net/http: finishRequest() aborts ongoing background reads

[JayNakrani]

[ continued from Issue #17547 ]

TL;DR: abortPendingRead() is called from finishRequest(), causing it to abort ongoing reads. This causes deadline_exceeded errors. finishRequest() should wait for reads to finish in normal case.

Consider following sequence of events from http.conn.serve():

1. startBackgroundRead() starts backgroundRead() in another goroutine.
2. Assume that backgroundRead() from above step does not finish until after the next step. (Simulated by adding sleep)
3. finishRequest() called, which calls abortPendingRead(). abortPendingRead() then proceeds to abort ongoing background reads by setting the read deadline to aLongTimeAgo. This causes the ongoing reads to report deadline_exceeded.

/cc @bradfitz @dsnet @mdempsky

[bradfitz]

I haven't looked into this yet. To clarify, are you saying this is a regression since faf882d (golang.org/cl/31173)?

[JayNakrani]

Regression: Yes. I think so.
I have a CL that should fix this. Sending in a bit.

[gopherbot]

CL https://golang.org/cl/31755 mentions this issue.

[JayNakrani]

From offline discussion with Brad & Joe: It's okay for finishRequest() to abort backgroundRead() every time. The imeplementation of net.Conn should respect this semantics.

This is WAI. So closing this.