The UEFI spec (8.2.2 Using the EFI_VARIABLE_AUTHENTICATION_2
descriptor) includes the following information about constructing
the PKCS#7 message for the authentication descriptor under
point 4(g):
SignedData.signerInfos shall be constructed as:
...
- SignerInfo.authenticatedAttributes shall not be present.
sbvarsign does not currently honour this, and generates a PKCS#7
message containing authenticated attributes. This is a snippet from
OpenSSL's printout of a message I reconstructed from an auth file:
signedAttrs:
object: contentType (1.2.840.113549.1.9.3)
set:
OBJECT:pkcs7-data (1.2.840.113549.1.7.1)
object: signingTime (1.2.840.113549.1.9.5)
set:
UTCTIME:Mar 2 11:20:21 2021 GMT
object: messageDigest (1.2.840.113549.1.9.4)
set:
OCTET STRING:
0000 - 99 58 87 86 82 82 b6 4b-c4 6a e4 e5 6b .X.....K.j..k
000d - 51 39 ac c3 b8 21 24 30-0c 28 e6 e3 aa Q9...!$0.(...
001a - 5c 33 c1 80 3f d1 \3..?.
Tell OpenSSL to stop adding attributes.
This also brings sbvarsign in to line with sign-efi-sig-list.
Signed-off-by: Daniel Axtens <dja@axtens.net>
Signed-off-by: James Bottomley <James.Bottomley@HansenPartnership.com>
If insert_new_keys() fails, say due to a full variable store, we currently
still exit(0). This can make it difficult to know something is wrong.
For example, Debian and Ubuntu implement a secureboot-db systemd service
to update the DB and DBX, which calls:
ExecStart=/usr/bin/sbkeysync --no-default-keystores --keystore /usr/share/secureboot/updates --verbose
But although this seemed to succeed on my system, looking at the logs shows
a different story:
Inserting key update /usr/share/secureboot/updates/dbx/dbxupdate_x64.bin into dbx
Error writing key update: Invalid argument
Error syncing keystore file /usr/share/secureboot/updates/dbx/dbxupdate_x64.bin
Signed-off-by: dann frazier <dann.frazier@canonical.com>
Signed-off-by: James Bottomley <James.Bottomley@HansenPartnership.com>
AKASHI Takahiro (1):
sbsign: allow for adding intermediate certificates
James Bottomley (8):
sbverify: fix verification with intermediate certificates
Tests: Add intermediate certificate tests to the sign-verify cases
Fix some openssl 1.1.0 deprecated functions
sbvarsign: remove unused global variable
sbverify: refer to unused function
Fix errors on 32 bit
Enable -Werror for builds
docs: add man page for sbkeysync
Signed-off-by: James Bottomley <James.Bottomley@HansenPartnership.com>
The function print_certificate_store_certs() is currently commented
out leading to an unused function warning. Make verbose a level and
call this function for levels > 1 (meaning you have to specify -v -v
to see it).
Signed-off-by: James Bottomley <James.Bottomley@HansenPartnership.com>
replace OPENSSL_config with OPENSSL_init_crypto and ASN1_STRING_data
with ASN1_STRING_get0_data
Signed-off-by: James Bottomley <James.Bottomley@HansenPartnership.com>
sbverify is currently failing if an intermediate certificate is added
on signing but the binary is verified with the singing certificate.
It fails with X509_V_ERR_UNABLE_TO_GET_ISSUER_CERT_LOCALLY.
This is happening because the x509_STORE only contains the signing
certificate but the pkcs7 bundle in the binary contains the issuer
certificate as well. Fix this by unconditionally approving any
locally missing certificates on verify.
Signed-off-by: James Bottomley <James.Bottomley@HansenPartnership.com>
SignedData can have multiple certificates, but the current
implementation of sbsign only allows a single one (as a signer).
With this patch, "-addcert" options will be available on command line to
specify a file in which any number of intermediate certificates in PEM
format can be concatenated.
$ sign --key <key> --cert <cert> --addcert <morecerts> [...] image_file
Background:
I'm working on implementing UEFI secure boot on U-Boot and want
to test my code against PE images with intermediate certificates
in certificate chain.
As far as I know, the only tool that supports it in signing is
Microsoft's signtool.exe. So I'd like to have some corresponding
tool on linux.
Signed-off-by: AKASHI Takahiro <takahiro.akashi@linaro.org>
Signed-off-by: James Bottomley <James.Bottomley@HansenPartnership.com>
James Bottomley (1):
README: update git location and add mailing list information
Laszlo Ersek (1):
sbvarsign: fix "EFI_VARIABLE_AUTHENTICATION_2.TimeStamp.Year" assignment
Steve McIntyre (1):
Fix PE/COFF checksum calculation
Signed-off-by: James Bottomley <James.Bottomley@HansenPartnership.com>
Now that a Mailing list is set up, update the README to point to it
and mention the new maintained git location for this fork.
