Merge branch 'sverker/trace-send-receive-matchspec/OTP-13507'

25 files changed, 618 insertions, 401 deletions

diff --git a/erts/emulator/beam/beam_bp.c b/erts/emulator/beam/beam_bp.c
index 2ee98ed7b5..1e30e8d8d1 100644
--- a/erts/emulator/beam/beam_bp.c
+++ b/erts/emulator/beam/beam_bp.c
@@ -248,7 +248,10 @@ erts_bp_match_export(BpFunctions* f, Eterm mfa[3], int specified)
 void
 erts_bp_free_matched_functions(BpFunctions* f)
 {
-    Free(f->matching);
+    if (f->matching) {
+	Free(f->matching);
+    }
+    else ASSERT(f->matched == 0);
 }
 
 void
@@ -1432,7 +1435,7 @@ set_function_break(BeamInstr *pc, Binary *match_spec, Uint break_flags,
     g = (GenericBp *) pc[-4];
     if (g == 0) {
 	int i;
-	if (count_op == erts_break_reset || count_op == erts_break_stop) {
+	if (count_op == ERTS_BREAK_RESTART || count_op == ERTS_BREAK_PAUSE) {
 	    /* Do not insert a new breakpoint */
 	    return;
 	}
@@ -1456,7 +1459,7 @@ set_function_break(BeamInstr *pc, Binary *match_spec, Uint break_flags,
 	MatchSetUnref(bp->meta_ms);
 	bp_meta_unref(bp->meta_tracer);
     } else if (common & ERTS_BPF_COUNT) {
-	if (count_op == erts_break_stop) {
+	if (count_op == ERTS_BREAK_PAUSE) {
 	    bp->flags &= ~ERTS_BPF_COUNT_ACTIVE;
 	} else {
 	    bp->flags |= ERTS_BPF_COUNT_ACTIVE;
@@ -1468,7 +1471,7 @@ set_function_break(BeamInstr *pc, Binary *match_spec, Uint break_flags,
 	BpDataTime* bdt = bp->time;
 	Uint i = 0;
 
-	if (count_op == erts_break_stop) {
+	if (count_op == ERTS_BREAK_PAUSE) {
 	    bp->flags &= ~ERTS_BPF_TIME_TRACE_ACTIVE;
 	} else {
 	    bp->flags |= ERTS_BPF_TIME_TRACE_ACTIVE;
diff --git a/erts/emulator/beam/beam_bp.h b/erts/emulator/beam/beam_bp.h
index 08641b86d6..9c2fc007a2 100644
--- a/erts/emulator/beam/beam_bp.h
+++ b/erts/emulator/beam/beam_bp.h
@@ -86,10 +86,10 @@ typedef struct generic_bp {
 #endif
 
 enum erts_break_op{
-    erts_break_nop   =  0, /* Must be false */
-    erts_break_set   = !0, /* Must be true */
-    erts_break_reset,
-    erts_break_stop
+    ERTS_BREAK_NOP   =  0, /* Must be false */
+    ERTS_BREAK_SET   = !0, /* Must be true */
+    ERTS_BREAK_RESTART,
+    ERTS_BREAK_PAUSE
 };
 
 typedef Uint32 ErtsBpIndex;
diff --git a/erts/emulator/beam/beam_emu.c b/erts/emulator/beam/beam_emu.c
index 59112e1e43..99eb26bec5 100644
--- a/erts/emulator/beam/beam_emu.c
+++ b/erts/emulator/beam/beam_emu.c
@@ -2200,7 +2200,7 @@ void process_main(void)
 
      PreFetch(0, next);
      if (IS_TRACED_FL(c_p, F_TRACE_RECEIVE)) {
-	 trace_receive(c_p, am_timeout);
+	 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 9d6dfefa3c..eed0702688 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 277e0668a2..09c83f1117 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 ff2018aa27..b65c0e303f 100644
--- a/erts/emulator/beam/erl_bif_trace.c
+++ b/erts/emulator/beam/erl_bif_trace.c
@@ -68,6 +68,9 @@ static struct {			/* Protected by code write permission */
 
 static Eterm
 trace_pattern(Process* p, Eterm MFA, Eterm Pattern, Eterm flaglist);
+static int
+erts_set_tracing_event_pattern(Eterm event, Binary*, int on);
+
 #ifdef ERTS_SMP
 static void smp_bp_finisher(void* arg);
 #endif
@@ -78,6 +81,8 @@ static void new_seq_trace_token(Process* p); /* help func for seq_trace_2*/
 static Eterm trace_info_pid(Process* p, Eterm pid_spec, Eterm key);
 static Eterm trace_info_func(Process* p, Eterm pid_spec, Eterm key);
 static Eterm trace_info_on_load(Process* p, Eterm key);
+static Eterm trace_info_event(Process* p, Eterm event, Eterm key);
+
 
 static void reset_bif_trace(void);
 static void setup_bif_trace(void);
@@ -85,14 +90,26 @@ static void install_exp_breakpoints(BpFunctions* f);
 static void uninstall_exp_breakpoints(BpFunctions* f);
 static void clean_export_entries(BpFunctions* f);
 
+ErtsTracingEvent erts_send_tracing[ERTS_NUM_BP_IX];
+ErtsTracingEvent erts_receive_tracing[ERTS_NUM_BP_IX];
+
 void
 erts_bif_trace_init(void)
 {
+    int i;
+
     erts_default_trace_pattern_is_on = 0;
     erts_default_match_spec = NULL;
     erts_default_meta_match_spec = NULL;
     erts_default_trace_pattern_flags = erts_trace_pattern_flags_off;
     erts_default_meta_tracer = erts_tracer_nil;
+
+    for (i=0; i<ERTS_NUM_BP_IX; i++) {
+        erts_send_tracing[i].on = 1;
+        erts_send_tracing[i].match_spec = NULL;
+	erts_receive_tracing[i].on = 1;
+	erts_receive_tracing[i].match_spec = NULL;
+    }
 }
 
 /*
@@ -137,15 +154,18 @@ trace_pattern(Process* p, Eterm MFA, Eterm Pattern, Eterm flaglist)
 	on = 1;
     } else if (Pattern == am_restart) {
 	match_prog_set = NULL;
-	on = erts_break_reset;
+	on = ERTS_BREAK_RESTART;
     } else if (Pattern == am_pause) {
 	match_prog_set = NULL;
-	on = erts_break_stop;
-    } else if ((match_prog_set = erts_match_set_compile(p, Pattern)) != NULL) {
-	MatchSetRef(match_prog_set);
-	on = 1;
-    } else{
-	goto error;
+	on = ERTS_BREAK_PAUSE;
+    } else {
+	match_prog_set = erts_match_set_compile(p, Pattern, MFA);
+	if (match_prog_set) {
+	    MatchSetRef(match_prog_set);
+	    on = 1;
+	} else{
+	    goto error;
+	}
     }
     
     is_global = 0;
@@ -314,6 +334,11 @@ trace_pattern(Process* p, Eterm MFA, Eterm Pattern, Eterm flaglist)
 	matches = erts_set_trace_pattern(p, mfa, specified,
 					 match_prog_set, match_prog_set,
 					 on, flags, meta_tracer, 0);
+    } else if (is_atom(MFA)) {
+        if (is_global || flags.breakpoint || on > ERTS_BREAK_SET) {
+            goto error;
+        }
+        matches = erts_set_tracing_event_pattern(MFA, match_prog_set, on);
     }
 
  error:
@@ -791,6 +816,8 @@ Eterm trace_info_2(BIF_ALIST_2)
 
     if (What == am_on_load) {
 	res = trace_info_on_load(p, Key);
+    } else if (What == am_send || What == am_receive) {
+        res = trace_info_event(p, What, Key);
     } else if (is_atom(What) || is_pid(What) || is_port(What)) {
 	res = trace_info_pid(p, What, Key);
     } else if (is_tuple(What)) {
@@ -1280,6 +1307,42 @@ trace_info_on_load(Process* p, Eterm key)
     }
 }
 
