Fix issue #1315 where filenames calculated with a base distance of zero
(ie the characters add up to 0(mod 256) aren't de-obfuscated on reading.
This was due to overloading of "0" to also mean "invalid UTF8; no rotation",
so we remove that double meaning
This is a simple "rotate" of the filename, with each file having a rot
distance based on the filename. We store the distance at the beginning
of the filename. So a file called "go" would become "37.KS".
This is not a strong encryption of filenames, but it should stop automated
scanning tools from picking up on filename patterns. As such it's an
intermediate between "off" and "standard". The advantage is that it
allows for longer path segment names.
We use the nameKey as an additional input to calculate the obfuscation
distance. This should mean that two different passwords will result
in two different keys
The obfuscation rotation works by splitting the ranges up and handle cases
0-9
A-Za-z
0xA0-0xFF
and anything greater in blocks of 256
This happened when the underlying reader returned io.ErrUnexpectedEOF.
The error handling for the call to io.ReadFull failed to take this
into account.
io.ErrUnexpectedEOF is reasonably common when SSL connections go wrong
when communicating with ACD, so it manifested itself as transfers from
non-encrypted ACD to encrypted ACD getting stuck.
Streams which truncated early with an EOF message would return a
"Failed to authenticate decrypted block" error. While technically
correct, this isn't a helpful error message as it masks the underlying
problem. This changes it to return "unexpected EOF" instead.
The rest of rclone knows it should retry such errors.
This was caused by failing to reset the internal buffer on seek so old
data was read first before the new data.
The unit tests didn't detect this because they were reading to the end
of the file to check integrity and thus emptying the internal buffer.
Both code and unit tests were fixed up.
