In cases where many small writes are made to a stream
lacking _writev, the array data structure backing the
WriteReq buffer would greatly increase GC pressure.
Specifically, in the fs.WriteStream case, the
clearBuffer routine would only clear a single WriteReq
from the buffer before exiting, but would cause the
entire backing array to be GC'd. Switching to [].shift
lessened pressure, but still the bulk of the time was
spent in memcpy.
This replaces that structure with a linked list-backed
queue so that adding and removing from the queue is O(1).
In the _writev case, collecting the buffer requires an
O(N) loop over the buffer, but that was already being
performed to collect callbacks, so slowdown should be
neglible.
PR-URL: https://github.com/joyent/node/pull/8826
Reviewed-by: Timothy J Fontaine <tjfontaine@gmail.com>
Reviewed-by: Trevor Norris <trev.norris@gmail.com>
Instead of simply creating a new v8::Object to contain the connection
information, instantiate a new instance of a FunctionTemplate. This will
allow future improvements for debugging and performance probes.
Additionally, the "provider" argument in the ReqWrap constructor is no
longer optional.
PR-URL: https://github.com/joyent/node/pull/8110
Signed-off-by: Trevor Norris <trev.norris@gmail.com>
Reviewed-by: Fedor Indutny <fedor@indutny.com>
Reviewed-by: Alexis Campailla <alexis@janeasystems.com>
Reviewed-by: Julien Gilli <julien.gilli@joyent.com>
Add address and/or port to errors where applicable for better reporting.
In the event the local address and port are accessible, it will also add
those to the error message.
See joyent/node#7005
Reviewed-By: Fedor Indutny <fedor@indutny.com>
Reviewed-By: Ben Noordhuis <info@bnoordhuis.nl>
PR-URL: https://github.com/iojs/io.js/pull/35
Instead of simply creating a new v8::Object to contain the connection
information, instantiate a new instance of a FunctionTemplate. This will
allow future improvements for debugging and performance probes.
Additionally, the "provider" argument in the ReqWrap constructor is no
longer optional.
PR-URL: https://github.com/joyent/node/pull/8110
Signed-off-by: Trevor Norris <trev.norris@gmail.com>
Reviewed-by: Fedor Indutny <fedor@indutny.com>
Reviewed-by: Alexis Campailla <alexis@janeasystems.com>
Reviewed-by: Julien Gilli <julien.gilli@joyent.com>
Turn on strict mode for the files in the lib/ directory. It helps
catch bugs and can have a positive effect on performance.
PR-URL: https://github.com/node-forward/node/pull/64
Reviewed-By: Colin Ihrig <cjihrig@gmail.com>
Reviewed-By: Fedor Indutny <fedor@indutny.com>
Currently when a server receives a new connection the underlying socket
handle begins reading data immediately. This causes problems when
sockets are passed between processes, as data can be read by the first
process and thus never read by the second process.
This commit allows sockets that are constructed with a handle to be
paused initially.
PR-URL: https://github.com/joyent/node/pull/8576
Fixes: https://github.com/joyent/node/issues/7905
Fixes: https://github.com/joyent/node/issues/7784
Reviewed-by: Trevor Norris <trev.norris@gmail.com>
The property server.connections should no longer be enumerable because
it has been deprecated. This will prevent deprecation warnings when
server objects are accessed by functions such as JSON.stringify.
Fixes: https://github.com/joyent/node/issues/8373
Reviewed-by: Trevor Norris <trev.norris@gmail.com>
Allow cluster workers to listen on exclusive ports for TCP and UDP,
instead of forcing all calls to go through the cluster master.
Fixes: #3856
Reviewed-by: Trevor Norris <trev.norris@gmail.com>
Reviewed-by: Fedor Indutny <fedor@indutny.com>
node::StringBytes::Write() has appropriate support to write strings with
'binary' encoding. So expose that API through StreamWrap and allow
inheriting classes to use it.
Signed-off-by: Trevor Norris <trev.norris@gmail.com>
net Sockets were calling read(0) to start reading, without
checking to see if they were paused first. This would result
in paused Socket objects keeping the event loop alive.
