Timing variability of any kind is problematic when working with potentially secret values such as
elliptic curve scalars, and such issues can potentially leak private keys and other secrets. Such a
problem was recently discovered in curve25519-dalek.
The Scalar29::sub (32-bit) and Scalar52::sub (64-bit) functions contained usage of a mask value
inside a loop where LLVM saw an opportunity to insert a branch instruction (jns on x86) to
conditionally bypass this code section when the mask value is set to zero as can be seen in godbolt:
- 32-bit (see L106): https://godbolt.org/z/zvaWxzvqv
- 64-bit (see L48): https://godbolt.org/z/PczYj7Pda
A similar problem was recently discovered in the Kyber reference implementation:
https://groups.google.com/a/list.nist.gov/g/pqc-forum/c/hqbtIGFKIpU/m/cnE3pbueBgAJ
As discussed on that thread, one portable solution, which is also used in this PR, is to introduce a volatile read as an optimization barrier, which prevents the compiler from optimizing it away.
The fix can be validated in godbolt here:
- 32-bit: https://godbolt.org/z/jc9j7eb8E
- 64-bit: https://godbolt.org/z/x8d46Yfah
The problem was discovered and the solution independently verified by Alexander Wagner [email protected] and Lea Themint [email protected] using their DATA tool: