Treat all unexpected socket errors as indicating close.

webrtc/base/physicalsocketserver.cc

Index: webrtc/base/physicalsocketserver.cc
diff --git a/webrtc/base/physicalsocketserver.cc b/webrtc/base/physicalsocketserver.cc
index e2ec589eb12d0229febab295fe420472957f3a9b..5acc94db66fc2119133babae7f35d964758d0f3a 100644
--- a/webrtc/base/physicalsocketserver.cc
+++ b/webrtc/base/physicalsocketserver.cc
@@ -706,18 +706,14 @@ bool SocketDispatcher::IsDescriptorClosed() {
       // Returned during ungraceful peer shutdown.
       case ECONNRESET:
         return true;
-      default:
-        // Assume that all other errors are just blocking errors, meaning the
-        // connection is still good but we just can't read from it right now.
-        // This should only happen when connecting (and at most once), because
-        // in all other cases this function is only called if the file
-        // descriptor is already known to be in the readable state. However,
-        // it's not necessary a problem if we spuriously interpret a
-        // "connection lost"-type error as a blocking error, because typically
-        // the next recv() will get EOF, so we'll still eventually notice that
-        // the socket is closed.
-        LOG_ERR(LS_WARNING) << "Assuming benign blocking error";
+      // This should only happen when connecting (and at most once), because
+      // in all other cases this function is only called if the file
+      // descriptor is already known to be in the readable state.
+      case EWOULDBLOCK:
         return false;
+      default:
+        LOG_ERR(LS_WARNING) << "Unexpected error from recv";
+        return true;

[Taylor Brandstetter, 2017/02/09 08:59:47]

I don't think this is safe. There are errors besides EWOULDBLOCK on different
platforms that actually are benign blocking errors, in my experience. So the
default should be to assume it's a benign error.

     }
   }
 }