This is a preliminary patch allowing for socket-type-specific
functionality in the I/O thread. For example, message format
can be checked asynchronously and misbehaved connections dropped
straight away.
Signed-off-by: Martin Sustrik <sustrik@250bpm.com>
So far the requests in req/rep pattern were delivered to and processed
by worker even though the original requester was dead. Thus,
the worker processing replies with noone to deliver results to.
This optimisation drops requests in two situations:
1. Queued inbound requests in XREP socket when peer disconnects.
2. Queued outbound requests in XREQ when socket is closed.
Signed-off-by: Martin Sustrik <sustrik@250bpm.com>
This patch introduces two changes:
1. 32-bit ID is used to identify the peer instead of UUID
2. REQ socket seeds the label stack with unique 32-bit request ID
It also drops any replies with non-matching request ID
Signed-off-by: Martin Sustrik <sustrik@250bpm.com>
So far there was no distinction between message parts used by 0MQ
and message parts used by user. Now, the message parts used by 0MQ
are marked as 'LABEL'.
Signed-off-by: Martin Sustrik <sustrik@250bpm.com>
So far, the pipe termination code was spread among socket type
classes, fair queuer, load balancer, etc. This patch moves
all the associated logic to a single place.
Signed-off-by: Martin Sustrik <sustrik@250bpm.com>
So far, there was a pair of unidirectional pipes between a socket
and a session (or an inproc peer). This resulted in complex
problems with half-closed states and tracking which inpipe
corresponds to which outpipe.
This patch doesn't add any functionality in itself, but is
essential for further work on features like subscription
forwarding.
Signed-off-by: Martin Sustrik <sustrik@250bpm.com>
When an inpipe terminated within XREP, it was erased from the array
and thus current_in (which is an index) pointed to a different
element in the array. This caused problems when we were in the
middle of reading a multipart message.
Signed-off-by: Martin Sustrik <sustrik@250bpm.com>
This patch addresses serveral issues:
1. It gathers message related functionality scattered over whole
codebase into a single class.
2. It makes zmq_msg_t an opaque datatype. Internals of the class
don't pollute zmq.h header file.
3. zmq_msg_t size decreases from 48 to 32 bytes. That saves ~33%
of memory in scenarios with large amount of small messages.
Signed-off-by: Martin Sustrik <sustrik@250bpm.com>
Threads were so far identified by integers called 'slots'.
This patch renames them to more comprehensible 'tid's (thread IDs).
Signed-off-by: Martin Sustrik <sustrik@250bpm.com>
1. ZMQ_LINGER option can be set/get
2. options are part of own_t base class rather than being declared
separately by individual objects
3. Linger option is propagated with "term" command so that the
newest value of it is used rather than the stored old one.
4. Session sets the linger timer if needed and terminates
as soon as it expires.
5. Corresponding documentation updated.
Signed-off-by: Martin Sustrik <sustrik@250bpm.com>
The patch was supposed to check that pipe writer sends messages
in atomic fashion. However, it prevented the user to read
half of a message and close the socket.
Signed-off-by: Martin Sustrik <sustrik@250bpm.com>
Specifically, shutdown of child objects is initiated *before* termination
handshake with socket object.
Signed-off-by: Martin Sustrik <sustrik@250bpm.com>
Sockets may now be migrated between OS threads; sockets may not be used by
more than one thread at any time. To migrate a socket to another thread the
caller must ensure that a full memory barrier is called before using the
socket from the target thread.
The new zmq_close() semantics implement the behaviour discussed at:
http://lists.zeromq.org/pipermail/zeromq-dev/2010-July/004244.html
Specifically, zmq_close() is now deterministic and while it still returns
immediately, it does not discard any data that may still be queued for
sending. Further, zmq_term() will now block until all outstanding data has
been sent.
TODO: Many bugs have been introduced, needs testing. Further, SO_LINGER or
an equivalent mechanism (possibly a configurable timeout to zmq_term())
needs to be implemented.
This commit introduces the necessary changes necessary
for implementing flow control. None of the socket types
implements the flow control yet. The code will crash when
the flow control is enabled and the thw lwm is reached.
The following commits will add flow-control support for
individual socket types.
