Don't use the relaxed accept() algorithm introduced in be2a217 unless
explicitly requested. It causes a 50+% performance drop on some node.js
benchmarks:
$ alias bench='out/Release/node benchmark/http_simple_auto.js \
-c 10 -n 50000 bytes/1 2>&1 | grep Req'
$ UV_TCP_SINGLE_ACCEPT=0 bench
Requests per second: 12331.84 [#/sec] (mean)
$ UV_TCP_SINGLE_ACCEPT=1 bench
Requests per second: 3944.63 [#/sec] (mean)
kqueue(2) on osx doesn't work (emits EINVAL error) with specific fds
(i.e. /dev/tty, /dev/null, etc). When given such descriptors - start
select(2) watcher thread that will emit io events.
Avoid the extra syscall, it's a no-op for non-listening sockets.
At least, it should be - it remains to be investigated if a FreeBSD kernel bug
affects ephemeral port allocation inside connect(). See [1] for details.
[1] http://www.freebsd.org/cgi/query-pr.cgi?pr=174087
Benchmarks demonstrated that the idle timer handle approach didn't balance the
load quite fair enough, the majority of new connections still ended up in one
or two processes.
The new approach voluntarily gives up a scheduler timeslice by calling
nanosleep() with a one nanosecond timeout.
Why not sched_yield()? Because on Linux (and this is probably true for other
Unices as well), sched_yield() only yields if there are other processes running
on the same CPU.
nanosleep() on the other hand always forces the process to sleep, which gives
other processes a chance to accept our pending connections.
Don't make the event loop spin when connect() returns EINPROGRESS.
Test case:
#include "uv.h"
static void connect_cb(uv_connect_t* req, int status) {
// ...
}
int main() {
uv_tcp_t handle;
uv_connect_t req;
struct sockaddr_in addr;
addr = uv_ip4_addr("8.8.8.8", 1234); // unreachable
uv_tcp_init(uv_default_loop(), &handle);
uv_tcp_connect(&req, (uv_stream_t*)&handle, addr, connect_cb);
uv_run(uv_default_loop()); // busy loops until connection times out
return 0;
}
After EINPROGRESS, there are zero active handles and one active request. That
in turn makes uv__poll_timeout() believe that it should merely poll, not block,
in epoll() / kqueue() / port_getn().
Sidestep that by artificially starting the handle on connect() and stopping it
again once the TCP handshake completes / is rejected / times out.
It's a slightly hacky approach because I don't want to change the ABI of the
stable branch. I'll address it properly in the master branch.
Incidentally fixes a rather obscure bug where uv_tcp_connect() reconnected
and leaked a file descriptor when the handle was already busy connecting,
handle->fd was zero (unlikely) and uv_tcp_connect() got called again.
