Merge branch 'lukas/erts/tracer-module-and-more/PR-1009/OTP-10268'

* lukas/erts/tracer-module-and-more/PR-1009/OTP-10268: (26 commits)
  erts: Don't trace on link events when port is dead
  erts: more logging in trace_bif_SUITE trace_bif_return
  erts: Make trace_delivered go via sys msg dispatcher again
  erts: Add comment about future trace optimizations
  erts: Optimize tracer reload test
  erts: Document erlang:match_spec_test/3
  runtime_tools: Make dbg work with erl_tracer modules
  runtime_rools: Allow new timestamp trace flags
  runtime_tools: Lots of dbg docs updates
  erts: Deallocate heap fragments from trace nif calls
  eprof: Fix tests after tracer module incompatabilities
  fprof: update to work with new spawned trace event
  erts: Add 'spawned' trace event to 'procs' trace flag
  erts: send and receive no longer need status lock
  erts: Do 'unregister' as "self-tracing"
  erts: Silence harmless valgrind warning in dec_term
  erts: Fix end_per_testcase crash in local_trace_SUITE
  observer: Update ttb to work with tracing on ports
  runtime_tools: Update dbg to work with tracing on ports
  erts: Add tracing examples in match spec docs
  ...

42 files changed, 2819 insertions, 2659 deletions

diff --git a/erts/emulator/beam/atom.names b/erts/emulator/beam/atom.names
index d466f00028..8c51f788c0 100644
--- a/erts/emulator/beam/atom.names
+++ b/erts/emulator/beam/atom.names
@@ -161,6 +161,7 @@ atom close
 atom closed
 atom code
 atom command
+atom commandv
 atom compact
 atom compat_rel
 atom compile
@@ -197,6 +198,7 @@ atom dirty_cpu_schedulers_online
 atom dirty_io
 atom disable_trace
 atom disabled
+atom discard
 atom display_items
 atom dist
 atom dist_cmd
@@ -231,6 +233,7 @@ atom exception_trace
 atom extended
 atom Eq='=:='
 atom Eqeq='=='
+atom erl_tracer
 atom erlang
 atom ERROR='ERROR'
 atom error_handler
@@ -243,6 +246,9 @@ atom exact_reductions
 atom exclusive
 atom exit_status
 atom existing
+atom existing_processes
+atom existing_ports
+atom existing
 atom exiting
 atom exports
 atom external
@@ -353,6 +359,7 @@ atom match
 atom match_limit
 atom match_limit_recursion
 atom match_spec
+atom match_spec_result
 atom max
 atom maximum
 atom max_tables max_processes
@@ -401,6 +408,8 @@ atom net_kernel_terminated
 atom never_utf
 atom new
 atom new_index
+atom new_processes
+atom new_ports
 atom new_uniq
 atom newline
 atom next
@@ -540,6 +549,7 @@ atom scheme
 atom scientific
 atom scope
 atom seconds
+atom send_to_non_existing_process
 atom sensitive
 atom sequential_tracer
 atom sequential_trace_token
@@ -561,6 +571,7 @@ atom size
 atom sl_alloc
 atom spawn_executable
 atom spawn_driver
+atom spawned
 atom ssl_tls
 atom stack_size
 atom start
@@ -569,6 +580,7 @@ atom static
 atom stderr_to_stdout
 atom stop
 atom stream
+atom strict_monotonic
 atom strict_monotonic_timestamp
 atom sunrm
 atom suspend
@@ -596,8 +608,9 @@ atom total_active_tasks
 atom total_heap_size
 atom total_run_queue_lengths
 atom tpkt
-atom trace trace_ts traced 
+atom trace trace_ts traced
 atom trace_control_word
+atom trace_status
 atom tracer
 atom trap_exit
 atom trim
diff --git a/erts/emulator/beam/beam_bif_load.c b/erts/emulator/beam/beam_bif_load.c
index 320dee6668..b10250dc49 100644
--- a/erts/emulator/beam/beam_bif_load.c
+++ b/erts/emulator/beam/beam_bif_load.c
@@ -719,13 +719,13 @@ set_default_trace_pattern(Eterm module)
     Binary *match_spec;
     Binary *meta_match_spec;
     struct trace_pattern_flags trace_pattern_flags;
-    Eterm meta_tracer_pid;
+    ErtsTracer meta_tracer;
 
     erts_get_default_trace_pattern(&trace_pattern_is_on,
 				   &match_spec,
 				   &meta_match_spec,
 				   &trace_pattern_flags,
-				   &meta_tracer_pid);
+				   &meta_tracer);
     if (trace_pattern_is_on) {
 	Eterm mfa[1];
 	mfa[0] = module;
@@ -733,7 +733,7 @@ set_default_trace_pattern(Eterm module)
 				      match_spec,
 				      meta_match_spec,
 				      1, trace_pattern_flags,
-				      meta_tracer_pid, 1);
+				      meta_tracer, 1);
     }
 }
 
@@ -1264,6 +1264,7 @@ delete_code(Module* modp)
     modp->curr.code_length = 0;
     modp->curr.catches = BEAM_CATCHES_NIL;
     modp->curr.nif = NULL;
+
 }
 
 
diff --git a/erts/emulator/beam/beam_bp.c b/erts/emulator/beam/beam_bp.c
index 0a13454951..2ee98ed7b5 100644
--- a/erts/emulator/beam/beam_bp.c
+++ b/erts/emulator/beam/beam_bp.c
@@ -92,15 +92,15 @@ get_mtime(Process *c_p)
 /*
 ** Helpers
 */
-static Eterm do_call_trace(Process* c_p, BeamInstr* I, Eterm* reg,
-			   int local, Binary* ms, Eterm tracer_pid);
+static ErtsTracer do_call_trace(Process* c_p, BeamInstr* I, Eterm* reg,
+                                int local, Binary* ms, ErtsTracer tracer);
 static void set_break(BpFunctions* f, Binary *match_spec, Uint break_flags,
-		      enum erts_break_op count_op, Eterm tracer_pid);
+		      enum erts_break_op count_op, ErtsTracer tracer);
 static void set_function_break(BeamInstr *pc,
 			       Binary *match_spec,
 			       Uint break_flags,
 			       enum erts_break_op count_op,
-			       Eterm tracer_pid);
+			       ErtsTracer tracer);
 
 static void clear_break(BpFunctions* f, Uint break_flags);
 static int clear_function_break(BeamInstr *pc, Uint break_flags);
@@ -108,7 +108,7 @@ static int clear_function_break(BeamInstr *pc, Uint break_flags);
 static BpDataTime* get_time_break(BeamInstr *pc);
 static GenericBpData* check_break(BeamInstr *pc, Uint break_flags);
 
-static void bp_meta_unref(BpMetaPid* bmp);
+static void bp_meta_unref(BpMetaTracer* bmt);
 static void bp_count_unref(BpCount* bcp);
 static void bp_time_unref(BpDataTime* bdt);
 static void consolidate_bp_data(Module* modp, BeamInstr* pc, int local);
@@ -302,7 +302,7 @@ consolidate_bp_data(Module* modp, BeamInstr* pc, int local)
 	MatchSetUnref(dst->local_ms);
     }
     if (flags & ERTS_BPF_META_TRACE) {
-	bp_meta_unref(dst->meta_pid);
+	bp_meta_unref(dst->meta_tracer);
 	MatchSetUnref(dst->meta_ms);
     }
     if (flags & ERTS_BPF_COUNT) {
@@ -343,8 +343,8 @@ consolidate_bp_data(Module* modp, BeamInstr* pc, int local)
 	MatchSetRef(dst->local_ms);
     }
     if (flags & ERTS_BPF_META_TRACE) {
-	dst->meta_pid = src->meta_pid;
-	erts_refc_inc(&dst->meta_pid->refc, 1);
+	dst->meta_tracer = src->meta_tracer;
+	erts_refc_inc(&dst->meta_tracer->refc, 1);
 	dst->meta_ms = src->meta_ms;
 	MatchSetRef(dst->meta_ms);
     }
@@ -436,13 +436,13 @@ uninstall_breakpoint(BeamInstr* pc)
 void
 erts_set_trace_break(BpFunctions* f, Binary *match_spec)
 {
-    set_break(f, match_spec, ERTS_BPF_LOCAL_TRACE, 0, am_true);
+    set_break(f, match_spec, ERTS_BPF_LOCAL_TRACE, 0, erts_tracer_true);
 }
 
 void
-erts_set_mtrace_break(BpFunctions* f, Binary *match_spec, Eterm tracer_pid)
+erts_set_mtrace_break(BpFunctions* f, Binary *match_spec, ErtsTracer tracer)
 {
-    set_break(f, match_spec, ERTS_BPF_META_TRACE, 0, tracer_pid);
+    set_break(f, match_spec, ERTS_BPF_META_TRACE, 0, tracer);
 }
 
 void
@@ -450,13 +450,13 @@ erts_set_call_trace_bif(BeamInstr *pc, Binary *match_spec, int local)
 {
     Uint flags = local ? ERTS_BPF_LOCAL_TRACE : ERTS_BPF_GLOBAL_TRACE;
 
-    set_function_break(pc, match_spec, flags, 0, NIL);
+    set_function_break(pc, match_spec, flags, 0, erts_tracer_nil);
 }
 
 void
-erts_set_mtrace_bif(BeamInstr *pc, Binary *match_spec, Eterm tracer_pid)
+erts_set_mtrace_bif(BeamInstr *pc, Binary *match_spec, ErtsTracer tracer)
 {
-    set_function_break(pc, match_spec, ERTS_BPF_META_TRACE, 0, tracer_pid);
+    set_function_break(pc, match_spec, ERTS_BPF_META_TRACE, 0, tracer);
 }
 
 void
@@ -464,7 +464,7 @@ erts_set_time_trace_bif(BeamInstr *pc, enum erts_break_op count_op)
 {
     set_function_break(pc, NULL,
 		       ERTS_BPF_TIME_TRACE|ERTS_BPF_TIME_TRACE_ACTIVE,
-		       count_op, NIL);
+		       count_op, erts_tracer_nil);
 }
 
 void
@@ -474,21 +474,21 @@ erts_clear_time_trace_bif(BeamInstr *pc) {
 
 void
 erts_set_debug_break(BpFunctions* f) {
-    set_break(f, NULL, ERTS_BPF_DEBUG, 0, NIL);
+    set_break(f, NULL, ERTS_BPF_DEBUG, 0, erts_tracer_nil);
 }
 
 void
 erts_set_count_break(BpFunctions* f, enum erts_break_op count_op)
 {
     set_break(f, 0, ERTS_BPF_COUNT|ERTS_BPF_COUNT_ACTIVE,
-	      count_op, NIL);
+	      count_op, erts_tracer_nil);
 }
 
 void
 erts_set_time_break(BpFunctions* f, enum erts_break_op count_op)
 {
     set_break(f, 0, ERTS_BPF_TIME_TRACE|ERTS_BPF_TIME_TRACE_ACTIVE,
-	      count_op, NIL);
+	      count_op, erts_tracer_nil);
 }
 
 void
@@ -625,19 +625,26 @@ erts_generic_breakpoint(Process* c_p, BeamInstr* I, Eterm* reg)
 
     if (bp_flags & ERTS_BPF_LOCAL_TRACE) {
 	ASSERT((bp_flags & ERTS_BPF_GLOBAL_TRACE) == 0);
-	(void) do_call_trace(c_p, I, reg, 1, bp->local_ms, am_true);
+	(void) do_call_trace(c_p, I, reg, 1, bp->local_ms, erts_tracer_true);
     } else if (bp_flags & ERTS_BPF_GLOBAL_TRACE) {
-	(void) do_call_trace(c_p, I, reg, 0, bp->local_ms, am_true);
+	(void) do_call_trace(c_p, I, reg, 0, bp->local_ms, erts_tracer_true);
     }
 
     if (bp_flags & ERTS_BPF_META_TRACE) {
-	Eterm old_pid;
-	Eterm new_pid;
-
-	old_pid = (Eterm) erts_smp_atomic_read_nob(&bp->meta_pid->pid);
-	new_pid = do_call_trace(c_p, I, reg, 1, bp->meta_ms, old_pid);
-	if (new_pid != old_pid) {
-	    erts_smp_atomic_set_nob(&bp->meta_pid->pid, new_pid);
+	ErtsTracer old_tracer, new_tracer;
+
+	old_tracer = erts_smp_atomic_read_nob(&bp->meta_tracer->tracer);
+
+	new_tracer = do_call_trace(c_p, I, reg, 1, bp->meta_ms, old_tracer);
+	if (!ERTS_TRACER_COMPARE(new_tracer, old_tracer)) {
+            if (old_tracer == erts_smp_atomic_cmpxchg_acqb(
+                    &bp->meta_tracer->tracer,
+                    (erts_aint_t)new_tracer,
+                    (erts_aint_t)old_tracer)) {
+                ERTS_TRACER_CLEAR(&old_tracer);
+            } else {
+                ERTS_TRACER_CLEAR(&new_tracer);
+            }
 	}
     }
 
@@ -645,7 +652,8 @@ erts_generic_breakpoint(Process* c_p, BeamInstr* I, Eterm* reg)
 	erts_smp_atomic_inc_nob(&bp->count->acount);
     }
 
-    if (bp_flags & ERTS_BPF_TIME_TRACE_ACTIVE && erts_is_tracer_proc_valid(c_p)) {
+    if (bp_flags & ERTS_BPF_TIME_TRACE_ACTIVE
+        && ERTS_TRACER_PROC_IS_ENABLED(c_p)) {
 	Eterm w;
 	erts_trace_time_call(c_p, I, bp->time);
 	w = (BeamInstr) *c_p->cp;
@@ -690,7 +698,7 @@ erts_bif_trace(int bif_index, Process* p, Eterm* args, BeamInstr* I)
     Eterm (*func)(Process*, Eterm*, BeamInstr*);
     Export* ep = bif_export[bif_index];
     Uint32 flags = 0, flags_meta = 0;
-    Eterm meta_tracer_pid = NIL;
+    ErtsTracer meta_tracer = erts_tracer_nil;
     int applying = (I == &(ep->code[3])); /* Yup, the apply code for a bif
 					   * is actually in the
 					   * export entry */
@@ -718,23 +726,32 @@ erts_bif_trace(int bif_index, Process* p, Eterm* args, BeamInstr* I)
 	IS_TRACED_FL(p, F_TRACE_CALLS)) {
 	int local = !!(bp_flags & ERTS_BPF_LOCAL_TRACE);
 	flags = erts_call_trace(p, ep->code, bp->local_ms, args,
-				local, &ERTS_TRACER_PROC(p));
+				local, &ERTS_TRACER(p));
     }
     if (bp_flags & ERTS_BPF_META_TRACE) {
-	Eterm tpid1, tpid2;
+	ErtsTracer old_tracer;
 
-	tpid1 = tpid2 =
-	    (Eterm) erts_smp_atomic_read_nob(&bp->meta_pid->pid);
+        meta_tracer = erts_smp_atomic_read_nob(&bp->meta_tracer->tracer);
+        old_tracer = meta_tracer;
 	flags_meta = erts_call_trace(p, ep->code, bp->meta_ms, args,
-				     0, &tpid2);
-	meta_tracer_pid = tpid2;
-	if (tpid1 != tpid2) {
-	    erts_smp_atomic_set_nob(&bp->meta_pid->pid, tpid2);
+				     0, &meta_tracer);
+
+	if (!ERTS_TRACER_COMPARE(old_tracer, meta_tracer)) {
+            ErtsTracer new_tracer = erts_tracer_nil;
+            erts_tracer_update(&new_tracer, meta_tracer);
+	    if (old_tracer == erts_smp_atomic_cmpxchg_acqb(
+                    &bp->meta_tracer->tracer,
+                    (erts_aint_t)new_tracer,
+                    (erts_aint_t)old_tracer)) {
+                ERTS_TRACER_CLEAR(&old_tracer);
+            } else {
+                ERTS_TRACER_CLEAR(&new_tracer);
+            }
 	}
     }
     if (bp_flags & ERTS_BPF_TIME_TRACE_ACTIVE &&
 	IS_TRACED_FL(p, F_TRACE_CALLS) &&
-	erts_is_tracer_proc_valid(p)) {
+	ERTS_TRACER_PROC_IS_ENABLED(p)) {
 	BeamInstr *pc = (BeamInstr *)ep->code+3;
 	erts_trace_time_call(p, pc, bp->time);
     }
@@ -778,8 +795,6 @@ erts_bif_trace(int bif_index, Process* p, Eterm* args, BeamInstr* I)
 	if (reason != TRAP) {
 	    Eterm class;
 	    Eterm value = p->fvalue;
-	    DeclareTmpHeapNoproc(nocatch,3);
-	    UseTmpHeapNoproc(3);
 	    /* Expand error value like in handle_error() */
 	    if (reason & EXF_ARGLIST) {
 		Eterm *tp;
@@ -788,7 +803,8 @@ erts_bif_trace(int bif_index, Process* p, Eterm* args, BeamInstr* I)
 		value = tp[1];
 	    }
 	    if ((reason & EXF_THROWN) && (p->catches <= 0)) {
-		value = TUPLE2(nocatch, am_nocatch, value);
+                Eterm *hp = HAlloc(p, 3);
+		value = TUPLE2(hp, am_nocatch, value);
 		reason = EXC_ERROR;
 	    }
 	    /* Note: expand_error_value() could theoretically
@@ -801,11 +817,11 @@ erts_bif_trace(int bif_index, Process* p, Eterm* args, BeamInstr* I)
 
 	    if (flags_meta & MATCH_SET_EXCEPTION_TRACE) {
 		erts_trace_exception(p, ep->code, class, value,
-				     &meta_tracer_pid);
+				     &meta_tracer);
 	    }
 	    if (flags & MATCH_SET_EXCEPTION_TRACE) {
 		erts_trace_exception(p, ep->code, class, value,
-				     &ERTS_TRACER_PROC(p));
+				     &ERTS_TRACER(p));
 	    }
 	    if ((flags & MATCH_SET_RETURN_TO_TRACE) && p->catches > 0) {
 		/* can only happen if(local)*/
@@ -827,7 +843,6 @@ erts_bif_trace(int bif_index, Process* p, Eterm* args, BeamInstr* I)
 		    }
 		}
 	    }
-	    UnUseTmpHeapNoproc(3);
 	    if ((flags_meta|flags) & MATCH_SET_EXCEPTION_TRACE) {
 		erts_smp_proc_lock(p, ERTS_PROC_LOCKS_ALL_MINOR);
 		ERTS_TRACE_FLAGS(p) |= F_EXCEPTION_TRACE;
@@ -836,11 +851,11 @@ erts_bif_trace(int bif_index, Process* p, Eterm* args, BeamInstr* I)
 	}
     } else {
 	if (flags_meta & MATCH_SET_RX_TRACE) {
-	    erts_trace_return(p, ep->code, result, &meta_tracer_pid);
+	    erts_trace_return(p, ep->code, result, &meta_tracer);
 	}
 	/* MATCH_SET_RETURN_TO_TRACE cannot occur if(meta) */
 	if (flags & MATCH_SET_RX_TRACE) {
-	    erts_trace_return(p, ep->code, result, &ERTS_TRACER_PROC(p));
+	    erts_trace_return(p, ep->code, result, &ERTS_TRACER(p));
 	}
 	if (flags & MATCH_SET_RETURN_TO_TRACE) {
 	    /* can only happen if(local)*/
@@ -857,9 +872,9 @@ erts_bif_trace(int bif_index, Process* p, Eterm* args, BeamInstr* I)
     return result;
 }
 
-static Eterm
+static ErtsTracer
 do_call_trace(Process* c_p, BeamInstr* I, Eterm* reg,
-	      int local, Binary* ms, Eterm tracer_pid)
+	      int local, Binary* ms, ErtsTracer tracer)
 {
     Eterm* cpp;
     int return_to_trace = 0;
@@ -899,7 +914,7 @@ do_call_trace(Process* c_p, BeamInstr* I, Eterm* reg,
 	ASSERT(is_CP(*cpp));
     }
     ERTS_SMP_UNREQ_PROC_MAIN_LOCK(c_p);
-    flags = erts_call_trace(c_p, I-3, ms, reg, local, &tracer_pid);
+    flags = erts_call_trace(c_p, I-3, ms, reg, local, &tracer);
     ERTS_SMP_REQ_PROC_MAIN_LOCK(c_p);
     if (cpp) {
 	c_p->cp = cp_save;
@@ -910,7 +925,7 @@ do_call_trace(Process* c_p, BeamInstr* I, Eterm* reg,
 	need += 1;
     }
     if (flags & MATCH_SET_RX_TRACE) {
-	need += 3;
+	need += 3 + size_object(tracer);
     }
     if (need) {
 	ASSERT(c_p->htop <= E && E <= c_p->hend);
@@ -926,14 +941,15 @@ do_call_trace(Process* c_p, BeamInstr* I, Eterm* reg,
 	E[0] = make_cp(c_p->cp);
 	c_p->cp = beam_return_to_trace;
     }
-    if (flags & MATCH_SET_RX_TRACE) {
+    if (flags & MATCH_SET_RX_TRACE)
+    {
 	E -= 3;
+        c_p->stop = E;
 	ASSERT(c_p->htop <= E && E <= c_p->hend);
 	ASSERT(is_CP((Eterm) (UWord) (I - 3)));
-	ASSERT(am_true == tracer_pid ||
-	       is_internal_pid(tracer_pid) || is_internal_port(tracer_pid));
+	ASSERT(IS_TRACER_VALID(tracer));
 	E[2] = make_cp(c_p->cp);
-	E[1] = tracer_pid;
+        E[1] = copy_object(tracer, c_p);
 	E[0] = make_cp(I - 3); /* We ARE at the beginning of an
 				  instruction,
 				  the funcinfo is above i. */
@@ -942,9 +958,9 @@ do_call_trace(Process* c_p, BeamInstr* I, Eterm* reg,
 	erts_smp_proc_lock(c_p, ERTS_PROC_LOCKS_ALL_MINOR);
 	ERTS_TRACE_FLAGS(c_p) |= F_EXCEPTION_TRACE;
 	erts_smp_proc_unlock(c_p, ERTS_PROC_LOCKS_ALL_MINOR);
-    }
-    c_p->stop = E;
-    return tracer_pid;
+    } else
+        c_p->stop = E;
+    return tracer;
 }
 
 void
@@ -1098,9 +1114,9 @@ erts_is_trace_break(BeamInstr *pc, Binary **match_spec_ret, int local)
     return 0;
 }
 
-int 
+int
 erts_is_mtrace_break(BeamInstr *pc, Binary **match_spec_ret,
-		     Eterm *tracer_pid_ret)
+		     ErtsTracer *tracer_ret)
 {
     GenericBpData* bp = check_break(pc, ERTS_BPF_META_TRACE);
     
@@ -1108,9 +1124,8 @@ erts_is_mtrace_break(BeamInstr *pc, Binary **match_spec_ret,
 	if (match_spec_ret) {
 	    *match_spec_ret = bp->meta_ms;
 	}
-	if (tracer_pid_ret) {
-	    *tracer_pid_ret =
-		(Eterm) erts_smp_atomic_read_nob(&bp->meta_pid->pid);
+	if (tracer_ret) {
+            *tracer_ret = erts_smp_atomic_read_nob(&bp->meta_tracer->tracer);
 	}
 	return 1;
     }
@@ -1391,7 +1406,7 @@ void erts_schedule_time_break(Process *p, Uint schedule) {
 
 static void
 set_break(BpFunctions* f, Binary *match_spec, Uint break_flags,
-	  enum erts_break_op count_op, Eterm tracer_pid)
+	  enum erts_break_op count_op, ErtsTracer tracer)
 {
     Uint i;
     Uint n;
@@ -1400,13 +1415,13 @@ set_break(BpFunctions* f, Binary *match_spec, Uint break_flags,
     for (i = 0; i < n; i++) {
 	BeamInstr* pc = f->matching[i].pc;
 	set_function_break(pc, match_spec, break_flags,
-			   count_op, tracer_pid);
+			   count_op, tracer);
     }
 }
 
 static void
 set_function_break(BeamInstr *pc, Binary *match_spec, Uint break_flags,
-		   enum erts_break_op count_op, Eterm tracer_pid)
+		   enum erts_break_op count_op, ErtsTracer tracer)
 {
     GenericBp* g;
     GenericBpData* bp;
@@ -1439,7 +1454,7 @@ set_function_break(BeamInstr *pc, Binary *match_spec, Uint break_flags,
 	MatchSetUnref(bp->local_ms);
     } else if (common & ERTS_BPF_META_TRACE) {
 	MatchSetUnref(bp->meta_ms);
-	bp_meta_unref(bp->meta_pid);
+	bp_meta_unref(bp->meta_tracer);
     } else if (common & ERTS_BPF_COUNT) {
 	if (count_op == erts_break_stop) {
 	    bp->flags &= ~ERTS_BPF_COUNT_ACTIVE;
@@ -1474,13 +1489,15 @@ set_function_break(BeamInstr *pc, Binary *match_spec, Uint break_flags,
 	MatchSetRef(match_spec);
 	bp->local_ms = match_spec;
     } else if (break_flags & ERTS_BPF_META_TRACE) {
-	BpMetaPid* bmp;
+	BpMetaTracer* bmt;
+        ErtsTracer meta_tracer = erts_tracer_nil;
 	MatchSetRef(match_spec);
 	bp->meta_ms = match_spec;
-	bmp = Alloc(sizeof(BpMetaPid));
-	erts_refc_init(&bmp->refc, 1);
-	erts_smp_atomic_init_nob(&bmp->pid, tracer_pid);
-	bp->meta_pid = bmp;
+	bmt = Alloc(sizeof(BpMetaTracer));
+	erts_refc_init(&bmt->refc, 1);
+        erts_tracer_update(&meta_tracer, tracer); /* copy tracer */
+	erts_smp_atomic_init_nob(&bmt->tracer, (erts_aint_t)meta_tracer);
+	bp->meta_tracer = bmt;
     } else if (break_flags & ERTS_BPF_COUNT) {
 	BpCount* bcp;
 
@@ -1544,7 +1561,7 @@ clear_function_break(BeamInstr *pc, Uint break_flags)
     }
     if (common & ERTS_BPF_META_TRACE) {
 	MatchSetUnref(bp->meta_ms);
-	bp_meta_unref(bp->meta_pid);
+	bp_meta_unref(bp->meta_tracer);
     }
     if (common & ERTS_BPF_COUNT) {
 	ASSERT((bp->flags & ERTS_BPF_COUNT_ACTIVE) == 0);
@@ -1560,10 +1577,12 @@ clear_function_break(BeamInstr *pc, Uint break_flags)
 }
 
 static void
-bp_meta_unref(BpMetaPid* bmp)
+bp_meta_unref(BpMetaTracer* bmt)
 {
-    if (erts_refc_dectest(&bmp->refc, 0) <= 0) {
-	Free(bmp);
+    if (erts_refc_dectest(&bmt->refc, 0) <= 0) {
+        ErtsTracer trc = erts_smp_atomic_read_nob(&bmt->tracer);
+        ERTS_TRACER_CLEAR(&trc);
+	Free(bmt);
     }
 }
 
diff --git a/erts/emulator/beam/beam_bp.h b/erts/emulator/beam/beam_bp.h
index f9eca94e2a..bb171be8a6 100644
--- a/erts/emulator/beam/beam_bp.h
+++ b/erts/emulator/beam/beam_bp.h
@@ -55,15 +55,15 @@ typedef struct {
 } BpCount;
 
 typedef struct {
-    erts_smp_atomic_t pid;
+    erts_smp_atomic_t tracer;
     erts_refc_t refc;
-} BpMetaPid;
+} BpMetaTracer;
 
 typedef struct generic_bp_data {
     Uint flags;
     Binary* local_ms;		/* Match spec for local call trace */
     Binary* meta_ms;		/* Match spec for meta trace */
-    BpMetaPid* meta_pid;	/* Meta trace pid */
+    BpMetaTracer* meta_tracer;	/* Meta tracer */
     BpCount* count;		/* For call count */
     BpDataTime* time;		/* For time trace */
 } GenericBpData;
@@ -132,10 +132,10 @@ void erts_set_call_trace_bif(BeamInstr *pc, Binary *match_spec, int local);
 void erts_clear_call_trace_bif(BeamInstr *pc, int local);
 
 void erts_set_mtrace_break(BpFunctions *f, Binary *match_spec,
-			  Eterm tracer_pid);
+			  ErtsTracer tracer);
 void erts_clear_mtrace_break(BpFunctions *f);
 void erts_set_mtrace_bif(BeamInstr *pc, Binary *match_spec,
-			 Eterm tracer_pid);
+			 ErtsTracer tracer);
 void erts_clear_mtrace_bif(BeamInstr *pc);
 
 void erts_set_debug_break(BpFunctions *f);
@@ -150,13 +150,13 @@ void erts_clear_export_break(Module *modp, BeamInstr* pc);
 
 BeamInstr erts_generic_breakpoint(Process* c_p, BeamInstr* I, Eterm* reg);
 BeamInstr erts_trace_break(Process *p, BeamInstr *pc, Eterm *args,
-		      Uint32 *ret_flags, Eterm *tracer_pid);
+                           Uint32 *ret_flags, ErtsTracer *tracer);
 
 int erts_is_trace_break(BeamInstr *pc, Binary **match_spec_ret, int local);
 int erts_is_mtrace_break(BeamInstr *pc, Binary **match_spec_ret,
-			 Eterm *tracer_pid_rte);
+			 ErtsTracer *tracer_ret);
 int erts_is_mtrace_bif(BeamInstr *pc, Binary **match_spec_ret,
-		       Eterm *tracer_pid_ret);
+		       ErtsTracer *tracer_ret);
 int erts_is_native_break(BeamInstr *pc);
 int erts_is_count_break(BeamInstr *pc, Uint *count_ret);
 int erts_is_time_break(Process *p, BeamInstr *pc, Eterm *call_time);
diff --git a/erts/emulator/beam/beam_emu.c b/erts/emulator/beam/beam_emu.c
index 901419c989..254bfe4ba0 100644
--- a/erts/emulator/beam/beam_emu.c
+++ b/erts/emulator/beam/beam_emu.c
@@ -3346,7 +3346,7 @@ do {						\
  OpCase(normal_exit): {
      SWAPOUT;
      c_p->freason = EXC_NORMAL;
-     c_p->arity = 0;		/* In case this process will ever be garbed again. */
+     c_p->arity = 0;		/* In case this process will never be garbed again. */
      ERTS_SMP_UNREQ_PROC_MAIN_LOCK(c_p);
      erts_do_exit_process(c_p, am_normal);
      ERTS_SMP_REQ_PROC_MAIN_LOCK(c_p);
@@ -3460,7 +3460,7 @@ do {						\
 		typedef Eterm NifF(struct enif_environment_t*, int argc, Eterm argv[]);
 		NifF* fp = vbf = (NifF*) I[1];
 		struct enif_environment_t env;
-		erts_pre_nif(&env, c_p, (struct erl_module_nif*)I[2]);
+		erts_pre_nif(&env, c_p, (struct erl_module_nif*)I[2], NULL);
 		live_hf_end = c_p->mbuf;
 		nif_bif_result = (*fp)(&env, bif_nif_arity, reg);
 		if (env.exception_thrown)
@@ -4588,7 +4588,7 @@ do {						\
      
      SWAPOUT;		/* Needed for shared heap */
      ERTS_SMP_UNREQ_PROC_MAIN_LOCK(c_p);
-     erts_trace_return(c_p, code, r(0), E+1/*Process tracer*/);
+     erts_trace_return(c_p, code, r(0), ERTS_TRACER_FROM_ETERM(E+1)/* tracer */);
      ERTS_SMP_REQ_PROC_MAIN_LOCK(c_p);
      SWAPIN;
      c_p->cp = NULL;
@@ -5228,7 +5228,8 @@ next_catch(Process* c_p, Eterm *reg) {
 	BeamInstr *cpp = c_p->cp;
 	if (cpp == beam_exception_trace) {
 	    erts_trace_exception(c_p, cp_val(ptr[0]),
-				 reg[1], reg[2], ptr+1);
+				 reg[1], reg[2],
+                                 ERTS_TRACER_FROM_ETERM(ptr+1));
 	    /* Skip return_trace parameters */
 	    ptr += 2;
 	} else if (cpp == beam_return_trace) {
@@ -5255,7 +5256,8 @@ next_catch(Process* c_p, Eterm *reg) {
 		}
 		if (cp_val(*prev) == beam_exception_trace) {
 		    erts_trace_exception(c_p, cp_val(ptr[0]),
-					 reg[1], reg[2], ptr+1);
+					 reg[1], reg[2],
+                                         ERTS_TRACER_FROM_ETERM(ptr+1));
 		}
 		/* Skip return_trace parameters */
 		ptr += 2;
@@ -6049,7 +6051,7 @@ erts_hibernate(Process* c_p, Eterm module, Eterm function, Eterm args, Eterm* re
 	    return -1;
 	}
 #else /* ERTS_SMP */
-	ERTS_SMP_MSGQ_MV_INQ2PRIVQ(c_p);
+        ERTS_SMP_MSGQ_MV_INQ2PRIVQ(c_p);
 	if (!c_p->msg.len)
 #endif
 	    erts_smp_atomic32_read_band_relb(&c_p->state, ~ERTS_PSFLG_ACTIVE);
diff --git a/erts/emulator/beam/bif.c b/erts/emulator/beam/bif.c
index 754e11f047..ed5b2983dd 100644
--- a/erts/emulator/beam/bif.c
+++ b/erts/emulator/beam/bif.c
@@ -123,15 +123,12 @@ static int insert_internal_link(Process* p, Eterm rpid)
 	erts_add_link(&ERTS_P_LINKS(p), LINK_PID, rp->common.id);
 	erts_add_link(&ERTS_P_LINKS(rp), LINK_PID, p->common.id);
 
-	ASSERT(is_nil(ERTS_TRACER_PROC(p))
-	       || is_internal_pid(ERTS_TRACER_PROC(p))
-	       || is_internal_port(ERTS_TRACER_PROC(p)));
+	ASSERT(IS_TRACER_VALID(ERTS_TRACER(p)));
 
 	if (IS_TRACED(p)) {
 	    if (ERTS_TRACE_FLAGS(p) & (F_TRACE_SOL|F_TRACE_SOL1))  {
 		ERTS_TRACE_FLAGS(rp) |= (ERTS_TRACE_FLAGS(p) & TRACEE_FLAGS);
-		ERTS_TRACER_PROC(rp) = ERTS_TRACER_PROC(p); /* maybe steal */
-
+                erts_tracer_replace(&rp->common, ERTS_TRACER(p));
 		if (ERTS_TRACE_FLAGS(p) & F_TRACE_SOL1) { /* maybe override */
 		    ERTS_TRACE_FLAGS(rp) &= ~(F_TRACE_SOL1 | F_TRACE_SOL);
 		    ERTS_TRACE_FLAGS(p) &= ~(F_TRACE_SOL1 | F_TRACE_SOL);
@@ -140,7 +137,8 @@ static int insert_internal_link(Process* p, Eterm rpid)
 	}
     }
     if (IS_TRACED_FL(rp, F_TRACE_PROCS))
-	trace_proc(p, rp, am_getting_linked, p->common.id);
+	trace_proc(p, p == rp ? rp_locks : ERTS_PROC_LOCK_MAIN|ERTS_PROC_LOCK_LINK,
+                   rp, am_getting_linked, p->common.id);
 
     if (p == rp)
 	erts_smp_proc_unlock(p, rp_locks & ~ERTS_PROC_LOCK_MAIN);
@@ -159,7 +157,7 @@ BIF_RETTYPE link_1(BIF_ALIST_1)
     DistEntry *dep;
 
     if (IS_TRACED_FL(BIF_P, F_TRACE_PROCS)) {
-	trace_proc(BIF_P, BIF_P, am_link, BIF_ARG_1);
+	trace_proc(BIF_P, ERTS_PROC_LOCK_MAIN, BIF_P, am_link, BIF_ARG_1);
     }
     /* check that the pid or port which is our argument is OK */
 
@@ -613,7 +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, NIL);
+    erts_queue_message(p, p_locksp, msgp, tup);
 }
 
 static BIF_RETTYPE
@@ -1001,6 +999,7 @@ BIF_RETTYPE unlink_1(BIF_ALIST_1)
     Process *rp;
     DistEntry *dep;
     ErtsLink *l = NULL, *rl = NULL;
+    ErtsProcLocks cp_locks = ERTS_PROC_LOCK_MAIN;
 
     /*
      * SMP specific note concerning incoming exit signals:
@@ -1015,7 +1014,7 @@ BIF_RETTYPE unlink_1(BIF_ALIST_1)
      */
 
     if (IS_TRACED_FL(BIF_P, F_TRACE_PROCS)) {
-	trace_proc(BIF_P, BIF_P, am_unlink, BIF_ARG_1);
+	trace_proc(BIF_P, cp_locks, BIF_P, am_unlink, BIF_ARG_1);
     }
 
     if (is_internal_port(BIF_ARG_1)) {
@@ -1120,10 +1119,10 @@ BIF_RETTYPE unlink_1(BIF_ALIST_1)
 
     erts_smp_proc_lock(BIF_P, ERTS_PROC_LOCK_LINK|ERTS_PROC_LOCK_STATUS);
 
+    cp_locks |= ERTS_PROC_LOCK_LINK|ERTS_PROC_LOCK_STATUS;
+
     /* get process struct */
-    rp = erts_pid2proc_opt(BIF_P, (ERTS_PROC_LOCK_MAIN
-				   | ERTS_PROC_LOCK_LINK
-				   | ERTS_PROC_LOCK_STATUS),
+    rp = erts_pid2proc_opt(BIF_P, cp_locks,
 			   BIF_ARG_1, ERTS_PROC_LOCK_LINK,
 			   ERTS_P2P_FLG_ALLOW_OTHER_X);
 
@@ -1149,14 +1148,17 @@ BIF_RETTYPE unlink_1(BIF_ALIST_1)
 	    erts_destroy_link(rl);
 
 	if (IS_TRACED_FL(rp, F_TRACE_PROCS) && rl != NULL) {
-	    trace_proc(BIF_P, rp, am_getting_unlinked, BIF_P->common.id);
+            erts_smp_proc_unlock(BIF_P, ERTS_PROC_LOCK_STATUS);
+            cp_locks &= ~ERTS_PROC_LOCK_STATUS;
+	    trace_proc(BIF_P, (ERTS_PROC_LOCK_MAIN | ERTS_PROC_LOCK_LINK),
+                       rp, am_getting_unlinked, BIF_P->common.id);
 	}
 
 	if (rp != BIF_P)
 	    erts_smp_proc_unlock(rp, ERTS_PROC_LOCK_LINK);
     }
  
-    erts_smp_proc_unlock(BIF_P, ERTS_PROC_LOCK_LINK|ERTS_PROC_LOCK_STATUS);
+    erts_smp_proc_unlock(BIF_P, cp_locks & ~ERTS_PROC_LOCK_MAIN);
 
     BIF_RET(am_true);
 
@@ -1909,7 +1911,7 @@ static Sint remote_send(Process *p, DistEntry *dep,
     }
 
     if (res >= 0) {
-	if (IS_TRACED(p))
+	if (IS_TRACED_FL(p, F_TRACE_SEND))
 	    trace_send(p, full_to, msg);
 	if (ERTS_PROC_GET_SAVED_CALLS_BUF(p))
 	    save_calls(p, &exp_send);
@@ -1928,7 +1930,7 @@ do_send(Process *p, Eterm to, Eterm msg, Eterm *refp, ErtsSendContext* ctx)
     Eterm* tp;
 
     if (is_internal_pid(to)) {
-	if (IS_TRACED(p))
+	if (IS_TRACED_FL(p, F_TRACE_SEND))
 	    trace_send(p, to, msg);
 	if (ERTS_PROC_GET_SAVED_CALLS_BUF(p))
 	    save_calls(p, &exp_send);
@@ -1957,7 +1959,7 @@ do_send(Process *p, Eterm to, Eterm msg, Eterm *refp, ErtsSendContext* ctx)
 
 	rp = erts_proc_lookup_raw(id);
 	if (rp) {
-	    if (IS_TRACED(p))
+	    if (IS_TRACED_FL(p, F_TRACE_SEND))
 		trace_send(p, to, msg);
 	    if (ERTS_PROC_GET_SAVED_CALLS_BUF(p))
 		save_calls(p, &exp_send);
@@ -1973,7 +1975,7 @@ do_send(Process *p, Eterm to, Eterm msg, Eterm *refp, ErtsSendContext* ctx)
 	    goto port_common;
 	}
 
-	if (IS_TRACED(p))
+	if (IS_TRACED_FL(p, F_TRACE_SEND))
 	    trace_send(p, to, msg);
 	if (ERTS_PROC_GET_SAVED_CALLS_BUF(p))
 	    save_calls(p, &exp_send);
@@ -2004,11 +2006,20 @@ do_send(Process *p, Eterm to, Eterm msg, Eterm *refp, ErtsSendContext* ctx)
 
       port_common:
 	ret_val = 0;
-        
+
 	if (pt) {
 	    int ps_flags = ctx->suspend ? 0 : ERTS_PORT_SIG_FLG_NOSUSPEND;
 	    *refp = NIL;
 
+            if (IS_TRACED_FL(p, F_TRACE_SEND)) 	/* trace once only !! */
+                trace_send(p, portid, msg);
+
+            if (have_seqtrace(SEQ_TRACE_TOKEN(p))) {
+                seq_trace_update_send(p);
+                seq_trace_output(SEQ_TRACE_TOKEN(p), msg,
+                                 SEQ_TRACE_SEND, portid, p);
+            }
+
 	    switch (erts_port_command(p, ps_flags, pt, msg, refp)) {
 	    case ERTS_PORT_OP_CALLER_EXIT:
 		/* We are exiting... */
@@ -2041,18 +2052,10 @@ do_send(Process *p, Eterm to, Eterm msg, Eterm *refp, ErtsSendContext* ctx)
 		break;
 	    }
 	}
-	
-	if (IS_TRACED(p)) 	/* trace once only !! */
-	    trace_send(p, portid, msg);
+
 	if (ERTS_PROC_GET_SAVED_CALLS_BUF(p))
 	    save_calls(p, &exp_send);
-	
-        if (have_seqtrace(SEQ_TRACE_TOKEN(p))) {
-	    seq_trace_update_send(p);
-	    seq_trace_output(SEQ_TRACE_TOKEN(p), msg, 
-			     SEQ_TRACE_SEND, portid, p);
-	}	    
-	
+
 	if (ERTS_PROC_IS_EXITING(p)) {
 	    KILL_CATCHES(p); /* Must exit */
 	    return SEND_USER_ERROR;
@@ -2075,7 +2078,7 @@ do_send(Process *p, Eterm to, Eterm msg, Eterm *refp, ErtsSendContext* ctx)
 	if (dep == erts_this_dist_entry) {
 	    Eterm id;
 	    erts_deref_dist_entry(dep);
-	    if (IS_TRACED(p))
+	    if (IS_TRACED_FL(p, F_TRACE_SEND))
 		trace_send(p, to, msg);
 	    if (ERTS_PROC_GET_SAVED_CALLS_BUF(p))
 		save_calls(p, &exp_send);
@@ -2106,7 +2109,7 @@ do_send(Process *p, Eterm to, Eterm msg, Eterm *refp, ErtsSendContext* ctx)
 	}
 	return ret;
     } else {
-	if (IS_TRACED(p)) /* XXX Is this really neccessary ??? */
+	if (IS_TRACED_FL(p, F_TRACE_SEND))
 	    trace_send(p, to, msg);
 	if (ERTS_PROC_GET_SAVED_CALLS_BUF(p))
 	    save_calls(p, &exp_send);
@@ -4337,12 +4340,34 @@ BIF_RETTYPE system_flag_2(BIF_ALIST_2)
     } else if (BIF_ARG_1 == am_trace_control_word) {
 	BIF_RET(db_set_trace_control_word(BIF_P, BIF_ARG_2));
     } else if (BIF_ARG_1 == am_sequential_tracer) {
-        Eterm old_value = erts_set_system_seq_tracer(BIF_P,
-						     ERTS_PROC_LOCK_MAIN,
-						     BIF_ARG_2);
-	if (old_value != THE_NON_VALUE) {
-	    BIF_RET(old_value);
-	}
+        ErtsTracer new_seq_tracer, old_seq_tracer;
+        Eterm ret;
+
+        if (BIF_ARG_2 == am_false)
+            new_seq_tracer = erts_tracer_nil;
+        else
+            new_seq_tracer = erts_term_to_tracer(THE_NON_VALUE, BIF_ARG_2);
+
+        if (new_seq_tracer == THE_NON_VALUE)
+            goto error;
+
+        old_seq_tracer = erts_set_system_seq_tracer(BIF_P,
+                                                    ERTS_PROC_LOCK_MAIN,
+                                                    new_seq_tracer);
+
+        ERTS_TRACER_CLEAR(&new_seq_tracer);
+
+        if (old_seq_tracer == THE_NON_VALUE)
+            goto error;
+
+        if (ERTS_TRACER_IS_NIL(old_seq_tracer))
+            BIF_RET(am_false);
+
+        ret = erts_tracer_to_term(BIF_P, old_seq_tracer);
+
+        ERTS_TRACER_CLEAR(&old_seq_tracer);
+
+        BIF_RET(ret);
     } else if (BIF_ARG_1 == make_small(1)) {
 	int i, max;
 	ErtsMessage* mp;
@@ -4679,7 +4704,7 @@ skip_current_msgq(Process *c_p)
 	res = 0;
     }
     else {
-	ERTS_SMP_MSGQ_MV_INQ2PRIVQ(c_p);
+        ERTS_SMP_MSGQ_MV_INQ2PRIVQ(c_p);
 	c_p->msg.save = c_p->msg.last;
 	res = 1;
     }
diff --git a/erts/emulator/beam/bif.tab b/erts/emulator/beam/bif.tab
index 9bb77190d6..58cd31cee9 100644
--- a/erts/emulator/beam/bif.tab
+++ b/erts/emulator/beam/bif.tab
@@ -181,9 +181,8 @@ bif erlang:port_set_data/2
 bif erlang:port_get_data/1
 
 # Tracing & debugging.
-bif erlang:trace_pattern/2
-bif erlang:trace_pattern/3
-bif erlang:trace/3
+bif erts_internal:trace_pattern/3
+bif erts_internal:trace/3
 bif erlang:trace_info/2
 bif erlang:trace_delivered/1
 bif erlang:seq_trace/2
diff --git a/erts/emulator/beam/code_ix.c b/erts/emulator/beam/code_ix.c
index 9da750e366..ec6267711b 100644
--- a/erts/emulator/beam/code_ix.c
+++ b/erts/emulator/beam/code_ix.c
@@ -94,6 +94,7 @@ void erts_commit_staging_code_ix(void)
     ix = (ix + 1) % ERTS_NUM_CODE_IX;
     erts_smp_atomic32_set_nob(&the_staging_code_index, ix);
     export_staging_unlock();
+    erts_tracer_nif_clear();
     CIX_TRACE("activate");
 }
 
diff --git a/erts/emulator/beam/dist.c b/erts/emulator/beam/dist.c
index 88449adb8e..52fd57e101 100644
--- a/erts/emulator/beam/dist.c
+++ b/erts/emulator/beam/dist.c
@@ -337,7 +337,7 @@ static void doit_link_net_exits_sub(ErtsLink *sublnk, void *vlnecp)
 	    erts_destroy_link(rlnk);
 	    if (xres >= 0 && IS_TRACED_FL(rp, F_TRACE_PROCS)) {
 		/* We didn't exit the process and it is traced */
-		trace_proc(NULL, rp, am_getting_unlinked, sublnk->pid);
+		trace_proc(NULL, 0, rp, am_getting_unlinked, sublnk->pid);
 	    }
 	}
 	erts_smp_proc_unlock(rp, rp_locks);
@@ -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, NIL);
+	    erts_queue_message(rp, &rp_locks, msgp, tup);
 	}
 	erts_smp_proc_unlock(rp, rp_locks);
     }
@@ -1275,7 +1275,7 @@ int erts_net_message(Port *prt,
 	erts_smp_de_links_unlock(dep);
 
 	if (IS_TRACED_FL(rp, F_TRACE_PROCS))
-	    trace_proc(NULL, rp, am_getting_linked, from);
+	    trace_proc(NULL, 0, rp, am_getting_linked, from);
 
 	erts_smp_proc_unlock(rp, ERTS_PROC_LOCK_LINK);
 	break;
@@ -1300,7 +1300,7 @@ int erts_net_message(Port *prt,
 	lnk = erts_remove_link(&ERTS_P_LINKS(rp), from);
 
 	if (IS_TRACED_FL(rp, F_TRACE_PROCS) && lnk != NULL) {
-	    trace_proc(NULL, rp, am_getting_unlinked, from);
+	    trace_proc(NULL, 0, rp, am_getting_unlinked, from);
 	}
 
 	erts_smp_proc_unlock(rp, ERTS_PROC_LOCK_LINK);
@@ -1628,7 +1628,11 @@ int erts_net_message(Port *prt,
 					     ERTS_XSIG_FLG_IGN_KILL);
 		if (xres >= 0 && IS_TRACED_FL(rp, F_TRACE_PROCS)) {
 		    /* We didn't exit the process and it is traced */
-		    trace_proc(NULL, rp, am_getting_unlinked, from);
+                    if (rp_locks & ERTS_PROC_LOCKS_XSIG_SEND) {
+                        erts_smp_proc_unlock(rp, ERTS_PROC_LOCKS_XSIG_SEND);
+                        rp_locks &= ~ERTS_PROC_LOCKS_XSIG_SEND;
+                    }
+		    trace_proc(NULL, 0, rp, am_getting_unlinked, from);
 		}
 	    }
 	    erts_smp_proc_unlock(rp, rp_locks);
@@ -1719,7 +1723,7 @@ decode_error:
     }
 data_error:
     UnUseTmpHeapNoproc(DIST_CTL_DEFAULT_SIZE);
-    erts_deliver_port_exit(prt, dep->cid, am_killed, 0);
+    erts_deliver_port_exit(prt, dep->cid, am_killed, 0, 1);
     ERTS_SMP_CHK_NO_PROC_LOCKS;
     return -1;
 }
@@ -2089,7 +2093,7 @@ erts_dist_command(Port *prt, int reds_limit)
     erts_smp_de_runlock(dep);
 
     if (status & ERTS_DE_SFLG_EXITING) {
-	erts_deliver_port_exit(prt, prt->common.id, am_killed, 0);
+	erts_deliver_port_exit(prt, prt->common.id, am_killed, 0, 1);
 	erts_deref_dist_entry(dep);
 	return reds + ERTS_PORT_REDS_DIST_CMD_EXIT;
     }
@@ -3313,7 +3317,7 @@ send_nodes_mon_msg(Process *rp,
     }
 
