rdrand.md

category	description
Gamedev	🎲 Get random numbers 🎲

rdrand

Audit

version	thoroughness	understanding	rating	notes
0.6.0	medium	medium	neutral	Diff review
0.5.2	medium	medium	neutral	Diff review
0.4.0	medium	medium	negative	Full review

0.6.0 / 0.5.2

Sound, although buggy AMD hardware makes implementing CryptoRng for RdRand/RdSeed a little unnerving to me ( nagisa/rust_rdrand#12 )

Reviewed

• benches\rdrand.rs: +1
• benches\rdseed.rs: +1
• benches\std.rs: +1
• src\changelog.rs: +1
• src\lib.rs: 0
  • All of $gen::try_fill_bytes::imp being unsafe is still concerning... but I believe it's sound now.
  • RdRand/RdSeed implement CrytoRng, which makes buggy AMD hardware concerning: nagisa/rust_rdrand#12

Concerns reduced since 0.4.0:

• https://github.com/nagisa/rust_rdrand/commit/7af432c6e315fde053d0056d1b7df893a865711a
  • unsafe blocks appear much larger than they need to be.

Concerns fixed since 0.4.0:

• https://github.com/nagisa/rust_rdrand/commit/26a0a2f9d885fbbb8e14fa47c8a48e366cf15455
  • mem::uninitialized() u32s
  • loop_rand! uses mem::uninitialized() for $el:ty, easy to misuse! Requires unsafe{} so technically sound. Not exported.
• https://github.com/nagisa/rust_rdrand/commit/7af432c6e315fde053d0056d1b7df893a865711a
  • $gen::try_fill_bytes: UNSOUND! word and buffer reference the same data. As both are &mut Ts, this is 100% undefined behavior.
  • ptr::copy_nonoverlapping: This should really use a slice copy which should be just as safe...? But maybe missing from core?
• Verified vs https://www.amd.com/system/files/TechDocs/24594.pdf
  • is_x86_feature_detected: I have not verified this is correct.

0.4.0

$gen::try_fill_bytes invokes undefined behavior (overlapping &mut u32 and &mut [u8]): nagisa/rust_rdrand#13. 0.5.x removed some use of uninitialized.

Reviewed

• benches\rdrand.rs: +1
• benches\rdseed.rs: +1
• benches\std.rs: +1
• src\changelog.rs: +1
• src\lib.rs: Concerns:
  • mem::uninitialized() u32s
  • is_x86_feature_detected: I have not verified this is correct.
  • loop_rand! uses mem::uninitialized() for $el:ty, easy to misuse! Requires unsafe{} so technically sound. Not exported.
  • $gen::try_fill_bytes: UNSOUND! word and buffer reference the same data. As both are &mut Ts, this is 100% undefined behavior.
  • unsafe blocks appear much larger than they need to be.
  • ptr::copy_nonoverlapping: This should really use a slice copy which should be just as safe...? But maybe missing from core?