The code was making two unsequenced reads from volatile locations.
This is undefined behavior. It was probably harmless because we didn't
care in what order the reads happened and the reads were from ordinary
memory, but UB is UB and IAR8 complained.
Get rid of the variable p. This makes it more apparent where the code
accesses the buffer at an offset whose value is sensitive.
No intended behavior change in this commit.
Rather than doing the quadratic-time constant-memory-trace on the
whole working buffer, do it on the section of the buffer where the
data to copy has to lie, which can be significantly smaller if the
output buffer is significantly smaller than the working buffer, e.g.
for TLS RSA ciphersuites (48 bytes vs MBEDTLS_MPI_MAX_SIZE).
In mbedtls_rsa_rsaes_pkcs1_v15_decrypt, use size_greater_than (which
is based on bitwise operations) instead of the < operator to compare
sizes when the values being compared must not leak. Some compilers
compile < to a branch at least under some circumstances (observed with
gcc 5.4 for arm-gnueabi -O9 on a toy program).
Replace memmove(to, to + offset, length) by a functionally equivalent
function that strives to make the same memory access patterns
regardless of the value of length. This fixes an information leak
through timing (especially timing of memory accesses via cache probes)
that leads to a Bleichenbacher-style attack on PKCS#1 v1.5 decryption
using the plaintext length as the observable.
mbedtls_rsa_rsaes_pkcs1_v15_decrypt takes care not to reveal whether
the padding is valid or not, even through timing or memory access
patterns. This is a defense against an attack published by
Bleichenbacher. The attacker can also obtain the same information by
observing the length of the plaintext. The current implementation
leaks the length of the plaintext through timing and memory access
patterns.
This commit is a first step towards fixing this leak. It reduces the
leak to a single memmove call inside the working buffer.
Make the function more robust by taking an arbitrary zero/nonzero
argument instead of insisting on zero/all-bits-one. Update and fix its
documentation.
mbedtls_rsa_rsaes_pkcs1_v15_decrypt took care of calculating the
padding length without leaking the amount of padding or the validity
of the padding. However it then skipped the copying of the data if the
padding was invalid, which could allow an adversary to find out
whether the padding was valid through precise timing measurements,
especially if for a local attacker who could observe memory access via
cache timings.
Avoid this leak by always copying from the decryption buffer to the
output buffer, even when the padding is invalid. With invalid padding,
copy the same amount of data as what is expected on valid padding: the
minimum valid padding size if this fits in the output buffer,
otherwise the output buffer size. To avoid leaking payload data from
an unsuccessful decryption, zero the decryption buffer before copying
if the padding was invalid.
`mbedtls_ssl_get_record_expansion()` is supposed to return the maximum
difference between the size of a protected record and the size of the
encapsulated plaintext.
Previously, it did not correctly estimate the maximum record expansion
in case of CBC ciphersuites in (D)TLS versions 1.1 and higher, in which
case the ciphertext is prefixed by an explicit IV.
This commit fixes this bug. Fixes#1914.
In `mbedtls_ccm_self_test()`, enforce input and output
buffers sent to the ccm API to be contigous and aligned,
by copying the test vectors to buffers on the stack.
- in x509_profile_check_pk_alg
- in x509_profile_check_md_alg
- in x509_profile_check_key
and in ssl_cli.c : unsigned char gets promoted to signed integer
In ecp_mul_comb(), if (!p_eq_g && grp->T == NULL) and then ecp_precompute_comb() fails (which can
happen due to OOM), then the new array of points T will be leaked (as it's newly allocated, but
hasn't been asigned to grp->T yet).
Symptom was a memory leak in ECDHE key exchange under low memory conditions.
The length to the debug message could conceivably leak through the time it
takes to print it, and that length would in turn reveal whether padding was
correct or not.
The basis for the Lucky 13 family of attacks is for an attacker to be able to
distinguish between (long) valid TLS-CBC padding and invalid TLS-CBC padding.
Since our code sets padlen = 0 for invalid padding, the length of the input to
the HMAC function, and the location where we read the MAC, give information
about that.
A local attacker could gain information about that by observing via a
cache attack whether the bytes at the end of the record (at the location of
would-be padding) have been read during MAC verification (computation +
comparison).
Let's make sure they're always read.
The basis for the Lucky 13 family of attacks is for an attacker to be able to
distinguish between (long) valid TLS-CBC padding and invalid TLS-CBC padding.
Since our code sets padlen = 0 for invalid padding, the length of the input to
the HMAC function gives information about that.
Information about this length (modulo the MD/SHA block size) can be deduced
from how much MD/SHA padding (this is distinct from TLS-CBC padding) is used.
If MD/SHA padding is read from a (static) buffer, a local attacker could get
information about how much is used via a cache attack targeting that buffer.
Let's get rid of this buffer. Now the only buffer used is the internal MD/SHA
one, which is always read fully by the process() function.
