Using IFUNC to patch sshd was kind of elegant, it achieved rootkit like behaviour with a pre-existing mechanism. And sure, it might be possible for a secure daemon like sshd to drop enough privileges that it could protect itself from a malicious dynamically linked library.
But IFUNC was not required, neither was systemd. The game was lost as soon as the attacker had arbitrary code installed in a semi-common library. It doesn't have to get linked directly with sshd, it only needed to be linked into any program running as root, at least one time.
Most programs make zero effort to sandbox themselves, and as soon as one of those links with the malicious library, it could do anything. Like indirectly targeting sshd by patching its binary on disk (optionally hiding it with a rootkit), or using debug APIs to patch sshd in memory.
IFUNC, systemd, and the patched openssh are all irrelevant to the issue, that was simply the route this attacker took to leverage their foothold in libxz. There are thousands of potential routes the attacker could have taken, and we simply can't defend from all of them.
> Most programs make zero effort to sandbox themselves, and as soon as one of those links with the malicious library, it could do anything. Like indirectly targeting sshd by patching its binary on disk (optionally hiding it with a rootkit), or using debug APIs to patch sshd in memory.
I do not understand how you even expected sshd to sandbox itself. Its entire purposes is to (a) daemonize , (b) allow incoming connections in and then (c) forward (possibly-root) shell statements. All 3 things are 100% required for sshd and would have already allowed an attack like this. Any talk about sandboxing here is wishful thinking.
I recently tried to make something properly sandboxed and, my goodness, we have basically crafted an ecosystem where everything needs access to everything. No wonder docker, despite all it's faults, is how everyone does it. You need an entire linux distro completely accessible in your sandbox.
It was not essential to the exploit, but that does not mean it was irrelevant. More commonly used libraries are watched harder. The exploit was made much, much, worse by its indirect use by way of systemd. Approximately nobody wanted that feature and it still went in. That's something we need to be able to discuss.
There is always selinux if we want to add protection against arbitrary code running as root. Just because something operate as root does not mean it must have privileged access to everything.
1) IFUNC is hardly the only way to run code before main.
2) The alternative they present is arguably less secure because the function pointer will remain writable for the life of the process, whereas with IFUNC the GOT will eventually be made immutable (or can be... not sure if that's the default behavior). In general function pointers aren't great for security unless you explicitly make the memory backing the pointer(s) unwritable, which at least is easier to do for a global table than it is for things like C++ vtables (because there's the extra indirection through data pointers involved to get to the table).
> The alternative they present is arguably less secure because the function pointer will remain writable for the life of the process
They also suggest an alternative to storing the function pointer, store the bit flags that decide which function to call. That restricts the call targets to only the legitimate ones intended.
Yeah, this blog is misguided. As a higher level criticism: it's confusing[1] the technical details with the payload with the exploit chain that deployed it.
The interesting thing is obviously not that you can get code to run at high privilege level by modifying a system component. I mean, duh, as it were.
The interesting thing is that the attackers (almost) got downstream Linux distros to suck down and deploy that malicious component for them. And that's not down to an oddball glibc feature, it happened because they got some human beings to trust a malicious actor. GNU glibc can't patch that!
> The alternative they present is arguably less secure because the function pointer will remain writable for the life of the process
The article mentions this, and also points to mprotect which you can use to protect the pointer.
Why people jump to criticize without reading first? BTW, you can ask an LLM to check your critique, before posting, if you don't want to read the text.
Yes, liblzma could have used multiple routes to take over sshd. Once you're running inside the process it's game over. The exact details, like how they used ifunc and an audit hook, are very interesting, but ultimately not that important.
Some solaris engineer got a little too clever and decided that the modular part of the the auth system needed to be dynamic libs. Now it's all in one process space, hard to understand, hard to debug and fragile.
I really like openbsd's bsdauth, I don't know if it is actually any better than pam but because it is moduler at the process level it is possible for mere mortals to debug and make custom auth plugins. Sadly only obsd actually uses it. https://man.openbsd.org/login.conf.5#AUTHENTICATION
IFUNC should be implemented by software itself,
like switching functions on runtime/compile checks.
Why bother having a slower, insecure version that is less
flexible than a function pointer? I have to agree with author.
Glibc is filled with even more nasty hacks ripe for new exploits.
