sequoia-openpgp v2.0.0-alpha.2 test failure on i686 (32-bit x86)

I found one test failure when doing a test build of sequoia-openpgp v2.0.0-alpha.2 on Fedora Rawhide, specific to 32-bit x86 (i686), i.e. on the i686-unknown-linux-gnu target.

I have re-run the same build twice, and the error persisted, so I don't think it's a fluke. The crash appears to happen in code from the argon2 crate.

Output from building with default features (i.e. the crypto-nettle backend):

---- cert::test::v6_minimal_cert stdout ----

thread 'cert::test::v6_minimal_cert' panicked at /usr/share/cargo/registry/argon2-0.5.3/src/lib.rs:230:26:
capacity overflow
note: run with `RUST_BACKTRACE=1` environment variable to display a backtrace

If it helps, I can re-run tests with RUST_BACKTRACE=1.

It's not clear whether this test also fails with other crypto backends, since my test builds test the Nettle backend first, and it fails the build before the tests are run with OpenSSL or RustCrypto backends.

Note that we also run the test suite of the argon2 crate on all architectures and I've not seen any issues there (but of course it's entirely possible that this code path is just not covered by its test suite).

Relevant environment:

• Rust / cargo version: 1.85.0
• Nettle 3.10.1, OpenSSL 3.2.4

changed the description

Closer investigation shows that this error happens due to Vec::with_capacity panicking because of a capacity overflow.

Argon2 needs a lot of memory, that is kinda the point of Argon2. 32-bit architectures may even run out of address space.

I tried tracing the error to its origin - it appears to be an integer overflow in requested size of an array:

https://doc.rust-lang.org/stable/src/core/alloc/layout.rs.html#462-469

Caused by this allocation here:

https://docs.rs/argon2/latest/src/argon2/lib.rs.html#230

I looked at the code for the cert::test::v6_minimal_cert test, and I don't see any special sauce there regarding Argon2 parameters, they are set here:

https://docs.rs/sequoia-openpgp/2.0.0-alpha.2/src/sequoia_openpgp/crypto/s2k.rs.html#192-217

Does this mean that all operations that involve Argon2 in sequoia-openpgp on 32-bit architectures would hard-crash the program?

Reading the argon2 crate source code, it looks like the block size depends on the "m" parameter, maybe that can be tweaked so it doesn't allocate so much memory that it overflows the maximum Vec size on 32-bit architectures?

mentioned in commit fd65d509

mentioned in commit 3325c74c

closed with merge request !1766 (merged)