[Libreoffice-commits] online.git: wsd/LOOLWSD.cpp

[Ashod Nakashian]

 wsd/LOOLWSD.cpp |   20 ++++++++++++++++++++
 1 file changed, 20 insertions(+)

New commits:
commit 2da90d0cfc4e73490e323cb3ab00f66dd4d25ab0
Author: Ashod Nakashian <ashod.nakashian at collabora.co.uk>
Date:   Sat Feb 11 18:03:14 2017 -0500

    wsd: explain the root of the document bifurcation bug
    
    Change-Id: I9ac290c94296ed18dc25ce3b2d9c4939e4eeea91
    Reviewed-on: https://gerrit.libreoffice.org/34192
    Reviewed-by: Ashod Nakashian <ashnakash at gmail.com>
    Tested-by: Ashod Nakashian <ashnakash at gmail.com>

diff --git a/wsd/LOOLWSD.cpp b/wsd/LOOLWSD.cpp
index d268d68..c910f98 100644
--- a/wsd/LOOLWSD.cpp
+++ b/wsd/LOOLWSD.cpp
@@ -1595,6 +1595,26 @@ inline ServerSocket* getServerSocket(int portNumber, bool reuseDetails)
         // that happens accidentally for developers only anyway, in production systemd takes care of
         // having just one, I hope.)
 
+        // Note: We really should _not_ need to do this at all. The reason we do is that,
+        // if we don't, when an instance of loolwsd is already running, a second instance
+        // would not fail to listen to the _same_ port, and start stealing connections.
+        // This would have the undesirable side-effect of bifurcating documents (meaning,
+        // clients loading a document that is already loaded in the first instance of
+        // wsd, if their socket is accepted by the second instance of wsd, will not
+        // be able to collaborate with their peers, who are in the first instance of wsd).
+        // This situation arises because Poco's Socket::bind() (when reuseAddress,
+        // the second arg, is true) enables reuse for both address _and_ port.
+        // The latter is damaging (and the root of this bug). It's designed (on Linux)
+        // to load-balance requests on a given port to all server instances.
+        // But this only works if requests are independent (as in web-page serving)
+        // and fail miserably for our purposes here (collaborative editing).
+
+        // The correct solution is to explicitly enable only address reuse
+        // before bind and pass reuseAddress=false (second arg) to bind.
+        // Reusing the address is fine, since that eliminates the wait
+        // when we recycle and the socket from the earlier run is still
+        // in TIME_WAIT.
+
         if (reuseDetails)
         {
             try