Forbidding nested calls to acquire() allows to remove the reference
counting logic and hence saving some bytes of code. This is valuable
because MBEDTLS_X509_ALWAYS_FLUSH is likely to be used on constrained
systems where code-size is limited.
Previously, reference counting for the CRT frames and PK contexts
handed out by mbedtls_x509_crt_{frame|pk}_acquire() was implemented
only in case threading support was enabled, which leaves the door
open for a potential use-after-free should a single-threaded application
use nested calls to mbedtls_x509_crt_acquire().
Since Mbed TLS itself does not use such nested calls, it might be
preferred long-term to forbid nesting of acquire calls on the API
level, and hence get rid of reference counting in the interest of
code-size benefits. However, this can be considered as an optimization
of X.509 on demand parsing, and for now this commit introduces
reference counting unconditionally to have a safe version of
on demand parsing to build further optimizations upon.
Previously, only one thread could access the parsing cache of an X.509 CRT
at a time. Firstly, this leads to significant performance penalties on
systems running many concurrent threads which share CRT structures --
for example, server threads sharing an SSL configuration containing the
server CRT. Secondly, the locking should be logically unnecessary, because
the threads are supposed to access the CRT frame and PK in a read-only,
or at least thread-safe manner.
This commit modifies the X.509 CRT cache implementation by allowing an
arbitrary number of concurrent readers, locking only the path of setting
up and clearing the cache.
We cannot move it to x509_crt.c because there are some static inline
function definitions in x509_crt.h which access members of
mbedtls_x509_crt_cache.
