[Bigdata-commit] SF.net SVN: bigdata:[7741] branches/MGC_1_3_0/bigdata/src/java/com/bigdata/ ha/QuorumPipelineImpl.java

[<mar...@us...>]

Revision: 7741
          http://bigdata.svn.sourceforge.net/bigdata/?rev=7741&view=rev
Author:   martyncutcher
Date:     2014-01-07 11:05:13 +0000 (Tue, 07 Jan 2014)
Log Message:
-----------
Change future checks to avoid deadlock if socket send blocks.  It has not been possible to test this locally since failure has only been observed in CI on EC2

Modified Paths:
--------------
    branches/MGC_1_3_0/bigdata/src/java/com/bigdata/ha/QuorumPipelineImpl.java

Modified: branches/MGC_1_3_0/bigdata/src/java/com/bigdata/ha/QuorumPipelineImpl.java
===================================================================
--- branches/MGC_1_3_0/bigdata/src/java/com/bigdata/ha/QuorumPipelineImpl.java	2014-01-06 21:12:19 UTC (rev 7740)
+++ branches/MGC_1_3_0/bigdata/src/java/com/bigdata/ha/QuorumPipelineImpl.java	2014-01-07 11:05:13 UTC (rev 7741)
@@ -1944,8 +1944,14 @@
                          * also ignored. We want to continue this loop until
                          * both Futures are done. Interrupts are not trapped, so
                          * an interrupt will still exit the loop.
+                         * 
+                         * It appears that it is possible for futSnd to be blocked
+                         * and not generate an error. If we do not exit the loop
+                         * and check the futRec future in this case then we coul loop
+                         * continuously.  This does rather beg the question of
+                         * whether we should only be checking futRec at this stage.
                          */
-                        while (!futSnd.isDone() || !futRec.isDone()) {
+                        while (!(futSnd.isDone() || futRec.isDone())) {
                             /*
                              * Make sure leader's quorum token remains valid for
                              * ALL writes.
@@ -2491,7 +2497,7 @@
 
         @Override
         public Void call() throws Exception {
-
+        	
             // wrap the messages together.
             final HAMessageWrapper wrappedMsg = new HAMessageWrapper(
                     req, snd, msg);
@@ -2524,8 +2530,11 @@
                          * also ignored. We want to continue this loop until
                          * both Futures are done. Interrupts are not trapped, so
                          * an interrupt will still exit the loop.
+                         * 
+                         * TODO: check the comparative logic with this and robustReplicate
+                         * to confirm the equivalence of checking the different futures.
                          */
-                        while (!futRec.isDone() || !futRep.isDone()) {
+                        while (!(futRec.isDone() || futRep.isDone())) {
                             /*
                              * The token must remain valid, even if this service
                              * is not joined with the met quorum. If fact,
@@ -2547,7 +2556,7 @@
                         }
                         
                         /*
-                         * Note: Both futures are DONE at this point. However,
+                         * Note: Both futures are DONE (or not - check condition above) at this point. However,
                          * we want to check the remote Future for the downstream
                          * service first in order to accurately report the
                          * service that was the source of a pipeline replication

This was sent by the SourceForge.net collaborative development platform, the world's largest Open Source development site.