+static Eterm
+trace_info_event(Process* p, Eterm event, Eterm key)
+{
+    ErtsTracingEvent* te;
+    Eterm retval;
+    Eterm* hp;
+
+    switch (event) {
+    case am_send:    te = erts_send_tracing;    break;
+    case am_receive: te = erts_receive_tracing; break;
+    default:
+        goto error;
+    }
+
+    if (key != am_match_spec)
+        goto error;
+
+    te = &te[erts_active_bp_ix()];
+
+    if (te->on) {
+        if (!te->match_spec)
+            retval = am_true;
+        else
+            retval = copy_object(MatchSetGetSource(te->match_spec), p);
+    }
+    else
+        retval = am_false;
+
+    hp = HAlloc(p, 3);
+    return TUPLE2(hp, key, retval);
+
+ error:
+    BIF_ERROR(p, BADARG);
+}
+
+
 #undef FUNC_TRACE_NOEXIST
 #undef FUNC_TRACE_UNTRACED
 #undef FUNC_TRACE_GLOBAL_TRACE
@@ -1307,7 +1370,7 @@ erts_set_trace_pattern(Process*p, Eterm* mfa, int specified,
 
     for (i = 0; i < n; i++) {
 	BeamInstr* pc = fp[i].pc;
-	Export* ep = (Export *)(((char *)(pc-3)) - offsetof(Export, code));
+	Export* ep = ErtsContainerStruct(pc, Export, code[3]);
 
 	if (on && !flags.breakpoint) {
 	    /* Turn on global call tracing */
@@ -1468,12 +1531,57 @@ erts_set_trace_pattern(Process*p, Eterm* mfa, int specified,
 }
 
 int
+erts_set_tracing_event_pattern(Eterm event, Binary* match_spec, int on)
+{
+    ErtsBpIndex ix = erts_staging_bp_ix();
+    ErtsTracingEvent* st;
+
+    switch (event) {
+    case am_send: st = &erts_send_tracing[ix]; break;
+    case am_receive: st = &erts_receive_tracing[ix]; break;
+    default: return -1;
+    }
+
+    MatchSetUnref(st->match_spec);
+
+    st->on = on;
+    st->match_spec = match_spec;
+    MatchSetRef(match_spec);
+
+    finish_bp.current = 1;  /* prepare phase not needed for event trace */
+    finish_bp.install = on;
+    finish_bp.e.matched = 0;
+    finish_bp.e.matching = NULL;
+    finish_bp.f.matched = 0;
+    finish_bp.f.matching = NULL;
+
+#ifndef ERTS_SMP
+    while (erts_finish_breakpointing()) {
+        /* Empty loop body */
+    }
+#endif
+    return 1;
+}
+
+static void
+consolidate_event_tracing(ErtsTracingEvent te[])
+{
+    ErtsTracingEvent* src = &te[erts_active_bp_ix()];
+    ErtsTracingEvent* dst = &te[erts_staging_bp_ix()];
+
+    MatchSetUnref(dst->match_spec);
+    dst->on = src->on;
+    dst->match_spec = src->match_spec;
+    MatchSetRef(dst->match_spec);
+}
+
+int
 erts_finish_breakpointing(void)
 {
     ERTS_SMP_LC_ASSERT(erts_has_code_write_permission());
 
     /*
-     * Memory barriers will be issued for all processes *before*
+     * Memory barriers will be issued for all schedulers *before*
      * each of the stages below. (Unless the other schedulers
      * are blocked, in which case memory barriers will be issued
      * when they are awaken.)
@@ -1542,6 +1650,8 @@ erts_finish_breakpointing(void)
 	erts_consolidate_bp_data(&finish_bp.f, 1);
 	erts_bp_free_matched_functions(&finish_bp.e);
 	erts_bp_free_matched_functions(&finish_bp.f);
+        consolidate_event_tracing(erts_send_tracing);
+	consolidate_event_tracing(erts_receive_tracing);
 	return 0;
     default:
 	ASSERT(0);
@@ -1556,11 +1666,10 @@ install_exp_breakpoints(BpFunctions* f)
     BpFunction* fp = f->matching;
     Uint ne = f->matched;
     Uint i;
-    Uint offset = offsetof(Export, code) + 3*sizeof(BeamInstr);
 
     for (i = 0; i < ne; i++) {
 	BeamInstr* pc = fp[i].pc;
-	Export* ep = (Export *) (((char *)pc)-offset);
+	Export* ep = ErtsContainerStruct(pc, Export, code[3]);
 
 	ep->addressv[code_ix] = pc;
     }
@@ -1573,11 +1682,10 @@ uninstall_exp_breakpoints(BpFunctions* f)
     BpFunction* fp = f->matching;
     Uint ne = f->matched;
     Uint i;
-    Uint offset = offsetof(Export, code) + 3*sizeof(BeamInstr);
 
     for (i = 0; i < ne; i++) {
 	BeamInstr* pc = fp[i].pc;
-	Export* ep = (Export *) (((char *)pc)-offset);
+	Export* ep = ErtsContainerStruct(pc, Export, code[3]);
 
 	if (ep->addressv[code_ix] != pc) {
 	    continue;
@@ -1594,11 +1702,10 @@ clean_export_entries(BpFunctions* f)
     BpFunction* fp = f->matching;
     Uint ne = f->matched;
     Uint i;
-    Uint offset = offsetof(Export, code) + 3*sizeof(BeamInstr);
 
     for (i = 0; i < ne; i++) {
 	BeamInstr* pc = fp[i].pc;
-	Export* ep = (Export *) (((char *)pc)-offset);
+	Export* ep = ErtsContainerStruct(pc, Export, code[3]);
 
 	if (ep->addressv[code_ix] == pc) {
 	    continue;
@@ -2277,7 +2384,7 @@ 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.c b/erts/emulator/beam/erl_db.c
index 615d23402b..acaca54e9a 100644
--- a/erts/emulator/beam/erl_db.c
+++ b/erts/emulator/beam/erl_db.c
@@ -2833,7 +2833,8 @@ BIF_RETTYPE ets_match_spec_run_r_3(BIF_ALIST_3)
 	    BIF_TRAP3(bif_export[BIF_ets_match_spec_run_r_3],
 		      BIF_P,lst,BIF_ARG_2,ret);
 	}
-	res = db_prog_match(BIF_P, mp, CAR(list_val(lst)), NULL, 0,
+	res = db_prog_match(BIF_P, BIF_P,
+                            mp, CAR(list_val(lst)), NULL, 0,
 			    ERTS_PAM_COPY_RESULT, &dummy);
 	if (is_value(res)) {
 	    hp = HAlloc(BIF_P, 2);
diff --git a/erts/emulator/beam/erl_db_util.c b/erts/emulator/beam/erl_db_util.c
index 76b96637ae..95b1cd0148 100644
--- a/erts/emulator/beam/erl_db_util.c
+++ b/erts/emulator/beam/erl_db_util.c
@@ -123,6 +123,9 @@ do {									\
 
 #define TermWords(t) (((t) / (sizeof(UWord)/sizeof(Eterm))) + !!((t) % (sizeof(UWord)/sizeof(Eterm))))
 
+#define add_dmc_err(EINFO, STR, VAR, TERM, SEV) \
+       vadd_dmc_err(EINFO, SEV, VAR, STR, TERM)
+
 
 static ERTS_INLINE Process *
 get_proc(Process *cp, Uint32 cp_locks, Eterm id, Uint32 id_locks)
@@ -411,17 +414,27 @@ get_match_pseudo_process(Process *c_p, Uint heap_size)
 {
     ErtsMatchPseudoProcess *mpsp;
 #ifdef ERTS_SMP
-    mpsp = (ErtsMatchPseudoProcess *) c_p->scheduler_data->match_pseudo_process;
-    if (mpsp)
+    ErtsSchedulerData *esdp;
+
+    esdp = c_p ? c_p->scheduler_data : erts_get_scheduler_data();
+
+    mpsp = esdp ? esdp->match_pseudo_process :
+        (ErtsMatchPseudoProcess*) erts_smp_tsd_get(match_pseudo_process_key);
+
+    if (mpsp) {
+        ASSERT(mpsp == erts_smp_tsd_get(match_pseudo_process_key));
+        ASSERT(mpsp->process.scheduler_data == esdp);
 	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;
+        if (esdp) {
+            esdp->match_pseudo_process = (void *) mpsp;
+        }
+        mpsp->process.scheduler_data = esdp;
 	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;
 #else
     mpsp = match_pseudo_process;
     cleanup_match_pseudo_process(mpsp, 0);
@@ -889,11 +902,7 @@ void db_match_dis(Binary *prog);
 #define TRACE /* Nothing */
 #define FENCE_PATTERN_SIZE 0
 #endif
-static void add_dmc_err(DMCErrInfo *err_info, 
-			   char *str,
-			   int variable,
-			   Eterm term,
-			   DMCErrorSeverity severity);
+static void vadd_dmc_err(DMCErrInfo*, DMCErrorSeverity, int var, const char *str, ...);
 
 static Eterm dpm_array_to_list(Process *psp, Eterm *arr, int arity);
 
@@ -989,12 +998,20 @@ Eterm erts_match_set_get_source(Binary *mpsp)
 }
 
 /* This one is for the tracing */
-Binary *erts_match_set_compile(Process *p, Eterm matchexpr) {
+Binary *erts_match_set_compile(Process *p, Eterm matchexpr, Eterm MFA) {
     Binary *bin;
     Uint sz;
     Eterm *hp;
+    Uint flags;
+
+    switch (MFA) {
+    case am_receive: flags = DCOMP_TRACE; break;
+    case am_send:    flags = DCOMP_TRACE | DCOMP_ALLOW_TRACE_OPS; break;
+    default:
+        flags = DCOMP_TRACE | DCOMP_CALL_TRACE | DCOMP_ALLOW_TRACE_OPS;
+    }
     