     ASSERT(hend == hp);
-    erts_queue_message(rp, rp_locksp, mp, msg, NIL);
+    erts_queue_message(rp, rp_locksp, mp, msg);
 }
 
 static void
diff --git a/erts/emulator/beam/erl_alloc.c b/erts/emulator/beam/erl_alloc.c
index cfe0bc3205..a6519bd9e4 100644
--- a/erts/emulator/beam/erl_alloc.c
+++ b/erts/emulator/beam/erl_alloc.c
@@ -3210,7 +3210,7 @@ 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, NIL);
+    erts_queue_message(rp, &rp_locks, mp, msg);
 
     if (air->req_sched == sched_id)
 	rp_locks &= ~ERTS_PROC_LOCK_MAIN;
diff --git a/erts/emulator/beam/erl_alloc.types b/erts/emulator/beam/erl_alloc.types
index 9482ab9265..d2fe440d47 100644
--- a/erts/emulator/beam/erl_alloc.types
+++ b/erts/emulator/beam/erl_alloc.types
@@ -270,6 +270,9 @@ type	PROC_SYS_TSK	SHORT_LIVED	PROCESSES	proc_sys_task
 type	PROC_SYS_TSK_QS	SHORT_LIVED	PROCESSES	proc_sys_task_queues
 type    NEW_TIME_OFFSET	SHORT_LIVED	SYSTEM		new_time_offset
 type	IOB_REQ		SHORT_LIVED	SYSTEM		io_bytes_request
+type    TRACER_NIF      LONG_LIVED      SYSTEM          tracer_nif
+type    TRACE_MSG_QUEUE SHORT_LIVED     SYSTEM          trace_message_queue
+type    SCHED_ASYNC_JOB SHORT_LIVED     SYSTEM          async_calls
 
 +if threads_no_smp
 # Need thread safe allocs, but std_alloc and fix_alloc are not;
diff --git a/erts/emulator/beam/erl_bif_ddll.c b/erts/emulator/beam/erl_bif_ddll.c
index a79ce11563..1e2db38442 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, am_undefined);
+    erts_queue_message(proc, &rp_locks, mp, mess);
     erts_smp_proc_unlock(proc, rp_locks);
     ERTS_SMP_CHK_NO_PROC_LOCKS;
 }
diff --git a/erts/emulator/beam/erl_bif_info.c b/erts/emulator/beam/erl_bif_info.c
index 68f6abfcdc..7a72b0d8cc 100644
--- a/erts/emulator/beam/erl_bif_info.c
+++ b/erts/emulator/beam/erl_bif_info.c
@@ -846,6 +846,7 @@ process_info_list(Process *c_p, Eterm pid, Eterm list, int always_wrap,
 
 	if (unlock_locks)
 	    erts_smp_proc_unlock(rp, unlock_locks);
+
     }
 
     /*
@@ -2162,8 +2163,8 @@ BIF_RETTYPE system_info_1(BIF_ALIST_1)
 	res = build_snifs_term(&hp, NULL, NIL);
 	BIF_RET(res);
     } else if (BIF_ARG_1 == am_sequential_tracer) {
-	val = erts_get_system_seq_tracer();
-	ASSERT(is_internal_pid(val) || is_internal_port(val) || val==am_false);
+	ErtsTracer seq_tracer = erts_get_system_seq_tracer();
+        val = erts_tracer_to_term(BIF_P, seq_tracer);
 	hp = HAlloc(BIF_P, 3);
 	res = TUPLE2(hp, am_sequential_tracer, val);
 	BIF_RET(res);
diff --git a/erts/emulator/beam/erl_bif_port.c b/erts/emulator/beam/erl_bif_port.c
index 71b6d78094..37f4e1de49 100644
--- a/erts/emulator/beam/erl_bif_port.c
+++ b/erts/emulator/beam/erl_bif_port.c
@@ -79,6 +79,7 @@ BIF_RETTYPE erts_internal_open_port_2(BIF_ALIST_2)
     }
 
     if (port->drv_ptr->flags & ERL_DRV_FLAG_USE_INIT_ACK) {
+
         /* Copied from erl_port_task.c */
         port->async_open_port = erts_alloc(ERTS_ALC_T_PRTSD,
                                            sizeof(*port->async_open_port));
@@ -109,6 +110,10 @@ BIF_RETTYPE erts_internal_open_port_2(BIF_ALIST_2)
     erts_add_link(&ERTS_P_LINKS(port), LINK_PID, BIF_P->common.id);
     erts_add_link(&ERTS_P_LINKS(BIF_P), LINK_PID, port->common.id);
 
+    if (IS_TRACED_FL(BIF_P, F_TRACE_PROCS))
+        trace_proc(BIF_P, ERTS_PROC_LOCK_MAIN|ERTS_PROC_LOCK_LINK, BIF_P,
+                   am_link, port->common.id);
+
     erts_smp_proc_unlock(BIF_P, ERTS_PROC_LOCK_LINK);
 
     erts_port_release(port);
@@ -339,8 +344,7 @@ BIF_RETTYPE erts_internal_port_close_1(BIF_ALIST_1)
     if (!prt)
 	BIF_RET(am_badarg);
 
-
-    switch (erts_port_exit(BIF_P, 0, prt, prt->common.id, am_normal, &ref)) {
+    switch (erts_port_exit(BIF_P, 0, prt, BIF_P->common.id, am_normal, &ref)) {
     case ERTS_PORT_OP_CALLER_EXIT:
     case ERTS_PORT_OP_BADARG:
     case ERTS_PORT_OP_DROPPED:
@@ -373,7 +377,7 @@ BIF_RETTYPE erts_internal_port_connect_2(BIF_ALIST_2)
     ref = NIL;
 #endif
 
-    switch (erts_port_connect(BIF_P, 0, prt, prt->common.id, BIF_ARG_2, &ref)) {
+    switch (erts_port_connect(BIF_P, 0, prt, BIF_P->common.id, BIF_ARG_2, &ref)) {
     case ERTS_PORT_OP_CALLER_EXIT:
     case ERTS_PORT_OP_BADARG:
     case ERTS_PORT_OP_DROPPED:
@@ -911,7 +915,7 @@ open_port(Process* p, Eterm name, Eterm settings, int *err_typep, int *err_nump)
     }
     
     if (IS_TRACED_FL(p, F_TRACE_SCHED_PROCS)) {
-        trace_virtual_sched(p, am_out);
+        trace_sched(p, ERTS_PROC_LOCK_MAIN, am_out);
     }
     
 
@@ -928,21 +932,22 @@ open_port(Process* p, Eterm name, Eterm settings, int *err_typep, int *err_nump)
         DTRACE3(port_open, process_str, name_buf, port_str);
     }
 #endif
+
+    if (port && IS_TRACED_FL(port, F_TRACE_PORTS))
+        trace_port(port, am_getting_linked, p->common.id);
+
     erts_smp_proc_lock(p, ERTS_PROC_LOCK_MAIN);
 
+    if (IS_TRACED_FL(p, F_TRACE_SCHED_PROCS)) {
+        trace_sched(p, ERTS_PROC_LOCK_MAIN, am_in);
+    }
+
     if (!port) {
 	DEBUGF(("open_driver returned (%d:%d)\n",
 		err_typep ? *err_typep : 4711,
 		err_nump ? *err_nump : 4711));
-    	if (IS_TRACED_FL(p, F_TRACE_SCHED_PROCS)) {
-            trace_virtual_sched(p, am_in);
-    	}
 	goto do_return;
     }
-    
-    if (IS_TRACED_FL(p, F_TRACE_SCHED_PROCS)) {
-        trace_virtual_sched(p, am_in);
-    }
 
     if (linebuf && port->linebuf == NULL){
 	port->linebuf = allocate_linebuf(linebuf);
diff --git a/erts/emulator/beam/erl_bif_trace.c b/erts/emulator/beam/erl_bif_trace.c
index a9443ee8df..ff2018aa27 100644
--- a/erts/emulator/beam/erl_bif_trace.c
+++ b/erts/emulator/beam/erl_bif_trace.c
@@ -52,7 +52,7 @@ static int                         erts_default_trace_pattern_is_on;
 static Binary                     *erts_default_match_spec;
 static Binary                     *erts_default_meta_match_spec;
 static struct trace_pattern_flags  erts_default_trace_pattern_flags;
-static Eterm                       erts_default_meta_tracer_pid;
+static ErtsTracer                  erts_default_meta_tracer;
 
 static struct {			/* Protected by code write permission */
     int current;
@@ -75,8 +75,6 @@ static BIF_RETTYPE
 system_monitor(Process *p, Eterm monitor_pid, Eterm list);
 
 static void new_seq_trace_token(Process* p); /* help func for seq_trace_2*/
-static int already_traced(Process *p, Process *tracee_p, Eterm tracer);
-static int port_already_traced(Process *p, Port *tracee_port, Eterm tracer);
 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);
@@ -94,21 +92,15 @@ erts_bif_trace_init(void)
     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_pid = NIL;
+    erts_default_meta_tracer = erts_tracer_nil;
 }
 
 /*
  * Turn on/off call tracing for the given function(s).
  */
-  
-Eterm
-trace_pattern_2(BIF_ALIST_2)
-{
-    return trace_pattern(BIF_P, BIF_ARG_1, BIF_ARG_2, NIL);
-}
 
 Eterm
-trace_pattern_3(BIF_ALIST_3)
+erts_internal_trace_pattern_3(BIF_ALIST_3)
 {
     return trace_pattern(BIF_P, BIF_ARG_1, BIF_ARG_2, BIF_ARG_3);
 }
@@ -125,11 +117,10 @@ trace_pattern(Process* p, Eterm MFA, Eterm Pattern, Eterm flaglist)
     Eterm l;
     struct trace_pattern_flags flags = erts_trace_pattern_flags_off;
     int is_global;
-    Process *meta_tracer_proc = p;
-    Eterm meta_tracer_pid = p->common.id;
+    ErtsTracer meta_tracer = erts_tracer_nil;
 
     if (!erts_try_seize_code_write_permission(p)) {
-	ERTS_BIF_YIELD3(bif_export[BIF_trace_pattern_3], p, MFA, Pattern, flaglist);
+	ERTS_BIF_YIELD3(bif_export[BIF_erts_internal_trace_pattern_3], p, MFA, Pattern, flaglist);
     }
     finish_bp.current = -1;
 
@@ -160,31 +151,11 @@ trace_pattern(Process* p, Eterm MFA, Eterm Pattern, Eterm flaglist)
     is_global = 0;
     for(l = flaglist; is_list(l); l = CDR(list_val(l))) {
 	if (is_tuple(CAR(list_val(l)))) {
-	    Eterm *tp = tuple_val(CAR(list_val(l)));
-	    
-	    if (arityval(tp[0]) != 2 || tp[1] != am_meta) {
-		goto error;
-	    }
-	    meta_tracer_pid = tp[2];
-	    if (is_internal_pid(meta_tracer_pid)) {
-		meta_tracer_proc = erts_pid2proc(NULL, 0, meta_tracer_pid, 0);
-		if (!meta_tracer_proc) {
-		    goto error;
-		}
-	    } else if (is_internal_port(meta_tracer_pid)) {
-		Port *meta_tracer_port;
-		meta_tracer_proc = NULL;
-		meta_tracer_port = (erts_port_lookup(
-					meta_tracer_pid,
-					ERTS_PORT_SFLGS_INVALID_TRACER_LOOKUP));
-		if (!meta_tracer_port)
-		    goto error;
-	    } else {
-		goto error;
-	    }
-	    if (is_global) {
-		goto error;
-	    }
+            meta_tracer = erts_term_to_tracer(am_meta, CAR(list_val(l)));
+            if (meta_tracer == THE_NON_VALUE) {
+                meta_tracer = erts_tracer_nil;
+                goto error;
+            }
 	    flags.breakpoint = 1;
 	    flags.meta       = 1;
 	} else {
@@ -202,6 +173,8 @@ trace_pattern(Process* p, Eterm MFA, Eterm Pattern, Eterm flaglist)
 		}
 		flags.breakpoint = 1;
 		flags.meta       = 1;
+                if (ERTS_TRACER_IS_NIL(meta_tracer))
+                    meta_tracer = erts_term_to_tracer(THE_NON_VALUE, p->common.id);
 		break;
 	    case am_global:
 		if (flags.breakpoint) {
@@ -252,14 +225,11 @@ trace_pattern(Process* p, Eterm MFA, Eterm Pattern, Eterm flaglist)
 	    MatchSetUnref(erts_default_meta_match_spec);
 	    erts_default_meta_match_spec = match_prog_set;
 	    MatchSetRef(erts_default_meta_match_spec);
-	    erts_default_meta_tracer_pid = meta_tracer_pid;
-	    if (meta_tracer_proc) {
-		ERTS_TRACE_FLAGS(meta_tracer_proc) |= F_TRACER;
-	    }
+            erts_tracer_update(&erts_default_meta_tracer, meta_tracer);
 	} else if (! flags.breakpoint) {
 	    MatchSetUnref(erts_default_meta_match_spec);
 	    erts_default_meta_match_spec = NULL;
-	    erts_default_meta_tracer_pid = NIL;
+	    ERTS_TRACER_CLEAR(&erts_default_meta_tracer);
 	}
 	if (erts_default_trace_pattern_flags.breakpoint &&
 	    flags.breakpoint) { 
@@ -340,20 +310,18 @@ trace_pattern(Process* p, Eterm MFA, Eterm Pattern, Eterm flaglist)
 	if (is_small(mfa[2])) {
 	    mfa[2] = signed_val(mfa[2]);
 	}
-    
-	if (meta_tracer_proc) {
-	    ERTS_TRACE_FLAGS(meta_tracer_proc) |= F_TRACER;
-	}
 
 	matches = erts_set_trace_pattern(p, mfa, specified,
 					 match_prog_set, match_prog_set,
-					 on, flags, meta_tracer_pid, 0);
+					 on, flags, meta_tracer, 0);
     }
 
  error:
     MatchSetUnref(match_prog_set);
     UnUseTmpHeap(3,p);
 
+    ERTS_TRACER_CLEAR(&meta_tracer);
+
 #ifdef ERTS_SMP
     if (finish_bp.current >= 0) {
 	ASSERT(matches >= 0);
@@ -404,7 +372,7 @@ erts_get_default_trace_pattern(int *trace_pattern_is_on,
 			       Binary **match_spec,
 			       Binary **meta_match_spec,
 			       struct trace_pattern_flags *trace_pattern_flags,
-			       Eterm *meta_tracer_pid)
+			       ErtsTracer *meta_tracer)
 {
     ERTS_SMP_LC_ASSERT(erts_has_code_write_permission() ||
 		       erts_smp_thr_progress_is_blocking());
@@ -416,8 +384,8 @@ erts_get_default_trace_pattern(int *trace_pattern_is_on,
 	*meta_match_spec = erts_default_meta_match_spec;
     if (trace_pattern_flags)
 	*trace_pattern_flags = erts_default_trace_pattern_flags;
-    if (meta_tracer_pid)
-	*meta_tracer_pid = erts_default_meta_tracer_pid;
+    if (meta_tracer)
+	*meta_tracer = erts_default_meta_tracer;
 }
 
 int erts_is_default_trace_enabled(void)
@@ -465,12 +433,12 @@ erts_trace_flag2bit(Eterm flag)
 ** occurred in the argument list.
 */
 int
-erts_trace_flags(Eterm List, 
-		 Uint *pMask, Eterm *pTracer, int *pCpuTimestamp)
+erts_trace_flags(Eterm List,
+                 Uint *pMask, ErtsTracer *pTracer, int *pCpuTimestamp)
 {
     Eterm list = List;
     Uint mask = 0;
-    Eterm tracer = NIL;
+    ErtsTracer tracer = erts_tracer_nil;
     int cpu_timestamp = 0;
     
     while (is_list(list)) {
@@ -483,33 +451,73 @@ erts_trace_flags(Eterm List,
 	    cpu_timestamp = !0;
 #endif
 	} else if (is_tuple(item)) {
-	    Eterm* tp = tuple_val(item);
-	    
-	    if (arityval(tp[0]) != 2 || tp[1] != am_tracer) goto error;
-	    if (is_internal_pid(tp[2]) || is_internal_port(tp[2])) {
-		tracer = tp[2];
-	    } else goto error;
+            tracer = erts_term_to_tracer(am_tracer, item);
+            if (tracer == THE_NON_VALUE)
+                goto error;
 	} else goto error;
 	list = CDR(list_val(list));
     }
     if (is_not_nil(list)) goto error;
     
-    if (pMask && mask)                  *pMask         = mask;
-    if (pTracer && tracer != NIL)       *pTracer       = tracer;
-    if (pCpuTimestamp && cpu_timestamp) *pCpuTimestamp = cpu_timestamp;
+    if (pMask && mask)                           *pMask         = mask;
+    if (pTracer && !ERTS_TRACER_IS_NIL(tracer))  *pTracer       = tracer;
+    if (pCpuTimestamp && cpu_timestamp)          *pCpuTimestamp = cpu_timestamp;
     return !0;
  error:
     return 0;
 }
 
-Eterm trace_3(BIF_ALIST_3)
+static ERTS_INLINE int
+start_trace(Process *c_p, ErtsTracer tracer,
+            ErtsPTabElementCommon *common,
+            int on, int mask)
+{
+    /* We can use the common part of both port+proc without checking what it is
+       In the code below port is used for both proc and port */
+    Port *port = (Port*)common;
+
+    /*
+     * SMP build assumes that either system is blocked or:
+     * * main lock is held on c_p
+     * * all locks are held on port common
+     */
+
+    if (!ERTS_TRACER_IS_NIL(tracer)) {
+        if ((ERTS_TRACE_FLAGS(port) & TRACEE_FLAGS)
+            && !ERTS_TRACER_COMPARE(ERTS_TRACER(port), tracer)) {
+            /* This tracee is already being traced, and not by the
+             * tracer to be */
+            if (erts_is_tracer_proc_enabled(c_p, ERTS_PROC_LOCKS_ALL,
+                                            common, am_trace_status)) {
+                /* The tracer is still in use */
+                return 1;
+            }
+            /* Current tracer now invalid */
+        }
+    }
+
+    if (on)
+        ERTS_TRACE_FLAGS(port) |= mask;
+    else
+        ERTS_TRACE_FLAGS(port) &= ~mask;
+
+    if ((ERTS_TRACE_FLAGS(port) & TRACEE_FLAGS) == 0) {
+        tracer = erts_tracer_nil;
+        erts_tracer_replace(common, erts_tracer_nil);
+    } else if (!ERTS_TRACER_IS_NIL(tracer))
+        erts_tracer_replace(common, tracer);
+
+    return 0;
+}
+
+Eterm erts_internal_trace_3(BIF_ALIST_3)
 {
     Process* p = BIF_P;
     Eterm pid_spec = BIF_ARG_1;
     Eterm how = BIF_ARG_2;
     Eterm list = BIF_ARG_3;
     int on;
-    Eterm tracer = NIL;
+    ErtsTracer tracer = erts_tracer_nil;
     int matches = 0;
     Uint mask = 0;
     int cpu_ts = 0;
@@ -522,41 +530,24 @@ Eterm trace_3(BIF_ALIST_3)
     }
 
     if (!erts_try_seize_code_write_permission(BIF_P)) {
-	ERTS_BIF_YIELD3(bif_export[BIF_trace_3], BIF_P, BIF_ARG_1, BIF_ARG_2, BIF_ARG_3);
+	ERTS_BIF_YIELD3(bif_export[BIF_erts_internal_trace_3],
+                        BIF_P, BIF_ARG_1, BIF_ARG_2, BIF_ARG_3);
     }
 
-    if (is_nil(tracer) || is_internal_pid(tracer)) {
-	Process *tracer_proc = erts_pid2proc(p,
-					     ERTS_PROC_LOCK_MAIN,
-					     is_nil(tracer) ? p->common.id : tracer,
-					     ERTS_PROC_LOCKS_ALL);
-	if (!tracer_proc)
-	    goto error;
-	ERTS_TRACE_FLAGS(tracer_proc) |= F_TRACER;
-	erts_smp_proc_unlock(tracer_proc,
-			     (tracer_proc == p
-			      ? ERTS_PROC_LOCKS_ALL_MINOR
-			      : ERTS_PROC_LOCKS_ALL));
-    } else if (is_internal_port(tracer)) {
-	Port *tracer_port = erts_id2port_sflgs(tracer,
-					       p,
-					       ERTS_PROC_LOCK_MAIN,
-					       ERTS_PORT_SFLGS_INVALID_TRACER_LOOKUP);
-	if (!tracer_port)
-	    goto error;
-	ERTS_TRACE_FLAGS(tracer_port) |= F_TRACER;
-	erts_port_release(tracer_port);
-    } else
-	goto error;
-
     switch (how) {
     case am_false: 
 	on = 0; 
 	break;
     case am_true: 
 	on = 1;
-	if (is_nil(tracer))
-	    tracer = p->common.id;
+        if (ERTS_TRACER_IS_NIL(tracer))
+            tracer = erts_term_to_tracer(am_tracer, p->common.id);
+
+        if (tracer == THE_NON_VALUE) {
+            tracer = erts_tracer_nil;
+            goto error;
+        }
+
 	break;
     default: 
 	goto error;
@@ -575,34 +566,20 @@ Eterm trace_3(BIF_ALIST_3)
 	}
 #endif
 
-	if (pid_spec == tracer)
-	    goto error;
-
 	tracee_port = erts_id2port_sflgs(pid_spec,
 					 p,
 					 ERTS_PROC_LOCK_MAIN,
 					 ERTS_PORT_SFLGS_INVALID_LOOKUP);
+
 	if (!tracee_port)
 	    goto error;
-	
-	if (tracer != NIL && port_already_traced(p, tracee_port, tracer)) {
+
+        if (start_trace(p, tracer, &tracee_port->common, on, mask)) {
 	    erts_port_release(tracee_port);
 	    goto already_traced;
-	}
-
-	if (on)
-	    ERTS_TRACE_FLAGS(tracee_port) |= mask;
-	else
-	    ERTS_TRACE_FLAGS(tracee_port) &= ~mask;
-
-	if (!ERTS_TRACE_FLAGS(tracee_port))
-	    ERTS_TRACER_PROC(tracee_port) = NIL;
-	else if (tracer != NIL)
-	    ERTS_TRACER_PROC(tracee_port) = tracer;
-
-	erts_port_release(tracee_port);
-
-	matches = 1;
+        }
+        erts_port_release(tracee_port);
+        matches = 1;
     } else if (is_pid(pid_spec)) {
 	Process *tracee_p;
 
@@ -615,33 +592,19 @@ Eterm trace_3(BIF_ALIST_3)
 	 * and not about to be tracing.
 	 */
 
-	if (pid_spec == tracer)
-	    goto error;
-
 	tracee_p = erts_pid2proc(p, ERTS_PROC_LOCK_MAIN,
 				 pid_spec, ERTS_PROC_LOCKS_ALL);
 	if (!tracee_p)
 	    goto error;
 
-	if (tracer != NIL && already_traced(p, tracee_p, tracer)) {
+        if (start_trace(tracee_p, tracer, &tracee_p->common, on, mask)) {
 	    erts_smp_proc_unlock(tracee_p,
 				 (tracee_p == p
 				  ? ERTS_PROC_LOCKS_ALL_MINOR
 				  : ERTS_PROC_LOCKS_ALL));
 	    goto already_traced;
-	}
-
-	if (on)
-	    ERTS_TRACE_FLAGS(tracee_p) |= mask;
-	else
-	    ERTS_TRACE_FLAGS(tracee_p) &= ~mask;
-
-	if ((ERTS_TRACE_FLAGS(tracee_p) & TRACEE_FLAGS) == 0)
-	    ERTS_TRACER_PROC(tracee_p) = NIL;
-	else if (tracer != NIL)
-	    ERTS_TRACER_PROC(tracee_p) = tracer;
-
-	erts_smp_proc_unlock(tracee_p,
+        }
+        erts_smp_proc_unlock(tracee_p,
 			     (tracee_p == p
 			      ? ERTS_PROC_LOCKS_ALL_MINOR
 			      : ERTS_PROC_LOCKS_ALL));
@@ -692,18 +655,27 @@ Eterm trace_3(BIF_ALIST_3)
 	}
 #endif
 	
-	if (pid_spec == am_all || pid_spec == am_existing) {
+	if (pid_spec == am_all || pid_spec == am_existing ||
+            pid_spec == am_ports || pid_spec == am_processes ||
+            pid_spec == am_existing_ports || pid_spec == am_existing_processes
+            ) {
 	    int i;
 	    int procs = 0;
 	    int ports = 0;
 	    int mods = 0;
 
 	    if (mask & (ERTS_PROC_TRACEE_FLAGS & ~ERTS_TRACEE_MODIFIER_FLAGS))
-		procs = 1;
+		procs = pid_spec != am_ports && pid_spec != am_existing_ports;
 	    if (mask & (ERTS_PORT_TRACEE_FLAGS & ~ERTS_TRACEE_MODIFIER_FLAGS))
-		ports = 1;
-	    if (mask & ERTS_TRACEE_MODIFIER_FLAGS)
-		mods = 1;
+		ports = pid_spec != am_processes && pid_spec != am_existing_processes;
+	    if (mask & ERTS_TRACEE_MODIFIER_FLAGS) {
+                if (pid_spec == am_ports || pid_spec == am_existing_ports)
+                    ports = 1;
+                else if (pid_spec == am_processes || pid_spec == am_existing_processes)
+                    procs = 1;
+                else
+                    mods = 1;
+            }
 
 #ifdef ERTS_SMP
 	    erts_smp_proc_unlock(p, ERTS_PROC_LOCK_MAIN);
@@ -719,22 +691,7 @@ Eterm trace_3(BIF_ALIST_3)
 		    Process* tracee_p = erts_pix2proc(i);
 		    if (! tracee_p) 
 			continue;
-		    if (tracer != NIL) {
-			if (tracee_p->common.id == tracer)
-			    continue;
-			if (already_traced(NULL, tracee_p, tracer))
-			    continue;
-		    }
-		    if (on) {
-			ERTS_TRACE_FLAGS(tracee_p) |= mask;
-		    } else {
-			ERTS_TRACE_FLAGS(tracee_p) &= ~mask;
-		    }
-		    if(!(ERTS_TRACE_FLAGS(tracee_p) & TRACEE_FLAGS)) {
-			ERTS_TRACER_PROC(tracee_p) = NIL;
-		    } else if (tracer != NIL) {
-			ERTS_TRACER_PROC(tracee_p) = tracer;
-		    }
+                    start_trace(p, tracer, &tracee_p->common, on, mask);
 		    matches++;
 		}
 	    }
@@ -749,33 +706,26 @@ Eterm trace_3(BIF_ALIST_3)
 		    state = erts_atomic32_read_nob(&tracee_port->state);
 		    if (state & ERTS_PORT_SFLGS_DEAD)
 			continue;
-		    if (tracer != NIL) {
-			if (tracee_port->common.id == tracer)
-			    continue;
-			if (port_already_traced(NULL, tracee_port, tracer))
-			    continue;
-		    }
-
-		    if (on) ERTS_TRACE_FLAGS(tracee_port) |= mask;
-		    else ERTS_TRACE_FLAGS(tracee_port) &= ~mask;
-		
-		    if (!(ERTS_TRACE_FLAGS(tracee_port) & TRACEE_FLAGS)) {
-			ERTS_TRACER_PROC(tracee_port) = NIL;
-		    } else if (tracer != NIL) {
-			ERTS_TRACER_PROC(tracee_port) = tracer;
-		    }
-		    /* matches are not counted for ports since it would violate compatibility */
-		    /* This could be a reason to modify this function or make a new one. */
+                    start_trace(p, tracer, &tracee_port->common, on, mask);
+                    matches++;
 		}
 	    }
 	}
 
-	if (pid_spec == am_all || pid_spec == am_new) {
-	    Uint def_flags = mask;
-	    Eterm def_tracer = tracer;
+	if (pid_spec == am_all || pid_spec == am_new
+            || pid_spec == am_ports || pid_spec == am_processes
+            || pid_spec == am_new_ports || pid_spec == am_new_processes
+            ) {
 
 	    ok = 1;
-	    erts_change_default_tracing(on, &def_flags, &def_tracer);
+            if (mask & ERTS_PROC_TRACEE_FLAGS &&
+                pid_spec != am_ports && pid_spec != am_new_ports)
+                erts_change_default_proc_tracing(
+                    on, mask & ERTS_PROC_TRACEE_FLAGS, tracer);
+            if (mask & ERTS_PORT_TRACEE_FLAGS &&
+                pid_spec != am_processes && pid_spec != am_new_processes)
+                erts_change_default_port_tracing(
+                    on, mask & ERTS_PORT_TRACEE_FLAGS, tracer);
 
 #ifdef HAVE_ERTS_NOW_CPU
 	    if (cpu_ts && !on) {
@@ -801,6 +751,7 @@ Eterm trace_3(BIF_ALIST_3)
     }
 #endif
     erts_release_code_write_permission();
+    ERTS_TRACER_CLEAR(&tracer);
 
     BIF_RET(make_small(matches));
 
@@ -810,6 +761,8 @@ Eterm trace_3(BIF_ALIST_3)
 
  error:
 
+    ERTS_TRACER_CLEAR(&tracer);
+
 #ifdef ERTS_SMP
     if (system_blocked) {
 	erts_smp_thr_progress_unblock();
@@ -821,88 +774,6 @@ Eterm trace_3(BIF_ALIST_3)
     BIF_ERROR(p, BADARG);
 }
 
-/* Check that the process to be traced is not already traced
- * by a valid other tracer than the tracer to be.
- */
-static int port_already_traced(Process *c_p, Port *tracee_port, Eterm tracer)
-{
-    /*
-     * SMP build assumes that either system is blocked or:
-     * * main lock is held on c_p
-     * * all locks are held on port tracee_p
-     */
-    if ((ERTS_TRACE_FLAGS(tracee_port) & TRACEE_FLAGS)
-	&& ERTS_TRACER_PROC(tracee_port) != tracer) {
-	/* This tracee is already being traced, and not by the 
-	 * tracer to be */
-	if (is_internal_port(ERTS_TRACER_PROC(tracee_port))) {
-	    if (!erts_is_valid_tracer_port(ERTS_TRACER_PROC(tracee_port))) {
-		/* Current trace port now invalid 
-		 * - discard it and approve the new. */
-		goto remove_tracer;
-	    } else
-		return 1;
-	}
-	else if(is_internal_pid(ERTS_TRACER_PROC(tracee_port))) {
-	    Process *tracer_p = erts_proc_lookup(ERTS_TRACER_PROC(tracee_port));
-	    if (!tracer_p) {
-		/* Current trace process now invalid
-		 * - discard it and approve the new. */
-		goto remove_tracer;
-	    } else
-		return 1;
-	}
-	else {
-	remove_tracer:
-	    ERTS_TRACE_FLAGS(tracee_port) &= ~TRACEE_FLAGS;
-	    ERTS_TRACER_PROC(tracee_port) = NIL;
-	}
-    }
-    return 0;
-}
-
-/* Check that the process to be traced is not already traced
- * by a valid other tracer than the tracer to be.
- */
-static int already_traced(Process *c_p, Process *tracee_p, Eterm tracer)
-{
-    /*
-     * SMP build assumes that either system is blocked or:
-     * * main lock is held on c_p
-     * * all locks multiple are held on tracee_p
-     */
-    if ((ERTS_TRACE_FLAGS(tracee_p) & TRACEE_FLAGS)
-	&& ERTS_TRACER_PROC(tracee_p) != tracer) {
-	/* This tracee is already being traced, and not by the 
-	 * tracer to be */
-	if (is_internal_port(ERTS_TRACER_PROC(tracee_p))) {
-	    if (!erts_is_valid_tracer_port(ERTS_TRACER_PROC(tracee_p))) {
-		/* Current trace port now invalid 
-		 * - discard it and approve the new. */
-		goto remove_tracer;
-	    } else
-		return 1;
-	}
-	else if(is_internal_pid(ERTS_TRACER_PROC(tracee_p))) {
-	    Process *tracer_p;
-
-	    tracer_p = erts_proc_lookup(ERTS_TRACER_PROC(tracee_p));
-	    if (!tracer_p) {
-		/* Current trace process now invalid
-		 * - discard it and approve the new. */
-		goto remove_tracer;
-	    } else
-		return 1;
-	}
-	else {
-	remove_tracer:
-	    ERTS_TRACE_FLAGS(tracee_p) &= ~TRACEE_FLAGS;
-	    ERTS_TRACER_PROC(tracee_p) = NIL;
-	}
-    }
-    return 0;
-}
-
 /*
  * Return information about a process or an external function being traced.
  */
@@ -920,7 +791,7 @@ Eterm trace_info_2(BIF_ALIST_2)
 
     if (What == am_on_load) {
 	res = trace_info_on_load(p, Key);
-    } else if (is_atom(What) || is_pid(What)) {
+    } else if (is_atom(What) || is_pid(What) || is_port(What)) {
 	res = trace_info_pid(p, What, Key);
     } else if (is_tuple(What)) {
 	res = trace_info_func(p, What, Key);
@@ -936,41 +807,52 @@ static Eterm
 trace_info_pid(Process* p, Eterm pid_spec, Eterm key)
 {
     Eterm tracer;
-    Uint trace_flags;
+    Uint trace_flags = am_false;
     Eterm* hp;
 
-    if (pid_spec == am_new) {
-	erts_get_default_tracing(&trace_flags, &tracer);
+    if (pid_spec == am_new || pid_spec == am_new_processes) {
+        ErtsTracer def_tracer;
+	erts_get_default_proc_tracing(&trace_flags, &def_tracer);
+        tracer = erts_tracer_to_term(p, def_tracer);
+        ERTS_TRACER_CLEAR(&def_tracer);
+    } else if (pid_spec == am_new_ports) {
+        ErtsTracer def_tracer;
+	erts_get_default_port_tracing(&trace_flags, &def_tracer);
+        tracer = erts_tracer_to_term(p, def_tracer);
+        ERTS_TRACER_CLEAR(&def_tracer);
+    } else if (is_internal_port(pid_spec)) {
+        Port *tracee;
+        tracee = erts_id2port_sflgs(pid_spec, p, ERTS_PROC_LOCK_MAIN,
+                                    ERTS_PORT_SFLGS_INVALID_LOOKUP);
+
+        if (!tracee)
+            return am_undefined;
+
+        if (!ERTS_TRACER_IS_NIL(ERTS_TRACER(tracee)))
+            erts_is_tracer_proc_enabled(NULL, 0, &tracee->common, am_trace_status);
+
+        tracer = erts_tracer_to_term(p, ERTS_TRACER(tracee));
+        trace_flags = ERTS_TRACE_FLAGS(tracee);
+
+        erts_port_release(tracee);
+
     } else if (is_internal_pid(pid_spec)) {
 	Process *tracee;
 	tracee = erts_pid2proc(p, ERTS_PROC_LOCK_MAIN,
-			       pid_spec, ERTS_PROC_LOCKS_ALL);
+			       pid_spec, ERTS_PROC_LOCK_MAIN);
 
-	if (!tracee) {
+	if (!tracee)
 	    return am_undefined;
-	} else {
-	    tracer = ERTS_TRACER_PROC(tracee);
-	    trace_flags = ERTS_TRACE_FLAGS(tracee);
-	}
 
-	if (is_internal_pid(tracer)) {
-	    if (!erts_proc_lookup(tracer)) {
-	    reset_tracer:
-		ERTS_TRACE_FLAGS(tracee) &= ~TRACEE_FLAGS;
-		trace_flags = ERTS_TRACE_FLAGS(tracee);
-		tracer = ERTS_TRACER_PROC(tracee) = NIL;
-	    }
-	}
-	else if (is_internal_port(tracer)) {
-	    if (!erts_is_valid_tracer_port(tracer))
-		goto reset_tracer;
-	}
-#ifdef ERTS_SMP
-	erts_smp_proc_unlock(tracee,
-			     (tracee == p
-			      ? ERTS_PROC_LOCKS_ALL_MINOR
-			      : ERTS_PROC_LOCKS_ALL));
-#endif
+        if (!ERTS_TRACER_IS_NIL(ERTS_TRACER(tracee)))
+            erts_is_tracer_proc_enabled(tracee, ERTS_PROC_LOCK_MAIN,
+                                        &tracee->common, am_trace_status);
+
+        tracer = erts_tracer_to_term(p, ERTS_TRACER(tracee));
+        trace_flags = ERTS_TRACE_FLAGS(tracee);
+
+        if (tracee != p)
+            erts_smp_proc_unlock(tracee, ERTS_PROC_LOCK_MAIN);
     } else if (is_external_pid(pid_spec)
 	       && external_pid_dist_entry(pid_spec) == erts_this_dist_entry) {
 	    return am_undefined;
@@ -1024,8 +906,10 @@ trace_info_pid(Process* p, Eterm pid_spec, Eterm key)
 	HRelease(p,limit,hp+3);
 	return TUPLE2(hp, key, flag_list);
     } else if (key == am_tracer) {
-	hp = HAlloc(p, 3);
-	return TUPLE2(hp, key, tracer); /* Local pid or port */
+        if (tracer == am_false)
+            tracer = NIL;
+        hp = HAlloc(p, 3);
+        return TUPLE2(hp, key, tracer);
     } else {
 	goto error;
     }
@@ -1054,11 +938,11 @@ trace_info_pid(Process* p, Eterm pid_spec, Eterm key)
  */
 static int function_is_traced(Process *p,
 			      Eterm mfa[3],
-			      Binary **ms,              /* out */
-			      Binary **ms_meta,         /* out */
-			      Eterm   *tracer_pid_meta, /* out */
-			      Uint    *count,           /* out */
-			      Eterm   *call_time)       /* out */
+			      Binary    **ms,              /* out */
+			      Binary    **ms_meta,         /* out */
+			      ErtsTracer *tracer_pid_meta, /* out */
+			      Uint       *count,           /* out */
+			      Eterm      *call_time)       /* out */
 {
     Export e;
     Export* ep;
@@ -1123,7 +1007,7 @@ trace_info_func(Process* p, Eterm func_spec, Eterm key)
     Eterm traced = am_false;
     Eterm match_spec = am_false;
     Eterm retval = am_false;
-    Eterm meta = am_false;
+    ErtsTracer meta = erts_tracer_nil;
     Eterm call_time = NIL;
     int r;
 
@@ -1193,7 +1077,10 @@ trace_info_func(Process* p, Eterm func_spec, Eterm key)
 	retval = match_spec;
 	break;
     case am_meta:
-	retval = meta;
+        retval = erts_tracer_to_term(p, meta);
+        if (retval == am_false)
+            /* backwards compatibility */
+            retval = NIL;
 	break;
     case am_meta_match_spec:
 	if (r & FUNC_TRACE_META_TRACE) {
@@ -1216,7 +1103,8 @@ trace_info_func(Process* p, Eterm func_spec, Eterm key)
 	}
 	break;
     case am_all: {
-	Eterm match_spec_meta = am_false, c = am_false, t, ct = am_false;
+	Eterm match_spec_meta = am_false, c = am_false, t, ct = am_false,
+            m = am_false;
 	
 	if (ms) {
 	    match_spec = MatchSetGetSource(ms);
@@ -1235,6 +1123,9 @@ trace_info_func(Process* p, Eterm func_spec, Eterm key)
 	if (r & FUNC_TRACE_TIME_TRACE) {
 	    ct = call_time;
 	}
+
+        m = erts_tracer_to_term(p, meta);
+
 	hp = HAlloc(p, (3+2)*6);
 	retval = NIL;
 	t = TUPLE2(hp, am_call_count, c); hp += 3;
@@ -1243,7 +1134,7 @@ trace_info_func(Process* p, Eterm func_spec, Eterm key)
 	retval = CONS(hp, t, retval); hp += 2;
 	t = TUPLE2(hp, am_meta_match_spec, match_spec_meta); hp += 3;
 	retval = CONS(hp, t, retval); hp += 2;
-	t = TUPLE2(hp, am_meta, meta); hp += 3;
+	t = TUPLE2(hp, am_meta, m); hp += 3;
 	retval = CONS(hp, t, retval); hp += 2;
 	t = TUPLE2(hp, am_match_spec, match_spec); hp += 3;
 	retval = CONS(hp, t, retval); hp += 2;
@@ -1306,7 +1197,8 @@ trace_info_on_load(Process* p, Eterm key)
     case am_meta:
 	hp = HAlloc(p, 3);
 	if (erts_default_trace_pattern_flags.meta) {
-	    return TUPLE2(hp, key, erts_default_meta_tracer_pid);
+            ASSERT(!ERTS_TRACER_IS_NIL(erts_default_meta_tracer));
+	    return TUPLE2(hp, key, erts_tracer_to_term(p, erts_default_meta_tracer));
 	} else {
 	    return TUPLE2(hp, key, am_false);
 	}
@@ -1345,7 +1237,7 @@ trace_info_on_load(Process* p, Eterm key)
 	}
     case am_all:
 	{
-	    Eterm match_spec = am_false, meta_match_spec = am_false, r = NIL, t;
+	    Eterm match_spec = am_false, meta_match_spec = am_false, r = NIL, t, m;
 	    
 	    if (erts_default_trace_pattern_flags.local ||
 		(! erts_default_trace_pattern_flags.breakpoint)) {
@@ -1363,6 +1255,8 @@ trace_info_on_load(Process* p, Eterm key)
 		    MatchSetGetSource(erts_default_meta_match_spec);
 		meta_match_spec = copy_object(meta_match_spec, p);
 	    }
+            m = (erts_default_trace_pattern_flags.meta
+                 ? erts_tracer_to_term(p, erts_default_meta_tracer) : am_false);
 	    hp = HAlloc(p, (3+2)*5 + 3);
 	    t = TUPLE2(hp, am_call_count, 
 		       (erts_default_trace_pattern_flags.call_count
@@ -1370,9 +1264,7 @@ trace_info_on_load(Process* p, Eterm key)
 	    r = CONS(hp, t, r); hp += 2;
 	    t = TUPLE2(hp, am_meta_match_spec, meta_match_spec); hp += 3;
 	    r = CONS(hp, t, r); hp += 2;
-	    t = TUPLE2(hp, am_meta, 
-		       (erts_default_trace_pattern_flags.meta
-			? erts_default_meta_tracer_pid : am_false)); hp += 3;
+	    t = TUPLE2(hp, am_meta, m); hp += 3;
 	    r = CONS(hp, t, r); hp += 2;
 	    t = TUPLE2(hp, am_match_spec, match_spec); hp += 3;
 	    r = CONS(hp, t, r); hp += 2;
@@ -1397,7 +1289,7 @@ int
 erts_set_trace_pattern(Process*p, Eterm* mfa, int specified,
 		       Binary* match_prog_set, Binary *meta_match_prog_set,
 		       int on, struct trace_pattern_flags flags,
-		       Eterm meta_tracer_pid, int is_blocking)
+		       ErtsTracer meta_tracer, int is_blocking)
 {
     const ErtsCodeIndex code_ix = erts_active_code_ix();
     int matches = 0;
@@ -1487,7 +1379,7 @@ erts_set_trace_pattern(Process*p, Eterm* mfa, int specified,
 		    }
 		    if (flags.meta) {
 			erts_set_mtrace_bif(pc, meta_match_prog_set,
-					    meta_tracer_pid);
+					    meta_tracer);
 			m = 1;
 		    }
 		    if (flags.call_time) {
@@ -1527,7 +1419,7 @@ erts_set_trace_pattern(Process*p, Eterm* mfa, int specified,
 	    }
 	    if (flags.meta) {
 		erts_set_mtrace_break(&finish_bp.f, meta_match_prog_set,
-				      meta_tracer_pid);
+				      meta_tracer);
 	    }
 	    if (flags.call_count) {
 		erts_set_count_break(&finish_bp.f, on);
@@ -2336,50 +2228,86 @@ BIF_RETTYPE system_profile_2(BIF_ALIST_2)
 }
 /* End: Trace for System Profiling */
 
-BIF_RETTYPE
-trace_delivered_1(BIF_ALIST_1)
+/* Trace delivered send an aux work message to all schedulers
+   and when all schedulers have acknowledged that they have seen
+   the message the message is sent to the requesting process.
+
+   IMPORTANT: We have to make sure that the all messages sent
+   using enif_send have been delivered before we send the message
+   to the caller.
+
+   There used to be a separate implementation for when only a pid
+   is passed in, but since this is not performance critical code
+   we now use the same approach for both.
+*/
+
+typedef struct {
+    Process *proc;
+    Eterm ref;
+    Eterm ref_heap[REF_THING_SIZE];
+    Eterm target;
+    erts_smp_atomic32_t refc;
+} ErtsTraceDeliveredAll;
+
+static void
+reply_trace_delivered_all(void *vtdarp)
 {
-    DECL_AM(trace_delivered);
-#ifdef ERTS_SMP
-    ErlHeapFragment *bp;
-#else
-    ErtsProcLocks locks = 0;
-#endif
-    Eterm *hp;
-    Eterm msg, ref, msg_ref;
-    Process *p;
-    if (BIF_ARG_1 == am_all) {
-	p = NULL;
-    } else if (! (p = erts_pid2proc(BIF_P, ERTS_PROC_LOCK_MAIN,
-				    BIF_ARG_1, ERTS_PROC_LOCKS_ALL))) {
-	if (is_not_internal_pid(BIF_ARG_1)) {
-	    BIF_ERROR(BIF_P, BADARG);
-	}
-    }
-    
-    ref = erts_make_ref(BIF_P);
+    ErtsTraceDeliveredAll *tdarp = (ErtsTraceDeliveredAll *) vtdarp;
 
+    if (erts_smp_atomic32_dec_read_nob(&tdarp->refc) == 0) {
+        Eterm ref_copy, msg;
+        Process *rp = tdarp->proc;
+        Eterm *hp = NULL;
+        ErlOffHeap *ohp;
 #ifdef ERTS_SMP
-    bp = new_message_buffer(REF_THING_SIZE + 4);
-    hp = &bp->mem[0];
-    msg_ref = STORE_NC(&hp, &bp->off_heap, ref);
+        ErlHeapFragment *bp;
+        bp = new_message_buffer(4 + NC_HEAP_SIZE(tdarp->ref));
+        hp = &bp->mem[0];
+        ohp = &bp->off_heap;
 #else
-    hp = HAlloc(BIF_P, 4);
-    msg_ref = ref;
+        ErtsProcLocks rp_locks = 0;
+        ErtsMessage *mp;
+        mp = erts_alloc_message_heap(
+            rp, &rp_locks, 4 + NC_HEAP_SIZE(tdarp->ref), &hp, &ohp);
 #endif
 
-    msg = TUPLE3(hp, AM_trace_delivered, BIF_ARG_1, msg_ref);
+        ref_copy = STORE_NC(&hp, ohp, tdarp->ref);
+        msg = TUPLE3(hp, am_trace_delivered, tdarp->target, ref_copy);
 