Less indirection means faster code. If the dynamic loader is already using a level of indirection and you patch into that same indirection instead of adding another, you're not making it slower.
I agree so much and wished this was the main focus of the debate. It's more a question of why does this exist in the first place and not of how did they abuse it. Building only from source is the minimum required transparency and a CI/CD pipeline able to manipulate the artifact before release takes this away. I remember the outrage, when serde (i think it was) wanted to ship parts as pre-compiled binaries for build performance reasons...
The entire argumentation here is ridiculous. There's a big jump from "IFUNC slightly undermines RELRO" to "IFUNC is the real culprit". You could have gotten all but the same effect spawning a thread from a plain init or C++ constructor. No one should think that any relro, r^x or aslr or anything like this is going to deter anyone who can literally control the contents of the libraries which are linked in. They could, literally, exec a copy of sshd with a patched config if necessary.
Yeah, that suggestion made me roll my eyes. It's the wrong granularity, there's no build system support, it's inconvenient (executable wrappers? require the user to understand all transitive deps?).
It also fails to mention glibc-hwcaps, which would've been a cleaner solution in the context.
> The short answer is that they have to. OpenSSH is developed by the OpenBSD community, for the OpenBSD community, and they do not give a flying Fedora about Linux.
What complete horseshit. I stopped reading there.
The OpenSSH Portable branch is maintained by OpenBSD developers and SystemD is a completely optional add-on so why on earth would they make it a dependency? If they didn't care about the Linux community they wouldn't develop this software *for free* for them. They can go write their own GNU SSH then.
It certainly doesn't help that there are 165+ definitions of what constitutes a "complete GNU+Linux system" some of which use SystemD and some which vow never to.
It's not the OpenBSD developers' fault some Linux distros use overly complex plumbing and can't agree on one standard for their OS unlike every other OS out there, including Windows.
The xz backdoor was a Debian and Red Hat issue because they maintained patches to fix problems of their own creation. No one else was affected. Why should the OpenBSD people care? It's not their problem.
The OP agrees with you... if you continue reading, they wrote
> These patches never went into Portable OpenSSH, because the Portable OpenSSH folks were ["not interested in taking a dependency on libsystemd"](link). And they never went into upstream OpenSSH, because OpenBSD doesn't have any need to support SystemD.
The language may have been harsher than it needed to and therefore could be more easily misunderstood, but I believe you are actually in agreement with them
It makes it sound even worse, cherry picking language like "not interested" as if the OpenBSD folks should shoulder blame for not being altruistic enough.
It reeks of trashing your benefactor, who gave you well-written free software, which you then made insecure with your own patches.
If you remove the roof of your car with a chainsaw and are inevitably injured later, is it the car manufacturer's fault they didn't offer that model as a convertible from the factory?
The better question is why are people still trying to assign blame all these years later? The IT world dodged a bullet but has moved on (and likely didn't learn from their mistakes as supply chain attacks are steadily increasing).
Okay. You could see it that way. Or you could read what the author wrote about who is to blame:
> No one person or team really made a mistake here, but with the benefit of hindsight it's clear the attackers perceived that the left hand of Debian/Fedora SSH did not know what the right hand of xz-utils was doing.
I guess it's up to interpretation, but I read it the complete opposite way, as in Linux distributions should not think so highly of themselves as to expect OpenBSD to conform and adapt to their mess, and OpenBSD rightfully should not be expected to "give a flying Fedora about Linux".
This is barking up the wrong tree.
Using IFUNC to patch sshd was kind of elegant, it achieved rootkit like behaviour with a pre-existing mechanism. And sure, it might be possible for a secure daemon like sshd to drop enough privileges that it could protect itself from a malicious dynamically linked library.
But IFUNC was not required, neither was systemd. The game was lost as soon as the attacker had arbitrary code installed in a semi-common library. It doesn't have to get linked directly with sshd, it only needed to be linked into any program running as root, at least one time.
Most programs make zero effort to sandbox themselves, and as soon as one of those links with the malicious library, it could do anything. Like indirectly targeting sshd by patching its binary on disk (optionally hiding it with a rootkit), or using debug APIs to patch sshd in memory.
IFUNC, systemd, and the patched openssh are all irrelevant to the issue, that was simply the route this attacker took to leverage their foothold in libxz. There are thousands of potential routes the attacker could have taken, and we simply can't defend from all of them.