-    bin = db_match_set_compile(p, matchexpr, DCOMP_TRACE);
+    bin = db_match_set_compile(p, matchexpr, flags);
     if (bin != NULL) {
 	MatchProg *prog = Binary2MatchProg(bin);
 	sz = size_object(matchexpr);
@@ -1124,8 +1141,8 @@ Eterm db_match_set_lint(Process *p, Eterm matchexpr, Uint flags)
     int i;
 
     if (!is_list(matchexpr)) {
-	add_dmc_err(err_info, "Match programs are not in a list.", 
-		    -1, 0UL, dmcError);
+	add_dmc_err(err_info, "Match programs are not in a list.",
+                    -1, 0UL, dmcError);
 	goto done;
     }
     num_heads = 0;
@@ -1133,9 +1150,8 @@ Eterm db_match_set_lint(Process *p, Eterm matchexpr, Uint flags)
 	++num_heads;
 
     if (l != NIL)  { /* proper list... */
-	add_dmc_err(err_info, "Match programs are not in a proper "
-		    "list.", 
-		    -1, 0UL, dmcError);
+	add_dmc_err(err_info, "Match programs are not in a proper list.",
+                     -1, 0UL, dmcError);
 	goto done;
     }
 
@@ -1202,30 +1218,37 @@ done:
     return ret;
 }
     
-Eterm erts_match_set_run(Process *p, Binary *mpsp, 
-			 Eterm *args, int num_args,
-			 enum erts_pam_run_flags in_flags,
-			 Uint32 *return_flags) 
+/* Returns
+ *   am_false      if no match or
+ *                 if {message,false} has been called,
+ *   am_true       if {message,_} has NOT been called or
+ *                 if {message,true} has been called,
+ *   Msg           if {message,Msg} has been called.
+ *
+ *   If return value is_not_immed
+ *   then erts_match_set_release_result_trace() must be called to release it.
+ */
+Eterm erts_match_set_run_trace(Process *c_p,
+                               Process *self,
+                               Binary *mpsp,
+                               Eterm *args, int num_args,
+                               enum erts_pam_run_flags in_flags,
+                               Uint32 *return_flags)
 {
     Eterm ret;
 
-    ret = db_prog_match(p, mpsp, NIL, args, num_args,
+    ret = db_prog_match(c_p, self, mpsp, NIL, args, num_args,
 			in_flags, return_flags);
-#if defined(HARDDEBUG)
-    if (is_non_value(ret)) {
-	erts_fprintf(stderr, "Failed\n");
-    } else {
-	erts_fprintf(stderr, "Returning : %T\n", ret);
+
+    ASSERT(!(is_non_value(ret) && *return_flags));
+
+    if (is_non_value(ret) || ret == am_false) {
+        erts_match_set_release_result(c_p);
+        return am_false;
     }
-#endif
+    if (is_immed(ret))
+        erts_match_set_release_result(c_p);
     return ret;
-    /* Returns 
-     *   THE_NON_VALUE if no match
-     *   am_false      if {message,false} has been called,
-     *   am_true       if {message,_} has not been called or
-     *                 if {message,true} has been called,
-     *   Msg           if {message,Msg} has been called.
-     */
 }
 
 static Eterm erts_match_set_run_ets(Process *p, Binary *mpsp,
@@ -1234,7 +1257,8 @@ static Eterm erts_match_set_run_ets(Process *p, Binary *mpsp,
 {
     Eterm ret;
 
-    ret = db_prog_match(p, mpsp, args, NULL, num_args,
+    ret = db_prog_match(p, p,
+                        mpsp, args, NULL, num_args,
 			ERTS_PAM_COPY_RESULT,
 			return_flags);
 #if defined(HARDDEBUG)
@@ -1730,7 +1754,9 @@ static Eterm dpm_array_to_list(Process *psp, Eterm *arr, int arity)
 ** the parameter 'arity' is only used if 'term' is actually an array,
 ** i.e. 'DCOMP_TRACE' was specified 
 */
-Eterm db_prog_match(Process *c_p, Binary *bprog,
+Eterm db_prog_match(Process *c_p,
+                    Process *self,
+                    Binary *bprog,
 		    Eterm term,
 		    Eterm *termp,
 		    int arity,
@@ -1743,10 +1769,10 @@ Eterm db_prog_match(Process *c_p, Binary *bprog,
     Eterm *esp;
     MatchVariable* variables;
     BeamInstr *cp;
-    UWord *pc = prog->text;
+    const UWord *pc = prog->text;
     Eterm *ehp;
     Eterm ret;
-    Uint n = 0; /* To avoid warning. */
+    Uint n;
     int i;
     unsigned do_catch;
     ErtsMatchPseudoProcess *mpsp;
@@ -1758,46 +1784,28 @@ Eterm db_prog_match(Process *c_p, Binary *bprog,
     Eterm (*bif)(Process*, ...);
     Eterm bif_args[3];
     int fail_label;
-    int atomic_trace;
 #ifdef DMC_DEBUG
     Uint *heap_fence;
     Uint *stack_fence;
     Uint save_op;
 #endif /* DMC_DEBUG */
 
+    ERTS_UNDEF(n,0);
+    ERTS_UNDEF(current_scheduled,NULL);
+
+    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);
-    ASSERT(esdp != NULL);
-    current_scheduled = esdp->current_process;
+    esdp = ERTS_GET_SCHEDULER_DATA_FROM_PROC(psp);
+    if (esdp)
+        current_scheduled = esdp->current_process;
     /* SMP: psp->scheduler_data is set by get_match_pseudo_process */
 