 #ifdef ERTS_SMP
-    erts_send_sys_msg_proc(BIF_P->common.id, BIF_P->common.id, msg, bp);
-    if (p)
-	erts_smp_proc_unlock(p,
-			     (BIF_P == p
-			      ? ERTS_PROC_LOCKS_ALL_MINOR
-			      : ERTS_PROC_LOCKS_ALL));
+        erts_send_sys_msg_proc(rp->common.id, rp->common.id, msg, bp);
 #else
-    erts_send_message(BIF_P, BIF_P, &locks, msg, ERTS_SND_FLG_NO_SEQ_TRACE);
+        erts_queue_message(rp, &rp_locks, mp, msg);
 #endif
 
-    BIF_RET(ref);
+	erts_free(ERTS_ALC_T_MISC_AUX_WORK, vtdarp);
+        erts_proc_dec_refc(rp);
+    }
+}
+
+BIF_RETTYPE
+trace_delivered_1(BIF_ALIST_1)
+{
+
+    if (BIF_ARG_1 == am_all || is_internal_pid(BIF_ARG_1)) {
+        Eterm *hp, ref;
+        ErtsTraceDeliveredAll *tdarp =
+            erts_alloc(ERTS_ALC_T_MISC_AUX_WORK, sizeof(ErtsTraceDeliveredAll));
+
+        tdarp->proc = BIF_P;
+        ref = erts_make_ref(BIF_P);
+        hp = &tdarp->ref_heap[0];
+        tdarp->ref = STORE_NC(&hp, NULL, ref);
+        tdarp->target = BIF_ARG_1;
+        erts_smp_atomic32_init_nob(&tdarp->refc,
+                                   (erts_aint32_t) erts_no_schedulers);
+        erts_proc_add_refc(BIF_P, 1);
+        erts_schedule_multi_misc_aux_work(0,
+                                          erts_no_schedulers,
+                                          reply_trace_delivered_all,
+                                          (void *) tdarp);
+        BIF_RET(ref);
+    } else {
+        BIF_ERROR(BIF_P, BADARG);
+    }
 }
diff --git a/erts/emulator/beam/erl_db_util.c b/erts/emulator/beam/erl_db_util.c
index be14386b14..76b96637ae 100644
--- a/erts/emulator/beam/erl_db_util.c
+++ b/erts/emulator/beam/erl_db_util.c
@@ -135,21 +135,22 @@ get_proc(Process *cp, Uint32 cp_locks, Eterm id, Uint32 id_locks)
 
 
 static Eterm
-set_tracee_flags(Process *tracee_p, Eterm tracer, Uint d_flags, Uint e_flags) {
+set_tracee_flags(Process *tracee_p, ErtsTracer tracer,
+                 Uint d_flags, Uint e_flags) {
     Eterm ret;
     Uint flags;
 
-    if (tracer == NIL) {
+    if (ERTS_TRACER_IS_NIL(tracer)) {
 	flags = ERTS_TRACE_FLAGS(tracee_p) & ~TRACEE_FLAGS;
     }  else {
 	flags = ((ERTS_TRACE_FLAGS(tracee_p) & ~d_flags) | e_flags);
-	if (! flags) tracer = NIL;
+	if (! flags) tracer = erts_tracer_nil;
     }
-    ret = ((ERTS_TRACER_PROC(tracee_p) != tracer
+    ret = ((!ERTS_TRACER_COMPARE(ERTS_TRACER(tracee_p),tracer)
 	    || ERTS_TRACE_FLAGS(tracee_p) != flags)
 	   ? am_true
 	   : am_false);
-    ERTS_TRACER_PROC(tracee_p) = tracer;
+    erts_tracer_replace(&tracee_p->common, tracer);
     ERTS_TRACE_FLAGS(tracee_p) = flags;
     return ret;
 }
@@ -163,40 +164,16 @@ set_tracee_flags(Process *tracee_p, Eterm tracer, Uint d_flags, Uint e_flags) {
 ** returns fail_term on failure. Fails if tracer pid or port is invalid.
 */
 static Eterm 
-set_match_trace(Process *tracee_p, Eterm fail_term, Eterm tracer,
+set_match_trace(Process *tracee_p, Eterm fail_term, ErtsTracer tracer,
 		Uint d_flags, Uint e_flags) {
-    Eterm ret = fail_term;
-    Process *tracer_p;
-    
-    ERTS_SMP_LC_ASSERT(ERTS_PROC_LOCKS_ALL == 
-		       erts_proc_lc_my_proc_locks(tracee_p));
-
-    if (is_internal_pid(tracer)
-	&& (tracer_p = 
-	    erts_pid2proc(tracee_p, ERTS_PROC_LOCKS_ALL,
-			  tracer, ERTS_PROC_LOCKS_ALL))) {
-	if (tracee_p != tracer_p) {
-	    ret = set_tracee_flags(tracee_p, tracer, d_flags, e_flags);
-	    ERTS_TRACE_FLAGS(tracer_p) |= (ERTS_TRACE_FLAGS(tracee_p)
-					   ? F_TRACER
-					   : 0);
-	    erts_smp_proc_unlock(tracer_p, ERTS_PROC_LOCKS_ALL);
-	}
-    } else if (is_internal_port(tracer)) {
-	Port *tracer_port = 
-	    erts_id2port_sflgs(tracer,
-			       tracee_p,
-			       ERTS_PROC_LOCKS_ALL,
-			       ERTS_PORT_SFLGS_INVALID_TRACER_LOOKUP);
-	if (tracer_port) {
-	    ret = set_tracee_flags(tracee_p, tracer, d_flags, e_flags);
-	    erts_port_release(tracer_port);
-	}
-    } else {
-	ASSERT(is_nil(tracer));
-	ret = set_tracee_flags(tracee_p, tracer, d_flags, e_flags);
-    }
-    return ret;
+
+    ERTS_SMP_LC_ASSERT(
+        ERTS_PROC_LOCKS_ALL == erts_proc_lc_my_proc_locks(tracee_p)
+        || erts_thr_progress_is_blocking());
+
+    if (ERTS_TRACER_IS_NIL(tracer) || erts_is_tracer_enabled(tracee_p, tracer))
+        return set_tracee_flags(tracee_p, tracer, d_flags, e_flags);
+    return fail_term;
 }
 
 /*
@@ -2358,7 +2335,7 @@ restart:
 	case matchEnableTrace:
 	    if ( (n = erts_trace_flag2bit(esp[-1]))) {
 		BEGIN_ATOMIC_TRACE(c_p);
-		set_tracee_flags(c_p, ERTS_TRACER_PROC(c_p), 0, n);
+		set_tracee_flags(c_p, ERTS_TRACER(c_p), 0, n);
 		esp[-1] = am_true;
 	    } else {
 		esp[-1] = FAIL_TERM;
@@ -2371,7 +2348,7 @@ restart:
 		BEGIN_ATOMIC_TRACE(c_p);
 		if ( (tmpp = get_proc(c_p, 0, esp[0], 0))) {
 		    /* Always take over the tracer of the current process */
-		    set_tracee_flags(tmpp, ERTS_TRACER_PROC(c_p), 0, n);
+		    set_tracee_flags(tmpp, ERTS_TRACER(c_p), 0, n);
 		    esp[-1] = am_true;
 		}
 	    }
@@ -2379,7 +2356,7 @@ restart:
 	case matchDisableTrace:
 	    if ( (n = erts_trace_flag2bit(esp[-1]))) {
 		BEGIN_ATOMIC_TRACE(c_p);
-		set_tracee_flags(c_p, ERTS_TRACER_PROC(c_p), n, 0);
+		set_tracee_flags(c_p, ERTS_TRACER(c_p), n, 0);
 		esp[-1] = am_true;
 	    } else {
 		esp[-1] = FAIL_TERM;
@@ -2392,7 +2369,7 @@ restart:
 		BEGIN_ATOMIC_TRACE(c_p);
 		if ( (tmpp = get_proc(c_p, 0, esp[0], 0))) {
 		    /* Always take over the tracer of the current process */
-		    set_tracee_flags(tmpp, ERTS_TRACER_PROC(c_p), n, 0);
+		    set_tracee_flags(tmpp, ERTS_TRACER(c_p), n, 0);
 		    esp[-1] = am_true;
 		}
 	    }
@@ -2428,7 +2405,7 @@ restart:
 	    {
 		/*    disable         enable                                */
 		Uint  d_flags  = 0,   e_flags  = 0;  /* process trace flags */
-		Eterm tracer = ERTS_TRACER_PROC(c_p);
+		ErtsTracer tracer = erts_tracer_nil;
 		/* XXX Atomicity note: Not fully atomic. Default tracer
 		 * is sampled from current process but applied to
 		 * tracee and tracer later after releasing main
@@ -2440,29 +2417,34 @@ restart:
 		 * {trace,[],[{{tracer,Tracer}}]} is much, much older.
 		 */
 		int   cputs = 0;
+                erts_tracer_update(&tracer, ERTS_TRACER(c_p));
 		
 		if (! erts_trace_flags(esp[-1], &d_flags, &tracer, &cputs) ||
 		    ! erts_trace_flags(esp[-2], &e_flags, &tracer, &cputs) ||
 		    cputs ) {
 		    (--esp)[-1] = FAIL_TERM;
+                    ERTS_TRACER_CLEAR(&tracer);
 		    break;
 		}
 		erts_smp_proc_lock(c_p, ERTS_PROC_LOCKS_ALL_MINOR);
 		(--esp)[-1] = set_match_trace(c_p, FAIL_TERM, tracer,
 					      d_flags, e_flags);
 		erts_smp_proc_unlock(c_p, ERTS_PROC_LOCKS_ALL_MINOR);
+                ERTS_TRACER_CLEAR(&tracer);
 	    }
 	    break;
 	case matchTrace3:
 	    {
 		/*    disable         enable                                */
 		Uint  d_flags  = 0,   e_flags  = 0;  /* process trace flags */
-		Eterm tracer = ERTS_TRACER_PROC(c_p);
+		ErtsTracer tracer = erts_tracer_nil;
 		/* XXX Atomicity note. Not fully atomic. See above. 
 		 * Above it could possibly be solved, but not here.
 		 */
 		int   cputs = 0;
 		Eterm tracee = (--esp)[0];
+
+                erts_tracer_update(&tracer, ERTS_TRACER(c_p));
 		
 		if (! erts_trace_flags(esp[-1], &d_flags, &tracer, &cputs) ||
 		    ! erts_trace_flags(esp[-2], &e_flags, &tracer, &cputs) ||
@@ -2470,6 +2452,7 @@ restart:
 		    ! (tmpp = get_proc(c_p, ERTS_PROC_LOCK_MAIN, 
 				       tracee, ERTS_PROC_LOCKS_ALL))) {
 		    (--esp)[-1] = FAIL_TERM;
+                    ERTS_TRACER_CLEAR(&tracer);
 		    break;
 		}
 		if (tmpp == c_p) {
@@ -2483,6 +2466,7 @@ restart:
 		    erts_smp_proc_unlock(tmpp, ERTS_PROC_LOCKS_ALL);
 		    erts_smp_proc_lock(c_p, ERTS_PROC_LOCK_MAIN);
 		}
+                ERTS_TRACER_CLEAR(&tracer);
 	    }
 	    break;
 	case matchCatch:  /* Match success, now build result */
@@ -5084,7 +5068,7 @@ static Eterm match_spec_test(Process *p, Eterm against, Eterm spec, int trace)
 	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);
     }
-    
+
     if (mps == NULL) {
 	hp = HAlloc(p,3);
 	ret = TUPLE2(hp, am_error, lint_res);
diff --git a/erts/emulator/beam/erl_gc.c b/erts/emulator/beam/erl_gc.c
index 62417fa05c..f33ade27f3 100644
--- a/erts/emulator/beam/erl_gc.c
+++ b/erts/emulator/beam/erl_gc.c
@@ -2233,9 +2233,7 @@ setup_rootset(Process *p, Eterm *objv, int nobj, Rootset *rootset)
 	n++;
     }
 #endif
-    ASSERT(is_nil(ERTS_TRACER_PROC(p)) ||
-	   is_internal_pid(ERTS_TRACER_PROC(p)) ||
-	   is_internal_port(ERTS_TRACER_PROC(p)));
+    ASSERT(IS_TRACER_VALID(ERTS_TRACER(p)));
 
     ASSERT(is_pid(follow_moved(p->group_leader, (Eterm) 0)));
     if (is_not_immed(p->group_leader)) {
@@ -2905,7 +2903,7 @@ reply_gc_info(void *vgcirp)
 	hpp = &hp;
     }
 
-    erts_queue_message(rp, &rp_locks, mp, msg, NIL);
+    erts_queue_message(rp, &rp_locks, mp, msg);
 
     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 1a621863bb..8e201d5711 100644
--- a/erts/emulator/beam/erl_hl_timer.c
+++ b/erts/emulator/beam/erl_hl_timer.c
@@ -1248,7 +1248,7 @@ hlt_bif_timer_timeout(ErtsHLTimer *tmr, Uint32 roflgs)
 	    ErtsMessage *mp = erts_alloc_message(0, NULL);
 	    mp->data.heap_frag = tmr->btm.bp;
 	    erts_queue_message(proc, &proc_locks, mp,
-			       tmr->btm.message, NIL);
+			       tmr->btm.message);
 	    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, NIL);
+	erts_queue_message(proc, &proc_locks, mp, msg);
 
 	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, NIL);
+	erts_queue_message(c_p, &proc_locks, mp, msg);
 
 	proc_locks &= ~ERTS_PROC_LOCK_MAIN;
 	if (proc_locks)
diff --git a/erts/emulator/beam/erl_lock_check.c b/erts/emulator/beam/erl_lock_check.c
index 9231fb1b34..39c0617143 100644
--- a/erts/emulator/beam/erl_lock_check.c
+++ b/erts/emulator/beam/erl_lock_check.c
@@ -97,6 +97,7 @@ static erts_lc_lock_order_t erts_lock_order[] = {
     {	"dist_entry_links",			"address"		},
     {   "code_write_permission",                NULL                    },
     {	"proc_status",				"pid"			},
+    {	"proc_trace",				"pid"			},
     {   "ports_snapshot",                       NULL                    },
     {	"meta_name_tab",	         	"address"		},
     {	"meta_main_tab_slot",			"address"		},
@@ -148,6 +149,7 @@ static erts_lc_lock_order_t erts_lock_order[] = {
     {	"dist_entry_out_queue",			"address"		},
     {	"port_sched_lock",			"port_id"		},
     {	"sys_msg_q", 				NULL			},
+    {	"tracer_mtx", 				NULL			},
     {   "port_table",                           NULL                    },
 #endif
     {	"mtrace_op",				NULL			},
@@ -160,9 +162,6 @@ static erts_lc_lock_order_t erts_lock_order[] = {
     {	"proclist_pre_alloc_lock",		"address"		},
     {	"xports_list_pre_alloc_lock",		"address"		},
     {	"inet_buffer_stack_lock",		NULL			},
-    {	"gc_info",				NULL			},
-    {	"io_wake",				NULL			},
-    {	"timer_wheel",				NULL			},
     {	"system_block",				NULL			},
     {	"timeofday",				NULL			},
     {	"get_time",				NULL			},
diff --git a/erts/emulator/beam/erl_map.h b/erts/emulator/beam/erl_map.h
index a40070d00d..7af9100906 100644
--- a/erts/emulator/beam/erl_map.h
+++ b/erts/emulator/beam/erl_map.h
@@ -66,7 +66,7 @@ typedef struct flatmap_s {
 
 
 /* erl_term.h stuff */
-#define flatmap_get_values(x)        (((Eterm *)(x)) + 3)
+#define flatmap_get_values(x)        (((Eterm *)(x)) + sizeof(flatmap_t)/sizeof(Eterm))
 #define flatmap_get_keys(x)          (((Eterm *)tuple_val(((flatmap_t *)(x))->keys)) + 1)
 #define flatmap_get_size(x)          (((flatmap_t*)(x))->size)
 
diff --git a/erts/emulator/beam/erl_message.c b/erts/emulator/beam/erl_message.c
index 7596747b91..9beff52835 100644
--- a/erts/emulator/beam/erl_message.c
+++ b/erts/emulator/beam/erl_message.c
@@ -324,7 +324,7 @@ erts_queue_dist_message(Process *rcvr,
 			tok_label, tok_lastcnt, tok_serial);
 	    }
 #endif
-	    erts_queue_message(rcvr, rcvr_locks, mp, msg, token);
+	    erts_queue_message(rcvr, rcvr_locks, mp, msg);
 	}
     }
     else {
@@ -349,7 +349,7 @@ erts_queue_dist_message(Process *rcvr,
         }
 #endif
 
-	LINK_MESSAGE(rcvr, mp);
+	LINK_MESSAGE(rcvr, mp, &mp->next, 1);
 
 	if (!(*rcvr_locks & ERTS_PROC_LOCK_MSGQ))
 	    erts_smp_proc_unlock(rcvr, ERTS_PROC_LOCK_MSGQ);
@@ -364,31 +364,34 @@ erts_queue_dist_message(Process *rcvr,
     }
 }
 
-/* Add a message last in message queue */
+/* Add messages last in message queue */
 static Sint
-queue_message(Process *c_p,
-	      Process* receiver,
-	      erts_aint32_t *receiver_state,
-	      ErtsProcLocks *receiver_locks,
-	      ErtsMessage* mp,
-	      Eterm message,
-	      Eterm seq_trace_token
-#ifdef USE_VM_PROBES
-		   , Eterm dt_utag
-#endif
-    )
+queue_messages(Process *c_p,
+               Process* receiver,
+               erts_aint32_t *receiver_state,
+               ErtsProcLocks *receiver_locks,
+               ErtsMessage* first,
+               ErtsMessage** last,
+               Uint len)
 {
     Sint res;
     int locked_msgq = 0;
     erts_aint32_t state;
 
-    ERTS_SMP_LC_ASSERT(*receiver_locks == erts_proc_lc_my_proc_locks(receiver));
+    ASSERT(is_value(ERL_MESSAGE_TERM(first)));
+    ASSERT(ERL_MESSAGE_TOKEN(first) == am_undefined ||
+           ERL_MESSAGE_TOKEN(first) == NIL ||
+           is_tuple(ERL_MESSAGE_TOKEN(first)));
 
 #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));
+#endif
 
     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 = ERTS_PROC_LOCK_MSGQ;
 
 	    if (receiver_state)
 		state = *receiver_state;
@@ -417,16 +420,10 @@ queue_message(Process *c_p,
 	/* Drop message if receiver is exiting or has a pending exit... */
 	if (locked_msgq)
 	    erts_smp_proc_unlock(receiver, ERTS_PROC_LOCK_MSGQ);
-	erts_cleanup_messages(mp);
+	erts_cleanup_messages(first);
 	return 0;
     }
 
-    ERL_MESSAGE_TERM(mp) = message;
-    ERL_MESSAGE_TOKEN(mp) = seq_trace_token;
-#ifdef USE_VM_PROBES
-    ERL_MESSAGE_DT_UTAG(mp) = dt_utag;
-#endif
-
     res = receiver->msg.len;
 #ifdef ERTS_SMP
     if (*receiver_locks & ERTS_PROC_LOCK_MAIN) {
@@ -440,75 +437,83 @@ queue_message(Process *c_p,
 	 */
 	res += receiver->msg_inq.len;
 	ERTS_SMP_MSGQ_MV_INQ2PRIVQ(receiver);
-	LINK_MESSAGE_PRIVQ(receiver, mp);
+        LINK_MESSAGE_PRIVQ(receiver, first, last, len);
     }
     else
 #endif
     {
-	LINK_MESSAGE(receiver, mp);
+	LINK_MESSAGE(receiver, first, last, len);
     }
 
+    if (IS_TRACED_FL(receiver, F_TRACE_RECEIVE)) {
+        ErtsMessage *msg = first;
+
 #ifdef USE_VM_PROBES
-    if (DTRACE_ENABLED(message_queued)) {
-        DTRACE_CHARBUF(receiver_name, DTRACE_TERM_BUF_SIZE);
-        Sint tok_label = 0;
-        Sint tok_lastcnt = 0;
-        Sint tok_serial = 0;
-
-        dtrace_proc_str(receiver, receiver_name);
-        if (seq_trace_token != NIL && is_tuple(seq_trace_token)) {
-            tok_label = signed_val(SEQ_TRACE_T_LABEL(seq_trace_token));
-            tok_lastcnt = signed_val(SEQ_TRACE_T_LASTCNT(seq_trace_token));
-            tok_serial = signed_val(SEQ_TRACE_T_SERIAL(seq_trace_token));
+        if (DTRACE_ENABLED(message_queued)) {
+            DTRACE_CHARBUF(receiver_name, DTRACE_TERM_BUF_SIZE);
+            Sint tok_label = 0;
+            Sint tok_lastcnt = 0;
+            Sint tok_serial = 0;
+            Eterm seq_trace_token = ERL_MESSAGE_TOKEN(msg);
+
+            dtrace_proc_str(receiver, receiver_name);
+            if (seq_trace_token != NIL && is_tuple(seq_trace_token)) {
+                tok_label = signed_val(SEQ_TRACE_T_LABEL(seq_trace_token));
+                tok_lastcnt = signed_val(SEQ_TRACE_T_LASTCNT(seq_trace_token));
+                tok_serial = signed_val(SEQ_TRACE_T_SERIAL(seq_trace_token));
+            }
+            DTRACE6(message_queued,
+                    receiver_name, size_object(ERL_MESSAGE_TERM(msg)),
+                    receiver->msg.len,
+                    tok_label, tok_lastcnt, tok_serial);
         }
-        DTRACE6(message_queued,
-                receiver_name, size_object(message), receiver->msg.len,
-                tok_label, tok_lastcnt, tok_serial);
-    }
 #endif
 
-    if (IS_TRACED_FL(receiver, F_TRACE_RECEIVE))
-	trace_receive(receiver, message);
+        while (msg) {
+            trace_receive(receiver, ERL_MESSAGE_TERM(msg));
+            msg = msg->next;
+        }
+
+    }
 
     if (locked_msgq)
 	erts_smp_proc_unlock(receiver, ERTS_PROC_LOCK_MSGQ);
 
-    erts_proc_notify_new_message(receiver,
 #ifdef ERTS_SMP
-				 *receiver_locks
+    erts_proc_notify_new_message(receiver, *receiver_locks);
 #else
-				 0
-#endif
-	);
-
-#ifndef ERTS_SMP
+    erts_proc_notify_new_message(receiver, 0);
     ERTS_HOLE_CHECK(receiver);
 #endif
     return res;
 }
 
-void
-#ifdef USE_VM_PROBES
-erts_queue_message_probe(Process* receiver, ErtsProcLocks *receiver_locks,
-                         ErtsMessage* mp,
-                         Eterm message, Eterm seq_trace_token, Eterm dt_utag)
-#else
+static Sint
+queue_message(Process *c_p,
+              Process* receiver,
+              erts_aint32_t *receiver_state,
+              ErtsProcLocks *receiver_locks,
+              ErtsMessage* mp, Eterm msg)
+{
+    ERL_MESSAGE_TERM(mp) = msg;
+    return queue_messages(c_p, receiver, receiver_state, receiver_locks,
+                          mp, &mp->next, 1 );
+}
+
+Sint
 erts_queue_message(Process* receiver, ErtsProcLocks *receiver_locks,
-                   ErtsMessage* mp,
-                   Eterm message, Eterm seq_trace_token)
-#endif
+                   ErtsMessage* mp, Eterm msg)
 {
-    queue_message(NULL,
-		  receiver,
-		  NULL,
-		  receiver_locks,
-		  mp,
-		  message,
-		  seq_trace_token
-#ifdef USE_VM_PROBES
-		  , dt_utag
-#endif
-	);
+    return queue_message(NULL, receiver, NULL, receiver_locks, mp, msg);
+}
+
+
+Sint
+erts_queue_messages(Process* receiver, ErtsProcLocks *receiver_locks,
+                    ErtsMessage* first, ErtsMessage** last, Uint len)
+{
+    return queue_messages(NULL, receiver, NULL, receiver_locks,
+                          first, last, len);
 }
 
 void
@@ -591,7 +596,7 @@ erts_try_alloc_message_on_heap(Process *pp,
 #if defined(ERTS_SMP)
 	*plp & ERTS_PROC_LOCK_MAIN
 #else
-	1
+	pp
 #endif
 	) {
 #ifdef ERTS_SMP
@@ -621,7 +626,7 @@ erts_try_alloc_message_on_heap(Process *pp,
 	*on_heap_p = !0;
     }
 #ifdef ERTS_SMP
-    else if (erts_smp_proc_trylock(pp, ERTS_PROC_LOCK_MAIN) == 0) {
+    else if (pp && erts_smp_proc_trylock(pp, ERTS_PROC_LOCK_MAIN) == 0) {
 	locked_main = 1;
 	*psp = erts_smp_atomic32_read_nob(&pp->state);
 	*plp |= ERTS_PROC_LOCK_MAIN;
@@ -821,17 +826,15 @@ erts_send_message(Process* sender,
 #endif
     }
 
+    ERL_MESSAGE_TOKEN(mp) = token;
+#ifdef USE_VM_PROBES
+    ERL_MESSAGE_DT_UTAG(mp) = utag;
+#endif
     res = queue_message(sender,
 			receiver,
 			&receiver_state,
 			receiver_locks,
-			mp,
-			message,
-			token
-#ifdef USE_VM_PROBES
-			, utag
-#endif
-	);
+			mp, message);
 
     BM_SWAP_TIMER(send,system);
     
@@ -887,7 +890,8 @@ erts_deliver_exit_message(Eterm from, Process *to, ErtsProcLocks *to_locksp,
 	/* the trace token must in this case be updated by the caller */
 	seq_trace_output(token, save, SEQ_TRACE_SEND, to->common.id, NULL);
 	temptoken = copy_struct(token, sz_token, &hp, ohp);
-	erts_queue_message(to, to_locksp, mp, save, temptoken);
+        ERL_MESSAGE_TOKEN(mp) = temptoken;
+	erts_queue_message(to, to_locksp, mp, save);
     } else {
 	sz_from = IS_CONST(from) ? 0 : size_object(from);
 #ifdef SHCOPY_SEND
@@ -909,7 +913,7 @@ 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, NIL);
+	erts_queue_message(to, to_locksp, mp, save);
     }
 }
 
@@ -1484,7 +1488,7 @@ erts_factory_message_create(ErtsHeapFactory* factory,
     int on_heap;
     erts_aint32_t state;
 
-    state = erts_smp_atomic32_read_nob(&proc->state);
+    state = proc ? erts_smp_atomic32_read_nob(&proc->state) : 0;
 
     if (state & ERTS_PSFLG_OFF_HEAP_MSGQ) {
 	msgp = erts_alloc_message(sz, &hp);
diff --git a/erts/emulator/beam/erl_message.h b/erts/emulator/beam/erl_message.h
index ad4e65274c..608cf552a2 100644
--- a/erts/emulator/beam/erl_message.h
+++ b/erts/emulator/beam/erl_message.h
@@ -182,48 +182,64 @@ typedef struct {
     Sint len;            /* queue length */
 } ErlMessageInQueue;
 
+typedef struct erl_trace_message_queue__ {
+    struct erl_trace_message_queue__ *next; /* point to the next receiver */
+    Eterm receiver;
+    ErtsMessage* first;
+    ErtsMessage** last;  /* point to the last next pointer */
+    Sint len;            /* queue length */
+} ErlTraceMessageQueue;
+
 #endif
 
 /* Get "current" message */
 #define PEEK_MESSAGE(p)  (*(p)->msg.save)
 
+#ifdef USE_VM_PROBES
+#define LINK_MESSAGE_DTAG(mp, dt) ERL_MESSAGE_DT_UTAG(mp) = dt
+#else
+#define LINK_MESSAGE_DTAG(mp, dt)
+#endif
 
-/* Add message last in private message queue */
-#define LINK_MESSAGE_PRIVQ(p, mp) do { \
-    *(p)->msg.last = (mp); \
-    (p)->msg.last = &(mp)->next; \
-    (p)->msg.len++; \
-} while (0)
-
+#define LINK_MESSAGE_IMPL(p, first_msg, last_msg, num_msgs, where) do { \
+        *(p)->where.last = (first_msg);                                 \
+        (p)->where.last = (last_msg);                                   \
+        (p)->where.len += (num_msgs);                                   \
+    } while(0)
 
 #ifdef ERTS_SMP
 
-/* Move in message queue to end of private message queue */
-#define ERTS_SMP_MSGQ_MV_INQ2PRIVQ(P)					\
-do {									\
-    if ((P)->msg_inq.first) {						\
-	*(P)->msg.last = (P)->msg_inq.first;				\
-	(P)->msg.last = (P)->msg_inq.last;				\
-	(P)->msg.len += (P)->msg_inq.len;				\
-	(P)->msg_inq.first = NULL;					\
-	(P)->msg_inq.last = &(P)->msg_inq.first;			\
-	(P)->msg_inq.len = 0;						\
-    }									\
-} while (0)
-
-/* Add message last in message queue */
-#define LINK_MESSAGE(p, mp) do { \
-    *(p)->msg_inq.last = (mp); \
-    (p)->msg_inq.last = &(mp)->next; \
-    (p)->msg_inq.len++; \
-} while(0)
+/* Add message last in private message queue */
+#define LINK_MESSAGE_PRIVQ(p, first_msg, last_msg, len)                 \
+    do {                                                                \
+        LINK_MESSAGE_IMPL(p, first_msg, last_msg, len, msg);            \
+    } while (0)
+
+/* Add message last_msg in message queue */
+#define LINK_MESSAGE(p, first_msg, last_msg, len)       \
+    LINK_MESSAGE_IMPL(p, first_msg, last_msg, len, msg_inq)
+
+#define ERTS_SMP_MSGQ_MV_INQ2PRIVQ(p)                   \
+    do {                                                \
+        if (p->msg_inq.first) {                         \
+            *p->msg.last = p->msg_inq.first;            \
+            p->msg.last = p->msg_inq.last;              \
+            p->msg.len += p->msg_inq.len;               \
+            p->msg_inq.first = NULL;                    \
+            p->msg_inq.last = &p->msg_inq.first;        \
+            p->msg_inq.len = 0;                         \
+        }                                               \
+    } while (0)
 
 #else
 
-#define ERTS_SMP_MSGQ_MV_INQ2PRIVQ(P)
+#define ERTS_SMP_MSGQ_MV_INQ2PRIVQ(p)
 
-/* Add message last in message queue */
-#define LINK_MESSAGE(p, mp) LINK_MESSAGE_PRIVQ((p), (mp))
+/* Add message last_msg in message queue */
+#define LINK_MESSAGE(p, first_msg, last_msg, len)                       \
+    do {                                                                \
+        LINK_MESSAGE_IMPL(p, first_msg, last_msg, len, msg);            \
+    } while(0)
 
 #endif
 
@@ -259,23 +275,30 @@ do {									\
 	(HEAP_FRAG_P)->off_heap.overhead = 0;				\
     } while (0)
 
+#ifdef USE_VM_PROBES
+#define ERL_MESSAGE_DT_UTAG_INIT(MP) ERL_MESSAGE_DT_UTAG(MP) = NIL
+#else
+#define ERL_MESSAGE_DT_UTAG_INIT(MP) do{ } while (0)
+#endif
+
+#define ERTS_INIT_MESSAGE(MP)                           \
+    do {                                                \
+        (MP)->next = NULL;                              \
+        ERL_MESSAGE_TERM(MP) = THE_NON_VALUE;           \
+        ERL_MESSAGE_TOKEN(MP) = NIL;                    \
+        ERL_MESSAGE_DT_UTAG_INIT(MP);                   \
+        MP->data.attached = NULL;                       \
+    } while (0)
+
 void init_message(void);
 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);
-#ifdef USE_VM_PROBES
-void erts_queue_message_probe(Process*, ErtsProcLocks*, ErtsMessage*,
-                              Eterm message, Eterm seq_trace_token, Eterm dt_utag);
-#define erts_queue_message(RP,RL,BP,Msg,SEQ) \
-    erts_queue_message_probe((RP),(RL),(BP),(Msg),(SEQ),NIL)
-#else
-void erts_queue_message(Process*, ErtsProcLocks*, ErtsMessage*,
-                        Eterm message, Eterm seq_trace_token);
-#define erts_queue_message_probe(RP,RL,BP,Msg,SEQ,TAG) \
-    erts_queue_message((RP),(RL),(BP),(Msg),(SEQ))
-#endif
+Sint erts_queue_message(Process*, ErtsProcLocks*,ErtsMessage*, Eterm);
+Sint erts_queue_messages(Process*, ErtsProcLocks*,
+                         ErtsMessage*, ErtsMessage**, Uint);
 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);
@@ -354,9 +377,7 @@ ERTS_GLB_FORCE_INLINE ErtsMessage *erts_alloc_message(Uint sz, Eterm **hpp)
 
     if (sz == 0) {
 	mp = erts_alloc_message_ref();
-	mp->next = NULL;
-	ERL_MESSAGE_TERM(mp) = NIL;
-	mp->data.attached = NULL;
+        ERTS_INIT_MESSAGE(mp);
 	if (hpp)
 	    *hpp = NULL;
 	return mp;
@@ -365,8 +386,7 @@ ERTS_GLB_FORCE_INLINE ErtsMessage *erts_alloc_message(Uint sz, Eterm **hpp)
     mp = erts_alloc(ERTS_ALC_T_MSG,
 		    sizeof(ErtsMessage) + (sz - 1)*sizeof(Eterm));
 
-    mp->next = NULL;
-    ERL_MESSAGE_TERM(mp) = NIL;
+    ERTS_INIT_MESSAGE(mp);
     mp->data.attached = ERTS_MSG_COMBINED_HFRAG;
     ERTS_INIT_HEAP_FRAG(&mp->hfrag, sz, sz);
 
diff --git a/erts/emulator/beam/erl_msacc.c b/erts/emulator/beam/erl_msacc.c
index 71e3fd8b6e..d0f305900a 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, NIL);
+    erts_queue_message(rp, &rp_locks, msgp, msg);
 
     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 1f5793f534..73c0eb8eba 100644
--- a/erts/emulator/beam/erl_nif.c
+++ b/erts/emulator/beam/erl_nif.c
@@ -92,11 +92,12 @@ static ERTS_INLINE Eterm* alloc_heap(ErlNifEnv* env, unsigned need)
 }
 
 static Eterm* alloc_heap_heavy(ErlNifEnv* env, unsigned need, Eterm* hp)
-{    
+{
     env->hp = hp;
-    if (env->heap_frag == NULL) {       
+    if (env->heap_frag == NULL) {
 	ASSERT(HEAP_LIMIT(env->proc) == env->hp_end);
-	HEAP_TOP(env->proc) = env->hp;	
+        ASSERT(env->hp + need > env->hp_end);
+	HEAP_TOP(env->proc) = env->hp;
     }
     else {
 	env->heap_frag->used_size = hp - env->heap_frag->mem;
@@ -120,7 +121,8 @@ static ERTS_INLINE void ensure_heap(ErlNifEnv* env, unsigned may_need)
 }
 #endif
 
-void erts_pre_nif(ErlNifEnv* env, Process* p, struct erl_module_nif* mod_nif)
+void erts_pre_nif(ErlNifEnv* env, Process* p, struct erl_module_nif* mod_nif,
+                  Process* tracee)
 {
     env->mod_nif = mod_nif;
     env->proc = p;
@@ -130,17 +132,7 @@ void erts_pre_nif(ErlNifEnv* env, Process* p, struct erl_module_nif* mod_nif)
     env->fpe_was_unmasked = erts_block_fpe();
     env->tmp_obj_list = NULL;
     env->exception_thrown = 0;
-}
-
-static void pre_nif_noproc(ErlNifEnv* env, struct erl_module_nif* mod_nif)
-{
-    env->mod_nif = mod_nif;
-    env->proc = NULL;
-    env->hp = NULL;
-    env->hp_end = NULL;
-    env->heap_frag = NULL;
-    env->fpe_was_unmasked = erts_block_fpe();
-    env->tmp_obj_list = NULL;
+    env->tracee = tracee;
 }
 
 /* Temporary object header, auto-deallocated when NIF returns
@@ -180,13 +172,6 @@ void erts_post_nif(ErlNifEnv* env)
     free_tmp_objs(env);
 }
 
-static void post_nif_noproc(ErlNifEnv* env)
-{
-    erts_unblock_fpe(env->fpe_was_unmasked);
-    free_tmp_objs(env);
-}
-
-
 /* Flush out our cached heap pointers to allow an ordinary HAlloc
 */
 static void flush_env(ErlNifEnv* env)
@@ -247,18 +232,20 @@ struct enif_msg_environment_t
     Process phony_proc;
 };
 
-ErlNifEnv* enif_alloc_env(void)
+static ERTS_INLINE void
+setup_nif_env(struct enif_msg_environment_t* msg_env,
+              struct erl_module_nif* mod,
+              Process* tracee)
 {
-    struct enif_msg_environment_t* msg_env =
-	erts_alloc_fnf(ERTS_ALC_T_NIF, sizeof(struct enif_msg_environment_t));
     Eterm* phony_heap = (Eterm*) msg_env; /* dummy non-NULL ptr */
-	
-    msg_env->env.hp = phony_heap; 
+
+    msg_env->env.hp = phony_heap;
     msg_env->env.hp_end = phony_heap;
     msg_env->env.heap_frag = NULL;
-    msg_env->env.mod_nif = NULL;
+    msg_env->env.mod_nif = mod;
     msg_env->env.tmp_obj_list = NULL;
     msg_env->env.proc = &msg_env->phony_proc;
+    msg_env->env.exception_thrown = 0;
     memset(&msg_env->phony_proc, 0, sizeof(Process));
     HEAP_START(&msg_env->phony_proc) = phony_heap;
     HEAP_TOP(&msg_env->phony_proc) = phony_heap;
@@ -270,6 +257,14 @@ ErlNifEnv* enif_alloc_env(void)
     msg_env->phony_proc.space_verified = 0;
     msg_env->phony_proc.space_verified_from = NULL;
 #endif
+    msg_env->env.tracee = tracee;
+}
+
+ErlNifEnv* enif_alloc_env(void)
+{
+    struct enif_msg_environment_t* msg_env =
+	erts_alloc_fnf(ERTS_ALC_T_NIF, sizeof(struct enif_msg_environment_t));
+    setup_nif_env(msg_env, NULL, NULL);
     return &msg_env->env;
 }
 void enif_free_env(ErlNifEnv* env)
@@ -278,6 +273,20 @@ void enif_free_env(ErlNifEnv* env)
     erts_free(ERTS_ALC_T_NIF, env);
 }
 
+static ERTS_INLINE void pre_nif_noproc(struct enif_msg_environment_t* msg_env,
+                                       struct erl_module_nif* mod,
+                                       Process* tracee)
+{
+    setup_nif_env(msg_env, mod, tracee);
+    msg_env->env.fpe_was_unmasked = erts_block_fpe();
+}
+
+static ERTS_INLINE void post_nif_noproc(struct enif_msg_environment_t* msg_env)
+{
+    erts_unblock_fpe(msg_env->env.fpe_was_unmasked);
+    enif_clear_env(&msg_env->env);
+}
+
 static ERTS_INLINE void clear_offheap(ErlOffHeap* oh)
 {
     oh->first = NULL;
@@ -304,22 +313,111 @@ void enif_clear_env(ErlNifEnv* env)
     ASSERT(!is_offheap(&MSO(p)));
     free_tmp_objs(env);
 }
-int enif_send(ErlNifEnv* env, const ErlNifPid* to_pid,
-	      ErlNifEnv* msg_env, ERL_NIF_TERM msg)
+
+#ifdef ERTS_SMP
+#ifdef DEBUG
+static int enif_send_delay = 0;
+#define ERTS_FORCE_ENIF_SEND_DELAY() (enif_send_delay++ % 2 == 0)
+#else
+#ifdef ERTS_PROC_LOCK_OWN_IMPL
+#define ERTS_FORCE_ENIF_SEND_DELAY() 0
+#else
+/*
+ * We always schedule messages if we do not use our own
+ * process lock implementation, as if we try to do a trylock on
+ * a lock that might already be locked by the same thread.
+ * And what happens then with different mutex implementations
+ * is not always guaranteed.
+ */
+#define ERTS_FORCE_ENIF_SEND_DELAY() 1
+#endif
+#endif
+
+int erts_flush_trace_messages(Process *c_p, ErtsProcLocks c_p_locks)
+{
+    ErlTraceMessageQueue *msgq, **last_msgq;
+    int reds = 0;
+
+    erts_smp_proc_lock(c_p, ERTS_PROC_LOCK_TRACE);
+
+    msgq = c_p->trace_msg_q;
+
+    if (!msgq)
+        goto error;
+
+    do {
+        Process* rp;
+        ErtsProcLocks rp_locks;
+        ErtsMessage *first, **last;
+        Uint len;
+
+        first = msgq->first;
+        last = msgq->last;
+        len = msgq->len;
+        msgq->first = NULL;
+        msgq->last = &msgq->first;
+        msgq->len = 0;
+        erts_smp_proc_unlock(c_p, ERTS_PROC_LOCK_TRACE);
+
+        ASSERT(len != 0);
+
+        rp = erts_proc_lookup(msgq->receiver);
+        if (rp) {
+            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);
+            if (rp->common.id == c_p->common.id)
+                rp_locks &= ~c_p_locks;
+            if (rp_locks)
+                erts_smp_proc_unlock(rp, rp_locks);
+            reds += len;
+        } else {
+            erts_cleanup_messages(first);
+        }
+        reds += 1;
+        erts_smp_proc_lock(c_p, ERTS_PROC_LOCK_TRACE);
+        msgq = msgq->next;
+    } while (msgq);
+
+    last_msgq = &c_p->trace_msg_q;
+
+    while (*last_msgq) {
+        msgq = *last_msgq;
+        if (msgq->len == 0) {
+            *last_msgq = msgq->next;
+            erts_free(ERTS_ALC_T_TRACE_MSG_QUEUE, msgq);
+        } else {
+            last_msgq = &msgq->next;
+        }
+    }
+
+error:
+    erts_smp_proc_unlock(c_p, ERTS_PROC_LOCK_TRACE);
+
+    return reds;
+}
+
+#endif
+
+int
+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;
+    ErtsProcLocks rp_locks = 0, lc_locks = 0, c_p_locks = ERTS_PROC_LOCK_MAIN;
     Process* rp;
     Process* c_p;
     ErtsMessage *mp;
     Eterm receiver = to_pid->pid;
     int flush_me = 0;
-    int scheduler = erts_get_scheduler_id() != 0;
+    ErtsSchedulerData *esdp = erts_get_scheduler_data();
+    int scheduler = esdp ? esdp->no : 0;
 
     if (env != NULL) {
 	c_p = env->proc;
 	if (receiver == c_p->common.id) {
-	    rp_locks = ERTS_PROC_LOCK_MAIN;
+	    rp_locks = c_p_locks;
 	    flush_me = 1;
 	}
     }
@@ -329,12 +427,13 @@ int enif_send(ErlNifEnv* env, const ErlNifPid* to_pid,
 #else
 	erts_exit(ERTS_ABORT_EXIT,"enif_send: env==NULL on non-SMP VM");
 #endif
-    }    
+    }
 
     rp = (scheduler
 	  ? erts_proc_lookup(receiver)
 	  : erts_pid2proc_opt(c_p, ERTS_PROC_LOCK_MAIN,
 			      receiver, rp_locks, ERTS_P2P_FLG_INC_REFC));
+
     if (rp == NULL) {
 	ASSERT(env == NULL || receiver != c_p->common.id);
 	return 0;
@@ -353,7 +452,6 @@ int enif_send(ErlNifEnv* env, const ErlNifPid* to_pid,
             menv->env.heap_frag = NULL;
             MBUF(&menv->phony_proc) = NULL;
         }
-        ASSERT(!is_offheap(&MSO(&menv->phony_proc)));
     } else {
         Uint sz = size_object(msg);
         Eterm *hp;
@@ -362,14 +460,91 @@ int enif_send(ErlNifEnv* env, const ErlNifPid* to_pid,
         ASSERT(hp == mp->hfrag.mem+mp->hfrag.used_size);
     }
 
-    if (flush_me) {	
-	flush_env(env); /* Needed for ERTS_HOLE_CHECK */ 
+    ERL_MESSAGE_TERM(mp) = msg;
+
+    if (flush_me) {
+	flush_env(env); /* Needed for ERTS_HOLE_CHECK */
+    }
+
+    if (!env || !env->tracee) {
+
+        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
     }
-    erts_queue_message(rp, &rp_locks, mp, msg, am_undefined);
+    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
+           so we have to do a try lock and if that fails we enqueue
+           the message in a special trace message output queue of the
+           tracee */
+        ErlTraceMessageQueue *msgq;
+        Process *t_p = env->tracee;
+
+
+        erts_smp_proc_lock(t_p, ERTS_PROC_LOCK_TRACE);
+
+        msgq = t_p->trace_msg_q;
+
+        while (msgq != NULL) {
+            if (msgq->receiver == receiver) {
+                break;
+            }
+            msgq = msgq->next;
+        }
+
+#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 ||
+            erts_smp_proc_trylock(rp, ERTS_PROC_LOCK_MSGQ) == EBUSY) {
+
+            if (!msgq) {
+
+                msgq = erts_alloc(ERTS_ALC_T_TRACE_MSG_QUEUE,
+                                  sizeof(ErlTraceMessageQueue));
+                msgq->receiver = receiver;
+                msgq->first = mp;
+                msgq->last = &mp->next;
+                msgq->len = 1;
+
+                /* Insert in linked list */
+                msgq->next = t_p->trace_msg_q;
+                t_p->trace_msg_q = msgq;
+
+                erts_smp_proc_unlock(t_p, ERTS_PROC_LOCK_TRACE);
+
+                erts_schedule_flush_trace_messages(t_p->common.id);
+
+            } else {
+                msgq->len++;
+                *msgq->last = mp;
+                msgq->last = &mp->next;
+                erts_smp_proc_unlock(t_p, ERTS_PROC_LOCK_TRACE);
+            }
+        } 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
+
     if (c_p == rp)
 	rp_locks &= ~ERTS_PROC_LOCK_MAIN;
-    if (rp_locks)
-	erts_smp_proc_unlock(rp, rp_locks);
+    if (rp_locks & ~lc_locks)
+	erts_smp_proc_unlock(rp, rp_locks & ~lc_locks);
     if (!scheduler)
 	erts_proc_dec_refc(rp);
     if (flush_me) {
@@ -398,6 +573,9 @@ enif_port_command(ErlNifEnv *env, const ErlNifPort* to_port,
     if (!prt)
         return 0;
 
+    if (IS_TRACED_FL(prt, F_TRACE_RECEIVE))
+        trace_port_receive(prt, env->proc->common.id, am_command, msg);
+
     return erts_port_output_async(prt, env->proc->common.id, msg);
 }
 
