19 files changed, 144 insertions, 123 deletions

diff --git a/erts/emulator/beam/beam_emu.c b/erts/emulator/beam/beam_emu.c
index 7eecd059a5..aab0baebea 100644
--- a/erts/emulator/beam/beam_emu.c
+++ b/erts/emulator/beam/beam_emu.c
@@ -2166,7 +2166,7 @@ void process_main(void)
 
      PreFetch(0, next);
      if (IS_TRACED_FL(c_p, F_TRACE_RECEIVE)) {
-	 trace_receive(c_p, c_p, am_timeout, NULL);
+	 trace_receive(c_p, am_clock_service, am_timeout, NULL);
      }
      if (ERTS_PROC_GET_SAVED_CALLS_BUF(c_p)) {
 	 save_calls(c_p, &exp_timeout);
diff --git a/erts/emulator/beam/bif.c b/erts/emulator/beam/bif.c
index ed5b2983dd..ac77fa96e1 100644
--- a/erts/emulator/beam/bif.c
+++ b/erts/emulator/beam/bif.c
@@ -611,7 +611,7 @@ erts_queue_monitor_message(Process *p,
     ref_copy    = copy_struct(ref, ref_size, &hp, ohp);
 
     tup = TUPLE5(hp, am_DOWN, ref_copy, type, item_copy, reason_copy);
-    erts_queue_message(p, p_locksp, msgp, tup);
+    erts_queue_message(p, *p_locksp, msgp, tup, am_system);
 }
 
 static BIF_RETTYPE
diff --git a/erts/emulator/beam/dist.c b/erts/emulator/beam/dist.c
index 52fd57e101..b020f16805 100644
--- a/erts/emulator/beam/dist.c
+++ b/erts/emulator/beam/dist.c
@@ -397,7 +397,7 @@ static void doit_node_link_net_exits(ErtsLink *lnk, void *vnecp)
 	    msgp = erts_alloc_message_heap(rp, &rp_locks,
 					   3, &hp, &ohp);
 	    tup = TUPLE2(hp, am_nodedown, name);
-	    erts_queue_message(rp, &rp_locks, msgp, tup);
+	    erts_queue_message(rp, rp_locks, msgp, tup, am_system);
 	}
 	erts_smp_proc_unlock(rp, rp_locks);
     }
@@ -1456,7 +1456,7 @@ int erts_net_message(Port *prt,
 		token = copy_struct(token, token_size, &hp, ohp);
 	    }
 
-	    erts_queue_dist_message(rp, &locks, ede_copy, token);
+	    erts_queue_dist_message(rp, locks, ede_copy, token, from);
 	    if (locks)
 		erts_smp_proc_unlock(rp, locks);
 	}
@@ -1505,7 +1505,7 @@ int erts_net_message(Port *prt,
 		token = copy_struct(token, token_size, &hp, ohp);
 	    }
 
-	    erts_queue_dist_message(rp, &locks, ede_copy, token);
+	    erts_queue_dist_message(rp, locks, ede_copy, token, tuple[2]);
 	    if (locks)
 		erts_smp_proc_unlock(rp, locks);
 	}
@@ -3317,7 +3317,7 @@ send_nodes_mon_msg(Process *rp,
     }
 
     ASSERT(hend == hp);
-    erts_queue_message(rp, rp_locksp, mp, msg);
+    erts_queue_message(rp, *rp_locksp, mp, msg, am_system);
 }
 
 static void
diff --git a/erts/emulator/beam/erl_alloc.c b/erts/emulator/beam/erl_alloc.c
index d04977b9ae..3dc12bac51 100644
--- a/erts/emulator/beam/erl_alloc.c
+++ b/erts/emulator/beam/erl_alloc.c
@@ -3286,7 +3286,7 @@ reply_alloc_info(void *vair)
     if (hp != hp_end)
 	erts_shrink_message_heap(&mp, rp, hp_start, hp, hp_end, &msg, 1);
 
-    erts_queue_message(rp, &rp_locks, mp, msg);
+    erts_queue_message(rp, rp_locks, mp, msg, am_system);
 
     if (air->req_sched == sched_id)
 	rp_locks &= ~ERTS_PROC_LOCK_MAIN;
diff --git a/erts/emulator/beam/erl_bif_ddll.c b/erts/emulator/beam/erl_bif_ddll.c
index 1e2db38442..ef77201544 100644
--- a/erts/emulator/beam/erl_bif_ddll.c
+++ b/erts/emulator/beam/erl_bif_ddll.c
@@ -1737,7 +1737,7 @@ static void notify_proc(Process *proc, Eterm ref, Eterm driver_name, Eterm type,
 	hp += REF_THING_SIZE;
 	mess = TUPLE5(hp,type,r,am_driver,driver_name,tag);
     }
-    erts_queue_message(proc, &rp_locks, mp, mess);
+    erts_queue_message(proc, rp_locks, mp, mess, am_system);
     erts_smp_proc_unlock(proc, rp_locks);
     ERTS_SMP_CHK_NO_PROC_LOCKS;
 }
diff --git a/erts/emulator/beam/erl_bif_trace.c b/erts/emulator/beam/erl_bif_trace.c
index c604053caa..ffc0afda58 100644
--- a/erts/emulator/beam/erl_bif_trace.c
+++ b/erts/emulator/beam/erl_bif_trace.c
@@ -2344,7 +2344,7 @@ reply_trace_delivered_all(void *vtdarp)
 #ifdef ERTS_SMP
         erts_send_sys_msg_proc(rp->common.id, rp->common.id, msg, bp);
 #else
-        erts_queue_message(rp, &rp_locks, mp, msg);
+        erts_queue_message(rp, rp_locks, mp, msg, am_system);
 #endif
 