-    atomic_trace = 0;
-#define BEGIN_ATOMIC_TRACE(p)                               \
-    do {                                                    \
-	if (! atomic_trace) {                               \
-            erts_refc_inc(&bprog->refc, 2);                 \
-	    erts_smp_proc_unlock((p), ERTS_PROC_LOCK_MAIN); \
-	    erts_smp_thr_progress_block();                  \
-            atomic_trace = !0;                              \
-	}                                                   \
-    } while (0)
-#define END_ATOMIC_TRACE(p)                               \
-    do {                                                  \
-	if (atomic_trace) {                               \
-            erts_smp_thr_progress_unblock();              \
-            erts_smp_proc_lock((p), ERTS_PROC_LOCK_MAIN); \
-            if (erts_refc_dectest(&bprog->refc, 0) == 0) {\
-                erts_bin_free(bprog);                     \
-	    }                                             \
-            atomic_trace = 0;                             \
-	}                                                 \
-    } while (0)
-
 #ifdef DMC_DEBUG
     save_op = 0;
     heap_fence = (Eterm*)((char*) mpsp->u.heap + prog->stack_offset) - 1;
@@ -2256,7 +2264,7 @@ restart:
 	    pc += n;
 	    break;
 	case matchSelf:
-	    *esp++ = c_p->common.id;
+	    *esp++ = self->common.id;
 	    break;
 	case matchWaste:
 	    --esp;
@@ -2266,6 +2274,7 @@ restart:
 	    break;
 	case matchProcessDump: {
 	    erts_dsprintf_buf_t *dsbufp = erts_create_tmp_dsbuf(0);
+            ASSERT(c_p == self);
 	    print_process_info(ERTS_PRINT_DSBUF, (void *) dsbufp, c_p);
 	    *esp++ = new_binary(build_proc, (byte *)dsbufp->str,
 				dsbufp->str_len);
@@ -2284,14 +2293,16 @@ restart:
 	    *return_flags |= MATCH_SET_EXCEPTION_TRACE;
 	    *esp++ = am_true;
 	    break;
-	case matchIsSeqTrace:
+        case matchIsSeqTrace:
+            ASSERT(c_p == self);
             if (have_seqtrace(SEQ_TRACE_TOKEN(c_p)))
 		*esp++ = am_true;
 	    else
 		*esp++ = am_false;
 	    break;
 	case matchSetSeqToken:
-	    t = erts_seq_trace(c_p, esp[-1], esp[-2], 0);
+            ASSERT(c_p == self);
+            t = erts_seq_trace(c_p, esp[-1], esp[-2], 0);
 	    if (is_non_value(t)) {
 		esp[-2] = FAIL_TERM;
 	    } else {
@@ -2299,7 +2310,8 @@ restart:
 	    }
 	    --esp;
 	    break;
-	case matchSetSeqTokenFake:
+        case matchSetSeqTokenFake:
+            ASSERT(c_p == self);
 	    t = seq_trace_fake(c_p, esp[-1]);
 	    if (is_non_value(t)) {
 		esp[-2] = FAIL_TERM;
@@ -2308,7 +2320,8 @@ restart:
 	    }
 	    --esp;
 	    break;
-	case matchGetSeqToken:
+        case matchGetSeqToken:
+            ASSERT(c_p == self);
             if (have_no_seqtrace(SEQ_TRACE_TOKEN(c_p)))
 		*esp++ = NIL;
 	    else {
@@ -2332,49 +2345,62 @@ restart:
 		ASSERT(is_immed(ehp[5]));
 	    } 
 	    break;
-	case matchEnableTrace:
+        case matchEnableTrace:
+            ASSERT(c_p == self);
 	    if ( (n = erts_trace_flag2bit(esp[-1]))) {
-		BEGIN_ATOMIC_TRACE(c_p);
+                erts_smp_proc_lock(c_p, ERTS_PROC_LOCKS_ALL_MINOR);
 		set_tracee_flags(c_p, ERTS_TRACER(c_p), 0, n);
+                erts_smp_proc_unlock(c_p, ERTS_PROC_LOCKS_ALL_MINOR);
 		esp[-1] = am_true;
 	    } else {
 		esp[-1] = FAIL_TERM;
 	    }
 	    break;
-	case matchEnableTrace2:
+        case matchEnableTrace2:
+            ASSERT(c_p == self);
 	    n = erts_trace_flag2bit((--esp)[-1]);
 	    esp[-1] = FAIL_TERM;
 	    if (n) {
-		BEGIN_ATOMIC_TRACE(c_p);
-		if ( (tmpp = get_proc(c_p, 0, esp[0], 0))) {
+		if ( (tmpp = get_proc(c_p, ERTS_PROC_LOCK_MAIN, esp[0], ERTS_PROC_LOCKS_ALL))) {
 		    /* Always take over the tracer of the current process */
 		    set_tracee_flags(tmpp, ERTS_TRACER(c_p), 0, n);
-		    esp[-1] = am_true;
+                    if (tmpp == c_p)
+                        erts_smp_proc_unlock(tmpp, ERTS_PROC_LOCKS_ALL_MINOR);
+                    else
+                        erts_smp_proc_unlock(tmpp, ERTS_PROC_LOCKS_ALL);
+                    esp[-1] = am_true;
 		}
 	    }
 	    break;
-	case matchDisableTrace:
+        case matchDisableTrace:
+            ASSERT(c_p == self);
 	    if ( (n = erts_trace_flag2bit(esp[-1]))) {
-		BEGIN_ATOMIC_TRACE(c_p);
+                erts_smp_proc_lock(c_p, ERTS_PROC_LOCKS_ALL_MINOR);
 		set_tracee_flags(c_p, ERTS_TRACER(c_p), n, 0);
+                erts_smp_proc_unlock(c_p, ERTS_PROC_LOCKS_ALL_MINOR);
 		esp[-1] = am_true;
 	    } else {
 		esp[-1] = FAIL_TERM;
 	    }
 	    break;
-	case matchDisableTrace2:
+        case matchDisableTrace2:
+            ASSERT(c_p == self);
 	    n = erts_trace_flag2bit((--esp)[-1]);
 	    esp[-1] = FAIL_TERM;
 	    if (n) {
-		BEGIN_ATOMIC_TRACE(c_p);
-		if ( (tmpp = get_proc(c_p, 0, esp[0], 0))) {
+		if ( (tmpp = get_proc(c_p, ERTS_PROC_LOCK_MAIN, esp[0], ERTS_PROC_LOCKS_ALL))) {
 		    /* Always take over the tracer of the current process */
 		    set_tracee_flags(tmpp, ERTS_TRACER(c_p), n, 0);
-		    esp[-1] = am_true;
+                    if (tmpp == c_p)
+                        erts_smp_proc_unlock(tmpp, ERTS_PROC_LOCKS_ALL_MINOR);
+                    else
+                        erts_smp_proc_unlock(tmpp, ERTS_PROC_LOCKS_ALL);
+                    esp[-1] = am_true;
 		}
 	    }
 	    break;
- 	case matchCaller:
+        case matchCaller:
+            ASSERT(c_p == self);
 	    if (!(c_p->cp) || !(cp = find_function_from_pc(c_p->cp))) {
  		*esp++ = am_undefined;
  	    } else {
@@ -2386,7 +2412,8 @@ restart:
 		ehp[3] = make_small((Uint) cp[2]);
 	    }
 	    break;
-	case matchSilent:
+        case matchSilent:
+            ASSERT(c_p == self);
 	    --esp;
 	    if (in_flags & ERTS_PAM_IGNORE_TRACE_SILENT)
 	      break;
@@ -2401,7 +2428,8 @@ restart:
 		erts_smp_proc_unlock(c_p, ERTS_PROC_LOCKS_ALL_MINOR);
 	    }
 	    break;
-	case matchTrace2:
+        case matchTrace2:
+            ASSERT(c_p == self);
 	    {
 		/*    disable         enable                                */
 		Uint  d_flags  = 0,   e_flags  = 0;  /* process trace flags */
@@ -2433,7 +2461,8 @@ restart:
                 ERTS_TRACER_CLEAR(&tracer);
 	    }
 	    break;
-	case matchTrace3:
+        case matchTrace3:
+            ASSERT(c_p == self);
 	    {
 		/*    disable         enable                                */
 		Uint  d_flags  = 0,   e_flags  = 0;  /* process trace flags */
@@ -2506,15 +2535,12 @@ success:
     }
 #endif
 
-    esdp->current_process = current_scheduled;
-
-    END_ATOMIC_TRACE(c_p);
+    if (esdp)
+        esdp->current_process = current_scheduled;
 
     return ret;
 #undef FAIL
 #undef FAIL_TERM
-#undef BEGIN_ATOMIC_TRACE
-#undef END_ATOMIC_TRACE
 }
 
 
@@ -3259,20 +3285,20 @@ int erts_db_is_compiled_ms(Eterm term)
 ** Utility to add an error
 */
 
-static void add_dmc_err(DMCErrInfo *err_info, 
-			char *str,
-			int variable,
-			Eterm term,
-			DMCErrorSeverity severity)
+static void vadd_dmc_err(DMCErrInfo *err_info,
+                         DMCErrorSeverity severity,
+                         int variable,
+                         const char *str,
+                         ...)
 {
+    DMCError *e;
+    va_list args;
+    va_start(args, str);
+
+
     /* Linked in in reverse order, to ease the formatting */
-    DMCError *e = erts_alloc(ERTS_ALC_T_DB_DMC_ERROR, sizeof(DMCError));
-    if (term != 0UL) {
-	erts_snprintf(e->error_string, DMC_ERR_STR_LEN, str, term);
-    } else {
-	strncpy(e->error_string, str, DMC_ERR_STR_LEN);
-	e->error_string[DMC_ERR_STR_LEN] ='\0';
-    }
+    e = erts_alloc(ERTS_ALC_T_DB_DMC_ERROR, sizeof(DMCError));
+    erts_vsnprintf(e->error_string, DMC_ERR_STR_LEN, str, args);
     e->variable = variable;
     e->severity = severity;
     e->next = err_info->first;
@@ -3282,8 +3308,11 @@ static void add_dmc_err(DMCErrInfo *err_info,
     err_info->first = e;
     if (severity >= dmcError)
 	err_info->error_added = 1;
+
+    va_end(args);
 }
     
+
 /*
 ** Handle one term in the match expression (not the guard) 
 */
@@ -3482,24 +3511,21 @@ static void do_emit_constant(DMCContext *context, DMC_STACK_TYPE(UWord) *text,
 	    context->stack_need = context->stack_used;
 }
 
-#define RETURN_ERROR_X(String, X, Y, ContextP, ConstantF)        \
-do {                                                            \
-if ((ContextP)->err_info != NULL) {				\
-    (ConstantF) = 0;						\
-    add_dmc_err((ContextP)->err_info, String, X, Y, dmcError);  \
-    return retOk;						\
-} else 								\
-  return retFail;                                                \
-} while(0)
+#define RETURN_ERROR_X(VAR, ContextP, ConstantF, String, ARG)            \
+    (((ContextP)->err_info != NULL)				         \
+     ? ((ConstantF) = 0,						 \
+        vadd_dmc_err((ContextP)->err_info, dmcError, VAR, String, ARG),  \
+        retOk)						                 \
+     : retFail)
 
 #define RETURN_ERROR(String, ContextP, ConstantF) \
-     RETURN_ERROR_X(String, -1, 0UL, ContextP, ConstantF)
+     return RETURN_ERROR_X(-1, ContextP, ConstantF, String, 0)
 
 #define RETURN_VAR_ERROR(String, N, ContextP, ConstantF) \
-     RETURN_ERROR_X(String, N, 0UL, ContextP, ConstantF)
+     return RETURN_ERROR_X(N, ContextP, ConstantF, String, 0)
 
 #define RETURN_TERM_ERROR(String, T, ContextP, ConstantF) \
-     RETURN_ERROR_X(String, -1, T, ContextP, ConstantF)
+     return RETURN_ERROR_X(-1, ContextP, ConstantF, String, T)
 
 #define WARNING(String, ContextP) \
 add_dmc_err((ContextP)->err_info, String, -1, 0UL, dmcWarning)
@@ -3765,7 +3791,7 @@ static DMCRet dmc_variable(DMCContext *context,
     Uint n = db_is_variable(t);
 
     if (n >= heap->vars_used || !heap->vars[n].is_bound) {
-	RETURN_VAR_ERROR("Variable $%d is unbound.", n, context, *constant);
+	RETURN_VAR_ERROR("Variable $%%d is unbound.", n, context, *constant);
     }
 
     dmc_add_pushv_variant(context, heap, text, n);
@@ -4097,7 +4123,30 @@ static DMCRet dmc_exception_trace(DMCContext *context,
     return retOk;
 }
 
