Add ML-DSA signing and verification for CMS/PKCS#7 SignedData, following
RFC 9882. ML-DSA is used in CMS "pure" mode: the signature is computed
over the complete message (the DER SET OF signed attributes, or the
eContent when none are present) with an empty context string and absent
signatureAlgorithm parameters, rather than over a pre-computed DigestInfo
as with RSA/ECDSA.
wolfcrypt/src/pkcs7.c:
- New ML-DSA helpers: wc_PKCS7_MlDsaLevelFromOID, wc_PKCS7_BuildPureSigMessage,
wc_PKCS7_MlDsaSign and wc_PKCS7_MlDsaVerify, wired into the per-algorithm
switch sites (GetSignSize, SignedDataGetEncAlgoId, SetPublicKeyOID,
CheckPublicKeyDer) and the sign/verify dispatchers. Only the final FIPS 204
ML-DSA OIDs are accepted; pre-standard draft Dilithium OIDs are not.
- GetSignSize derives the ML-DSA signature length from the parameter set.
- InitWithCert copies the signer public key into the RSA-sized publicKey buffer
only for RSA/ECC certs (the raw-sign callback consumers); large PQC keys such
as ML-DSA would overflow it and are never read back, so publicKeySz stays 0.
- wc_MlDsaKey is always heap allocated (it embeds multi-KB key buffers); the
accompanying DecodedCert uses the WC_DECLARE_VAR/WC_ALLOC_VAR_EX macros for
stack-vs-heap handling under WOLFSSL_SMALL_STACK.
- wc_PKCS7_SignedDataBuildSignature skips building the DigestInfo for ML-DSA,
which signs the full message in pure mode and never consumes it.
- wc_PKCS7_MlDsaSign wraps the ML-DSA private-key decode in
PRIVATE_KEY_UNLOCK/PRIVATE_KEY_LOCK. Unlike RSA/ECC, the FIPS module gates
wc_MlDsaKey_PrivateKeyDecode behind the private-key read lock, so signing
would otherwise fail with FIPS_PRIVATE_KEY_LOCKED_E under --enable-fips. The
macros are no-ops in non-FIPS builds.
wolfssl/wolfcrypt/pkcs7.h:
- Document that the fixed-size signer public key buffer (publicKey/publicKeySz)
holds only RSA/ECC keys; it stays RSA-sized.
wolfcrypt/src/hash.c:
- Map the SHAKE128/SHAKE256 OIDs to their hash types in wc_OidGetHash().
certs/mldsa:
- Add expanded-only PKCS#8 DER private keys (mldsa44/65/87-key.der) matching
the self-signed ML-DSA certificates, with README and include.am updates.
The expanded-only shape (no seed) decodes via wc_MlDsaKey_ImportPrivRaw
without keygen-from-seed or the ASN template, so pkcs7signed_mldsa_test also
passes in WOLFSSL_MLDSA_NO_MAKE_KEY and non-WOLFSSL_ASN_TEMPLATE builds.
certs/renewcerts.sh:
- Generate the mldsa<N>-key.der files from the matching mldsa<N>-key.pem in the
expanded-only shape (openssl pkey -provparam ml-dsa.output_formats=priv), so
a regeneration keeps the DER key in step with the cert. The OpenSSL detection
probe now requires both ML-DSA keygen and that conversion across all three
levels, so the block runs fully (matched cert+key) or is skipped entirely
rather than aborting mid-way.
wolfcrypt/test/test.c:
- Add pkcs7signed_mldsa_test(): round-trip encode/verify of SignedData across
ML-DSA-44/65/87, with and without signed attributes, including a check that
the digest algorithm parameters are encoded as expected. The message-digest
OID is selected from the enabled hash set (SHA-512, else SHA-256, else SHA-1)
so the test builds when SHA-512 is disabled. A negative case confirms ML-DSA
rejects a caller-supplied pre-computed content hash with BAD_FUNC_ARG.
For each product (wolfSSH, wolfCLU, wolfTPM, wolfMQTT, wolfPKCS11, wolfProvider) this builds wolfSSL and then builds the product against it, at both the product's latest release tag and its master branch. It only checks that they compile, it does not run any tests. If a wolfSSL change breaks a product's latest release on purpose, you say so in a commit message with breaks-<product>=<tag>. A break on a product's master is not allowed and has to be fixed.
breaks-wolfssh=v1.5.0-stable
Note: wolfSSH v1.5.0-stable does not currently compile against wolfSSL. This commit did not break it, it adds the cross-library check that discovered the break. The token above declares it so the new check tracks it as a known break instead of failing red, until wolfSSH ships a fixed release.