Did anyone look into whether OpenBSD has privilege separated sshd enough for that exploit to not be possible?
> Most programs make zero effort to sandbox themselves, and as soon as one of those links with the malicious library, it could do anything. Like indirectly targeting sshd by patching its binary on disk (optionally hiding it with a rootkit), or using debug APIs to patch sshd in memory.
I do not understand how you even expected sshd to sandbox itself. Its entire purposes is to (a) daemonize , (b) allow incoming connections in and then (c) forward (possibly-root) shell statements. All 3 things are 100% required for sshd and would have already allowed an attack like this. Any talk about sandboxing here is wishful thinking.
I recently tried to make something properly sandboxed and, my goodness, we have basically crafted an ecosystem where everything needs access to everything. No wonder docker, despite all it's faults, is how everyone does it. You need an entire linux distro completely accessible in your sandbox.
It was not essential to the exploit, but that does not mean it was irrelevant. More commonly used libraries are watched harder. The exploit was made much, much, worse by its indirect use by way of systemd. Approximately nobody wanted that feature and it still went in. That's something we need to be able to discuss.
But the fact that it was done indirectly from systemd has nothing to do with IFUNC.
There is always selinux if we want to add protection against arbitrary code running as root. Just because something operate as root does not mean it must have privileged access to everything.
1) IFUNC is hardly the only way to run code before main.
2) The alternative they present is arguably less secure because the function pointer will remain writable for the life of the process, whereas with IFUNC the GOT will eventually be made immutable (or can be... not sure if that's the default behavior). In general function pointers aren't great for security unless you explicitly make the memory backing the pointer(s) unwritable, which at least is easier to do for a global table than it is for things like C++ vtables (because there's the extra indirection through data pointers involved to get to the table).
> The alternative they present is arguably less secure because the function pointer will remain writable for the life of the process
They also suggest an alternative to storing the function pointer, store the bit flags that decide which function to call. That restricts the call targets to only the legitimate ones intended.
Yeah, this blog is misguided. As a higher level criticism: it's confusing[1] the technical details with the payload with the exploit chain that deployed it.
The interesting thing is obviously not that you can get code to run at high privilege level by modifying a system component. I mean, duh, as it were.
The interesting thing is that the attackers (almost) got downstream Linux distros to suck down and deploy that malicious component for them. And that's not down to an oddball glibc feature, it happened because they got some human beings to trust a malicious actor. GNU glibc can't patch that!
[1] Incorrectly, as you point out.
> The alternative they present is arguably less secure because the function pointer will remain writable for the life of the process
The article mentions this, and also points to mprotect which you can use to protect the pointer.
Why people jump to criticize without reading first? BTW, you can ask an LLM to check your critique, before posting, if you don't want to read the text.
Yes but at best their "solution" is equally secure, not any better.
They argue, and I tend to agree, that their solution is more secure.
1. It impiles some function pointers to be writable temporarily, not all of them.
2. It doesn't hide writable pointers from a cursory glance not familiar with IFUNC.
The GOT has to be initially writable regardless of ifunc, even with relro, to apply relocations.
Would xz still have been able to alter opensshd without IFUNC?
Yes, liblzma could have used multiple routes to take over sshd. Once you're running inside the process it's game over. The exact details, like how they used ifunc and an audit hook, are very interesting, but ultimately not that important.
Isn't it the same problem with pam.
Some solaris engineer got a little too clever and decided that the modular part of the the auth system needed to be dynamic libs. Now it's all in one process space, hard to understand, hard to debug and fragile.
I really like openbsd's bsdauth, I don't know if it is actually any better than pam but because it is moduler at the process level it is possible for mere mortals to debug and make custom auth plugins. Sadly only obsd actually uses it. https://man.openbsd.org/login.conf.5#AUTHENTICATION
IFUNC should be implemented by software itself, like switching functions on runtime/compile checks. Why bother having a slower, insecure version that is less flexible than a function pointer? I have to agree with author. Glibc is filled with even more nasty hacks ripe for new exploits.
Less indirection means faster code. If the dynamic loader is already using a level of indirection and you patch into that same indirection instead of adding another, you're not making it slower.
