OpenSSL Feature Requests and Priorities
All,
We have been asked by the BAC to provide a list of features or capabilities that the large business community (LBC) would like OpenSSL to deliver (or focus on).
In that spirit can each of you provide a prioritized list of things that your large business would like? For extra credit, why it's important would also be valuable.
For example, Cisco and Comcast believe that DTLS1.3 should be prioritized and delivered ASAP. DTLS1.3 is heavily used in VPN as well as streaming platforms where the installed base cannot be easily migrated to QUIC. Since PQ cannot be supported on DTLS1.2 the OpenSSL doesn't fully support PQ until DTLS1.3 is delivered. So PQ and DTLS1.3 are very important to Cisco as well as Comcast.
Maybe this isn't the case for your large business. But what is important to your business? Is it performance? Is it support for some new RFC? What would be valuable to your business.
This forum provides the opportunity to have your say into the OpenSSL roadmap. Please, help OpenSSL by having your voice heard!
thx,
-jj
No problem at all! I'm looking forward to some turkey as well! Have a good one!!
For Red Hat, we're interested in:
DTLS 1.3, for PQC in non-HTTP protocols that have not or will not move to QUIC
Performance; not necessarily because we have big problems with the current performance, but because we really want to avoid more upstream projects deciding to chose some other fork as their default that we'd then have to revert downstream
A better way to phase out SHA-1. We have downstream patches, but a better upstream method would be good. See #17662 and #19084 that were an attempt to get that done, but haven't gotten anywhere since we couldn't really agree on how to best implement this.
Improving the contribution experience. I believe this helps us all long-term.
This is just a quick list, haven't had a discussion about it with the team yet. I'll ask around for some more input.
@Clemens Lang Thanks so much Clemens! Nice to hear from you! If I was able to have a meeting with LBC along with OpenSSL would you be willing to participate?
thx,
-jj
@Jeff Johnson Sure, if it works time-wise for me, I'll be happy to join.
@Clemens LangI would add the improvement of EVP_SKEY support to the list
Some preliminary thoughts.
Support for the BoringSSL QUIC support API. We need a full featured QUIC stack so use Google Quiche and having to maintain our own patches on top of OpenSSL for this is increasingly untenable. Migration to the OpenSSL one is a nonstarter given the degree of integration with our main connection processor we have.
Better tracking of backwards compatibility issues. Surprises with changes to FIPS requirements forcing support for some option shouldn't happen: we should be able to alert our internal consumers what's happening ahead of time so they can migrate.
Performance at scale absolutely matters (many hundreds of thousands of connections a second, gigabytes of bandwidth), and we don't need dynamic loading/unloading of providers. Providers were helpful in getting PQ ready early.
Better FIPS policy: we often end up having to avoid FIPS compliant constructions with an impact to security because of what OpenSSL has chosen. AES-GCM in particular has been an issue, where the security policy limits it to TLS.
Build system migration to CMake or similar. Right now we have to wrap a bunch of changes around to handle the integration with our internal packaging and development cycle, and the integration never really works right. With CMake we'd get that for free.
@Watson Ladd Thanks! I really appreciate your input! I'll compile the total list and provide them back to OpenSSL. Maybe a meeting to hear the concerns of the LBC would be good? Would you participate in such a meeting?
thx,
-jj
@Jeff Johnson Yeah, we would participate
@Watson Ladd For "Support for the BoringSSL QUIC support API. We need a full featured QUIC stack so use Google Quiche and having to maintain our own patches on top of OpenSSL for this is increasingly untenable."
You should direct that request to the Google Quiche maintainers where it properly belongs - to add support for the officially documented interface in OpenSSL for external QUIC TLS usage. This is what the other independent QUIC stacks have done - including msquic and ngtcp2. nginx also has a pure OpenSSL solution which predates the official API for external QUIC stacks to use. The old BoringSSL QUIC APIs aren't going to be added into base OpenSSL - that ship has long since sailed. And we aren't going to backfit those APIs into earlier releases (the other related request that has been raised).
Most QUIC implementations support a range of TLS crypto interfaces - the concept that there was just the one single API for BoringSSL and everyone used it and life was simple and no other changes were required isn't reality. The set of interfaces are varied and QUIC stack maintainers deal with this issue.
If you chose to sit on top of BoringSSL or a using a BoringSSL dependent package, then you have to be willing to pay the costs of doing so (something which Google makes abundantly clear in their documentation). We have never seen this as an "OpenSSL" issue to solve and we respect that other projects make other decisions about what to depend on. If you have a collection of companies that also face this same issue in terms of increasing costs of maintaining your own OpenSSL fork, then perhaps you can collectively work to contribute patches to Quiche - assuming you can convince the Quiche maintainers that this is something they should do.