Server-side PKCS#7 encode improvements that let downstream EST/SCEP enrollment
code (wolfCert) drive the existing encoder through the public API rather than
hand-rolling DER. Everything is gated under the existing HAVE_PKCS7 — no new
build options and no new public functions; the convenience wrappers live
caller-side.
Allow degenerate (certs-only) SignedData encode
Relax the hashOID != 0 requirement in PKCS7_EncodeSigned() when
sidType == DEGENERATE_SID, so a caller can produce a certs-only bundle (no
signer, attributes, or eContent — the form used by EST /cacerts and SCEP
GetCACert) by selecting DEGENERATE_SID via wc_PKCS7_SetSignerIdentifierType()
and calling wc_PKCS7_EncodeSignedData(). The output round-trips through
wc_PKCS7_VerifySignedData().
Size the signed-attribute array to the actual count
The SignerInfo attribute working array is now sized to the real attribute
count instead of a fixed [7] array. An inline buffer (sized
MAX_SIGNED_ATTRIBS_SZ, the historical footprint) covers the common
allocation-free case; a heap buffer is used only when the count exceeds it.
The default-attribute count comes from a single helper
(wc_PKCS7_GetDefaultSignedAttribCount) so the sizing matches the emission
logic exactly, and the canned-attribute write is bound-checked against the
array capacity. This also fixes a latent overflow where the backing array was
hardcoded [7] while the bound check used MAX_SIGNED_ATTRIBS_SZ. The macro is
retained for source compatibility but no longer caps the count.
Document the decoded-attribute value shape
Documented the stable shape of PKCS7DecodedAttrib.value (the contents of the
SET OF AttributeValue, outer SET tag stripped) so callers can rely on it. No
behavior change.
Fix multi-certificate decode in non-streaming builds
Bound the additional-certificate loop in wc_PKCS7_VerifySignedData against the
absolute end of the certificate set (idx + length) rather than the relative
length. In NO_PKCS7_STREAM builds the old bound dropped trailing certificates
(all but the first when a large eContent preceded the set), failing
verification when the signer cert was among those dropped. Streaming builds
were unaffected.
Tests
Added coverage in pkcs7signed_test: degenerate certs-only encode via the
public API, nine-attribute encode (beyond the inline capacity), decoded
attribute value shape for PrintableString and OCTET STRING, and a
multi-certificate decode regression with large content that triggers the
bound bug under NO_PKCS7_STREAM. Added a signed-attribute selection
round-trip covering a messageDigest-only subset and the no-attributes case
via wc_PKCS7_SetDefaultSignedAttribs/wc_PKCS7_NoDefaultSignedAttribs, a
WOLFSSL_NO_MALLOC over-capacity case that must return BUFFER_E instead of
overrunning the inline buffer, and a malformed certificate-set length that
exercises the certSetEnd clamp in the verifier. Config-sensitive cases are
guarded.
The socat suite is sleep-bound and slow run serially. Drive it through
parallel-make-check.py as ~6 shards per CPU, 2 running per CPU at once: each
shard runs a round-robin slice of the tests in its own bwrap network
namespace (so parallel shards don't collide on ports) and its own build-dir
copy. The work is almost all waiting, so the oversubscription just overlaps
the waits.
Install bubblewrap so the netns isolation actually happens (without it the
runner silently shares one namespace and the shards collide). Each fresh
netns is IPv4-loopback only, so re-create IPv6 loopback (CAP_NET_ADMIN) for
the ::1 / dual-stack tests, and add non-loopback placeholders (fc00::1,
192.0.2.1) so glibc's AI_ADDRCONFIG still returns both families - without
them socat's getaddrinfo fails on numeric non-loopback addresses, e.g. the
multicast tests. Relax the AppArmor unprivileged-userns restriction so the
bwrap netns + CAP_NET_ADMIN work on ubuntu-24.04.
Support AES-XTS AVX512/VAES
Support AES-GCM AVX512/VAES
Support AES-ECB/CBC/CTR AVX512/VAES/AVX1/AES-NI.
