a fair comparison -_-'
This is a really great article about binary compatibility!
I disagree with their idea for fixing it by splitting up glibc. I think it's a bad idea because it doesn't actually fix the problems that lead to compat breakage, and it's bad because it's harder than it seems.
They cite these compat bugs as part of their reasoning for why glibc should be split up:
- https://sourceware.org/bugzilla/show_bug.cgi?id=29456
- https://sourceware.org/bugzilla/show_bug.cgi?id=32653
- https://sourceware.org/bugzilla/show_bug.cgi?id=32786
I don't see how a single one of these would be fixed by splitting up glibc. If their proposed libdl or libthread were updated and had one of these regressions, it would cause just as much of a bug as if a monolithic libc updates with one of these regressions.
So, splitting up glibc wouldn't fix the issue.
Also, splitting up glibc would be super nasty because of how the threading, loading, and syscall parts of libc are coupled (some syscalls are implemented with deep threading awareness, like the setxid calls, threads need to know about the loader and vice-versa, and other issues).
I think the problem here is how releases are cut. In an ideal world, glibc devs would have caught all three of those bugs before shipping 2.41. Big corpos like Microsoft manage that by having a binary compatibility team that runs All The Apps on every new version of the OS. I'm guessing that glibc doesn't have (as much of) that kind of process.
Windows having multiple C libraries has its own pain points, in particular it's difficult to ship binary libraries that return allocated memory to their consumer (you either need to have the library consumer allocate the memory, which probably explains why so many Win32 APIs have this behaviour, or allow alloc/free functions to be registered). Not to mention different C libraries having their own file handle, TLS, etc state. Unsurprisingly Microsoft now ships the Universal CRT (UCRT) as part of Windows.
Having consumer allocate the memory where possible has other benefits in that the consumer can use more efficient allocation techniques (stack, arenas etc), and given that Win32 API design goes all the way back to the earliest version of Windows running on very resource-limited PCs, that would be the main reason why.
However, in cases where this wasn't consistently feasible - e.g. COM - Windows instead mandates the use of a common API to manage memory: CoGetMalloc etc.
Zig CC + Vulkan would solve most of these issues:
1. Zig's toolchain statically links with musl libc, producing binaries that depend only on the Linux kernel syscall ABI, not any specific glibc version.
2. This eliminates all the symbol versioning nightmares (`GLIBC_2.xx not found`) that plague distribution across different Linux systems.
3. Vulkan provides a standardized, vendor-neutral GPU API that's supported across all modern graphics hardware, eliminating driver-specific dependencies.
4. The resulting binary is completely self-contained - no external library dependencies, no version mismatches, no containerization needed.
5. You get forward AND backward compatibility - the binary will run on both ancient and cutting-edge distros without modification.
The only real limitation is for NVIDIA CUDA-specific workloads, where you'd still need their proprietary stack.
Furthermore, for those who prefer a higher-level API, Zig CC + WebGPU offers similar benefits with a simpler programming model, though with slightly less mature implementations and a possible small performance overhead.
> Windows maintains such strong binary compatibility
The REAL reason windows maintains binary compatibility is because it is commercial and nobody ships source code.
In fact, many applications ship a whole boatload of DLLs, which I think is the commercial equivalent of static linking.
It kinda is, and kinda isn't.
It is, in the sense that the package is bigger, and the package ships "everything it needs".
It isn't in the sense that those parts can be updated independently as long as the DLL interface is backward compatible.
For example, I ship OpenSSL dlls with my app. Which means swapping in a later (compatible) OpenSSL can be done (by the user if necessary.)
If I'm making a small utility I static link it - and I still use utilities daily I compiled 25 years ago. Obviously those dynamically link to KERNEL etc, but Microsoft has insane levels of compatibility there.
And perhaps that's the fundamental root of the issue. Windows has one provider, very committed to the longevity of software. Linux, well, does not.
That's OK. The world has room for different philosophies. And each one will have strengths and weaknesses.