Fixes#8200
Reviewed-by: Trevor Norris <trev.norris@gmail.com>
This is the Node side of the fix for Node's cluster module on Windows.
https://github.com/joyent/node/issues/7691
The other required part is
https://github.com/joyent/libuv/pull/1384
Windows and Unix return certain socket errors (i.e. EADDRINUSE) at
different times: bind on Windows, and listen on Unix.
In an effort to hide this difference, libuv on Windows stores such
errors in the bind_error field of uv_tcp_t, to defer raising it at
listen time.
This worked fine except for the case in which a socket is shared in
a Node cluster and a bind error occurs.
A previous attempt to fix this (
d1e6be14603da36fe00e
) was flawed becaused in an attempt to relay the error at the JS level
it caused the master to start accepting connections.
With this new approach, libuv itself is relaying the bind errors,
providing for a uniform behavior of uv_tcp_listen.
Reviewed-By: Fedor Indutny <fedor@indutny.com>
Currently the address resolution family defaults to IPv4. Instead remove
the preference and instead resolve to a family suitable for the host.
Expose the getaddrinfo flags and allow them to be passed.
Add documentation about new flags.
Reviewed-by: Trevor Norris <trev.norris@gmail.com>
This is the Node side of the fix for Node's cluster module on Windows.
https://github.com/joyent/node/issues/7691
The other required part is
https://github.com/joyent/libuv/pull/1384
Windows and Unix return certain socket errors (i.e. EADDRINUSE) at
different times: bind on Windows, and listen on Unix.
In an effort to hide this difference, libuv on Windows stores such
errors in the bind_error field of uv_tcp_t, to defer raising it at
listen time.
This worked fine except for the case in which a socket is shared in
a Node cluster and a bind error occurs.
A previous attempt to fix this (
d1e6be14603da36fe00e
) was flawed becaused in an attempt to relay the error at the JS level
it caused the master to start accepting connections.
With this new approach, libuv itself is relaying the bind errors,
providing for a uniform behavior of uv_tcp_listen.
Reviewed-By: Fedor Indutny <fedor@indutny.com>
Switch condition order to check for null before calling isNaN().
Also remove two unnecessary calls to isNaN() that are already
covered by calls to isFinite(). This commit targets v0.10, as
opposed to #7891, which targets master (suggested by
@bnoordhuis). Closes#7840.
Signed-off-by: Fedor Indutny <fedor@indutny.com>
Fix for `error` events emitting only once when reconnecting
a single instance of net.Socket.
Fixesjoyent/node#7888
Signed-off-by: Fedor Indutny <fedor@indutny.com>
See https://github.com/joyent/node/issues/7675
net.server.listen() behaves inconsistently depending on whether the port
number is provided.
1. port === 0 && host == '' (i.e. false-y), node creates an AF_INET
socket but does not call bind().
2. port > 0 && host == '', node creates an AF_INET6 socket and calls
bind().
The fix makes 1 consistent with 2.
Signed-off-by: Fedor Indutny <fedor@indutny.com>
When close() is called on a non-listening server, a synchronous
error is thrown. This commit causes the error to be passed to
the asynchronous callback function instead.
Signed-off-by: Fedor Indutny <fedor@indutny.com>
Expose `setBlocking` on Pipe's and if a pipe is being created for stdio
on windows then make the pipes blocking.
This fixes test-stream2-stderr-sync.js on Windows.
Fixes#3584
Expose localPort for binding to a specific port for outbound
connections.
If localAddress is not specified '0.0.0.0' is used for ip4 and '::'
for ip6 connections.
Fixes#7092
Before sending a socket from a cluster master to a worker,
we would call listen in UV but not handle the error.
I made createServerHandle call listen on Windows so we get a chance
the handle any bind/listen errors early.
This fix is 100% windows specific.
It fixes test-cluster-bind-twice and
test-cluster-shared-handle-bind-error on Windows.
This is a slightly modified revert of bc39bdd.
Getting domains to use AsyncListeners became too much of a challenge
with many edge cases. While this is still a goal, it will have to be
deferred for now until more test coverage can be provided.
The domain module has been switched over to use the domain module API as
much as currently possible. There are still some hooks in the
EventEmitter, but hopefully we can remove those in the future.
When `tls.connect()` is called with `socket` option, it should try to
reuse hostname previously passed to `net.connect()` and only after that
fall back to `'localhost'`.
fix#6409
Fix "Assertion failed" when trying to connect to non-int ports:
Assertion failed: (args[2]->Uint32Value()), function Connect,
file ../src/tcp_wrap.cc, line 379.
