net/http: fix Transport race returning bodyless responses and reusing conns
The Transport had a delicate protocol between its readLoop goroutine
and the goroutine calling RoundTrip. The basic concern is that the
caller's RoundTrip goroutine wants to wait for either a
connection-level error (the conn dying) or the response. But sometimes
both happen: there's a valid response (without a body), but the conn
is also going away. Both goroutines' logic dealing with this had grown
large and complicated with hard-to-follow comments over the years.
Simplify and document. Pull some bits into functions and do all
bodyless stuff in one place (it's special enough), rather than having
a bunch of conditionals scattered everywhere. One test is no longer
even applicable since the race it tested is no longer possible (the
code doesn't exist).
The bug that this fixes is that when the Transport reads a bodyless
response from a server, it was returning that response before
returning the persistent connection to the idle pool. As a result,
~1/1000 of serial requests would end up creating a new connection
rather than re-using the just-used connection due to goroutine
scheduling chance. Instead, this now adds bodyless responses'
connections back to the idle pool first, then sends the response to
the RoundTrip goroutine, but making sure that the RoundTrip goroutine
is outside of its select on the connection dying.
There's a new buffered channel involved now, which is a minor
complication, but it's much more self-contained and well-documented
than the previous complexity. (The alternative of making the
responseAndError channel itself unbuffered is too invasive and risky
at this point; it would require a number of changes to avoid
deadlocked goroutines in error cases)
In any case, flakes look to be gone now. We'll see if trybots agree.