diff -r 6eca3ecc3deb src/sys/threadcomm/interface/dlregisthreadcomm.c
--- a/src/sys/threadcomm/interface/dlregisthreadcomm.c	Mon Jul 16 18:15:42 2012 -0500
+++ b/src/sys/threadcomm/interface/dlregisthreadcomm.c	Tue Jul 17 03:07:30 2012 -0500
@@ -2,6 +2,7 @@
 
 static PetscBool PetscThreadCommPackageInitialized = PETSC_FALSE;
 
+PETSC_EXTERN PetscErrorCode PetscThreadCommDetach(MPI_Comm);
 #undef __FUNCT__
 #define __FUNCT__ "PetscThreadCommFinalizePackage"
 /*@C
@@ -16,18 +17,12 @@
 PetscErrorCode PetscThreadCommFinalizePackage(void)
 {
   PetscErrorCode ierr;
-  MPI_Comm comm;
 
   PetscFunctionBegin;
   ierr = PetscThreadCommRegisterDestroy();CHKERRQ(ierr);
 
-#if !defined(PETSC_HAVE_MPIUNI)
-  comm = PETSC_COMM_WORLD;      /* Release reference from PetscThreadCommInitialize */
-  ierr = PetscCommDestroy(&comm);CHKERRQ(ierr);
-#endif
-
-  comm = PETSC_COMM_SELF;       /* Release reference from PetscThreadCommInitialize */
-  ierr = PetscCommDestroy(&comm);CHKERRQ(ierr);
+  ierr = PetscThreadCommDetach(PETSC_COMM_WORLD);CHKERRQ(ierr);
+  ierr = PetscThreadCommDetach(PETSC_COMM_SELF);CHKERRQ(ierr);
 
   ierr = MPI_Keyval_free(&Petsc_ThreadComm_keyval);CHKERRQ(ierr);
   PetscThreadCommPackageInitialized = PETSC_FALSE;
diff -r 6eca3ecc3deb src/sys/threadcomm/interface/threadcomm.c
--- a/src/sys/threadcomm/interface/threadcomm.c	Mon Jul 16 18:15:42 2012 -0500
+++ b/src/sys/threadcomm/interface/threadcomm.c	Tue Jul 17 03:07:30 2012 -0500
@@ -116,7 +116,7 @@
   *tcomm = PETSC_NULL;
 
   ierr = PetscNew(struct _p_PetscThreadComm,&tcommout);CHKERRQ(ierr);
-  tcommout->refct = 1;
+  tcommout->refct = 0;
   tcommout->nworkThreads =  -1;
   tcommout->affinities = PETSC_NULL;
   ierr = PetscNew(struct _PetscThreadCommOps,&tcommout->ops);CHKERRQ(ierr);
@@ -139,33 +139,34 @@
 
 .seealso: PetscThreadCommCreate()
 */
-PetscErrorCode PetscThreadCommDestroy(PetscThreadComm tcomm)
+PetscErrorCode PetscThreadCommDestroy(PetscThreadComm *tcomm)
 {
   PetscErrorCode ierr;
   PetscInt       i;
 
   PetscFunctionBegin;
 
-  if (!tcomm || --tcomm->refct > 0) PetscFunctionReturn(0);
-
-  /* Destroy the implementation specific data struct */
-  if(tcomm->ops->destroy) {
-    (*tcomm->ops->destroy)(tcomm);
-  } 
-
-  ierr = PetscFree(tcomm->affinities);CHKERRQ(ierr);
-  ierr = PetscFree(tcomm->ops);CHKERRQ(ierr);
-  for(i=0;i<PETSC_KERNELS_MAX;i++) {
+  if (!*tcomm) PetscFunctionReturn(0);
+  if(!--(*tcomm)->refct) {
+    /* Destroy the implementation specific data struct */
+    if((*tcomm)->ops->destroy) {
+      (*(*tcomm)->ops->destroy)(*tcomm);
+    } 
+    ierr = PetscFree((*tcomm)->affinities);CHKERRQ(ierr);
+    ierr = PetscFree((*tcomm)->ops);CHKERRQ(ierr);
+    for(i=0;i<PETSC_KERNELS_MAX;i++) {
 #if defined(PETSC_HAVE_MEMALIGN)
-    free(PetscJobQueue->jobs[i]->job_status);
+      free(PetscJobQueue->jobs[i]->job_status);
 #else
-    ierr = PetscFree(PetscJobQueue->jobs[i]->job_status);CHKERRQ(ierr);
+      ierr = PetscFree(PetscJobQueue->jobs[i]->job_status);CHKERRQ(ierr);
 #endif
-    ierr = PetscFree(PetscJobQueue->jobs[i]);CHKERRQ(ierr);
+      ierr = PetscFree(PetscJobQueue->jobs[i]);CHKERRQ(ierr);
+    }
+    ierr = PetscFree(PetscJobQueue);CHKERRQ(ierr);
+    ierr = PetscThreadCommReductionDestroy((*tcomm)->red);CHKERRQ(ierr);
+    ierr = PetscFree((*tcomm));CHKERRQ(ierr);
   }
-  ierr = PetscFree(PetscJobQueue);CHKERRQ(ierr);
-  ierr = PetscThreadCommReductionDestroy(tcomm->red);CHKERRQ(ierr);
-  ierr = PetscFree(tcomm);CHKERRQ(ierr);
+  *tcomm = PETSC_NULL;
   PetscFunctionReturn(0);
 }
 