-
+static int check_trace(const char* op,
+                       DMCContext *context,
+                       int *constant,
+                       int need_cflags,
+                       int allow_in_guard,
+                       DMCRet* retp)
+{
+    if (!(context->cflags & DCOMP_TRACE)) {
+	*retp = RETURN_ERROR_X(-1, context, *constant, "Special form '%s' "
+                               "used in wrong dialect.", op);
+        return 0;
+    }
+    if ((context->cflags & need_cflags) != need_cflags) {
+        *retp = RETURN_ERROR_X(-1, context, *constant, "Special form '%s' "
+                               "not allow for this trace event.", op);
+        return 0;
+    }
+    if (context->is_guard && !allow_in_guard) {
+        *retp = RETURN_ERROR_X(-1, context, *constant, "Special form '%s' "
+                               "called in guard context.", op);
+        return 0;
+    }
+    return 1;
+}
 
 static DMCRet dmc_is_seq_trace(DMCContext *context,
 			       DMCHeap *heap,
@@ -4107,12 +4156,11 @@ static DMCRet dmc_is_seq_trace(DMCContext *context,
 {
     Eterm *p = tuple_val(t);
     Uint a = arityval(*p);
+    DMCRet ret;
     
-    if (!(context->cflags & DCOMP_TRACE)) {
-	RETURN_ERROR("Special form 'is_seq_trace' used in wrong dialect.",
-		     context, 
-		     *constant);
-    }
+    if (!check_trace("is_seq_trace", context, constant, DCOMP_ALLOW_TRACE_OPS, 1, &ret))
+        return ret;
+
     if (a != 1) {
 	RETURN_TERM_ERROR("Special form 'is_seq_trace' called with "
 			  "arguments in %T.", t, context, *constant);
@@ -4136,16 +4184,8 @@ static DMCRet dmc_set_seq_token(DMCContext *context,
     DMCRet ret;
     int c;
     
-
-    if (!(context->cflags & DCOMP_TRACE)) {
-	RETURN_ERROR("Special form 'set_seq_token' used in wrong dialect.",
-		     context, 
-		     *constant);
-    }
-    if (context->is_guard) {
-	RETURN_ERROR("Special form 'set_seq_token' called in "
-		     "guard context.", context, *constant);
-    }
+    if (!check_trace("set_seq_trace", context, constant, DCOMP_ALLOW_TRACE_OPS, 0, &ret))
+        return ret;
 
     if (a != 3) {
 	RETURN_TERM_ERROR("Special form 'set_seq_token' called with wrong "
@@ -4182,16 +4222,11 @@ static DMCRet dmc_get_seq_token(DMCContext *context,
 {
     Eterm *p = tuple_val(t);
     Uint a = arityval(*p);
+    DMCRet ret;
+
+    if (!check_trace("get_seq_token", context, constant, DCOMP_ALLOW_TRACE_OPS, 0, &ret))
+        return ret;
 
-    if (!(context->cflags & DCOMP_TRACE)) {
-	RETURN_ERROR("Special form 'get_seq_token' used in wrong dialect.",
-		     context, 
-		     *constant);
-    }
-    if (context->is_guard) {
-	RETURN_ERROR("Special form 'get_seq_token' called in "
-		     "guard context.", context, *constant);
-    }
     if (a != 1) {
 	RETURN_TERM_ERROR("Special form 'get_seq_token' called with "
 			  "arguments in %T.", t, context, 
@@ -4255,16 +4290,10 @@ static DMCRet dmc_process_dump(DMCContext *context,
 {
     Eterm *p = tuple_val(t);
     Uint a = arityval(*p);
-    
-    if (!(context->cflags & DCOMP_TRACE)) {
-	RETURN_ERROR("Special form 'process_dump' used in wrong dialect.",
-		     context, 
-		     *constant);
-    }
-    if (context->is_guard) {
-	RETURN_ERROR("Special form 'process_dump' called in "
-		     "guard context.", context, *constant);
-    }
+    DMCRet ret;
+
+    if (!check_trace("process_dump", context, constant, DCOMP_ALLOW_TRACE_OPS, 0, &ret))
+        return ret;
 
     if (a != 1) {
 	RETURN_TERM_ERROR("Special form 'process_dump' called with "
@@ -4288,17 +4317,8 @@ static DMCRet dmc_enable_trace(DMCContext *context,
     DMCRet ret;
     int c;
     
-
-    if (!(context->cflags & DCOMP_TRACE)) {
-	RETURN_ERROR("Special form 'enable_trace' used in wrong dialect.",
-		     context, 
-		     *constant);
-    }
-    if (context->is_guard) {
-	RETURN_ERROR("Special form 'enable_trace' called in guard context.",
-		     context, 
-		     *constant);
-    }
+    if (!check_trace("enable_trace", context, constant, DCOMP_ALLOW_TRACE_OPS, 0, &ret))
+        return ret;
 
     switch (a) {
     case 2:
@@ -4347,18 +4367,9 @@ static DMCRet dmc_disable_trace(DMCContext *context,
     Uint a = arityval(*p);
     DMCRet ret;
     int c;
-    
 
-    if (!(context->cflags & DCOMP_TRACE)) {
-	RETURN_ERROR("Special form 'disable_trace' used in wrong dialect.",
-		     context, 
-		     *constant);
-    }
-    if (context->is_guard) {
-	RETURN_ERROR("Special form 'disable_trace' called in guard context.",
-		     context, 
-		     *constant);
-    }
+    if (!check_trace("disable_trace", context, constant, DCOMP_ALLOW_TRACE_OPS, 0, &ret))
+        return ret;
 
     switch (a) {
     case 2:
@@ -4408,17 +4419,8 @@ static DMCRet dmc_trace(DMCContext *context,
     DMCRet ret;
     int c;
     
-
-    if (!(context->cflags & DCOMP_TRACE)) {
-	RETURN_ERROR("Special form 'trace' used in wrong dialect.",
-		     context, 
-		     *constant);
-    }
-    if (context->is_guard) {
-	RETURN_ERROR("Special form 'trace' called in guard context.",
-		     context, 
-		     *constant);
-    }
+    if (!check_trace("trace", context, constant, DCOMP_ALLOW_TRACE_OPS, 0, &ret))
+        return ret;
 
     switch (a) {
     case 3:
@@ -4479,16 +4481,11 @@ static DMCRet dmc_caller(DMCContext *context,
 {
     Eterm *p = tuple_val(t);
     Uint a = arityval(*p);
+    DMCRet ret;
      
-    if (!(context->cflags & DCOMP_TRACE)) {
-	RETURN_ERROR("Special form 'caller' used in wrong dialect.",
-		     context, 
-		     *constant);
-    }
-    if (context->is_guard) {
- 	RETURN_ERROR("Special form 'caller' called in "
- 		     "guard context.", context, *constant);
-    }
+    if (!check_trace("caller", context, constant,
+                     (DCOMP_CALL_TRACE|DCOMP_ALLOW_TRACE_OPS), 0, &ret))
+        return ret;
   
     if (a != 1) {
  	RETURN_TERM_ERROR("Special form 'caller' called with "
@@ -4514,15 +4511,8 @@ static DMCRet dmc_silent(DMCContext *context,
     DMCRet ret;
     int c;
      
-    if (!(context->cflags & DCOMP_TRACE)) {
-	RETURN_ERROR("Special form 'silent' used in wrong dialect.",
-		     context, 
-		     *constant);
-    }
-    if (context->is_guard) {
- 	RETURN_ERROR("Special form 'silent' called in "
- 		     "guard context.", context, *constant);
-    }
+    if (!check_trace("silent", context, constant, DCOMP_ALLOW_TRACE_OPS, 0, &ret))
+        return ret;
   
     if (a != 2) {
 	RETURN_TERM_ERROR("Special form 'silent' called with wrong "
@@ -5062,11 +5052,14 @@ static Eterm match_spec_test(Process *p, Eterm against, Eterm spec, int trace)
 	return THE_NON_VALUE;
     }
     if (trace) {
-	lint_res = db_match_set_lint(p, spec, DCOMP_TRACE | DCOMP_FAKE_DESTRUCTIVE);
-	mps = db_match_set_compile(p, spec, DCOMP_TRACE | DCOMP_FAKE_DESTRUCTIVE);
+        const Uint cflags = (DCOMP_TRACE | DCOMP_FAKE_DESTRUCTIVE |
+                             DCOMP_CALL_TRACE | DCOMP_ALLOW_TRACE_OPS);
+	lint_res = db_match_set_lint(p, spec, cflags);
+	mps = db_match_set_compile(p, spec, cflags);
     } else {
-	lint_res = db_match_set_lint(p, spec, DCOMP_TABLE | DCOMP_FAKE_DESTRUCTIVE);
-	mps = db_match_set_compile(p, spec, DCOMP_TABLE | DCOMP_FAKE_DESTRUCTIVE);
+        const Uint cflags = (DCOMP_TABLE | DCOMP_FAKE_DESTRUCTIVE);
+	lint_res = db_match_set_lint(p, spec, cflags);
+	mps = db_match_set_compile(p, spec, cflags);
     }
 
     if (mps == NULL) {
@@ -5099,7 +5092,8 @@ static Eterm match_spec_test(Process *p, Eterm against, Eterm spec, int trace)
 	    }
 	    save_cp = p->cp;
 	    p->cp = NULL;
-	    res = erts_match_set_run(p, mps, arr, n,
+	    res = erts_match_set_run_trace(p, p,
+                      mps, arr, n,
 		      ERTS_PAM_COPY_RESULT|ERTS_PAM_IGNORE_TRACE_SILENT,
 		      &ret_flags);
 	    p->cp = save_cp;
@@ -5182,7 +5176,8 @@ Eterm db_match_dbterm(DbTableCommon* tb, Process* c_p, Binary* bprog,
 	obj = db_alloc_tmp_uncompressed(tb, obj);
     }
 
