Mathias,
I'm seeing some anomalous results with the "by8" AVX CTR optimization in
3.18.
the patch you're replying to actually *disabled* the "by8" variant for
v3.17 as it had another bug related to wrong counter handling in GCM.
The fix for that particular issue only made it to v3.18, so the code
got re-enabled only for v3.18. But it looks like that there's yet
another bug :/
Right, I should have clarified that I initially suspected the "by8"
variant was to blame because your patch that disables it resolves the
discrepancy.
In particular, crypto_aead_encrypt appears to produce different ciphertext
from the same plaintext depending on whether or not the optimization is
enabled.
See the attached patch to tcrypt that demonstrates the discrepancy.
I can reproduce your observations, so I can confirm the difference,
when using the "by8" variant compared to other AES implementations.
When applying this very patch (commit 7da4b29d496b ("crypto: aesni -
disable "by8" AVX CTR optimization")) -- the patch that disables the
"by8" variant -- on top of v3.18 the discrepancies are gone. So the
behavior is bound to the "by8" optimization, only.
Right -- this is exactly what I'm seeing as well.
As it was Chandramouli, who contributed the code, maybe he has a clue
what's wrong here. Chandramouli?
A few more observations:
* Encryption produces bad ciphertext only when the size of plaintext
exceeds a certain threshold. In test_aead_encrypt_consistency in the
tcrypt patch, I found that data_size must be >= 128 to produce bad
ciphertext.
* Encrypting then decrypting data always gets back to the original
plaintext, no matter what the size.
* The bad ciphertext from encryption is only evident when the same
encrypt operation is performed on a different AES implementation and the
ciphertexts are compared.
* When the encrypt operation produces bad ciphertext, the generated auth
tag is actually correct, so another AES implementation that decrypts the
ciphertext will end up with corrupted plaintext that succeeds
authentication.
James
--
To unsubscribe from this list: send the line "unsubscribe linux-crypto" in
the body of a message to majordomo@xxxxxxxxxxxxxxx
More majordomo info at http://vger.kernel.org/majordomo-info.html
