This imports upstream's scrypt implementation, though it's been heavily
revised. I lost track of words vs. blocks vs. bigger blocks too many
times in the original code and introduced a typedef for the fixed-width
Salsa20 blocks. The downside is going from bytes to blocks is a bit
trickier, so I took advantage of our little-endian assumption.
This also adds an missing check for N < 2^32. Upstream's code is making
this assumption in Integerify. I'll send that change back upstream. I've
also removed the weird edge case where a NULL out_key parameter means to
validate N/r/p against max_mem and nothing else. That's just in there to
get a different error code out of their PKCS#12 code.
Performance-wise, the cleanup appears to be the same (up to what little
precision I was able to get here), but an optimization to use bitwise
AND rather than modulus makes us measurably faster. Though scrypt isn't
a fast operation to begin with, so hopefully it isn't anyone's
bottleneck.
This CL does not route scrypt up to the PKCS#12 code, though we could
write our own version of that if we need to later.
BUG=chromium:731993
Change-Id: Ib2f43344017ed37b6bafd85a2c2b103d695020b8
Reviewed-on: https://boringssl-review.googlesource.com/17084
Reviewed-by: Adam Langley <agl@google.com>
b529253bea