AESNI version of CCM+CBC-MAC

[sef]

Summary

This adds the much faster AESNI version of the CCM and CBC encryption/authentication.

Test Plan

cryptocheck -A 0 -a aes-ccm -d aesni 1024
cryptocheck -A 10 -a aes-ccm -d aesni 1024

[cem]

sys/crypto/aesni/aesni_ccm.c

70

This file is more or less new code; it's free to follow style(9) from the beginning.

So nitpicks here: Blank line between variable declarations and first statement; and return (retval);.

82

ditto the blank line thing

99

Seems like retval could be named something more descriptive, like B0 or MAC.

And temp_block is maybe auth_block or something (staging_block would match software CBC-MAC).

109

Does __m128i foo = 0; generate different code than __m128i foo = _mm_setzero_si128();? It isn't important, I'm just vaguely curious.

Edit: Ahhh, __m128i is literally a vector, rather than integral, type, according to the compiler. So it cannot simply be initialized with an integer.

129–130

Could just use

be16enc(&temp_block, auth_len);

134–136

What do you mean by calling convention? The immediate function takes a size_t, which can represent larger values.

I understand that internally OCF deals only in int sizes, but I'd suggest either implementing the 3rd mode or asserting we're within range of the 2nd mode anyway.

140–143

Simplified:

be16enc(&temp_block, 0xfffe);
be32enc((char *)&temp_block + 2, auth_len);

162

style nit: sizeof(temp_block), for consistency with the object we're acting on

185

typo: abytes

186

usually spelled "4 GB" or "4 gigabytes" but not "4gbytes"

188–190

I don't know what these pre-processor error directives (there are a few) are intended to be, but probably remove them before commit. (Figure out how to incorporate them into comments or code, if necessary.)

193–196

Ugh. I know you just inherited this mess from GCM, but what an obnoxious API :P.

205–206

Is it impossible to generate a tag for AAD without any ciphered input?

230–231

You don't need explicit_bzero for ordinary initializations. Here _mm_setzero_si128 would be fine. explicit_bzero is for deleting private keys when you release memory.

Also, current_block is never used?

258

So really last_block is the rolling MAC value. I'd suggest a name incorporating MAC, for clarity.

259–262

Right, ok, we're generating a keystream from an encrypted series of counter blocks.

262

style nit: Usually we at declare variables at the beginning of some scope, if not function scope.

272

This step doesn't have any dependency on the data, right? We could do it earlier, as soon as we have saved a copy in s_x for use in generating the keystream.

275

style: Empty return

Here is where you would explicit_bzero() sensitive data, like maybe last_block, s0, temp_block, and X.

302

maybe this is just phabricator, but it looks like there are extra spaces between __m128i and s0

315

style(9) nit: pointers aren't booleans; compare with null in expressions. Ditto below.

334

It's not a hash. It's a MAC, or tag.

342–344

You already zero padded temp_block on lines 322-325.

346

Not that they're different types, but since this acts on pad_block, sizeof(pad_block) is slightly more appropriate

362

empty return

probably also the last block of decrypted plaintext is sensitive (temp_block)? And perhaps the tag, hash_block.

382–383

No verifying of the tag over the AAD?

418

s_x is never used?

426

timingsafe_bcmp

427

style(9) nit: return (0);

[sef]

In D19298#425490, @cem wrote:

In this implementation, I'm not sure the 2-pass decrypt buys us anything. We already pre-allocate space (if input is non-contiguous) — checking the tag first doesn't save us from a DoS in terms of memory use. If the (unexpected case) tag verification fails, we'd need to do extra work to clobber anything we output, maybe? I might be forgetting something about CCM.

The reason I still did it that was was because AESNI_CCM_decrypt doesn't allocate the memory -- the caller does, and if something else were to call it (or the aesni wrapping code ever changes how it gets the buffers), it'd expose data it shouldn't.

[sef]

(I'll go through the rest of the comments this weekend.)

[cem]

In D19298#425522, @sef wrote:

In D19298#425490, @cem wrote:

In this implementation, I'm not sure the 2-pass decrypt buys us anything. We already pre-allocate space (if input is non-contiguous) — checking the tag first doesn't save us from a DoS in terms of memory use. If the (unexpected case) tag verification fails, we'd need to do extra work to clobber anything we output, maybe? I might be forgetting something about CCM.

The reason I still did it that was was because AESNI_CCM_decrypt doesn't allocate the memory -- the caller does, and if something else were to call it (or the aesni wrapping code ever changes how it gets the buffers), it'd expose data it shouldn't.