@@ -1485,10 +1663,10 @@ static void close_lib(struct erl_module_nif* lib)
     ASSERT(erts_refc_read(&lib->rt_dtor_cnt,0) == 0);
 
     if (lib->entry != NULL && lib->entry->unload != NULL) {
-	ErlNifEnv env;
-	pre_nif_noproc(&env, lib);
-	lib->entry->unload(&env, lib->priv_data);
-	post_nif_noproc(&env);
+	struct enif_msg_environment_t msg_env;
+        pre_nif_noproc(&msg_env, lib, NULL);
+	lib->entry->unload(&msg_env.env, lib->priv_data);
+        post_nif_noproc(&msg_env);
     }
     if (!erts_is_static_nif(lib->handle))
       erts_sys_ddll_close(lib->handle);
@@ -1634,10 +1812,10 @@ static void nif_resource_dtor(Binary* bin)
     ASSERT(ERTS_MAGIC_BIN_DESTRUCTOR(bin) == &nif_resource_dtor);
 
     if (type->dtor != NULL) {
-	ErlNifEnv env;
-	pre_nif_noproc(&env, type->owner);
-	type->dtor(&env,resource->data);
-	post_nif_noproc(&env);
+        struct enif_msg_environment_t msg_env;
+        pre_nif_noproc(&msg_env, type->owner, NULL);
+	type->dtor(&msg_env.env, resource->data);
+        post_nif_noproc(&msg_env);
     }
     if (erts_refc_dectest(&type->refc, 0) == 0) {
 	ASSERT(type->next == NULL);
@@ -2797,7 +2975,7 @@ BIF_RETTYPE load_nif_2(BIF_ALIST_2)
 	    }
 	    old_func = next_func(mod->curr.nif->entry, &old_incr, old_func);
 	}
-	erts_pre_nif(&env, BIF_P, lib);
+	erts_pre_nif(&env, BIF_P, lib, NULL);
 	veto = entry->reload(&env, &lib->priv_data, BIF_ARG_2);
 	erts_post_nif(&env);
 	if (veto) {
@@ -2818,7 +2996,7 @@ BIF_RETTYPE load_nif_2(BIF_ALIST_2)
 		ret = load_nif_error(BIF_P, upgrade, "Upgrade not supported by this NIF library.");
 		goto error;
 	    }
-	    erts_pre_nif(&env, BIF_P, lib);
+	    erts_pre_nif(&env, BIF_P, lib, NULL);
 	    veto = entry->upgrade(&env, &lib->priv_data, &mod->old.nif->priv_data, BIF_ARG_2);
 	    erts_post_nif(&env);
 	    if (veto) {
@@ -2829,7 +3007,7 @@ BIF_RETTYPE load_nif_2(BIF_ALIST_2)
 		commit_opened_resource_types(lib);
 	}
 	else if (entry->load != NULL) { /********* Initial load ***********/
-	    erts_pre_nif(&env, BIF_P, lib);
+	    erts_pre_nif(&env, BIF_P, lib, NULL);
 	    veto = entry->load(&env, &lib->priv_data, BIF_ARG_2);
 	    erts_post_nif(&env);
 	    if (veto) {
@@ -2916,6 +3094,9 @@ erts_unload_nif(struct erl_module_nif* lib)
     ASSERT(erts_smp_thr_progress_is_blocking());
     ASSERT(lib != NULL);
     ASSERT(lib->mod != NULL);
+
+    erts_tracer_nif_clear();
+
     for (rt = resource_type_list.next;
 	 rt != &resource_type_list;
 	 rt = next) {
@@ -2958,6 +3139,76 @@ void erl_nif_init()
     resource_type_list.owner = NULL;
     resource_type_list.module = THE_NON_VALUE;
     resource_type_list.name = THE_NON_VALUE;
+
+}
+
+int erts_nif_get_funcs(struct erl_module_nif* mod,
+                       ErlNifFunc **funcs)
+{
+    *funcs = mod->entry->funcs;
+    return mod->entry->num_of_funcs;
+}
+
+Eterm erts_nif_call_function(Process *p, Process *tracee,
+                             struct erl_module_nif* mod,
+                             ErlNifFunc *fun, int argc, Eterm *argv)
+{
+    Eterm nif_result;
+#ifdef DEBUG
+    /* Verify that function is part of this module */
+    int i;
+    for (i = 0; i < mod->entry->num_of_funcs; i++)
+        if (fun == mod->entry->funcs+i)
+            break;
+    ASSERT(i < mod->entry->num_of_funcs);
+    if (p)
+        ERTS_SMP_LC_ASSERT(erts_proc_lc_my_proc_locks(p) & ERTS_PROC_LOCK_MAIN
+                           || erts_smp_thr_progress_is_blocking());
+#endif
+    if (p) {
+        /* This is almost a normal nif call like in beam_emu,
+           except that any heap fragment created in the nif will be
+           discarded without checking if anything in it is live.
+           This is because we cannot do a GC here as we don't know
+           the number of live registers that have to be preserved.
+           This means that any heap part of the returned term may
+           not be used outside this function. */
+        struct enif_environment_t env;
+        ErlHeapFragment *orig_hf = MBUF(p);
+        ErlOffHeap orig_oh = MSO(p);
+        ASSERT(is_internal_pid(p->common.id));
+        MBUF(p) = NULL;
+        clear_offheap(&MSO(p));
+
+        erts_pre_nif(&env, p, mod, tracee);
+        nif_result = (*fun->fptr)(&env, argc, argv);
+        if (env.exception_thrown)
+            nif_result = THE_NON_VALUE;
+        erts_post_nif(&env);
+
+        /* Free any offheap and heap fragments created in nif */
+        if (MSO(p).first) {
+            erts_cleanup_offheap(&MSO(p));
+            clear_offheap(&MSO(p));
+        }
+        if (MBUF(p))
+            free_message_buffer(MBUF(p));
+
+        /* restore original heap fragment list */
+        MBUF(p) = orig_hf;
+        MSO(p) = orig_oh;
+    } else {
+        /* Nif call was done without a process context,
+           so we create a phony one. */
+        struct enif_msg_environment_t msg_env;
+        pre_nif_noproc(&msg_env, mod, tracee);
+        nif_result = (*fun->fptr)(&msg_env.env, argc, argv);
+        if (msg_env.env.exception_thrown)
+            nif_result = THE_NON_VALUE;
+        post_nif_noproc(&msg_env);
+    }
+
+    return nif_result;
 }
 
 #ifdef USE_VM_PROBES
diff --git a/erts/emulator/beam/erl_nif.h b/erts/emulator/beam/erl_nif.h
index 44a2581436..3964f7f679 100644
--- a/erts/emulator/beam/erl_nif.h
+++ b/erts/emulator/beam/erl_nif.h
@@ -87,16 +87,16 @@ typedef ErlNifSInt64 ErlNifTime;
 #define ERL_NIF_TIME_ERROR ((ErlNifSInt64) ERTS_NAPI_TIME_ERROR__)
 
 typedef enum {
-    ERL_NIF_SEC = ERTS_NAPI_SEC__,
-    ERL_NIF_MSEC = ERTS_NAPI_MSEC__,
-    ERL_NIF_USEC = ERTS_NAPI_USEC__,
-    ERL_NIF_NSEC = ERTS_NAPI_NSEC__
+    ERL_NIF_SEC    = ERTS_NAPI_SEC__,
+    ERL_NIF_MSEC   = ERTS_NAPI_MSEC__,
+    ERL_NIF_USEC   = ERTS_NAPI_USEC__,
+    ERL_NIF_NSEC   = ERTS_NAPI_NSEC__
 } ErlNifTimeUnit;
 
 struct enif_environment_t;
 typedef struct enif_environment_t ErlNifEnv;
 
-typedef struct
+typedef struct enif_func_t
 {
     const char* name;
     unsigned arity;
diff --git a/erts/emulator/beam/erl_port.h b/erts/emulator/beam/erl_port.h
index 59aa034f48..c2588e718d 100644
--- a/erts/emulator/beam/erl_port.h
+++ b/erts/emulator/beam/erl_port.h
@@ -944,7 +944,7 @@ erts_schedule_proc2port_signal(Process *,
 			       ErtsPortTaskHandle *,
 			       ErtsProc2PortSigCallback);
 
-int erts_deliver_port_exit(Port *, Eterm, Eterm, int);
+int erts_deliver_port_exit(Port *, Eterm, Eterm, int, int);
 
 /*
  * Port signal flags
diff --git a/erts/emulator/beam/erl_port_task.c b/erts/emulator/beam/erl_port_task.c
index 30ce181ebb..3102e44c11 100644
--- a/erts/emulator/beam/erl_port_task.c
+++ b/erts/emulator/beam/erl_port_task.c
@@ -1766,6 +1766,8 @@ erts_port_task_execute(ErtsRunQueue *runq, Port **curr_port_pp)
 		if (!(state & ERTS_PORT_SFLGS_DEAD)) {
 		    DTRACE_DRIVER(driver_timeout, pp);
 		    LTTNG_DRIVER(driver_timeout, pp);
+                    if (IS_TRACED_FL(pp, F_TRACE_RECEIVE))
+                        trace_port(pp, am_receive, am_timeout);
 		    (*pp->drv_ptr->timeout)((ErlDrvData) pp->drv_data);
 		}
 	    }
diff --git a/erts/emulator/beam/erl_process.c b/erts/emulator/beam/erl_process.c
index d222e79f56..a520005c35 100644
--- a/erts/emulator/beam/erl_process.c
+++ b/erts/emulator/beam/erl_process.c
@@ -444,7 +444,8 @@ int erts_system_profile_ts_type = ERTS_TRACE_FLG_NOW_TIMESTAMP;
 typedef enum {
     ERTS_PSTT_GC,	/* Garbage Collect */
     ERTS_PSTT_CPC,	/* Check Process Code */
-    ERTS_PSTT_COHMQ     /* Change off heap message queue */
+    ERTS_PSTT_COHMQ,    /* Change off heap message queue */
+    ERTS_PSTT_FTMQ      /* Flush trace msg queue */
 } ErtsProcSysTaskType;
 
 #define ERTS_MAX_PROC_SYS_TASK_ARGS 2
@@ -1138,7 +1139,7 @@ reply_sched_wall_time(void *vswtrp)
 	hpp = &hp;
     }
 
-    erts_queue_message(rp, &rp_locks, mp, msg, NIL);
+    erts_queue_message(rp, &rp_locks, mp, msg);
 
     if (swtrp->req_sched == esdp->no)
 	rp_locks &= ~ERTS_PROC_LOCK_MAIN;
@@ -1217,7 +1218,7 @@ reply_system_check(void *vscrp)
     hpp = &hp;
     msg = STORE_NC(hpp, ohp, scrp->ref);
 
-    erts_queue_message(rp, &rp_locks, mp, msg, NIL);
+    erts_queue_message(rp, &rp_locks, mp, msg);
 
     if (scrp->req_sched == esdp->no)
 	rp_locks &= ~ERTS_PROC_LOCK_MAIN;
@@ -2311,7 +2312,7 @@ handle_reap_ports(ErtsAuxWorkData *awdp, erts_aint32_t aux_work, int waiting)
 						    ERTS_PORT_SFLG_HALT);
 		erts_smp_atomic32_inc_nob(&erts_halt_progress);
 		if (!(state & (ERTS_PORT_SFLG_EXITING|ERTS_PORT_SFLG_CLOSING)))
-		    erts_deliver_port_exit(prt, prt->common.id, am_killed, 0);
+		    erts_deliver_port_exit(prt, prt->common.id, am_killed, 0, 1);
 	    }
 
 	    erts_port_release(prt);
@@ -9315,6 +9316,7 @@ Process *schedule(Process *p, int calls)
 	esdp = erts_scheduler_data;
 	ASSERT(esdp->current_process == p);
 #endif
+
 	reds = actual_reds = calls - esdp->virtual_reds;
 	if (reds < ERTS_PROC_MIN_CONTEXT_SWITCH_REDS_COST)
 	    reds = ERTS_PROC_MIN_CONTEXT_SWITCH_REDS_COST;
@@ -9327,27 +9329,39 @@ Process *schedule(Process *p, int calls)
 
 	p->reds += actual_reds;
 
-	erts_smp_proc_lock(p, ERTS_PROC_LOCK_STATUS);
+#ifdef ERTS_SMP
+	erts_smp_proc_lock(p, ERTS_PROC_LOCK_TRACE);
+        if (p->trace_msg_q) {
+            erts_smp_proc_unlock(p, ERTS_PROC_LOCK_TRACE);
+            erts_schedule_flush_trace_messages(p->common.id);
+        } else
+            erts_smp_proc_unlock(p, ERTS_PROC_LOCK_TRACE);
+#endif
 
-	state = erts_smp_atomic32_read_acqb(&p->state);
+	state = erts_smp_atomic32_read_nob(&p->state);
 
 	if (IS_TRACED(p)) {
 	    if (IS_TRACED_FL(p, F_TRACE_CALLS) && !(state & ERTS_PSFLG_FREE))
 		erts_schedule_time_break(p, ERTS_BP_CALL_TIME_SCHEDULE_OUT);
-	    if (state & (ERTS_PSFLG_FREE|ERTS_PSFLG_EXITING)) {
+	    if ((state & (ERTS_PSFLG_FREE|ERTS_PSFLG_EXITING)) == ERTS_PSFLG_EXITING) {
 		if (ARE_TRACE_FLAGS_ON(p, F_TRACE_SCHED_EXIT))
-		    trace_sched(p, ((state & ERTS_PSFLG_FREE)
-				    ? am_out_exited
-				    : am_out_exiting));
+		    trace_sched(p, ERTS_PROC_LOCK_MAIN,
+                                ((state & ERTS_PSFLG_FREE)
+                                 ? am_out_exited
+                                 : am_out_exiting));
 	    }
 	    else {
-		if (ARE_TRACE_FLAGS_ON(p, F_TRACE_SCHED))
-		    trace_sched(p, am_out);
-		else if (ARE_TRACE_FLAGS_ON(p, F_TRACE_SCHED_PROCS))
-		    trace_virtual_sched(p, am_out);
+		if (ARE_TRACE_FLAGS_ON(p, F_TRACE_SCHED) ||
+                    ARE_TRACE_FLAGS_ON(p, F_TRACE_SCHED_PROCS))
+		    trace_sched(p, ERTS_PROC_LOCK_MAIN, am_out);
 	    }
 	}
 
+	erts_smp_proc_lock(p, ERTS_PROC_LOCK_STATUS);
+
+        /* have to re-read state after taking lock */
+        state = erts_smp_atomic32_read_nob(&p->state);
+
 #ifdef ERTS_SMP
 	if (state & ERTS_PSFLG_PENDING_EXIT)
 	    erts_handle_pending_exit(p, (ERTS_PROC_LOCK_MAIN
@@ -9844,27 +9858,27 @@ Process *schedule(Process *p, int calls)
 
 	p->fcalls = reds;
 
+	erts_smp_proc_unlock(p, ERTS_PROC_LOCK_STATUS);
+
 	if (IS_TRACED(p)) {
 	    if (state & ERTS_PSFLG_EXITING) {
 		if (ARE_TRACE_FLAGS_ON(p, F_TRACE_SCHED_EXIT))
-		    trace_sched(p, am_in_exiting);
+		    trace_sched(p, ERTS_PROC_LOCK_MAIN, am_in_exiting);
 	    }
 	    else {
-		if (ARE_TRACE_FLAGS_ON(p, F_TRACE_SCHED))
-		    trace_sched(p, am_in);
-		else if (ARE_TRACE_FLAGS_ON(p, F_TRACE_SCHED_PROCS))
-		    trace_virtual_sched(p, am_in);
+		if (ARE_TRACE_FLAGS_ON(p, F_TRACE_SCHED) ||
+                    ARE_TRACE_FLAGS_ON(p, F_TRACE_SCHED_PROCS))
+		    trace_sched(p, ERTS_PROC_LOCK_MAIN, am_in);
 	    }
 	    if (IS_TRACED_FL(p, F_TRACE_CALLS)) {
 		erts_schedule_time_break(p, ERTS_BP_CALL_TIME_SCHEDULE_IN);
 	    }
 	}
 
-	erts_smp_proc_unlock(p, ERTS_PROC_LOCK_STATUS);
 
 #ifdef ERTS_SMP
-	if (is_not_nil(ERTS_TRACER_PROC(p)))
-	    erts_check_my_tracer_proc(p);
+        /* Clears tracer if it has been removed */
+        (void)ERTS_TRACER_PROC_IS_ENABLED(p);
 #endif
 
 	if (state & ERTS_PSFLG_RUNNING_SYS) {
@@ -9996,7 +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, NIL);
+	erts_queue_message(rp, &rp_locks, mp, msg);
 
 	if (c_p == rp)
 	    rp_locks &= ~ERTS_PROC_LOCK_MAIN;
@@ -10164,8 +10178,7 @@ execute_sys_tasks(Process *c_p, erts_aint32_t *statep, int in_reds)
 {
     int garbage_collected = 0;
     erts_aint32_t state = *statep;
-    int max_reds = in_reds;
-    int reds = 0;
+    int reds = in_reds;
     int qmask = 0;
 
     ERTS_SMP_LC_ASSERT(erts_proc_lc_my_proc_locks(c_p) == ERTS_PROC_LOCK_MAIN);
@@ -10193,12 +10206,12 @@ execute_sys_tasks(Process *c_p, erts_aint32_t *statep, int in_reds)
 	    if (c_p->flags & F_DISABLE_GC) {
 		save_gc_task(c_p, st, st_prio);
 		st = NULL;
-		reds++;
+		reds--;
 	    }
 	    else {
 		if (!garbage_collected) {
 		    FLAGS(c_p) |= F_NEED_FULLSWEEP;
-		    reds += erts_garbage_collect_nobump(c_p,
+		    reds -= erts_garbage_collect_nobump(c_p,
 							0,
 							c_p->arg_reg,
 							c_p->arity);
@@ -10207,21 +10220,30 @@ execute_sys_tasks(Process *c_p, erts_aint32_t *statep, int in_reds)
 		st_res = am_true;
 	    }
 	    break;
-	case ERTS_PSTT_CPC:
+	case ERTS_PSTT_CPC: {
+            int cpc_reds = 0;
 	    st_res = erts_check_process_code(c_p,
 					     st->arg[0],
                                              unsigned_val(st->arg[1]),
-					     &reds);
+					     &cpc_reds);
+            reds -= cpc_reds;
 	    if (is_non_value(st_res)) {
 		/* Needed gc, but gc was disabled */
 		save_gc_task(c_p, st, st_prio);
 		st = NULL;
 	    }
 	    break;
+        }
 	case ERTS_PSTT_COHMQ:
-	    reds += erts_complete_off_heap_message_queue_change(c_p);
+	    reds -= erts_complete_off_heap_message_queue_change(c_p);
+	    st_res = am_true;
+	    break;
+#ifdef ERTS_SMP
+        case ERTS_PSTT_FTMQ:
+	    reds -= erts_flush_trace_messages(c_p, ERTS_PROC_LOCK_MAIN);
 	    st_res = am_true;
 	    break;
+#endif
 	default:
 	    ERTS_INTERNAL_ERROR("Invalid process sys task type");
 	    st_res = am_false;
@@ -10231,11 +10253,11 @@ execute_sys_tasks(Process *c_p, erts_aint32_t *statep, int in_reds)
 	    reds += notify_sys_task_executed(c_p, st, st_res);
 
 	state = erts_smp_atomic32_read_acqb(&c_p->state);
-    } while (qmask && reds < max_reds);
+    } while (qmask && reds > 0);
 
     *statep = state;
 
-    return reds;
+    return in_reds - reds;
 }
 
 static int
@@ -10267,6 +10289,12 @@ cleanup_sys_tasks(Process *c_p, erts_aint32_t in_state, int in_reds)
 	case ERTS_PSTT_COHMQ:
 	    st_res = am_false;
 	    break;
+#ifdef ERTS_SMP
+        case ERTS_PSTT_FTMQ:
+	    reds -= erts_flush_trace_messages(c_p, ERTS_PROC_LOCK_MAIN);
+	    st_res = am_true;
+	    break;
+#endif
 	default:
 	    ERTS_INTERNAL_ERROR("Invalid process sys task type");
 	    st_res = am_false;
@@ -10402,8 +10430,8 @@ badarg:
     BIF_ERROR(BIF_P, BADARG);
 }
 
-void
-erts_schedule_complete_off_heap_message_queue_change(Eterm pid)
+static void
+erts_schedule_generic_sys_task(Eterm pid, ErtsProcSysTaskType type)
 {
     Process *rp = erts_proc_lookup(pid);
     if (rp) {
@@ -10413,7 +10441,7 @@ erts_schedule_complete_off_heap_message_queue_change(Eterm pid)
 
 	st = erts_alloc(ERTS_ALC_T_PROC_SYS_TSK,
 			ERTS_PROC_SYS_TASK_SIZE(0));
-	st->type = ERTS_PSTT_COHMQ;
+	st->type = type;
 	st->requester = NIL;
 	st->reply_tag = NIL;
 	st->req_id = NIL;
@@ -10428,6 +10456,18 @@ erts_schedule_complete_off_heap_message_queue_change(Eterm pid)
     }
 }
 
+void
+erts_schedule_complete_off_heap_message_queue_change(Eterm pid)
+{
+    erts_schedule_generic_sys_task(pid, ERTS_PSTT_COHMQ);
+}
+
+void
+erts_schedule_flush_trace_messages(Eterm pid)
+{
+    erts_schedule_generic_sys_task(pid, ERTS_PSTT_FTMQ);
+}
+
 static void
 save_gc_task(Process *c_p, ErtsProcSysTask *st, int prio)
 {
@@ -10886,6 +10926,7 @@ erl_create_process(Process* parent, /* Parent of process (default group leader).
     Eterm res = THE_NON_VALUE;
     erts_aint32_t state = 0;
     erts_aint32_t prio = (erts_aint32_t) PRIORITY_NORMAL;
+    ErtsProcLocks locks = ERTS_PROC_LOCKS_ALL;
 #ifdef SHCOPY_SPAWN
     erts_shcopy_t info;
     INITIALIZE_SHCOPY(info);
@@ -11062,7 +11103,7 @@ erl_create_process(Process* parent, /* Parent of process (default group leader).
 	    : STORE_NC(&p->htop, &p->off_heap, parent->group_leader);
     }
 
-    erts_get_default_tracing(&ERTS_TRACE_FLAGS(p), &ERTS_TRACER_PROC(p));
+    erts_get_default_proc_tracing(&ERTS_TRACE_FLAGS(p), &ERTS_TRACER(p));
 
     p->msg.first = NULL;
     p->msg.last = &p->msg.first;
@@ -11098,21 +11139,58 @@ erl_create_process(Process* parent, /* Parent of process (default group leader).
     p->last_old_htop = NULL;
 #endif
 
+#ifdef ERTS_SMP
+    p->trace_msg_q = NULL;
+    p->scheduler_data = NULL;
+    p->suspendee = NIL;
+    p->pending_suspenders = NULL;
+    p->pending_exit.reason = THE_NON_VALUE;
+    p->pending_exit.bp = NULL;
+#endif
+
+#if !defined(NO_FPE_SIGNALS) || defined(HIPE)
+    p->fp_exception = 0;
+#endif
+
     if (IS_TRACED(parent)) {
 	if (ERTS_TRACE_FLAGS(parent) & F_TRACE_SOS) {
 	    ERTS_TRACE_FLAGS(p) |= (ERTS_TRACE_FLAGS(parent) & TRACEE_FLAGS);
-	    ERTS_TRACER_PROC(p) = ERTS_TRACER_PROC(parent);
+            erts_tracer_replace(&p->common, ERTS_TRACER(parent));
 	}
-	if (ARE_TRACE_FLAGS_ON(parent, F_TRACE_PROCS)) {
-	    trace_proc_spawn(parent, p->common.id, mod, func, args);
-	}
-	if (ERTS_TRACE_FLAGS(parent) & F_TRACE_SOS1) {
+        if (ERTS_TRACE_FLAGS(parent) & F_TRACE_SOS1) {
 	    /* Overrides TRACE_CHILDREN */
 	    ERTS_TRACE_FLAGS(p) |= (ERTS_TRACE_FLAGS(parent) & TRACEE_FLAGS);
-	    ERTS_TRACER_PROC(p) = ERTS_TRACER_PROC(parent);
+            erts_tracer_replace(&p->common, ERTS_TRACER(parent));
 	    ERTS_TRACE_FLAGS(p) &= ~(F_TRACE_SOS1 | F_TRACE_SOS);
 	    ERTS_TRACE_FLAGS(parent) &= ~(F_TRACE_SOS1 | F_TRACE_SOS);
 	}
+        if (so->flags & SPO_LINK && ERTS_TRACE_FLAGS(parent) & (F_TRACE_SOL|F_TRACE_SOL1)) {
+		ERTS_TRACE_FLAGS(p) |= (ERTS_TRACE_FLAGS(parent)&TRACEE_FLAGS);
+                erts_tracer_replace(&p->common, ERTS_TRACER(parent));
+		if (ERTS_TRACE_FLAGS(parent) & F_TRACE_SOL1) {/*maybe override*/
+		    ERTS_TRACE_FLAGS(p) &= ~(F_TRACE_SOL1 | F_TRACE_SOL);
+		    ERTS_TRACE_FLAGS(parent) &= ~(F_TRACE_SOL1 | F_TRACE_SOL);
+		}
+        }
+        if (ARE_TRACE_FLAGS_ON(parent, F_TRACE_PROCS)) {
+            locks &= ~(ERTS_PROC_LOCK_STATUS|ERTS_PROC_LOCK_TRACE);
+            erts_smp_proc_unlock(p, ERTS_PROC_LOCK_STATUS|ERTS_PROC_LOCK_TRACE);
+            erts_smp_proc_unlock(parent, ERTS_PROC_LOCK_STATUS|ERTS_PROC_LOCK_TRACE);
+            trace_proc_spawn(parent, am_spawn, p->common.id, mod, func, args);
+            if (so->flags & SPO_LINK)
+                trace_proc(parent, locks, parent, am_link, p->common.id);
+        }
+    }
+
+    if (IS_TRACED_FL(p, F_TRACE_PROCS)) {
+        if (locks & (ERTS_PROC_LOCK_STATUS|ERTS_PROC_LOCK_TRACE)) {
+            locks &= ~(ERTS_PROC_LOCK_STATUS|ERTS_PROC_LOCK_TRACE);
+            erts_smp_proc_unlock(p, locks & (ERTS_PROC_LOCK_STATUS|ERTS_PROC_LOCK_TRACE));
+            erts_smp_proc_unlock(parent, locks & (ERTS_PROC_LOCK_STATUS|ERTS_PROC_LOCK_TRACE));
+        }
+        trace_proc_spawn(p, am_spawned, parent->common.id, mod, func, args);
+        if (so->flags & SPO_LINK)
+            trace_proc(p, locks, p, am_getting_linked, parent->common.id);
     }
 
     /*
@@ -11123,10 +11201,6 @@ erl_create_process(Process* parent, /* Parent of process (default group leader).
 #ifdef DEBUG
 	int ret;
 #endif
-	if (IS_TRACED_FL(parent, F_TRACE_PROCS)) {
-	    trace_proc(parent, parent, am_link, p->common.id);
-	}
-
 #ifdef DEBUG
 	ret = erts_add_link(&ERTS_P_LINKS(parent),  LINK_PID, p->common.id);
 	ASSERT(ret == 0);
@@ -11137,17 +11211,6 @@ erl_create_process(Process* parent, /* Parent of process (default group leader).
 	erts_add_link(&ERTS_P_LINKS(p), LINK_PID, parent->common.id);
 #endif
 
-	if (IS_TRACED(parent)) {
-	    if (ERTS_TRACE_FLAGS(parent) & (F_TRACE_SOL|F_TRACE_SOL1)) {
-		ERTS_TRACE_FLAGS(p) |= (ERTS_TRACE_FLAGS(parent)&TRACEE_FLAGS);
-		ERTS_TRACER_PROC(p) = ERTS_TRACER_PROC(parent); /*maybe steal*/
-
-		if (ERTS_TRACE_FLAGS(parent) & F_TRACE_SOL1) {/*maybe override*/
-		    ERTS_TRACE_FLAGS(p) &= ~(F_TRACE_SOL1 | F_TRACE_SOL);
-		    ERTS_TRACE_FLAGS(parent) &= ~(F_TRACE_SOL1 | F_TRACE_SOL);
-		}
-	    }
-	}
     }
 
     /*
@@ -11162,19 +11225,7 @@ erl_create_process(Process* parent, /* Parent of process (default group leader).
 	so->mref = mref;
     }
 
-#ifdef ERTS_SMP
-    p->scheduler_data = NULL;
-    p->suspendee = NIL;
-    p->pending_suspenders = NULL;
-    p->pending_exit.reason = THE_NON_VALUE;
-    p->pending_exit.bp = NULL;
-#endif
-
-#if !defined(NO_FPE_SIGNALS) || defined(HIPE)
-    p->fp_exception = 0;
-#endif
-
-    erts_smp_proc_unlock(p, ERTS_PROC_LOCKS_ALL);
+    erts_smp_proc_unlock(p, locks);
 
     res = p->common.id;
 
@@ -11198,7 +11249,7 @@ erl_create_process(Process* parent, /* Parent of process (default group leader).
 
  error:
 
-    erts_smp_proc_unlock(parent, ERTS_PROC_LOCKS_ALL_MINOR);
+    erts_smp_proc_unlock(parent, locks & ERTS_PROC_LOCKS_ALL_MINOR);
 
     return res;
 }
@@ -11223,7 +11274,7 @@ void erts_init_empty_process(Process *p)
     p->rcount = 0;
     p->common.id = ERTS_INVALID_PID;
     p->reds = 0;
-    ERTS_TRACER_PROC(p) = NIL;
+    ERTS_TRACER(p) = erts_tracer_nil;
     ERTS_TRACE_FLAGS(p) = F_INITIAL_TRACE_FLAGS;
     p->group_leader = ERTS_INVALID_PID;
     p->flags = 0;
@@ -11343,7 +11394,7 @@ erts_debug_verify_clean_empty_process(Process* p)
     ASSERT(p->live_hf_end == ERTS_INVALID_HFRAG_PTR);
     ASSERT(p->heap == NULL);
     ASSERT(p->common.id == ERTS_INVALID_PID);
-    ASSERT(ERTS_TRACER_PROC(p) == NIL);
+    ASSERT(ERTS_TRACER_IS_NIL(ERTS_TRACER(p)));
     ASSERT(ERTS_TRACE_FLAGS(p) == F_INITIAL_TRACE_FLAGS);
     ASSERT(p->group_leader == ERTS_INVALID_PID);
     ASSERT(p->next == NULL);
@@ -11690,7 +11741,7 @@ 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, NIL);
+	erts_queue_message(to, to_locksp, mp, mess);
     } else {
 	Eterm temp_token;
 	Uint sz_token;
@@ -11708,9 +11759,10 @@ send_exit_message(Process *to, ErtsProcLocks *to_locksp,
 	mess = copy_struct(exit_term, term_size, &hp, ohp);
 #endif
 	/* the trace token must in this case be updated by the caller */
-	seq_trace_output(token, mess, SEQ_TRACE_SEND, to->common.id, NULL);
+	seq_trace_output(token, mess, SEQ_TRACE_SEND, to->common.id, to);
 	temp_token = copy_struct(token, sz_token, &hp, ohp);
-	erts_queue_message(to, to_locksp, mp, mess, temp_token);
+        ERL_MESSAGE_TOKEN(mp) = temp_token;
+	erts_queue_message(to, to_locksp, mp, mess);
     }
 }
 
@@ -11819,6 +11871,9 @@ send_exit_signal(Process *c_p,		/* current process if and only
 
     if ((state & ERTS_PSFLG_TRAP_EXIT)
 	&& (reason != am_kill || (flags & ERTS_XSIG_FLG_IGN_KILL))) {
+        /* have to release the status lock in order to send the exit message */
+        erts_smp_proc_unlock(rp, *rp_locks & ERTS_PROC_LOCKS_XSIG_SEND);
+        *rp_locks &= ~ERTS_PROC_LOCKS_XSIG_SEND;
         if (have_seqtrace(token) && token_update)
 	    seq_trace_update_send(token_update);
 	if (is_value(exit_tuple))
@@ -12116,6 +12171,7 @@ static void doit_exit_link(ErtsLink *lnk, void *vpcontext)
     DistEntry *dep;
     Process *rp;
 
+
     switch(lnk->type) {
     case LINK_PID:
 	if(is_internal_port(item)) {
@@ -12163,7 +12219,11 @@ static void doit_exit_link(ErtsLink *lnk, void *vpcontext)
 		    if (xres >= 0 && IS_TRACED_FL(rp, F_TRACE_PROCS)) {
 			/* We didn't exit the process and it is traced */
 			if (IS_TRACED_FL(rp, F_TRACE_PROCS)) {
-			    trace_proc(p, rp, am_getting_unlinked, p->common.id);
+                            if (rp_locks & ERTS_PROC_LOCKS_XSIG_SEND) {
+                                erts_smp_proc_unlock(rp, ERTS_PROC_LOCKS_XSIG_SEND);
+                                rp_locks &= ~ERTS_PROC_LOCKS_XSIG_SEND;
+                            }
+			    trace_proc(NULL, 0, rp, am_getting_unlinked, p->common.id);
 			}
 		    }
 		}
@@ -12280,8 +12340,6 @@ erts_do_exit_process(Process* p, Eterm reason)
 	if (IS_TRACED_FL(p, F_TRACE_CALLS))
 	    erts_schedule_time_break(p, ERTS_BP_CALL_TIME_SCHEDULE_EXITING);
 
-	if (IS_TRACED_FL(p,F_TRACE_PROCS))
-	    trace_proc(p, p, am_exit, reason);
     }
 
     erts_trace_check_exiting(p->common.id);
@@ -12297,6 +12355,10 @@ erts_do_exit_process(Process* p, Eterm reason)
 
     erts_smp_proc_unlock(p, ERTS_PROC_LOCKS_ALL_MINOR);
 
+    if (IS_TRACED_FL(p,F_TRACE_PROCS))
+        trace_proc(p, ERTS_PROC_LOCK_MAIN, p, am_exit, reason);
+
+
     /*
      * p->u.initial of this process can *not* be used anymore;
      * will be overwritten by misc termination data.
@@ -12436,6 +12498,9 @@ erts_continue_exit_process(Process *p)
 	ASSERT(!p->common.u.alive.reg);
     }
 
+    if (IS_TRACED_FL(p, F_TRACE_SCHED_EXIT))
+        trace_sched(p, curr_locks, am_out_exited);
+
     erts_smp_proc_lock(p, ERTS_PROC_LOCKS_ALL_MINOR);
     curr_locks = ERTS_PROC_LOCKS_ALL;
 
@@ -12559,6 +12624,12 @@ erts_continue_exit_process(Process *p)
     if (nif_export)
 	erts_destroy_nif_export(nif_export);
 
+#ifdef ERTS_SMP
+    erts_flush_trace_messages(p, 0);
+#endif
+
+    ERTS_TRACER_CLEAR(&ERTS_TRACER(p));
+
     delete_process(p);
 
 #ifdef ERTS_SMP
diff --git a/erts/emulator/beam/erl_process.h b/erts/emulator/beam/erl_process.h
index c2303eea49..61acf5924b 100644
--- a/erts/emulator/beam/erl_process.h
+++ b/erts/emulator/beam/erl_process.h
@@ -1041,6 +1041,7 @@ struct process {
 
 #ifdef ERTS_SMP
     ErlMessageInQueue msg_inq;
+    ErlTraceMessageQueue *trace_msg_q;
     ErtsPendExit pending_exit;
     erts_proc_lock_t lock;
     ErtsSchedulerData *scheduler_data;
@@ -1358,7 +1359,6 @@ extern int erts_system_profile_ts_type;
 #define F_TRACE_FLAG(N)      (1 << (ERTS_TRACE_TS_TYPE_BITS + (N)))
 
 /* process trace_flags */
-
 #define F_NOW_TS             (ERTS_TRACE_FLG_NOW_TIMESTAMP \
 			      << ERTS_TRACE_FLAGS_TS_TYPE_SHIFT)
 #define F_STRICT_MON_TS      (ERTS_TRACE_FLG_STRICT_MONOTONIC_TIMESTAMP \
@@ -1380,8 +1380,7 @@ extern int erts_system_profile_ts_type;
 #define F_TRACE_ARITY_ONLY   F_TRACE_FLAG(12)
 #define F_TRACE_RETURN_TO    F_TRACE_FLAG(13) /* Return_to trace when breakpoint tracing */
 #define F_TRACE_SILENT       F_TRACE_FLAG(14) /* No call trace msg suppress */
-#define F_TRACER             F_TRACE_FLAG(15) /* May be (has been) tracer */
-#define F_EXCEPTION_TRACE    F_TRACE_FLAG(16) /* May have exception trace on stack */
+#define F_EXCEPTION_TRACE    F_TRACE_FLAG(15) /* May have exception trace on stack */
 
 /* port trace flags, currently the same as process trace flags */
 #define F_TRACE_SCHED_PORTS  F_TRACE_FLAG(17) /* Trace of port scheduling */
@@ -1410,12 +1409,14 @@ extern int erts_system_profile_ts_type;
 		     | F_TRACE_SCHED_PORTS | F_TRACE_SCHED_NO \
 		     | F_TRACE_SCHED_EXIT)
 
+
 #define ERTS_TRACEE_MODIFIER_FLAGS \
-  (F_TRACE_SILENT | F_TIMESTAMP_MASK | F_TRACE_SCHED_NO)
-#define ERTS_PORT_TRACEE_FLAGS \
-  (ERTS_TRACEE_MODIFIER_FLAGS | F_TRACE_PORTS | F_TRACE_SCHED_PORTS)
+    (F_TRACE_SILENT | F_TIMESTAMP_MASK | F_TRACE_SCHED_NO \
+     | F_TRACE_RECEIVE | F_TRACE_SEND)
+#define ERTS_PORT_TRACEE_FLAGS                                     \
+    (ERTS_TRACEE_MODIFIER_FLAGS | F_TRACE_PORTS | F_TRACE_SCHED_PORTS)
 #define ERTS_PROC_TRACEE_FLAGS \
-  ((TRACEE_FLAGS & ~ERTS_PORT_TRACEE_FLAGS) | ERTS_TRACEE_MODIFIER_FLAGS)
+    ((TRACEE_FLAGS & ~ERTS_PORT_TRACEE_FLAGS) | ERTS_TRACEE_MODIFIER_FLAGS)
 
 #define SEQ_TRACE_FLAG(N)        (1 << (ERTS_TRACE_TS_TYPE_BITS + (N)))
 
@@ -1717,6 +1718,8 @@ void erts_schedule_thr_prgr_later_cleanup_op(void (*)(void *),
 					     ErtsThrPrgrLaterOp *,
 					     UWord);
 void erts_schedule_complete_off_heap_message_queue_change(Eterm pid);
+void erts_schedule_flush_trace_messages(Eterm pid);
+int erts_flush_trace_messages(Process *c_p, ErtsProcLocks locks);
 
 #if defined(ERTS_SMP) && defined(ERTS_ENABLE_LOCK_CHECK)
 int erts_dbg_check_halloc_lock(Process *p);
@@ -1996,7 +1999,6 @@ erts_psd_set(Process *p, int ix, void *data)
 #define ERTS_PROC_SET_NIF_TRAP_EXPORT(P, NTE) \
     erts_psd_set((P), ERTS_PSD_NIF_TRAP_EXPORT, (void *) (NTE))
 
-
 ERTS_GLB_INLINE Eterm erts_proc_get_error_handler(Process *p);
 ERTS_GLB_INLINE Eterm erts_proc_set_error_handler(Process *p, Eterm handler);
 
@@ -2326,14 +2328,17 @@ erts_alloc_message_heap_state(Process *pp,
 			      ErlOffHeap **ohpp)
 {
     int on_heap;
+    ErtsMessage *mp;
 
     if ((*psp) & ERTS_PSFLG_OFF_HEAP_MSGQ) {
-	ErtsMessage *mp = erts_alloc_message(sz, hpp);
+	mp = erts_alloc_message(sz, hpp);
 	*ohpp = sz == 0 ? NULL : &mp->hfrag.off_heap;
 	return mp;
     }
 
-    return erts_try_alloc_message_on_heap(pp, psp, plp, sz, hpp, ohpp, &on_heap);
+    mp = erts_try_alloc_message_on_heap(pp, psp, plp, sz, hpp, ohpp, &on_heap);
+    ASSERT(pp || !on_heap);
+    return mp;
 }
 
 ERTS_GLB_INLINE ErtsMessage *
@@ -2343,7 +2348,7 @@ erts_alloc_message_heap(Process *pp,
 			Eterm **hpp,
 			ErlOffHeap **ohpp)
 {
-    erts_aint32_t state = erts_smp_atomic32_read_nob(&pp->state);
+    erts_aint32_t state = pp ? erts_smp_atomic32_read_nob(&pp->state) : 0;
     return erts_alloc_message_heap_state(pp, &state, plp, sz, hpp, ohpp);
 }
 
diff --git a/erts/emulator/beam/erl_process_lock.c b/erts/emulator/beam/erl_process_lock.c
index e1f470d381..a69185bc5c 100644
--- a/erts/emulator/beam/erl_process_lock.c
+++ b/erts/emulator/beam/erl_process_lock.c
@@ -106,6 +106,7 @@ static struct {
     Sint16 proc_lock_msgq;
     Sint16 proc_lock_btm;
     Sint16 proc_lock_status;
+    Sint16 proc_lock_trace;
 } lc_id;
 #endif
 
@@ -152,6 +153,7 @@ erts_init_proc_lock(int cpus)
     lc_id.proc_lock_msgq	= erts_lc_get_lock_order_id("proc_msgq");
     lc_id.proc_lock_btm		= erts_lc_get_lock_order_id("proc_btm");
     lc_id.proc_lock_status	= erts_lc_get_lock_order_id("proc_status");
+    lc_id.proc_lock_trace	= erts_lc_get_lock_order_id("proc_trace");
 #endif
 }
 
@@ -1057,6 +1059,11 @@ erts_proc_lock_init(Process *p)
 #ifdef ERTS_ENABLE_LOCK_CHECK
     erts_lc_trylock(1, &p->lock.status.lc);
 #endif
+    erts_mtx_init_x(&p->lock.trace, "proc_trace", p->common.id, do_lock_count);
+    ethr_mutex_lock(&p->lock.trace.mtx);
+#ifdef ERTS_ENABLE_LOCK_CHECK
+    erts_lc_trylock(1, &p->lock.trace.lc);
+#endif
 #endif
 #ifdef ERTS_PROC_LOCK_DEBUG
     for (i = 0; i <= ERTS_PROC_LOCK_MAX_BIT; i++)
@@ -1078,6 +1085,7 @@ erts_proc_lock_fin(Process *p)
     erts_mtx_destroy(&p->lock.msgq);
     erts_mtx_destroy(&p->lock.btm);
     erts_mtx_destroy(&p->lock.status);
+    erts_mtx_destroy(&p->lock.trace);
 #endif
 #if defined(ERTS_ENABLE_LOCK_COUNT) && defined(ERTS_SMP)
     erts_lcnt_proc_lock_destroy(p);
@@ -1100,26 +1108,29 @@ void erts_lcnt_enable_proc_lock_count(int enable) {
 void erts_lcnt_proc_lock_init(Process *p) {
     if (!(erts_lcnt_rt_options & ERTS_LCNT_OPT_PROCLOCK)) {
         erts_lcnt_init_lock_empty(&(p->lock.lcnt_main));
+        erts_lcnt_init_lock_empty(&(p->lock.lcnt_link));
         erts_lcnt_init_lock_empty(&(p->lock.lcnt_msgq));
         erts_lcnt_init_lock_empty(&(p->lock.lcnt_btm));
-        erts_lcnt_init_lock_empty(&(p->lock.lcnt_link));
         erts_lcnt_init_lock_empty(&(p->lock.lcnt_status));
+        erts_lcnt_init_lock_empty(&(p->lock.lcnt_trace));
     } else { /* now the common case */
         Eterm pid = (p->common.id != ERTS_INVALID_PID) ? p->common.id : NIL;
         erts_lcnt_init_lock_x(&(p->lock.lcnt_main),  "proc_main",  ERTS_LCNT_LT_PROCLOCK, pid);
+        erts_lcnt_init_lock_x(&(p->lock.lcnt_link),  "proc_link",  ERTS_LCNT_LT_PROCLOCK, pid);
         erts_lcnt_init_lock_x(&(p->lock.lcnt_msgq),  "proc_msgq",  ERTS_LCNT_LT_PROCLOCK, pid);
         erts_lcnt_init_lock_x(&(p->lock.lcnt_btm),   "proc_btm",   ERTS_LCNT_LT_PROCLOCK, pid);
-        erts_lcnt_init_lock_x(&(p->lock.lcnt_link),  "proc_link",  ERTS_LCNT_LT_PROCLOCK, pid);
         erts_lcnt_init_lock_x(&(p->lock.lcnt_status),"proc_status",ERTS_LCNT_LT_PROCLOCK, pid);
+        erts_lcnt_init_lock_x(&(p->lock.lcnt_trace), "proc_trace", ERTS_LCNT_LT_PROCLOCK, pid);
     } /* the lock names should really be aligned to four characters */
 } /* logic reversed */
 
 void erts_lcnt_proc_lock_destroy(Process *p) {
     erts_lcnt_destroy_lock(&(p->lock.lcnt_main));
+    erts_lcnt_destroy_lock(&(p->lock.lcnt_link));
     erts_lcnt_destroy_lock(&(p->lock.lcnt_msgq));
     erts_lcnt_destroy_lock(&(p->lock.lcnt_btm));
-    erts_lcnt_destroy_lock(&(p->lock.lcnt_link));
     erts_lcnt_destroy_lock(&(p->lock.lcnt_status));
+    erts_lcnt_destroy_lock(&(p->lock.lcnt_trace));
 }
 
 static void lcnt_enable_proc_lock_count(Process *proc, int enable) {
@@ -1138,10 +1149,11 @@ static void lcnt_enable_proc_lock_count(Process *proc, int enable) {
 void erts_lcnt_proc_lock(erts_proc_lock_t *lock, ErtsProcLocks locks) {
     if (!(erts_lcnt_rt_options & ERTS_LCNT_OPT_PROCLOCK)) return;
     if (locks & ERTS_PROC_LOCK_MAIN) { erts_lcnt_lock(&(lock->lcnt_main)); }
+    if (locks & ERTS_PROC_LOCK_LINK) { erts_lcnt_lock(&(lock->lcnt_link)); }
     if (locks & ERTS_PROC_LOCK_MSGQ) { erts_lcnt_lock(&(lock->lcnt_msgq)); }
     if (locks & ERTS_PROC_LOCK_BTM) { erts_lcnt_lock(&(lock->lcnt_btm)); }
-    if (locks & ERTS_PROC_LOCK_LINK) { erts_lcnt_lock(&(lock->lcnt_link)); }
     if (locks & ERTS_PROC_LOCK_STATUS) { erts_lcnt_lock(&(lock->lcnt_status)); }
+    if (locks & ERTS_PROC_LOCK_TRACE) { erts_lcnt_lock(&(lock->lcnt_trace)); }
 }
 
 void erts_lcnt_proc_lock_post_x(erts_proc_lock_t *lock, ErtsProcLocks locks,
@@ -1150,44 +1162,50 @@ void erts_lcnt_proc_lock_post_x(erts_proc_lock_t *lock, ErtsProcLocks locks,
     if (locks & ERTS_PROC_LOCK_MAIN) {
 	erts_lcnt_lock_post_x(&(lock->lcnt_main), file, line);
     }
+    if (locks & ERTS_PROC_LOCK_LINK) {
+	erts_lcnt_lock_post_x(&(lock->lcnt_link), file, line);
+    }
     if (locks & ERTS_PROC_LOCK_MSGQ) {
         erts_lcnt_lock_post_x(&(lock->lcnt_msgq), file, line);
     }
     if (locks & ERTS_PROC_LOCK_BTM) {
         erts_lcnt_lock_post_x(&(lock->lcnt_btm), file, line);
     }
-    if (locks & ERTS_PROC_LOCK_LINK) {
-	erts_lcnt_lock_post_x(&(lock->lcnt_link), file, line);
-    }
     if (locks & ERTS_PROC_LOCK_STATUS) {
 	erts_lcnt_lock_post_x(&(lock->lcnt_status), file, line);
     }
+    if (locks & ERTS_PROC_LOCK_TRACE) {
+        erts_lcnt_lock_post_x(&(lock->lcnt_trace), file, line);
+    }
 }
 
 void erts_lcnt_proc_lock_unaquire(erts_proc_lock_t *lock, ErtsProcLocks locks) {
     if (!(erts_lcnt_rt_options & ERTS_LCNT_OPT_PROCLOCK)) return;
     if (locks & ERTS_PROC_LOCK_MAIN) { erts_lcnt_lock_unaquire(&(lock->lcnt_main)); }
+    if (locks & ERTS_PROC_LOCK_LINK) { erts_lcnt_lock_unaquire(&(lock->lcnt_link)); }
     if (locks & ERTS_PROC_LOCK_MSGQ) { erts_lcnt_lock_unaquire(&(lock->lcnt_msgq)); }
     if (locks & ERTS_PROC_LOCK_BTM) { erts_lcnt_lock_unaquire(&(lock->lcnt_btm)); }
-    if (locks & ERTS_PROC_LOCK_LINK) { erts_lcnt_lock_unaquire(&(lock->lcnt_link)); }
     if (locks & ERTS_PROC_LOCK_STATUS) { erts_lcnt_lock_unaquire(&(lock->lcnt_status)); }
+    if (locks & ERTS_PROC_LOCK_TRACE) { erts_lcnt_lock_unaquire(&(lock->lcnt_trace)); }
 }
 
 void erts_lcnt_proc_unlock(erts_proc_lock_t *lock, ErtsProcLocks locks) {
     if (!(erts_lcnt_rt_options & ERTS_LCNT_OPT_PROCLOCK)) return;
     if (locks & ERTS_PROC_LOCK_MAIN) { erts_lcnt_unlock(&(lock->lcnt_main)); }
+    if (locks & ERTS_PROC_LOCK_LINK) { erts_lcnt_unlock(&(lock->lcnt_link)); }
     if (locks & ERTS_PROC_LOCK_MSGQ) { erts_lcnt_unlock(&(lock->lcnt_msgq)); }
     if (locks & ERTS_PROC_LOCK_BTM) { erts_lcnt_unlock(&(lock->lcnt_btm)); }
-    if (locks & ERTS_PROC_LOCK_LINK) { erts_lcnt_unlock(&(lock->lcnt_link)); }
     if (locks & ERTS_PROC_LOCK_STATUS) { erts_lcnt_unlock(&(lock->lcnt_status)); }
+    if (locks & ERTS_PROC_LOCK_TRACE) { erts_lcnt_unlock(&(lock->lcnt_trace)); }
 }
 void erts_lcnt_proc_trylock(erts_proc_lock_t *lock, ErtsProcLocks locks, int res) {
     if (!(erts_lcnt_rt_options & ERTS_LCNT_OPT_PROCLOCK)) return;
     if (locks & ERTS_PROC_LOCK_MAIN) { erts_lcnt_trylock(&(lock->lcnt_main), res); }
+    if (locks & ERTS_PROC_LOCK_LINK) { erts_lcnt_trylock(&(lock->lcnt_link), res); }
     if (locks & ERTS_PROC_LOCK_MSGQ) { erts_lcnt_trylock(&(lock->lcnt_msgq), res); }
     if (locks & ERTS_PROC_LOCK_BTM) { erts_lcnt_trylock(&(lock->lcnt_btm), res); }
-    if (locks & ERTS_PROC_LOCK_LINK) { erts_lcnt_trylock(&(lock->lcnt_link), res); }
     if (locks & ERTS_PROC_LOCK_STATUS) { erts_lcnt_trylock(&(lock->lcnt_status), res); }
+    if (locks & ERTS_PROC_LOCK_TRACE) { erts_lcnt_trylock(&(lock->lcnt_trace), res); }
 } /* reversed logic */
 #endif /* ERTS_ENABLE_LOCK_COUNT */
 
