commit ed3b986d63294d732cabe689e049b6e76d312980
author deadbeef <deadbeef@webrtc.org> Fri Jun 02 10:33:16 2017 -0700
committer Commit Bot <commit-bot@chromium.org> Fri Jun 02 17:33:16 2017 +0000
tree 86a1408b68abe7c328f1e9467f1158ea25d46b38
parent 25fd62bcc00dd4dbf725963ca861d9c70e7b2170

Fixing SSL error that occurs when underlying socket is blocked.

BoringSSL (or OpenSSL) require that when SSL_write fails due to the
underlying socket being blocked, it's retried with the same parameters
until it succeeds. But we weren't doing this, and our socket
abstraction doesn't have an equivalent requirement. So when this was
occurring, we would just end up trying to send the next RTP or STUN
packet (instead of the packet that couldn't be sent), and BoringSSL
doesn't like that.

So, when this condition occurs now, we'll simply enter a "pending write"
mode and buffer the data that couldn't be completely sent. When the
underlying socket becomes writable again, or if Send is called again
before that happens, we retry sending the buffered data. Making both
BoringSSL and the upper layer of code that expects normal TCP socket
behavior happy.

Also adding some more logging, and fixing an issue with VirtualSocketServer
that made it behave slightly differently than PhysicalSocketServer when a
TCP packet is only partially read.

BUG=webrtc:7753

Review-Url: https://codereview.webrtc.org/2915243002
Cr-Commit-Position: refs/heads/master@{#18416}