transport_check_fds and transport_read_pdu had almost the same
functionality: reading and validating one pdu at a time.
Now transport_read_pdu reads one pdu from the transport layer and verifies
that the pdu data is valid - as before.
transport_read_pdu also ensures that the stream is sealed and
rewound when the pdu is received completely.
transport_check_fds just uses transport_read_pdu and does *not* do
the verification a second time based on the stream.
Besides the clean up this fixes the following problems:
* transport_read always read 4 bytes. Fast-path input synchronize pdus
are only 3 bytes long. In this case on byte got lost in the stream
buffer which lead to "de-synchronization" of server and
client.
* Size check in tpdu_read_connection_confirm - already read bytes
weren't taken into account.
select() has the major drawback that it cannot handle file descriptor
that are bigger than 1024. This patch makes use of poll() instead of
select() when poll() support is available.
This big patch allows to have non-blocking writes. To achieve
this, it slightly changes the way transport is handled. The misc transport
layers are handled with OpenSSL BIOs. In the chain we insert a
bufferedBIO that will bufferize write calls that couldn't be honored.
For an access with Tls security the BIO chain would look like this:
FreeRdp Code ===> SSL bio ===> buffered BIO ===> socket BIO
The buffered BIO will store bytes that couldn't be send because of
blocking write calls.
This patch also rework TSG so that it would look like this in the
case of SSL security with TSG:
(TSG in)
> SSL BIO => buffered BIO ==> socket BIO
/
FreeRdp => SSL BIO => TSG BIO
\
> SSL BIO => buffered BIO ==> socket BIO
(TSG out)
So from the FreeRDP point of view sending something is only BIO_writing
on the frontBio (last BIO on the left).
This patch fixes an issue with hung clients connected to FreeRDP server:
- Removed the completely useless inner while loop which used the transport
receive buffer stream position as break contition: The transport receive
buffer stream is replaced after each iteration of this loop with a fresh
instance from a pool which has the position set to 0. Thus it was
technically impossible that this loop would ever be run twice.
- We must not always return if transport_read_nonblocking returns 0:
transport_read_nonblocking() is also called in transport_write() and
therefore it is possible that the stream position of the transport
receive buffer is already > 0 when entering transport_check_fds.
