The RAM address and data scrambling API was called twice, once before filling
RAM with random values, and once after. Since moving to a significantly
better PRNG (xorwow) this is now deemed unnecessary. See issue #225.
This changes both FPGA and firmware hashes.
Modify the loop to zeroise the FW-RAM instead of the
RAM. RAM is filled with random data at the start of main().
Changes firmware and bitstream digests.
Signed-off-by: Joachim Strömbergson <joachim@assured.se>
xorwow provides significantly better random data, compared to previously
used function. Making it harder to predict what data RAM is filled with.
It adds a startup time of approx 80 ms, but can be compensated with
optimising other parts of the startup routine.
This changes both firmware and fpga hashes.
Signed-off-by: Joachim Strömbergson <joachim@assured.se>
The memset() responsible for the zeroisation of the secure_ctx under
the compute_cdi() function in FW's main.c, was optimised away by the
compiler. Instead of using memset(), secure_wipe() is introduced
which uses a volatile keyword to prevent the compiler to try to
optimise it. Secure_wipe() is now used on all locations handling
removal of sensitive data.
Use _RAM_ADDR_RAND instead of _RAM_ASLR since this is not OS-level
ASLR we're talking about. It's address randomization as seen from
outside of the CPU, not from the process running inside it. Ordinary
ASLR is visible from the CPU.
Since UDS is not byte-readable we copy it by word to local_uds. Now
UDS lives for a short while in local_uds on the stack in FW_RAM and in
the internal buffer of the blake2s context (also in FW_RAM) but is
very soon overwritten.
Add clang-tidy and splint static analytics check. For now, we use only
the cert-* warnings on clang-tidy and run splint with a lot of flags
to allow more things.
Changes to silence these analytics:
- Stop returning stuff from our debug print functions. We don't check
them anyway and we don't have any way of detecting transmission
failure.
- Declare more things static that isn't used outside of a file.
- Change types to be more consistent, typically to size_t or
something or to uint32_t.
We don't use any .data or .bss segment at all to keep all the firmware
variables in the stack in protected fw_ram.
Signed-off-by: Daniel Lublin <daniel@lublin.se>
This means firmware's stack shouldn't be accessible to programs
running in app_mode.
It also means we don't need to take special care of secure_ctx which
can now be an ordinary stack variable.
Nonetheless we zero out secure_ctx after final use and inline some
assembler to zero out the entire fw_ram after use, just before
switching to app_mode.
Signed-off-by: Daniel Lublin <daniel@lublin.se>
For every state, define a constant bitstring with allowed commands and
check incoming command agains that.
Signed-off-by: Daniel Lublin <daniel@lublin.se>
- We always assert on allowed commands in a state.
- We don't allow FW_CMD_LOAD_APP to be used twice.
- Enter fail state on read buffer overrun, header endpoint not for us,
header parse error, and unknown firmware command.
Signed-off-by: Daniel Lublin <daniel@lublin.se>
UDS is now byte readable (but not writable).
Use UDS and USS directly in a blake2s_update() instead of
concatenating them into fw_ram. UDS will still live for a short while
in fw_ram in the blake2s context buffer but will soon be overwritten.
Signed-off-by: Daniel Lublin <daniel@lublin.se>
We define macros for them that expand to nothing or to a constant to
avoid any extra function calls to dummy functions when running on real
hardware with no console.
Signed-off-by: Daniel Lublin <daniel@lublin.se>
Introduce memcpy_s() and wordcpy_s() that takes the destination buffer
size as an argument. Use assert() which aborts our program to an
eternal loop if we hit problems.
Sprinkle asserts elsewhere as well.
Signed-off-by: Daniel Lublin <daniel@lublin.se>
In firmware we store the address to firmware blake2s() function at
TK1_MMIO_TK1_BLAKE2S so app can use this firmware function sort of
like a system call but without context switch.
