"On x86-64, there are two CPU settings which control the kernel’s ability to access memory."
There are a couple more than two, even in 2021.
Memory Protection Keys come to mind, as do the NPT/EPT tables when virtualization is in play. SEV and SGX also have their own ways of preventing the kernel from writing to memory. The CPU also has range registers that protect certain special physical address ranges, like the TDX module's range. You can't write there either.
That's all that comes to mind at the moment. It's definitely a fun question!
a thought: do MPK actually control the kernel's ability to access memory? on intel, i think if you try to read that memory, a page fault wont be thrown. although with PKS, kernel reads will cause a page fault.
so can the kernel (ring0) freely read/write to memory encrypted with MPK? I think so, yes. good luck with whatever happens next tho lol
But the point here is that userspace can use this to bypass kernel protections that would otherwise prevent it from mutating R^X pages for example, not that the kernel can bypass its own.
not entirely, IOMMU is a thing, that is IIRC how Amazon and other hyperscalers can promise you virtual machines whose memory cannot be touched even in the case the host is compromised (and, by extension, also if the feds arrive to v& your server).
>how Amazon and other hyperscalers can promise you virtual machines whose memory cannot be touched even in the case the host is compromised (and, by extension, also if the feds arrive to v& your server).
Even if we take those promises at face value, it practically doesn't mean much because every server still needs to handle reboots, which is when they can inject their evil code.
TL;DR: when a user writes to /proc/self/mem, the kernel bypasses the MMU and hardware address translation, opting to emulate it in software (including emulated page faults!), which allows it to disregard any memory protection that is currently setup in the page tables.
It doesn't bypass it exactly, it's still accessing it via virtual memory and the page tables. It's just that the kernel maintains one big linear memory map of RAM that's writable.
"On x86-64, there are two CPU settings which control the kernel’s ability to access memory."
There are a couple more than two, even in 2021.
Memory Protection Keys come to mind, as do the NPT/EPT tables when virtualization is in play. SEV and SGX also have their own ways of preventing the kernel from writing to memory. The CPU also has range registers that protect certain special physical address ranges, like the TDX module's range. You can't write there either.
That's all that comes to mind at the moment. It's definitely a fun question!
a thought: do MPK actually control the kernel's ability to access memory? on intel, i think if you try to read that memory, a page fault wont be thrown. although with PKS, kernel reads will cause a page fault.
so can the kernel (ring0) freely read/write to memory encrypted with MPK? I think so, yes. good luck with whatever happens next tho lol
I'm still surprised I was the first one to notice when Linus tried to change this - I always thought it was a pretty well known behavior.
The kernel owns the page tables. It can always find another way in.
But the point here is that userspace can use this to bypass kernel protections that would otherwise prevent it from mutating R^X pages for example, not that the kernel can bypass its own.
> The kernel owns the page tables.
not entirely, IOMMU is a thing, that is IIRC how Amazon and other hyperscalers can promise you virtual machines whose memory cannot be touched even in the case the host is compromised (and, by extension, also if the feds arrive to v& your server).
>how Amazon and other hyperscalers can promise you virtual machines whose memory cannot be touched even in the case the host is compromised (and, by extension, also if the feds arrive to v& your server).
Even if we take those promises at face value, it practically doesn't mean much because every server still needs to handle reboots, which is when they can inject their evil code.
MK-TME allows having memory encrypted at run time, and the platform TPM signs an attestation saying the memory was not altered.
Malicious code can't be injected at boot without breaking that TPM.
Subject to the huge caveat that the attacker does not have physical access. https://tee.fail/
An interesting implementation flaw, but not a conceptual problem with the design.
If your threat model is being v& by feds, maybe you should keep your server at home behind Tor.
TL;DR: when a user writes to /proc/self/mem, the kernel bypasses the MMU and hardware address translation, opting to emulate it in software (including emulated page faults!), which allows it to disregard any memory protection that is currently setup in the page tables.
It doesn't bypass it exactly, it's still accessing it via virtual memory and the page tables. It's just that the kernel maintains one big linear memory map of RAM that's writable.
Thank You.