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.
Adjust the size/performance trade-off:
* Reduces size of sha256_process() from 7.4KB to 2KB on ARMv7-M
* Reduces performance by less than 14% on Cortex-M4
* Seems to even improve performance on my Core i7
* mbedtls-1.3:
Use link-time garbage collection in memory.sh
scripts/memory.sh only work on Linux
Add missing 'const' on selftest data
Use only headers for doxygen (no doc in C files)
Add missing extern "C" guard in aesni.h
Fix compile error with renego disabled
Remove slow PKCS5 test
Stop checking key-cert match systematically
Make tests/*.sh runnable from anywhere
Update visual C files