 	erts_free(ERTS_ALC_T_MISC_AUX_WORK, vtdarp);
diff --git a/erts/emulator/beam/erl_db_util.c b/erts/emulator/beam/erl_db_util.c
index ae2cc40bfa..7d64529250 100644
--- a/erts/emulator/beam/erl_db_util.c
+++ b/erts/emulator/beam/erl_db_util.c
@@ -414,17 +414,19 @@ get_match_pseudo_process(Process *c_p, Uint heap_size)
 {
     ErtsMatchPseudoProcess *mpsp;
 #ifdef ERTS_SMP
-    mpsp = (ErtsMatchPseudoProcess *) c_p->scheduler_data->match_pseudo_process;
+    ErtsSchedulerData *esdp = c_p ? c_p->scheduler_data : erts_get_scheduler_data();
+
+    mpsp = (ErtsMatchPseudoProcess *) esdp->match_pseudo_process;
     if (mpsp)
 	cleanup_match_pseudo_process(mpsp, 0);
     else {
 	ASSERT(erts_smp_tsd_get(match_pseudo_process_key) == NULL);
 	mpsp = create_match_pseudo_process();
-	c_p->scheduler_data->match_pseudo_process = (void *) mpsp;
+	esdp->match_pseudo_process = (void *) mpsp;
 	erts_smp_tsd_set(match_pseudo_process_key, (void *) mpsp);
     }
     ASSERT(mpsp == erts_smp_tsd_get(match_pseudo_process_key));
-    mpsp->process.scheduler_data = c_p->scheduler_data;
+    mpsp->process.scheduler_data = esdp;
 #else
     mpsp = match_pseudo_process;
     cleanup_match_pseudo_process(mpsp, 0);
@@ -1750,7 +1752,7 @@ Eterm db_prog_match(Process *c_p,
     Eterm *esp;
     MatchVariable* variables;
     BeamInstr *cp;
-    UWord *pc = prog->text;
+    const UWord *pc = prog->text;
     Eterm *ehp;
     Eterm ret;
     Uint n;
@@ -1773,13 +1775,15 @@ Eterm db_prog_match(Process *c_p,
 
     ERTS_UNDEF(n,0);
 
+    ASSERT(c_p || !(in_flags & ERTS_PAM_COPY_RESULT));
+
     mpsp = get_match_pseudo_process(c_p, prog->heap_size);
     psp = &mpsp->process;
 
     /* We need to lure the scheduler into believing in the pseudo process, 
        because of floating point exceptions. Do *after* mpsp is set!!! */
 
-    esdp = ERTS_GET_SCHEDULER_DATA_FROM_PROC(c_p);
+    esdp = ERTS_GET_SCHEDULER_DATA_FROM_PROC(psp);
     ASSERT(esdp != NULL);
     current_scheduled = esdp->current_process;
     /* SMP: psp->scheduler_data is set by get_match_pseudo_process */
diff --git a/erts/emulator/beam/erl_gc.c b/erts/emulator/beam/erl_gc.c
index df5d0f4918..85a1467a52 100644
--- a/erts/emulator/beam/erl_gc.c
+++ b/erts/emulator/beam/erl_gc.c
@@ -2924,7 +2924,7 @@ reply_gc_info(void *vgcirp)
 	hpp = &hp;
     }
 
-    erts_queue_message(rp, &rp_locks, mp, msg);
+    erts_queue_message(rp, rp_locks, mp, msg, am_system);
 
     if (gcirp->req_sched == esdp->no)
 	rp_locks &= ~ERTS_PROC_LOCK_MAIN;
diff --git a/erts/emulator/beam/erl_hl_timer.c b/erts/emulator/beam/erl_hl_timer.c
index 8e201d5711..c418762578 100644
--- a/erts/emulator/beam/erl_hl_timer.c
+++ b/erts/emulator/beam/erl_hl_timer.c
@@ -1247,8 +1247,8 @@ hlt_bif_timer_timeout(ErtsHLTimer *tmr, Uint32 roflgs)
 	if (!ERTS_PROC_IS_EXITING(proc)) {
 	    ErtsMessage *mp = erts_alloc_message(0, NULL);
 	    mp->data.heap_frag = tmr->btm.bp;
-	    erts_queue_message(proc, &proc_locks, mp,
-			       tmr->btm.message);
+	    erts_queue_message(proc, proc_locks, mp,
+			       tmr->btm.message, am_clock_service);
 	    erts_smp_proc_unlock(proc, ERTS_PROC_LOCKS_MSG_SEND);
 	    queued_message = 1;
 	    proc_locks &= ~ERTS_PROC_LOCKS_MSG_SEND;
@@ -1980,7 +1980,7 @@ access_sched_local_btm(Process *c_p, Eterm pid,
 
 	ERTS_HLT_ASSERT(hp + (async ? 4 : 3) == hp_end);
 
-	erts_queue_message(proc, &proc_locks, mp, msg);
+	erts_queue_message(proc, proc_locks, mp, msg, am_clock_service);
 
 	if (c_p)
 	    proc_locks &= ~ERTS_PROC_LOCK_MAIN;
@@ -2111,7 +2111,7 @@ try_access_sched_remote_btm(ErtsSchedulerData *esdp,
 
 	msg = TUPLE3(hp, tag, tref, res);
 
-	erts_queue_message(c_p, &proc_locks, mp, msg);
+        erts_queue_message(c_p, proc_locks, mp, msg, am_clock_service);
 
 	proc_locks &= ~ERTS_PROC_LOCK_MAIN;
 	if (proc_locks)
diff --git a/erts/emulator/beam/erl_message.c b/erts/emulator/beam/erl_message.c
index ee8a7702ce..9bb6e40a11 100644
--- a/erts/emulator/beam/erl_message.c
+++ b/erts/emulator/beam/erl_message.c
@@ -252,9 +252,10 @@ erts_realloc_shrink_message(ErtsMessage *mp, Uint sz, Eterm *brefs, Uint brefs_s
 
 void
 erts_queue_dist_message(Process *rcvr,
-			ErtsProcLocks *rcvr_locks,
+			ErtsProcLocks rcvr_locks,
 			ErtsDistExternal *dist_ext,
-			Eterm token)
+			Eterm token,
+                        Eterm from)
 {
     ErtsMessage* mp;
 #ifdef USE_VM_PROBES
@@ -266,7 +267,7 @@ erts_queue_dist_message(Process *rcvr,
     erts_aint_t state;
 #endif
 