-    res = db_prog_match(c_p, bprog, make_tuple(obj->tpl), NULL, 0,
+    res = db_prog_match(c_p, c_p,
+                        bprog, make_tuple(obj->tpl), NULL, 0,
 			ERTS_PAM_COPY_RESULT|ERTS_PAM_CONTIGUOUS_TUPLE, &dummy);
 
     if (is_value(res) && hpp!=NULL) {
diff --git a/erts/emulator/beam/erl_db_util.h b/erts/emulator/beam/erl_db_util.h
index 5d7946a525..60f7067d70 100644
--- a/erts/emulator/beam/erl_db_util.h
+++ b/erts/emulator/beam/erl_db_util.h
@@ -425,6 +425,11 @@ typedef struct dmc_err_info {
 #define DCOMP_FAKE_DESTRUCTIVE ((Uint) 8) /* When this is active, no setting of
 					     trace control words or seq_trace tokens will be done. */
 
+/* Allow lock seizing operations on the tracee and 3rd party processes */
+#define DCOMP_ALLOW_TRACE_OPS ((Uint) 0x10)
+
+/* This is call trace */
+#define DCOMP_CALL_TRACE ((Uint) 0x20)
 
 Binary *db_match_compile(Eterm *matchexpr, Eterm *guards,
 			 Eterm *body, int num_matches, 
@@ -435,7 +440,8 @@ Binary *db_match_compile(Eterm *matchexpr, Eterm *guards,
 Eterm db_match_dbterm(DbTableCommon* tb, Process* c_p, Binary* bprog,
 		      int all, DbTerm* obj, Eterm** hpp, Uint extra);
 
-Eterm db_prog_match(Process *p, Binary *prog, Eterm term,
+Eterm db_prog_match(Process *p, Process *self,
+                    Binary *prog, Eterm term,
 		    Eterm *termp, int arity,
 		    enum erts_pam_run_flags in_flags,
 		    Uint32 *return_flags /* Zeroed on enter */);
diff --git a/erts/emulator/beam/erl_gc.c b/erts/emulator/beam/erl_gc.c
index e21fba6bf1..bed7e668d7 100644
--- a/erts/emulator/beam/erl_gc.c
+++ b/erts/emulator/beam/erl_gc.c
@@ -2938,7 +2938,7 @@ 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 9beff52835..9bb6e40a11 100644
--- a/erts/emulator/beam/erl_message.c
+++ b/erts/emulator/beam/erl_message.c
@@ -32,6 +32,7 @@
 #include "erl_process.h"
 #include "erl_binary.h"
 #include "dtrace-wrapper.h"
+#include "beam_bp.h"
 
 ERTS_SCHED_PREF_QUICK_ALLOC_IMPL(message_ref,
 				 ErtsMessageRef,
@@ -251,9 +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
@@ -265,7 +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;
@@ -280,10 +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;
 	    }
@@ -293,7 +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);
@@ -301,10 +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);
@@ -324,7 +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 {
@@ -351,12 +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
@@ -366,14 +371,15 @@ 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,
+               ErtsProcLocks receiver_locks,
                ErtsMessage* first,
                ErtsMessage** last,
-               Uint len)
+               Uint len,
+               Eterm from)
 {
+    ErtsTracingEvent* te;
     Sint res;
     int locked_msgq = 0;
     erts_aint32_t state;
@@ -386,12 +392,12 @@ queue_messages(Process *c_p,
 #ifdef ERTS_SMP
 #ifdef ERTS_ENABLE_LOCK_CHECK
     ERTS_SMP_LC_ASSERT(erts_proc_lc_my_proc_locks(receiver) < ERTS_PROC_LOCK_MSGQ ||
-                       *receiver_locks == erts_proc_lc_my_proc_locks(receiver));
+                       receiver_locks == erts_proc_lc_my_proc_locks(receiver));
 #endif
 
-    if (!(*receiver_locks & ERTS_PROC_LOCK_MSGQ)) {
+    if (!(receiver_locks & ERTS_PROC_LOCK_MSGQ)) {
 	if (erts_smp_proc_trylock(receiver, ERTS_PROC_LOCK_MSGQ) == EBUSY) {
-            ErtsProcLocks need_locks = ERTS_PROC_LOCK_MSGQ;
+            ErtsProcLocks need_locks;
 
 	    if (receiver_state)
 		state = *receiver_state;
@@ -400,10 +406,11 @@ queue_messages(Process *c_p,
 	    if (state & (ERTS_PSFLG_EXITING|ERTS_PSFLG_PENDING_EXIT))
 		goto exiting;
 
-	    if (*receiver_locks & ERTS_PROC_LOCK_STATUS) {
-		erts_smp_proc_unlock(receiver, ERTS_PROC_LOCK_STATUS);
-		need_locks |= ERTS_PROC_LOCK_STATUS;
+            need_locks = receiver_locks & (ERTS_PROC_LOCK_STATUS|ERTS_PROC_LOCK_TRACE);
+	    if (need_locks) {
+		erts_smp_proc_unlock(receiver, need_locks);
 	    }
+            need_locks |= ERTS_PROC_LOCK_MSGQ;
 	    erts_smp_proc_lock(receiver, need_locks);
 	}
 	locked_msgq = 1;
@@ -426,7 +433,7 @@ queue_messages(Process *c_p,
 
     res = receiver->msg.len;
 #ifdef ERTS_SMP
-    if (*receiver_locks & ERTS_PROC_LOCK_MAIN) {
+    if (receiver_locks & ERTS_PROC_LOCK_MAIN) {
 	/*
 	 * We move 'in queue' to 'private queue' and place
 	 * message at the end of 'private queue' in order
@@ -445,7 +452,10 @@ queue_messages(Process *c_p,
 	LINK_MESSAGE(receiver, first, last, len);
     }
 
-    if (IS_TRACED_FL(receiver, F_TRACE_RECEIVE)) {
+    if (IS_TRACED_FL(receiver, F_TRACE_RECEIVE)
+        && (te = &erts_receive_tracing[erts_active_bp_ix()],
+            te->on)) {
+
         ErtsMessage *msg = first;
 
 #ifdef USE_VM_PROBES
@@ -468,19 +478,18 @@ queue_messages(Process *c_p,
                     tok_label, tok_lastcnt, tok_serial);
         }
 #endif
-
         while (msg) {
-            trace_receive(receiver, ERL_MESSAGE_TERM(msg));
+            trace_receive(receiver, from, ERL_MESSAGE_TERM(msg), te);
             msg = msg->next;
         }
 
     }
-
-    if (locked_msgq)
+    if (locked_msgq) {
 	erts_smp_proc_unlock(receiver, ERTS_PROC_LOCK_MSGQ);
+    }
 
 #ifdef ERTS_SMP
-    erts_proc_notify_new_message(receiver, *receiver_locks);
+    erts_proc_notify_new_message(receiver, receiver_locks);
 #else
     erts_proc_notify_new_message(receiver, 0);
     ERTS_HOLE_CHECK(receiver);
@@ -489,31 +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
@@ -830,11 +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);
     
@@ -891,7 +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
@@ -913,7 +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 dc83a780a2..a695a028ba 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 bc32f3f167..36b0af0c1a 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;
@@ -10024,7 +10024,7 @@ 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;
@@ -11766,7 +11766,7 @@ 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;
@@ -11787,7 +11787,7 @@ 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 436b4aca21..1c0fc0a11f 100644
--- a/erts/emulator/beam/erl_trace.c
+++ b/erts/emulator/beam/erl_trace.c
@@ -238,7 +238,6 @@ write_timestamp(ErtsTraceTimeStamp *tsp, Eterm **hpp)
 }
 
 #ifdef ERTS_SMP
-#define PATCH_TS_SIZE(p) patch_ts_size(TFLGS_TS_TYPE(p))
 
 static ERTS_INLINE Uint
 patch_ts_size(int ts_type)
@@ -258,7 +257,7 @@ patch_ts_size(int ts_type)
 	return 0;
     }
 }
-#endif
+#endif /* ERTS_SMP */
 
 /*
  * Write a timestamp. The timestamp MUST be the last
@@ -394,7 +393,8 @@ static ERTS_INLINE int
 send_to_tracer_nif(Process *c_p, ErtsPTabElementCommon *t_p,
                    Eterm t_p_id, ErtsTracerNif *tnif,
                    enum ErtsTracerOpt topt,
-                   Eterm tag, Eterm msg, Eterm extra);
+                   Eterm tag, Eterm msg, Eterm extra,
+                   Eterm pam_result);
 static ERTS_INLINE Eterm
 call_enabled_tracer(Process *c_p, const ErtsTracer tracer,
                     ErtsTracerNif **tnif_ref,
@@ -739,7 +739,7 @@ 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);
     }
 }
 
