This is the same game as with the
generichash construction. We want
to protect it with a mutex so
different processes can safely do
work on the same resource.
While here, also move the _update
function onto the dirty scheduler.
It is by far the most expensive
operation, and why it wasn't there
in the first place is odd. This should
unblock the scheduler on long
sign-checks. It also move the
possible mutex block onto the
dirty scheduler thread, away from
the core schedulers, improving
latency in the system as a result.
While sodium is thread-safe, our
resources are not. Furthermore,
we might have an update call going
when someone decides to call
finalize and so on. It is not clever
to do so, but on the other hand
I want to protect against this.
While here, mark the mutexed
calls as dirty CPU. This avoids them
blocking the main scheduler and
only messes with the background
dirty threads, which is somewhat
more safe.
The consequence is that order
access to the resource is now
serialized. I don't think you should
do it, but it is now possible.
Since the chacha20poly1305 constructions were the IETF variants,
we renamed those so they follow the official library better. While
here, we also fixed the argument order of the files.
sign and generichash failed to release their resources under failure.
This can lead to subtle memory leaks in the very unlikely event
we can't initialize.
At least according to:
https://libsodium.gitbook.io/doc/hashing/generic_hashing
We noticed crashes when it was not 16-byte aligned - probably is
architecture dependent. This makes the safe choice and always 64-byte
align it.
The API for pwhash_str returns a cstring in the output buffer. These
are null terminated. However, we return the full buffer as a binary
back to Erlang. This means that we have a buffer with 0'es in the end.
The tests take this buffer and passes it back in as is. Hence all the
tests pass. However, it is conceivable that if we write said buffer to
disk somewhere, we are not going to write those 0's out.
When we then load the ASCII-armored Argon2 string into memory again,
it is not 0-terminated as a cstring should be, and this produces
errors all over the place.
The fix is twofold:
* Return the full buffer to Erlang, but use binary:split/2 to create a
subbinary with the relevant part.
* Add a 0 in the end of ASCII Argon2 string before passing it to
libsodium
Since we are looking at pwhashing, and Argon2, we expect the
computational problem to be memory bound. Thus, spending a bit more
work in memory is not going to have any considerable impact on the
speed of this system.
These variables are being initialized via calls to `enif_get_uint`,
so it's safer to declare them as unsigned int's rather than size_t's.
Their being used in calls to `enif_alloc_binary`, which takes a size_t
as its size.
However, the resulting ErlNifBinary keeps its size as an unsigned int,
so asking for a size that's an unsigned int should be safe.
This would be problematic in the case where sizeof(size_t) <
sizeof(unsigned), which would mean we're getting fewer bytes allocated
than expected.
Perhaps an explicit check for, for example, `hashSize > MAX_SIZE` would
be good here?