Abort trap: 6
`dns.lookup` defaults to selecting IPv4 record even if IPv6 is available
for the desired zone. Generally, this approach works, but if IPv4
address is unavailable - there'll be no other way to opt-out and connect using
IPv6 address than calling `dns.lookup` and passing it to `.connect()`
directly.
This commit adds `family` option to `net.connect` method to figure out
this issue.
Don't use v8::Object::SetHiddenValue() to keep a reference alive to the
buffer, we can just as easily do that from JS land and it's a lot faster
to boot.
Because the buffer is now a visible property of the write request
object, it's essential that we do *not* log it - we'd be effectively
serializing the whole buffer to a pretty-printed string.
Libuv now returns errors directly. Make everything in src/ and lib/
follow suit.
The changes to lib/ are not strictly necessary but they remove the need
for the abominations that are process._errno and node::SetErrno().
This is a big commit that touches just about every file in the src/
directory. The V8 API has changed in significant ways. The most
important changes are:
* Binding functions take a const v8::FunctionCallbackInfo<T>& argument
rather than a const v8::Arguments& argument.
* Binding functions return void rather than v8::Handle<v8::Value>. The
return value is returned with the args.GetReturnValue().Set() family
of functions.
* v8::Persistent<T> no longer derives from v8::Handle<T> and no longer
allows you to directly dereference the object that the persistent
handle points to. This means that the common pattern of caching
oft-used JS values in a persistent handle no longer quite works,
you first need to reconstruct a v8::Local<T> from the persistent
handle with the Local<T>::New(isolate, persistent) factory method.
A handful of (internal) convenience classes and functions have been
added to make dealing with the new API a little easier.
The most visible one is node::Cached<T>, which wraps a v8::Persistent<T>
with some template sugar. It can hold arbitrary types but so far it's
exclusively used for v8::Strings (which was by far the most commonly
cached handle type.)
This is only relevant for CentOS 6.3 using kernel version 2.6.32.
On other linuxes and darwin, the `read` call gets an ECONNRESET in that
case. On sunos, the `write` call fails with EPIPE.
However, old CentOS will occasionally send an EOF instead of a
ECONNRESET or EPIPE when the client has been destroyed abruptly.
Make sure we don't keep trying to write or read more in that case.
Fixes#5504
However, there is still the question of what libuv should do when it
gets an EOF. Apparently in this case, it will continue trying to read,
which is almost certainly the wrong thing to do.
That should be fixed in libuv, even though this works around the issue.
Buffer.byteLength() works only for string inputs. Thus, when connection
has pending Buffer to write, it should just use it's length instead of
throwing exception.
Consider this example:
// fd 3 is a bound tcp socket
var s = net.createServer(cb);
s.listen({ fd: 3 });
console.log(s.address()); // prints null
This commit makes net.Server#address() print the actual address.
Ditto for non-listen sockets; properties like net.Socket#localAddress
and net.Socket#remoteAddress now return the correct value.
Fixes#5009.
Don't emit the 'close' event with process.nextTick.
Closing a handle is an operation that usually *but not always* completes
on the next tick of the event loop, hence using process.nextTick is not
reliable.
Use a proper handle close callback and emit the 'close' event from
inside the callback.
Update tests that depend on the intricacies of the old model.
Fixes#3459.
1. Get rid of unnecessary 'finishing' flag
2. Dont check both ending and ended. Extraneous.
Also: Remove extraneous 'finishing' flag, and don't check both 'ending'
and 'ended', since checking just 'ending' is sufficient.
This commit fixes a bug where the cluster module fails to propagate
EADDRINUSE errors.
When a worker starts a (net, http) server, it requests the listen socket
from its master who then creates and binds the socket.
Now, OS X and Windows don't always signal EADDRINUSE from bind() but
instead defer the error until a later syscall. libuv mimics this
behaviour to provide consistent behaviour across platforms but that
means the worker could end up with a socket that is not actually bound
to the requested addresss.
That's why the worker now checks if the socket is bound, raising
EADDRINUSE if that's not the case.
Fixes#2721.
The stock writable stream "write after end" message is overly vague, if
you have clearly not called end() yourself yet.
