WOLFPSA_AES_FAST replaced with WOLFSSL_AES_TOUCH_LINES per the wolfPSA
5.9.4 constant-time policy; DICE.md and the MCXN verifier comment now say
ESP256 and state the RFC 9783 legacy-ID compatibility note.
WC_ALLOW_ECC_ZERO_HASH and WOLFPSA_AES_FAST per the new psa_config.h
checks (bitsliced AES blows the GCM stack budget), and exclude the new
psa_store_zephyr.c from the custom-store build.
wolfSSL v5.9.4-stable, wolfPSA v5.9.4, wolfCOSE v2.0.0; wolfHSM
left at v1.5.0-8-g86dd6df. wolfPSA v5.9.4 changes the store Close
return to int; wolfCOSE v2.0.0 splits the source into 15 files and
drops the deprecated ES256 alg ID, so DICE now signs with ESP256.
mb2_parse_info_request_tag() (DEBUG_MB2) still divided by
sizeof(uint16_t) after the u32 revert, doubling the entry count and
reading (size-8) bytes past the tag. Same width, one more call site.
The F-14470 fix changed the info-request entry width to 16-bit on a
misreading of the spec: Multiboot2 v2.0 defines mbi_tag_types as
u32[n] ("an array of u32's"), and the original code already divided
by sizeof(uint32_t). A 32-bit request {4} was misread as {4, 0}, so
the follow-up zero-skip was only papering over the misparse; real
FSP images with a mandatory request tag stopped booting (fsp_qemu
CI). Restore the u32 width, drop the zero-skip, and rebuild the
test fixtures with 32-bit entries.
F-14470 corrected the request-list entry width to the spec's 16-bit,
which doubled the entry count. Intel FSP pads the list so the last
entry is zero (the end-tag type), and the old 32-bit width never
reached it. A zero entry is padding, not a mandatory request: skip
it, so FSP images keep booting. New test covers a zero-padded list.
F-14487/F-14488 scrubbed the TRNG staging buffers with wc_ForceZero,
which links the hal objects to wolfCrypt's memory.c. The OTP keystore
primer links the hal without wolfCrypt and failed to build
(undefined reference to wc_ForceZero) in the stm32h5_tz_dualbank_otp
and trustzone-emulator configs. Scrub with a local volatile loop, the
same pattern nvm_cache_scrub uses for the same reason.
The secure-mode TRNG seed helper (shared by the STM32L5/U5/H5
TrustZone ports) staged each TRNG_DR word in a stack local and
returned without clearing it. Force-zero the local before the
return, mirroring the existing scrub in hal/cm4.c.
The secure-mode TRNG seed helper copied each 24-byte CryptoCell EHR
batch into a stack array and returned without clearing it, leaving
raw TRNG output in the retired secure frame. Force-zero the batch at
the end of each iteration, mirroring the existing scrub in hal/cm4.c.
loadFile() in the TPM policy-signing tool read the private key
through a fully buffered stdio stream, leaving a libc-owned copy of
the key material that outlives the call. Open the file unbuffered
(setvbuf _IONBF), the same pattern sign.c already uses for its key
reads.
nsc_mech_prepare() read the non-secure legacy ulPasswordLen pointer
twice: once to size the secure password copy, once again (via nsc_in)
to fill the secure length word the library dereferences. A concurrent
non-secure write between the two reads could desync the length from
the password buffer size. Take a single snapshot and write it into
the secure length word directly. Add the missing test for the legacy
branch: the stub now consumes the length the way wolfPKCS11 does
(through the bounced pointer) and checks scrubbing.
mb2_build_boot_info_header() rejected a valid OS image header when
the optional information-request tag was missing, and rejected any
request type it did not implement without checking the tag's
optional flag (bit 0) the spec defines for skipping. A well-formed
header without the tag now requests nothing, optional unsupported
requests are skipped, and only a mandatory unsupported one rejects
the image. The finder returns NULL for a missing tag and for a
malformed tag list alike, so the absent case walks the list to tell
the two apart: any corrupt tag rejects the header. Tests updated to
the spec semantics (optionality driven by the tag flags) plus
malformed-list cases.
qspi_quad_enable() memset the status register it had just read before
setting the QE bit, so the write-back zeroed every other writable bit
(SRL/CMP in SR2, BP/SRWD on the SR1 path) on the first probe of a part
with QE clear. Drop the memset so the write is a read-modify-write.
Test pins the preserved-bits write and the no-write sunny path.
wolfBoot_record_verify_failure() read the image version from
img->hdr, which for an external-flash image is a device offset, not
a CPU address: the persistent record stored a garbage version, and
the read can fault where the offset is unmapped. Read the version
through wolfBoot_get_image_version(part), which fetches external
headers via ext_flash_check_read. No unit test compiles
update_flash.c with WOLFBOOT_PERSIST_FAILURE_STATUS, so add a
compile check for that ext-flash configuration to the gate.
wolfBoot_get_partition_state() only rejected PART_NONE, so any other
part byte (SWAP, SELF, unknown) passed through the NSC veneer to a
NULL dereference of get_partition_magic() in secure code. Reject
every part without a trailer at the shared entry; the NSC veneer
delegates here, so a non-secure caller can no longer crash the
secure world. The mock get_trailer_at() now matches the real one
(NULL for parts without a trailer), which is what lets the test
prove the deref.
The parser read the SEQUENCE length without checking it, skipped the
AlgorithmIdentifier without comparing it, and ignored trailing bytes,
so only the digest length was pinned. Any free byte in the message is
forgeable with a low-exponent key (e=3) via Bleichenbacher 2006, and
keygen -i imports exponents unchecked. Require the SEQUENCE to span
the whole payload, the AlgorithmIdentifier to match the configured
hash OID byte for byte, and the OCTET STRING to end at the payload
end. Test pins the exact wc_EncodeSignature() encoding and rejects
trailing bytes and the empty-algoid forge shape.
nvm_select_fresh_sector() gets the PART_UPDATE flag base from
PART_UPDATE_ENDFLAGS (UPDATE size) but computed addrErase with the
BOOT size, a copy-paste from the BOOT branch. The two are equal in
every compilable write-once build today (asymmetric sizes are
#error'd without SELF_UPDATE_MONOLITHIC, which is incompatible with
write-once), so align the erase sector with the base.
David Garske
Mattia Moffa
Daniele Lacamera