I mistakenly believed that only RDSEED could fail. However, the Intel
manuals state that RDRAND can fail too.
I can't actually observe it failing, even with all cores running RDRAND in a
tight loop. In any case, the ChaCha20 masking means that it wouldn't be
a big deal anyway.
Still, this change tests the carry flag after RDRAND and the code falls
back to |CRYPTO_sysrand| if RDRAND has a hiccup. (The Intel manuals
suggest[1] calling RDRAND in a loop, ten times, before considering it to
have failed. But a single failure appears to be such a rare event that
the complexity in the asm code doesn't seem worth it.)
This change also adds an asm function to fill a buffer with random data.
Otherwise the overhead of calling |CRYPTO_rdrand|, and bouncing the data
in and out of memory starts to add up.
Thanks to W. Mark Kubacki, who may have reported this. (There's some
confusion in the bug report.)
Before:
Did 6148000 RNG (16 bytes) operations in 1000080us: 98.4 MB/s
Did 649000 RNG (256 bytes) operations in 1000281us: 166.1 MB/s
Did 22000 RNG (8192 bytes) operations in 1033538us: 174.4 MB/s
After:
Did 6573000 RNG (16 bytes) operations in 1000002us: 105.2 MB/s
Did 693000 RNG (256 bytes) operations in 1000127us: 177.4 MB/s
Did 24000 RNG (8192 bytes) operations in 1028466us: 191.2 MB/s
[1] Intel Reference Manual, section 7.3.17.1.
Change-Id: Iba7f82e844ebacef535472a31f2dd749aad1190a
Reviewed-on: https://boringssl-review.googlesource.com/5180
Reviewed-by: Adam Langley <agl@google.com>
2cac3506fa