When we receive a FIN from the other side, and call destroySoon() as a
result, then generate an EPIPE error (which is what would happen if you
did actually write to the socket), with a message explaining what
actually happened.
This makes it so that `stream.push(chunk)` is the only way to signal the
end of reading, removing the confusing disparity between the
callback-style _read method, and the fact that most real-world streams
do not have a 1:1 corollation between the "please give me data" event,
and the actual arrival of a chunk of data.
It is still possible, of course, to implement a `CallbackReadable` on
top of this. Simply provide a method like this as the callback:
function readCallback(er, chunk) {
if (er)
stream.emit('error', er);
else
stream.push(chunk);
}
However, *only* fs streams actually would behave in this way, so it
makes not a lot of sense to make TCP, TLS, HTTP, and all the rest have
to bend into this uncomfortable paradigm.
Don't emit a 'connect' event on sockets that are handed off to
net.Server 'connection' event listeners.
1. It's superfluous because the connection has already been established
at that point.
2. The implementation is arguably wrong because the event is emitted on
the same tick of the event loop while the rule of thumb is to always
emit it on the next one.
This has been tried before in commit f0a440d but was reverted again in
ede1acc because the change was incomplete (at least one test hadn't
been updated).
Fixes#1047 (again).
Previously, we were only destroying sockets on end if their readable
side had already been ended. This causes a problem for non-readable
streams, since we don't expect to ever see an 'end' event from those.
Treat the lack of a 'readable' flag the same as if it was an ended
readable stream.
Fix#4751
Let ECONNRESET network errors bubble up so clients can detect them.
Commit c4454d2e suppressed and turned them into regular end-of-stream
events to fix the then-failing simple/test-regress-GH-1531 test. See
also issue #1571 for (scant) details.
It turns out that special handling is no longer necessary. Remove the
special casing and let the error bubble up naturally.
pummel/test-https-ci-reneg-attack and pummel/test-tls-ci-reneg-attack
are updated because they expected an EPIPE error code that is now an
ECONNRESET. Suppression of the ECONNRESET prevented the test from
detecting that the connection has been severed whereupon the next
write would fail with an EPIPE.
Fixes#1776.
This commit breaks simple/test-stream2-stderr-sync. Need to figure out
a better way, or just accept that `(function W(){stream.write(b,W)})()`
is going to be noisy. People should really be using the `'drain'` event
for this use-case anyway.
This reverts commit 02f7d1bfd8.
Always defer the _write callback. The optimization here was only
relevant in some oddball edge cases that we don't actually care about.
Our benchmarks confirm that just always deferring the Socket._write cb
is perfectly fine to do, and in some cases, even slightly more
performant.
TCPWrap::Initialize() and PipeWrap::Initialize() should be called before
any data will be read from received socket. But, because of lazy
initialization of these bindings, Initialize() method isn't called.
Init bindings manually upon socket receiving.
See #4669
Keeping list of all sockets that were sent to child process causes memory
leak and thus unacceptable (see #4587). However `server.close()` should
still work properly.
This commit introduces two options:
* child.send(socket, { track: true }) - will send socket and track its status.
You should use it when you want to receive `close` event on sent sockets.
* child.send(socket) - will send socket without tracking it status. This
performs much better, because of smaller number of RTT between master and
child.
With both of these options `server.close()` will wait for all sent
sockets to get closed.
Keeping list of all sockets that were sent to child process causes memory
leak and thus unacceptable (see #4587). However `server.close()` should
still work properly.
This commit introduces two options:
* child.send(socket, { track: true }) - will send socket and track its status.
You should use it when you want `server.connections` to be a reliable
number, and receive `close` event on sent sockets.
* child.send(socket) - will send socket without tracking it status. This
performs much better, because of smaller number of RTT between master and
child.
With both of these options `server.close()` will wait for all sent
sockets to get closed.
socket.readyState, .readable, and .writable behavior changed as
a result of the new streaming interfaces. Updated to be backwards
compatible with current API and adds regression test.
closes#4461
Fix a bug where calling .end() on a socket without calling .connect() first
throws a TypeError:
TypeError: Cannot read property 'shutdown' of undefined
at Socket.onSocketFinish (net.js:194:20)
at Socket.EventEmitter.emit (events.js:91:17)
at Socket.Writable.end (_stream_writable.js:281:10)
at Socket.end (net.js:352:31)
Fixes#4463.