-    ERTS_SMP_LC_ASSERT(*rcvr_locks == erts_proc_lc_my_proc_locks(rcvr));
+    ERTS_SMP_LC_ASSERT(rcvr_locks == erts_proc_lc_my_proc_locks(rcvr));
 
     mp = erts_alloc_message(0, NULL);
     mp->data.dist_ext = dist_ext;
@@ -281,10 +282,10 @@ erts_queue_dist_message(Process *rcvr,
 	ERL_MESSAGE_TOKEN(mp) = token;
 
 #ifdef ERTS_SMP
-    if (!(*rcvr_locks & ERTS_PROC_LOCK_MSGQ)) {
+    if (!(rcvr_locks & ERTS_PROC_LOCK_MSGQ)) {
 	if (erts_smp_proc_trylock(rcvr, ERTS_PROC_LOCK_MSGQ) == EBUSY) {
 	    ErtsProcLocks need_locks = ERTS_PROC_LOCK_MSGQ;
-	    if (*rcvr_locks & ERTS_PROC_LOCK_STATUS) {
+	    if (rcvr_locks & ERTS_PROC_LOCK_STATUS) {
 		erts_smp_proc_unlock(rcvr, ERTS_PROC_LOCK_STATUS);
 		need_locks |= ERTS_PROC_LOCK_STATUS;
 	    }
@@ -294,7 +295,7 @@ erts_queue_dist_message(Process *rcvr,
 
     state = erts_smp_atomic32_read_acqb(&rcvr->state);
     if (state & (ERTS_PSFLG_PENDING_EXIT|ERTS_PSFLG_EXITING)) {
-	if (!(*rcvr_locks & ERTS_PROC_LOCK_MSGQ))
+	if (!(rcvr_locks & ERTS_PROC_LOCK_MSGQ))
 	    erts_smp_proc_unlock(rcvr, ERTS_PROC_LOCK_MSGQ);
 	/* Drop message if receiver is exiting or has a pending exit ... */
 	erts_cleanup_messages(mp);
@@ -302,10 +303,13 @@ erts_queue_dist_message(Process *rcvr,
     else
 #endif
     if (IS_TRACED_FL(rcvr, F_TRACE_RECEIVE)) {
+        if (from == am_Empty)
+            from = dist_ext->dep->sysname;
+
 	/* Ahh... need to decode it in order to trace it... */
-	if (!(*rcvr_locks & ERTS_PROC_LOCK_MSGQ))
+	if (!(rcvr_locks & ERTS_PROC_LOCK_MSGQ))
 	    erts_smp_proc_unlock(rcvr, ERTS_PROC_LOCK_MSGQ);
-	if (!erts_decode_dist_message(rcvr, *rcvr_locks, mp, 0))
+	if (!erts_decode_dist_message(rcvr, rcvr_locks, mp, 0))
 	    erts_free_message(mp);
 	else {
 	    Eterm msg = ERL_MESSAGE_TERM(mp);
@@ -325,7 +329,7 @@ erts_queue_dist_message(Process *rcvr,
 			tok_label, tok_lastcnt, tok_serial);
 	    }
 #endif
-	    erts_queue_message(rcvr, rcvr_locks, mp, msg);
+	    erts_queue_message(rcvr, rcvr_locks, mp, msg, from);
 	}
     }
     else {
@@ -352,12 +356,12 @@ erts_queue_dist_message(Process *rcvr,
 
 	LINK_MESSAGE(rcvr, mp, &mp->next, 1);
 
-	if (!(*rcvr_locks & ERTS_PROC_LOCK_MSGQ))
+	if (!(rcvr_locks & ERTS_PROC_LOCK_MSGQ))
 	    erts_smp_proc_unlock(rcvr, ERTS_PROC_LOCK_MSGQ);
 
 	erts_proc_notify_new_message(rcvr,
 #ifdef ERTS_SMP
-				     *rcvr_locks
+				     rcvr_locks
 #else
 				     0
 #endif
@@ -367,13 +371,13 @@ erts_queue_dist_message(Process *rcvr,
 
 /* Add messages last in message queue */
 static Sint
-queue_messages(Process *c_p,
-               Process* receiver,
+queue_messages(Process* receiver,
                erts_aint32_t *receiver_state,
                ErtsProcLocks receiver_locks,
                ErtsMessage* first,
                ErtsMessage** last,
-               Uint len)
+               Uint len,
+               Eterm from)
 {
     ErtsTracingEvent* te;
     Sint res;
@@ -475,7 +479,7 @@ queue_messages(Process *c_p,
         }
 #endif
         while (msg) {
-            trace_receive(c_p, receiver, ERL_MESSAGE_TERM(msg), te);
+            trace_receive(receiver, from, ERL_MESSAGE_TERM(msg), te);
             msg = msg->next;
         }
 
@@ -494,31 +498,31 @@ queue_messages(Process *c_p,
 }
 
 static Sint
-queue_message(Process *c_p,
-              Process* receiver,
+queue_message(Process* receiver,
               erts_aint32_t *receiver_state,
-              ErtsProcLocks *receiver_locks,
-              ErtsMessage* mp, Eterm msg)
+              ErtsProcLocks receiver_locks,
+              ErtsMessage* mp, Eterm msg, Eterm from)
 {
     ERL_MESSAGE_TERM(mp) = msg;
-    return queue_messages(c_p, receiver, receiver_state, *receiver_locks,
-                          mp, &mp->next, 1 );
+    return queue_messages(receiver, receiver_state, receiver_locks,
+                          mp, &mp->next, 1, from);
 }
 
 Sint
-erts_queue_message(Process* receiver, ErtsProcLocks *receiver_locks,
-                   ErtsMessage* mp, Eterm msg)
+erts_queue_message(Process* receiver, ErtsProcLocks receiver_locks,
+                   ErtsMessage* mp, Eterm msg, Eterm from)
 {
-    return queue_message(NULL, receiver, NULL, receiver_locks, mp, msg);
+    return queue_message(receiver, NULL, receiver_locks, mp, msg, from);
 }
 
