memory_buffer_alloc should use a sufficient default alignment

[gilles-peskine-arm]

Description

The MEMORY_BUFFER_ALLOC module defaults to aligning blocks on a 4-byte boundary.

include/mbedtls/memory_buffer_alloc.h:#define MBEDTLS_MEMORY_ALIGN_MULTIPLE       4 /**< Align on multiples of this value */

On architectures that require 8- or 16-byte alignment for some data, this results in buggy code.

The fix is to default MBEDTLS_MEMORY_ALIGN_MULTIPLE to something sensible. In C11, that's _Alignof(max_align_t). In older C dialects, should we use sizeof(char*) (backward compatible) or 2 * sizeof(char*) (I don't know of a machine where it's required, but it can improve performance with vector instructions)?

To test, we should run a job in all.sh with MBEDTLS_MEMORY_BUFFER_ALLOC_C enabled and UBSan enabled (it detects misaligned pointer accesses). At the moment, MBEDTLS_MEMORY_BUFFER_ALLOC_C is tested via config.pl full. In Mbed TLS, we've changed that to a dedicated component_test_memory_buffer_allocator. We should wait until this change is sideported to crypto, then modify the test component to enable UBSan.

To reproduce:

scripts/config.pl set MBEDTLS_MEMORY_BUFFER_ALLOC_C
scripts/config.pl set MBEDTLS_MEMORY_DEBUG
scripts/config.pl set MBEDTLS_PLATFORM_MEMORY
make CFLAGS='-fsanitize=undefined' LDFLAGS='-fsanitize=undefined' lib tests
make test

Issue request type

[ ] Question
[ ] Enhancement
[x] Bug

[ciarmcom]

Internal Jira reference: https://jira.arm.com/browse/IOTCRYPT-967