Remove code from aes_asm.S/aes_asm.asm
Add CPU defines for AVX512 and VAES
Updated ASM files with new defines for AVX512.
Added support for printing out the new CPU Id flags in benchmark.
Added new files to Windows projects.
aes.c: Supports ECB/CBC/CTR in assembly. Supports calling AVX512/VAES assembly.
Make every --enable-tinytls13 spelling build and pass locally, and grow the
CI matrix to cover them. These are fixes found while testing the configs the
CI workflow had not actually exercised.
- internal.h, internal.c, ssl_load.c: include ML-DSA and Falcon in the
pkCurveOID member and producer guards so the PSK plus ML-DSA build compiles.
- tls13.c: gate the DoTls13CertificateVerify definition on NO_CERTS to match
its call site.
- settings.h: let the AES-256 adder survive the floor, default the
user_settings path to the SHA-256 floor, make WOLFSSL_NO_MALLOC opt-in so
the test suite still runs, and keep ML-DSA ASN.1 for the cert profile.
- configure.ac: drive ENABLED_ASM and emit WOLFSSL_NO_ASM for the small C
floor, restrict SP math to P-256, strip ML-DSA ASN.1 only on the PSK floor,
and print a notice for the reduced security cert verify.
- examples: guard the cert loading paths for NO_CERTS and treat NO_CERTS as
PSK mode in echoserver and echoclient.
- Add examples/configs/tinytls13_smoke.c, an in memory TLS 1.3 handshake test
that drives PSK, ECDSA, ML-DSA-65 and RSA-PSS chain verify, plus forced
cipher suites, for builds with no example or unit test harness.
- certs: add ECDSA leaves signed by the ML-DSA-65 and RSA-PSS CAs so the cert
profiles drive a real PQC and PSS chain verify in CI.
- .github/workflows/tinytls13.yml: cover every profile and adder, run the
smoke handshake on the build verified configs, and least privilege the
workflow token.
The draft guard skips the job on draft PRs, but the pull_request
trigger used the default types (no ready_for_review), so marking a
draft ready did not re-run the job and it stayed skipped. Add the
standard types, matching the other workflows, so it re-runs when the
PR becomes ready.
With the cache save restricted to master, a cold-cache PR or release
run can no longer restore in the test job what the build job just built
(the per-PR cache scope is gone), so mbedtls/nss were compiled twice.
Upload the build as an artifact on a cache miss and download it in the
test job instead of recompiling, matching the handoff hostap-vm already
uses. master still restores from the shared cache, so it never uses the
artifact.
The v6.4.3_rel version was repeated in the cache path, cache key,
download URL and extract command. Define it once as a workflow-level
env var and reference it everywhere.
GitHub Actions caches are branch-scoped: an entry written by a
pull_request run lives under refs/pull/<N>/merge and is invisible to
other PRs. The haproxy, mbedtls, nss, ntp, threadx and hostap-vm
workflows used combined actions/cache with fixed keys, so every PR
re-saved its own copy of the same dependency, yielding one duplicate
cache entry per PR.
Split each into actions/cache/restore (always) plus actions/cache/save
gated to refs/heads/master, and add a daily schedule so a master run
reseeds the single shared entry that all PRs restore. mbedtls/nss save
in their build job only; the test jobs restore-only.
Disable the setup-msys2 package cache: the action only toggles caching
on/off and cannot save on master while restoring on PRs.
A single stalled apt mirror connection hung the ubuntu-24.04-full /
ubuntu-22.04-full download for ~20 min (they normally finish in a few),
tripping the 20-min job timeout and leaving those tags stale. The per-package
retry() only re-runs on a non-zero exit, so a hang never tripped it.
- apt drops a stalled connection after 30s and retries it
(Acquire::http/https::Timeout, Acquire::Retries).
- each apt-get is wrapped in `timeout` so a wedged process is hard-killed and
retry() re-runs it from scratch.
- raise the build job timeout 20 -> 60 min as a final backstop.
arduino.yml's per-core actions/cache layer stored the installed cores and
toolchains (~/.arduino15) - several GB, dominated by the esp32 and mbed
cores - in the 10 GB Actions cache. For esp32 it was also ineffective: the
disk-cleanup step deletes the esp32 toolchain before actions/cache saves it,
so esp32 re-downloaded every run anyway.