@@ -1224,6 +1242,10 @@ erts_proc_lc_lock(Process *p, ErtsProcLocks locks, char *file, unsigned int line
 	lck.id = lc_id.proc_lock_status;
 	erts_lc_lock_x(&lck,file,line);
     }
+    if (locks & ERTS_PROC_LOCK_TRACE) {
+	lck.id = lc_id.proc_lock_trace;
+	erts_lc_lock_x(&lck,file,line);
+    }
 }
 
 void
@@ -1253,6 +1275,10 @@ erts_proc_lc_trylock(Process *p, ErtsProcLocks locks, int locked,
 	lck.id = lc_id.proc_lock_status;
 	erts_lc_trylock_x(locked, &lck, file, line);
     }
+    if (locks & ERTS_PROC_LOCK_TRACE) {
+	lck.id = lc_id.proc_lock_trace;
+	erts_lc_trylock_x(locked, &lck, file, line);
+    }
 }
 
 void
@@ -1261,6 +1287,10 @@ erts_proc_lc_unlock(Process *p, ErtsProcLocks locks)
     erts_lc_lock_t lck = ERTS_LC_LOCK_INIT(-1,
 					   p->common.id,
 					   ERTS_LC_FLG_LT_PROCLOCK);
+    if (locks & ERTS_PROC_LOCK_TRACE) {
+	lck.id = lc_id.proc_lock_trace;
+	erts_lc_unlock(&lck);
+    }
     if (locks & ERTS_PROC_LOCK_STATUS) {
 	lck.id = lc_id.proc_lock_status;
 	erts_lc_unlock(&lck);
@@ -1292,6 +1322,10 @@ erts_proc_lc_might_unlock(Process *p, ErtsProcLocks locks)
     erts_lc_lock_t lck = ERTS_LC_LOCK_INIT(-1,
 					   p->common.id,
 					   ERTS_LC_FLG_LT_PROCLOCK);
+    if (locks & ERTS_PROC_LOCK_TRACE) {
+	lck.id = lc_id.proc_lock_trace;
+	erts_lc_might_unlock(&lck);
+    }
     if (locks & ERTS_PROC_LOCK_STATUS) {
 	lck.id = lc_id.proc_lock_status;
 	erts_lc_might_unlock(&lck);
@@ -1323,6 +1357,8 @@ erts_proc_lc_might_unlock(Process *p, ErtsProcLocks locks)
 	erts_lc_might_unlock(&p->lock.btm.lc);
     if (locks & ERTS_PROC_LOCK_STATUS)
 	erts_lc_might_unlock(&p->lock.status.lc);
+    if (locks & ERTS_PROC_LOCK_TRACE)
+	erts_lc_might_unlock(&p->lock.trace.lc);
 #endif
 }
 
@@ -1354,6 +1390,10 @@ erts_proc_lc_require_lock(Process *p, ErtsProcLocks locks, char *file,
 	lck.id = lc_id.proc_lock_status;
 	erts_lc_require_lock(&lck, file, line);
     }
+    if (locks & ERTS_PROC_LOCK_TRACE) {
+	lck.id = lc_id.proc_lock_trace;
+	erts_lc_require_lock(&lck, file, line);
+    }
 #elif ERTS_PROC_LOCK_RAW_MUTEX_IMPL
     if (locks & ERTS_PROC_LOCK_MAIN)
 	erts_lc_require_lock(&p->lock.main.lc, file, line);
@@ -1365,6 +1405,8 @@ erts_proc_lc_require_lock(Process *p, ErtsProcLocks locks, char *file,
 	erts_lc_require_lock(&p->lock.btm.lc, file, line);
     if (locks & ERTS_PROC_LOCK_STATUS)
 	erts_lc_require_lock(&p->lock.status.lc, file, line);
+    if (locks & ERTS_PROC_LOCK_TRACE)
+	erts_lc_require_lock(&p->lock.trace.lc, file, line);
 #endif
 }
 
@@ -1375,6 +1417,10 @@ erts_proc_lc_unrequire_lock(Process *p, ErtsProcLocks locks)
     erts_lc_lock_t lck = ERTS_LC_LOCK_INIT(-1,
 					   p->common.id,
 					   ERTS_LC_FLG_LT_PROCLOCK);
+    if (locks & ERTS_PROC_LOCK_TRACE) {
+	lck.id = lc_id.proc_lock_trace;
+	erts_lc_unrequire_lock(&lck);
+    }
     if (locks & ERTS_PROC_LOCK_STATUS) {
 	lck.id = lc_id.proc_lock_status;
 	erts_lc_unrequire_lock(&lck);
@@ -1406,6 +1452,8 @@ erts_proc_lc_unrequire_lock(Process *p, ErtsProcLocks locks)
 	erts_lc_unrequire_lock(&p->lock.btm.lc);
     if (locks & ERTS_PROC_LOCK_STATUS)
 	erts_lc_unrequire_lock(&p->lock.status.lc);
+    if (locks & ERTS_PROC_LOCK_TRACE)
+	erts_lc_unrequire_lock(&p->lock.trace.lc);
 #endif
 }
 
@@ -1429,6 +1477,8 @@ erts_proc_lc_trylock_force_busy(Process *p, ErtsProcLocks locks)
 	    lck.id = lc_id.proc_lock_btm;
 	else if (locks & ERTS_PROC_LOCK_STATUS)
 	    lck.id = lc_id.proc_lock_status;
+	else if (locks & ERTS_PROC_LOCK_TRACE)
+	    lck.id = lc_id.proc_lock_trace;
 	else
 	    erts_lc_fail("Unknown proc lock found");
 
@@ -1441,14 +1491,7 @@ erts_proc_lc_trylock_force_busy(Process *p, ErtsProcLocks locks)
 
 void erts_proc_lc_chk_only_proc_main(Process *p)
 {
-#if ERTS_PROC_LOCK_OWN_IMPL
-    erts_lc_lock_t proc_main = ERTS_LC_LOCK_INIT(lc_id.proc_lock_main,
-						 p->common.id,
-						 ERTS_LC_FLG_LT_PROCLOCK);
-    erts_lc_check_exact(&proc_main, 1);
-#elif ERTS_PROC_LOCK_RAW_MUTEX_IMPL
-    erts_lc_check_exact(&p->lock.main.lc, 1);
-#endif
+    erts_proc_lc_chk_only_proc(p, ERTS_PROC_LOCK_MAIN);
 }
 
 #if ERTS_PROC_LOCK_OWN_IMPL
@@ -1456,14 +1499,67 @@ void erts_proc_lc_chk_only_proc_main(Process *p)
   ERTS_LC_LOCK_INIT(-1, THE_NON_VALUE, ERTS_LC_FLG_LT_PROCLOCK)
 #endif /* ERTS_PROC_LOCK_OWN_IMPL */
 
+void erts_proc_lc_chk_only_proc(Process *p, ErtsProcLocks locks)
+{
+    int have_locks_len = 0;
+#if ERTS_PROC_LOCK_OWN_IMPL
+    erts_lc_lock_t have_locks[6] = {ERTS_PROC_LC_EMPTY_LOCK_INIT,
+				    ERTS_PROC_LC_EMPTY_LOCK_INIT,
+				    ERTS_PROC_LC_EMPTY_LOCK_INIT,
+				    ERTS_PROC_LC_EMPTY_LOCK_INIT,
+				    ERTS_PROC_LC_EMPTY_LOCK_INIT,
+                                    ERTS_PROC_LC_EMPTY_LOCK_INIT};
+    if (locks & ERTS_PROC_LOCK_MAIN) {
+	have_locks[have_locks_len].id = lc_id.proc_lock_main;
+	have_locks[have_locks_len++].extra = p->common.id;
+    }
+    if (locks & ERTS_PROC_LOCK_LINK) {
+	have_locks[have_locks_len].id = lc_id.proc_lock_link;
+	have_locks[have_locks_len++].extra = p->common.id;
+    }
+    if (locks & ERTS_PROC_LOCK_MSGQ) {
+	have_locks[have_locks_len].id = lc_id.proc_lock_msgq;
+	have_locks[have_locks_len++].extra = p->common.id;
+    }
+    if (locks & ERTS_PROC_LOCK_BTM) {
+	have_locks[have_locks_len].id = lc_id.proc_lock_btm;
+	have_locks[have_locks_len++].extra = p->common.id;
+    }
+    if (locks & ERTS_PROC_LOCK_STATUS) {
+	have_locks[have_locks_len].id = lc_id.proc_lock_status;
+	have_locks[have_locks_len++].extra = p->common.id;
+    }
+    if (locks & ERTS_PROC_LOCK_TRACE) {
+	have_locks[have_locks_len].id = lc_id.proc_lock_trace;
+	have_locks[have_locks_len++].extra = p->common.id;
+    }
+#elif ERTS_PROC_LOCK_RAW_MUTEX_IMPL
+    erts_lc_lock_t have_locks[6];
+    if (locks & ERTS_PROC_LOCK_MAIN)
+	have_locks[have_locks_len++] = p->lock.main.lc;
+    if (locks & ERTS_PROC_LOCK_LINK)
+	have_locks[have_locks_len++] = p->lock.link.lc;
+    if (locks & ERTS_PROC_LOCK_MSGQ)
+	have_locks[have_locks_len++] = p->lock.msgq.lc;
+    if (locks & ERTS_PROC_LOCK_BTM)
+	have_locks[have_locks_len++] = p->lock.btm.lc;
+    if (locks & ERTS_PROC_LOCK_STATUS)
+	have_locks[have_locks_len++] = p->lock.status.lc;
+    if (locks & ERTS_PROC_LOCK_TRACE)
+	have_locks[have_locks_len++] = p->lock.trace.lc;
+#endif
+    erts_lc_check_exact(have_locks, have_locks_len);
+}
+
 void
 erts_proc_lc_chk_have_proc_locks(Process *p, ErtsProcLocks locks)
 {
     int have_locks_len = 0;
 #if ERTS_PROC_LOCK_OWN_IMPL
-    erts_lc_lock_t have_locks[5] = {ERTS_PROC_LC_EMPTY_LOCK_INIT,
+    erts_lc_lock_t have_locks[6] = {ERTS_PROC_LC_EMPTY_LOCK_INIT,
 				    ERTS_PROC_LC_EMPTY_LOCK_INIT,
 				    ERTS_PROC_LC_EMPTY_LOCK_INIT,
+                                    ERTS_PROC_LC_EMPTY_LOCK_INIT,
 				    ERTS_PROC_LC_EMPTY_LOCK_INIT,
 				    ERTS_PROC_LC_EMPTY_LOCK_INIT};
     if (locks & ERTS_PROC_LOCK_MAIN) {
@@ -1486,8 +1582,12 @@ erts_proc_lc_chk_have_proc_locks(Process *p, ErtsProcLocks locks)
 	have_locks[have_locks_len].id = lc_id.proc_lock_status;
 	have_locks[have_locks_len++].extra = p->common.id;
     }
+    if (locks & ERTS_PROC_LOCK_TRACE) {
+	have_locks[have_locks_len].id = lc_id.proc_lock_trace;
+	have_locks[have_locks_len++].extra = p->common.id;
+    }
 #elif ERTS_PROC_LOCK_RAW_MUTEX_IMPL
-    erts_lc_lock_t have_locks[5];
+    erts_lc_lock_t have_locks[6];
     if (locks & ERTS_PROC_LOCK_MAIN)
 	have_locks[have_locks_len++] = p->lock.main.lc;
     if (locks & ERTS_PROC_LOCK_LINK)
@@ -1498,6 +1598,8 @@ erts_proc_lc_chk_have_proc_locks(Process *p, ErtsProcLocks locks)
 	have_locks[have_locks_len++] = p->lock.btm.lc;
     if (locks & ERTS_PROC_LOCK_STATUS)
 	have_locks[have_locks_len++] = p->lock.status.lc;
+    if (locks & ERTS_PROC_LOCK_TRACE)
+	have_locks[have_locks_len++] = p->lock.trace.lc;
 #endif
     erts_lc_check(have_locks, have_locks_len, NULL, 0);
 }
@@ -1508,12 +1610,14 @@ erts_proc_lc_chk_proc_locks(Process *p, ErtsProcLocks locks)
     int have_locks_len = 0;
     int have_not_locks_len = 0;
 #if ERTS_PROC_LOCK_OWN_IMPL
-    erts_lc_lock_t have_locks[5] = {ERTS_PROC_LC_EMPTY_LOCK_INIT,
+    erts_lc_lock_t have_locks[6] = {ERTS_PROC_LC_EMPTY_LOCK_INIT,
 				    ERTS_PROC_LC_EMPTY_LOCK_INIT,
+                                    ERTS_PROC_LC_EMPTY_LOCK_INIT,
 				    ERTS_PROC_LC_EMPTY_LOCK_INIT,
 				    ERTS_PROC_LC_EMPTY_LOCK_INIT};
-    erts_lc_lock_t have_not_locks[5] = {ERTS_PROC_LC_EMPTY_LOCK_INIT,
+    erts_lc_lock_t have_not_locks[6] = {ERTS_PROC_LC_EMPTY_LOCK_INIT,
 					ERTS_PROC_LC_EMPTY_LOCK_INIT,
+                                        ERTS_PROC_LC_EMPTY_LOCK_INIT,
 					ERTS_PROC_LC_EMPTY_LOCK_INIT,
 					ERTS_PROC_LC_EMPTY_LOCK_INIT};
 
@@ -1557,9 +1661,17 @@ erts_proc_lc_chk_proc_locks(Process *p, ErtsProcLocks locks)
 	have_not_locks[have_not_locks_len].id = lc_id.proc_lock_status;
 	have_not_locks[have_not_locks_len++].extra = p->common.id;
     }
+    if (locks & ERTS_PROC_LOCK_TRACE) {
+	have_locks[have_locks_len].id = lc_id.proc_lock_trace;
+	have_locks[have_locks_len++].extra = p->common.id;
+    }
+    else {
+	have_not_locks[have_not_locks_len].id = lc_id.proc_lock_trace;
+	have_not_locks[have_not_locks_len++].extra = p->common.id;
+    }
 #elif ERTS_PROC_LOCK_RAW_MUTEX_IMPL
-    erts_lc_lock_t have_locks[5];
-    erts_lc_lock_t have_not_locks[5];
+    erts_lc_lock_t have_locks[6];
+    erts_lc_lock_t have_not_locks[6];
 
     if (locks & ERTS_PROC_LOCK_MAIN)
 	have_locks[have_locks_len++] = p->lock.main.lc;
@@ -1581,6 +1693,10 @@ erts_proc_lc_chk_proc_locks(Process *p, ErtsProcLocks locks)
 	have_locks[have_locks_len++] = p->lock.status.lc;
     else
 	have_not_locks[have_not_locks_len++] = p->lock.status.lc;
+    if (locks & ERTS_PROC_LOCK_TRACE)
+	have_locks[have_locks_len++] = p->lock.trace.lc;
+    else
+	have_not_locks[have_not_locks_len++] = p->lock.trace.lc;
 #endif
 
     erts_lc_check(have_locks, have_locks_len,
@@ -1590,10 +1706,10 @@ erts_proc_lc_chk_proc_locks(Process *p, ErtsProcLocks locks)
 ErtsProcLocks
 erts_proc_lc_my_proc_locks(Process *p)
 {
-    int resv[5];
+    int resv[6];
     ErtsProcLocks res = 0;
 #if ERTS_PROC_LOCK_OWN_IMPL
-    erts_lc_lock_t locks[5] = {ERTS_LC_LOCK_INIT(lc_id.proc_lock_main,
+    erts_lc_lock_t locks[6] = {ERTS_LC_LOCK_INIT(lc_id.proc_lock_main,
 						 p->common.id,
 						 ERTS_LC_FLG_LT_PROCLOCK),
 			       ERTS_LC_LOCK_INIT(lc_id.proc_lock_link,
@@ -1607,16 +1723,20 @@ erts_proc_lc_my_proc_locks(Process *p)
 						 ERTS_LC_FLG_LT_PROCLOCK),
 			       ERTS_LC_LOCK_INIT(lc_id.proc_lock_status,
 						 p->common.id,
+						 ERTS_LC_FLG_LT_PROCLOCK),
+			       ERTS_LC_LOCK_INIT(lc_id.proc_lock_trace,
+						 p->common.id,
 						 ERTS_LC_FLG_LT_PROCLOCK)};
 #elif ERTS_PROC_LOCK_RAW_MUTEX_IMPL
-    erts_lc_lock_t locks[5] = {p->lock.main.lc,
+    erts_lc_lock_t locks[6] = {p->lock.main.lc,
 			       p->lock.link.lc,
 			       p->lock.msgq.lc,
 			       p->lock.btm.lc,
-			       p->lock.status.lc};
+			       p->lock.status.lc,
+			       p->lock.trace.lc};
 #endif
 
-    erts_lc_have_locks(resv, locks, 5);
+    erts_lc_have_locks(resv, locks, 6);
     if (resv[0])
 	res |= ERTS_PROC_LOCK_MAIN;
     if (resv[1])
@@ -1627,6 +1747,8 @@ erts_proc_lc_my_proc_locks(Process *p)
 	res |= ERTS_PROC_LOCK_BTM;
     if (resv[4])
 	res |= ERTS_PROC_LOCK_STATUS;
+    if (resv[5])
+	res |= ERTS_PROC_LOCK_TRACE;
 
     return res;
 }
@@ -1634,14 +1756,15 @@ erts_proc_lc_my_proc_locks(Process *p)
 void
 erts_proc_lc_chk_no_proc_locks(char *file, int line)
 {
-    int resv[5];
-    int ids[5] = {lc_id.proc_lock_main,
+    int resv[6];
+    int ids[6] = {lc_id.proc_lock_main,
 		  lc_id.proc_lock_link,
 		  lc_id.proc_lock_msgq,
 		  lc_id.proc_lock_btm,
-		  lc_id.proc_lock_status};
-    erts_lc_have_lock_ids(resv, ids, 5);
-    if (!ERTS_IS_CRASH_DUMPING && (resv[0] || resv[1] || resv[2] || resv[3] || resv[4])) {
+		  lc_id.proc_lock_status,
+		  lc_id.proc_lock_trace};
+    erts_lc_have_lock_ids(resv, ids, 6);
+    if (!ERTS_IS_CRASH_DUMPING && (resv[0] || resv[1] || resv[2] || resv[3] || resv[4] || resv[5])) {
 	erts_lc_fail("%s:%d: Thread has process locks locked when expected "
 		     "not to have any process locks locked",
 		     file, line);
diff --git a/erts/emulator/beam/erl_process_lock.h b/erts/emulator/beam/erl_process_lock.h
index ac324c6f3e..2cccf0697a 100644
--- a/erts/emulator/beam/erl_process_lock.h
+++ b/erts/emulator/beam/erl_process_lock.h
@@ -66,7 +66,7 @@
 
 #endif
 
-#define ERTS_PROC_LOCK_MAX_BIT 4
+#define ERTS_PROC_LOCK_MAX_BIT 5
 
 typedef erts_aint32_t ErtsProcLocks;
 
@@ -84,6 +84,7 @@ typedef struct erts_proc_lock_t_ {
     erts_lcnt_lock_t lcnt_msgq;
     erts_lcnt_lock_t lcnt_btm;
     erts_lcnt_lock_t lcnt_status;
+    erts_lcnt_lock_t lcnt_trace;
 #endif
 #elif ERTS_PROC_LOCK_RAW_MUTEX_IMPL
     erts_mtx_t main;
@@ -91,6 +92,7 @@ typedef struct erts_proc_lock_t_ {
     erts_mtx_t msgq;
     erts_mtx_t btm;
     erts_mtx_t status;
+    erts_mtx_t trace;
 #else
 #  error "no implementation"
 #endif
@@ -137,9 +139,18 @@ typedef struct erts_proc_lock_t_ {
  *   Protects the following fields in the process structure:
  *   * pending_suspenders
  *   * suspendee
+ *   * sys_tasks
  *   * ...
  */
-#define ERTS_PROC_LOCK_STATUS		(((ErtsProcLocks) 1) << ERTS_PROC_LOCK_MAX_BIT)
+#define ERTS_PROC_LOCK_STATUS		(((ErtsProcLocks) 1) << 4)
+
+/*
+ * Trace message lock:
+ *   Protects the order in which messages are sent
+ *   from trace nifs. This lock is taken inside enif_send.
+ *
+ */
+#define ERTS_PROC_LOCK_TRACE            (((ErtsProcLocks) 1) << ERTS_PROC_LOCK_MAX_BIT)
 
 /*
  * Special fields:
@@ -154,8 +165,8 @@ typedef struct erts_proc_lock_t_ {
  *   all process locks are held, and are allowed to be read if
  *   at least one process lock (whichever one doesn't matter)
  *   is held:
- *     * tracer_proc
- *     * tracer_flags
+ *     * common.tracer
+ *     * common.trace_flags
  *
  *   The following fields are only allowed to be accessed if
  *   both the schedule queue lock and at least one process lock
@@ -198,17 +209,15 @@ typedef struct erts_proc_lock_t_ {
 
 /* ERTS_PROC_LOCKS_* are combinations of process locks */
 
-#define ERTS_PROC_LOCKS_MSG_RECEIVE	(ERTS_PROC_LOCK_MSGQ		\
-					 | ERTS_PROC_LOCK_STATUS)
-#define ERTS_PROC_LOCKS_MSG_SEND	(ERTS_PROC_LOCK_MSGQ		\
-					 | ERTS_PROC_LOCK_STATUS)
+#define ERTS_PROC_LOCKS_MSG_RECEIVE	ERTS_PROC_LOCK_MSGQ
+#define ERTS_PROC_LOCKS_MSG_SEND	ERTS_PROC_LOCK_MSGQ
 #define ERTS_PROC_LOCKS_XSIG_SEND	ERTS_PROC_LOCK_STATUS
 
 #define ERTS_PROC_LOCKS_ALL \
   ((((ErtsProcLocks) 1) << (ERTS_PROC_LOCK_MAX_BIT + 1)) - 1)
 
 #define ERTS_PROC_LOCKS_ALL_MINOR	(ERTS_PROC_LOCKS_ALL \
-					 & ~ERTS_PROC_LOCK_MAIN)
+                                         & ~ERTS_PROC_LOCK_MAIN)
 
 
 #define ERTS_PIX_LOCKS_BITS		10
@@ -260,6 +269,7 @@ void erts_proc_lc_might_unlock(Process *p, ErtsProcLocks locks);
 void erts_proc_lc_chk_have_proc_locks(Process *p, ErtsProcLocks locks);
 void erts_proc_lc_chk_proc_locks(Process *p, ErtsProcLocks locks);
 void erts_proc_lc_chk_only_proc_main(Process *p);
+void erts_proc_lc_chk_only_proc(Process *p, ErtsProcLocks locks);
 void erts_proc_lc_chk_no_proc_locks(char *file, int line);
 ErtsProcLocks erts_proc_lc_my_proc_locks(Process *p);
 int erts_proc_lc_trylock_force_busy(Process *p, ErtsProcLocks locks);
@@ -477,9 +487,15 @@ erts_smp_proc_raw_trylock__(Process *p, ErtsProcLocks locks)
     if (locks & ERTS_PROC_LOCK_STATUS)
 	if (erts_mtx_trylock(&p->lock.status) == EBUSY)
 	    goto busy_status;
+    if (locks & ERTS_PROC_LOCK_TRACE)
+	if (erts_mtx_trylock(&p->lock.trace) == EBUSY)
+	    goto busy_trace;
 
     return 0;
 
+busy_trace:
+    if (locks & ERTS_PROC_LOCK_TRACE)
+	erts_mtx_unlock(&p->lock.trace);
 busy_status:
     if (locks & ERTS_PROC_LOCK_BTM)
 	erts_mtx_unlock(&p->lock.btm);
@@ -568,6 +584,8 @@ erts_smp_proc_lock__(Process *p,
 	erts_mtx_lock(&p->lock.btm);
     if (locks & ERTS_PROC_LOCK_STATUS)
 	erts_mtx_lock(&p->lock.status);
+    if (locks & ERTS_PROC_LOCK_TRACE)
+	erts_mtx_lock(&p->lock.trace);
 
 #ifdef ERTS_PROC_LOCK_DEBUG
     erts_proc_lock_op_debug(p, locks, 1);
@@ -653,6 +671,8 @@ erts_smp_proc_unlock__(Process *p,
     erts_proc_lock_op_debug(p, locks, 0);
 #endif
 
+    if (locks & ERTS_PROC_LOCK_TRACE)
+	erts_mtx_unlock(&p->lock.trace);
     if (locks & ERTS_PROC_LOCK_STATUS)
 	erts_mtx_unlock(&p->lock.status);
     if (locks & ERTS_PROC_LOCK_BTM)
diff --git a/erts/emulator/beam/erl_ptab.h b/erts/emulator/beam/erl_ptab.h
index 97df8bc3fc..a5931ffc25 100644
--- a/erts/emulator/beam/erl_ptab.h
+++ b/erts/emulator/beam/erl_ptab.h
@@ -37,14 +37,16 @@
 #include "erl_alloc.h"
 #include "erl_monitors.h"
 
-#define ERTS_TRACER_PROC(P) 	((P)->common.tracer_proc)
+#define ERTS_TRACER(P)          ((P)->common.tracer)
+#define ERTS_TRACER_MODULE(T) 	(CAR(list_val(T)))
+#define ERTS_TRACER_STATE(T) 	(CDR(list_val(T)))
 #define ERTS_TRACE_FLAGS(P)	((P)->common.trace_flags)
 
 #define ERTS_P_LINKS(P)		((P)->common.u.alive.links)
 #define ERTS_P_MONITORS(P)	((P)->common.u.alive.monitors)
 
 #define IS_TRACED(p) \
-    (ERTS_TRACER_PROC((p)) != NIL)
+    (ERTS_TRACER(p) != NIL)
 #define ARE_TRACE_FLAGS_ON(p,tf) \
     ((ERTS_TRACE_FLAGS((p)) & (tf|F_SENSITIVE)) == (tf))
 #define IS_TRACED_FL(p,tf) \
@@ -56,7 +58,7 @@ typedef struct {
 	erts_atomic_t atmc;
 	Sint sint;
     } refc;
-    Eterm tracer_proc;
+    ErtsTracer tracer;
     Uint trace_flags;
     erts_smp_atomic_t timer;
     union {
diff --git a/erts/emulator/beam/erl_time_sup.c b/erts/emulator/beam/erl_time_sup.c
index 947def7398..346404fe2a 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, NIL);
+		    erts_queue_message(rp, &rp_locks, mp, message);
 		}
 		erts_smp_proc_unlock(rp, rp_locks);
 	    }
diff --git a/erts/emulator/beam/erl_trace.c b/erts/emulator/beam/erl_trace.c
index 1654ea58d9..bd88769dfc 100644
--- a/erts/emulator/beam/erl_trace.c
+++ b/erts/emulator/beam/erl_trace.c
@@ -20,6 +20,19 @@
 
 /*
  * Support functions for tracing.
+ *
+ * Ideas for future speed improvements in tracing framework:
+ *  * Move ErtsTracerNif into ErtsTracer
+ *     + Removes need for locking
+ *     + Removes hash lookup overhead
+ *     + Use a refc on the ErtsTracerNif to know when it can
+ *       be freed. We don't want to allocate a separate
+ *       ErtsTracerNif for each module used.
+ *  * Optimize GenericBp for cache locality by reusing equivalent
+ *    GenericBp and GenericBpData in multiple tracer points.
+ *     + Possibly we want to use specialized instructions for different
+ *       types of trace so that the knowledge of which struct is used
+ *       can be in the instruction.
  */
 
 #ifdef HAVE_CONFIG_H
@@ -39,6 +52,7 @@
 #include "erl_bits.h"
 #include "erl_thr_progress.h"
 #include "erl_bif_unique.h"
+#include "erl_map.h"
 
 #if 0
 #define DEBUG_PRINTOUTS
@@ -46,17 +60,15 @@
 #undef DEBUG_PRINTOUTS
 #endif
 
-extern BeamInstr beam_return_to_trace[1];   /* OpCode(i_return_to_trace) */
-extern BeamInstr beam_return_trace[1];      /* OpCode(i_return_trace) */
-extern BeamInstr beam_return_time_trace[1]; /* OpCode(i_return_time_trace) */
-
 /* Pseudo export entries. Never filled in with data, only used to
    yield unique pointers of the correct type. */
 Export exp_send, exp_receive, exp_timeout;
 
-static Eterm system_seq_tracer;
-static Uint default_trace_flags;
-static Eterm default_tracer;
+static ErtsTracer system_seq_tracer;
+static Uint default_proc_trace_flags;
+static ErtsTracer default_proc_tracer;
+static Uint default_port_trace_flags;
+static ErtsTracer default_port_tracer;
 
 static Eterm system_monitor;
 static Eterm system_profile;
@@ -70,8 +82,6 @@ static erts_smp_rwmtx_t sys_trace_rwmtx;
 
 enum ErtsSysMsgType {
     SYS_MSG_TYPE_UNDEFINED,
-    SYS_MSG_TYPE_TRACE,
-    SYS_MSG_TYPE_SEQTRACE,
     SYS_MSG_TYPE_SYSMON,
     SYS_MSG_TYPE_ERRLGR,
     SYS_MSG_TYPE_PROC_MSG,
@@ -298,43 +308,6 @@ write_ts(int ts_type, Eterm *hp, ErlHeapFragment *bp, Process *tracer)
     return res;
 }
 
-/*
- * Patch a timestamp into a tuple. The tuple MUST be the last thing
- * built on the heap before the call, and the timestamp MUST be
- * the last thing after the call. This since patch_ts() might adjust
- * the size of the used area.
- */
-
-#define PATCH_TS__(Type, Tuple, Hp, Bp, Tracer)			\
-    do {							\
-	int ts_type__ = (Type);					\
-	if (ts_type__)						\
-	    patch_ts(ts_type__, (Tuple), (Hp), (Bp), (Tracer));	\
-    } while (0)
-
-#ifdef ERTS_SMP
-#define PATCH_TS(Type, Tuple, Hp, Bp, Tracer) \
-    PATCH_TS__((Type), (Tuple), (Hp), (Bp), NULL)
-#else
-#define PATCH_TS(Type, Tuple, Hp, Bp, Tracer) \
-    PATCH_TS__((Type), (Tuple), (Hp), (Bp), (Tracer))
-#endif
-
-static ERTS_INLINE void
-patch_ts(int ts_type, Eterm tuple, Eterm* hp, ErlHeapFragment *bp, Process *tracer)
-{
-    Eterm *tptr = tuple_val(tuple);
-    int arity = arityval(*tptr);
-
-    ASSERT(ts_type);
-    ASSERT((tptr+arity+1) == hp);
-
-    tptr[0] = make_arityval(arity+1);
-    tptr[1] = am_trace_ts;
-
-    *hp = write_ts(ts_type, hp+1, bp, tracer);
-}
-
 #ifdef ERTS_SMP
 static void enqueue_sys_msg_unlocked(enum ErtsSysMsgType type,
 				     Eterm from,
@@ -349,6 +322,14 @@ static void enqueue_sys_msg(enum ErtsSysMsgType type,
 static void init_sys_msg_dispatcher(void);
 #endif
 
+static void init_tracer_nif(void);
+static int tracer_cmp_fun(void*, void*);
+static HashValue tracer_hash_fun(void*);
+static void *tracer_alloc_fun(void*);
+static void tracer_free_fun(void*);
+
+typedef struct ErtsTracerNif_ ErtsTracerNif;
+
 void erts_init_trace(void) {
     erts_smp_rwmtx_opt_t rwmtx_opts = ERTS_SMP_RWMTX_OPT_DEFAULT_INITER;
     rwmtx_opts.type = ERTS_SMP_RWMTX_TYPE_EXTREMELY_FREQUENT_READ;
@@ -362,80 +343,41 @@ void erts_init_trace(void) {
     erts_bif_trace_init();
     erts_system_monitor_clear(NULL);
     erts_system_profile_clear(NULL);
-    default_trace_flags = F_INITIAL_TRACE_FLAGS;
-    default_tracer = NIL;
-    system_seq_tracer = am_false;
+    default_proc_trace_flags = F_INITIAL_TRACE_FLAGS;
+    default_proc_tracer = erts_tracer_nil;
+    default_port_trace_flags = F_INITIAL_TRACE_FLAGS;
+    default_port_tracer = erts_tracer_nil;
+    system_seq_tracer = erts_tracer_nil;
 #ifdef ERTS_SMP
     init_sys_msg_dispatcher();
 #endif
+    init_tracer_nif();
 }
 
-static Eterm system_seq_tracer;
-
 #define ERTS_ALLOC_SYSMSG_HEAP(SZ, BPP, OHPP, UNUSED) \
   (*(BPP) = new_message_buffer((SZ)), \
    *(OHPP) = &(*(BPP))->off_heap, \
    (*(BPP))->mem)
 
-#ifdef ERTS_SMP
-#define ERTS_ENQ_TRACE_MSG(FPID, TPID, MSG, BP) \
-do { \
-    ERTS_LC_ASSERT(erts_smp_lc_mtx_is_locked(&smq_mtx)); \
-    enqueue_sys_msg_unlocked(SYS_MSG_TYPE_TRACE, (FPID), (TPID), (MSG), (BP)); \
-} while(0)
-#else
-#define ERTS_ENQ_TRACE_MSG(FPID, TPROC, MSG, BP)		\
-    do {							\
-	ErtsMessage *mp__ = erts_alloc_message(0, NULL);		\
-	mp__->data.heap_frag = (BP);				\
-	erts_queue_message((TPROC), NULL, mp__, (MSG), NIL);	\
-    } while (0)
-#endif
-
-/*
- * NOTE that the ERTS_GET_TRACER_REF() returns from the function  (!!!)
- * using it, and resets the parameters used if the tracer is invalid, i.e.,
- * use it with extreme care!
- */
-#ifdef ERTS_SMP
-#define ERTS_NULL_TRACER_REF NIL
-#define ERTS_TRACER_REF_TYPE Eterm
- /* In the smp case, we never find the tracer invalid here (the sys
-    message dispatcher thread takes care of that). */
-#define ERTS_GET_TRACER_REF(RES, TPID, TRACEE_FLGS) \
-do { (RES) = (TPID); } while(0)
-int
-erts_is_tracer_proc_valid(Process* p)
-{
-    return 1;
-}
-#else
-#define ERTS_NULL_TRACER_REF NULL
-#define ERTS_TRACER_REF_TYPE Process *
-#define ERTS_GET_TRACER_REF(RES, TPID, TRACEE_FLGS) \
-do { \
-    (RES) = erts_proc_lookup((TPID)); \
-    if (!(RES) || !(ERTS_TRACE_FLAGS((RES)) & F_TRACER)) { \
-	(TPID) = NIL; \
-	(TRACEE_FLGS) &= ~TRACEE_FLAGS; \
-	return; \
-    } \
-} while (0)
-int
-erts_is_tracer_proc_valid(Process* p)
-{
-    Process* tracer;
+static ERTS_INLINE int
+send_to_tracer_nif_raw(Process *c_p, Process *tracee, const ErtsTracer tracer,
+                       Uint trace_flags, Eterm t_p_id, ErtsTracerNif *tnif,
+                       Eterm tag, Eterm msg, Eterm extra, Eterm pam_result);
+static ERTS_INLINE int
+send_to_tracer_nif(Process *c_p, ErtsPTabElementCommon *t_p,
+                   Eterm t_p_id, ErtsTracerNif *tnif, Eterm tag,
+                   Eterm msg, Eterm extra);
+static ERTS_INLINE Eterm
+call_enabled_tracer(Process *c_p, const ErtsTracer tracer,
+                    ErtsTracerNif **tnif_ref, Eterm tag, Eterm t_p_id);
+static int
+is_tracer_proc_enabled(Process* c_p, ErtsProcLocks c_p_locks,
+                       ErtsPTabElementCommon *t_p,
+                       ErtsTracerNif **tnif_ret, Eterm tag);
 
-    tracer = erts_proc_lookup(ERTS_TRACER_PROC(p));
-    if (tracer && ERTS_TRACE_FLAGS(tracer) & F_TRACER) {
-	return 1;
-    } else {
-	ERTS_TRACER_PROC(p) = NIL;
-	ERTS_TRACE_FLAGS(p) = ~TRACEE_FLAGS;
-	return 0;
-    }
-}
-#endif
+#define SEND_TO_TRACER(c_p, tag, msg)                                   \
+    send_to_tracer_nif(c_p, &(c_p)->common, (c_p)->common.id, NULL, tag, \
+                       msg, THE_NON_VALUE)
 
 static Uint active_sched;
 
@@ -450,19 +392,6 @@ static void
 exiting_reset(Eterm exiting)
 {
     erts_smp_rwmtx_rwlock(&sys_trace_rwmtx);
-    if (exiting == default_tracer) {
-	default_tracer = NIL;
-	default_trace_flags &= TRACEE_FLAGS;
-#ifdef DEBUG
-	default_trace_flags |= F_INITIAL_TRACE_FLAGS;
-#endif
-    }
-    if (exiting == system_seq_tracer) {
-#ifdef DEBUG_PRINTOUTS
-	erts_fprintf(stderr, "seq tracer %T exited\n", exiting);
-#endif
-	system_seq_tracer = am_false;
-    }
     if (exiting == system_monitor) {
 #ifdef ERTS_SMP
 	system_monitor = NIL;
@@ -487,11 +416,7 @@ erts_trace_check_exiting(Eterm exiting)
 {
     int reset = 0;
     erts_smp_rwmtx_rlock(&sys_trace_rwmtx);
-    if (exiting == default_tracer)
-	reset = 1;
-    else if (exiting == system_seq_tracer)
-	reset = 1;
-    else if (exiting == system_monitor)
+    if (exiting == system_monitor)
 	reset = 1;
     else if (exiting == system_profile)
 	reset = 1;
@@ -500,23 +425,26 @@ erts_trace_check_exiting(Eterm exiting)
 	exiting_reset(exiting);
 }
 
-static ERTS_INLINE int
-is_valid_tracer(Eterm tracer)
+ErtsTracer
+erts_set_system_seq_tracer(Process *c_p, ErtsProcLocks c_p_locks, ErtsTracer new)
 {
-    return erts_proc_lookup(tracer) || erts_is_valid_tracer_port(tracer);
-}
-
-Eterm
-erts_set_system_seq_tracer(Process *c_p, ErtsProcLocks c_p_locks, Eterm new)
-{
-    Eterm old;
-
-    if (new != am_false && !is_valid_tracer(new))
-	return THE_NON_VALUE;
+    ErtsTracer old;
+
+    if (!ERTS_TRACER_IS_NIL(new)) {
+        Eterm nif_result = call_enabled_tracer(
+            NULL, new, NULL,
+            am_trace_status, am_undefined);
+        switch (nif_result) {
+        case am_trace: break;
+        default:
+            return THE_NON_VALUE;
+        }
+    }
 
     erts_smp_rwmtx_rwlock(&sys_trace_rwmtx);
     old = system_seq_tracer;
-    system_seq_tracer = new;
+    system_seq_tracer = erts_tracer_nil;
+    erts_tracer_update(&system_seq_tracer, new);
 
 #ifdef DEBUG_PRINTOUTS
     erts_fprintf(stderr, "set seq tracer new=%T old=%T\n", new, old);
@@ -525,66 +453,132 @@ erts_set_system_seq_tracer(Process *c_p, ErtsProcLocks c_p_locks, Eterm new)
     return old;
 }
 
-Eterm
+ErtsTracer
 erts_get_system_seq_tracer(void)
 {
-    Eterm st;
+    ErtsTracer st;
     erts_smp_rwmtx_rlock(&sys_trace_rwmtx);
     st = system_seq_tracer;
 #ifdef DEBUG_PRINTOUTS
     erts_fprintf(stderr, "get seq tracer %T\n", st);
 #endif
     erts_smp_rwmtx_runlock(&sys_trace_rwmtx);
+
+    if (st != erts_tracer_nil &&
+        call_enabled_tracer(NULL, st, NULL, am_trace_status, am_undefined) == am_remove) {
+        erts_set_system_seq_tracer(NULL, 0, erts_tracer_nil);
+        st = erts_tracer_nil;
+    }
+
     return st;
 }
 
 static ERTS_INLINE void
-get_default_tracing(Uint *flagsp, Eterm *tracerp)
-{
-    if (!(default_trace_flags & TRACEE_FLAGS))
-	default_tracer = NIL;
-
-    if (is_nil(default_tracer)) {
-	default_trace_flags &= ~TRACEE_FLAGS;
-    } else if (is_internal_pid(default_tracer)) {
-	if (!erts_proc_lookup(default_tracer)) {
-	reset_tracer:
-	    default_trace_flags &= ~TRACEE_FLAGS;
-	    default_tracer = NIL;
-	}
+get_default_tracing(Uint *flagsp, ErtsTracer *tracerp,
+                    Uint *default_trace_flags,
+                    ErtsTracer *default_tracer)
+{
+    if (!(*default_trace_flags & TRACEE_FLAGS))
+	ERTS_TRACER_CLEAR(default_tracer);
+
+    if (ERTS_TRACER_IS_NIL(*default_tracer)) {
+	*default_trace_flags &= ~TRACEE_FLAGS;
     } else {
-	ASSERT(is_internal_port(default_tracer));
-	if (!erts_is_valid_tracer_port(default_tracer))
-	    goto reset_tracer;
+        Eterm nif_result = call_enabled_tracer(
+            NULL, *default_tracer, NULL,
+            am_trace_status, am_undefined);
+        switch (nif_result) {
+        case am_trace: break;
+        default: {
+            ErtsTracer curr_default_tracer = *default_tracer;
+            if (tracerp) {
+                /* we only have a rlock, so we have to unlock and then rwlock */
+                erts_smp_rwmtx_runlock(&sys_trace_rwmtx);
+                erts_smp_rwmtx_rwlock(&sys_trace_rwmtx);
+            }
+            /* check if someone else changed default tracer
+               while we got the write lock, if so we don't do
+               anything. */
+            if (curr_default_tracer == *default_tracer) {
+                *default_trace_flags &= ~TRACEE_FLAGS;
+                ERTS_TRACER_CLEAR(default_tracer);
+            }
+            if (tracerp) {
+                erts_smp_rwmtx_rwunlock(&sys_trace_rwmtx);
+                erts_smp_rwmtx_rlock(&sys_trace_rwmtx);
+            }
+        }
+        }
     }
 
     if (flagsp)
-	*flagsp = default_trace_flags;
-    if (tracerp)
-	*tracerp = default_tracer;
+	*flagsp = *default_trace_flags;
+    if (tracerp) {
+	erts_tracer_update(tracerp,*default_tracer);
+    }
+}
+
+static ERTS_INLINE void
+erts_change_default_tracing(int setflags, Uint flags,
+                            const ErtsTracer tracer,
+                            Uint *default_trace_flags,
+                            ErtsTracer *default_tracer)
+{
+    if (setflags)
+        *default_trace_flags |= flags;
+    else
+        *default_trace_flags &= ~flags;
+
+    erts_tracer_update(default_tracer, tracer);
+
+    get_default_tracing(NULL, NULL, default_trace_flags, default_tracer);
 }
 
 void
-erts_change_default_tracing(int setflags, Uint *flagsp, Eterm *tracerp)
+erts_change_default_proc_tracing(int setflags, Uint flagsp,
+                                 const ErtsTracer tracer)
 {
     erts_smp_rwmtx_rwlock(&sys_trace_rwmtx);
-    if (flagsp) {
-	if (setflags)
-	    default_trace_flags |= *flagsp;
-	else
-	    default_trace_flags &= ~(*flagsp);
-    }
-    if (tracerp)
-	default_tracer = *tracerp;
-    get_default_tracing(flagsp, tracerp);
+    erts_change_default_tracing(
+        setflags, flagsp, tracer,
+        &default_proc_trace_flags,
+        &default_proc_tracer);
     erts_smp_rwmtx_rwunlock(&sys_trace_rwmtx);
 }
 
 void
-erts_get_default_tracing(Uint *flagsp, Eterm *tracerp)
+erts_change_default_port_tracing(int setflags, Uint flagsp,
+                                 const ErtsTracer tracer)
+{
+    erts_smp_rwmtx_rwlock(&sys_trace_rwmtx);
+    erts_change_default_tracing(
+        setflags, flagsp, tracer,
+        &default_port_trace_flags,
+        &default_port_tracer);
+    erts_smp_rwmtx_rwunlock(&sys_trace_rwmtx);
+}
+
+void
+erts_get_default_proc_tracing(Uint *flagsp, ErtsTracer *tracerp)
 {
     erts_smp_rwmtx_rlock(&sys_trace_rwmtx);
-    get_default_tracing(flagsp, tracerp);
+    *tracerp = erts_tracer_nil; /* initialize */
+    get_default_tracing(
+        flagsp, tracerp,
+        &default_proc_trace_flags,
+        &default_proc_tracer);
+    erts_smp_rwmtx_runlock(&sys_trace_rwmtx);
+}
+
+void
+erts_get_default_port_tracing(Uint *flagsp, ErtsTracer *tracerp)
+{
+    erts_smp_rwmtx_rlock(&sys_trace_rwmtx);
+    *tracerp = erts_tracer_nil; /* initialize */
+    get_default_tracing(
+        flagsp, tracerp,
+        &default_port_trace_flags,
+        &default_port_tracer);
     erts_smp_rwmtx_runlock(&sys_trace_rwmtx);
 }
 
@@ -622,29 +616,21 @@ erts_get_system_profile(void) {
     return profile;
 }
 
-#ifdef ERTS_SMP
-static void
-do_send_to_port(Eterm to,
-		Port* unused_port,
-		Eterm from,
-		enum ErtsSysMsgType type,
-		Eterm message)
-{
-    Uint sz = size_object(message);
-    ErlHeapFragment *bp = new_message_buffer(sz);
-    Uint *hp = bp->mem;
-    Eterm msg = copy_struct(message, sz, &hp, &bp->off_heap);
-
-    enqueue_sys_msg_unlocked(type, from, to, msg, bp);
-}
 
-#define WRITE_SYS_MSG_TO_PORT write_sys_msg_to_port
+#ifdef HAVE_ERTS_NOW_CPU
+#  define GET_NOW(m, s, u) \
+do { \
+    if (erts_cpu_timestamp) \
+	erts_get_now_cpu(m, s, u); \
+    else \
+	get_now(m, s, u); \
+} while (0)
 #else
-#define WRITE_SYS_MSG_TO_PORT do_send_to_port
+#  define GET_NOW(m, s, u) do {get_now(m, s, u);} while (0)
 #endif
 
 static void
-WRITE_SYS_MSG_TO_PORT(Eterm unused_to,
+write_sys_msg_to_port(Eterm unused_to,
 		      Port* trace_port,
 		      Eterm unused_from,
 		      enum ErtsSysMsgType unused_type,
@@ -671,150 +657,6 @@ WRITE_SYS_MSG_TO_PORT(Eterm unused_to,
     erts_free(ERTS_ALC_T_TMP, (void *) buffer);
 }
 
