Do not send empty TCP packet in _http_outgoing.js

[indutny]

After 99943e1 we send empty TCP packets on res.end() in quite many cases. I don't think that it is a serious thing, but it is definitely worth fixing.

[indutny]

See #3803 for reference

[jasnell]

+1 I'd noticed that before too but hadn't managed to get around to fixing it.
/cc @nodejs/http

[indutny]

Yeah, it is kind of a way to emit finish event only after completion of all previous writes.

[trevnorris]

Is this a feature request or a bug?

[indutny]

More like a feature request with the elements of bug fixing.

[Trott]

Is there a way to reproduce this not-exactly-a-bug reliably?

[indutny]

I think yes.

[Trott]

Probably don't want to close this, right? But no one's actively working on it either? Is someone available to mentor an interested potential-contributor who might want to try to implement this?

[bnoordhuis]

Has this been fixed? I added a judicious assert (see diff) and I can't reproduce.

diff --git a/deps/uv/src/unix/stream.c b/deps/uv/src/unix/stream.c
index 7b23d16..80805e5 100644
--- a/deps/uv/src/unix/stream.c
+++ b/deps/uv/src/unix/stream.c
@@ -1444,6 +1444,7 @@ int uv_write2(uv_write_t* req,
   memcpy(req->bufs, bufs, nbufs * sizeof(bufs[0]));
   req->nbufs = nbufs;
   req->write_index = 0;
+  assert(0 != uv__count_bufs(bufs, nbufs));
   stream->write_queue_size += uv__count_bufs(bufs, nbufs);
 
   /* Append the request to write_queue. */

(It does break some REPL and child process tests but that's expected. The salient part is that no zero-sized TCP writes seem to take place.)

[apapirovski]

This seems to be addressed from what I can tell. Please feel free to re-open if you think I'm wrong or this needs to remain open to tackle some edge cases.