1094522 - Investigate why HTTP/2 test runs over 2 different connections

Status: RESOLVED FIXED

(Core :: Networking: HTTP, defect)

Description

[u408661]

I noticed this while debugging my tests for bug 1091263. As it turns out, we end up with two different HTTP/2 sessions when running test_http2.js. The first connection is used by the first test, and the second connection is used by all subsequent tests (unless there's a test that explicitly causes a new connection). This happens regardless of which test is first, and regardless of how many tests are in the suite. This was really annoying when I was running just a new subset of the tests that were supposed to share a connection, and they didn't. Once I re-enabled the rest of the tests (that were run before my new tests), everything was fine.

This isn't the end of the world, but we should see if we can figure out why this happens, and also if it ever happens in the real world.

Comment 1

[Patrick McManus [:mcmanus]]

I think I diagnosed this on the test_spdy side, which sees a similar thing.

A sepculative connect starts a connection, and the tcp handshake completes before the connection manager has a transaction for it yet. (in this case that's cause the connect is to localhost - so its wicked fast). So the connection manager drives the ssl handshake with a null transaciton.

but before the null transaction has completed the http request arrives in the connection manager and that starts a new connection up because there are no idle or active-spdy connections.

give it a few milliseconds and you've got 2 sessions.

There actually is code in makenewconnections() that looks for a sepculative half open connection.. and if it finds such a thing it claims it for itself rather than making a new connection. I think the issue can be fixed by extending that logic to active null transactions as well.

Comment 2

[Patrick McManus [:mcmanus]]

https://treeherder.mozilla.org/#/jobs?repo=try&revision=dbf8c15ff4ba

Comment 3

[Patrick McManus [:mcmanus]]

Attachment: claim conns with null transactions instead of starting new ones

Comment 4

[Patrick McManus [:mcmanus]]

https://hg.mozilla.org/integration/mozilla-inbound/rev/050e9b432799

Comment 5

[Patrick McManus [:mcmanus]]

Attachment: only claim null transactions for your connection if KA eligible

Comment 6

[u408661]

remote:   https://hg.mozilla.org/integration/mozilla-inbound/rev/c4cccb7acc1b

Comment 7

[Wes Kocher (:KWierso) (Not reading bugmail; email directly if needed)]

https://hg.mozilla.org/mozilla-central/rev/050e9b432799
https://hg.mozilla.org/mozilla-central/rev/c4cccb7acc1b