-
-#ifndef ERTS_SMP
-/* Send        {trace_ts, Pid, out, 0, Timestamp}
- * followed by {trace_ts, Pid, in, 0, NewTimestamp}
- *
- * 'NewTimestamp' through patch_ts().
- */
-static void 
-do_send_schedfix_to_port(Port *trace_port, Eterm pid, Eterm timestamp, int ts_type) {
-#define LOCAL_HEAP_SIZE (5+5+ERTS_TRACE_PATCH_TS_MAX_SIZE)
-    DeclareTmpHeapNoproc(local_heap,LOCAL_HEAP_SIZE);
-    Eterm message;
-    Eterm *hp;
-    Eterm mfarity;
-
-    UseTmpHeapNoproc(LOCAL_HEAP_SIZE);
-    ASSERT(is_pid(pid));
-    ASSERT(is_tuple(timestamp));
-    ASSERT(*tuple_val(timestamp) == make_arityval(3));
-    
-    hp = local_heap;
-    mfarity = make_small(0);
-    message = TUPLE5(hp, am_trace_ts, pid, am_out, mfarity, timestamp);
-    /* Note, hp is deliberately NOT incremented since it will be reused */
-
-    do_send_to_port(trace_port->common.id,
-		    trace_port,
-		    pid,
-		    SYS_MSG_TYPE_UNDEFINED,
-		    message);
-
-
-    message = TUPLE5(hp, am_trace_ts, pid, am_in, mfarity,
-		     NIL /* Will be overwritten by timestamp */);
-    hp += 6;
-    hp[-1] = write_ts(ts_type, hp, NULL, NULL);
-
-    do_send_to_port(trace_port->common.id,
-		    trace_port,
-		    pid,
-		    SYS_MSG_TYPE_UNDEFINED,
-		    message);
-    UnUseTmpHeapNoproc(LOCAL_HEAP_SIZE);
-#undef LOCAL_HEAP_SIZE
-}
-#endif
-
-/* If (c_p != NULL), a fake schedule out/in message pair will be sent,
- * if the driver so requests. 
- * It is assumed that 'message' is not an 'out' message.
- *
- * 'c_p' is the currently executing process, "tracee" is the traced process
- * which 'message' concerns => if (*tracee_flags & F_TIMESTAMP_MASK), 
- * 'message' must contain a timestamp.
- */
-static void
-send_to_port(Process *c_p, Eterm message, 
-	     Eterm *tracer_pid, Uint *tracee_flags) {
-    Port* trace_port;
-#ifndef ERTS_SMP
-    int ts_type;
-#define LOCAL_HEAP_SIZE ERTS_TRACE_PATCH_TS_MAX_SIZE
-    Eterm ts;
-    DeclareTmpHeapNoproc(local_heap,LOCAL_HEAP_SIZE);
-#endif
-
-    ASSERT(is_internal_port(*tracer_pid));
-#ifdef ERTS_SMP
-    if (is_not_internal_port(*tracer_pid))
-	return;
-
-    trace_port = NULL;
-#else
-
-    trace_port = erts_id2port_sflgs(*tracer_pid,
-				    NULL,
-				    0,
-				    ERTS_PORT_SFLGS_INVALID_TRACER_LOOKUP);
-
-    if (!trace_port) {
-	*tracee_flags &= ~TRACEE_FLAGS;
-	*tracer_pid = NIL;
-	return;
-    }
-
-    /*
-     * Make a fake schedule only if the current process is traced
-     * with 'running' and 'timestamp'.
-     */
-
-    if ( c_p == NULL || 
-	 (! IS_TRACED_FL(c_p, F_TRACE_SCHED | F_TIMESTAMP_MASK))) { 
-#endif
-	do_send_to_port(*tracer_pid,
-			trace_port,
-			c_p ? c_p->common.id : NIL,
-			SYS_MSG_TYPE_TRACE,
-			message);
-#ifndef ERTS_SMP
-	erts_port_release(trace_port);
-	return;
-    }
-
-    /*
-     * Note that the process being traced for some type of trace messages
-     * (e.g. getting_linked) need not be the current process. That other
-     * process might not have timestamps enabled.
-     */
-    UseTmpHeapNoproc(LOCAL_HEAP_SIZE);
-
-    /* A fake schedule might be needed.
-     * Create a dummy trace message with timestamp to be
-     * passed to do_send_schedfix_to_port().
-     */
-    ts_type = TFLGS_TS_TYPE(c_p);
-    ts = write_ts(ts_type, local_heap, NULL, NULL);
-
-    trace_port->control_flags &= ~PORT_CONTROL_FLAG_HEAVY;
-    do_send_to_port(*tracer_pid,
-		    trace_port,
-		    c_p ? c_p->common.id : NIL,
-		    SYS_MSG_TYPE_TRACE,
-		    message);
-
-    if (trace_port->control_flags & PORT_CONTROL_FLAG_HEAVY) {
-	/* The driver has just informed us that the last write took a 
-	 * non-neglectible amount of time.
-	 *
-	 * We need to fake some trace messages to compensate for the time the
-	 * current process had to sacrifice for the writing of the previous
-	 * trace message. We pretend that the process got scheduled out
-	 * just after writning the real trace message, and now gets scheduled
-	 * in again.
-	 */
-	do_send_schedfix_to_port(trace_port, c_p->common.id, ts, ts_type);
-    }
-
-    erts_port_release(trace_port);
-
-    UnUseTmpHeapNoproc(LOCAL_HEAP_SIZE);
-#undef LOCAL_HEAP_SIZE
-#endif
-}
-
 #ifndef ERTS_SMP
 /* Profile send
  * Checks if profiler is port or process
@@ -843,11 +685,11 @@ profile_send(Eterm from, Eterm message) {
 					   0,
 					   ERTS_PORT_SFLGS_INVALID_TRACER_LOOKUP);
 	if (profiler_port) {
-	    do_send_to_port(profiler,
-			    profiler_port,
-			    NIL, /* or current process->common.id */
-			    SYS_MSG_TYPE_SYSPROF,
-			    message);
+	    write_sys_msg_to_port(profiler,
+                                  profiler_port,
+                                  NIL, /* or current process->common.id */
+                                  SYS_MSG_TYPE_SYSPROF,
+                                  message);
 	    erts_port_release(profiler_port);
 	}
     	
@@ -867,154 +709,26 @@ 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, NIL);
+        erts_queue_message(profile_p, NULL, mp, msg);
     }
 }
 
 #endif
 
-
-/* A fake schedule out/in message pair will be sent,
- * if the driver so requests.
- *
- * 'c_p' is the currently executing process, may be NULL.
- */
 static void
-seq_trace_send_to_port(Process *c_p,
-		       Eterm seq_tracer,
-		       Eterm message)
+trace_sched_aux(Process *p, ErtsProcLocks locks, Eterm what)
 {
-    Port* trace_port;
-#ifndef ERTS_SMP
-    int ts_type;
-    Eterm ts;
-#define LOCAL_HEAP_SIZE ERTS_TRACE_PATCH_TS_MAX_SIZE
-    DeclareTmpHeapNoproc(local_heap,LOCAL_HEAP_SIZE);
-    UseTmpHeapNoproc(LOCAL_HEAP_SIZE);
-#endif
-
-    ASSERT(is_internal_port(seq_tracer));
-#ifdef ERTS_SMP
-    if (is_not_internal_port(seq_tracer))
-	return;
-
-    trace_port = NULL;
-#else
-    trace_port = erts_id2port_sflgs(seq_tracer,
-				    NULL,
-				    0,
-				    ERTS_PORT_SFLGS_INVALID_TRACER_LOOKUP);
-    if (!trace_port) {
-	system_seq_tracer = am_false;
-#ifndef ERTS_SMP
-	UnUseTmpHeapNoproc(LOCAL_HEAP_SIZE);
-#endif
-	return;
-    }
-
-    if (c_p == NULL
-	|| (! IS_TRACED_FL(c_p, F_TRACE_SCHED | F_TIMESTAMP_MASK))) {
-#endif
-	do_send_to_port(seq_tracer,
-			trace_port,
-			c_p ? c_p->common.id : NIL,
-			SYS_MSG_TYPE_SEQTRACE,
-			message);
-
-#ifndef ERTS_SMP
-	erts_port_release(trace_port);
-	UnUseTmpHeapNoproc(LOCAL_HEAP_SIZE);
-	return;
-    }
-    /* Make a fake schedule only if the current process is traced
-     * with 'running' and 'timestamp'.
-     */
-
-    /* A fake schedule might be needed.
-     * Create a dummy trace message with timestamp to be
-     * passed to do_send_schedfix_to_port().
-     */
-    ts_type = TFLGS_TS_TYPE(c_p);
-    ts = write_ts(ts_type, local_heap, NULL, NULL);
-
-    trace_port->control_flags &= ~PORT_CONTROL_FLAG_HEAVY;
-    do_send_to_port(seq_tracer,
-		    trace_port,
-		    c_p ? c_p->common.id : NIL,
-		    SYS_MSG_TYPE_SEQTRACE,
-		    message);
-
-    if (trace_port->control_flags & PORT_CONTROL_FLAG_HEAVY) {
-	/* The driver has just informed us that the last write took a 
-	 * non-neglectible amount of time.
-	 *
-	 * We need to fake some trace messages to compensate for the time the
-	 * current process had to sacrifice for the writing of the previous
-	 * trace message. We pretend that the process got scheduled out
-	 * just after writing the real trace message, and now gets scheduled
-	 * in again.
-	 */
-	do_send_schedfix_to_port(trace_port, c_p->common.id, ts, ts_type);
-    }
-
-    erts_port_release(trace_port);
-
-    UnUseTmpHeapNoproc(LOCAL_HEAP_SIZE);
-#undef LOCAL_HEAP_SIZE
-#endif
-}
-
-static ERTS_INLINE void
-send_to_tracer(Process *tracee,
-	       ERTS_TRACER_REF_TYPE tracer_ref,
-	       Eterm msg,
-	       Eterm **hpp,
-	       ErlHeapFragment *bp,
-	       int no_fake_sched)
-{
-    ERTS_SMP_LC_ASSERT(erts_proc_lc_my_proc_locks(tracee));
-
-    erts_smp_mtx_lock(&smq_mtx);
-
-    if (is_internal_pid(ERTS_TRACER_PROC(tracee))) {
-	PATCH_TS(TFLGS_TS_TYPE(tracee), msg, *hpp, bp, tracer_ref);
-	ERTS_ENQ_TRACE_MSG(tracee->common.id, tracer_ref, msg, bp);
-    }
-    else {
-	ASSERT(is_internal_port(ERTS_TRACER_PROC(tracee)));
-	PATCH_TS(TFLGS_TS_TYPE(tracee), msg, *hpp, NULL, NULL);
-	send_to_port(no_fake_sched ? NULL : tracee,
-		     msg,
-		     &ERTS_TRACER_PROC(tracee),
-		     &ERTS_TRACE_FLAGS(tracee));
-    }
-
-    erts_smp_mtx_unlock(&smq_mtx);
-
-}
-
-static void
-trace_sched_aux(Process *p, Eterm what, int never_fake_sched)
-{
-#define LOCAL_HEAP_SIZE (5+4+1+ERTS_TRACE_PATCH_TS_MAX_SIZE)
-    DeclareTmpHeap(local_heap,LOCAL_HEAP_SIZE,p);
-    Eterm tmp, mess, *hp;
-    ErlHeapFragment *bp = NULL;
-    ErlOffHeap *off_heap;
-    ERTS_TRACER_REF_TYPE tracer_ref = ERTS_NULL_TRACER_REF;
-    int sched_no, curr_func, to_port, no_fake_sched;
+    Eterm tmp, *hp;
+    int curr_func;
+    ErtsTracerNif *tnif = NULL;
 
-    if (is_nil(ERTS_TRACER_PROC(p)))
+    if (ERTS_TRACER_IS_NIL(ERTS_TRACER(p)))
 	return;
 
-    no_fake_sched = never_fake_sched;
-
     switch (what) {
     case am_out:
     case am_out_exiting:
     case am_out_exited:
-	no_fake_sched = 1;
-	break;
     case am_in:
     case am_in_exiting:
 	break;
@@ -1023,16 +737,8 @@ trace_sched_aux(Process *p, Eterm what, int never_fake_sched)
 	break;
     }
 
-    sched_no = IS_TRACED_FL(p, F_TRACE_SCHED_NO);
-    to_port = is_internal_port(ERTS_TRACER_PROC(p));
-
-    if (!to_port) {
-	ASSERT(is_internal_pid(ERTS_TRACER_PROC(p)));
-
-	ERTS_GET_TRACER_REF(tracer_ref,
-			    ERTS_TRACER_PROC(p),
-			    ERTS_TRACE_FLAGS(p));
-    }
+    if (!is_tracer_proc_enabled(p, locks, &p->common, &tnif, what))
+        return;
 
     if (ERTS_PROC_IS_EXITING(p))
 	curr_func = 0;
@@ -1042,44 +748,16 @@ trace_sched_aux(Process *p, Eterm what, int never_fake_sched)
 	curr_func = p->current != NULL;
     }
 
-    UseTmpHeap(LOCAL_HEAP_SIZE,p);
-
-    if (to_port)
-	hp = local_heap;
-    else {
-	Uint size = 5;
-	if (curr_func)
-	    size += 4;
-	if (sched_no)
-	    size += 1;
-	size += PATCH_TS_SIZE(p);
-	hp = ERTS_ALLOC_SYSMSG_HEAP(size, &bp, &off_heap, tracer_ref);
-    }
-
     if (!curr_func) {
 	tmp = make_small(0);
     } else {
+        hp = HAlloc(p, 4);
 	tmp = TUPLE3(hp,p->current[0],p->current[1],make_small(p->current[2]));
 	hp += 4;
     }
 
-    if (!sched_no) {
-	mess = TUPLE4(hp, am_trace, p->common.id, what, tmp);
-	hp += 5;
-    }
-    else {
-#ifdef ERTS_SMP
-	Eterm sched_id = make_small(p->scheduler_data->no);
-#else
-	Eterm sched_id = make_small(1);
-#endif
-	mess = TUPLE5(hp, am_trace, p->common.id, what, sched_id, tmp);
-	hp += 6;
-    }
-
-    send_to_tracer(p, tracer_ref, mess, &hp, bp, no_fake_sched);
-    UnUseTmpHeap(LOCAL_HEAP_SIZE,p);
-#undef LOCAL_HEAP_SIZE
+    send_to_tracer_nif(p, &p->common, p->common.id, tnif,
+                       what, tmp, THE_NON_VALUE);
 }
 
 /* Send {trace_ts, Pid, What, {Mod, Func, Arity}, Timestamp}
@@ -1089,9 +767,9 @@ trace_sched_aux(Process *p, Eterm what, int never_fake_sched)
  * 'out_exiting', or 'out_exited'.
  */
 void
-trace_sched(Process *p, Eterm what)
+trace_sched(Process *p, ErtsProcLocks locks, Eterm what)
 {
-    trace_sched_aux(p, what, 0);
+    trace_sched_aux(p, locks, what);
 }
 
 /* Send {trace_ts, Pid, Send, Msg, DestPid, Timestamp}
@@ -1102,140 +780,51 @@ trace_sched(Process *p, Eterm what)
 void
 trace_send(Process *p, Eterm to, Eterm msg)
 {
-    Eterm operation;
-    unsigned sz_msg;
-    unsigned sz_to;
-    Eterm* hp;
-    Eterm mess;
-    
-    if (!ARE_TRACE_FLAGS_ON(p, F_TRACE_SEND)) {
-	return;
-    }
+    Eterm operation = am_send;
+    ErtsTracerNif *tnif = NULL;
+
+    ASSERT(ARE_TRACE_FLAGS_ON(p, F_TRACE_SEND));
 
-    operation = am_send;
     if (is_internal_pid(to)) {
 	if (!erts_proc_lookup(to))
 	    goto send_to_non_existing_process;
     }
     else if(is_external_pid(to)
 	    && external_pid_dist_entry(to) == erts_this_dist_entry) {
-	char *s;
     send_to_non_existing_process:
-	s = "send_to_non_existing_process";
-	operation = am_atom_put(s, sys_strlen(s));
+	operation = am_send_to_non_existing_process;
     }
 
-    if (is_internal_port(ERTS_TRACER_PROC(p))) {
-#define LOCAL_HEAP_SIZE (6 + ERTS_TRACE_PATCH_TS_MAX_SIZE)
-	DeclareTmpHeapNoproc(local_heap,LOCAL_HEAP_SIZE);
-	UseTmpHeapNoproc(LOCAL_HEAP_SIZE);
-
-	hp = local_heap;
-	mess = TUPLE5(hp, am_trace, p->common.id, operation, msg, to);
-	hp += 6;
-	erts_smp_mtx_lock(&smq_mtx);
-	PATCH_TS(TFLGS_TS_TYPE(p), mess, hp, NULL, NULL);
-	send_to_port(p, mess, &ERTS_TRACER_PROC(p), &ERTS_TRACE_FLAGS(p));
-	UnUseTmpHeapNoproc(LOCAL_HEAP_SIZE);
-#undef LOCAL_HEAP_SIZE
-	erts_smp_mtx_unlock(&smq_mtx);
-    } else {
-	Uint need;
-	ErlHeapFragment *bp;
-	ErlOffHeap *off_heap;
-	ERTS_TRACER_REF_TYPE tracer_ref;
-
-	ASSERT(is_internal_pid(ERTS_TRACER_PROC(p)));
-
-	ERTS_GET_TRACER_REF(tracer_ref,
-			    ERTS_TRACER_PROC(p),
-			    ERTS_TRACE_FLAGS(p));
-
-	sz_msg = size_object(msg);
-	sz_to  = size_object(to);
-	need = sz_msg + sz_to + 6 + PATCH_TS_SIZE(p);
-	
-	hp = ERTS_ALLOC_SYSMSG_HEAP(need, &bp, &off_heap, tracer_ref);
-
-	to = copy_struct(to,
-			 sz_to,
-			 &hp,
-			 off_heap);
-	msg = copy_struct(msg,
-			  sz_msg,
-			  &hp,
-			  off_heap);
-	mess = TUPLE5(hp, am_trace, p->common.id, operation, msg, to);
-	hp += 6;
-
-	erts_smp_mtx_lock(&smq_mtx);
-
-	PATCH_TS(TFLGS_TS_TYPE(p), mess, hp, bp, tracer_ref);
-	ERTS_ENQ_TRACE_MSG(p->common.id, tracer_ref, mess, bp);
-	erts_smp_mtx_unlock(&smq_mtx);
-    }
+    if (is_tracer_proc_enabled(p, ERTS_PROC_LOCK_MAIN, &p->common, &tnif, operation))
+        send_to_tracer_nif(p, &p->common, p->common.id, tnif, operation, msg, to);
 }
 
 /* Send {trace_ts, Pid, receive, Msg, Timestamp}
  * or   {trace, Pid, receive, Msg}
  */
 void
-trace_receive(Process *rp, Eterm msg)
+trace_receive(Process *c_p, Eterm msg)
 {
-    Eterm mess;
-    size_t sz_msg;
-    Eterm* hp;
-
-    if (is_internal_port(ERTS_TRACER_PROC(rp))) {
-#define LOCAL_HEAP_SIZE (5+ERTS_TRACE_PATCH_TS_MAX_SIZE)
-	DeclareTmpHeapNoproc(local_heap,LOCAL_HEAP_SIZE);
-	UseTmpHeapNoproc(LOCAL_HEAP_SIZE);
-
-	hp = local_heap;
-	mess = TUPLE4(hp, am_trace, rp->common.id, am_receive, msg);
-	hp += 5;
-	erts_smp_mtx_lock(&smq_mtx);
-	PATCH_TS(TFLGS_TS_TYPE(rp), mess, hp, NULL, NULL);
-	send_to_port(rp, mess, &ERTS_TRACER_PROC(rp), &ERTS_TRACE_FLAGS(rp));
-	UnUseTmpHeapNoproc(LOCAL_HEAP_SIZE);
-#undef LOCAL_HEAP_SIZE
-	erts_smp_mtx_unlock(&smq_mtx);
-    } else {
-	Uint hsz;
-	ErlHeapFragment *bp;
-	ErlOffHeap *off_heap;
-	ERTS_TRACER_REF_TYPE tracer_ref;
-
-	ASSERT(is_internal_pid(ERTS_TRACER_PROC(rp)));
-
-	ERTS_GET_TRACER_REF(tracer_ref,
-			    ERTS_TRACER_PROC(rp),
-			    ERTS_TRACE_FLAGS(rp));
-
-	sz_msg = size_object(msg);
-
-	hsz = sz_msg + 5 + PATCH_TS_SIZE(rp);
-
-	hp = ERTS_ALLOC_SYSMSG_HEAP(hsz, &bp, &off_heap, tracer_ref);
-
-	msg = copy_struct(msg, sz_msg, &hp, off_heap);
-	mess = TUPLE4(hp, am_trace, rp->common.id, am_receive, msg);
-	hp += 5;
-
-	erts_smp_mtx_lock(&smq_mtx);
-
-	PATCH_TS(TFLGS_TS_TYPE(rp), mess, hp, bp, tracer_ref);
-	ERTS_ENQ_TRACE_MSG(rp->common.id, tracer_ref, mess, bp);
-	erts_smp_mtx_unlock(&smq_mtx);
-    }
+    ErtsTracerNif *tnif = NULL;
+    if (is_tracer_proc_enabled(NULL, 0, &c_p->common,
+                               &tnif, am_receive))
+        send_to_tracer_nif(NULL, &c_p->common, c_p->common.id,
+                           tnif, am_receive, msg, THE_NON_VALUE);
 }
 
 int
 seq_trace_update_send(Process *p)
 {
-    Eterm seq_tracer = erts_get_system_seq_tracer();
+    ErtsTracer seq_tracer = erts_get_system_seq_tracer();
     ASSERT((is_tuple(SEQ_TRACE_TOKEN(p)) || is_nil(SEQ_TRACE_TOKEN(p))));
-    if ((p->common.id == seq_tracer) || have_no_seqtrace(SEQ_TRACE_TOKEN(p))) {
+    if (have_no_seqtrace(SEQ_TRACE_TOKEN(p)) ||
+        (seq_tracer != NIL &&
+         call_enabled_tracer( NULL, seq_tracer, NULL, am_trace_status,
+                              p ? p->common.id : am_undefined) != am_trace)
+#ifdef USE_VM_PROBES
+	 || (SEQ_TRACE_TOKEN(p) == am_have_dt_utag)
+#endif
+	 ) {
 	return 0;
     }
     SEQ_TRACE_TOKEN_SENDER(p) = p->common.id;
@@ -1263,20 +852,28 @@ seq_trace_output_generic(Eterm token, Eterm msg, Uint type,
 			 Eterm receiver, Process *process, Eterm exitfrom)
 {
     Eterm mess;
-    ErlHeapFragment* bp;
     Eterm* hp;
     Eterm label;
     Eterm lastcnt_serial;
     Eterm type_atom;
-    int sz_exit;
-    Eterm seq_tracer;
-    int ts_type;
+    ErtsTracer seq_tracer;
+    int seq_tracer_flags = 0;
+#define LOCAL_HEAP_SIZE (64)
+    DeclareTmpHeapNoproc(local_heap,LOCAL_HEAP_SIZE);
 
     seq_tracer = erts_get_system_seq_tracer();
 
     ASSERT(is_tuple(token) || is_nil(token));
-    if (SEQ_TRACE_T_SENDER(token) == seq_tracer || token == NIL ||
-	(process && ERTS_TRACE_FLAGS(process) & F_SENSITIVE)) {
+    if (token == NIL || (process && ERTS_TRACE_FLAGS(process) & F_SENSITIVE) ||
+        ERTS_TRACER_IS_NIL(seq_tracer) ||
+        call_enabled_tracer(
+            NULL, seq_tracer, NULL, am_trace_status,
+            process ? process->common.id : am_undefined) != am_trace) {
+	return;
+    }
+
+    if ((unsigned_val(SEQ_TRACE_T_FLAGS(token)) & type) == 0) {
+	/* No flags set, nothing to do */
 	return;
     }
 
@@ -1289,151 +886,29 @@ seq_trace_output_generic(Eterm token, Eterm msg, Uint type,
 	return;			/* To avoid warning */
     }
 
-    if ((unsigned_val(SEQ_TRACE_T_FLAGS(token)) & type) == 0) {
-	/* No flags set, nothing to do */
-	return;
-    }
+    UseTmpHeapNoproc(LOCAL_HEAP_SIZE);
 
-    if (seq_tracer == am_false) {
-	return;			/* no need to send anything */
+    hp = local_heap;
+    label = SEQ_TRACE_T_LABEL(token);
+    lastcnt_serial = TUPLE2(hp, SEQ_TRACE_T_LASTCNT(token),
+                            SEQ_TRACE_T_SERIAL(token));
+    hp += 3;
+    if (exitfrom != NIL) {
+        msg = TUPLE3(hp, am_EXIT, exitfrom, msg);
+        hp += 4;
     }
+    mess = TUPLE5(hp, type_atom, lastcnt_serial, SEQ_TRACE_T_SENDER(token),
+                  receiver, msg);
+    hp += 6;
 
-    ts_type = ERTS_SEQTFLGS2TSTYPE(unsigned_val(SEQ_TRACE_T_FLAGS(token)));
+    seq_tracer_flags |=  ERTS_SEQTFLGS2TFLGS(unsigned_val(SEQ_TRACE_T_FLAGS(token)));
 
-    if (is_internal_port(seq_tracer)) {
-#define LOCAL_HEAP_SIZE (60 + ERTS_TRACE_PATCH_TS_MAX_SIZE)
-	DeclareTmpHeapNoproc(local_heap,LOCAL_HEAP_SIZE);
-	UseTmpHeapNoproc(LOCAL_HEAP_SIZE);
+    send_to_tracer_nif_raw(NULL, process, seq_tracer, seq_tracer_flags,
+                           label, NULL, am_seq_trace, mess,
+                           THE_NON_VALUE, am_true);
 
-	hp = local_heap;
-	label = SEQ_TRACE_T_LABEL(token);
-	lastcnt_serial = TUPLE2(hp, SEQ_TRACE_T_LASTCNT(token),
-				SEQ_TRACE_T_SERIAL(token));
-	hp += 3;
-	if (exitfrom != NIL) {
-	    msg = TUPLE3(hp, am_EXIT, exitfrom, msg);
-	    hp += 4;
-	}
-	mess = TUPLE5(hp, type_atom, lastcnt_serial, SEQ_TRACE_T_SENDER(token),
-		      receiver, msg);
-	hp += 6;
-
-	erts_smp_mtx_lock(&smq_mtx);
-	if (!ts_type) {
-	    mess = TUPLE3(hp, am_seq_trace, label, mess);
-	    seq_trace_send_to_port(NULL, seq_tracer, mess);
-	} else {
-	    mess = TUPLE4(hp, am_seq_trace, label, mess,
-			  NIL /* Will be overwritten by timestamp */);
-	    hp += 5;
-	    hp[-1] = write_ts(ts_type, hp, NULL, NULL);
-	    seq_trace_send_to_port(process, seq_tracer, mess);
-	}
-	UnUseTmpHeapNoproc(LOCAL_HEAP_SIZE);
+    UnUseTmpHeapNoproc(LOCAL_HEAP_SIZE);
 #undef LOCAL_HEAP_SIZE
-	erts_smp_mtx_unlock(&smq_mtx);
-    } else {
-#ifndef ERTS_SMP
-	Process* tracer;
-#endif
-	Eterm sender_copy;
-	Eterm receiver_copy;
-	Eterm m2;
-	Uint sz_label, sz_lastcnt_serial, sz_msg, sz_ts, sz_sender,
-	    sz_exitfrom, sz_receiver;
-
-	ASSERT(is_internal_pid(seq_tracer));
-
-#ifndef ERTS_SMP
-
-	tracer = erts_proc_lookup(seq_tracer);
-	if (!tracer) {
-	    system_seq_tracer = am_false;
-	    return; /* no need to send anything */
-	}
-#endif
-	if (receiver == seq_tracer) {
-	    return; /* no need to send anything */
-	}
-
-	sz_label = size_object(SEQ_TRACE_T_LABEL(token));
-	sz_sender = size_object(SEQ_TRACE_T_SENDER(token));
-	sz_receiver = size_object(receiver);
-	sz_lastcnt_serial = 3; /* TUPLE2 */
-	sz_msg = size_object(msg);
-
-	sz_ts = patch_ts_size(ts_type);
-	if (exitfrom != NIL) {
-	    sz_exit = 4; /* create {'EXIT',exitfrom,msg} */
-	    sz_exitfrom = size_object(exitfrom);
-	}
-	else {
-	    sz_exit = 0;
-	    sz_exitfrom = 0;
-	}
-	bp = new_message_buffer(4 /* TUPLE3 */ + sz_ts + 6 /* TUPLE5 */ 
-				+ sz_lastcnt_serial + sz_label + sz_msg
-				+ sz_exit + sz_exitfrom
-				+ sz_sender + sz_receiver);
-	hp = bp->mem;
-	label = copy_struct(SEQ_TRACE_T_LABEL(token), sz_label, &hp, &bp->off_heap);
-	lastcnt_serial = TUPLE2(hp,SEQ_TRACE_T_LASTCNT(token),SEQ_TRACE_T_SERIAL(token));
-	hp += 3;
-	m2 = copy_struct(msg, sz_msg, &hp, &bp->off_heap);
-	if (sz_exit) {
-	    Eterm exitfrom_copy = copy_struct(exitfrom,
-					      sz_exitfrom,
-					      &hp,
-					      &bp->off_heap);
-	    m2 = TUPLE3(hp, am_EXIT, exitfrom_copy, m2);
-	    hp += 4;
-	}
-	sender_copy = copy_struct(SEQ_TRACE_T_SENDER(token),
-				  sz_sender,
-				  &hp,
-				  &bp->off_heap);
-	receiver_copy = copy_struct(receiver,
-				    sz_receiver,
-				    &hp,
-				    &bp->off_heap);
-	mess = TUPLE5(hp,
-		      type_atom,
-		      lastcnt_serial,
-		      sender_copy,
-		      receiver_copy,
-		      m2);
-	hp += 6;
-
-	erts_smp_mtx_lock(&smq_mtx);
-
-	if (!ts_type)
-	    mess = TUPLE3(hp, am_seq_trace, label, mess);
-	else {
-	    mess = TUPLE4(hp, am_seq_trace, label, mess,
-			  NIL /* Will be overwritten by timestamp */);
-	    hp += 5;
-	    /* Write timestamp in element 6 of the 'msg' tuple */
-	    hp[-1] = write_ts(ts_type, hp, bp,
-#ifndef ERTS_SMP
-			      tracer
-#else
-			      NULL
-#endif
-		);
-	}
-
-#ifdef ERTS_SMP
-	enqueue_sys_msg_unlocked(SYS_MSG_TYPE_SEQTRACE, NIL, NIL, mess, bp);
-	erts_smp_mtx_unlock(&smq_mtx);
-#else
-        /* trace_token must be NIL here */
-	{
-	    ErtsMessage *mp = erts_alloc_message(0, NULL);
-	    mp->data.heap_frag = bp;
-	    erts_queue_message(tracer, NULL, mp, mess, NIL);
-	}
-#endif
-    }
 }
 
 /* Send {trace_ts, Pid, return_to, {Mod, Func, Arity}, Timestamp}
@@ -1442,63 +917,19 @@ seq_trace_output_generic(Eterm token, Eterm msg, Uint type,
 void 
 erts_trace_return_to(Process *p, BeamInstr *pc)
 {
-#define LOCAL_HEAP_SIZE (4+5+ERTS_TRACE_PATCH_TS_MAX_SIZE)
-    Eterm* hp;
     Eterm mfa;
-    Eterm mess;
-    DeclareTmpHeapNoproc(local_heap,LOCAL_HEAP_SIZE);
 
     BeamInstr *code_ptr = find_function_from_pc(pc);
 
 
-    UseTmpHeapNoproc(LOCAL_HEAP_SIZE);
-
-    hp = local_heap;
-
     if (!code_ptr) {
 	mfa = am_undefined;
     } else {
+        Eterm *hp = HAlloc(p, 4);
 	mfa = TUPLE3(hp, code_ptr[0], code_ptr[1], make_small(code_ptr[2]));
-	hp += 4;
     }
-	
-    mess = TUPLE4(hp, am_trace, p->common.id, am_return_to, mfa);
-    hp += 5;
-
-    erts_smp_mtx_lock(&smq_mtx);
-
-    PATCH_TS(TFLGS_TS_TYPE(p), mess, hp, NULL, NULL);
-
-    if (is_internal_port(ERTS_TRACER_PROC(p))) {
-	send_to_port(p, mess, &ERTS_TRACER_PROC(p), &ERTS_TRACE_FLAGS(p));
-    } else {
-	ErlHeapFragment *bp;
-	ErlOffHeap *off_heap;
-	ERTS_TRACER_REF_TYPE tracer_ref;
-	unsigned size;
 
-	/*
-	 * Find the tracer.
-	 */
-	ASSERT(is_internal_pid(ERTS_TRACER_PROC(p)));
-
-	ERTS_GET_TRACER_REF(tracer_ref,
-			    ERTS_TRACER_PROC(p),
-			    ERTS_TRACE_FLAGS(p));
-
-	size = size_object(mess);
-
-	hp = ERTS_ALLOC_SYSMSG_HEAP(size, &bp, &off_heap, tracer_ref);
-	
-	/*
-	 * Copy the trace message into the buffer and enqueue it.
-	 */
-	mess = copy_struct(mess, size, &hp, off_heap);
-	ERTS_ENQ_TRACE_MSG(p->common.id, tracer_ref, mess, bp);
-    }
-    UnUseTmpHeapNoproc(LOCAL_HEAP_SIZE);
-#undef LOCAL_HEAP_SIZE
-    erts_smp_mtx_unlock(&smq_mtx);
+    SEND_TO_TRACER(p, am_return_to, mfa);
 }
 
 
@@ -1506,114 +937,53 @@ erts_trace_return_to(Process *p, BeamInstr *pc)
  * or   {trace, Pid, return_from, {Mod, Name, Arity}, Retval}
  */
 void
-erts_trace_return(Process* p, BeamInstr* fi, Eterm retval, Eterm *tracer_pid)
+erts_trace_return(Process* p, BeamInstr* fi, Eterm retval, ErtsTracer *tracer)
 {
     Eterm* hp;
-    Eterm mfa;
-    Eterm mess;
-    Eterm mod, name;
+    Eterm mfa, mod, name;
     int arity;
     Uint meta_flags, *tracee_flags;
-    int ts_type;
-#ifdef ERTS_SMP
-    Eterm tracee;
-#endif
-    
-    ASSERT(tracer_pid);
-    if (*tracer_pid == am_true) {
+
+    ASSERT(tracer);
+    if (ERTS_TRACER_COMPARE(*tracer, erts_tracer_true)) {
 	/* Breakpoint trace enabled without specifying tracer =>
 	 *   use process tracer and flags
 	 */
-	tracer_pid = &ERTS_TRACER_PROC(p);
+	tracer = &ERTS_TRACER(p);
     }
-    if (is_nil(*tracer_pid)) {
+    if (ERTS_TRACER_IS_NIL(*tracer)) {
 	/* Trace disabled */
 	return;
     }
-    ASSERT(is_internal_pid(*tracer_pid) || is_internal_port(*tracer_pid));
-    if (*tracer_pid == p->common.id) {
-	/* Do not generate trace messages to oneself */
-	return;
-    }
-    if (tracer_pid == &ERTS_TRACER_PROC(p)) {
+    ASSERT(IS_TRACER_VALID(*tracer));
+    if (tracer == &ERTS_TRACER(p)) {
 	/* Tracer specified in process structure =>
 	 *   non-breakpoint trace =>
 	 *     use process flags
 	 */
 	tracee_flags = &ERTS_TRACE_FLAGS(p);
-#ifdef ERTS_SMP
-	tracee = p->common.id;
-#endif
+        if (! (*tracee_flags & F_TRACE_CALLS)) {
+            return;
+        }
     } else {
 	/* Tracer not specified in process structure =>
 	 *   tracer specified in breakpoint =>
 	 *     meta trace =>
 	 *       use fixed flag set instead of process flags
-	 */	    
+	 */
 	meta_flags = F_TRACE_CALLS | F_NOW_TS;
 	tracee_flags = &meta_flags;
-#ifdef ERTS_SMP
-	tracee = NIL;
-#endif
-    }
-    if (! (*tracee_flags & F_TRACE_CALLS)) {
-	return;
     }
-    
+
     mod = fi[0];
     name = fi[1];
     arity = fi[2];
-    
-    ts_type = ERTS_TFLGS2TSTYPE(*tracee_flags);
-
-    if (is_internal_port(*tracer_pid)) {
-#define LOCAL_HEAP_SIZE (4+6+ERTS_TRACE_PATCH_TS_MAX_SIZE)
-	DeclareTmpHeapNoproc(local_heap,LOCAL_HEAP_SIZE);
-	UseTmpHeapNoproc(LOCAL_HEAP_SIZE);
-	hp = local_heap;
-	mfa = TUPLE3(hp, mod, name, make_small(arity));
-	hp += 4;
-	mess = TUPLE5(hp, am_trace, p->common.id, am_return_from, mfa, retval);
-	hp += 6;
-	erts_smp_mtx_lock(&smq_mtx);
-	PATCH_TS(ts_type, mess, hp, NULL, NULL);
-	send_to_port(p, mess, tracer_pid, tracee_flags);
-	UnUseTmpHeapNoproc(LOCAL_HEAP_SIZE);
-#undef LOCAL_HEAP_SIZE
-	erts_smp_mtx_unlock(&smq_mtx);
-    } else {
-	ErlHeapFragment *bp;
-	ErlOffHeap *off_heap;
-	ERTS_TRACER_REF_TYPE tracer_ref;
-	unsigned size;
-	unsigned retval_size;
-
-	ASSERT(is_internal_pid(*tracer_pid));
-
-	ERTS_GET_TRACER_REF(tracer_ref, *tracer_pid, *tracee_flags);
-
-	retval_size = size_object(retval);
-	size = 6 + 4 + retval_size + patch_ts_size(ts_type);
-
-	hp = ERTS_ALLOC_SYSMSG_HEAP(size, &bp, &off_heap, tracer_ref);
-
-	/*
-	 * Build the trace tuple and put it into receive queue of the tracer process.
-	 */
-	
-	mfa = TUPLE3(hp, mod, name, make_small(arity));
-	hp += 4;
-	retval = copy_struct(retval, retval_size, &hp, off_heap);
-	mess = TUPLE5(hp, am_trace, p->common.id, am_return_from, mfa, retval);
-	hp += 6;
 
-	erts_smp_mtx_lock(&smq_mtx);
-
-	PATCH_TS(ts_type, mess, hp, bp, tracer_ref);
-
-	ERTS_ENQ_TRACE_MSG(tracee, tracer_ref, mess, bp);
-	erts_smp_mtx_unlock(&smq_mtx);
-    }
+    hp = HAlloc(p, 4);
+    mfa = TUPLE3(hp, mod, name, make_small(arity));
+    hp += 4;
+    send_to_tracer_nif_raw(p, NULL, *tracer, *tracee_flags, p->common.id,
+                           NULL, am_return_from, mfa, retval, am_true);
 }
 
 /* Send {trace_ts, Pid, exception_from, {Mod, Name, Arity}, {Class,Value}, 
@@ -1625,116 +995,50 @@ erts_trace_return(Process* p, BeamInstr* fi, Eterm retval, Eterm *tracer_pid)
  */
 void
 erts_trace_exception(Process* p, BeamInstr mfa[3], Eterm class, Eterm value,
-		     Eterm *tracer_pid)
+		     ErtsTracer *tracer)
 {
     Eterm* hp;
-    Eterm mfa_tuple;
-    Eterm cv;
-    Eterm mess;
+    Eterm mfa_tuple, cv;
     Uint meta_flags, *tracee_flags;
-    int ts_type;
-#ifdef ERTS_SMP
-    Eterm tracee;
-#endif
-    
-    ASSERT(tracer_pid);
-    if (*tracer_pid == am_true) {
+
+    ASSERT(tracer);
+    if (ERTS_TRACER_COMPARE(*tracer, erts_tracer_true)) {
 	/* Breakpoint trace enabled without specifying tracer =>
 	 *   use process tracer and flags
 	 */
-	tracer_pid = &ERTS_TRACER_PROC(p);
+	tracer = &ERTS_TRACER(p);
     }
-    if (is_nil(*tracer_pid)) {
+    if (ERTS_TRACER_IS_NIL(*tracer)) {
 	/* Trace disabled */
 	return;
     }
-    ASSERT(is_internal_pid(*tracer_pid) || is_internal_port(*tracer_pid));
-    if (*tracer_pid == p->common.id) {
-	/* Do not generate trace messages to oneself */
-	return;
-    }
-    if (tracer_pid == &ERTS_TRACER_PROC(p)) {
+    ASSERT(IS_TRACER_VALID(*tracer));
+    if (tracer == &ERTS_TRACER(p)) {
 	/* Tracer specified in process structure =>
 	 *   non-breakpoint trace =>
 	 *     use process flags
 	 */
 	tracee_flags = &ERTS_TRACE_FLAGS(p);
-#ifdef ERTS_SMP
-	tracee = p->common.id;
-#endif
-	if (! (*tracee_flags & F_TRACE_CALLS)) {
-	    return;
-	}
+        if (! (*tracee_flags & F_TRACE_CALLS)) {
+            return;
+        }
     } else {
 	/* Tracer not specified in process structure =>
 	 *   tracer specified in breakpoint =>
 	 *     meta trace =>
 	 *       use fixed flag set instead of process flags
-	 */	    
+	 */
 	meta_flags = F_TRACE_CALLS | F_NOW_TS;
 	tracee_flags = &meta_flags;
-#ifdef ERTS_SMP
-	tracee = NIL;
-#endif
     }
-    
-    ts_type = ERTS_TFLGS2TSTYPE(*tracee_flags);
-
-    if (is_internal_port(*tracer_pid)) {
-#define LOCAL_HEAP_SIZE (4+3+6+ERTS_TRACE_PATCH_TS_MAX_SIZE)
-	DeclareTmpHeapNoproc(local_heap,LOCAL_HEAP_SIZE);
-	UseTmpHeapNoproc(LOCAL_HEAP_SIZE);
 
-	hp = local_heap;
-	mfa_tuple = TUPLE3(hp, (Eterm) mfa[0], (Eterm) mfa[1], make_small((Eterm)mfa[2]));
-	hp += 4;
-	cv = TUPLE2(hp, class, value);
-	hp += 3;
-	mess = TUPLE5(hp, am_trace, p->common.id, am_exception_from, mfa_tuple, cv);
-	hp += 6;
-	ASSERT((hp - local_heap) <= LOCAL_HEAP_SIZE);
-	erts_smp_mtx_lock(&smq_mtx);
-	PATCH_TS(ts_type, mess, hp, NULL, NULL);
-	send_to_port(p, mess, tracer_pid, tracee_flags);
-	UnUseTmpHeapNoproc(LOCAL_HEAP_SIZE);
-#undef LOCAL_HEAP_SIZE
-	erts_smp_mtx_unlock(&smq_mtx);
-    } else {
-	ErlHeapFragment *bp;
-	ErlOffHeap *off_heap;
-	ERTS_TRACER_REF_TYPE tracer_ref;
-	unsigned size;
-	unsigned value_size;
-
-	ASSERT(is_internal_pid(*tracer_pid));
-
-	ERTS_GET_TRACER_REF(tracer_ref, *tracer_pid, *tracee_flags);
-	
-	value_size = size_object(value);
-	size = 6 + 4 + 3 + value_size + patch_ts_size(ts_type);
-
-	hp = ERTS_ALLOC_SYSMSG_HEAP(size, &bp, &off_heap, tracer_ref);
-
-	/*
-	 * Build the trace tuple and put it into receive queue of the tracer process.
-	 */
-	
-	mfa_tuple = TUPLE3(hp, (Eterm) mfa[0], (Eterm) mfa[1], make_small((Eterm) mfa[2]));
-	hp += 4;
-	value = copy_struct(value, value_size, &hp, off_heap);
-	cv = TUPLE2(hp, class, value);
-	hp += 3;
-	mess = TUPLE5(hp, am_trace, p->common.id, 
-		      am_exception_from, mfa_tuple, cv);
-	hp += 6;
-
-	erts_smp_mtx_lock(&smq_mtx);
-
-	PATCH_TS(ts_type, mess, hp, bp, tracer_ref);
-
-	ERTS_ENQ_TRACE_MSG(tracee, tracer_ref, mess, bp);
-	erts_smp_mtx_unlock(&smq_mtx);
-    }
+    hp = HAlloc(p, 7);;
+    mfa_tuple = TUPLE3(hp, (Eterm) mfa[0], (Eterm) mfa[1], make_small((Eterm)mfa[2]));
+    hp += 4;
+    cv = TUPLE2(hp, class, value);
+    hp += 3;
+    send_to_tracer_nif_raw(p, NULL, *tracer, *tracee_flags, p->common.id,
+                           NULL, am_exception_from, mfa_tuple, cv, am_true);
 }
 
 /*
@@ -1753,7 +1057,7 @@ erts_trace_exception(Process* p, BeamInstr mfa[3], Eterm class, Eterm value,
  */
 Uint32
 erts_call_trace(Process* p, BeamInstr mfa[3], Binary *match_spec,
-		Eterm* args, int local, Eterm *tracer_pid)
+		Eterm* args, int local, ErtsTracer *tracer)
 {
     Eterm* hp;
     Eterm mfa_tuple;
@@ -1761,55 +1065,50 @@ erts_call_trace(Process* p, BeamInstr mfa[3], Binary *match_spec,
     int i;
     Uint32 return_flags;
     Eterm pam_result = am_true;
-    Eterm mess;
     Uint meta_flags, *tracee_flags;
-    int ts_type;
-#ifdef ERTS_SMP
-    Eterm tracee;
-#endif
+    ErtsTracerNif *tnif = NULL;
     Eterm transformed_args[MAX_ARG];
-    DeclareTypedTmpHeap(ErlSubBin,sub_bin_heap,p);
+    ErtsTracer pre_ms_tracer = erts_tracer_nil;
 
-    ASSERT(tracer_pid);
-    if (*tracer_pid == am_true) {
-	/* Breakpoint trace enabled without specifying tracer =>
+    ASSERT(tracer);
+    if (ERTS_TRACER_COMPARE(*tracer, erts_tracer_true)) {
+        /* Breakpoint trace enabled without specifying tracer =>
 	 *   use process tracer and flags
 	 */
-	tracer_pid = &ERTS_TRACER_PROC(p);
-    } 
-    if (is_nil(*tracer_pid)) {
-	/* Trace disabled */
-	return 0;
+        tracer = &ERTS_TRACER(p);
     }
-    ASSERT(is_internal_pid(*tracer_pid) || is_internal_port(*tracer_pid));
-    if (*tracer_pid == p->common.id) {
-	/* Do not generate trace messages to oneself */
+    if (ERTS_TRACER_IS_NIL(*tracer)) {
+	/* Trace disabled */
 	return 0;
     }
-    if (tracer_pid == &ERTS_TRACER_PROC(p)) {
+    ASSERT(IS_TRACER_VALID(*tracer));
+    if (tracer == &ERTS_TRACER(p)) {
 	/* Tracer specified in process structure =>
 	 *   non-breakpoint trace =>
 	 *     use process flags
 	 */
 	tracee_flags = &ERTS_TRACE_FLAGS(p);
-#ifdef ERTS_SMP
-	tracee = p->common.id;
-#endif
+        if (!is_tracer_proc_enabled(p, ERTS_PROC_LOCK_MAIN, &p->common, &tnif, am_call)) {
+            return 0;
+        }
     } else {
 	/* Tracer not specified in process structure =>
 	 *   tracer specified in breakpoint =>
 	 *     meta trace =>
 	 *       use fixed flag set instead of process flags
-	 */	    
-	if (ERTS_TRACE_FLAGS(p) & F_SENSITIVE) {
-	    /* No trace messages for sensitive processes. */
-	    return 0;
-	}
+	 */
+        if (ERTS_TRACE_FLAGS(p) & F_SENSITIVE) {
+            /* No trace messages for sensitive processes. */
+            return 0;
+        }
 	meta_flags = F_TRACE_CALLS | F_NOW_TS;
 	tracee_flags = &meta_flags;
-#ifdef ERTS_SMP
-	tracee = NIL;
-#endif
+        switch (call_enabled_tracer(p, *tracer, &tnif, am_call, p->common.id)) {
+        default:
+        case am_remove: *tracer = erts_tracer_nil;
+        case am_discard: return 0;
+        case am_trace: break;
+        }
     }
 
     /*
@@ -1820,18 +1119,13 @@ erts_call_trace(Process* p, BeamInstr mfa[3], Binary *match_spec,
      * temporarily convert any match contexts to sub binaries.
      */
     arity = (Eterm) mfa[2];
-    UseTmpHeap(ERL_SUB_BIN_SIZE,p);
-#ifdef DEBUG
-    sub_bin_heap->thing_word = 0;
-#endif
     for (i = 0; i < arity; i++) {
 	Eterm arg = args[i];
 	if (is_boxed(arg) && header_is_bin_matchstate(*boxed_val(arg))) {
 	    ErlBinMatchState* ms = (ErlBinMatchState *) boxed_val(arg);
 	    ErlBinMatchBuffer* mb = &ms->mb;
 	    Uint bit_size;
-
-	    ASSERT(sub_bin_heap->thing_word == 0); /* At most one of match context */
+            ErlSubBin *sub_bin_heap = (ErlSubBin *)HAlloc(p, ERL_SUB_BIN_SIZE);
 
 	    bit_size = mb->size - mb->offset;
 	    sub_bin_heap->thing_word = HEADER_SUB_BIN;
@@ -1848,275 +1142,98 @@ erts_call_trace(Process* p, BeamInstr mfa[3], Binary *match_spec,
     }
     args = transformed_args;
 