This is a combination of 6 commits.
* XXX net fixup lcase stream
* net: Refactor to use streams2
Use 'socket.resume()' in many tests to trigger old-mode behavior.
* net: Call destroy() if shutdown() is not provided
This is important for TTY wrap streams
* net: Call .end() in socket.destroySoon if necessary
This makes the http 1.0 keepAlive test pass, also.
* net wtf-ish stuff kinda busted
* net fixup
* Added isIP method to make use of inet_pton to cares_wrap.cc
* Modified net.isIP() to make use of new C++ isIP method.
* Added new tests to test-net-isip.js.
This is a back-port of commit fb6377e from the master branch.
* Added isIP method to make use of inet_pton to cares_wrap.cc
* Modified net.isIP() to make use of new C++ isIP method.
* Added new tests to test-net-isip.js.
This commit reverts the following commits (in reverse chronological order):
74d076c errnoException must be done immediately
ddb02b9 net: support Server.listen(Pipe)
085a098 cluster: do not use internal server API
d138875 net: lazy listen on handler
Commit d138875 introduced a backwards incompatible change that broke the
simple/test-net-socket-timeout and simple/test-net-lazy-listen tests - it
defers listening on the target port until the `net.Server` instance has at
least one 'connection' event listener.
The other patches had to be reverted in order to revert d138875.
Fixes#3832.
This fixes the problem that calling pause() on a socket would not
actually prevent 'data' events from being emitted. It also replaces
the existing test by a more elaborate one.
Ref: #3118
Problem: calling `server.listen()` (no port) on a net.Server triggered the
following libuv assertion:
node: ../deps/uv/src/unix/stream.c:406: uv__write: Assertion `fd_to_send >= 0'
failed.
Cause: uv_tcp_t handles are lazily initialized. Omitting the port made the
handle get initialized even more lazily. Too lazily - it wasn't initialized
when the handle was sent over to the child process.
Solution: implicitly bind to a random port in listen() when the port number
is omitted, it forces the handle to initialize. This is not a change in
behavior, listen() has always been identical to listen(0).
Fixes#3325.
Fix#3455.
The remoteAddress and remotePort properties are
dynamically retrieved from _getpeername().
While _getpeername() checks if the _handle is
null, it is also possible for the tcp_wrapped
_handle.getpeername() to return null on error.
Such a condition happens when the remote closes
and one of these properties is accessed before
_handle is set to null.
This implements server.listen({ fd: <filedescriptor> }). The fd should
refer to an underlying resource that is already bound and listening, and
causes the new server to also accept connections on it.
Not supported on Windows. Raises ENOTSUP.
The server 'close' event was emitted before the last client 'close' event. Not
exactly fatal but potentially confusing.
Before this commit the order looked something like [client, server, client],
now it looks like [client, client, server].
See #3340 for more details.
In case a worker would spawn a new subprocess with process.env, NODE_UNIQUE_ID
would have been a part of the env. Making the new subprocess believe it is a
worker, this would result in some confusion if the subprocess where to listen to
a port, since the server handle request would then be relayed to the worker.
This patch removes the NODE_UNIQUE_ID flag from process.env on startup so any
subprocess spawned by a worker is a normal process with no cluster stuff.
child_process.fork() support sending native hander object, this patch add support for sending
net.Server and net.Socket object by converting the object to a native handle object and back
to a useful object again.
Note when sending a Socket there was emitted by a net Server object, the server.connections
property becomes null, because it is no longer possible to known when it is destroyed.
This commit fixes a bug where the cluster module failed to propagate EADDRINUSE
errors.
When a worker starts a (net, http) server, it requests the listen socket from
its master who then creates and binds the socket.
Now, OS X and Windows don't always signal EADDRINUSE from bind() but instead
defer the error until a later syscall. libuv mimics this behaviour to provide
consistent behaviour across platforms but that means the worker could end up
with a socket that is not actually bound to the requested addresss.
That's why the worker now checks if the socket is bound, raising EADDRINUSE if
that's not the case.
Fixes#2721.
Passing a non-buffer or non-string argument to Socket.prototype.write triggered
an assert:
Assertion failed: (Buffer::HasInstance(args[0])), function Write,
file ../src/stream_wrap.cc, line 289.
