HTTP agent: remove excess calls to removeSocket

[davidvgalbraith]

socket.destroy() triggers a 'close' event from the socket
which triggers the onClose handler of HTTPAgent which calls
self.removeSocket(). So by calling self.removeSocket() prior
to socket.destroy() we end up with two calls to self.removeSocket().

If there are pending requests, removeSocket ends up creating
a new socket. So if there are pending requests, each time
a request completes, we tear down one socket and create two
more. So the total number of sockets grows exponentially and
without regard for any maxSockets settings. This was noticed in
#4050. Let's get rid of the extra calls to removeSocket

so we only call it once per completed request.

[mscdex]

CI was all green except unrelated, flaky tests on Windows and pi1-raspbian-wheezy had this test failure:

not ok 46 test-http-agent-maxsockets-regress-4050.js
# events.js:143
# throw er; // Unhandled 'error' event
# ^
# # Error: socket hang up
# at createHangUpError (_http_client.js:210:15)
# at Socket.socketCloseListener (_http_client.js:242:23)
# at emitOne (events.js:84:20)
# at Socket.emit (events.js:171:7)
# at TCP._onclose (net.js:470:12)

I'm not sure if it's a flaky test or what?

[davidvgalbraith]

Found it -- calling abort() on a request before it has triggered the response event gives you a socket hangup. Replacing that flimsy timeout with a response event handler ought to make it bulletproof.

[davidvgalbraith]

Hey! Been a few days, anything else I can do to give this one a kick?

[mscdex]

CI again: https://ci.nodejs.org/job/node-test-pull-request/974/

[mscdex]

LGTM, the couple of tests that did fail are known to be flaky.

[jasnell]

@nodejs/http

[jasnell]

LGTM but would like additional review from @nodejs/http

[indutny]

I see one possible problem here. close event is emitted on a next libuv tick. But considering that the .sockets is used just for bookkeeping, it should be safe.

LGTM

[dougwilson]

LGTM

[davidvgalbraith]

Universal acclaim! Let's get this in?

[davidvgalbraith]

Bump! (hope it's not too pestilential to bump)

[indutny]

@davidvgalbraith I would like to land this patch, but may I ask you to review this contributing guidelines first: https://github.com/nodejs/node/blob/master/CONTRIBUTING.md#step-3-commit, and make sure that the commit log matches them?

Namely there are at least two issues with the messages right now:

• Subsystem. There is no subsystem called http agent, please use http
• Line length. Commit log should consist of both title and description, both wrapped on appropriate line limits as per https://github.com/nodejs/node/blob/master/CONTRIBUTING.md#step-3-commit

Cheers!

[davidvgalbraith]

Ok, I changed the subsystem in the title to HTTP and made the body lines ~72 characters each.

[indutny]

Landed in 6e11e22, thank you!