I agree so much and wished this was the main focus of the debate. It's more a question of why does this exist in the first place and not of how did they abuse it. Building only from source is the minimum required transparency and a CI/CD pipeline able to manipulate the artifact before release takes this away. I remember the outrage, when serde (i think it was) wanted to ship parts as pre-compiled binaries for build performance reasons...
The entire argumentation here is ridiculous. There's a big jump from "IFUNC slightly undermines RELRO" to "IFUNC is the real culprit". You could have gotten all but the same effect spawning a thread from a plain init or C++ constructor. No one should think that any relro, r^x or aslr or anything like this is going to deter anyone who can literally control the contents of the libraries which are linked in. They could, literally, exec a copy of sshd with a patched config if necessary.
The title is just clickbait.
seeing LD_PRELOAD in the "less-exploitable alternatives to GNU IFUNC" section was kind of funny
Yeah, that suggestion made me roll my eyes. It's the wrong granularity, there's no build system support, it's inconvenient (executable wrappers? require the user to understand all transitive deps?).
It also fails to mention glibc-hwcaps, which would've been a cleaner solution in the context.
> By letting the linker run arbitrary code before main
As I know C++ allows running arbitrary code before main too - for constructors of global variables. Does it bring security risks too?
It does not, it's actually arbitrary code running during the dynamic loading process, i.e. before _start.
But what if I have a C++ dynamic library? Does it call constructors for global variables before _start function in the main program starts?
_start takes care of calling the global initializers and register the atexit callback for the finalizers.
(In practice _start calls __libc_start_main, a libc function that handles all of that).
False dichotomy. There was a series of blatant process failures from Github maintainer through Debian package maintainers. IFUNC also bad.
Well, also Unicode identifiers, a C11 spec bug, nobody cares to fix. Still in C26, because "users expect Unicode stability", esp. it's bugs.
> Why do Linux Distros modify OpenSSH?
> The short answer is that they have to. OpenSSH is developed by the OpenBSD community, for the OpenBSD community, and they do not give a flying Fedora about Linux.
What complete horseshit. I stopped reading there.
The OpenSSH Portable branch is maintained by OpenBSD developers and SystemD is a completely optional add-on so why on earth would they make it a dependency? If they didn't care about the Linux community they wouldn't develop this software *for free* for them. They can go write their own GNU SSH then.
It certainly doesn't help that there are 165+ definitions of what constitutes a "complete GNU+Linux system" some of which use SystemD and some which vow never to.
It's not the OpenBSD developers' fault some Linux distros use overly complex plumbing and can't agree on one standard for their OS unlike every other OS out there, including Windows.
The xz backdoor was a Debian and Red Hat issue because they maintained patches to fix problems of their own creation. No one else was affected. Why should the OpenBSD people care? It's not their problem.
The OP agrees with you... if you continue reading, they wrote
> These patches never went into Portable OpenSSH, because the Portable OpenSSH folks were ["not interested in taking a dependency on libsystemd"](link). And they never went into upstream OpenSSH, because OpenBSD doesn't have any need to support SystemD.
The language may have been harsher than it needed to and therefore could be more easily misunderstood, but I believe you are actually in agreement with them
It makes it sound even worse, cherry picking language like "not interested" as if the OpenBSD folks should shoulder blame for not being altruistic enough.
It reeks of trashing your benefactor, who gave you well-written free software, which you then made insecure with your own patches.
If you remove the roof of your car with a chainsaw and are inevitably injured later, is it the car manufacturer's fault they didn't offer that model as a convertible from the factory?
The better question is why are people still trying to assign blame all these years later? The IT world dodged a bullet but has moved on (and likely didn't learn from their mistakes as supply chain attacks are steadily increasing).
Okay. You could see it that way. Or you could read what the author wrote about who is to blame:
> No one person or team really made a mistake here, but with the benefit of hindsight it's clear the attackers perceived that the left hand of Debian/Fedora SSH did not know what the right hand of xz-utils was doing.
with OpenBSD not even being mentioned here
I guess it's up to interpretation, but I read it the complete opposite way, as in Linux distributions should not think so highly of themselves as to expect OpenBSD to conform and adapt to their mess, and OpenBSD rightfully should not be expected to "give a flying Fedora about Linux".
Yes, that nefarious nation-state threat actor known as GNU IFUNC!
Curses, thwarted again!
xz-tools should scrap and reimplemented the code to the safer one , current one have safety and performance issue.