Signed-off-by: James Bottomley <James.Bottomley@HansenPartnership.com>
According to UEFI-2.8, section 8.3 "Time Services" / GetTime(), the
"EFI_TIME.Year" field must be in the range [1900, 9999] (both bounds
inclusive). It is not stated or even implied that "EFI_TIME.Year" would
not be an absolute year number.
According to POSIX, the "tm_year" field of "struct tm" is defined as
"Years since 1900". In other words, "tm_year" is relative to 1900.
In set_timestamp(), time() and gmtime() are suitable for populating
"EFI_VARIABLE_AUTHENTICATION_2.TimeStamp", as the UEFI spec specifically
requires a stamp expressed in the GMT (UTC) zone. But we still need to
offset "tm->tm_year" by 1900 for filling in "timestamp->Year". So let's do
that now.
While this issue does not seem to affect upstream edk2, SetVariable()
calls with payloads containing an invalid
"EFI_VARIABLE_AUTHENTICATION_2.TimeStamp.Year" value do seem to be
rejected at least on some Dell Inspiron machines (using a UEFI
implementation from AMI).
Reported-by: Eugene Khoruzhenko <ekhoruzhenko@absolute.com>
Reported-by: Paulo Henrique Lacerda de Amorim <phlamorim@riseup.net>
Ref: https://edk2.groups.io/g/devel/message/49402
Fixes: 953b00481f
Signed-off-by: Laszlo Ersek <lersek@redhat.com>
Signed-off-by: James Bottomley <James.Bottomley@HansenPartnership.com>
Only count the cert_table header once when performing the calculation
and counting buffer sizes.
The problem entered because of a mismerge of multiple signature
support and "be1f3d8 Update the PE checksum field using the
somewhat-underdocumented algorithm, so that we match the Microsoft
implementation in our signature generation.". Originally
image->cert_table held the full certificate table including the
Microsoft _WINH_CERTIFICATE header and image->sigbuf pointed to the
pkcs11 signature inside, so the two had to be checksummed separately.
After multiple signature support, image->sigbuf points to the full
certificate table because we now need the headers to decide where one
signature ends and the next begins, so the correct checksum only needs
to sum over the entire image->sigbuf.
Signed-off-by: Steve McIntyre <93sam@debian.org>
Signed-off-by: James Bottomley <James.Bottomley@HansenPartnership.com>
Fix fedora build
Fix variable signing for current tianocore
Fix image processing not to invalidate existing signatures
Signed-off-by: James Bottomley <James.Bottomley@HansenPartnership.com>
The old code forced region size to be aligned to the PECOFF file
alignment parameter, which is correct according to the spec. However,
the major UEFI platforms do not align up when checking the signature,
so if the PECOFF binary being signed already contains a signature,
realigning the sections will make the existing signature invalid. Fix
this by relaxing the rule about aligning up (also eliminates
complaints about some pecoff sections being misaligned).
Signed-off-by: James Bottomley <James.Bottomley@HansenPartnership.com>
The EFI standard is ambiguous about which one to use for variable
updates (it is definite about using PKCS7 for signed binaries). Until
recently, the reference platform, tianocore, accepted both. However
after patch
commit c035e37335ae43229d7e68de74a65f2c01ebc0af
Author: Zhang Lubo <lubo.zhang@intel.com>
Date: Thu Jan 5 14:58:05 2017 +0800
SecurityPkg: enhance secure boot Config Dxe & Time Based AuthVariable.
The acceptance of PKCS7 got broken. This breakage seems to be
propagating to the UEFI ecosystem, so update the variable signing
tools to emit the SignedData type (which all previous EFI
implementations accepted).
Signed-off-by: James Bottomley <James.Bottomley@HansenPartnership.com>
Add the ability to specify an engine to read the keyfile. For safety,
we don't do the full dynamic engine support, but only use engines
configured for use by the platform.
Signed-off-by: James Bottomley <James.Bottomley@HansenPartnership.com>
This version builds correctly on openssl 1.1 and also includes
functional autotests for every architecture.
Signed-off-by: James Bottomley <James.Bottomley@HansenPartnership.com>
Older versions of openssl 1.0.0 don't have X509_STORE_CTX_get0_store
so define that as well.
Signed-off-by: James Bottomley <James.Bottomley@HansenPartnership.com>
The current test infrastructure is tied to x86/amd64. This means the
tests always fail on a non-x86 architecture (like aarch64). Fix this
by generating the efi binary directly from C code and removing the
architectural restrictions in the Makefile.am. One of the
consequences of this is that we no longer test ia32 on x86_64, but the
difficulty of detecting which architectures can support 32 bit
variants and generating them correctly from EFI c code is too great.
We also need to exclude tests involving objdump from aarch64 since its
bfd still doesn't have an efi_app_aarch64 target.
Signed-off-by: James Bottomley <James.Bottomley@HansenPartnership.com>
pecoff for i386 can be too short, so it gets padded for an accurate
signature. Make sure the size comparison takes this into account to
avoid spurious failures.
Signed-off-by: James Bottomley <James.Bottomley@HansenPartnership.com>
newer versions of gcc have contained an efi target for a while so use
it instead of hacking a linker script.
