Shuffle things around a bit inside mbedtls_mpi_random() in preparation for
breaking out mbedtls_mpi_core_random().
Signed-off-by: Gilles Peskine <Gilles.Peskine@arm.com>
Rewrite the minimum bound comparison to avoid a local allocation. This costs
a bit of code size, but saves RAM. This is in preparation for moving the
bulk of the function to the bignum_core module where allocation is not
permitted.
Signed-off-by: Gilles Peskine <Gilles.Peskine@arm.com>
The first half of the table is not used, let's reuse index 0 for the
result instead of appending it in the end.
Signed-off-by: Janos Follath <janos.follath@arm.com>
The table size was set before the configured window size bound was
applied which lead to out of bounds access when the configured window
size bound is less.
Signed-off-by: Janos Follath <janos.follath@arm.com>
The window size starts giving diminishing returns around 6 on most
platforms and highly unlikely to be more than 31 in practical use cases.
Still, compilers and static analysers might complain about this and
better to be pedantic.
Co-authored-by: Gilles Peskine <gilles.peskine@arm.com>
Signed-off-by: Janos Follath <janos.follath@arm.com>
With small exponents (for example, when doing RSA-1024 with CRT, each
prime is 512 bits and we'll use wsize = 5 which may be smaller that the
maximum - or even worse when doing public RSA operations which typically
have a 16-bit exponent so we'll use wsize = 1) the usage of W will have
pre-computed values, then empty space, then the accumulator at the very
end.
Move X next to the precomputed values to make accesses more efficient
and intuitive.
Signed-off-by: Janos Follath <janos.follath@arm.com>
Elements of W didn't all have the same owner: all were owned by this
function, except W[x_index]. It is more robust if we make a proper copy
of X.
Signed-off-by: Janos Follath <janos.follath@arm.com>
Out of window zeroes were doing squaring on the output variable
directly. This leaks the position of windows and the out of window
zeroes.
Loading the output variable from the table in constant time removes this
leakage.
Signed-off-by: Janos Follath <janos.follath@arm.com>
When x is the most negative value of a two's complement type,
`(unsigned_type)(-x)` has undefined behavior, whereas `-(unsigned_type)x`
has well-defined behavior and does what was intended.
Signed-off-by: Gilles Peskine <Gilles.Peskine@arm.com>
Fix undefined behavior (typically harmless in practice) of
mbedtls_mpi_add_mpi(), mbedtls_mpi_add_abs() and mbedtls_mpi_add_int() when
both operands are 0 and the left operand is represented with 0 limbs.
Signed-off-by: Gilles Peskine <Gilles.Peskine@arm.com>
In mbedtls_mpi_add_mpi() and mbedtls_mpi_sub_mpi(), and by extention
mbedtls_mpi_add_int() and mbedtls_mpi_sub_int(), when the resulting value
was zero, the sign bit of the result was incorrectly set to -1 when the
left-hand operand was negative. This is not a valid mbedtls_mpi
representation. Fix this: always set the sign to +1 when the result is 0.
Signed-off-by: Gilles Peskine <Gilles.Peskine@arm.com>
mbedtls_mpi_add_mpi() and mbedtls_mpi_sub_mpi() have the same logic, just
with one bit to flip in the sign calculation. Move the shared logic to a new
auxiliary function. This slightly reduces the code size (if the compiler
doesn't inline) and reduces the maintenance burden.
Signed-off-by: Gilles Peskine <Gilles.Peskine@arm.com>
Turn mpi_fill_random_internal() into mbedtls_mpi_core_fill_random(). It
had basically the right code except for how X is passed to the function.
Write unit tests.
Signed-off-by: Gilles Peskine <Gilles.Peskine@arm.com>
This commit contains the function prototype for mbedtls_mpi_core_shift_r,
and the implementation minimally modified from mbedtls_mpi_shift_r.
Signed-off-by: Gilles Peskine <Gilles.Peskine@arm.com>
We used to include platform.h only when MBEDTLS_PLATFORM_C was enabled, and
to define ad hoc replacements for mbedtls_xxx functions on a case-by-case
basis when MBEDTLS_PLATFORM_C was disabled. The only reason for this
complication was to allow building individual source modules without copying
platform.h. This is not something we support or recommend anymore, so get
rid of the complication: include platform.h unconditionally.
There should be no change in behavior since just including the header should
not change the behavior of a program.
This commit replaces most occurrences of conditional inclusion of
platform.h, using the following code:
```
perl -i -0777 -pe 's!#if.*\n#include "mbedtls/platform.h"\n(#else.*\n(#define (mbedtls|MBEDTLS)_.*\n|#include <(stdarg|stddef|stdio|stdlib|string|time)\.h>\n)*)?#endif.*!#include "mbedtls/platform.h"!mg' $(git grep -l '#include "mbedtls/platform.h"')
```
Signed-off-by: Gilles Peskine <Gilles.Peskine@arm.com>
The unit tests were created by capturing runs of the existing function during
execution of existing unit tests.
Signed-off-by: Tom Cosgrove <tom.cosgrove@arm.com>
Co-authored-by: Tom Cosgrove <81633263+tom-cosgrove-arm@users.noreply.github.com>
Co-authored-by: Werner Lewis <werner.wmlewis@gmail.com>
Signed-off-by: Janos Follath <janos.follath@arm.com>
This reverts commit 62c5901f0a5061a8825e19c77f88c91fea235078.
Reverting commit due the macros are meant to be local and not following the
naming convention.