@@ -787,7 +787,7 @@ trace_sched_aux(Process *p, ErtsProcLocks locks, Eterm what)
     }
 
     send_to_tracer_nif(p, &p->common, p->common.id, tnif, TRACE_FUN_T_SCHED_PROC,
-                       what, tmp, THE_NON_VALUE);
+                       what, tmp, THE_NON_VALUE, am_true);
 }
 
 /* Send {trace_ts, Pid, What, {Mod, Func, Arity}, Timestamp}
@@ -812,9 +812,32 @@ trace_send(Process *p, Eterm to, Eterm msg)
 {
     Eterm operation = am_send;
     ErtsTracerNif *tnif = NULL;
+    ErtsTracingEvent* te;
+    Eterm pam_result;
 
     ASSERT(ARE_TRACE_FLAGS_ON(p, F_TRACE_SEND));
 
+    te = &erts_send_tracing[erts_active_bp_ix()];
+    if (!te->on) {
+	return;
+    }
+    if (te->match_spec) {
+	Eterm args[2];
+	Uint32 return_flags;
+	args[0] = to;
+	args[1] = msg;
+	pam_result = erts_match_set_run_trace(p, p,
+                                              te->match_spec, args, 2,
+                                              ERTS_PAM_TMP_RESULT, &return_flags);
+	if (pam_result == am_false)
+            return;
+        if (ERTS_TRACE_FLAGS(p) & F_TRACE_SILENT) {
+            erts_match_set_release_result_trace(p, pam_result);
+	    return;
+	}
+    } else
+        pam_result = am_true;
+
     if (is_internal_pid(to)) {
 	if (!erts_proc_lookup(to))
 	    goto send_to_non_existing_process;
@@ -826,23 +849,63 @@ trace_send(Process *p, Eterm to, Eterm msg)
     }
 
     if (is_tracer_enabled(p, ERTS_PROC_LOCK_MAIN, &p->common, &tnif,
-                TRACE_FUN_E_SEND, operation))
+                          TRACE_FUN_E_SEND, operation)) {
         send_to_tracer_nif(p, &p->common, p->common.id, tnif, TRACE_FUN_T_SEND,
-                operation, msg, to);
+                           operation, msg, to, pam_result);
+    }
+    erts_match_set_release_result_trace(p, pam_result);
 }
 
 /* Send {trace_ts, Pid, receive, Msg, Timestamp}
  * or   {trace, Pid, receive, Msg}
  */
 void
-trace_receive(Process *c_p, Eterm msg)
+trace_receive(Process* receiver,
+              Eterm from,
+              Eterm msg, ErtsTracingEvent* te)
 {
     ErtsTracerNif *tnif = NULL;
-    if (is_tracer_enabled(NULL, 0, &c_p->common, &tnif,
-                TRACE_FUN_E_RECEIVE, am_receive))
-        send_to_tracer_nif(NULL, &c_p->common, c_p->common.id,
+    Eterm pam_result;
+
+    if (!te) {
+        te = &erts_receive_tracing[erts_active_bp_ix()];
+        if (!te->on)
+            return;
+    }
+    else ASSERT(te->on);
+
+    if (te->match_spec) {
+        Eterm args[3];
+        Uint32 return_flags;
+        if (is_pid(from)) {
+            args[0] = pid_node_name(from);
+            args[1] = from;
+        }
+        else {
+            ASSERT(is_atom(from));
+            args[0] = from;  /* node name or other atom (e.g 'system') */
+            args[1] = am_undefined;
+        }
+        args[2] = msg;
+        pam_result = erts_match_set_run_trace(NULL, receiver,
+                                              te->match_spec, args, 3,
+                                              ERTS_PAM_TMP_RESULT, &return_flags);
+        if (pam_result == am_false)
+            return;
+        if (ERTS_TRACE_FLAGS(receiver) & F_TRACE_SILENT) {
+            erts_match_set_release_result_trace(NULL, pam_result);
+            return;
+        }
+    } else
+        pam_result = am_true;
+
+    if (is_tracer_enabled(NULL, 0, &receiver->common, &tnif,
+                          TRACE_FUN_E_RECEIVE, am_receive)) {
+        send_to_tracer_nif(NULL, &receiver->common, receiver->common.id,
                            tnif, TRACE_FUN_T_RECEIVE,
-                           am_receive, msg, THE_NON_VALUE);
+                           am_receive, msg, THE_NON_VALUE, pam_result);
+    }
+    erts_match_set_release_result_trace(NULL, pam_result);
 }
 
 int
@@ -964,7 +1027,7 @@ erts_trace_return_to(Process *p, BeamInstr *pc)
     }
 
     send_to_tracer_nif(p, &p->common, p->common.id, NULL, TRACE_FUN_T_CALL,
-                       am_return_to, mfa, THE_NON_VALUE);
+                       am_return_to, mfa, THE_NON_VALUE, am_true);
 }
 
 
@@ -1202,20 +1265,14 @@ erts_call_trace(Process* p, BeamInstr mfa[3], Binary *match_spec,
            may remove it, and we still want to generate a trace message */
         erts_tracer_update(&pre_ms_tracer, *tracer);
         tracer = &pre_ms_tracer;
-        pam_result = erts_match_set_run(p, match_spec, args, arity,
-                                        ERTS_PAM_TMP_RESULT, &return_flags);
-        if (is_non_value(pam_result)) {
-            erts_match_set_release_result(p);
-            UnUseTmpHeap(ERL_SUB_BIN_SIZE,p);
-            ERTS_TRACER_CLEAR(&pre_ms_tracer);
-            return 0;
-        }
+        pam_result = erts_match_set_run_trace(p, p,
+                                              match_spec, args, arity,
+                                              ERTS_PAM_TMP_RESULT, &return_flags);
     }
 
     if (tracee_flags == &meta_flags) {
         /* Meta trace */
         if (pam_result == am_false) {
-            erts_match_set_release_result(p);
             UnUseTmpHeap(ERL_SUB_BIN_SIZE,p);
             ERTS_TRACER_CLEAR(&pre_ms_tracer);
             return return_flags;
@@ -1223,13 +1280,12 @@ erts_call_trace(Process* p, BeamInstr mfa[3], Binary *match_spec,
     } else {
         /* Non-meta trace */
         if (*tracee_flags & F_TRACE_SILENT) {
-            erts_match_set_release_result(p);
+            erts_match_set_release_result_trace(p, pam_result);
             UnUseTmpHeap(ERL_SUB_BIN_SIZE,p);
             ERTS_TRACER_CLEAR(&pre_ms_tracer);
             return 0;
         }
         if (pam_result == am_false) {
-            erts_match_set_release_result(p);
             UnUseTmpHeap(ERL_SUB_BIN_SIZE,p);
             ERTS_TRACER_CLEAR(&pre_ms_tracer);
             return return_flags;
@@ -1265,11 +1321,14 @@ erts_call_trace(Process* p, BeamInstr mfa[3], Binary *match_spec,
      * Build the trace tuple and send it to the port.
      */
     send_to_tracer_nif_raw(p, NULL, *tracer, *tracee_flags, p->common.id,
-                           tnif, TRACE_FUN_T_CALL, am_call, mfa_tuple, THE_NON_VALUE, pam_result);
-    erts_match_set_release_result(p);
+                           tnif, TRACE_FUN_T_CALL, am_call, mfa_tuple,
+                           THE_NON_VALUE, pam_result);
 
-    if (match_spec && tracer == &pre_ms_tracer)
-        ERTS_TRACER_CLEAR(&pre_ms_tracer);
+    if (match_spec) {
+        erts_match_set_release_result_trace(p, pam_result);
+        if (tracer == &pre_ms_tracer)
+            ERTS_TRACER_CLEAR(&pre_ms_tracer);
+    }
 
     return return_flags;
 }
@@ -1290,7 +1349,7 @@ trace_proc(Process *c_p, ErtsProcLocks c_p_locks,
     if (is_tracer_enabled(c_p, c_p_locks, &t_p->common, &tnif,
                 TRACE_FUN_E_PROCS, what))
         send_to_tracer_nif(c_p, &t_p->common, t_p->common.id, tnif, TRACE_FUN_T_PROCS,
-                           what, data, THE_NON_VALUE);
+                           what, data, THE_NON_VALUE, am_true);
 }
 
 
@@ -1316,7 +1375,7 @@ trace_proc_spawn(Process *p, Eterm what, Eterm pid,
         mfa = TUPLE3(hp, mod, func, args);
         hp += 4;
         send_to_tracer_nif(p, &p->common, p->common.id, tnif, TRACE_FUN_T_PROCS,
-                           what, pid, mfa);
+                           what, pid, mfa, am_true);
     }
 }
 
@@ -1368,7 +1427,7 @@ trace_gc(Process *p, Eterm what, Uint size)
         msg = CONS(hp, tup, msg); hp += 2;
 
         send_to_tracer_nif(p, &p->common, p->common.id, tnif, TRACE_FUN_T_GC,
-                what, msg, am_undefined);
+                what, msg, am_undefined, am_true);
     }
 }
 