 
 Sint
-erts_queue_messages(Process* receiver, ErtsProcLocks *receiver_locks,
-                    ErtsMessage* first, ErtsMessage** last, Uint len)
+erts_queue_messages(Process* receiver, ErtsProcLocks receiver_locks,
+                    ErtsMessage* first, ErtsMessage** last, Uint len,
+                    Eterm from)
 {
-    return queue_messages(NULL, receiver, NULL, *receiver_locks,
-                          first, last, len);
+    return queue_messages(receiver, NULL, receiver_locks,
+                          first, last, len, from);
 }
 
 void
@@ -835,11 +839,11 @@ erts_send_message(Process* sender,
 #ifdef USE_VM_PROBES
     ERL_MESSAGE_DT_UTAG(mp) = utag;
 #endif
-    res = queue_message(sender,
-			receiver,
+    res = queue_message(receiver,
 			&receiver_state,
-			receiver_locks,
-			mp, message);
+			*receiver_locks,
+			mp, message,
+                        sender->common.id);
 
     BM_SWAP_TIMER(send,system);
     
@@ -896,7 +900,7 @@ erts_deliver_exit_message(Eterm from, Process *to, ErtsProcLocks *to_locksp,
 	seq_trace_output(token, save, SEQ_TRACE_SEND, to->common.id, NULL);
 	temptoken = copy_struct(token, sz_token, &hp, ohp);
         ERL_MESSAGE_TOKEN(mp) = temptoken;
-	erts_queue_message(to, to_locksp, mp, save);
+	erts_queue_message(to, *to_locksp, mp, save, am_system);
     } else {
 	sz_from = IS_CONST(from) ? 0 : size_object(from);
 #ifdef SHCOPY_SEND
@@ -918,7 +922,7 @@ erts_deliver_exit_message(Eterm from, Process *to, ErtsProcLocks *to_locksp,
 		     ? from
 		     : copy_struct(from, sz_from, &hp, ohp));
 	save = TUPLE3(hp, am_EXIT, from_copy, mess);
-	erts_queue_message(to, to_locksp, mp, save);
+	erts_queue_message(to, *to_locksp, mp, save, am_system);
     }
 }
 
diff --git a/erts/emulator/beam/erl_message.h b/erts/emulator/beam/erl_message.h
index 608cf552a2..851ac37fda 100644
--- a/erts/emulator/beam/erl_message.h
+++ b/erts/emulator/beam/erl_message.h
@@ -295,10 +295,10 @@ ErlHeapFragment* new_message_buffer(Uint);
 ErlHeapFragment* erts_resize_message_buffer(ErlHeapFragment *, Uint,
 					    Eterm *, Uint);
 void free_message_buffer(ErlHeapFragment *);
-void erts_queue_dist_message(Process*, ErtsProcLocks*, ErtsDistExternal *, Eterm);
-Sint erts_queue_message(Process*, ErtsProcLocks*,ErtsMessage*, Eterm);
-Sint erts_queue_messages(Process*, ErtsProcLocks*,
-                         ErtsMessage*, ErtsMessage**, Uint);
+void erts_queue_dist_message(Process*, ErtsProcLocks, ErtsDistExternal *, Eterm, Eterm);
+Sint erts_queue_message(Process*, ErtsProcLocks,ErtsMessage*, Eterm, Eterm);
+Sint erts_queue_messages(Process*, ErtsProcLocks,
+                         ErtsMessage*, ErtsMessage**, Uint, Eterm);
 void erts_deliver_exit_message(Eterm, Process*, ErtsProcLocks *, Eterm, Eterm);
 Sint erts_send_message(Process*, Process*, ErtsProcLocks*, Eterm, unsigned);
 void erts_link_mbuf_to_proc(Process *proc, ErlHeapFragment *bp);
diff --git a/erts/emulator/beam/erl_msacc.c b/erts/emulator/beam/erl_msacc.c
index d0f305900a..0e625f213b 100644
--- a/erts/emulator/beam/erl_msacc.c
+++ b/erts/emulator/beam/erl_msacc.c
@@ -257,7 +257,7 @@ static void send_reply(ErtsMsAcc *msacc, ErtsMSAccReq *msaccrp) {
 
     if (msacc->unmanaged) erts_mtx_unlock(&msacc->mtx);
 
-    erts_queue_message(rp, &rp_locks, msgp, msg);
+    erts_queue_message(rp, rp_locks, msgp, msg, am_system);
 
     if (esdp && msaccrp->req_sched == esdp->no)
 	rp_locks &= ~ERTS_PROC_LOCK_MAIN;
diff --git a/erts/emulator/beam/erl_nif.c b/erts/emulator/beam/erl_nif.c
index 941f44b9ec..bd3e7ebe54 100644
--- a/erts/emulator/beam/erl_nif.c
+++ b/erts/emulator/beam/erl_nif.c
@@ -366,7 +366,7 @@ int erts_flush_trace_messages(Process *c_p, ErtsProcLocks c_p_locks)
             rp_locks = 0;
             if (rp->common.id == c_p->common.id)
                 rp_locks = c_p_locks;
-            erts_queue_messages(rp, &rp_locks, first, last, len);
+            erts_queue_messages(rp, rp_locks, first, last, len, c_p->common.id);
             if (rp->common.id == c_p->common.id)
                 rp_locks &= ~c_p_locks;
             if (rp_locks)
@@ -405,7 +405,10 @@ enif_send(ErlNifEnv* env, const ErlNifPid* to_pid,
           ErlNifEnv* msg_env, ERL_NIF_TERM msg)
 {
     struct enif_msg_environment_t* menv = (struct enif_msg_environment_t*)msg_env;
-    ErtsProcLocks rp_locks = 0, lc_locks = 0, c_p_locks = ERTS_PROC_LOCK_MAIN;
+    ErtsProcLocks rp_locks = 0;
+#ifdef ERTS_SMP
+    ErtsProcLocks lc_locks = 0;
+#endif
     Process* rp;
     Process* c_p;
     ErtsMessage *mp;
@@ -417,7 +420,7 @@ enif_send(ErlNifEnv* env, const ErlNifPid* to_pid,
     if (env != NULL) {
 	c_p = env->proc;
 	if (receiver == c_p->common.id) {
-	    rp_locks = c_p_locks;
+	    rp_locks = ERTS_PROC_LOCK_MAIN;
 	    flush_me = 1;
 	}
     }
@@ -470,14 +473,8 @@ enif_send(ErlNifEnv* env, const ErlNifPid* to_pid,
 