Signed-off-by: Gabor Mezei <gabor.mezei@arm.com>
Fix a null pointer dereference when performing some operations on zero
represented with 0 limbs: mbedtls_mpi_mod_int() dividing by 2, or
mbedtls_mpi_write_string() in base 2.
Signed-off-by: Gilles Peskine <Gilles.Peskine@arm.com>
Compilers are likely to generate shorter assembly for loops of the
form `while( cnt-- ) { ... }` rather than
`for( ; count >= X; count -= X ) { ... }`. (E.g. the latter needs
a subtract+compare+branch after each loop, while the former only
needs decrement+branch).
Signed-off-by: Hanno Becker <hanno.becker@arm.com>
Removing the trimming has significant memory impact. While it is clearly what
we want to do eventually for constant-time'ness, it should be fixed alongside
a strategy to contain the ramifications on memory usage.
Signed-off-by: Hanno Becker <hanno.becker@arm.com>
A previous commit has changed the signature of mpi_mul_hlp, making the length
of the output explicit. This commit adjusts mpi_montmul() accordingly.
It also fixes a comment on the required size of the temporary value
passed to mpi_montmul() (but does not change the call-sites).
Signed-off-by: Hanno Becker <hanno.becker@arm.com>
A previous commit has changed the signature of mpi_mul_hlp(), making
the length of the output explicit.
This commit adjusts mbedtls_mpi_mul_int() to this change.
Along the way, we make the code simpler and more secure by not calculating
the minimal limb-size of A. A previous comment indicated that this was
functionally necessary because of the implementation of mpi_mul_hlp() --
if it ever was, it isn't anymore.
Signed-off-by: Hanno Becker <hanno.becker@arm.com>
The previous commit has changed the signature of mpi_mul_hlp(),
making the length of the output explicit.
This commit adjusts the call-site in mbedtls_mpi_mul_mpi() to
this new signature.
A notable change to the multiplication strategy had to be made:
mbedtls_mpi_mul_mpi() performs a simple row-wise schoolbook
multiplication, which however was so far computed iterating
rows from top to bottom. This leads to the undesirable consequence
that as lower rows are calculated and added to the temporary
result, carry chains can grow. It is simpler and faster to
iterate from bottom to top instead, as it is guaranteed that
there will be no carry when adding the next row to the previous
temporary result: The length of the output in each iteration
can be fixed to len(B)+1.
Signed-off-by: Hanno Becker <hanno.becker@arm.com>
The helper `mpi_mul_hlp()` performs a multiply-accumulate
operation `d += s * b`, where `d,b` are MPIs and `b` is a scalar.
Previously, only the length of `s` was specified, while `d` was
assumed to be 0-terminated of unspecified length.
This was leveraged at the end of the core multiplication steps
computingg the first `limb(s)` limbs of `d + s*b`: Namely, the
routine would keep on adding the current carry to `d` until none
was left. This can, in theory, run for an arbitrarily long time
if `d` has a tail of `0xFF`s, and hence the assumption of
`0`-termination.
This solution is both fragile and insecure -- the latter because
the carry-loop depends on the result of the multiplication.
This commit changes the signature of `mpi_mul_hlp()` to receive
the length of the output buffer, which must be greater or equal
to the length of the input buffer.
It is _not_ assumed that the output buffer is strictly larger
than the input buffer -- instead, the routine will simply return
any carry that's left. This will be useful in some applications
of this function. It is the responsibility of the caller to either
size the output appropriately so that no carry will be left, or
to handle the carry.
NOTE: The commit leaves the library in a state where it cannot
be compiled since the call-sites of mpi_mul_hlp() have
not yet been adjusted. This will be done in the subsequent
commits.
Signed-off-by: Hanno Becker <hanno.becker@arm.com>
This commit removes code from the Montgomery multiplication routine
`mpi_montmul()` which seems to serve no purpose.
Details: `mpi_montmul()` uses a temporary storage `T` for intermediate
results which is assumed to be of twice the size as the inputs to be
multiplied, and which is used as follows: After the i-th (i=0,1,...)
iteration, the n-limb word starting at `T->p + i + 1` contains the
Montgomery multiplication of B with the limbs 0,..,i of A, and the
variable `d` points to `T->p + i + 1`. In particular, after `n` iterations,
`T->p + n` holds the full multiplication
(subject to conditional subtraction).
As a consequence of this way of using the temporary `T`, the contents
of `{T->p, ..., T->p + i}` are irrelevant after the i-th iteration. Nonetheless,
the code copies `A[i]` to `T->p[i]` at the end of the i-th iterations, which is
redundant and can be removed.
Signed-off-by: Hanno Becker <hanno.becker@arm.com>
Elinimate macros defined by modules locally in the functions that are
moving to the new constant-time module.
Signed-off-by: gabor-mezei-arm <gabor.mezei@arm.com>
There were multiple functions called mbedtls_cf_size_bool_eq. They had exactly
the same behavior, so move the one in bignum.c and remove the other.
Signed-off-by: gabor-mezei-arm <gabor.mezei@arm.com>
The loop exits early iff there is a nonzero limb, so i==0 means that
all limbs are 0, whether the number of limbs is 0 or not.
Signed-off-by: Gilles Peskine <Gilles.Peskine@arm.com>
Fix a bug introduced in "Fix multiplication producing a negative zero" that
caused the sign to be forced to +1 when A > 0, B < 0 and B's low-order limb
is 0.
Add a non-regression test. More generally, systematically test combinations
of leading zeros, trailing zeros and signs.
Signed-off-by: Gilles Peskine <Gilles.Peskine@arm.com>
