This change also tries to optimize Emit by looping through the listeners in
C++. The javascript version of this function is still there and being used,
but only by javascript code. Not an ideal solution - there should only be
one implementation - however for now it seems to help.
This doesn't solve all of the memory leaks that we're experiencing, there
seems to be another subtle problem.
Solution is to manually add Attach() to OnConnection.
For client side it seems there is no Detach() being called after NS
resolution? Otherwise I would have removed it. That was another bug.
Note: We don't want to modify evnet's behavior to have on_connect called
directly when the socket is accepted. evnet needs to support SSL, and
on_connect is supposed to signal that the SSL connection is established. The
point here is that being "connected" and being "attached" to the event loop
are two different things. SSL stuff may be transmitted when a socket is not
"connected" but it must always be attached.
The problem was that Connection::on_close was calling Detach() directly
after executing the "disconnect" event. Since we had a boolean attach count,
this was leaving sockets detached even if they had reattached in during the
event.
* Added many asserts in http.cc and net.cc to ensure that sockets are
connected when they should be.
* Changed ObjectWrap to use a reference count instead of boolean attached_
value.
* Fixed similar bug in Timer.
Simplify and inline ObjectWrap::Wrap
Inline and clean up ObjectWrap::Unwrap
Move ObjectWrap into its own file.
Remove handle from ObjectWrap constructor. add obj->Wrap(handle)
Simplify Attach/Detach in ObjectWrap
Remove ObjectWrap::InformV8ofAllocation. (Too messy/complex.)
This is a rather large refactor! Mostly for the better side. I've had to
remove some functionality like req.interrupt(). A lot of other work is left
messy or incomplete.
The constructor for TCP servers can no longer take a connection handler for
purely technical reasons. (The constructor for EventEmitter is implemented
in C++ but addListener is in javascript, and I don't want to make too many
C++ -> Javascript references.) Thus I introduce new constructor methods to
ease the creation of the servers:
node.tcp.createServer()
node.http.createServer()
These work almost the same as the old constructors.
In general we're working towards a future where no constructors are
publicly exposed or take arguments.
The HTTP events like "on_uri" are not yet using the event interface.
onMessage still is a constructor - but this will change soon.
This requires that onExit() is not called immediately upon receiving a
SIGCHLD. There could still be data in the pipez. So, instead just set a
flag and invoke the pipe watchers.
Sometimes one will not receive an EOF from pipes because the process was
killed by a SIGTERM, or something. If SIGCHLD has been recved but we are
getting EAGAIN, the pipez need to be closed too.
Tests were silently failing since f56309d...
Since the eio_watcher is now not counted in the list of active watchers, we
need to explicitly tell the ev_loop not to exit when entering the thread
pool.
onExit() is similar to the onLoad() callback. onExit() is called on each
module just before the process exits. This can be used to check state in
unit tests, but not to perform I/O. The process will forcibly exit as soon
as all of the onExit callbacks are made.
There is one major API change in the refactor: filename extensions are now
required when requiring or including modules.
Added extra test to test-module-loading.js.
If users do not send transfer-encoding or content-length headers, then I
will not add any additional. Content-Length: 0 is assumed if there aren't
other headers and chunked encoding is rare.
All the c++ code is now reduced to simple wrappers. The node.fs.File object
is defined entirely in javascript now. As is the actionQueue methods.
This makes the boundaries much cleaner. There is still some thought that
needs to go into how exactly the API should behave but this simplification
is a first step.
The error codes still remain for the two general file system operations:
rename and stat.
Additionally I've removed the actionQueue for file system operations. They
are sent directly into the thread pool.
Instead we're going to just get a single callback for the URI. This can be
parsed additionally in javascript using parseuri:
http://blog.stevenlevithan.com/archives/parseuri
I haven't added that yet, but it will come soon.
Encoding UTF-16 (the native string representation) to UTF-8 is rather
expensive, however just chopping off the second bit to convert UTF-16 to
ASCII is rather fast. I've noticed major performance issues with
String::WriteUtf8 and thus I'm going to explicitly separate in the API.
Still need interfaces to this for the web server.
This is sloppy: after each ObjectWrap allocation the user needs to
call ObjectWrap::InformV8ofAllocation(). In addition each class deriving
from ObjectWrap needs to implement the virtual method size() which should
return the size of the derived class. If I was better at C++ I could
possibly make this less ugly. For now this is how it is.
Memory usage looks much better after this commit.
Instead servers are passed a function which gets called on connection (like
in the original design) which has one argument, the connecting socket. The
user sets up callbacks on that. It's pretty much how I had it originally.
Encoding is now set via v8 getter/setter and can be changed dynamically.
The timeout for all sockets is fixed at 60 seconds for now. Need to fix
that.
The LowLevelServer is a direct interface to the parser given people access
to things like partially received headers. This could be used to implement
an extremely optimized server which acts before parsing is complete.
Most people will be using node.http.Server which is still rather low-level
compared to other http interfaces, but does take care of some details for
you.
Connections were being garbage collected while they were still in progress
since the object would leave scope. This commit adds ObjectWrap::Attach()
and ObjectWrap::Detach() to tell v8 that an object is currently on the event
loop and will be needed in the future.
Other changes to oi_socket.c and net.cc are to fix bugs encountered while
running the HTTP server.
For server-side sockets, no longer pass the server object to the
js constructor. This is set later with SetAcceptor.
I think the change is a bit strage and convoluted but it allows one give
protocol /classes/ to the c++ constructors instead of protocol instances.
This is nice because derived classes (like HTTP) don't need to copy the
protocol instanciation code.
Here I massively change both the external and internal API of the TCP
sockets and servers.
This change introduces the concept of a protocol object like is found in
Twisted Python. I believe this allows for a much cleaner description of how
a socket behaves. What was once a single object "client" or "connection" is
now represented by two objects: a "connection" and a "protocol".
Well - I don't want to ramble too much because neither API is yet public or
documented. Look the diff of test/test-pingpong.js to see how things have
changed.
Also changed Init_net() to Socket::Initialize() and Server::Initialize().
Seems more object oriented, but I'm unsure how this will play when I want
load modules dynamically with dlopen(). I'll sit with it for a while and
see how it feels.
This might need to be changed in the future if ev is needed in thread pools
or extension libraries. However for now it makes sense to just use a single
loop.