Signed-off-by: James Bottomley <James.Bottomley@HansenPartnership.com>
The original tests to warn about overwriting signatures have never
worked after the multiple signature code was added (because we add a
new signature instead of overwriting the old one) update the tests to
check instead for the signature addition.
Signed-off-by: James Bottomley <James.Bottomley@HansenPartnership.com>
Most structure definitions in OpenSSL are now opaque and we must call
the appropriate accessor functions to get information from them.
Not all the accessors are available in older versions, so define the
missing accessors as macros.
The X509_retrieve_match() function is no longer usable, as we cannot
initialise an X509_OBJECT ourselves. Instead, iterate over the
certificate store and use X509_OBJECT_get_type and X509_cmp to
compare certificates.
autotest is very finicky. The environment can't be set up in
SH_LOG_COMPILER, but have to be done in AM_TESTS_ENVIRONMENT instead,
so fix this.
Signed-off-by: James Bottomley <James.Bottomley@HansenPartnership.com>
In the current framework for using engine based keys, the engine has
to be loaded and initialised as part of the default engines. The only
way this can happen for the TPM engine is if it is named in a config
secion, so all the tools must read and act on the config file to be
able to use TPM based keys.
Signed-off-by: James Bottomley <James.Bottomley@HansenPartnership.com>
It causes the ARM build to crash (because of directives) and it's
unnecessary in this file.
Signed-off-by: James Bottomley <James.Bottomley@HansenPartnership.com>
arm has a variety of uname -m forms, all beginning with arm, so use
this to determine the EFI architecture
Signed-off-by: James Bottomley <James.Bottomley@HansenPartnership.com>
OpenSSL 1.0.2e now actively checks for both data and contents being present
for a certificate. Clear out contents so that we have only data, and run a
chance of actually verifying the signature.
Bug-Ubuntu: https://launchpad.net/bugs/1526959
Signed-off-by: James Bottomley <James.Bottomley@HansenPartnership.com>
Buffers of odd length can be passed to the checksum, for example signatures.
csum_bytes uses a uint16_t so change the function to prevent overflowing the
buffer, while taking the extra byte into account if the length is odd.
Signed-off-by: Linn Crosetto <linn@hpe.com>
Signed-off-by: James Bottomley <James.Bottomley@HansenPartnership.com>
Newer versions of openssl return a different error with alternate
certificate chains; update for compatibility.
Signed-off-by: Marc Deslauriers <marc.deslauriers@canonical.com>
Bug-Ubuntu: https://bugs.launchpad.net/bugs/1474541
Signed-off-by: James Bottomley <James.Bottomley@HansenPartnership.com>
Note that for the ARM case, we are using IMAGE_FILE_MACHINE_THUMB (0x1c2)
rather than IMAGE_FILE_MACHINE_ARM (0x1c0), as the latter refers to
an older calling convention that is incompatible with Tianocore UEFI.
Signed-off-by: Ard Biesheuvel <ard.biesheuvel@linaro.org>
Signed-off-by: James Bottomley <James.Bottomley@HansenPartnership.com>
The loop that iterates over the PE/COFF sections correctly skips zero
sized sections, but still increments the loop index 'i'. This results in
subsequent iterations poking into unallocated memory.
Signed-off-by: Ard Biesheuvel <ard.biesheuvel@linaro.org>
Signed-off-by: James Bottomley <James.Bottomley@HansenPartnership.com>
Original patch from: Steve Langasek <steve.langasek@canonical.com>
The ubuntu version of the signature expiry patch ignores serveral more errors,
so add them.
Bug-Ubuntu: https://bugs.launchpad.net/bugs/1234649.
Signed-off-by: James Bottomley <James.Bottomley@HansenPartnership.com>
algorithm, so that we match the Microsoft implementation in our
signature generation.
[jejb: add endian to autogen.sh and fix for multi-sign]
Signed-off-by: Jeremy Kerr <jeremy.kerr@canonical.com>
Signed-off-by: James Bottomley <James.Bottomley@HansenPartnership.com>
sbsign will sign an already signed binary (adding a signature at the end)
sbverify has a new mode --list, for listing all the signatures and sbattach
takes a --signum argument for --remove or --detach.
Signed-off-by: James Bottomley <JBottomley@Parallels.com>
This prevents a FIPS failure message if no FIPS module is loaded.
Plus add -v as short form for --verbose in sbverify
Signed-off-by: James Bottomley <JBottomley@Parallels.com>
The new Tianocore multi-sign code fails now for images signed with
sbsigntools. The reason is that we don't actually align the signature table,
we just slap it straight after the binary data. Unfortunately, the new
multi-signature code checks that our alignment offsets are correct and fails
the signature for this reason. Fix by adding junk to the end of the image to
align the signature section.
Signed-off-by: James Bottomley <JBottomley@Parallels.com>
Not zeroing the image after talloc occasionally leads to a segfault because
the programme thinks it has a signature when in reality it just has a junk
pointer and segfaults.
Signed-off-by: James Bottomley <JBottomley@Parallels.com>