I note that there are no open issues on quiche to support the official OpenSSL interfaces that I can see with a quick look.
Tim,
I'm trying to understand why you couldn't adopt the BoringSSL QUIC API instead of the more complex OpenSSL API proposal put forward. In general there's been a lot of work done by them to make safer APIs for common operations: it would be good if you could learn and use it in ways that help people writing code that need to use both.
@Watson Ladd can you elaborate more on FIPS policy and also about AES-GCM?
About AES-GCM. It is approved. One can trivially use it with internally-generated IVs => ask fips provider to generate IV, encrypt things, query the IV => pass-along IV, decrypt with externally provided IV. If for any reason you need to externally set IV (because it is actually a shared secret, is rotated, is unique, etc similar to TLS but not TLS) one too can use that in approved mode. There will be a dynamic indicator raise of non-approved but allowed usage with security claim.
# openssl list -cipher-algorithms -propquery fips=yes | grep GCM
{ 2.16.840.1.101.3.4.1.6, aes-128-gcm, id-aes128-GCM } @ fips
{ 2.16.840.1.101.3.4.1.46, aes-256-gcm, id-aes256-GCM } @ fips
{ 2.16.840.1.101.3.4.1.26, aes-192-gcm, id-aes192-GCM } @ fipsNote I have tried to provide approved way of doing aes-gcm as a code example demo, but it has not yet been merged.
See https://github.com/openssl/openssl/pull/25623 and i sort of gave up on merging that. Should this be provided as a demo?
@Dimitri John Ledkov (Chainguard) It's not the case for the 3.0 series modules: the security policy only applies to AES-GCM in TLS as I understand.
AES-GCM is not restricted in this way in the security policy. You are free to use it anywhere.
There is a restriction on the IV being internally generated (i.e. inside the FIPS boundary). There are exceptions for TLS, SSH, IPsec and one other protocol (these are specified in one of the IGs).
@Paul Dale We didn't read the relevant text in the security policy as covering every case in the IGs. Could easily have been our unfamiliarity with what exactly are the counturs.
From FIPS policy point of view - what do you want? More access to legacy, deprecated, disallowed constructs? or inverse access to future constructs compatible with draft future deprecation? (i.e. APIs that already enforce the upcoming 128 bit transition)
I too dislike the current OpenSSL FIPS policy stance - as I very much prefer forward-looking implementation, even if it causes breakage to existing usage.
Implement KDF SNMP in Default and FIPS providers
Implement KDF IKEv2 in Default and FIPS providers
Implement PQC SSH KDFs in the Default and FIPS providers and/or contribute to openssh-portable to use existing openssl primitives if possible
As far as I can tell openssh-portable has own implementation for all the PQC algorithms, and for it to be FIPS validated one now once again needs to validate openssh as a module (and/or pqc libraries it uses) or to disable PQC or to do something on either openssh and/or openssl side to implement PQC KDF / KEM etc to satisfy the needs of openssh to offer all the PQC goodies. Is anybody looking at doing openssh with PQC powered by openssl fips providers? ( i guess at @Clemens Lang )
@Dimitri John Ledkov (Chainguard) Yes, we're patching this downstream. We're having a hard time convincing the openssh-portable project to take those patches, but maybe we need to try harder (this is all work in progress at the moment).
If that fails, maybe there's a space for an openssh-portable² project where distributions collaborate to make OpenSSH use algorithms from OpenSSL?
@Clemens Lang I mean if openssh-portable is not able to use algorithms exclusively from openssl, it sort of defeats the purpose. I have had pushback on my patches to use OpenSSLv3 APIs; and also moving away from SHA1 to SHA256 for non-security-usage to create session-id. It is annoying that openssh has grown cryptographic implementations - one had to put openssh through FIPS validation; and there was a period of time when one didn't need to do that; and now we are back there.
@Dimitri John Ledkov (Chainguard) My team at Cisco is doing this as well (using OpenSSL Provider or FIPS Provider). We also have an RFC to use pure PQ algs as well that has received major pushback (basically a no way), even though NIAP requires these. The openssh hybrids (IIRC) aren't FIPS'able either (x25519) so we have done our own via OpenSSL there as well.
@Jeff Johnson We've been dealing with this at Akamai using some patches from Red Hat: a colleague of mine Bill Ahern was involved deeply in this.