- New arduino-cores-image workflow resolves each of the 9 distinct cores and
publishes a tar of ~/.arduino15 + ~/Arduino/libraries to
ghcr.io/<owner>/wolfssl-ci-arduino:<core>. It runs monthly: esp32, the
fastest-moving core, releases ~monthly and the rest far less often.
- New install-arduino-core composite action restores that bundle offline and
verifies the core is present, falling back to `arduino-cli core install`
when the bundle is unavailable - so nothing breaks until the image is first
published and made public.
- arduino.yml calls the action in place of the inline core install and the
actions/cache step.
This takes the flaky espressif / esp8266.com / pjrc.com downloads off the PR
critical path and frees the Actions cache of the largest binaries it held.
setup-alire@v5 caches the gnat_native+gprbuild toolchain via actions/cache
(key alr[1][2.1.0][...]), holding ~1.26 GiB - 3x the 428 MiB toolchain, one
copy per ref - against the repo's 10 GiB cache cap. On a miss the toolchain
is only a ~17s pull from github.com (alire-project releases), so the cache
saved ~20-30s on a ~6.5min Ada job (dominated by gnatprove). Not worth the
space; install it fresh each run.
Addresses PR review feedback. The kernel-tracking linuxkm bundle treated a
failed --download-only as a warning and still published, so a transient
mirror error could ship a partial bundle. Because the daily job skips
rebuilds while the kernel label matches, such a partial bundle would
persist until the kernel next changes (~monthly), forcing consumers to fall
back to apt the whole time.
The linuxkm set is small and entirely required, so resolve it as one
closure and let a failure fail the job; we push only on success, so the
last good bundle stays in place. The static -full/-minimal bundles keep
their per-package skip-and-warn - they serve many independent consumer
subsets and rebuild weekly, so maximizing coverage is the right trade-off
there.
Extends the ghcr offline-install path to every install-apt-deps consumer
that was still on plain apt, and publishes the bundles they need.
New bundles built by ci-deps-image:
- ubuntu-24.04-embedded: the membrowse ARM cross-toolchain (~0.5 GB), kept
out of -full so it does not bloat the interop workflows' pull.
- ubuntu-24.04-linuxkm: linux-headers-$(uname -r) + the kernel-module build
toolchain. linux-headers tracks the runner's running kernel, so a daily
job rebuilds it only when uname -r changed (recorded as an image label);
a mismatch during a runner-image rollout just falls back to apt.
Consumers now passing ghcr-debs-tag:
- sssd -> ubuntu-24.04-full (its deps added to that list)
- hostap-vm -> ubuntu-22.04-full (its deps added to that list)
- membrowse targets -> ubuntu-24.04-embedded; the two linuxkm targets ->
ubuntu-24.04-linuxkm (new per-target matrix.ghcr_tag)
- linuxkm.yml -> ubuntu-24.04-linuxkm (pinned to ubuntu-24.04 so the
bundle's headers match the runner kernel)
Each consumer still falls back to apt when its bundle is unavailable, so
nothing breaks until ci-deps-image first publishes the new tags.
The ci-cache-offload work added a ghcr .deb bundle path to
install-apt-deps, making the actions/cache apt-archive layer redundant.
Remove it so no apt-deps-* cache entries are produced. Apt packages now
install either offline from the ghcr bundle (when ghcr-debs-tag is set)
or via plain apt-get with the existing retry/backoff.
- Strip the Compute/Restore/Pre-seed/Collect/Save cache steps and the
cache-hit fast path; drop the now-unused 'cache' input.
- Update callers that passed 'cache': membrowse-onboard, membrowse-report
(and the apt_cache matrix key in membrowse-targets.json), and sssd.
The ghcr offline path and the ccache actions/cache usage are untouched.
Rebasing onto master (which migrated JS actions to Node.js 24 runtimes)
left a few action refs that this branch added in new steps still on the
old major versions. Bring them in line with master:
- ccache-setup read-only restore: actions/cache/restore@v4 -> @v5
- smoke-test / os-check ccache save: actions/cache/save@v4 -> @v5
- ci-deps-image checkout: actions/checkout@v4 -> @v5
Zackery
Tobias Frauenschläger
David Garske
Sean Parkinson
Zackery Backman
Emma Stensland
JacobBarthelmeh
Yosuke Shimizu
Paul Adelsbach
aidan garske
Juliusz Sosinowicz
Eric Blankenhorn
Aidan Garske