         if (c_p && IS_TRACED_FL(c_p, F_TRACE_SEND))
             trace_send(c_p, receiver, msg);
-
-#ifndef ERTS_SMP
     }
-#endif
-
-        erts_queue_message(rp, &rp_locks, mp, msg);
 #ifdef ERTS_SMP
-    }
     else {
         /* This clause is taken when the nif is called in the context
            of a traced process. We do not know which locks we have
@@ -502,8 +499,6 @@ enif_send(ErlNifEnv* env, const ErlNifPid* to_pid,
 #ifdef ERTS_ENABLE_LOCK_CHECK
         lc_locks = erts_proc_lc_my_proc_locks(rp);
         rp_locks |= lc_locks;
-        if (receiver == c_p->common.id)
-            c_p_locks |= lc_locks;
 #endif
         if (ERTS_FORCE_ENIF_SEND_DELAY() || msgq ||
             rp_locks & ERTS_PROC_LOCK_MSGQ ||
@@ -532,19 +527,25 @@ enif_send(ErlNifEnv* env, const ErlNifPid* to_pid,
                 msgq->last = &mp->next;
                 erts_smp_proc_unlock(t_p, ERTS_PROC_LOCK_TRACE);
             }
+            goto done;
         } else {
             erts_smp_proc_unlock(t_p, ERTS_PROC_LOCK_TRACE);
             rp_locks &= ~ERTS_PROC_LOCK_TRACE;
             rp_locks |= ERTS_PROC_LOCK_MSGQ;
-            erts_queue_message(rp, &rp_locks, mp, msg);
         }
     }
-#endif
+#endif /* ERTS_SMP */
+
+    erts_queue_message(rp, rp_locks, mp, msg,
+                       c_p ? c_p->common.id : am_undefined);
 
+#ifdef ERTS_SMP
+done:
     if (c_p == rp)
 	rp_locks &= ~ERTS_PROC_LOCK_MAIN;
     if (rp_locks & ~lc_locks)
 	erts_smp_proc_unlock(rp, rp_locks & ~lc_locks);
+#endif
     if (!scheduler)
 	erts_proc_dec_refc(rp);
     if (flush_me) {
diff --git a/erts/emulator/beam/erl_process.c b/erts/emulator/beam/erl_process.c
index d485affa3b..da2559839d 100644
--- a/erts/emulator/beam/erl_process.c
+++ b/erts/emulator/beam/erl_process.c
@@ -1139,7 +1139,7 @@ reply_sched_wall_time(void *vswtrp)
 	hpp = &hp;
     }
 
-    erts_queue_message(rp, &rp_locks, mp, msg);
+    erts_queue_message(rp, rp_locks, mp, msg, am_system);
 
     if (swtrp->req_sched == esdp->no)
 	rp_locks &= ~ERTS_PROC_LOCK_MAIN;
@@ -1218,7 +1218,7 @@ reply_system_check(void *vscrp)
     hpp = &hp;
     msg = STORE_NC(hpp, ohp, scrp->ref);
 
-    erts_queue_message(rp, &rp_locks, mp, msg);
+    erts_queue_message(rp, rp_locks, mp, msg, am_system);
 
     if (scrp->req_sched == esdp->no)
 	rp_locks &= ~ERTS_PROC_LOCK_MAIN;
@@ -10010,7 +10010,7 @@ notify_sys_task_executed(Process *c_p, ErtsProcSysTask *st, Eterm st_result)
 	ASSERT(hp_start + hsz == hp);
 #endif
 
-	erts_queue_message(rp, &rp_locks, mp, msg);
+	erts_queue_message(rp, rp_locks, mp, msg, c_p->common.id);
 
 	if (c_p == rp)
 	    rp_locks &= ~ERTS_PROC_LOCK_MAIN;
@@ -11744,7 +11744,7 @@ send_exit_message(Process *to, ErtsProcLocks *to_locksp,
 	mp = erts_alloc_message_heap(to, to_locksp, term_size, &hp, &ohp);
 	mess = copy_struct(exit_term, term_size, &hp, ohp);
 #endif
-	erts_queue_message(to, to_locksp, mp, mess);
+	erts_queue_message(to, *to_locksp, mp, mess, am_system);
     } else {
 	Eterm temp_token;
 	Uint sz_token;
@@ -11765,7 +11765,7 @@ send_exit_message(Process *to, ErtsProcLocks *to_locksp,
 	seq_trace_output(token, mess, SEQ_TRACE_SEND, to->common.id, to);
 	temp_token = copy_struct(token, sz_token, &hp, ohp);
         ERL_MESSAGE_TOKEN(mp) = temp_token;
-	erts_queue_message(to, to_locksp, mp, mess);
+	erts_queue_message(to, *to_locksp, mp, mess, am_system);
     }
 }
 
diff --git a/erts/emulator/beam/erl_time_sup.c b/erts/emulator/beam/erl_time_sup.c
index 346404fe2a..fadbd704bd 100644
--- a/erts/emulator/beam/erl_time_sup.c
+++ b/erts/emulator/beam/erl_time_sup.c
@@ -1966,7 +1966,7 @@ send_time_offset_changed_notifications(void *new_offsetp)
 		    *patch_refp = ref;
 		    ASSERT(hsz == size_object(message_template));
 		    message = copy_struct(message_template, hsz, &hp, ohp);
-		    erts_queue_message(rp, &rp_locks, mp, message);
+		    erts_queue_message(rp, rp_locks, mp, message, am_clock_service);
 		}
 		erts_smp_proc_unlock(rp, rp_locks);
 	    }
diff --git a/erts/emulator/beam/erl_trace.c b/erts/emulator/beam/erl_trace.c
index 755deda9c1..682f08f667 100644
--- a/erts/emulator/beam/erl_trace.c
+++ b/erts/emulator/beam/erl_trace.c
@@ -738,7 +738,7 @@ profile_send(Eterm from, Eterm message) {
 	else
 	    msg = copy_struct(message, sz, &hp, &mp->hfrag.off_heap);
 