What about the security issues of the old operating systems people keep around for their mission critical software that has no upgrade path?
isn't this tu quoque?
Isn't the compatibility promise what syscalls essentially are?
Historically, they (almost) never break and they are steadily incremented to prevent overlaps in differences of parameters.
As WASI is also implementing syscalls for WASM, I'd argue that the binary format doesn't really matter as long as it's using the same syscalls in the end.
I understand this topic is mostly focussing on glibc/muslc problems, but if you want to develop stable software, use CGo free Go binaries. They likely will run in 10 years the same way they do today.
C ABI compatibility is a mess on linux mostly because upstream maintainers don't give a damn about semantic versioning. Just take a look at the SO file headers, and how they differ from upstream "semantic" versions of the library. As long as shared objects differ in versions due to breaking changes, and as long as the C ecosystem doesn't enforce correct versioning, this won't change.
There’s no reason to believe that widespread Linux adoption would not irrevocably damage the experience. It would end up looking something like Android. The same thing happened to the internet. It was ruined by mass adoption. You can call this gatekeeping, but nothing positive has come from getting most of the world onto social media.
I'm inclined to agree with you here. As much as I'd love to see native apps from some big vendors (Adobe, in particular) if what it takes to get there is mass adoption of Linux on the desktop, I'm not sure I want it.
In a way, we already have it in the form of ChromeOS, and I certainly don't want ChromeOS or Android to be the default experience, or for the community to start to cater to that model.
All of the things Linux would need to become to get Windows level marketshare would strip away what makes it appealing in the first place, in particular, full user control. I don't want a nanny desktop, but that's exactly what it would become.
Linux can stay niche for those who appreciate it as it is.
> GLIBC is an example of a "system library" that cannot be bundled with your application because it includes the dynamic linker itself. This linker is responsible for loading other libraries, some of which may also depend on GLIBC—but not always.
Running WordPerfect on modern Linux is done by shipping both of those components:
https://github.com/taviso/wpunix
As an end user I often patch the glibc version incompatibility away with https://github.com/corsix/polyfill-glibc
This often leads to discovering new version incompatibilities in other libs. But as the article says others usually can be statically compiled.Wow. This is huge. This patch doesn’t risk weird runtime errors due to missing polyfills or anything? Does it risk polyfilling irrelevant functions that aren’t referenced and bloating?
Very cool. Glad someone finally made this.
golang seems to manage being 100% statically linked (unless you go out of your way, by using plugins). It just forgoes the stuff that the article mentions glibc doing dynamic linking for, and does it the simple and direct way.
What do you use golang for? If your answer is only web-dev and if you don't touch anything GUI, system libraries, PAM or VPNs, you simply haven't compiled something complex enough.
Related video about releasing games on linux, i.e. dlopen() all the things https://www.youtube.com/watch?v=MeMPCSqQ-34
Why don’t the authors release a distro that actually implements their own proposal?
Say, based on (perhaps on pieces of) musl, to be more straightforward to get things going?
This is simply too complex. There are quite a bit dependencies on Glibc and other GNU-based behavior.
What you're actually asking from them is creating a new Android or Chrome OS i.e. recreating the entire userspace. "Just split up musl" will not work. How the system headers work needs to be redefined. This is basically a complete redefinition of entire API of all Linux programs and libraries. Which library will supply unistd.h or any of the sys/whatever.h? Any interdependencies? I think even POSIX itself can be hostile against this. So the fix will have to deviate a lot from that (that's probably why Windows cannot be made POSIX easily either).
> shipping software on Linux
That's a surprisingly hard nut to crack when containers won't work for your use case. We found https://github.com/silitics/rugix to work well in that situation.
Really the title should be something like “the difficulty of releasing binaries on Linux and how to work around it.” It isn’t really an atrocious state, the goal of a distro should be to get useful software into the repos. Software not distributed in an open source format doesn’t really help there.
I think the problem is that people treat Linux as OS instead of just a kernel.
