Android crash after removal of ring dependency

[zzorba]

We have pulled in a recent version of the matrix-rust-sdk into our app, and have noted that the android app seems to now crash on initialization (on emulators, and perhaps real devices). After some digging, it appears to be related to the recent change to remove the 'ring' dependency: https://github.com/matrix-org/matrix-rust-sdk/pull/6334/changes

Specifically, it looks like now that reqwest pulls in rustls which defaults to aws-lc-rs. It seems like the aws-lc-rs crate has known issues on certain Android targets/emulators — it can fail to initialize or produce invalid code on arm64 emulators with 16KB pages. The crash we got seemed to indicate an issue with 16kb page sizes specifically, and restoring ring on android seemed to fix the issue.

Not sure on the best way to fix this or not, but figured a post about the issue we're experiencing might help others in a similar situation.

[poljar]

Is this different from #6368?

Is there a log or trace output for this?

[jmartinesp]

it can fail to initialize or produce invalid code on arm64 emulators with 16KB pages.

AFAIK, the fact that something uses 16KB pages or not is only based on the NDK version your Rust code is built with, since we build everything from source and there shouldn't be any pre-built binaries. It shouldn't matter which library you're using under the hood.

[mgoldenberg]

Tagging aws/aws-lc-rs#969, as it seems like it might be relevant.

[zzorba]

So I'm not certain, but I'm pretty sure I'm directing my compiler/linker to use the correct 16kb page boundaries, however the aws-lc-rs issue @mgoldenberg tagged there is still causing an issue. I've confirmed the page boundaries in the produced binaries are correct.

Here's what I've found:

1. Crash: SIGSEGV — "trying to execute non-executable memory" from .bss
2. Device: sdk_gphone16k_arm64 — this is the 16KB page size emulator
3. aws-lc-rs symbols are linked into libdiode.so (we confirmed AWSLC_* symbols earlier)
4. aws-lc-rs uses runtime code generation that copies trampolines to .bss and tries to execute them — on 16KB pages, the page protection boundaries differ, so the trampoline lands in non-executable memory

The ELF segment alignment we fixed is correct (0x4000), but aws-lc-rs has an internal issue where its assembly code assumes 4KB pages when setting up executable trampolines at runtime. The linker flag can't fix that — it's a bug in aws-lc-rs's assembly/C code.

Excluding aws-lc-rs in favor of the old ring library seems to be a workable solution while we wait on the upstream changes