@@ -646,6 +647,7 @@
   PetscFunctionReturn(0);
 }
 
+
 EXTERN_C_BEGIN
 #undef __FUNCT__
 #define __FUNCT__ "Petsc_DelThreadComm"
@@ -665,7 +667,7 @@
   PetscFunctionBegin;
   ierr = MPI_Attr_get(comm,keyval,(PetscThreadComm*)&tcomm,&flg);CHKERRQ(ierr);
   if(flg) {
-    ierr = PetscThreadCommDestroy((PetscThreadComm)tcomm);CHKERRQ(ierr);
+    ierr = PetscThreadCommDestroy((PetscThreadComm*)&tcomm);CHKERRQ(ierr);
     ierr = PetscInfo1(0,"Deleting thread communicator data in an MPI_Comm %ld\n",(long)comm);if (ierr) PetscFunctionReturn((PetscMPIInt)ierr);
   }
   PetscFunctionReturn(0);
@@ -673,6 +675,60 @@
 EXTERN_C_END
 
 #undef __FUNCT__
+#define __FUNCT__ "PetscThreadCommDetach" 
+PetscErrorCode PetscThreadCommDetach(MPI_Comm comm)
+{
+  PetscErrorCode ierr;
+  MPI_Comm       icomm=comm;
+  PetscMPIInt    flg;
+  void           *ptr;
+
+  PetscFunctionBegin;
+  ierr = MPI_Attr_get(comm,Petsc_InnerComm_keyval,&ptr,&flg);CHKERRQ(ierr);
+  if(flg) {
+    ierr = PetscMemcpy(&icomm,&ptr,sizeof(MPI_Comm));CHKERRQ(ierr);
+    ierr = MPI_Attr_get(icomm,Petsc_ThreadComm_keyval,&ptr,&flg);CHKERRQ(ierr);
+    if(flg) {
+      ierr = PetscThreadCommDestroy((PetscThreadComm*)&ptr);CHKERRQ(ierr);
+      ierr = MPI_Attr_delete(icomm,Petsc_ThreadComm_keyval);CHKERRQ(ierr);
+      /* Release reference from PetscThreadCommInitialize */
+      ierr = PetscCommDestroy(&comm);CHKERRQ(ierr);
+    }
+  }
+  PetscFunctionReturn(0);
+}
+
+#undef __FUNCT__
+#define __FUNCT__ "PetscThreadCommAttach"
+PetscErrorCode PetscThreadCommAttach(MPI_Comm comm,PetscThreadComm tcomm)
+{
+  PetscErrorCode ierr;
+  MPI_Comm       icomm;
+  PetscMPIInt    flg,flg1,flg2;
+  void           *ptr;
+  PetscCommCounter *counter;
+
+  PetscFunctionBegin;
+  ierr = MPI_Attr_get(comm,Petsc_Counter_keyval,&counter,&flg);CHKERRQ(ierr);
+  if(!flg) { /* Communicator not initialized yet */
+    ierr = PetscCommDuplicate(comm,&icomm,PETSC_NULL);CHKERRQ(ierr);
+    tcomm->refct++;
+    ierr = MPI_Attr_put(icomm,Petsc_ThreadComm_keyval,tcomm);CHKERRQ(ierr);
+  } else {
+    ierr = MPI_Attr_get(comm,Petsc_InnerComm_keyval,&ptr,&flg1);CHKERRQ(ierr);
+    if(flg1) {
+      ierr = PetscMemcpy(&ptr,&icomm,sizeof(MPI_Comm));CHKERRQ(ierr);
+      ierr = MPI_Attr_get(icomm,Petsc_ThreadComm_keyval,&ptr,&flg2);CHKERRQ(ierr);
+      if(!flg2) {
+	tcomm->refct++;
+	ierr = MPI_Attr_put(icomm,Petsc_ThreadComm_keyval,tcomm);CHKERRQ(ierr);
+      }
+    }
+  }
+  PetscFunctionReturn(0);
+}
+
+#undef __FUNCT__
 #define __FUNCT__ "PetscThreadCommInitialize"
 /*
   PetscThreadCommInitialize - Initializes the thread communicator object 
@@ -684,7 +740,6 @@
 {
   PetscErrorCode  ierr;
   PetscThreadComm tcomm;
-  MPI_Comm        icomm;
   PetscInt        i,j;
 
   PetscFunctionBegin;
@@ -708,17 +763,9 @@
   PetscJobQueue->kernel_ctr  = 0;
   tcomm->job_ctr     = 0;
 
-  /* PETSC_COMM_SELF = PETSC_COMM_WORLD for MPIUNI */
-#if !defined(PETSC_HAVE_MPIUNI)
-  ierr = PetscCommDuplicate(PETSC_COMM_WORLD,&icomm,PETSC_NULL);CHKERRQ(ierr);
-  ierr = MPI_Attr_put(icomm,Petsc_ThreadComm_keyval,(void*)tcomm);CHKERRQ(ierr);
-  tcomm->refct++;               /* Share the threadcomm with PETSC_COMM_SELF */
-#endif
-
-  ierr = PetscCommDuplicate(PETSC_COMM_SELF,&icomm,PETSC_NULL);CHKERRQ(ierr);
-  ierr = MPI_Attr_put(icomm,Petsc_ThreadComm_keyval,(void*)tcomm);CHKERRQ(ierr);
-
   /* This routine leaves extra references to the inner comms. They are released in PetscThreadCommFinalizePackage(). */
+  ierr = PetscThreadCommAttach(PETSC_COMM_WORLD,tcomm);CHKERRQ(ierr);
+  ierr = PetscThreadCommAttach(PETSC_COMM_SELF,tcomm);CHKERRQ(ierr);
 
   ierr = PetscThreadCommSetType(tcomm,NOTHREAD);CHKERRQ(ierr);
   ierr = PetscThreadCommReductionCreate(tcomm,&tcomm->red);CHKERRQ(ierr);