Age | Commit message (Collapse) | Author |
|
This commit implements atomic operations for the armv7a architecture
using gcc inline assembly. This offers higher performance compared to
pthread mutexes.
Tested on an am3517 evm, clocked at 600MHz:
./inproc_thr 200 1000000
------------------------
53-60K messages / sec, pthread mutexes
73-90K messages / sec, assembly atomic ops
./inproc_lat 200 1000000
------------------------
average latency: 42.234 [us], pthread mutexes
average latency: 35.496 [us], assembly atomic ops
|
|
|
|
This patch is based on patch by Martin Hurton from 0MQ project.
Signed-off-by: Martin Sustrik <sustrik@250bpm.com>
|
|
Signed-off-by: Martin Sustrik <sustrik@250bpm.com>
|
|
This patch is a preliminary work for mergine TCP and IPC connecters
and listeners. The ultimated goal is to avoid code duplication
occuring in these classes.
Signed-off-by: Martin Sustrik <sustrik@250bpm.com>
|
|
Optimisation. Up to now new fd_set was initialised in
each signaler_wait call. Now the fd_set is initialised
once when signaler is created. This is useful espacially on
Windows where fd_set is list of pointers rather than bitset
and thus can be rather large.
Signed-off-by: Martin Sustrik <sustrik@250bpm.com>
|
|
This patch provides a single place for selecting polling mechanisms
(polling.hpp). Up to now the selection was spread among the build
system and several source files.
Signed-off-by: Martin Sustrik <sustrik@250bpm.com>
|
|
Signed-off-by: Martin Sustrik <sustrik@250bpm.com>
|
|
This patch allows for partial shutdown of the socket.
Signed-off-by: Martin Sustrik <sustrik@250bpm.com>
|
|
SURVEYOR socket now uses clock instance from socket_base_t
for measuring survey timeout. This allows for better performance
in tight xs_recv() loops.
Signed-off-by: Martin Sustrik <sustrik@250bpm.com>
|
|
This fixes a critical bug in zmq_getsockopt.
|
|
This patch instantiate a clock_t instance for each XS
socket. Thus, it is shared between subsequent calls
to xs_recv (and xs_send). That in turn significantly
limits the number of invocations of getimeofday (or similar)
when timeouts are used and recv/send is called in a
tight loop.
Signed-off-by: Martin Sustrik <sustrik@250bpm.com>
|
|
This is an optimisation. The idea is that when surveyor disappears
it makes no sense to process any surveys it have sent. There are
two cases handled:
- when (X)SURVEYOR socket is closed any unsent outbound messages
are discarded
- when (X)RESPONDENT detects that its peer have disconnected it
drops any unprocessed messages that have been received from it
Signed-off-by: Martin Sustrik <sustrik@250bpm.com>
|
|
It can be used to timeout the survey. Value is in milliseconds
and -1 means infinite (default).
Signed-off-by: Martin Sustrik <sustrik@250bpm.com>
|
|
Survey pattern is "multicast with reply". There are two roles:
surveyor and respondent. Surveyor publishes a survey which gets
delivered to all connected respondents. Each repondent can send
a response to the survey. All the responses are delivered to
the original surveyor. Once the surveyor decides that the survey
is over (e.g. deadline was reached) it can send initiate survey.
Late responses from old surveys are automatically discarded by
the surveyor socket.
Socket types: SURVEYOR, XSURVEYOR, RESPONDENT, XRESPONDENT
Patch also includes a test program with surveoyr, two respondents
and an intermediary device.
Signed-off-by: Martin Sustrik <sustrik@250bpm.com>
|
|
- couple of comments adjusted
- empty 'stack delimiter' message correctly closed
- non-reachable function xrep_t::rollback() removed
Signed-off-by: Martin Sustrik <sustrik@250bpm.com>
|
|
Signed-off-by: Martin Sustrik <sustrik@250bpm.com>
|
|
Signed-off-by: Gabriele Svelto <gabriele.svelto@gmail.com>
|
|
Signed-off-by: Martin Sustrik <sustrik@250bpm.com>
|
|
While --with-poller configure option could be used for selecting
the polling mechanism for I/O threads (asynchronous), synchronous
polling (xs_poll, signaler_t) was not affected by the option.
This patch uses selected mechanism is it is either 'select' or 'poll'.
Othrwise it chooses the mechanism according to the platform.
Signed-off-by: Martin Sustrik <sustrik@250bpm.com>
|
|
thread_t is one of the classes where C++ syntax creates more
problems then it solves. This patch converts it into simple
C-style pseudo-class.
Signed-off-by: Martin Sustrik <sustrik@250bpm.com>
|
|
Till now the subscription, as specified by the user, was send
upstream.
This patch allows SUB-side filter to transform the user subscription
into wire subscription. For example, only a has can be sent upstream
instead of a perfect subscription.