That's totally reasonable.

[sef]

sys/crypto/aesni/aesni_ccm.c

205–206

Huh. I hadn't thought about that, but looks like it can. Thanks! (I'm working on the other feedback as well.)

[jhb]

Just a couple of thoughts from what I've seen flying by:

1. I think it's correct to decrypt twice (but I think everyone agrees on that now)
2. I have patches to make ccr(4) do CCM.
3. It certainly seems like you should be able to do a CCM request that only has AAD and no payload. GCM supports this and the NIST-KAT tests for GCM include examples of this.
4. As part of my ccr(4) work I'm working on adding the CCM NIST-KAT vectors to the security/nist-kat port and hooking those up to the python tester (I thought we had asked for that earlier as part of the CCM changes, but I don't see them in the tree and haven't seen them in a review, so I'm just going to work on it as part of testing ccr(4)).

[sef]

New diff coming right after this.

sys/crypto/aesni/aesni_ccm.c

114

I find memcpy backwards, bcopy sane. But it's also kernel code, and bcopy/bzero is more prevalent than memcpy/memset.

129–130

ha. I did not know about those. Much nicer.

186

I deal with bits (networking stuff) so often that my habit is to specify Xbits or Xbytes as appropriate.

188–190

Old debugging code, I've gotten rid of all of them.

193–196

I hate any function with more than 6 arguments. That's from compiler work decades ago ;).

258

I am, *always*, horrible at naming things. I changed it to "rolling_mac".

272

I put it there because of how the spec was written.

302

There was in at least one place, fixed.

342–344

Done, I believe.

362

Is the tag sensitive? I've got it in the new changes, but since it's part of the message, I'm not sure it's sensitive?

382–383

I changed both of them to check for message and aad len being zero for early exit.

418

Good catch, holdover from the older code.

[sef]

Review feedback incorporated.

[cem]

Looking better, thanks

sys/crypto/aesni/aesni_ccm.c

114

bcopy/bzero is more prevalent than memcpy/memset.

How did you arrive at that conclusion? Recursive grep suggests the opposite is true.

136

I think we lost 0xfffe?

186

Being explicit about bytes is harmless, but we're missing a space and a full SI-prefix.

258

Thanks

362

This is the correct tag, as opposed to the one provided by the attacker in the message. It's sensitive, no?

[sef]

I lost the length descriptor prefix in the last change. I've put it back, and run cryptocheck with -A lengths of 0, 13, 16, 32, 192102, and 127091. (User-space can't test more than 256k unfortunately.)

[sef]

sys/crypto/aesni/aesni_ccm.c

114

Older check. "_was_ more prevalent" it turns out.

136

Yeah, put back and explicitly tested the larger AAD lengths.

[sef]

Poke?

[jhb]

Only some small suggestions / nits from me. I think it looks good overall. Does this pass the limited subset of CCM vectors the latest python script test runs?

sys/crypto/aesni/aesni.c

134

FWIW, I've sorted these alphabetically in the manpage for ccr and think it wouldn't hurt to do the same here.

138

Unrelated, it seems like SHA224 is missing (that might be a cem@ bug and a separate commit to fix).

852

So we do have AES_CBC_MAC_HASH_LEN which is what I used in ccr(4) as the equivalent of GMAC_DIGEST_LEN. I prefer the explicit constants as right now this is hardcoding the assumption that CCM and GCM have the same tags. We could instead add a _Static_assert() that 'sizeof(tag) >= GMAC_DIGEST_LEN' for the GCM case and something similar for CCM.

sys/crypto/aesni/aesni_ccm.c

69

Nit: s/put/Put/

100

FWIW, this will be the 3rd copy of this code in the tree (plain software generates the B0 block, ccr has a 'generate_ccm_b0' function). Someday we should export a shared helper function, but we don't have to do that as part of this change.

212

For actual CCM I think the limits are 7 and 13.

[sef]

I'm updating and build testing before uploading new diffs.

sys/crypto/aesni/aesni.c

138

I didn't touch that -- but it's possible I'm out of date. I'll do an update before I upload these modified diffs.

sys/crypto/aesni/aesni_ccm.c

69

Keyboard issues.

212

That may be right, the spec just said "15-L octets"; it clearly can't be 0, since it says it must also be unique. But the other constraints on L aren't as clear to me.

[sef]

The promised updated diffs.