You should assume every Linux distro is a different OS so when you are shipping your app for Linux you're actually shipping your app for Debian, Fedora, Ubuntu etc
This is definitely the root of it. Linux isn't the OS, your distro is.
Somewhere along the way we seem to have collectively forgotten that the distro, via it's package manager, is responsible for delivering software, ensuring compatibility, and integrating it all - at least for open source software. That was one of Debian's biggest selling points "back in the day" - that they had a huge repository and big team of maintainers packaging everything under the sun so it works on Debian.
Now we have app image, flatpak, snap, etc. All means of getting developers to ship binaries directly to users, bypassing the package managers, instead of shipping code to distros, but we shouldn't expect devs to ship to the 100s of different OSes in this ecosystem. They ship the code, distros pick it up (or don't) and make it work on their distro.
The shift to direct-to-user binaries doesn't make sense for OSS. It should be packaged by the distro, not the dev (of course, the devs can help with packaging if they are so inclined).
Obviously this doesn't apply to proprietary software, but like you said, you can just target on of the LTS enterprise distros and get a well-defined, stable base (for 10 years with Ubuntu and Red hat).
That's the sane way to tackle it. If you're the vendor, just target the top N (whatever value of N you can cope with).
I don't mean disrespect towards people running Alpine (hi), Arch, or Gentoo, but you wouldn't be running these distros if you aren't ready to handle their quirks.
TFA mostly talks about binary compat. Even if you can get away with statically linking everything, you still have to cope with the mess that is userspace fragmentation: <https://tailscale.com/blog/sisyphean-dns-client-linux>
So yeah, supporting the top N gets you approximately sqrt(N/(N+1))% of the way. (Assuming desktop Linux market share is about 1%.)
>I don't mean disrespect towards people running Alpine (hi), Arch, or Gentoo, but you wouldn't be running these distros if you aren't ready to handle their quirks.
You'd think so, but you'd be surprised how many gamers switching from Windows to Linux get recommended Arch and Manjaro
I see Bazzite recommended more than Arch these days, but Arch is confusingly a very popular choice.
I wonder if Arch being the basis for SteamOS is responsible.
No discussion of Cosmopolitan or APE?
Software and binary formats with a 100% chance of being blocked as malware doesn't count.
This article missed a critical point which is "the right way" to select a glibc ABI version: see binutils ld documentation, second part of the page related to VERSION support. This must include glibc internal symbols.
This will allow to craft ELF binaries on a modern distro which will run on "older" distros. This is critical for games and game engines. There is an significant upfront only-once work in order to select an "old" glibc ABI.
The quick and dirty alternative being having a toolchain configured to link with an "old" glibc on the side.
This article missed the -static-libstdc++ critical option for c++ applications (the c++ ABI is hell on earth), but did not miss the -static-libgcc and the dynamic loading of system interface shared libs.
One of the features Zig provides is ability to target any glibc version. See https://github.com/ziglang/glibc-abi-tool/ for more details on how this is solved.
You’re my hero for this.
One of my side projects is building a toolchain to enable C++ cross-compile using the Zig header/source libs.
I didn’t love Zig as a Clang++ replacement because it has a bit too much magic. And it might go ahead? But the underlying library code is a God send
Could you point to an example of a project doing it the right way? I'm a little unclear on the type of incantation needed for the "version script".
Is there a reason glibc can't just do a better job at keeping some legacy symbols around? It's not like it's big stuff. They're things like legacy string functions. We're talking a few kilobytes of code in most cases.
The Linux kernel goes to a lot of effort to not break user space, at least for non-exotic core features and syscalls. It seems like a lot of user-space in Linux-land does not make the same effort.
It's particularly bad when it's the C library doing this, since that's at the center of the dependency graph for almost everything.
The problem is the opposite: they are trying to run executables built using a newer glibc in a system that has an older glibc. glibc keeps all the old function definitions since practically forever.