Signed-off-by: Martin Sustrik <sustrik@250bpm.com>
|
|
Signed-off-by: Martin Sustrik <sustrik@250bpm.com>
|
|
This patch propoagates the error from signaler and mailbox
initialisation up the stack.
To achieve this signaler and mailbox classes were re-written
is C-like syntax.
Finally, shutdown_stress test now ignores EMFILE/ENFILE errors.
Thus, the tests should pass even on OSX which sets the max
number of file descriptors pretty low by default.
Signed-off-by: Martin Sustrik <sustrik@250bpm.com>
|
|
- subscription forwarding is handled by XSUB socket
- filtering is handled by SUB sockets
- subscriptions are decoupled from filter engines
- filter doesn't have to be able to enumarate the subscriptions
(no sf_enumerate function)
Signed-off-by: Martin Sustrik <sustrik@250bpm.com>
|
|
* src/stream_engine.cpp (stream_engine_t::stream_engine_t):
Use
#ifdef SO_NOSIGPIPE
instead of
#if defined XS_HAVE_OSX || defined XS_HAVE_FREEBSD
|
|
Signed-off-by: Martin Sustrik <sustrik@250bpm.com>
|
|
This patch consolidates the up-to-now scattered code that
sets different options on newly created sockets.
There are open_socket and open_tcp_socket functions that
do the tuning automatically.
In case the socket is not created but got from elsewhere
(such as accept() call) there are tune_socket and tune_tcp_socket
functions that will do the tuning.
Signed-off-by: Martin Sustrik <sustrik@250bpm.com>
|
|
Signed-off-by: Martin Sustrik <sustrik@250bpm.com>
|
|
Signed-off-by: Martin Sustrik <sustrik@250bpm.com>
|
|
On Windows, HANDLE_FLAG_INHERIT is equivalent to
O_CLOEXEC on UNIX.
Signed-off-by: Martin Sustrik <sustrik@250bpm.com>
|
|
on laggy networks
|
|
Signed-off-by: Martin Sustrik <sustrik@250bpm.com>
|
|
Signed-off-by: Gabriele Svelto <gabriele.svelto@gmail.com>
|
|
While TCP and IPC socket duplicates are correctly closed on
fork+exec, file descriptors used for internal communication
within libxs are not. This patch fixes the problem.
Signed-off-by: Martin Sustrik <sustrik@250bpm.com>
|
|
If accept() call fails due to insuffient OS resources
the new connection is rejected.
Signed-off-by: Martin Sustrik <sustrik@250bpm.com>
|
|
Signed-off-by: Martin Sustrik <sustrik@250bpm.com>
|
|
The socket connecting using the inproc transport never
received the identity of the remote peer.
|
|
The scoket implementation for inproc transfer failed to flush
identity message. The result was that the identity message
was not delivered until after the user sent the first message.
The identity message was never delivered if the user
used the socket only to receive messages.
|
|
|
|
libxs (as opposed to libzmq) requires at least one I/O thread per
context. Thus, "no I/O thread available" error doesn't make sense
anyome.
This patch removes the code and the documentation.
It keeps the definition of EMTHREAD itself for backward compatibility.
Signed-off-by: Martin Sustrik <sustrik@250bpm.com>
|
|
Signed-off-by: Martin Sustrik <sustrik@250bpm.com>
|
|
Previous patch have broken flushing a delimiter in pipe_t::terminate()
function. This patch allows the flush to happen even if pipe state
is set to 'terminated'.
Signed-off-by: Martin Sustrik <sustrik@250bpm.com>
|
|
* src/pipe.cpp (flush): Fix bogus typo: '&&' was used instead of '||'
|
|
This patch introduces following features:
- XS_PLUGIN context option to add plugins to libxs
- XS_FILTER option to switch between different filter types
- Automatic loading of plug-ins is *not* implemented.
From the implementation point of view:
- standard prefix filter is implemented as a pluggable filter
- trie_t and mtrie_t are joined into a single class
- the code for 0MQ/3.1 compatibility is left in in the form of comments
- new test for testing re-subscriptions is added
Signed-off-by: Martin Sustrik <sustrik@250bpm.com>
|
|
Although not a POSIX behaviour, FreeBSD can fail with ECONNRESET.
Signed-off-by: Martin Sustrik <sustrik@250bpm.com>
|
|
- XS_PROTOCOL option added
- libxs ignores when unused flags are set to 1 (0MQ/2.1 bug)
- compatibility tests added
Signed-off-by: Martin Sustrik <sustrik@250bpm.com>
|
|
This option allows to turn on TCP keepalives on the underlying connections.
Signed-off-by: Martin Sustrik <sustrik@250bpm.com>
|
|
When connection with duplicate identity arrives it is ignored.
Till now it resulted in assertion.
Signed-off-by: Martin Sustrik <sustrik@250bpm.com>
|
|
Signed-off-by: Martin Sustrik <sustrik@250bpm.com>
|