-        erts_queue_message(profile_p, NULL, mp, msg);
+        erts_queue_message(profile_p, 0, mp, msg, from);
     }
 }
 
@@ -859,8 +859,8 @@ trace_send(Process *p, Eterm to, Eterm msg)
  * or   {trace, Pid, receive, Msg}
  */
 void
-trace_receive(Process *c_p,
-              Process* receiver,
+trace_receive(Process* receiver,
+              Eterm from,
               Eterm msg, ErtsTracingEvent* te)
 {
     ErtsTracerNif *tnif = NULL;
@@ -875,15 +875,15 @@ trace_receive(Process *c_p,
     if (te->match_spec) {
         Eterm args[2];
         Uint32 return_flags;
-        args[0] = am_undefined; /* ToDo: from who? */
+        args[0] = from;
         args[1] = msg;
-        pam_result = erts_match_set_run(c_p, receiver,
+        pam_result = erts_match_set_run(NULL, receiver,
                                         te->match_spec, args, 2,
                                         ERTS_PAM_TMP_RESULT, &return_flags);
         if (is_non_value(pam_result)
             || pam_result == am_false
             || (ERTS_TRACE_FLAGS(receiver) & F_TRACE_SILENT)) {
-            erts_match_set_release_result(c_p);
+            erts_match_set_release_result(NULL);
             return;
         }
     } else
@@ -895,7 +895,7 @@ trace_receive(Process *c_p,
                            tnif, TRACE_FUN_T_RECEIVE,
                            am_receive, msg, THE_NON_VALUE, pam_result);
     }
-    erts_match_set_release_result(c_p);
+    erts_match_set_release_result(NULL);
 }
 
 int
@@ -1483,7 +1483,7 @@ monitor_long_schedule_proc(Process *p, BeamInstr *in_fp, BeamInstr *out_fp, Uint
     {
 	ErtsMessage *mp = erts_alloc_message(0, NULL);
 	mp->data.heap_frag = bp;
-	erts_queue_message(monitor_p, NULL, mp, msg);
+	erts_queue_message(monitor_p, 0, mp, msg, am_system);
     }
 #endif
 }
@@ -1548,7 +1548,7 @@ monitor_long_schedule_port(Port *pp, ErtsPortTaskType type, Uint time)
     {
 	ErtsMessage *mp = erts_alloc_message(0, NULL);
 	mp->data.heap_frag = bp;
-	erts_queue_message(monitor_p, NULL, mp, msg);
+	erts_queue_message(monitor_p, 0, mp, msg, am_system);
     }
 #endif
 }
@@ -1623,7 +1623,7 @@ monitor_long_gc(Process *p, Uint time) {
     {
 	ErtsMessage *mp = erts_alloc_message(0, NULL);
 	mp->data.heap_frag = bp;
-	erts_queue_message(monitor_p, NULL, mp, msg);
+	erts_queue_message(monitor_p, 0, mp, msg, am_system);
     }
 #endif
 }
@@ -1698,7 +1698,7 @@ monitor_large_heap(Process *p) {
     {
 	ErtsMessage *mp = erts_alloc_message(0, NULL);
 	mp->data.heap_frag = bp;
-	erts_queue_message(monitor_p, NULL, mp, msg);
+	erts_queue_message(monitor_p, 0, mp, msg, am_system);
     }
 #endif
 }
@@ -1730,7 +1730,7 @@ monitor_generic(Process *p, Eterm type, Eterm spec) {
     {
 	ErtsMessage *mp = erts_alloc_message(0, NULL);
 	mp->data.heap_frag = bp;
-	erts_queue_message(monitor_p, NULL, mp, msg);
+	erts_queue_message(monitor_p, 0, mp, msg, am_system);
     }
 #endif
 
@@ -2514,7 +2514,7 @@ sys_msg_dispatcher_func(void *unused)
 		queue_proc_msg:
 		    mp = erts_alloc_message(0, NULL);
 		    mp->data.heap_frag = smqp->bp;
-		    erts_queue_message(proc,&proc_locks,mp,smqp->msg);
+		    erts_queue_message(proc,proc_locks,mp,smqp->msg,am_system);
 #ifdef DEBUG_PRINTOUTS
 		    erts_fprintf(stderr, "delivered\n");
 #endif
diff --git a/erts/emulator/beam/erl_trace.h b/erts/emulator/beam/erl_trace.h
index 040aa296a1..de4beadb7e 100644
--- a/erts/emulator/beam/erl_trace.h
+++ b/erts/emulator/beam/erl_trace.h
@@ -100,7 +100,7 @@ void erts_send_sys_msg_proc(Eterm, Eterm, Eterm, ErlHeapFragment *);
 #endif
 
 void trace_send(Process*, Eterm, Eterm);
-void trace_receive(Process*, Process*, Eterm, ErtsTracingEvent*);
+void trace_receive(Process*, Eterm, Eterm, ErtsTracingEvent*);
 Uint32 erts_call_trace(Process *p, BeamInstr mfa[], struct binary *match_spec,
                        Eterm* args, int local, ErtsTracer *tracer);
 void erts_trace_return(Process* p, BeamInstr* fi, Eterm retval,
diff --git a/erts/emulator/beam/io.c b/erts/emulator/beam/io.c
index b14ca77a04..1d307ae15e 100644
--- a/erts/emulator/beam/io.c
+++ b/erts/emulator/beam/io.c
@@ -1432,10 +1432,11 @@ finalize_force_imm_drv_call(ErtsTryImmDrvCallState *sp)
 
 static ERTS_INLINE void
 queue_port_sched_op_reply(Process *rp,
-			  ErtsProcLocks *rp_locksp,
+			  ErtsProcLocks rp_locks,
                           ErtsHeapFactory* factory,
 			  Uint32 *ref_num,
-			  Eterm msg)
+			  Eterm msg,
+			  Port* prt)
 {
     Eterm* hp = erts_produce_heap(factory, ERTS_QUEUE_PORT_SCHED_OP_REPLY_SIZE, 0);
     Eterm ref;
@@ -1448,11 +1449,12 @@ queue_port_sched_op_reply(Process *rp,
 
     erts_factory_trim_and_close(factory, &msg, 1);
 