Frankly, I do not understand who would think glibc symbols themselves would be the challenge in this case. Even if you statically link glibc there's zero guarantee the syscalls will be present in the older Linux (cue .ABI-tag failures). Or even damn ELF format changes (e.g. gnu-style hashes). The simple solution is to build in the older Linux (&glibc).
In my long experience with ancient binaries, glibc has almost never been the problem, and its ability to _run_ ancient binaries is all but excellent; even Linux is more of a problem than glibc is (for starters paths to everywhere in /proc, /sys change every other half-decade).
> executables built and using a newer glibc
It’s an abomination that Linux uses system libraries when building. Catastrophically terrible and stupid decision.
It should be trivial for any program to compile and specify any arbitrary previous version of glibc as the target.
Linux got this so incredibly wildly wrong. It’s a shame.
Ehm... you _can_ use the older toolchain even in the newer "Linux", in the same way you have to use a older Visual Studio to target Windows 8 from Windows 10.
So maybe the difference here is between can and actually doing it?
I tend to stay on the oldest supported version of Windows until they drop support and haven't ever seen an application that wouldn't run because it's built on a newer version of Windows.
I am saying that compiler toolchains on Linux should never ever under any circumstances ever rely on anything on the system for compiling. Compiling based on the system global version of glibc is stupid, bad, wrong, and Linus should be ashamed for letting it happen.
It should be trivial for Windows to cross-compile for Linux for any distro and for any ancient version of glibc.
It is not trivial.
Here is a post describing the mountain range of bullshit that Zig had to move to enable trivial cross-compile and backwards targeting. https://andrewkelley.me/post/zig-cc-powerful-drop-in-replace...
Linux is far and away the worst offender out of Linux, Mac, and Windows. By leaps and bounds.
The legacy symbols are still around. It's just that the linker won't use them by default.
> a lot of user-space in Linux-land does not make the same effort
I believe that, what the article misses is that glibc is maintained and extended with an entirely different community and development model. Windows remains compatible over decades because Microsoft (a) is the sole distributor, and (b) puts an immense effort towards backwards compat. In Linux userspace, it's simply a non-goal across distributions. If you want to ship a binary for a particular distro, you need to build the binary on / for that distro; even within a distro, a major release bump (or especially a major release downgrade) may break a binary.
Ultimately, it's a consequence of Conway’s Law. Microsoft is the sole distributor of Windows, so they can enforce compatibilty with an iron fist, and there are people working for Microsoft whose pay depends on said compatibility. With "Linux" in general, there is no common authority to appeal to, and (again) most vendors don't even promise a seamless userspace upgrade path from one major release to another.
This is unfixable; it will never change -- as long as independent parties are permitted to distribute different operating systems yet call them all "Linux".
Ship multiple binaries, or distribute the source code (and let users build it).
EDIT: you'll notice that "ship multiple binaries" is what distros (especially commercial distros) do. They maintain separate branches, backport fixes to old branches, and employ software maintenance engineers to focus on this kind of work. If you want to target multiple major releases, this is what you have to do, too.
If you (as a commercial ISV) target a commercial distro with long-term support, and can convince your users to use / license the same distro, you'll have a good, stable development experience. You only need to port like once every decade, when you jump major releases.
The Linux user base / the Linux market is fragmented; that's the whole goal. The technical proliferation / inconsistency is just a consequence. Unless you take away the freedom of users to run their own flavors of "Linux", there won't be a uniform Linux target.
In a way, it's funny to even expect otherwise. Why do you expect to ship the same binaries when the foundations are diverse, with no standardization authority that all Linux distributors recognize as such? And even POSIX is an API spec, not an ABI spec.
And, any authority that controls binary aspects will immediately accrue political capital. This is exactly what shouldn't happen in Linux. The fact that anyone can fork (or start) a distro, and contribute to the chaos, is good for freedom.
> If you (as a commercial ISV) target a commercial distro with long-term support, and can convince your users to use / license the same distro, you'll have a good, stable development experience. You only need to port like once every decade, when you jump major releases.