-    ts_type = ERTS_TFLGS2TSTYPE(*tracee_flags);
-
-    if (is_internal_port(*tracer_pid)) {
-	Eterm local_heap[64+ERTS_TRACE_PATCH_TS_MAX_SIZE+MAX_ARG];
-	hp = local_heap;
+    /*
+     * If there is a PAM program, run it.  Return if it fails.
+     *
+     * Some precedence rules:
+     *
+     * - No proc flags, e.g 'silent' or 'return_to'
+     *   has any effect on meta trace.
+     * - The 'silent' process trace flag silences all call
+     *   related messages, e.g 'call', 'return_to' and 'return_from'.
+     * - The {message,_} PAM function does not affect {return_trace}.
+     * - The {message,false} PAM function shall give the same
+     *   'call' trace message as no PAM match.
+     * - The {message,true} PAM function shall give the same
+     *   'call' trace message as a nonexistent PAM program.
+     */
 
-	if (!erts_is_valid_tracer_port(*tracer_pid)) {
-#ifdef ERTS_SMP
-	    ASSERT(is_nil(tracee) || tracer_pid == &ERTS_TRACER_PROC(p));
-	    if (is_not_nil(tracee))
-		erts_smp_proc_lock(p, ERTS_PROC_LOCKS_ALL_MINOR);
-#endif
-	    *tracee_flags &= ~TRACEE_FLAGS;
-	    *tracer_pid = NIL;
-#ifdef ERTS_SMP
-	    if (is_not_nil(tracee))
-		erts_smp_proc_unlock(p, ERTS_PROC_LOCKS_ALL_MINOR);
-#endif
-	    UnUseTmpHeap(ERL_SUB_BIN_SIZE,p);
-	    return 0;
-	}
-	
-	/*
-	 * If there is a PAM program, run it.  Return if it fails.
-	 *
-	 * Some precedence rules:
-	 * 
-	 * - No proc flags, e.g 'silent' or 'return_to' 
-	 *   has any effect on meta trace.
-	 * - The 'silent' process trace flag silences all call 
-	 *   related messages, e.g 'call', 'return_to' and 'return_from'.
-	 * - The {message,_} PAM function does not affect {return_trace}.
-	 * - The {message,false} PAM function shall give the same
-	 *   'call' trace message as no PAM match.
-	 * - The {message,true} PAM function shall give the same
-	 *   'call' trace message as a nonexistent PAM program.
-	 */
-	
-	/* BEGIN this code should be the same for port and pid trace */
-	return_flags = 0;
-	if (match_spec) {
-	    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);
-		return 0;
-	    }
-	}
-	if (tracee_flags == &meta_flags) {
-	    /* Meta trace */
-	    if (pam_result == am_false) {
-		erts_match_set_release_result(p);
-		UnUseTmpHeap(ERL_SUB_BIN_SIZE,p);
-		return return_flags;
-	    }
-	} else {
-	    /* Non-meta trace */
-	    if (*tracee_flags & F_TRACE_SILENT) { 
-		erts_match_set_release_result(p);
-		UnUseTmpHeap(ERL_SUB_BIN_SIZE,p);
-		return 0;
-	    }
-	    if (pam_result == am_false) {
-		erts_match_set_release_result(p);
-		UnUseTmpHeap(ERL_SUB_BIN_SIZE,p);
-		return return_flags;
-	    }
-	    if (local && (*tracee_flags & F_TRACE_RETURN_TO)) {
-		return_flags |= MATCH_SET_RETURN_TO_TRACE;
-	    }
-	}
-	/* END this code should be the same for port and pid trace */
-	
-	/*
-	 * Build the the {M,F,A} tuple in the local heap. 
-	 * (A is arguments or arity.)
-	 */
-	
-	if (*tracee_flags & F_TRACE_ARITY_ONLY) {
-	    mfa_tuple = make_small(arity);
-	} else {
-	    mfa_tuple = NIL;
-	    for (i = arity-1; i >= 0; i--) {
-		mfa_tuple = CONS(hp, args[i], mfa_tuple);
-		hp += 2;
-	    }
-	}
-	mfa_tuple = TUPLE3(hp, (Eterm) mfa[0], (Eterm) mfa[1], mfa_tuple);
-	hp += 4;
-	
-	/*
-	 * Build the trace tuple and send it to the port.
-	 */
-	
-	mess = TUPLE4(hp, am_trace, p->common.id, am_call, mfa_tuple);
-	hp += 5;
-	if (pam_result != am_true) {
-	    hp[-5] = make_arityval(5);
-	    *hp++ = pam_result;
-	}
-	erts_smp_mtx_lock(&smq_mtx);
-	PATCH_TS(ts_type, mess, hp, NULL, NULL);
-	send_to_port(p, mess, tracer_pid, tracee_flags);
-	erts_smp_mtx_unlock(&smq_mtx);
-	erts_match_set_release_result(p);
-	UnUseTmpHeap(ERL_SUB_BIN_SIZE,p);
-	return *tracer_pid == NIL ? 0 : return_flags;
+    return_flags = 0;
+    if (match_spec) {
+        /* we have to make a copy of the tracer here as the 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;
+        }
+    }
 
+    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;
+        }
     } else {
-	ErlHeapFragment *bp;
-	ErlOffHeap *off_heap;
-	Process *tracer;
-	ERTS_TRACER_REF_TYPE tracer_ref;
-#ifdef ERTS_SMP
-	Eterm tpid;
-#endif
-	unsigned size;
-	unsigned sizes[MAX_ARG];
-	unsigned pam_result_size = 0;
-	int invalid_tracer;
-
-	ASSERT(is_internal_pid(*tracer_pid));
-	
-	tracer = erts_pid2proc(p, ERTS_PROC_LOCK_MAIN,
-			       *tracer_pid, ERTS_PROC_LOCK_STATUS);
-	if (!tracer)
-	    invalid_tracer = 1;
-	else {
-	    invalid_tracer = !(ERTS_TRACE_FLAGS(tracer) & F_TRACER);
-	    erts_smp_proc_unlock(tracer, ERTS_PROC_LOCK_STATUS);
-	}
+        /* Non-meta trace */
+        if (*tracee_flags & F_TRACE_SILENT) {
+            erts_match_set_release_result(p);
+            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;
+        }
+        if (local && (*tracee_flags & F_TRACE_RETURN_TO)) {
+            return_flags |= MATCH_SET_RETURN_TO_TRACE;
+        }
+    }
 
-	if (invalid_tracer) {
-#ifdef ERTS_SMP
-	    ASSERT(is_nil(tracee)
-		   || tracer_pid == &ERTS_TRACER_PROC(p));
-	    if (is_not_nil(tracee))
-		erts_smp_proc_lock(p, ERTS_PROC_LOCKS_ALL_MINOR);
-#endif
-	    *tracee_flags &= ~TRACEE_FLAGS;
-	    *tracer_pid = NIL;
-#ifdef ERTS_SMP
-	    if (is_not_nil(tracee))
-		erts_smp_proc_unlock(p, ERTS_PROC_LOCKS_ALL_MINOR);
-#endif
-	    UnUseTmpHeap(ERL_SUB_BIN_SIZE,p);
-	    return 0;
-	}
+    ASSERT(!ERTS_TRACER_IS_NIL(*tracer));
 
-#ifdef ERTS_SMP
-	tpid = *tracer_pid; /* Need to save tracer pid,
-			       since *tracer_pid might
-			       be reset by erts_match_set_run() */
-	tracer_ref = tpid;
-#else
-	tracer_ref = tracer;
-#endif
-	
-	/*
-	 * If there is a PAM program, run it.  Return if it fails.
-	 *
-	 * See the rules above in the port trace code.
-	 */
-	
-	/* BEGIN this code should be the same for port and pid trace */
-	return_flags = 0;
-	if (match_spec) {
-	    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);
-		return 0;
-	    }
-	}
-	if (tracee_flags == &meta_flags) {
-	    /* Meta trace */
-	    if (pam_result == am_false) {
-		erts_match_set_release_result(p);
-		UnUseTmpHeap(ERL_SUB_BIN_SIZE,p);
-		return return_flags;
-	    }
-	} else {
-	    /* Non-meta trace */
-	    if (*tracee_flags & F_TRACE_SILENT) { 
-		erts_match_set_release_result(p);
-		UnUseTmpHeap(ERL_SUB_BIN_SIZE,p);
-		return 0;
-	    }
-	    if (pam_result == am_false) {
-		erts_match_set_release_result(p);
-		UnUseTmpHeap(ERL_SUB_BIN_SIZE,p);
-		return return_flags;
-	    }
-	    if (local && (*tracee_flags & F_TRACE_RETURN_TO)) {
-		return_flags |= MATCH_SET_RETURN_TO_TRACE;
-	    }
-	}
-	/* END this code should be the same for port and pid trace */
-	
-	/*
-	 * Calculate number of words needed on heap.
-	 */
-	
-	size = 4 + 5;		/* Trace tuple + MFA tuple. */
-	if (! (*tracee_flags & F_TRACE_ARITY_ONLY)) {
-	    size += 2*arity;
-	    for (i = arity-1; i >= 0; i--) {
-		sizes[i] = size_object(args[i]);
-		size += sizes[i];
-	    }
-	}
-	size += patch_ts_size(ts_type);
-	if (pam_result != am_true) {
-	    pam_result_size = size_object(pam_result);
-	    size += 1 + pam_result_size;
-	    /* One element in trace tuple + term size. */
-	}
-	
-	hp = ERTS_ALLOC_SYSMSG_HEAP(size, &bp, &off_heap, tracer_ref);
+    /*
+     * Build the the {M,F,A} tuple in the local heap.
+     * (A is arguments or arity.)
+     */
 
-	/*
-	 * Build the the {M,F,A} tuple in the message buffer. 
-	 * (A is arguments or arity.)
-	 */
-	
-	if (*tracee_flags & F_TRACE_ARITY_ONLY) {
-	    mfa_tuple = make_small(arity);
-	} else {
-	    mfa_tuple = NIL;
-	    for (i = arity-1; i >= 0; i--) {
-		Eterm term = copy_struct(args[i], sizes[i], &hp, off_heap);
-		mfa_tuple = CONS(hp, term, mfa_tuple);
-		hp += 2;
-	    }
-	}
-	mfa_tuple = TUPLE3(hp, (Eterm) mfa[0], (Eterm) mfa[1], mfa_tuple);
-	hp += 4;
-	
-	/*
-	 * Copy the PAM result (if any) onto the heap.
-	 */
-	
-	if (pam_result != am_true) {
-	    pam_result = copy_struct(pam_result, pam_result_size, &hp, off_heap);
-	}
 
-	erts_match_set_release_result(p);
+    if (*tracee_flags & F_TRACE_ARITY_ONLY) {
+        hp = HAlloc(p, 4);
+        mfa_tuple = make_small(arity);
+    } else {
+        hp = HAlloc(p, 4 + arity * 2);
+        mfa_tuple = NIL;
+        for (i = arity-1; i >= 0; i--) {
+            mfa_tuple = CONS(hp, args[i], mfa_tuple);
+            hp += 2;
+        }
+    }
+    mfa_tuple = TUPLE3(hp, (Eterm) mfa[0], (Eterm) mfa[1], mfa_tuple);
+    hp += 4;
 
-	/*
-	 * Build the trace tuple and enqueue it.
-	 */
-	
-	mess = TUPLE4(hp, am_trace, p->common.id, am_call, mfa_tuple);
-	hp += 5;
-	if (pam_result != am_true) {
-	    hp[-5] = make_arityval(5);
-	    *hp++ = pam_result;
-	}
+    /*
+     * Build the trace tuple and send it to the port.
+     */
+    send_to_tracer_nif_raw(p, NULL, *tracer, *tracee_flags, p->common.id,
+                           tnif, am_call, mfa_tuple, THE_NON_VALUE, pam_result);
+    erts_match_set_release_result(p);
 
-	erts_smp_mtx_lock(&smq_mtx);
+    if (match_spec && tracer == &pre_ms_tracer)
+        ERTS_TRACER_CLEAR(&pre_ms_tracer);
 
-	PATCH_TS(ts_type, mess, hp, bp, tracer_ref);
-	ERTS_ENQ_TRACE_MSG(tracee, tracer_ref, mess, bp);
-	erts_smp_mtx_unlock(&smq_mtx);
-	UnUseTmpHeap(ERL_SUB_BIN_SIZE,p);
-	return return_flags;
-    }
+    return return_flags;
 }
 
 /* Sends trace message:
@@ -2128,69 +1245,13 @@ erts_call_trace(Process* p, BeamInstr mfa[3], Binary *match_spec,
  * 't_p' is the traced process.
  */
 void
-trace_proc(Process *c_p, Process *t_p, Eterm what, Eterm data)
+trace_proc(Process *c_p, ErtsProcLocks c_p_locks,
+           Process *t_p, Eterm what, Eterm data)
 {
-    Eterm mess;
-    Eterm* hp;
-    int need;
-
-    ERTS_SMP_LC_ASSERT((erts_proc_lc_my_proc_locks(t_p) != 0)
-		       || erts_thr_progress_is_blocking());
-    if (is_internal_port(ERTS_TRACER_PROC(t_p))) {
-#define LOCAL_HEAP_SIZE (5+5)
-	DeclareTmpHeapNoproc(local_heap,LOCAL_HEAP_SIZE);
-	UseTmpHeapNoproc(LOCAL_HEAP_SIZE);
-
-
-	hp = local_heap;
-	mess = TUPLE4(hp, am_trace, t_p->common.id, what, data);
-	hp += 5;
-	erts_smp_mtx_lock(&smq_mtx);
-	PATCH_TS(TFLGS_TS_TYPE(t_p), mess, hp, NULL, NULL);
-	send_to_port(
-#ifndef ERTS_SMP
-	    /* No fake schedule out and in again after an exit */
-	    what == am_exit ? NULL : c_p,
-#else
-	    /* Fake schedule out and in are never sent when smp enabled */
-	    c_p,
-#endif
-	    mess,
-	    &ERTS_TRACER_PROC(t_p),
-	    &ERTS_TRACE_FLAGS(t_p));
-	UnUseTmpHeapNoproc(LOCAL_HEAP_SIZE);
-#undef LOCAL_HEAP_SIZE
-	erts_smp_mtx_unlock(&smq_mtx);
-    } else {
-	Eterm tmp;
-	ErlHeapFragment *bp;
-	ErlOffHeap *off_heap;
-	ERTS_TRACER_REF_TYPE tracer_ref;
-	size_t sz_data;
-
-	ASSERT(is_internal_pid(ERTS_TRACER_PROC(t_p)));
-
-	ERTS_GET_TRACER_REF(tracer_ref,
-			    ERTS_TRACER_PROC(t_p),
-			    ERTS_TRACE_FLAGS(t_p));
-
-	sz_data = size_object(data);
-
-	need = sz_data + 5 + PATCH_TS_SIZE(t_p);
-
-	hp = ERTS_ALLOC_SYSMSG_HEAP(need, &bp, &off_heap, tracer_ref);
-
-	tmp = copy_struct(data, sz_data, &hp, off_heap);
-	mess = TUPLE4(hp, am_trace, t_p->common.id, what, tmp);
-	hp += 5;
-
-	erts_smp_mtx_lock(&smq_mtx);
-
-	PATCH_TS(TFLGS_TS_TYPE(t_p), mess, hp, bp, tracer_ref);
-
-	ERTS_ENQ_TRACE_MSG(t_p->common.id, tracer_ref, mess, bp);
-	erts_smp_mtx_unlock(&smq_mtx);
-    }
+    ErtsTracerNif *tnif = NULL;
+    if (is_tracer_proc_enabled(c_p, c_p_locks, &t_p->common, &tnif, what))
+        send_to_tracer_nif(c_p, &t_p->common, t_p->common.id, tnif,
+                           what, data, THE_NON_VALUE);
 }
 
 
@@ -2202,62 +1263,20 @@ trace_proc(Process *c_p, Process *t_p, Eterm what, Eterm data)
  * and 'args' may be a deep term.
  */
 void
-trace_proc_spawn(Process *p, Eterm pid, 
+trace_proc_spawn(Process *p, Eterm what, Eterm pid,
 		 Eterm mod, Eterm func, Eterm args)
 {
-    Eterm mfa;
-    Eterm mess;
-    Eterm* hp;
-
-    if (is_internal_port(ERTS_TRACER_PROC(p))) {
-#define LOCAL_HEAP_SIZE (4+6+5)
-	DeclareTmpHeapNoproc(local_heap,LOCAL_HEAP_SIZE);
-	UseTmpHeapNoproc(LOCAL_HEAP_SIZE);
+    ErtsTracerNif *tnif = NULL;
+    if (is_tracer_proc_enabled(p, ERTS_PROC_LOCKS_ALL &
+                               ~(ERTS_PROC_LOCK_STATUS|ERTS_PROC_LOCK_TRACE),
+                               &p->common, &tnif, what)) {
+        Eterm mfa;
+        Eterm* hp;
 
-	hp = local_heap;
-	mfa = TUPLE3(hp, mod, func, args);
-	hp += 4;
-	mess = TUPLE5(hp, am_trace, p->common.id, am_spawn, pid, mfa);
-	hp += 6;
-	erts_smp_mtx_lock(&smq_mtx);
-	PATCH_TS(TFLGS_TS_TYPE(p), mess, hp, NULL, NULL);
-	send_to_port(p, mess, &ERTS_TRACER_PROC(p), &ERTS_TRACE_FLAGS(p));
-	UnUseTmpHeapNoproc(LOCAL_HEAP_SIZE);
-#undef LOCAL_HEAP_SIZE
-	erts_smp_mtx_unlock(&smq_mtx);
-    } else {
-	Eterm tmp;
-	ErlHeapFragment *bp;
-	ErlOffHeap *off_heap;
-	ERTS_TRACER_REF_TYPE tracer_ref;
-	size_t sz_args, sz_pid;
-	Uint need;
-
-	ASSERT(is_internal_pid(ERTS_TRACER_PROC(p)));
-
-	ERTS_GET_TRACER_REF(tracer_ref,
-			    ERTS_TRACER_PROC(p),
-			    ERTS_TRACE_FLAGS(p));
-
-	sz_args = size_object(args);
-	sz_pid = size_object(pid);
-	need = sz_args + 4 + 6 + PATCH_TS_SIZE(p);
-
-	hp = ERTS_ALLOC_SYSMSG_HEAP(need, &bp, &off_heap, tracer_ref);
-
-	tmp = copy_struct(args, sz_args, &hp, off_heap);
-	mfa = TUPLE3(hp, mod, func, tmp);
-	hp += 4;
-	tmp = copy_struct(pid, sz_pid, &hp, off_heap);
-	mess = TUPLE5(hp, am_trace, p->common.id, am_spawn, tmp, mfa);
-	hp += 6;
-
-	erts_smp_mtx_lock(&smq_mtx);
-
-	PATCH_TS(TFLGS_TS_TYPE(p), mess, hp, bp, tracer_ref);
-
-	ERTS_ENQ_TRACE_MSG(p->common.id, tracer_ref, mess, bp);
-	erts_smp_mtx_unlock(&smq_mtx);
+        hp = HAlloc(p, 4);
+        mfa = TUPLE3(hp, mod, func, args);
+        hp += 4;
+        send_to_tracer_nif(p, &p->common, p->common.id, tnif, what, pid, mfa);
     }
 }
 
@@ -2291,66 +1310,22 @@ void save_calls(Process *p, Export *e)
 void
 trace_gc(Process *p, Eterm what)
 {
-    ErlHeapFragment *bp = NULL;
-    ErlOffHeap *off_heap;
-    ERTS_TRACER_REF_TYPE tracer_ref = ERTS_NULL_TRACER_REF;	/* Initialized
-								   to eliminate
-								   compiler
-								   warning */
+    ErtsTracerNif *tnif = NULL;
     Eterm* hp;
     Eterm msg = NIL;
-    Uint size;
+    Uint size = 0;
 
-#define LOCAL_HEAP_SIZE						\
-    (ERTS_PROCESS_GC_INFO_MAX_SIZE) +                           \
-	5/*4-tuple */ + ERTS_TRACE_PATCH_TS_MAX_SIZE
-    DeclareTmpHeap(local_heap,LOCAL_HEAP_SIZE,p);
+    if (is_tracer_proc_enabled(
+            p, ERTS_PROC_LOCK_MAIN, &p->common, &tnif, what)) {
 
-    UseTmpHeap(LOCAL_HEAP_SIZE,p);
-
-    if (is_internal_port(ERTS_TRACER_PROC(p))) {
-	hp = local_heap;
-#ifdef DEBUG
-	size = 0;
         (void) erts_process_gc_info(p, &size, NULL);
+        hp = HAlloc(p, size);
 
-	size += 5/*4-tuple*/ + PATCH_TS_SIZE(p);
-#endif
-    } else {
-	ASSERT(is_internal_pid(ERTS_TRACER_PROC(p)));
-
-	ERTS_GET_TRACER_REF(tracer_ref,
-			    ERTS_TRACER_PROC(p),
-			    ERTS_TRACE_FLAGS(p));
-
-	size = 0;
-        (void) erts_process_gc_info(p, &size, NULL);
-
-	size += 5/*4-tuple*/ + PATCH_TS_SIZE(p);
-
-	hp = ERTS_ALLOC_SYSMSG_HEAP(size, &bp, &off_heap, tracer_ref);
-    }
-
-    ASSERT(size <= LOCAL_HEAP_SIZE);
+        msg = erts_process_gc_info(p, NULL, &hp);
 
-    msg = erts_process_gc_info(p, NULL, &hp);
-
-    msg = TUPLE4(hp, am_trace, p->common.id, what, msg);
-    hp += 5;
-
-    erts_smp_mtx_lock(&smq_mtx);
-
-    if (is_internal_port(ERTS_TRACER_PROC(p))) {
-	PATCH_TS(TFLGS_TS_TYPE(p), msg, hp, NULL, NULL);
-	send_to_port(p, msg, &ERTS_TRACER_PROC(p), &ERTS_TRACE_FLAGS(p));
-    }
-    else {
-	PATCH_TS(TFLGS_TS_TYPE(p), msg, hp, bp, tracer_ref);
-	ERTS_ENQ_TRACE_MSG(p->common.id, tracer_ref, msg, bp);
+        send_to_tracer_nif(p, &p->common, p->common.id, tnif, what,
+                           msg, THE_NON_VALUE);
     }
-    erts_smp_mtx_unlock(&smq_mtx);
-    UnUseTmpHeap(LOCAL_HEAP_SIZE,p);
-#undef LOCAL_HEAP_SIZE
 }
 
 void 
@@ -2412,7 +1387,7 @@ 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, NIL);
+	erts_queue_message(monitor_p, NULL, mp, msg);
     }
 #endif
 }
@@ -2477,7 +1452,7 @@ 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, NIL);
+	erts_queue_message(monitor_p, NULL, mp, msg);
     }
 #endif
 }
@@ -2552,7 +1527,7 @@ 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, NIL);
+	erts_queue_message(monitor_p, NULL, mp, msg);
     }
 #endif
 }
@@ -2627,7 +1602,7 @@ 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, NIL);
+	erts_queue_message(monitor_p, NULL, mp, msg);
     }
 #endif
 }
@@ -2659,7 +1634,7 @@ 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, NIL);
+	erts_queue_message(monitor_p, NULL, mp, msg);
     }
 #endif
 
@@ -2756,71 +1731,15 @@ profile_scheduler_q(Eterm scheduler_id, Eterm state, Eterm no_schedulers, Uint M
 
 }
 
-
-/* Send {trace_ts, Pid, What, {Mod, Func, Arity}, Timestamp}
- * or   {trace, Pid, What, {Mod, Func, Arity}}
- *
- * where 'What' is supposed to be 'in' or 'out'.
- *
- * Virtual scheduling do not fake scheduling for ports.
- */
-
-
-void trace_virtual_sched(Process *p, Eterm what)
-{
-    trace_sched_aux(p, what, 1);
-}
-
 /* Port profiling */
 
 void
 trace_port_open(Port *p, Eterm calling_pid, Eterm drv_name) {
-    Eterm mess;
-    Eterm* hp;
-
-    if (is_internal_port(ERTS_TRACER_PROC(p))) {
-#define LOCAL_HEAP_SIZE (6+ERTS_TRACE_PATCH_TS_MAX_SIZE)
-	DeclareTmpHeapNoproc(local_heap,LOCAL_HEAP_SIZE);
-	UseTmpHeapNoproc(LOCAL_HEAP_SIZE);
-
-	hp = local_heap;
-
-	mess = TUPLE5(hp, am_trace, calling_pid, am_open, p->common.id, drv_name);
-	hp += 6;
-	erts_smp_mtx_lock(&smq_mtx);
-	PATCH_TS(TFLGS_TS_TYPE(p), mess, hp, NULL, NULL);
-	/* No fake schedule */
-	send_to_port(NULL, mess, &ERTS_TRACER_PROC(p), &ERTS_TRACE_FLAGS(p));
-	UnUseTmpHeapNoproc(LOCAL_HEAP_SIZE);
-#undef LOCAL_HEAP_SIZE
-	erts_smp_mtx_unlock(&smq_mtx);
-    } else {
-	ErlHeapFragment *bp;
-	ErlOffHeap *off_heap;
-	size_t sz_data;
-	ERTS_TRACER_REF_TYPE tracer_ref;
-	
-	ASSERT(is_internal_pid(ERTS_TRACER_PROC(p)));
-
-	sz_data = 6 + PATCH_TS_SIZE(p);
-
-	ERTS_GET_TRACER_REF(tracer_ref,
-			    ERTS_TRACER_PROC(p),
-			    ERTS_TRACE_FLAGS(p));
-
-	hp = ERTS_ALLOC_SYSMSG_HEAP(sz_data, &bp, &off_heap, tracer_ref);
-
-	mess = TUPLE5(hp, am_trace, calling_pid, am_open, p->common.id, drv_name);
-	hp += 6;
-
-	erts_smp_mtx_lock(&smq_mtx);
-
-	PATCH_TS(TFLGS_TS_TYPE(p), mess, hp, bp, tracer_ref);
-
-	ERTS_ENQ_TRACE_MSG(p->common.id, tracer_ref, mess, bp);
-	erts_smp_mtx_unlock(&smq_mtx);
-    }
-
+    ErtsTracerNif *tnif = NULL;
+    ERTS_SMP_CHK_NO_PROC_LOCKS;
+    if (is_tracer_proc_enabled(NULL, 0, &p->common, &tnif, am_open))
+        send_to_tracer_nif(NULL, &p->common, p->common.id, tnif, am_open,
+                           calling_pid, drv_name);
 }
 
 /* Sends trace message:
@@ -2832,52 +1751,208 @@ trace_port_open(Port *p, Eterm calling_pid, Eterm drv_name) {
  */
 void
 trace_port(Port *t_p, Eterm what, Eterm data) {
-    Eterm mess;
-    Eterm* hp;
 
+    ErtsTracerNif *tnif = NULL;
     ERTS_SMP_LC_ASSERT(erts_lc_is_port_locked(t_p)
 		       || erts_thr_progress_is_blocking());
+    ERTS_SMP_CHK_NO_PROC_LOCKS;
+    if (is_tracer_proc_enabled(NULL, 0, &t_p->common, &tnif, what))
+        send_to_tracer_nif(NULL, &t_p->common, t_p->common.id, tnif,
+                           what, data, THE_NON_VALUE);
+}
 
-    if (is_internal_port(ERTS_TRACER_PROC(t_p))) {
-#define LOCAL_HEAP_SIZE (5+ERTS_TRACE_PATCH_TS_MAX_SIZE)
-	DeclareTmpHeapNoproc(local_heap,LOCAL_HEAP_SIZE);
-	UseTmpHeapNoproc(LOCAL_HEAP_SIZE);
 
-	hp = local_heap;
-	mess = TUPLE4(hp, am_trace, t_p->common.id, what, data);
-	hp += 5;
-	erts_smp_mtx_lock(&smq_mtx);
-	PATCH_TS(TFLGS_TS_TYPE(t_p), mess, hp, NULL, NULL);
-	/* No fake schedule */
-	send_to_port(NULL,mess,&ERTS_TRACER_PROC(t_p),&ERTS_TRACE_FLAGS(t_p));
-	UnUseTmpHeapNoproc(LOCAL_HEAP_SIZE);
-#undef LOCAL_HEAP_SIZE
-	erts_smp_mtx_unlock(&smq_mtx);
+static Eterm
+trace_port_tmp_binary(char *bin, Sint sz, Binary **bptrp, Eterm **hp)
+{
+    if (sz <= ERL_ONHEAP_BIN_LIMIT) {
+        ErlHeapBin *hb = (ErlHeapBin *)*hp;
+        hb->thing_word = header_heap_bin(sz);
+        hb->size = sz;
+        sys_memcpy(hb->data, bin, sz);
+        *hp += heap_bin_size(sz);
+        return make_binary(hb);
     } else {
-	ErlHeapFragment *bp;
-	ErlOffHeap *off_heap;
-	size_t sz_data;
-	ERTS_TRACER_REF_TYPE tracer_ref;
+        ProcBin* pb = (ProcBin *)*hp;
+        Binary *bptr = erts_bin_nrml_alloc(sz);
+        erts_refc_init(&bptr->refc, 1);
+        sys_memcpy(bptr->orig_bytes, bin, sz);
+        pb->thing_word = HEADER_PROC_BIN;
+        pb->size = sz;
+        pb->next = NULL;
+        pb->val = bptr;
+        pb->bytes = (byte*) bptr->orig_bytes;
+        pb->flags = 0;
+        *bptrp = bptr;
+        *hp += PROC_BIN_SIZE;
+        return make_binary(pb);
+    }
+}
 
-	ASSERT(is_internal_pid(ERTS_TRACER_PROC(t_p)));
+/* Sends trace message:
+ *    {trace, PortPid, 'receive', {pid(), {command, iolist()}}}
+ *    {trace, PortPid, 'receive', {pid(), {control, pid()}}}
+ *    {trace, PortPid, 'receive', {pid(), exit}}
+ *
+ */
+void
+trace_port_receive(Port *t_p, Eterm caller, Eterm what, ...)
+{
+    ErtsTracerNif *tnif = NULL;
+    ERTS_SMP_LC_ASSERT(erts_lc_is_port_locked(t_p)
+		       || erts_thr_progress_is_blocking());
+    ERTS_SMP_CHK_NO_PROC_LOCKS;
+    if (is_tracer_proc_enabled(NULL, 0, &t_p->common, &tnif, am_receive)) {
+        /* We can use a stack heap here, as the nif is called in the
+           context of a port */
+#define LOCAL_HEAP_SIZE (3 + 3 + heap_bin_size(ERL_ONHEAP_BIN_LIMIT) + 3)
+        DeclareTmpHeapNoproc(local_heap,LOCAL_HEAP_SIZE);
+
+        Eterm *hp, data, *orig_hp = NULL;
+        Binary *bptr = NULL;
+        va_list args;
+        UseTmpHeapNoproc(LOCAL_HEAP_SIZE);
+        hp = local_heap;
+
+        if (what == am_close) {
+            data = what;
+        } else {
+            Eterm arg;
+            va_start(args, what);
+            if (what == am_command) {
+                char *bin = va_arg(args, char *);
+                Sint sz = va_arg(args, Sint);
+                va_end(args);
+                arg = trace_port_tmp_binary(bin, sz, &bptr, &hp);
+            } else if (what == am_call || what == am_control) {
+                unsigned int command = va_arg(args, unsigned int);
+                char *bin = va_arg(args, char *);
+                Sint sz = va_arg(args, Sint);
+                Eterm cmd;
+                va_end(args);
+                arg = trace_port_tmp_binary(bin, sz, &bptr, &hp);
+#if defined(ARCH_32)
+                if (!IS_USMALL(0, command)) {
+                    *hp = make_pos_bignum_header(1);
+                    BIG_DIGIT(hp, 0) = (Uint)command;
+                    cmd = make_big(hp);
+                    hp += 2;
+                } else
+#endif
+                {
+                    cmd = make_small((Sint)command);
+                }
+                arg = TUPLE2(hp, cmd, arg);
+                hp += 3;
+            } else if (what == am_commandv) {
+                ErlIOVec *evp = va_arg(args, ErlIOVec*);
+                int i;
+                va_end(args);
+                if ((6 + evp->vsize * (2+PROC_BIN_SIZE+ERL_SUB_BIN_SIZE)) > LOCAL_HEAP_SIZE) {
+                    hp = erts_alloc(ERTS_ALC_T_TMP,
+                                    (6 + evp->vsize * (2+PROC_BIN_SIZE+ERL_SUB_BIN_SIZE)) * sizeof(Eterm));
+                    orig_hp = hp;
+                }
+                arg = NIL;
+                /* Convert each element in the ErlIOVec to a sub bin that points
+                   to a procbin. We don't have to increment the proc bin refc as
+                   the port task keeps the reference alive. */
+                for (i = evp->vsize-1; i >= 0; i--) {
+                    if (evp->iov[i].iov_len) {
+                        ProcBin* pb = (ProcBin*)hp;
+                        ErlSubBin *sb;
+                        ASSERT(evp->binv[i]);
+                        pb->thing_word = HEADER_PROC_BIN;
+                        pb->val = ErlDrvBinary2Binary(evp->binv[i]);
+                        pb->size = pb->val->orig_size;
+                        pb->next = NULL;
+                        pb->bytes = (byte*) pb->val->orig_bytes;
+                        pb->flags = 0;
+                        hp += PROC_BIN_SIZE;
+
+                        sb = (ErlSubBin*) hp;
+                        sb->thing_word = HEADER_SUB_BIN;
+                        sb->size = evp->iov[i].iov_len;
+                        sb->offs = (byte*)(evp->iov[i].iov_base) - pb->bytes;
+                        sb->orig = make_binary(pb);
+                        sb->bitoffs = 0;
+                        sb->bitsize = 0;
+                        sb->is_writable = 0;
+                        hp += ERL_SUB_BIN_SIZE;
+
+                        arg = CONS(hp, make_binary(sb), arg);
+                        hp += 2;
+                    }
+                }
+                what = am_command;
+            } else {
+                arg = va_arg(args, Eterm);
+                va_end(args);
+            }
+            data = TUPLE2(hp, what, arg);
+            hp += 3;
+        }
 
-	sz_data = 5 + PATCH_TS_SIZE(t_p);
+        data = TUPLE2(hp, caller, data);
+        send_to_tracer_nif(NULL, &t_p->common, t_p->common.id, tnif,
+                           am_receive, data, THE_NON_VALUE);
 
-	ERTS_GET_TRACER_REF(tracer_ref,
-			    ERTS_TRACER_PROC(t_p),
-			    ERTS_TRACE_FLAGS(t_p));
+        if (bptr && erts_refc_dectest(&bptr->refc, 1) == 0)
+            erts_bin_free(bptr);
 
-	hp = ERTS_ALLOC_SYSMSG_HEAP(sz_data, &bp, &off_heap, tracer_ref);
+        if (orig_hp)
+            erts_free(ERTS_ALC_T_TMP, orig_hp);
 
-	mess = TUPLE4(hp, am_trace, t_p->common.id, what, data);
-	hp += 5;
+        UnUseTmpHeapNoproc(LOCAL_HEAP_SIZE);
+    }
+#undef LOCAL_HEAP_SIZE
+}
 
-	erts_smp_mtx_lock(&smq_mtx);
+void
+trace_port_send(Port *t_p, Eterm receiver, Eterm msg, int exists)
+{
+    ErtsTracerNif *tnif = NULL;
+    Eterm op = exists ? am_send : am_send_to_non_existing_process;
+    ERTS_SMP_LC_ASSERT(erts_lc_is_port_locked(t_p)
+		       || erts_thr_progress_is_blocking());
+    ERTS_SMP_CHK_NO_PROC_LOCKS;
+    if (is_tracer_proc_enabled(NULL, 0, &t_p->common, &tnif, op))
+        send_to_tracer_nif(NULL, &t_p->common, t_p->common.id, tnif,
+                           op, msg, receiver);
+}
 
-	PATCH_TS(TFLGS_TS_TYPE(t_p), mess, hp, bp, tracer_ref);
+void trace_port_send_binary(Port *t_p, Eterm to, Eterm what, char *bin, Sint sz)
+{
+    ErtsTracerNif *tnif = NULL;
+    ERTS_SMP_LC_ASSERT(erts_lc_is_port_locked(t_p)
+		       || erts_thr_progress_is_blocking());
+    ERTS_SMP_CHK_NO_PROC_LOCKS;
+    if (is_tracer_proc_enabled(NULL, 0, &t_p->common, &tnif, am_send)) {
+        Eterm msg;
+        Binary* bptr = NULL;
+#define LOCAL_HEAP_SIZE (3 + 3 + heap_bin_size(ERL_ONHEAP_BIN_LIMIT))
+        DeclareTmpHeapNoproc(local_heap,LOCAL_HEAP_SIZE);
 
-	ERTS_ENQ_TRACE_MSG(t_p->common.id, tracer_ref, mess, bp);
-	erts_smp_mtx_unlock(&smq_mtx);
+        Eterm *hp;
+
+        ERTS_CT_ASSERT(heap_bin_size(ERL_ONHEAP_BIN_LIMIT) >= PROC_BIN_SIZE);
+        UseTmpHeapNoproc(LOCAL_HEAP_SIZE);
+        hp = local_heap;
+
+        msg = trace_port_tmp_binary(bin, sz, &bptr, &hp);
+
+        msg = TUPLE2(hp, what, msg);
+        hp += 3;
+        msg = TUPLE2(hp, t_p->common.id, msg);
+        hp += 3;
+
+        send_to_tracer_nif(NULL, &t_p->common, t_p->common.id, tnif,
+                           am_send, msg, to);
+        if (bptr && erts_refc_dectest(&bptr->refc, 1) == 0)
+            erts_bin_free(bptr);
+
+        UnUseTmpHeapNoproc(LOCAL_HEAP_SIZE);
+#undef LOCAL_HEAP_SIZE
     }
 }
 
@@ -2891,83 +1966,18 @@ trace_port(Port *t_p, Eterm what, Eterm data) {
 
 void
 trace_sched_ports(Port *p, Eterm what) {
-    trace_sched_ports_where(p,what, make_small(0));
+    trace_sched_ports_where(p, what, make_small(0));
 }
 
-void 
-trace_sched_ports_where(Port *p, Eterm what, Eterm where) {
-    Eterm mess;
-    Eterm* hp;
-    int ws = 5;	
-    Eterm sched_id = am_undefined;
-    
-    if (is_internal_port(ERTS_TRACER_PROC(p))) {
-#define LOCAL_HEAP_SIZE (6+ERTS_TRACE_PATCH_TS_MAX_SIZE)
-	DeclareTmpHeapNoproc(local_heap,LOCAL_HEAP_SIZE);
-	UseTmpHeapNoproc(LOCAL_HEAP_SIZE);
-
-	hp = local_heap;
-	
-	if (IS_TRACED_FL(p, F_TRACE_SCHED_NO)) {
-#ifdef ERTS_SMP
-		ErtsSchedulerData *esd = erts_get_scheduler_data();
-		if (esd) sched_id = make_small(esd->no);
-		else sched_id = am_undefined;
-#else
-		sched_id = make_small(1);
-#endif
-		mess = TUPLE5(hp, am_trace, p->common.id, what, sched_id, where);
-		ws = 6;
-	} else {
-		mess = TUPLE4(hp, am_trace, p->common.id, what, where);
-		ws = 5;
-	}
-	hp += ws;
-	
-	erts_smp_mtx_lock(&smq_mtx);
-
-	PATCH_TS(TFLGS_TS_TYPE(p), mess, hp, NULL, NULL);
-	
-	/* No fake scheduling */
-	send_to_port(NULL, mess, &ERTS_TRACER_PROC(p), &ERTS_TRACE_FLAGS(p));
-	UnUseTmpHeapNoproc(LOCAL_HEAP_SIZE);
-#undef LOCAL_HEAP_SIZE
-	erts_smp_mtx_unlock(&smq_mtx);
-    } else  {
-	ErlHeapFragment *bp;
-    	ErlOffHeap *off_heap;
-	ERTS_TRACER_REF_TYPE tracer_ref;
-
-	ASSERT(is_internal_pid(ERTS_TRACER_PROC(p)));
-	
-	if (IS_TRACED_FL(p, F_TRACE_SCHED_NO)) ws = 6; /* Make place for scheduler id */
-
-	ERTS_GET_TRACER_REF(tracer_ref,
-			    ERTS_TRACER_PROC(p),
-			    ERTS_TRACE_FLAGS(p));
-
-	hp = ERTS_ALLOC_SYSMSG_HEAP(ws+PATCH_TS_SIZE(p), &bp, &off_heap, tracer_ref);
-
-	if (IS_TRACED_FL(p, F_TRACE_SCHED_NO)) {
-#ifdef ERTS_SMP
-		ErtsSchedulerData *esd = erts_get_scheduler_data();
-		if (esd) sched_id = make_small(esd->no);
-		else sched_id = am_undefined;
-#else
-		sched_id = make_small(1);
-#endif
-		mess = TUPLE5(hp, am_trace, p->common.id, what, sched_id, where);
-	} else {
-		mess = TUPLE4(hp, am_trace, p->common.id, what, where);
-	}
-	hp += ws;
-
-	erts_smp_mtx_lock(&smq_mtx);
-
-	PATCH_TS(TFLGS_TS_TYPE(p), mess, hp, bp, tracer_ref);
-	ERTS_ENQ_TRACE_MSG(p->common.id, tracer_ref, mess, bp);
-	erts_smp_mtx_unlock(&smq_mtx);
-    }
+void
+trace_sched_ports_where(Port *t_p, Eterm what, Eterm where) {
+    ErtsTracerNif *tnif = NULL;
+    ERTS_SMP_LC_ASSERT(erts_lc_is_port_locked(t_p)
+		       || erts_thr_progress_is_blocking());
+    ERTS_SMP_CHK_NO_PROC_LOCKS;
+    if (is_tracer_proc_enabled(NULL, 0, &t_p->common, &tnif, what))
+        send_to_tracer_nif(NULL, &t_p->common, t_p->common.id,
+                           tnif, what, where, THE_NON_VALUE);
 }
 
 /* Port profiling */
@@ -3083,28 +2093,6 @@ profile_runnable_proc(Process *p, Eterm status){
 
 #ifdef ERTS_SMP
 
-void
-erts_check_my_tracer_proc(Process *p)
-{
-    if (is_internal_pid(ERTS_TRACER_PROC(p))) {
-	Process *tracer = erts_pid2proc(p,
-					ERTS_PROC_LOCK_MAIN,
-					ERTS_TRACER_PROC(p),
-					ERTS_PROC_LOCK_STATUS);
-	int invalid_tracer = (!tracer
-			      || !(ERTS_TRACE_FLAGS(tracer) & F_TRACER));
-	if (tracer)
-	    erts_smp_proc_unlock(tracer, ERTS_PROC_LOCK_STATUS);
-	if (invalid_tracer) {
-	    erts_smp_proc_lock(p, ERTS_PROC_LOCKS_ALL_MINOR);
-	    ERTS_TRACE_FLAGS(p) &= ~TRACEE_FLAGS;
-	    ERTS_TRACER_PROC(p) = NIL;
-	    erts_smp_proc_unlock(p, ERTS_PROC_LOCKS_ALL_MINOR);
-	}
-    }
-}
-
-
 typedef struct ErtsSysMsgQ_ ErtsSysMsgQ;
 struct  ErtsSysMsgQ_ {
     ErtsSysMsgQ *next;
@@ -3182,12 +2170,6 @@ static void
 print_msg_type(ErtsSysMsgQ *smqp)
 {
     switch (smqp->type) {
-    case SYS_MSG_TYPE_TRACE:
-	erts_fprintf(stderr, "TRACE ");
-	break;
-    case SYS_MSG_TYPE_SEQTRACE:
-	erts_fprintf(stderr, "SEQTRACE ");
-	break;
     case SYS_MSG_TYPE_SYSMON:
 	erts_fprintf(stderr, "SYSMON ");
 	break;
@@ -3198,8 +2180,8 @@ print_msg_type(ErtsSysMsgQ *smqp)
 	erts_fprintf(stderr, "ERRLGR ");
 	break;
     case SYS_MSG_TYPE_PROC_MSG:
-	erts_fprintf(stderr, "PROC_MSG ");
-	break;
+       erts_fprintf(stderr, "PROC_MSG ");
+       break;
     default:
 	erts_fprintf(stderr, "??? ");
 	break;
@@ -3211,17 +2193,6 @@ static void
 sys_msg_disp_failure(ErtsSysMsgQ *smqp, Eterm receiver)
 {
     switch (smqp->type) {
-    case SYS_MSG_TYPE_TRACE:
-	/* Invalid tracer_proc's are removed when processes
-	   are scheduled in. */
-	break;
-    case SYS_MSG_TYPE_SEQTRACE:
-	/* Reset seq_tracer if it hasn't changed */
-	erts_smp_rwmtx_rwlock(&sys_trace_rwmtx);
-	if (system_seq_tracer == receiver)
-	    system_seq_tracer = am_false;
-	erts_smp_rwmtx_rwunlock(&sys_trace_rwmtx);
-	break;
     case SYS_MSG_TYPE_SYSMON:
 	if (receiver == NIL
 	    && !erts_system_monitor_long_gc
@@ -3282,7 +2253,7 @@ sys_msg_disp_failure(ErtsSysMsgQ *smqp, Eterm receiver)
 	break;
     }
     case SYS_MSG_TYPE_PROC_MSG:
-	break;
+        break;
     default:
 	ASSERT(0);
     }
@@ -3400,13 +2371,9 @@ sys_msg_dispatcher_func(void *unused)
 	    print_msg_type(smqp);
 #endif
 	    switch (smqp->type) {
-	    case SYS_MSG_TYPE_TRACE:
-	    case SYS_MSG_TYPE_PROC_MSG:
-		receiver = smqp->to;
-		break;
-	    case SYS_MSG_TYPE_SEQTRACE:
-		receiver = erts_get_system_seq_tracer();
-		break;
+            case SYS_MSG_TYPE_PROC_MSG:
+                receiver = smqp->to;
+                break;
 	    case SYS_MSG_TYPE_SYSMON:
 		receiver = erts_get_system_monitor();
 		if (smqp->from == receiver) {
@@ -3441,16 +2408,8 @@ sys_msg_dispatcher_func(void *unused)
 
 	    if (is_internal_pid(receiver)) {
 		proc = erts_pid2proc(NULL, 0, receiver, proc_locks);
-		if (!proc
-		    || (smqp->type == SYS_MSG_TYPE_TRACE
-			&& !(ERTS_TRACE_FLAGS(proc) & F_TRACER))) {
+		if (!proc) {
 		    /* Bad tracer */
-#ifdef DEBUG_PRINTOUTS
-		    if (smqp->type == SYS_MSG_TYPE_TRACE && proc)
-			erts_fprintf(stderr,
-				     "<tracer alive but missing  "
-				     "F_TRACER flag> ");
-#endif
 		    goto failure;
 		}
 		else {
@@ -3458,7 +2417,7 @@ 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,NIL);
+		    erts_queue_message(proc,&proc_locks,mp,smqp->msg);
 #ifdef DEBUG_PRINTOUTS
 		    erts_fprintf(stderr, "delivered\n");
 #endif
@@ -3524,12 +2483,6 @@ erts_foreach_sys_msg_in_q(void (*func)(Eterm,
     for (sm = sys_message_queue; sm; sm = sm->next) {
 	Eterm to;
 	switch (sm->type) {
-	case SYS_MSG_TYPE_TRACE:
-	    to = sm->to;
-	    break;
-	case SYS_MSG_TYPE_SEQTRACE:
-	    to = erts_get_system_seq_tracer();
-	    break;
 	case SYS_MSG_TYPE_SYSMON:
 	    to = erts_get_system_monitor();
 	    break;
@@ -3567,3 +2520,498 @@ init_sys_msg_dispatcher(void)
 }
 