@@ -1431,7 +1490,7 @@ 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
 }
@@ -1496,7 +1555,7 @@ 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
 }
@@ -1571,7 +1630,7 @@ 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
 }
@@ -1646,7 +1705,7 @@ 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
 }
@@ -1678,7 +1737,7 @@ 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
 
@@ -1748,7 +1807,7 @@ trace_port_open(Port *p, Eterm calling_pid, Eterm drv_name) {
     ERTS_SMP_CHK_NO_PROC_LOCKS;
     if (is_tracer_enabled(NULL, 0, &p->common, &tnif, TRACE_FUN_E_PORTS, am_open))
         send_to_tracer_nif(NULL, &p->common, p->common.id, tnif, TRACE_FUN_T_PORTS,
-                am_open, calling_pid, drv_name);
+                am_open, calling_pid, drv_name, am_true);
 }
 
 /* Sends trace message:
@@ -1767,7 +1826,7 @@ trace_port(Port *t_p, Eterm what, Eterm data) {
     ERTS_SMP_CHK_NO_PROC_LOCKS;
     if (is_tracer_enabled(NULL, 0, &t_p->common, &tnif, TRACE_FUN_E_PORTS, what))
         send_to_tracer_nif(NULL, &t_p->common, t_p->common.id, tnif, TRACE_FUN_T_PORTS,
-                           what, data, THE_NON_VALUE);
+                           what, data, THE_NON_VALUE, am_true);
 }
 
 
@@ -1907,7 +1966,7 @@ trace_port_receive(Port *t_p, Eterm caller, Eterm what, ...)
         ASSERT(hp <= (local_heap + LOCAL_HEAP_SIZE) || orig_hp);
         send_to_tracer_nif(NULL, &t_p->common, t_p->common.id, tnif,
                            TRACE_FUN_T_RECEIVE,
-                           am_receive, data, THE_NON_VALUE);
+                           am_receive, data, THE_NON_VALUE, am_true);
 
         if (bptr && erts_refc_dectest(&bptr->refc, 1) == 0)
             erts_bin_free(bptr);
@@ -1930,7 +1989,7 @@ trace_port_send(Port *t_p, Eterm receiver, Eterm msg, int exists)
     ERTS_SMP_CHK_NO_PROC_LOCKS;
     if (is_tracer_enabled(NULL, 0, &t_p->common, &tnif, TRACE_FUN_E_SEND, op))
         send_to_tracer_nif(NULL, &t_p->common, t_p->common.id, tnif, TRACE_FUN_T_SEND,
-                           op, msg, receiver);
+                           op, msg, receiver, am_true);
 }
 
 void trace_port_send_binary(Port *t_p, Eterm to, Eterm what, char *bin, Sint sz)
@@ -1959,7 +2018,7 @@ void trace_port_send_binary(Port *t_p, Eterm to, Eterm what, char *bin, Sint sz)
         hp += 3;
 
         send_to_tracer_nif(NULL, &t_p->common, t_p->common.id, tnif, TRACE_FUN_T_SEND,
-                           am_send, msg, to);
+                           am_send, msg, to, am_true);
         if (bptr && erts_refc_dectest(&bptr->refc, 1) == 0)
             erts_bin_free(bptr);
 
@@ -1990,7 +2049,7 @@ trace_sched_ports_where(Port *t_p, Eterm what, Eterm where) {
     if (is_tracer_enabled(NULL, 0, &t_p->common, &tnif, TRACE_FUN_E_SCHED_PORT, what))
         send_to_tracer_nif(NULL, &t_p->common, t_p->common.id,
                            tnif, TRACE_FUN_T_SCHED_PORT,
-                           what, where, THE_NON_VALUE);
+                           what, where, THE_NON_VALUE, am_true);
 }
 
 /* Port profiling */
@@ -2430,7 +2489,7 @@ 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
@@ -2813,7 +2872,7 @@ send_to_tracer_nif_raw(Process *c_p, Process *tracee,
 static ERTS_INLINE int
 send_to_tracer_nif(Process *c_p, ErtsPTabElementCommon *t_p,
                    Eterm t_p_id, ErtsTracerNif *tnif, enum ErtsTracerOpt topt,
-                   Eterm tag, Eterm msg, Eterm extra)
+                   Eterm tag, Eterm msg, Eterm extra, Eterm pam_result)
 {
 #if defined(ERTS_SMP) && defined(ERTS_ENABLE_LOCK_CHECK)
     if (c_p) {
@@ -2833,7 +2892,7 @@ send_to_tracer_nif(Process *c_p, ErtsPTabElementCommon *t_p,
                                   is_internal_pid(t_p->id) ? (Process*)t_p : NULL,
                                   t_p->tracer, t_p->trace_flags,
                                   t_p_id, tnif, topt, tag, msg, extra,
-                                  am_true);
+                                  pam_result);
 }
 
 static ERTS_INLINE Eterm
diff --git a/erts/emulator/beam/erl_trace.h b/erts/emulator/beam/erl_trace.h
index 9a007e62ec..de4beadb7e 100644
--- a/erts/emulator/beam/erl_trace.h
+++ b/erts/emulator/beam/erl_trace.h
@@ -56,6 +56,15 @@
 
 struct binary;
 
+typedef struct
+{
+    int on;
+    struct binary* match_spec;
+} ErtsTracingEvent;
+
+extern ErtsTracingEvent erts_send_tracing[];
+extern ErtsTracingEvent erts_receive_tracing[];
+
 /* erl_bif_trace.c */
 Eterm erl_seq_trace_info(Process *p, Eterm arg1);
 void erts_system_monitor_clear(Process *c_p);
@@ -91,7 +100,7 @@ void erts_send_sys_msg_proc(Eterm, Eterm, Eterm, ErlHeapFragment *);
 #endif
 
 void trace_send(Process*, Eterm, Eterm);
-void trace_receive(Process *, Eterm);
+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/global.h b/erts/emulator/beam/global.h
index 49eec44053..550db95ba8 100644
--- a/erts/emulator/beam/global.h
+++ b/erts/emulator/beam/global.h
@@ -1483,9 +1483,19 @@ do {								\
 
 #define MatchSetGetSource(MPSP) erts_match_set_get_source(MPSP)
 
-extern Binary *erts_match_set_compile(Process *p, Eterm matchexpr);
+extern Binary *erts_match_set_compile(Process *p, Eterm matchexpr, Eterm MFA);
 Eterm erts_match_set_lint(Process *p, Eterm matchexpr); 
 extern void erts_match_set_release_result(Process* p);
+ERTS_GLB_INLINE void erts_match_set_release_result_trace(Process* p, Eterm);
+
+#if ERTS_GLB_INLINE_INCL_FUNC_DEF
+ERTS_GLB_INLINE
+void erts_match_set_release_result_trace(Process* p, Eterm pam_result)
+{
+    if (is_not_immed(pam_result))
+        erts_match_set_release_result(p);
+}
+#endif /* ERTS_GLB_INLINE_INCL_FUNC_DEF */
 
 enum erts_pam_run_flags {
     ERTS_PAM_TMP_RESULT=1,
@@ -1493,10 +1503,12 @@ enum erts_pam_run_flags {
     ERTS_PAM_CONTIGUOUS_TUPLE=4,
     ERTS_PAM_IGNORE_TRACE_SILENT=8
 };
-extern Eterm erts_match_set_run(Process *p, Binary *mpsp, 
-				Eterm *args, int num_args,
-				enum erts_pam_run_flags in_flags,
-				Uint32 *return_flags);
+extern Eterm erts_match_set_run_trace(Process *p,
+                                      Process *self,
+                                      Binary *mpsp,
+                                      Eterm *args, int num_args,
+                                      enum erts_pam_run_flags in_flags,
+                                      Uint32 *return_flags);
 extern Eterm erts_match_set_get_source(Binary *mpsp);
 extern void erts_match_prog_foreach_offheap(Binary *b,
 					    void (*)(ErlOffHeap *, void *),
diff --git a/erts/emulator/beam/io.c b/erts/emulator/beam/io.c
index 5c2595c69d..b9f6ab04c6 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/sys.h b/erts/emulator/beam/sys.h
index 748fba15c7..dda3ba565c 100644
--- a/erts/emulator/beam/sys.h
+++ b/erts/emulator/beam/sys.h
@@ -92,6 +92,9 @@
 #define ErtsInArea(ptr,start,nbytes) \
     ((UWord)((char*)(ptr) - (char*)(start)) < (nbytes))
 
+#define ErtsContainerStruct(ptr, type, member) \
+  (type *)((char *)(1 ? (ptr) : &((type *)0)->member) - offsetof(type, member))
+
 #if defined (__WIN32__)
 #  include "erl_win_sys.h"
 #else
diff --git a/erts/emulator/beam/utils.c b/erts/emulator/beam/utils.c
index 68006e7ef3..cedc88e5fe 100644
--- a/erts/emulator/beam/utils.c
+++ b/erts/emulator/beam/utils.c
@@ -2263,7 +2263,7 @@ 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
 }