Conversation

TIL: The most popular 64-bit architecture in the world has a small address space. Kills any hope of never reusing malloc addresses.

Quote Tweet

Daniel Micay

@DanielMicay

Mar 22, 2021

Replying to @DanielMicay @pcwalton and 2 others

On arm64, you typically only have a 39-bit address space although it can be 48-bit like x86_64. The kernel takes half so it's really 38-bit or 47-bit. 38-bit is already far too small to use hardened_malloc with the normal configuration. 47-bit is 128 TiB which isn't limitless.

Replying to

Wait... is this physical or virtual?

Replying to

and

The per-process virtual memory. Typically: 128TiB for userspace on x86_64 and only 256GiB for userspace on arm64. It may still seem like a lot but consider ASLR offsets, all the use cases for shadow data (CFI, ASan, etc.), memory mapped files including databases, stacks, etc.

Replying to

and

No, I think that seems low. Anyway aarch64 is still predominantly on what was traditionally considered embedded platforms. I don’t expect this to remain unchanged but we will see as the desktop use and perhaps even HPC cases increase. I get around 1073 GFLOPS on my Mac mini.

Replying to

and

For ARMv8.3 and beyond, increasing it means giving up on pointer authentication working properly. Existing platforms tend to already support using a substantially larger address space such as 48-bit via 4-level page tables and 4k pages. It just isn't being done in practice.

Replying to

and

Apple uses pointer authentication and they aren't going to want to give up on it to have a larger address space. I don't expect them to do that. The feature inherently needs a bunch of otherwise unused bits in pointers so using larger pages + deeper pages tables doesn't help.

Replying to

and

github.com/torvalds/linux 3-level page tables + 4k pages is near universal for Linux. We're using 4-level page tables + 4k pages for 48-bit address space (like typical x86_64) on GrapheneOS for ARMv8.2 and below. We'll go back for ARMv8.3 and above due to pointer auth though...

github.com

linux/Kconfig at 79cc2ed5a716544621b11a3f90550e5c7d314306 · torvalds/linux

Linux kernel source tree. Contribute to torvalds/linux development by creating an account on GitHub.

Replying to

and

For HPC, using the larger address space would be expected, but on a server I think most people are going to want to have pointer authentication. I think people are going to need to get used to address space not being nearly as plentiful as they actually think it is on 64-bit.

11:36 AM · Mar 23, 2021Twitter Web App

Replying to

and

I agree with your overall assessment.