net/http: tighten protocol between Transport.roundTrip and persistConn.readLoop
In debugging the flaky test in #13825, I discovered that my previous
change to tighten and simplify the communication protocol between
Transport.roundTrip and persistConn.readLoop in
https://golang.org/cl/17890 wasn't complete.
This change simplifies it further: the buffered-vs-unbuffered
complexity goes away, and we no longer need to re-try channel reads in
the select case. It was trying to prioritize channels in the case that
two were readable in the select. (it was only failing in the race builder
because the race builds randomize select scheduling)
The problem was that in the bodyless response case we had to return
the idle connection before replying to roundTrip. But putIdleConn
previously both added it to the free list (which we wanted), but also
closed the connection, which made the caller goroutine
(Transport.roundTrip) have two readable cases: pc.closech, and the
response. We guarded against similar conditions in the caller's select
for two readable channels, but such a fix wasn't possible here, and would
be overly complicated.
Instead, switch to unbuffered channels. The unbuffered channels were only
to prevent goroutine leaks, so address that differently: add a "callerGone"
channel closed by the caller on exit, and select on that during any unbuffered
sends.
As part of the fix, split putIdleConn into two halves: a part that
just returns to the freelist, and a part that also closes. Update the
four callers to the variants each wanted.
Incidentally, the connections were closing on return to the pool due
to MaxIdleConnsPerHost (somewhat related: #13801), but this bug
could've manifested for plenty of other reasons.