Changes since version 0.10.27:
* windows: fix handling closed socket while poll handle is closing (Saúl
Ibarra Corretgé)
* unix: return system error on EAI_SYSTEM (Saúl Ibarra Corretgé)
* unix: fix bogus structure field name (Saúl Ibarra Corretgé)
* darwin: invoke `mach_timebase_info` only once (Fedor Indutny)
This is the libuv side of the fix for Node's cluster module on Windows.
https://github.com/joyent/node/issues/7691
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.
If libuv is loaded as a DLL and is later unloaded deadlocks can happen
when running atexit handlers, so we can't use synchronization
priomitives or join threads there.
For reference see https://github.com/saghul/pyuv/issues/171
After 41891222bc landed it's possible that uv__udp_sendmsg is called
even if there are no pending write nor write completed requests:
1. User calls uv_udp_send and the request is sent immediately. The
request is the added to the completed queue and we 'feed' the uv__io
handle so that we process the completed request in the next
iteration.
2. User calls uv_udp_send again but the request is not completed
immediately, so it's queued in the write_queue.
3. The uv__io handle gets a UV__POLLOUT event and uv__udp_sendmsg is
run, which completes the send request and puts it in the
write_completed_queue. Afterwards, uv__udp_run_completed is executed
and the write_completed queue is drained.
4. At this point, the uv__io handle was made pending in step 3, in
uv__udp_sendmsg, but we no longer have requests to write or to complete,
so we skip processing.
It's possible that recv_cb_called is bigger than the number of sockets,
because if all sends succeed the recv callback is called twice: once
with the actual data, and another time with 0.
This makes libuv more tolerant to the properties of the pipes
that it can use without any issue. This is necessary because Cygwin
(and hence Mintty) opens STDIN without FILE_WRITE_ATTRIBUTES.
This functionality is present in stream and uv_udp_t has a queue
as well so it makes sense for udp to have a send_write_size.
Since udp sends entire messages atomically, the send_queue_count field
lets the user determine how many messages are there left to send.
Debug build failed on mingw32 because CRT assertion disable code was stubbed out.
Replace __declspec(thread) with UV_THREAD_LOCAL which is defined as __thread on GCC.
According to @aktau, the call to `mach_timebase_info` costs 90% of the
total execution time of `uv_hrtime()`. The result of the call is static
on all existing platforms, so there is no need in invoking it multiple
times.
Signed-off-by: Fedor Indutny <fedor@indutny.com>