-    erts_queue_message(rp, rp_locksp, factory->message, msg);
+    erts_queue_message(rp, rp_locks, factory->message, msg,
+		       prt ? prt->common.id : am_undefined);
 }
 
 static void
-port_sched_op_reply(Eterm to, Uint32 *ref_num, Eterm msg)
+port_sched_op_reply(Eterm to, Uint32 *ref_num, Eterm msg, Port* prt)
 {
     Process *rp = erts_proc_lookup_raw(to);
     if (rp) {
@@ -1478,10 +1480,11 @@ port_sched_op_reply(Eterm to, Uint32 *ref_num, Eterm msg)
 				  factory.off_heap));
 
 	queue_port_sched_op_reply(rp,
-				  &rp_locks,
+				  rp_locks,
                                   &factory,
 				  ref_num,
-				  msg_copy);
+				  msg_copy,
+				  prt);
 
 	if (rp_locks)
 	    erts_smp_proc_unlock(rp, rp_locks);
@@ -1651,7 +1654,7 @@ port_badsig(Port *prt, erts_aint32_t state, int op,
 			state,
 			sigdp->flags & ERTS_P2P_SIG_DATA_FLG_BAD_OUTPUT);
     if (sigdp->flags & ERTS_P2P_SIG_DATA_FLG_REPLY)
-	port_sched_op_reply(sigdp->caller, sigdp->ref, am_badarg);
+	port_sched_op_reply(sigdp->caller, sigdp->ref, am_badarg, prt);
     return ERTS_PORT_REDS_BADSIG;
 } /* port_badsig */
 /* bad_port_signal() will
@@ -1820,7 +1823,7 @@ port_sig_outputv(Port *prt, erts_aint32_t state, int op, ErtsProc2PortSigData *s
     }
 
     if (sigdp->flags & ERTS_P2P_SIG_DATA_FLG_REPLY)
-	port_sched_op_reply(sigdp->caller, sigdp->ref, reply);
+	port_sched_op_reply(sigdp->caller, sigdp->ref, reply, prt);
 
     cleanup_scheduled_outputv(sigdp->u.outputv.evp,
 			      sigdp->u.outputv.cbinp);
@@ -1928,7 +1931,7 @@ port_sig_output(Port *prt, erts_aint32_t state, int op, ErtsProc2PortSigData *si
     }
 
     if (sigdp->flags & ERTS_P2P_SIG_DATA_FLG_REPLY)
-	port_sched_op_reply(sigdp->caller, sigdp->ref, reply);
+	port_sched_op_reply(sigdp->caller, sigdp->ref, reply, prt);
 
     cleanup_scheduled_output(sigdp->u.output.bufp);
 
@@ -2636,7 +2639,7 @@ port_sig_exit(Port *prt,
     if (sigdp->u.exit.bp)
 	free_message_buffer(sigdp->u.exit.bp);
     if (sigdp->flags & ERTS_P2P_SIG_DATA_FLG_REPLY)
-	port_sched_op_reply(sigdp->caller, sigdp->ref, msg);
+	port_sched_op_reply(sigdp->caller, sigdp->ref, msg, prt);
 
     return ERTS_PORT_REDS_EXIT;
 }
@@ -2829,7 +2832,7 @@ port_sig_connect(Port *prt, erts_aint32_t state, int op, ErtsProc2PortSigData *s
 	    msg = am_true;
     }
     if (sigdp->flags & ERTS_P2P_SIG_DATA_FLG_REPLY)
-	port_sched_op_reply(sigdp->caller, sigdp->ref, msg);
+	port_sched_op_reply(sigdp->caller, sigdp->ref, msg, prt);
     return ERTS_PORT_REDS_CONNECT;
 }
 
@@ -2912,7 +2915,7 @@ port_sig_unlink(Port *prt, erts_aint32_t state, int op, ErtsProc2PortSigData *si
     if (op == ERTS_PROC2PORT_SIG_EXEC)
 	port_unlink(prt, sigdp->u.unlink.from);
     if (sigdp->flags & ERTS_P2P_SIG_DATA_FLG_REPLY)
-	port_sched_op_reply(sigdp->caller, sigdp->ref, am_true);
+	port_sched_op_reply(sigdp->caller, sigdp->ref, am_true, prt);
     return ERTS_PORT_REDS_UNLINK;
 }
 
@@ -3007,7 +3010,7 @@ port_sig_link(Port *prt, erts_aint32_t state, int op, ErtsProc2PortSigData *sigd
 	port_link_failure(sigdp->u.link.port, sigdp->u.link.to);
     }
     if (sigdp->flags & ERTS_P2P_SIG_DATA_FLG_REPLY)
-	port_sched_op_reply(sigdp->caller, sigdp->ref, am_true);
+	port_sched_op_reply(sigdp->caller, sigdp->ref, am_true, prt);
     return ERTS_PORT_REDS_LINK;
 }
 
@@ -3064,7 +3067,8 @@ init_ack_send_reply(Port *port, Eterm resp)
     }
     port_sched_op_reply(port->async_open_port->to,
                         port->async_open_port->ref,
-                        resp);
+                        resp,
+			port);
 
     erts_free(ERTS_ALC_T_PRTSD, port->async_open_port);
     port->async_open_port = NULL;
@@ -3461,7 +3465,7 @@ deliver_result(Port *prt, Eterm sender, Eterm pid, Eterm res)
 				     sz_res + 3, &hp, &ohp);
 	res = copy_struct(res, sz_res, &hp, ohp);
 	tuple = TUPLE2(hp, sender, res);
-	erts_queue_message(rp, &rp_locks, mp, tuple);
+	erts_queue_message(rp, rp_locks, mp, tuple, sender);
 
 	if (rp_locks)
 	    erts_smp_proc_unlock(rp, rp_locks);
@@ -3562,7 +3566,7 @@ static void deliver_read_message(Port* prt, erts_aint32_t state, Eterm to,
         trace_port_send(prt, to, tuple, 1);
 
