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>
In line with the verification process in firmware, update our verify
callback to explicitly trust all certificates that we load to our cert
store.
Signed-off-by: Jeremy Kerr <jeremy.kerr@canonical.com>
Proposed changes to the kernel will establish /sys/firmware/efi/efivars
as the canonical mountpoint for the efivars filesystem.
Signed-off-by: Jeremy Kerr <jeremy.kerr@canonical.com>
Rather than overrunning the heap, explicitly allocate the pad area for
cases where we've aligned-up the section table sizes.
Signed-off-by: Jeremy Kerr <jeremy.kerr@canonical.com>
Since we write the certificate table starting at data_size (not size),
use this value when generating the cert table header.
Signed-off-by: Jeremy Kerr <jeremy.kerr@canonical.com>
The PE/COFF spec allows variable-sized data directories, which reduce
the size of the optional header. While GNU ld always produces
maximum-sized headers, the kernel's EFI_STUB code generates a smaller
header size, which causes the image parsing code to abort.
This change allows variable-sized optional headers, but checks for at
least enough of an optional header to contain a CERT_TABLE data
directory entry.
We also rename struct image's aouthdr to opthdr, as it contains more
than just the a.out fields.
Signed-off-by: Jeremy Kerr <jeremy.kerr@canonical.com>
We were updating siglist before incrementing i, and so aborting the
siglist iteration earlier than necessary.
Signed-off-by: Jeremy Kerr <jeremy.kerr@canonical.com>
Rather than having three sets of (firmware, filesystem) key databases,
refactor into two sets of (kdk, db, dbx) databases. This allows us to
add the PK later.
Signed-off-by: Jeremy Kerr <jeremy.kerr@canonical.com>