@Watson Ladd Since we all seem to be dealing with the same issue, maybe it would be good to meet and provide a united front or plan? Thoughts?
@Jeff Johnson We're not super concerned about NIAP's requirement of pure PQ algos for key exchange. If pure isn't available (and it doesn't seem like a pure PQ proposal would ever fly in the IETF SSH working group), I bet they'll settle for the next best thing. Alternatively, they can try working without SSH.
x25519 with ML-KEM 768 should be OK for FIPS iff your ML-KEM is certified, see SP 800-227 section 4.6 (it even mentions X25519MLKEM).
We at Red Hat would certainly be open to sharing the load of writing and maintaining these OpenSSH patches.
Re SHA-1 and MD5 usage, at Chainguard, we see clearly that both SHA-1 and MD5 are used extensively for non-security purposes as a generic digest. And we are adding both of these to the base provider without any `fips=` value, meaning one can get to them with `?fips=yes` and `-fips` and many existing projects already know how to do that (python dotnet) and others can be taught to do that. We are trying to only allow simple digest, and block HMAC and signatures. This is based on FIPS 140-3 IG 2.4.A. See detailed explanation of what we are doing there at https://edu.chainguard.dev/chainguard/fips/non-approved-algorithms/ if others are interested, we can introduce this wider in OpenSSL. Possibly as some sort of "fips-unapproved" provider, which is intentionally outside of the FIPS provider boundary; does not need to go through FIPS validation; can be loaded along side base and fips providers, and intentionally is there to support computing digests only - essentially as a CRC functionality only.
Provide source code compat with openssl forks
Specifically, things like EVP_sha1 can become macros only, in a compat header or some such - or activated by compat define. Such that there is more source code compatibility with boringcrypto/aws-lc/libressl. Similarly we can do that even for complex HMAC -> EVP_MAC transition. Intention is that ABI does not have such functions, but source code can compile against either openssl or boringcrypto unmodified and use "the right" modern EVP APIs.
Please support RFC 7093 Section 2(1) SHA256/160 for subjectKeyIdentifier=hash
This is for SHA1 deprecation and matches what already letsencrypt does by default, as well as most other libraries.
Please support SHA-256 for OCSP
https://datatracker.ietf.org/doc/html/draft-ietf-lamps-rfc5019bis-12 and other drafts are in a queue to be published soon/eventually. It would be nice if openssl was able to support OCSP with SHA-256. Or somehow else support unapproved SHA-1 when operating with fips provider.
For Versa Networks, we are interested in:
PQC Backporting: Request to backport new PQC features from recent and future non-LTS releases to LTS OpenSSL 3.5 as new drafts, RFCs, FIPS, and NIAP requirements evolve. Since OpenSSL 3.5 is the latest LTS including many PQC features, it will likely become the de facto LTS version for PQC. This adoption will likely be driven by CNSA 2.0 requirements for various DoD use cases over the next 1–3 years.
LMS Verification: Request to backport the LMS-verification algorithm to LTS OpenSSL 3.5. While we might not need LMS for many use cases, if ML-DSA is explicitly confirmed soon in the NIAP protection profile for DoD, we must cover worst-case scenarios—specifically where BIOS/firmware vendors support or use only LMS for signing.
DTLS 1.3: Request for DTLS 1.3 support.
QKD Provider: Request for a QKD provider compliant with ETSI-014.
Runtime Control APIs: Request for an equivalent to the ENG_CTRL/ENG_CTRL_CMD_STRING() APIs to control the runtime behavior of the provider (formerly "crypto engine"). These APIs, which allowed for finer control over cryptographic operations, were deprecated during the migration to 3.x.
Performance: Scale and performance improvements for TLS and QUIC.
@Suresh Keisam Backporting major features into an LTS release would be a significant change. I don't think OpenSSL has ever done that before, and I'm not convinced it should start to do that now.
I've a dim recollection that it was done once. I agree, it shouldn't become a common thing and each case should be carefully considered with a bias to not doing it.
@Paul Dale @Clemens Lang I agree that it would be a significant change and shouldn't become a common thing. However, PQC is a once-in-a-lifetime event/change, and it presents a great opportunity for all of us. It would be great if we could make at least the main features and algorithms (ML-DSA and LMS for now) available in a LTS release, as the payoff will be huge for everyone in the long term.
ML-DSA is included in the 3.5 release. LMS is not.
It's Thanksgiving this week, we'll respond longer next week as a result.