If things go well, it's even better than that: If you target ex. RHEL 8, there's a very good chance that your binaries will work on RHEL 9 and a decent shot at RHEL 10 with zero changes (though of course you should test all versions you want to work). And the same for Ubuntu 20.04/22.04/24.04/... and Debian/SUSE/whatever. Backwards incompatibilities can happen, but within a single stable distro they're not super common so the lazy ISV can probably only really port forward after more than a decade if they really want.
(Incidentally, this isn't a hypothetical: I once had the joy of working on software that targeted RHEL 5, and those binaries ran on RHEL/CentOS 7 without any problems.)
> Ship multiple binaries, or distribute the source code (and let users build it).
And that's why we have package managers and distro maintainers/packagers. You'll get no help from the community if your stuff is proprietary, just the way it is. Ship the code, distros will pick it up and do the packaging for you to make it available in their distro. It's part of the free software culture that surrounds the Linux ecosystem.
If you absolutely must ship proprietary software, then target an enterprise distro. Ship it for RHEL or Ubuntu LTS and you get, at least, 10 years of a stable base.
The big Linux distros (EL, Ubuntu) in fact have the opposite incentive, to get proprietary vendors to rather their distribution specifically.
This theory would check out if a proprietary vendor could easily get away with shipping a single binary package for all supported versions of, say, Ubuntu.
Having to build and maintain a binary packege separately for each version of the same distro probably isn't that appealing to them.
For Ubuntu they would only target LTS releases, most likely.
On EL it's easier, now you would just support 2 or 3 of EL7, EL8, and EL9.
As an example of something I use, Xfdtd only officially supports one version of Ubuntu and 2 versions of EL https://www.remcom.com/system-requirements#xfdtd-system-requ...
In practice, it wasn't too hard to get it running on EL9 or Fedora either...
Which works if you use binutils ld. Does it work with mold or gold? And then how do you use this with languages other than c++/c like Go or Rust?
I thought that Go invoked syscalls directly instead of going through libc.
It's Go's best feature, a simple CGO_ENABLED=0 gives you freedom from the tyranny of libc.
Rust toolchain has already a bug since 2015: it cannot link statically libgcc, namely it does not have "-static-libgcc" option since 2015.
I got the bug with "TinyGlade" video game (extremely good BTW), which is written in rust, and with the dev we hit that bug. Namely... better have a libgcc with the right ABI... and I can tell you, this has a been a HUGE issue since valve started to distribute games more than a decade ago.
As in you have to build on an old distro? Or something else?
Does Nix have any relevance here?
Here's a thought: just distribute source code. ABI issues should be mostly fixed. Most computers can compile source code fast enough for the user not to notice and cache the results so that it's never a problem again. If you want optimised code, you can do a source to source optimisation then zip and minify the file. You could compile such a library to approximately native speeds without much user-end lag using modern JIT methods, and maybe even run LTO in a background thread so that the exectuables outdo dynamically linked ones.
I hate compiling. 9/10 something goes wrong. Sometimes I can fix it, other times I just abandon the effort and use something else. These days I just use packages or docker images and if that doesn't work out I'm moving on, ain't nobody got time for this. You really can't expect people who just want to use their computers and don't even know what a compiler is to get involved in a process like that.
No, end users need not get involved, it could/should be handled by the operating system.
Gentoo? Compiling big packages takes ages.
More than 20 years, with a single-core Pentium 4, it could take indeed something like 3 days of continuous compilation to compile an entire Gentoo distribution, in order to have a personal computer with every application that one might want.
However, already after the appearance of the first dual-core AMD Athlon64, 20 years ago, that time could be reduced to not much more than a half of day, while nowadays, with a decent desktop CPU from 5 years ago, most Gentoo packages can be compiled and installed in less than a minute.
There are only a few packages whose compilation and installation can take a noticeable time, of up to tens of minutes, depending on the chosen options and on the number of cores of the CPU, e.g. Firefox, LibreOffice, LLVM.