[cem]

I think most feedback has been incorporated, but I think a few things weren't addressed yet:

sys/crypto/aesni/aesni.c

138

Right, I didn't enumerate SHA224 independent of SHA256 (nor the HMAC varieties) in the device string. It's really just SHA256 with a different initial vector and truncated result. Maybe it should be included (unrelated to CCM).

sys/crypto/aesni/aesni_ccm.c

186

still present

212

NIST 800-38c §A.1 states n ∈ [7, 13].

310

I still find hashp and hash_block problematic here; these have specific cryptographic meaning that AES-CBC-MAC does not satisfy. macp/mac_block (or tagp/tag_block) would be totally fine.

[sef]

I'll get to those tomorrow, thanks! (Various stuff going on today has thrown my schedule off.)

[sef]

Doing testing with the changes, then I'll upload the diffs.

[sef]

A bit late, but the updated diff promised. Thanks cem & jhb!

[cem]

Thanks! I believe all my concerns are addressed. (And thanks again for your continued patience.)

[ngie]

1. This broke the build with gcc:

11:40:34 --- all_subdir_aesni ---
11:40:34 /workspace/src/sys/crypto/aesni/aesni_ccm.c: In function 'xor_and_encrypt':
11:40:34 /workspace/src/sys/crypto/aesni/aesni_ccm.c:61:18: error: incompatible types when initializing type '__m128 {aka __vector(4) float}' using type '__m128i {aka __vector(2) long long int}'
11:40:34   __m128 retval = _mm_xor_si128(a, b);
11:40:34                   ^~~~~~~~~~~~~
11:40:34 /workspace/src/sys/crypto/aesni/aesni_ccm.c:63:21: error: incompatible type for argument 3 of 'aesni_enc'
11:40:34   retval = AESNI_ENC(retval, k, nr);
11:40:34                      ^
11:40:34 /workspace/src/sys/crypto/aesni/aesni_ccm.c:47:64: note: in definition of macro 'AESNI_ENC'
11:40:34  #define AESNI_ENC(d, k, nr) aesni_enc(nr-1, (const __m128i*)k, d)
11:40:34                                                                 ^
11:40:34 In file included from /workspace/src/sys/crypto/aesni/aesni_ccm.c:46:0:
11:40:34 /workspace/src/sys/crypto/aesni/aesencdec.h:115:1: note: expected '__m128i {aka const __vector(2) long long int}' but argument is of type '__m128 {aka __vector(4) float}'
11:40:34  aesni_enc(int rounds, const __m128i *keysched, const __m128i from)
11:40:34  ^~~~~~~~~
11:40:34 /workspace/src/sys/crypto/aesni/aesni_ccm.c:64:9: error: incompatible types when returning type '__m128 {aka __vector(4) float}' but '__m128i {aka __vector(2) long long int}' was expected
11:40:34   return (retval);
11:40:34          ^
11:40:34 /workspace/src/sys/crypto/aesni/aesni_ccm.c:65:1: error: control reaches end of non-void function [-Werror=return-type]
11:40:34  }
11:40:34  ^

2. Why was this committed without signoff from someone on secteam (CC: secteam)?

CC: @lwhsu

[cem]

In D19298#440806, @ngie wrote:

1. This broke the build with gcc:

I just noticed the same thing :-). Should be fixed in r348293.

2. Why was this committed without signoff from someone on secteam?

Very rarely do we have secteam sign off on new cipher implementations. This cipher is intended to allow importing encrypted zpools from other/older OpenZFS implementations, but should be discouraged from use for literally any other purpose. This is part of a series; arguably the same question could be raised for rS346650, rS346649, rS344142, rS344141 (D19090), and rS344140.

In my experience, Secteam is way overloaded (it's just delphij and gordon). I would love it if someone on secteam would take a critical eye to this series of changes, but I don't know if it is the best use of their limited time.

[lwhsu]

In D19298#440809, @cem wrote:

In D19298#440806, @ngie wrote:

1. This broke the build with gcc:

I just noticed the same thing :-). Should be fixed in r348293.

Thanks for the fix! I was going to file a ticket for asking help. :-)

2. Why was this committed without signoff from someone on secteam?

Very rarely do we have secteam sign off on new cipher implementations. This cipher is intended to allow importing encrypted zpools from other/older OpenZFS implementations, but should be discouraged from use for literally any other purpose. This is part of a series; arguably the same question could be raised for rS346650, rS346649, rS344142, rS344141 (D19090), and rS344140.

In my experience, Secteam is way overloaded (it's just delphij and gordon). I would love it if someone on secteam would take a critical eye to this series of changes, but I don't know if it is the best use of their limited time.

I am not familiar with the process here, but I guess it would be nice to have a post-commit notification for them, in case we missed anything.