tree abe9305227662edad00cb78d44258d3d7c42da4b
parent f1c79f8dac61141a518b87361cfef0619e3ef49e
author Bernie Innocenti <codewiz@google.com> 1526298049 +0900
committer Bernie Innocenti <codewiz@google.com> 1526384088 +0900

Avoid spinning endlessly on packet socket errors

In the event the packet socket has an error, it is likely not cleared
and clatd can spin in a tight loop. This may happen when a network is
going down (ENETDOWN, ENETUNREACH, ...). Ordinarily the
ConnectivityService/Nat464Xlat/ClatdController control plane would
tear down clatd in this case.  However, if the control plane is
deadlocked, clatd will chew up CPU indefinitely.

This fix consists of detecting when poll() exits due to a socket error
and trying to clear the error by issuing a no-op recv() on the packet
socket and a 0-byte read() for the /dev/tun socket.

Test: manually, on a marlyn device (kernel 3.18.70)
Bug: 78602493
Change-Id: Ic23f999712a674df11e981a4314ad371e8d5fb6d