There is only a single package whose compilation may take ages unless you have an expensive CPU and enough memory per core: Google Chromium (including its derivatives that use the same code base).
I can effortlessly compile Debian packages on my phone.
With some limits of course. I can't compile Chromium even on my laptop. But most of stuff - I can.
Not Gentoo. The system I'm suggesting is one where you replace the linker with one which accepts a platform agnostic assembly code with similar semantics to C, and perhaps some additional tools for stack traversal to allow the implementation of garbage collected languages and exceptions. Compiling would be fast because the code you distribute would have already had optimisations applied to it in a pre-compilation step. It would also be automatic in the same way the present linker system is automatic.
The idea is that software would only be shipped if it actually works, and part of this would be ensuring it compiles. I'm not suggesting we use Make files or some other foul approach like that. Code should ideally be distributed as a single zip file with an executable header that triggers it to be unpacked in some cache directory by a program analogous to the linker, and compiled by the simple process of compiling all source files. Additionally this could be sped up drastically by having the compiler be a single executable with uniform flags across files. Header files could be parsed once and shared between build threads in parallel, which alone would safe much work on the part of the compiler. If this still proves insufficient, a language with superior semantics to C could be chosen which can be compiled faster. Most code these days shipped is JS which is compiled on the user's computer. This would be an order of magnitude improvement on that from a performance perspective.
the first launch of firefox would take a few hours.
let alone the first boot of the linux kernel... :)
This is somewhat exaggerated.
The compilation of firefox could take a few hours on some laptop dual-core Skylake CPU from 10 years ago.
Nowadays, on any decent dektop CPU with many cores the compilation of Firefox should take significantly less than an hour, though it remains one of the handful of open-source applications with a really long and non-negligible compilation time.
The Linux kernel is normally compiled much faster than Firefox, except when one would enable the compilation of all existing kernel modules, for all the hardware that could be supported by Linux, even if almost all of that is not present and it would never be present on the target computer system.
Just because you are used to slow compilers doesn't mean fast ones are impossible. As I said in the original post, code optimisation can be done before compilation on the developers machine, so all that need be done on the target is a simple debug build. An example is Jonathon Blow's JAI compiler which compiles around 250,000 lines of code per second. Even on very slow hardware this is reduced to perhaps 80,000LoC/s, two minutes for 10 million lines of code. Most users would be tolerant of an "install" box with a progress bar that takes at most two minutes for only the largest application on the computer. Or perhaps the browser people would opt not to distribute their software in this way. Advanced software may have advanced needs, beyond that of the average application. Most code is already written in Javascript these days, meaning it begins life being interpreted and is JIT compiled on the user's computer. Few seem to notice the compilation but many complain about the general sluggishness.
If you offered users a deal: spend 1 minute installing the software and it is snappy and responsive when you use it, I suspect all would accept.
> An example is Jonathon Blow's JAI compiler
I thought you were advocating "just distribute source code" – JAI is a closed-source language that, in its decade of development, has never been used for a significant project.
I fail to see how that relates to my using it as an example of a fast compiler. For that matter, my suggested approach doesn't need to to release source code, just some kind of programming language code. You could minify it, pre-optimise it, or do any number of things that would mean it isn't source code but instead becomes derived code (as I said in my original comment). Distributing textual code is not the same as being open source. I'm really just saying you should distribute something that isn't a platform specific blob with a fixed binary interface. You could, for instance, use LLVM IR code to achieve this, which is little more than a kind of portable assembly. LLVM can apply heavy optimisations to this code on the developer's machine, then take only the last step of code generation on the user's machine.
a.k.a Rust approach
Slow and usually pointless for minor updates.
Absolutely ridiculous that scrollbars are disabled for this page. The kids are too cool for such pedestrian web parts! Grow up and stop forcing your shitty UX on people.
Edit: Found it - it's black on black - even worse!