Why you can’t trust CPUID | Hacker News Reader

Why you can’t trust CPUID

chipsandcheese.com

64 pointsrbanffy24 comments

Don't forget that the likely reason why the capability to change the CPUID was introduced by AMD in its processors is that Intel (mis)used the CPUID feature via its compiler making the compiled program run faster if it runs on an Intel CPU (detected via whether CPUID is GenuineIntel, not by checking for CPU features):

> https://www.agner.org/optimize/blog/read.php?i=73

I don't understand why this matters.. if journalists are trusting random screenshot (or database entry) it could obviously be faked! Forget advanced MSR manipulation, one can just photoshop the screenshot, or open in debugger and edit window elements, or binary-patch the executable. Same goes with database entry - I bet the result submission is some HTTPs call which can be mitmproxy-ed and edited.

So the real title should have been: "You thought there are 5 ways to fake test results? Nope, there are 6." Not nearly as exciting.

CPUID is not designed to be secure, but, if every CPU model had a (directly inaccessible) unique key, it could be used to sign a challenge and prove that CPU is what it's telling the program it is.

Yeah we don't want that. Faking the feature list is quite useful for VMs, if say you have Zen 2 and Zen 3 CPUs in your cluster but still want VMs to be live-migrable between the two

Which leads back to the issue that the current top comment mentions. Intel has been caught using vendor Id checks to undermine performance on competing CPUs. Hence keeping the identity mutable may be necessary to protect against very real and very well known bad actors.

I suppose, but what would be the use case of this? It definitely won't in the "reliable benchmark" case, as the check could be just patched out.

> as the check could be just patched out

You could still sign the result with the CPU ID key.

But, then, the result can be completely bogus. The benchmark itself would need to be something that can't be faked, but a proof-of-work benchmark would end up being a very synthetic one, and very biased towards a limited set of functions.

The same functionality could and would be abused to make sure you were running a specific version of Internet Explorer on a certain version of Microsoft Windows (or equivalent locked down system) to get access to a website. Freedom relies on the the authority boundaries created by open protocols, and such "remote attestation" schemes would be the death of it.

For those who are not using Windows as the main daily driver: On Linux, you can just run the `lscpu` command, which reads info from the /proc/cpuinfo file, but there is a CPU-X[1] GUI app too. Free software gathers information on the CPU, motherboard and more. Similar to CPU-Z for Windows but works on Linux and FreeBSD. It isn't possible to trick /proc/cpuinfo on Linux, it always gave me the correct info.

[1] https://x0rg.github.io/CPU-X/

> It isn't possible to trick /proc/cpuinfo on Linux, it always gave me the correct info.

Is this because Linux reads this once at boot time, before you have a chance to modify it? If so, I assume you could still modify the string from the firmware, bootloader, or simply by modifying the kernel.

I don't know exactly how it's maintained, but you're 100% right -- it can be modified elsewhere

For example, on my AMD Threadripper system I can make a VM that pretends to be essentially any Xeon

It may perform horribly where features/extensions don't line up, but it'll report whatever.

For physical devices it's possible too, just not nearly as common

Here is what:

     ls -l /proc/cpuinfo

It says (basically it is a pseudo-filesystem with read-only permission, so I don't think so it is possible without making any changes at the Linux kernel/driver level):

     -r--r--r-- 1 root root 0 Oct 28 09:38 /proc/cpuinfo

Indeed, the proc filesystem is well maintained by the kernel. I understand that part, less so if it's properly immutable/irreplaceable

The thing is, it has to get the information from somewhere -- and depending on the age/type of the hardware (physical or virtualized), masking this is either easy or nearly impossible

I remember changing this information (DMI?) was more common back in the like Pentium 3 era, for example.

Others have helped me understand this is part of the CPUID thing -- you know, the core of the topic for the overall post!

I appreciate it everyone, save your fingers :)

I've seen this done for physical chips before too, but I can't recall anything - hence vagueness

I wonder if one could bind-mount some other proc file (say /proc/$EVIL_PID/cmdline) over cpuinfo to fake content. Would meet root and be visible in mount output though.

You definitely can and things like that are precisely why bind mounts (usually) require root, because POSIX is built very much on the premise of one filesystem tree seen by both privileged and unprivileged users.

Other systems do better by not depending on that.

It would be easier to intercept the open() for the /proc/cpuinfo file with a LD_PRELOAD library. Of course that would still be detectable by the benchmark. You could also use a modified libc to spoof it. Then their only way out would be static linking.

You could even just modify the kernel to report arbitrary strings in /proc/cpuinfo.

So they use the CPUID instruction instead. The CPUID instruction can be trapped to throw a SIGSEGV instead of returning real values on x86 with arch_prctl(ARCH_SET_CPUID, 0). So an injected SIGSEGV handler could then spoof it.

But even then, you could also trap the CPUID instruction in the kernel, and spoof it from there, which would be even harder to detect from user space.

In the end, the benchmark program always needs to trust the kernel. Is it really worth trying to detect spoofing?

It's because hypervisor intercepts the call for CPUID and returns what you told you to.

It's usually used to downgrade a set of features if you want VM to be able to be live-migrated between machines with different CPUs, we migrated a bunch of stuff off Intel to AMD that way. There is a performance hit on that as you're basically making virtual "lowest common denominator" CPU

Thank you! I was aware of the downgrading for live-migration, but not that there was a particular call for it

That's a VM. It will be trapping CPUID at the instruction level and returning whatever the VM is set up as.

I'm not sure why anyone would trust random Geekbench database entries in any case, there's no way to authenticate anything. I suspect you can just edit the process memory after running a benchmark and before submitting the results, to save yourself the trouble of reverse engineering their web API.

Yeah why do you even need to change the CPU ID, this would be much easier.

This was reaffirmed however long ago virtualization came about

This already has been a problem in the 80ies when Intel VM86 mode came into use - real mode programs tested the unused bits in the status register and then made assumptions about the CPU (as recommended by Intel in the programming manuals). When the program was running inside a VM86 environment (virtual real mode inside of protected mode), the instruction for this triggered an protection fault, and the "hypervisor" filled in the results as if it got executed correctly, potentially confusing programs about which CPU they were running on.

Reply on news.ycombinator.com