     ERL_MESSAGE_TOKEN(mp) = am_undefined;
-    erts_queue_message(rp, &rp_locks, mp, tuple);
+    erts_queue_message(rp, rp_locks, mp, tuple, prt->common.id);
     if (rp_locks)
 	erts_smp_proc_unlock(rp, rp_locks);
     if (!scheduler)
@@ -3734,7 +3738,7 @@ deliver_vec_message(Port* prt,			/* Port */
         trace_port_send(prt, to, tuple, 1);
 
     ERL_MESSAGE_TOKEN(mp) = am_undefined;
-    erts_queue_message(rp, &rp_locks, mp, tuple);
+    erts_queue_message(rp, rp_locks, mp, tuple, prt->common.id);
     erts_smp_proc_unlock(rp, rp_locks);
     if (!scheduler)
 	erts_proc_dec_refc(rp);
@@ -4388,10 +4392,11 @@ port_sig_control(Port *prt,
 					    &factory.hp,
 					    factory.off_heap);
 	    queue_port_sched_op_reply(rp,
-				      &rp_locks,
+				      rp_locks,
                                       &factory,
 				      sigdp->ref,
-				      msg);
+				      msg,
+				      prt);
 
 	    if (rp_locks)
 		erts_smp_proc_unlock(rp, rp_locks);
@@ -4402,7 +4407,7 @@ port_sig_control(Port *prt,
     /* failure */
 
     if (sigdp->caller != ERTS_INVALID_PID)
-        port_sched_op_reply(sigdp->caller, sigdp->ref, am_badarg);
+        port_sched_op_reply(sigdp->caller, sigdp->ref, am_badarg, prt);
 
 done:
 
@@ -4739,10 +4744,11 @@ port_sig_call(Port *prt,
 		    msg = TUPLE2(hp, am_ok, msg);
 
 		    queue_port_sched_op_reply(rp,
-					      &rp_locks,
+					      rp_locks,
                                               &factory,
 					      sigdp->ref,
-					      msg);
+					      msg,
+					      prt);
 
 		    if (rp_locks)
 			erts_smp_proc_unlock(rp, rp_locks);
@@ -4754,7 +4760,7 @@ port_sig_call(Port *prt,
 	}
     }
 
-    port_sched_op_reply(sigdp->caller, sigdp->ref, am_badarg);
+    port_sched_op_reply(sigdp->caller, sigdp->ref, am_badarg, prt);
 
 done:
 
@@ -4969,7 +4975,7 @@ port_sig_info(Port *prt,
 {
     ASSERT(sigdp->flags & ERTS_P2P_SIG_DATA_FLG_REPLY);
     if (op != ERTS_PROC2PORT_SIG_EXEC)
-	port_sched_op_reply(sigdp->caller, sigdp->ref, am_undefined);
+	port_sched_op_reply(sigdp->caller, sigdp->ref, am_undefined, prt);
     else {
 	Eterm *hp, *hp_start;
 	Uint hsz;
@@ -4995,10 +5001,11 @@ port_sig_info(Port *prt,
 	    mp->data.heap_frag = bp;
             erts_factory_selfcontained_message_init(&factory, mp, hp);
 	    queue_port_sched_op_reply(rp,
-				      &rp_locks,
+				      rp_locks,
                                       &factory,
 				      sigdp->ref,
-				      value);
+				      value,
+				      prt);
 	}
 	if (rp_locks)
 	    erts_smp_proc_unlock(rp, rp_locks);
@@ -5115,7 +5122,7 @@ reply_io_bytes(void *vreq)
 
 	msg = TUPLE4(hp, ref, make_small(sched_id), ein, eout);
 
-	erts_queue_message(rp, &rp_locks, mp, msg);
+	erts_queue_message(rp, rp_locks, mp, msg, am_system);
 
 	if (req->sched_id == sched_id)
 	    rp_locks &= ~ERTS_PROC_LOCK_MAIN;
@@ -5613,7 +5620,7 @@ void driver_report_exit(ErlDrvPort ix, int status)
         trace_port_send(prt, pid, tuple, 1);
 
    ERL_MESSAGE_TOKEN(mp) = am_undefined;
-   erts_queue_message(rp, &rp_locks, mp, tuple);
+   erts_queue_message(rp, rp_locks, mp, tuple, prt->common.id);
 
    erts_smp_proc_unlock(rp, rp_locks);
    if (!scheduler)
@@ -6217,15 +6224,20 @@ driver_deliver_term(Port *prt, Eterm to, ErlDrvTermData* data, int len)
  done:
 
     if (res > 0) {
+	Eterm from = am_undefined;
 	mess = ESTACK_POP(stack);  /* get resulting value */
 	erts_factory_trim_and_close(&factory, &mess, 1);
 
-        if (prt && IS_TRACED_FL(prt, F_TRACE_SEND))
-            trace_port_send(prt, to, mess, 1);
+	if (prt) {
+	    if (IS_TRACED_FL(prt, F_TRACE_SEND)) {
+		trace_port_send(prt, to, mess, 1);
+	    }
+	    from = prt->common.id;
+	}
 
 	/* send message */
         ERL_MESSAGE_TOKEN(factory.message) = am_undefined;
-	erts_queue_message(rp, &rp_locks, factory.message, mess);
+	erts_queue_message(rp, rp_locks, factory.message, mess, from);
     }
     else if (res == -2) {
         /* this clause only happens when we were requested to
diff --git a/erts/emulator/beam/utils.c b/erts/emulator/beam/utils.c
index b280995488..74114626be 100644
--- a/erts/emulator/beam/utils.c
+++ b/erts/emulator/beam/utils.c
@@ -2298,7 +2298,7 @@ static void do_send_logger_message(Eterm *hp, ErlOffHeap *ohp, ErlHeapFragment *
     {
 	ErtsMessage *mp = erts_alloc_message(0, NULL);
 	mp->data.heap_frag = bp;
-	erts_queue_message(p, NULL /* only used for smp build */, mp, message);
+	erts_queue_message(p, 0, mp, message, am_system);
     }
 #endif
 }