Fixes#2532.
This makes it so that the stdin TTY-wrap stream gets ref'ed on
.resume() and unref'ed on .pause()
The semantics of the names "pause" and "resume" are a bit weird, but the
important thing is that this corrects an API change from 0.4 -> 0.6
which made it impossible to read from stdin multiple times, without
knowing when it might end up being closed. If no one has it open, this
lets the process die naturally.
LGTM'd by @ry
Don't allow `socket.destroy()` to run twice. The self-destruct sequence itself
is idempotent but it makes the 'close' and 'error' events fire more than once,
which may confuse listeners.
Fixes#2223.
Just a syntactic sugar for doing, for example:
var server = net.createServer(function (c) {
c.end('goodbye, cruel world!\r\n');
server.close().on('close', function () {
console.log('really, goodbye!');
});
}).listen(1337);
Fixes#1922.
This commit fixes two bugs in the handling of write requests when the connect()
call is still in progress.
1. The deferred write request's size was counted twice towards `.bytesWritten`.
2. The callback was not called. After connecting, `Socket.write()` was called
with three arguments (data, encoding, cb) but it ignored the third argument.
Coincidentally fixes test/simple/test-net-connect-buffer.js.
Only register once for listening when passing a callback to Server.listen(),
this prevents servers recycled using close() from invoking the callback when
Server.listen() is called later.
Sockets emitted by the 'connection' event are always connected, having
them emit the 'connect' event makes no sense. It only confused people,
as it's not clear if you have to listen to 'connect' or not.
That try..catch block was also very scary. It would silently swallow
exceptions in 'connect' listeners and destroy the socket. Makes no
sense.
Fixes#1047.
setImplementationMethods checks the type of a socket and defines different
behavior based on the type, so auto detect it if type not implicitly
specified.
Change to end() behavior in 33c339 was breaking it. end() should wait for
connection before dumping. Changed test-net-connect-timeout to use destroy()
instead.
Users too often would forget to add
socket.on('end', function () {
socket.end();
});
Which is a mistake. Therefore we default to this behavior and
only optionally let people handle the 'end' case themselves.
When a server hit EMFILE it would continue to try to accept new connections
from the queue. This patch introduces a timeout of one second where it will
stop trying to accept new files. After the second is over it tries again.
This is a rather serious bug that has been effecting many highly concurrent
programs. It was introduced in 4593c0, version v0.2.0.
TODO: A test for this situation. Currently I test it like this
termA% cd projects/node
termA% ulimit -n 256
termA% ./node benchmark/idle_server.js
termB% cd projects/node
termB% ./node benchmark/idle_clients.js
And watch how the server process behaves.
- Without this, recvMsg can be invoked before the event emitter gets a
chance to run. In this case, recvMsg.fd will be overwritten and the
original caller can end up emitting null.
This is needed in case the provided socket is not the default 'tcp4' type
(i.e. and needs different read/write/etc methods). With this patch, one can
call listenFD(sock, 'unix') to bind to existing UNIX domain sockets.
Now that FD passing is in master, it'd be great to be able to use a received
socket (which has already had bind(2) and listen(2) called on it) to fire up a
new net.Server instance. This patch adds a net.Server.listenFD() method which
will start up the accept watcher on the provided FD.
a) create a layer of indirection in net.Stream to allow swapping in
different read/write implementations and
b) emit an 'fd' event when file descriptors are received over a UNIX pipe,
as finally as a tangential benefit
c) remove a bunch of conditionals from the primary codepaths for
ease-of-reading.
There is a difference between errors which happen to a socket - like
receiving EPIPE - an exceptional situation but ultimately okay and the
situation where code throws in a callback - which is not okay.
Fixes test/simple/test-http-exceptions.js
TODO: explain this in docs.
- setTimeout should active the timeout too. (test-net-set-timeout tests
this.)
- 'timeout' event is not automatically followed by an 'error' event. That
is the user is now responsible for destroying the stream if there is an
idle timeout.
When making a TCP connection, readyState returns 'opening' while resolving
the host. However between the resolving period and the establishing a
connection period, it would return 'closed'. This fixes it.
This change also ensures that the socket is closed before the 'end' event is
emitted in the case that the socket was previously shutdown.