 #endif
+
+#include "erl_nif.h"
+
+struct ErtsTracerNif_ {
+    HashBucket hb;
+    Eterm module;
+    struct erl_module_nif* nif_mod;
+    ErlNifFunc *enabled;
+    ErlNifFunc *trace;
+};
+
+static Hash *tracer_hash = NULL;
+static erts_smp_rwmtx_t tracer_mtx;
+
+static ErtsTracerNif *
+load_tracer_nif(const ErtsTracer tracer)
+{
+    Module* mod = erts_get_module(ERTS_TRACER_MODULE(tracer),
+                                  erts_active_code_ix());
+    struct erl_module_instance *instance;
+    ErlNifFunc *funcs;
+    int num_of_funcs;
+    ErtsTracerNif tnif_tmpl, *tnif;
+    int i;
+
+    if (mod && mod->curr.nif != NULL) {
+        instance = &mod->curr;
+    } else {
+        return NULL;
+    }
+
+    tnif_tmpl.enabled = NULL;
+    tnif_tmpl.trace = NULL;
+    tnif_tmpl.nif_mod = instance->nif;
+    tnif_tmpl.module = ERTS_TRACER_MODULE(tracer);
+
+    num_of_funcs = erts_nif_get_funcs(instance->nif, &funcs);
+
+    for(i = 0; i < num_of_funcs; i++) {
+        if (strcmp("enabled",funcs[i].name) == 0 && funcs[i].arity == 3) {
+            tnif_tmpl.enabled = funcs + i;
+        } else if (strcmp("trace",funcs[i].name) == 0 && funcs[i].arity == 6) {
+            tnif_tmpl.trace = funcs + i;
+        }
+    }
+
+    if (tnif_tmpl.enabled == NULL ||
+        tnif_tmpl.trace   == NULL ) {
+        return NULL;
+    }
+
+    erts_smp_rwmtx_rwlock(&tracer_mtx);
+    tnif = hash_put(tracer_hash, &tnif_tmpl);
+    erts_smp_rwmtx_rwunlock(&tracer_mtx);
+
+    return tnif;
+}
+
+static ERTS_INLINE ErtsTracerNif *
+lookup_tracer_nif(const ErtsTracer tracer)
+{
+    ErtsTracerNif tnif_tmpl;
+    ErtsTracerNif *tnif;
+    tnif_tmpl.module = ERTS_TRACER_MODULE(tracer);
+    erts_smp_rwmtx_rlock(&tracer_mtx);
+    if ((tnif = hash_get(tracer_hash, &tnif_tmpl)) == NULL) {
+        erts_smp_rwmtx_runlock(&tracer_mtx);
+        tnif = load_tracer_nif(tracer);
+        ASSERT(!tnif || tnif->nif_mod);
+        return tnif;
+    }
+    erts_smp_rwmtx_runlock(&tracer_mtx);
+    ASSERT(tnif->nif_mod);
+    return tnif;
+}
+
+/* This function converts an Erlang tracer term to ErtsTracer.
+   It returns THE_NON_VALUE if an invalid tracer term was given.
+   Accepted input is:
+     pid() || port() || {prefix, pid()} || {prefix, port()} ||
+     {prefix, atom(), term()} || {atom(), term()}
+ */
+ErtsTracer
+erts_term_to_tracer(Eterm prefix, Eterm t)
+{
+    ErtsTracer tracer = erts_tracer_nil;
+    ASSERT(is_atom(prefix) || prefix == THE_NON_VALUE);
+    if (!is_nil(t)) {
+        Eterm module = am_erl_tracer, state = THE_NON_VALUE;
+        Eterm hp[2];
+        if (is_tuple(t)) {
+            Eterm *tp = tuple_val(t);
+            if (prefix != THE_NON_VALUE) {
+                if (arityval(tp[0]) == 2 && tp[1] == prefix)
+                    t = tp[2];
+                else if (arityval(tp[0]) == 3 && tp[1] == prefix && is_atom(tp[2])) {
+                    module = tp[2];
+                    state = tp[3];
+                }
+            } else {
+                if (arityval(tp[0]) == 2 && is_atom(tp[2])) {
+                    module = tp[1];
+                    state = tp[2];
+                }
+            }
+        }
+        if (state == THE_NON_VALUE && (is_internal_pid(t) || is_internal_port(t)))
+            state = t;
+        if (state == THE_NON_VALUE)
+            return THE_NON_VALUE;
+        erts_tracer_update(&tracer, CONS(hp, module, state));
+    }
+    if (!lookup_tracer_nif(tracer)) {
+        ASSERT(ERTS_TRACER_MODULE(tracer) != am_erl_tracer);
+        ERTS_TRACER_CLEAR(&tracer);
+        return THE_NON_VALUE;
+    }
+    return tracer;
+}
+
+Eterm
+erts_tracer_to_term(Process *p, ErtsTracer tracer)
+{
+    if (ERTS_TRACER_IS_NIL(tracer))
+        return am_false;
+    if (ERTS_TRACER_MODULE(tracer) == am_erl_tracer)
+        /* Have to manage these specifically in order to be
+           backwards compatible */
+        return ERTS_TRACER_STATE(tracer);
+    else {
+        Eterm *hp = HAlloc(p, 3);
+        return TUPLE2(hp, ERTS_TRACER_MODULE(tracer),
+                      copy_object(ERTS_TRACER_STATE(tracer), p));
+    }
+}
+
+
+static ERTS_INLINE int
+send_to_tracer_nif_raw(Process *c_p, Process *tracee,
+                       const ErtsTracer tracer, Uint tracee_flags,
+                       Eterm t_p_id, ErtsTracerNif *tnif, Eterm tag, Eterm msg,
+                       Eterm extra, Eterm pam_result)
+{
+    if (tnif || (tnif = lookup_tracer_nif(tracer)) != NULL) {
+#define MAP_SIZE 3
+        Eterm argv[6],
+            local_heap[3+MAP_SIZE /* values */+(MAP_SIZE+1 /* keys */)];
+        flatmap_t *map = (flatmap_t*)(local_heap+(MAP_SIZE+1));
+        Eterm *map_values = flatmap_get_values(map);
+
+        int argc = 6;
+
+        argv[0] = tag;
+        argv[1] = ERTS_TRACER_STATE(tracer);
+        argv[2] = t_p_id;
+        argv[3] = msg;
+        argv[4] = extra == THE_NON_VALUE ? am_undefined : extra;
+        argv[5] = make_flatmap(map);
+
+        map->thing_word = MAP_HEADER_FLATMAP;
+        map->size = MAP_SIZE;
+        map->keys = TUPLE3(local_heap, am_match_spec_result, am_scheduler_id,
+                           am_timestamp);
+
+        *map_values++ = pam_result;
+        if (tracee_flags & F_TRACE_SCHED_NO)
+            *map_values++ = make_small(erts_get_scheduler_id());
+        else
+            *map_values++ = am_undefined;
+        if (tracee_flags & F_NOW_TS)
+#ifdef HAVE_ERTS_NOW_CPU
+            if (erts_cpu_timestamp)
+                *map_values++ = am_cpu_timestamp;
+            else
+#endif
+                *map_values++ = am_timestamp;
+        else if (tracee_flags & F_STRICT_MON_TS)
+            *map_values++ = am_strict_monotonic;
+        else if (tracee_flags & F_MON_TS)
+            *map_values++ = am_monotonic;
+        else
+            *map_values++ = am_undefined;
+
+#undef MAP_SIZE
+        erts_nif_call_function(c_p, tracee ? tracee : c_p,
+                               tnif->nif_mod, tnif->trace, argc, argv);
+    }
+    return 1;
+}
+
+
+static ERTS_INLINE int
+send_to_tracer_nif(Process *c_p, ErtsPTabElementCommon *t_p,
+                   Eterm t_p_id, ErtsTracerNif *tnif, Eterm tag,
+                   Eterm msg, Eterm extra)
+{
+#if defined(ERTS_SMP) && defined(ERTS_ENABLE_LOCK_CHECK)
+    if (c_p) {
+        /* We have to hold the main lock of the currently executing process */
+        erts_proc_lc_chk_have_proc_locks(c_p, ERTS_PROC_LOCK_MAIN);
+    }
+    if (is_internal_pid(t_p->id)) {
+        /* We have to have at least one lock */
+        ERTS_SMP_LC_ASSERT(erts_proc_lc_my_proc_locks((Process*)t_p) & ERTS_PROC_LOCKS_ALL);
+    } else {
+        ASSERT(is_internal_port(t_p->id));
+        ERTS_SMP_LC_ASSERT(erts_lc_is_port_locked((Port*)t_p));
+    }
+#endif
+
+    return send_to_tracer_nif_raw(c_p,
+                                  is_internal_pid(t_p->id) ? (Process*)t_p : NULL,
+                                  t_p->tracer, t_p->trace_flags,
+                                  t_p_id, tnif, tag, msg, extra,
+                                  am_true);
+}
+
+static ERTS_INLINE Eterm
+call_enabled_tracer(Process *c_p, const ErtsTracer tracer,
+                    ErtsTracerNif **tnif_ret, Eterm tag, Eterm t_p_id)
+{
+    ErtsTracerNif *tnif = lookup_tracer_nif(tracer);
+    if (tnif) {
+        Eterm argv[] = {tag, ERTS_TRACER_STATE(tracer), t_p_id};
+        if (tnif_ret) *tnif_ret = tnif;
+        return erts_nif_call_function(
+            c_p, NULL, tnif->nif_mod, tnif->enabled, 3, argv);
+    }
+    return am_remove;
+}
+
+static int
+is_tracer_proc_enabled(Process* c_p, ErtsProcLocks c_p_locks,
+                       ErtsPTabElementCommon *t_p,
+                       ErtsTracerNif **tnif_ret, Eterm tag)
+{
+    Eterm nif_result;
+
+#if defined(ERTS_SMP) && defined(ERTS_ENABLE_LOCK_CHECK)
+    if (c_p)
+        ERTS_SMP_LC_ASSERT(erts_proc_lc_my_proc_locks(c_p) == c_p_locks
+                           || erts_thr_progress_is_blocking());
+    if (is_internal_pid(t_p->id)) {
+        /* We have to have at least one lock */
+        ERTS_SMP_LC_ASSERT(erts_proc_lc_my_proc_locks((Process*)t_p) & ERTS_PROC_LOCKS_ALL
+                           || erts_thr_progress_is_blocking());
+    } else {
+        ASSERT(is_internal_port(t_p->id));
+        ERTS_SMP_LC_ASSERT(erts_lc_is_port_locked((Port*)t_p)
+                           || erts_thr_progress_is_blocking());
+    }
+#endif
+
+    nif_result = call_enabled_tracer(c_p, t_p->tracer, tnif_ret, tag, t_p->id);
+    switch (nif_result) {
+    case am_discard: return 0;
+    case am_trace: return 1;
+    case THE_NON_VALUE:
+    case am_remove: break;
+    default:
+        /* only am_remove should be returned, but if
+           something else is returned we fall-through
+           and remove the tracer. */
+        ASSERT(0);
+    }
+
+    /* Only remove tracer on self() and ports */
+    if (is_internal_port(t_p->id) || (c_p && c_p->common.id == t_p->id)) {
+        ErtsProcLocks c_p_xlocks = 0;
+        if (is_internal_pid(t_p->id)) {
+            ERTS_SMP_LC_ASSERT(erts_proc_lc_my_proc_locks(c_p) & ERTS_PROC_LOCK_MAIN);
+            if (c_p_locks != ERTS_PROC_LOCKS_ALL) {
+                c_p_xlocks = ~c_p_locks & ERTS_PROC_LOCKS_ALL;
+                if (erts_smp_proc_trylock(c_p, c_p_xlocks) == EBUSY) {
+                    erts_smp_proc_unlock(c_p, c_p_locks & ~ERTS_PROC_LOCK_MAIN);
+                    erts_smp_proc_lock(c_p, ERTS_PROC_LOCKS_ALL_MINOR);
+                }
+            }
+        }
+        erts_tracer_replace(t_p, erts_tracer_nil);
+        t_p->trace_flags &= ~TRACEE_FLAGS;
+
+        if (c_p_xlocks)
+            erts_smp_proc_unlock(c_p, c_p_xlocks);
+    }
+
+
+    return 0;
+}
+
+int erts_is_tracer_proc_enabled(Process* c_p, ErtsProcLocks c_p_locks,
+                                ErtsPTabElementCommon *t_p, Eterm type)
+{
+    return is_tracer_proc_enabled(c_p, c_p_locks, t_p, NULL, am_trace_status);
+}
+
+int erts_is_tracer_enabled(Process *c_p, const ErtsTracer tracer)
+{
+    ErtsTracerNif *tnif = lookup_tracer_nif(tracer);
+    if (tnif) {
+        Eterm nif_result = call_enabled_tracer(c_p, tracer, &tnif,
+                                               am_trace_status,
+                                               c_p->common.id);
+        switch (nif_result) {
+        case am_discard:
+        case am_trace: return 1;
+        default:
+            break;
+        }
+    }
+    return 0;
+}
+
+void erts_tracer_replace(ErtsPTabElementCommon *t_p, const ErtsTracer tracer)
+{
+#if defined(ERTS_SMP) && defined(ERTS_ENABLE_LOCK_CHECK)
+    if (is_internal_pid(t_p->id) && !erts_thr_progress_is_blocking()) {
+        erts_proc_lc_chk_have_proc_locks((Process*)t_p, ERTS_PROC_LOCKS_ALL);
+    } else if (is_internal_port(t_p->id)) {
+        ERTS_LC_ASSERT(erts_lc_is_port_locked((Port*)t_p)
+                       || erts_thr_progress_is_blocking());
+    }
+#endif
+    if (ERTS_TRACER_COMPARE(t_p->tracer, tracer))
+        return;
+
+    erts_tracer_update(&t_p->tracer, tracer);
+}
+
+static void free_tracer(void *p)
+{
+    ErtsTracer tracer = (ErtsTracer)p;
+
+    if (is_immed(ERTS_TRACER_STATE(tracer))) {
+        erts_free(ERTS_ALC_T_HEAP_FRAG, ptr_val(tracer));
+    } else {
+        ErlHeapFragment *hf = (void*)((char*)(ptr_val(tracer)) - offsetof(ErlHeapFragment, mem));
+        free_message_buffer(hf);
+    }
+}
+
+/* un-define erts_tracer_update before implementation */
+#ifdef erts_tracer_update
+#undef erts_tracer_update
+#endif
+
+/*
+ * ErtsTracer is either NIL, 'true' or [Mod | State]
+ *
+ * - If State is immediate then the memory for
+ *   the cons cell is just two words + sizeof(ErtsThrPrgrLaterOp) large.
+ * - If State is a complex term then the cons cell
+ *   is allocated in an ErlHeapFragment where the cons
+ *   ptr points to the mem field. So in order to get the
+ *   ptr to the fragment you do this:
+ *   (char*)(ptr_val(tracer)) - offsetof(ErlHeapFragment, mem)
+ *   Normally you shouldn't have to care about this though
+ *   as erts_tracer_update takes care of it for you.
+ *
+ *   When ErtsTracer is stored in the stack as part of a
+ *   return trace, the cons cell is stored on the heap of
+ *   the process.
+ *
+ *   The cons cell is not always stored on the heap as:
+ *     1) for port/meta tracing there is no heap
+ *     2) we would need the main lock in order to
+ *        read the tracer which is undesirable.
+ *
+ * One way to optimize this (memory wise) is to keep an refc and only bump
+ * the refc when *tracer is NIL.
+ */
+void
+erts_tracer_update(ErtsTracer *tracer, const ErtsTracer new_tracer)
+{
+    ErlHeapFragment *hf;
+
+    if (is_not_nil(*tracer)) {
+        Uint offs = 2;
+        UWord size = 2 * sizeof(Eterm) + sizeof(ErtsThrPrgrLaterOp);
+        ASSERT(is_list(*tracer));
+        if (is_not_immed(ERTS_TRACER_STATE(*tracer))) {
+            hf = (void*)(((char*)(ptr_val(*tracer)) - offsetof(ErlHeapFragment, mem)));
+            offs = hf->used_size;
+            size = hf->alloc_size * sizeof(Eterm) + sizeof(ErlHeapFragment);
+            ASSERT(offs == size_object(*tracer));
+        }
+        /* We schedule the free:ing of the tracer until after a thread progress
+           has been made so that we know that no schedulers have any references
+           to it. Because we do this, it is possible to release all locks of a
+           process/port and still use the ErtsTracer of that port/process
+           without having to worry if it is free'd.
+        */
+        erts_schedule_thr_prgr_later_cleanup_op(
+            free_tracer, (void*)(*tracer),
+            (ErtsThrPrgrLaterOp*)(ptr_val(*tracer) + offs),
+            size);
+    }
+
+    if (is_nil(new_tracer)) {
+        *tracer = new_tracer;
+    } else if (is_immed(ERTS_TRACER_STATE(new_tracer))) {
+        /* If tracer state is an immediate we only allocate a 2 Eterm heap.
+           Not sure if it is worth it, we save 4 words (sizeof(ErlHeapFragment))
+           per tracer. */
+        Eterm *hp = erts_alloc(ERTS_ALC_T_HEAP_FRAG,
+                               2*sizeof(Eterm) + sizeof(ErtsThrPrgrLaterOp));
+        *tracer = CONS(hp, ERTS_TRACER_MODULE(new_tracer),
+                       ERTS_TRACER_STATE(new_tracer));
+    } else {
+        Eterm *hp, tracer_state = ERTS_TRACER_STATE(new_tracer),
+            tracer_module = ERTS_TRACER_MODULE(new_tracer);
+        Uint sz = size_object(tracer_state);
+        hf = new_message_buffer(sz + 2  /* cons cell */ + (sizeof(ErtsThrPrgrLaterOp)+sizeof(Eterm)-1)/sizeof(Eterm));
+        hp = hf->mem + 2;
+        hf->used_size -= (sizeof(ErtsThrPrgrLaterOp)+sizeof(Eterm)-1)/sizeof(Eterm);
+        *tracer = copy_struct(tracer_state, sz, &hp, &hf->off_heap);
+        *tracer = CONS(hf->mem, tracer_module, *tracer);
+        ASSERT((void*)(((char*)(ptr_val(*tracer)) - offsetof(ErlHeapFragment, mem))) == hf);
+    }
+}
+
+static void init_tracer_nif()
+{
+    erts_smp_rwmtx_opt_t rwmtx_opt = ERTS_SMP_RWMTX_OPT_DEFAULT_INITER;
+    rwmtx_opt.type = ERTS_SMP_RWMTX_TYPE_EXTREMELY_FREQUENT_READ;
+    rwmtx_opt.lived = ERTS_SMP_RWMTX_LONG_LIVED;
+    erts_smp_rwmtx_init_opt(&tracer_mtx, &rwmtx_opt, "tracer_mtx");
+
+    erts_tracer_nif_clear();
+
+}
+
+int erts_tracer_nif_clear()
+{
+
+    erts_smp_rwmtx_rlock(&tracer_mtx);
+    if (!tracer_hash || tracer_hash->nobjs) {
+
+        HashFunctions hf;
+        hf.hash = tracer_hash_fun;
+        hf.cmp = tracer_cmp_fun;
+        hf.alloc = tracer_alloc_fun;
+        hf.free = tracer_free_fun;
+        hf.meta_alloc = (HMALLOC_FUN) erts_alloc;
+        hf.meta_free = (HMFREE_FUN) erts_free;
+        hf.meta_print = (HMPRINT_FUN) erts_print;
+
+        erts_smp_rwmtx_runlock(&tracer_mtx);
+        erts_smp_rwmtx_rwlock(&tracer_mtx);
+
+        if (tracer_hash)
+            hash_delete(tracer_hash);
+
+        tracer_hash = hash_new(ERTS_ALC_T_TRACER_NIF, "tracer_hash", 10, hf);
+
+        erts_smp_rwmtx_rwunlock(&tracer_mtx);
+        return 1;
+    }
+
+    erts_smp_rwmtx_runlock(&tracer_mtx);
+    return 0;
+}
+
+static int tracer_cmp_fun(void* a, void* b)
+{
+    return ((ErtsTracerNif*)a)->module != ((ErtsTracerNif*)b)->module;
+}
+
+static HashValue tracer_hash_fun(void* obj)
+{
+    return make_internal_hash(((ErtsTracerNif*)obj)->module);
+}
+
+static void *tracer_alloc_fun(void* tmpl)
+{
+    ErtsTracerNif *obj = erts_alloc(ERTS_ALC_T_TRACER_NIF,
+                                    sizeof(ErtsTracerNif) +
+                                    sizeof(ErtsThrPrgrLaterOp));
+    memcpy(obj, tmpl, sizeof(*obj));
+    return obj;
+}
+
+static void tracer_free_fun_cb(void* obj)
+{
+    erts_free(ERTS_ALC_T_TRACER_NIF, obj);
+}
+
+static void tracer_free_fun(void* obj)
+{
+    ErtsTracerNif *tnif = obj;
+    erts_schedule_thr_prgr_later_op(
+        tracer_free_fun_cb, obj,
+        (ErtsThrPrgrLaterOp*)(tnif + 1));
+
+}
diff --git a/erts/emulator/beam/erl_trace.h b/erts/emulator/beam/erl_trace.h
index 1da27ee128..177fd373a6 100644
--- a/erts/emulator/beam/erl_trace.h
+++ b/erts/emulator/beam/erl_trace.h
@@ -46,6 +46,8 @@
 #define ERTS_SEQTFLGS2TSTYPE(SEQTFLGS)				\
     ((int) (((SEQTFLGS) >> ERTS_SEQ_TRACE_FLAGS_TS_TYPE_SHIFT)	\
 	    & ERTS_TRACE_TS_TYPE_MASK))
+#define ERTS_SEQTFLGS2TFLGS(SEQTFLGS)                                   \
+    (ERTS_SEQTFLGS2TSTYPE(SEQTFLGS) << ERTS_TRACE_FLAGS_TS_TYPE_SHIFT)
 
 #endif /* ERL_TRACE_H__FLAGS__ */
 
@@ -62,15 +64,19 @@ void erts_system_profile_clear(Process *c_p);
 /* erl_trace.c */
 void erts_init_trace(void);
 void erts_trace_check_exiting(Eterm exiting);
-Eterm erts_set_system_seq_tracer(Process *c_p,
-				 ErtsProcLocks c_p_locks,
-				 Eterm new);
-Eterm erts_get_system_seq_tracer(void);
-void erts_change_default_tracing(int setflags, Uint *flagsp, Eterm *tracerp);
-void erts_get_default_tracing(Uint *flagsp, Eterm *tracerp);
+ErtsTracer erts_set_system_seq_tracer(Process *c_p,
+                                      ErtsProcLocks c_p_locks,
+                                      ErtsTracer new);
+ErtsTracer erts_get_system_seq_tracer(void);
+void erts_change_default_proc_tracing(int setflags, Uint flagsp,
+                                      const ErtsTracer tracerp);
+void erts_get_default_proc_tracing(Uint *flagsp, ErtsTracer *tracerp);
+void erts_change_default_port_tracing(int setflags, Uint flagsp,
+                                      const ErtsTracer tracerp);
+void erts_get_default_port_tracing(Uint *flagsp, ErtsTracer *tracerp);
 void erts_set_system_monitor(Eterm monitor);
 Eterm erts_get_system_monitor(void);
-int erts_is_tracer_proc_valid(Process* p);
+int erts_is_tracer_valid(Process* p);
 
 #ifdef ERTS_SMP
 void erts_check_my_tracer_proc(Process *);
@@ -81,28 +87,32 @@ void erts_foreach_sys_msg_in_q(void (*func)(Eterm,
 					    Eterm,
 					    ErlHeapFragment *));
 void erts_queue_error_logger_message(Eterm, Eterm, ErlHeapFragment *);
+void erts_send_sys_msg_proc(Eterm, Eterm, Eterm, ErlHeapFragment *);
 #endif
 
-void erts_send_sys_msg_proc(Eterm, Eterm, Eterm, ErlHeapFragment *);
 void trace_send(Process*, Eterm, Eterm);
-void trace_receive(Process*, Eterm);
-Uint32 erts_call_trace(Process *p, BeamInstr mfa[], struct binary *match_spec, Eterm* args,
-		       int local, Eterm *tracer_pid);
-void erts_trace_return(Process* p, BeamInstr* fi, Eterm retval, Eterm *tracer_pid);
+void trace_receive(Process *, Eterm);
+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,
+                       ErtsTracer *tracer);
 void erts_trace_exception(Process* p, BeamInstr mfa[], Eterm class, Eterm value,
-			  Eterm *tracer);
+                          ErtsTracer *tracer);
 void erts_trace_return_to(Process *p, BeamInstr *pc);
-void trace_sched(Process*, Eterm);
-void trace_proc(Process*, Process*, Eterm, Eterm);
-void trace_proc_spawn(Process*, Eterm pid, Eterm mod, Eterm func, Eterm args);
+void trace_sched(Process*, ErtsProcLocks, Eterm);
+void trace_proc(Process*, ErtsProcLocks, Process*, Eterm, Eterm);
+void trace_proc_spawn(Process*, Eterm what, Eterm pid, Eterm mod, Eterm func, Eterm args);
 void save_calls(Process *p, Export *);
 void trace_gc(Process *p, Eterm what);
 /* port tracing */
-void trace_virtual_sched(Process*, Eterm);
+void trace_virtual_sched(Process*, ErtsProcLocks, Eterm);
 void trace_sched_ports(Port *pp, Eterm);
 void trace_sched_ports_where(Port *pp, Eterm, Eterm);
 void trace_port(Port *, Eterm what, Eterm data);
 void trace_port_open(Port *, Eterm calling_pid, Eterm drv_name);
+void trace_port_receive(Port *, Eterm calling_pid, Eterm tag, ...);
+void trace_port_send(Port *, Eterm to, Eterm msg, int exists);
+void trace_port_send_binary(Port *, Eterm to, Eterm what, char *bin, Sint sz);
 
 /* system_profile */
 void erts_set_system_profile(Eterm profile);
@@ -121,7 +131,7 @@ void monitor_large_heap(Process *p);
 void monitor_generic(Process *p, Eterm type, Eterm spec);
 Uint erts_trace_flag2bit(Eterm flag);
 int erts_trace_flags(Eterm List, 
-		 Uint *pMask, Eterm *pTracer, int *pCpuTimestamp);
+		 Uint *pMask, ErtsTracer *pTracer, int *pCpuTimestamp);
 Eterm erts_bif_trace(int bif_index, Process* p, Eterm* args, BeamInstr *I);
 
 #ifdef ERTS_SMP
@@ -161,15 +171,53 @@ int erts_set_trace_pattern(Process*p, Eterm* mfa, int specified,
 			   struct binary* match_prog_set,
 			   struct binary *meta_match_prog_set,
 			   int on, struct trace_pattern_flags,
-			   Eterm meta_tracer_pid, int is_blocking);
+			   ErtsTracer meta_tracer, int is_blocking);
 void
 erts_get_default_trace_pattern(int *trace_pattern_is_on,
 			       struct binary **match_spec,
 			       struct binary **meta_match_spec,
 			       struct trace_pattern_flags *trace_pattern_flags,
-			       Eterm *meta_tracer_pid);
+			       ErtsTracer *meta_tracer);
 int erts_is_default_trace_enabled(void);
 void erts_bif_trace_init(void);
 int erts_finish_breakpointing(void);
 
+/* Nif tracer functions */
+int erts_is_tracer_proc_enabled(Process *c_p, ErtsProcLocks c_p_locks,
+                                ErtsPTabElementCommon *t_p, Eterm type);
+int erts_is_tracer_enabled(Process *c_p, const ErtsTracer tracer);
+Eterm erts_tracer_to_term(Process *p, ErtsTracer tracer);
+ErtsTracer erts_term_to_tracer(Eterm prefix, Eterm term);
+void erts_tracer_replace(ErtsPTabElementCommon *t_p,
+                         const ErtsTracer new_tracer);
+void erts_tracer_update(ErtsTracer *tracer, const ErtsTracer new_tracer);
+int erts_tracer_nif_clear(void);
+
+#define erts_tracer_update(t,n) do { if (*(t) != (n)) erts_tracer_update(t,n); } while(0)
+#define ERTS_TRACER_CLEAR(t) erts_tracer_update(t, erts_tracer_nil)
+
+static const ErtsTracer
+ERTS_DECLARE_DUMMY(erts_tracer_true) = am_true;
+
+static const ErtsTracer
+ERTS_DECLARE_DUMMY(erts_tracer_nil) = NIL;
+
+#define ERTS_TRACER_COMPARE(t1, t2)                     \
+    (EQ((t1), (t2)))
+
+#define ERTS_TRACER_IS_NIL(t1) ERTS_TRACER_COMPARE(t1, erts_tracer_nil)
+
+#define IS_TRACER_VALID(tracer)                                         \
+    (ERTS_TRACER_COMPARE(tracer,erts_tracer_true)                       \
+     || ERTS_TRACER_IS_NIL(tracer)                                      \
+     || (is_list(tracer) && is_atom(CAR(list_val(tracer)))))
+
+#define ERTS_TRACER_FROM_ETERM(termp) \
+    ((ErtsTracer*)(termp))
+
+#define ERTS_TRACER_PROC_IS_ENABLED(PROC)                               \
+    (!ERTS_TRACER_IS_NIL(ERTS_TRACER(PROC))                             \
+     && erts_is_tracer_proc_enabled(PROC, ERTS_PROC_LOCK_MAIN,          \
+                                    &(PROC)->common, am_trace_status))
+
 #endif /* ERL_TRACE_H__ */
diff --git a/erts/emulator/beam/erlang_dtrace.d b/erts/emulator/beam/erlang_dtrace.d
index 73ef5a108a..237889e0f5 100644
--- a/erts/emulator/beam/erlang_dtrace.d
+++ b/erts/emulator/beam/erlang_dtrace.d
@@ -700,6 +700,35 @@ provider erlang {
      */
     probe efile_drv__return(int, int, char *, int, int, int);
 
+
+/*
+ * The set of probes called by the erlang tracer nif backend. In order
+ * to receive events on these you both have to enable tracing in erlang
+ * using the trace bifs and also from dtrace/systemtap.
+ */
+
+
+    /**
+     * A trace message of type `event` was triggered by process `p`.
+     *
+     *
+     * @param p the PID (string form) of the process
+     * @param event the event that was triggered (i.e. call or spawn)
+     * @param state the state of the tracer nif as a string
+     * @param arg1 first argument to the trace event
+     * @param arg2 second argument to the trace event
+     */
+    probe trace(char *p, char *event, char *state, char *arg1, char *arg2);
+
+    /**
+     * A sequence trace message of type `label` was triggered.
+     *
+     * @param state the state of the tracer nif as a string
+     * @param label the seq trace label
+     * @param seq_info the seq trace info tuple as a string
+     */
+    probe trace_seq(char *state, char *label, char *seq_info);
+
 /*
  * NOTE:
  * For formatting int64_t arguments within a D script, see:
diff --git a/erts/emulator/beam/external.c b/erts/emulator/beam/external.c
index 9f43240b7e..723c25ff77 100644
--- a/erts/emulator/beam/external.c
+++ b/erts/emulator/beam/external.c
@@ -3958,9 +3958,13 @@ error:
      * Must unlink all off-heap objects that may have been
      * linked into the process. 
      */
-    if (factory->hp < hp) { /* Sometimes we used hp and sometimes factory->hp */
-	factory->hp = hp;   /* the largest must be the freshest */
+    if (factory->mode != FACTORY_CLOSED) {
+	if (factory->hp < hp) { /* Sometimes we used hp and sometimes factory->hp */
+	    factory->hp = hp;   /* the largest must be the freshest */
+	}
     }
+    else ASSERT(factory->hp == hp);
+
 error_hamt:
     erts_factory_undo(factory);
     PSTACK_DESTROY(hamt_array);
diff --git a/erts/emulator/beam/global.h b/erts/emulator/beam/global.h
index 1d3704f0af..a7bc990deb 100644
--- a/erts/emulator/beam/global.h
+++ b/erts/emulator/beam/global.h
@@ -44,6 +44,8 @@
 #include "erl_port.h"
 #include "erl_gc.h"
 
+struct enif_func_t;
+
 struct enif_environment_t /* ErlNifEnv */
 {
     struct erl_module_nif* mod_nif;
@@ -54,14 +56,21 @@ struct enif_environment_t /* ErlNifEnv */
     int fpe_was_unmasked;
     struct enif_tmp_obj_t* tmp_obj_list;
     int exception_thrown; /* boolean */
+    Process *tracee;
 };
 extern void erts_pre_nif(struct enif_environment_t*, Process*,
-			 struct erl_module_nif*);
+			 struct erl_module_nif*, Process* tracee);
 extern void erts_post_nif(struct enif_environment_t* env);
 extern Eterm erts_nif_taints(Process* p);
 extern void erts_print_nif_taints(int to, void* to_arg);
 void erts_unload_nif(struct erl_module_nif* nif);
 extern void erl_nif_init(void);
+extern int erts_nif_get_funcs(struct erl_module_nif*,
+                              struct enif_func_t **funcs);
+extern Eterm erts_nif_call_function(Process *p, Process *tracee,
+                                    struct erl_module_nif*,
+                                    struct enif_func_t *,
+                                    int argc, Eterm *argv);
 
 /* Driver handle (wrapper for old plain handle) */
 #define ERL_DE_OK      0
@@ -1532,10 +1541,15 @@ ERTS_GLB_INLINE void dtrace_fun_decode(Process *process,
 ERTS_GLB_INLINE void
 dtrace_pid_str(Eterm pid, char *process_buf)
 {
-    erts_snprintf(process_buf, DTRACE_TERM_BUF_SIZE, "<%lu.%lu.%lu>",
-                  pid_channel_no(pid),
-                  pid_number(pid),
-                  pid_serial(pid));
+    if (is_pid(pid))
+        erts_snprintf(process_buf, DTRACE_TERM_BUF_SIZE, "<%lu.%lu.%lu>",
+                      pid_channel_no(pid),
+                      pid_number(pid),
+                      pid_serial(pid));
+    else if (is_port(pid))
+        erts_snprintf(process_buf, DTRACE_TERM_BUF_SIZE, "#Port<%lu.%lu>",
+                      port_channel_no(pid),
+                      port_number(pid));
 }
 
 ERTS_GLB_INLINE void
@@ -1547,9 +1561,7 @@ dtrace_proc_str(Process *process, char *process_buf)
 ERTS_GLB_INLINE void
 dtrace_port_str(Port *port, char *port_buf)
 {
-    erts_snprintf(port_buf, DTRACE_TERM_BUF_SIZE, "#Port<%lu.%lu>",
-                  port_channel_no(port->common.id),
-                  port_number(port->common.id));
+    dtrace_pid_str(port->common.id, port_buf);
 }
 
 ERTS_GLB_INLINE void
diff --git a/erts/emulator/beam/io.c b/erts/emulator/beam/io.c
index de73da8e22..b14ca77a04 100644
--- a/erts/emulator/beam/io.c
+++ b/erts/emulator/beam/io.c
@@ -88,7 +88,7 @@ int erts_port_parallelism = 0;
 static erts_atomic64_t bytes_in;
 static erts_atomic64_t bytes_out;
 
-static void deliver_result(Eterm sender, Eterm pid, Eterm res);
+static void deliver_result(Port *p, Eterm sender, Eterm pid, Eterm res);
 static int init_driver(erts_driver_t *, ErlDrvEntry *, DE_Handle *);
 static void terminate_port(Port *p);
 static void pdl_init(void);
@@ -212,6 +212,7 @@ static ERTS_INLINE void
 kill_port(Port *pp)
 {
     ERTS_SMP_LC_ASSERT(erts_lc_is_port_locked(pp));
+    ERTS_TRACER_CLEAR(&ERTS_TRACER(pp));
     erts_ptab_delete_element(&erts_port, &pp->common); /* Time of death */
     erts_port_task_free_port(pp);
     /* In non-smp case the port structure may have been deallocated now */
@@ -404,7 +405,7 @@ static Port *create_port(char *name,
     prt->os_pid = -1;
 
     /* Set default tracing */
-    erts_get_default_tracing(&ERTS_TRACE_FLAGS(prt), &ERTS_TRACER_PROC(prt));
+    erts_get_default_port_tracing(&ERTS_TRACE_FLAGS(prt), &ERTS_TRACER(prt));
 
     ERTS_CT_ASSERT(offsetof(Port,common) == 0);
 
@@ -709,7 +710,7 @@ erts_open_driver(erts_driver_t* driver,	/* Pointer to driver. */
     if (driver->start) {
         ERTS_MSACC_PUSH_STATE_M();
 	if (IS_TRACED_FL(port, F_TRACE_SCHED_PORTS)) {
-	    trace_sched_ports_where(port, am_in, am_start);
+	    trace_sched_ports_where(port, am_in, am_open);
 	}
 	port->caller = pid;
 #ifdef USE_VM_PROBES
@@ -753,7 +754,7 @@ erts_open_driver(erts_driver_t* driver,	/* Pointer to driver. */
 	ERTS_MSACC_POP_STATE_M();
 	port->caller = NIL;
 	if (IS_TRACED_FL(port, F_TRACE_SCHED_PORTS)) {
-	    trace_sched_ports_where(port, am_out, am_start);
+	    trace_sched_ports_where(port, am_out, am_open);
 	}
 #ifdef ERTS_SMP
 	if (port->xports)
@@ -1300,7 +1301,7 @@ try_imm_drv_call(ErtsTryImmDrvCallState *sp)
 	reds_left_in = CONTEXT_REDS/10;
     else {
 	if (IS_TRACED_FL(c_p, F_TRACE_SCHED_PROCS))
-	    trace_virtual_sched(c_p, am_out);
+	    trace_sched(c_p, ERTS_PROC_LOCK_MAIN, am_out);
 	/*
 	 * No status lock held while sending runnable
 	 * proc trace messages. It is however not needed
@@ -1384,7 +1385,7 @@ finalize_imm_drv_call(ErtsTryImmDrvCallState *sp)
 	}
 
 	if (IS_TRACED_FL(c_p, F_TRACE_SCHED_PROCS))
-	    trace_virtual_sched(c_p, am_in);
+	    trace_sched(c_p, ERTS_PROC_LOCK_MAIN, am_in);
 	/*
 	 * No status lock held while sending runnable
 	 * proc trace messages. It is however not needed
@@ -1447,7 +1448,7 @@ queue_port_sched_op_reply(Process *rp,
 
     erts_factory_trim_and_close(factory, &msg, 1);
 
-    erts_queue_message(rp, rp_locksp, factory->message, msg, NIL);
+    erts_queue_message(rp, rp_locksp, factory->message, msg);
 }
 
 static void
@@ -1535,9 +1536,8 @@ erts_schedule_proc2port_signal(Process *c_p,
 	ERTS_SMP_MSGQ_MV_INQ2PRIVQ(c_p);
 	c_p->msg.save = c_p->msg.last;
 
-	erts_smp_proc_unlock(c_p,
-			     (ERTS_PROC_LOCK_MAIN
-			      | ERTS_PROC_LOCKS_MSG_RECEIVE));
+        erts_smp_proc_unlock(c_p, (ERTS_PROC_LOCKS_MSG_RECEIVE
+                                   | ERTS_PROC_LOCK_MAIN));
     }
 
 
@@ -1742,6 +1742,10 @@ call_driver_outputv(int bang_op,
 	ERTS_SMP_LC_ASSERT(erts_lc_is_port_locked(prt)	
 			   || ERTS_IS_CRASH_DUMPING);
 
+
+        if (IS_TRACED_FL(prt, F_TRACE_RECEIVE))
+            trace_port_receive(prt, caller, am_commandv, evp);
+
 #ifdef USE_VM_PROBES
 	if (DTRACE_ENABLED(driver_outputv)) {
 	    DTRACE_FORMAT_COMMON_PID_AND_PORT(caller, prt);
@@ -1868,6 +1872,9 @@ call_driver_output(int bang_op,
         }
 #endif
 
+        if (IS_TRACED_FL(prt, F_TRACE_RECEIVE))
+            trace_port_receive(prt, caller, am_command, bufp, size);
+
 	prt->caller = caller;
 	(*drv->output)((ErlDrvData) prt->drv_data, bufp, size);
 	prt->caller = NIL;
@@ -2589,7 +2596,10 @@ call_deliver_port_exit(int bang_op,
 	    return ERTS_PORT_OP_DROPPED;
     }
 
-    if (!erts_deliver_port_exit(prt, from, reason, bang_op))
+    if (IS_TRACED_FL(prt, F_TRACE_RECEIVE))
+        trace_port_receive(prt, from, am_close);
+
+    if (!erts_deliver_port_exit(prt, from, reason, bang_op, broken_link))
 	return ERTS_PORT_OP_DROPPED;
 
 #ifdef USE_VM_PROBES
@@ -2660,13 +2670,13 @@ erts_port_exit(Process *c_p,
 					       ERTS_PORT_SFLGS_INVALID_LOOKUP,
 					       0,
 					       !refp,
-					       am_exit);
+					       am_close);
 
 
 	switch (try_imm_drv_call(&try_call_state)) {
 	case ERTS_TRY_IMM_DRV_CALL_OK: {
 	    res = call_deliver_port_exit(flags & ERTS_PORT_SIG_FLG_BANG_OP,
-					 from,
+					 c_p ? c_p->common.id : from,
 					 prt,
 					 try_call_state.state,
 					 reason,
@@ -2745,8 +2755,11 @@ set_port_connected(int bang_op,
 	    return ERTS_PORT_OP_DROPPED;
 	}
 
+        if (IS_TRACED_FL(prt, F_TRACE_RECEIVE))
+            trace_port_receive(prt, from, am_connect, connect);
+
 	ERTS_PORT_SET_CONNECTED(prt, connect);
-	deliver_result(prt->common.id, from, am_connected);
+	deliver_result(prt, prt->common.id, from, am_connected);
 
 #ifdef USE_VM_PROBES
 	if(DTRACE_ENABLED(port_command)) {
@@ -2768,10 +2781,22 @@ set_port_connected(int bang_op,
 	erts_add_link(&ERTS_P_LINKS(rp), LINK_PID, prt->common.id);
 	erts_add_link(&ERTS_P_LINKS(prt), LINK_PID, connect);
 
+        if (IS_TRACED_FL(rp, F_TRACE_PROCS))
+            trace_proc(NULL, 0, rp, am_getting_linked, prt->common.id);
+
 	ERTS_PORT_SET_CONNECTED(prt, connect);
 
 	erts_smp_proc_unlock(rp, ERTS_PROC_LOCK_LINK);
 
+        if (IS_TRACED_FL(prt, F_TRACE_PORTS))
+            trace_port(prt, am_getting_linked, connect);
+        if (IS_TRACED_FL(prt, F_TRACE_RECEIVE))
+            trace_port_receive(prt, from, am_connect, connect);
+        if (IS_TRACED_FL(prt, F_TRACE_SEND)) {
+            Eterm hp[3];
+            trace_port_send(prt, from, TUPLE2(hp, prt->common.id, am_connected), 1);
+        }
+
 #ifdef USE_VM_PROBES
 	if (DTRACE_ENABLED(port_connect)) {
 	    DTRACE_CHARBUF(process_str, DTRACE_TERM_BUF_SIZE);
@@ -2874,8 +2899,11 @@ static void
 port_unlink(Port *prt, Eterm from)
 {
     ErtsLink *lnk = erts_remove_link(&ERTS_P_LINKS(prt), from);
-    if (lnk)
+    if (lnk) {
+        if (IS_TRACED_FL(prt, F_TRACE_PORTS))
+            trace_port(prt, am_getting_unlinked, from);
 	erts_destroy_link(lnk);
+    }
 }
 
 static int
@@ -2945,10 +2973,10 @@ port_link_failure(Eterm port_id, Eterm linker)
 					     NIL,
 					     NULL,
 					     0);
-	    if (xres >= 0 && IS_TRACED_FL(rp, F_TRACE_PROCS)) {
+	    if (xres >= 0) {
 		/* We didn't exit the process and it is traced */
 		if (IS_TRACED_FL(rp, F_TRACE_PROCS))
-		    trace_proc(NULL, rp, am_getting_unlinked, port_id);
+		    trace_proc(NULL, 0, rp, am_getting_unlinked, port_id);
 	    }
 	    if (rp_locks)
 		erts_smp_proc_unlock(rp, rp_locks);
@@ -2959,10 +2987,15 @@ port_link_failure(Eterm port_id, Eterm linker)
 static void
 port_link(Port *prt, erts_aint32_t state, Eterm to)
 {
-    if (!(state & ERTS_PORT_SFLGS_INVALID_LOOKUP))
+    if (IS_TRACED_FL(prt, F_TRACE_PORTS))
+        trace_port(prt, am_getting_linked, to);
+    if (!(state & ERTS_PORT_SFLGS_INVALID_LOOKUP)) {
 	erts_add_link(&ERTS_P_LINKS(prt), LINK_PID, to);
-    else
+    } else {
 	port_link_failure(prt->common.id, to);
+        if (IS_TRACED_FL(prt, F_TRACE_PORTS))
+            trace_port(prt, am_unlink, to);
+    }
 }
 
 static int
@@ -2970,8 +3003,9 @@ port_sig_link(Port *prt, erts_aint32_t state, int op, ErtsProc2PortSigData *sigd
 {
     if (op == ERTS_PROC2PORT_SIG_EXEC)
 	port_link(prt, state, sigdp->u.link.to);
-    else
+    else {
 	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);
     return ERTS_PORT_REDS_LINK;
@@ -3391,7 +3425,7 @@ static int read_linebuf(LineBufContext *bp)
 }
 
 static void
-deliver_result(Eterm sender, Eterm pid, Eterm res)
+deliver_result(Port *prt, Eterm sender, Eterm pid, Eterm res)
 {
     Process *rp;
     ErtsProcLocks rp_locks = 0;
@@ -3399,12 +3433,22 @@ deliver_result(Eterm sender, Eterm pid, Eterm res)
 
     ERTS_SMP_CHK_NO_PROC_LOCKS;
 
+    ASSERT(!prt || prt->common.id == sender);
+#ifdef ERTS_SMP
+    ASSERT(!prt || erts_lc_is_port_locked(prt));
+#endif
+
     ASSERT(is_internal_port(sender) && is_internal_pid(pid));
 
     rp = (scheduler
 	  ? erts_proc_lookup(pid)
 	  : erts_pid2proc_opt(NULL, 0, pid, 0, ERTS_P2P_FLG_INC_REFC));
 
+    if (prt && IS_TRACED_FL(prt, F_TRACE_SEND)) {
+        Eterm hp[3];
+        trace_port_send(prt, pid, TUPLE2(hp, sender, res), !!rp);
+    }
+
     if (rp) {
 	Eterm tuple;
 	ErtsMessage *mp;
@@ -3417,7 +3461,7 @@ deliver_result(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, NIL);
+	erts_queue_message(rp, &rp_locks, mp, tuple);
 
 	if (rp_locks)
 	    erts_smp_proc_unlock(rp, rp_locks);
@@ -3449,6 +3493,7 @@ static void deliver_read_message(Port* prt, erts_aint32_t state, Eterm to,
     ErlOffHeap *ohp;
     ErtsProcLocks rp_locks = 0;
     int scheduler = erts_get_scheduler_id() != 0;
+    int trace_send = IS_TRACED_FL(prt, F_TRACE_SEND);
 
     ERTS_SMP_LC_ASSERT(erts_lc_is_port_locked(prt));    
     ERTS_SMP_CHK_NO_PROC_LOCKS;
@@ -3471,7 +3516,7 @@ static void deliver_read_message(Port* prt, erts_aint32_t state, Eterm to,
     if (!rp)
 	return;
 
-    mp = erts_alloc_message_heap(rp, &rp_locks, need, &hp, &ohp);
+    mp = erts_alloc_message_heap(trace_send ? NULL : rp, &rp_locks, need, &hp, &ohp);
 
     listp = NIL;
     if ((state & ERTS_PORT_SFLG_BINARY_IO) == 0) {
@@ -3513,7 +3558,11 @@ static void deliver_read_message(Port* prt, erts_aint32_t state, Eterm to,
     tuple = TUPLE2(hp, prt->common.id, tuple);
     hp += 3;
 
-    erts_queue_message(rp, &rp_locks, mp, tuple, am_undefined);
+    if (trace_send)
+        trace_port_send(prt, to, tuple, 1);
+
+    ERL_MESSAGE_TOKEN(mp) = am_undefined;
+    erts_queue_message(rp, &rp_locks, mp, tuple);
     if (rp_locks)
 	erts_smp_proc_unlock(rp, rp_locks);
     if (!scheduler)
@@ -3592,6 +3641,7 @@ deliver_vec_message(Port* prt,			/* Port */
     ErtsProcLocks rp_locks = 0;
     int scheduler = erts_get_scheduler_id() != 0;
     erts_aint32_t state;
+    int trace_send = IS_TRACED_FL(prt, F_TRACE_SEND);
 
     ERTS_SMP_LC_ASSERT(erts_lc_is_port_locked(prt));
     ERTS_SMP_CHK_NO_PROC_LOCKS;
@@ -3619,7 +3669,7 @@ deliver_vec_message(Port* prt,			/* Port */
 	need += (hlen+csize)*2;
     }
 
-    mp = erts_alloc_message_heap(rp, &rp_locks, need, &hp, &ohp);
+    mp = erts_alloc_message_heap(trace_send ? NULL : rp, &rp_locks, need, &hp, &ohp);
 
     listp = NIL;
     iov += vsize;
@@ -3680,7 +3730,11 @@ deliver_vec_message(Port* prt,			/* Port */
     tuple = TUPLE2(hp, prt->common.id, tuple);
     hp += 3;
 
-    erts_queue_message(rp, &rp_locks, mp, tuple, am_undefined);
+    if (IS_TRACED_FL(prt, F_TRACE_SEND))
+        trace_port_send(prt, to, tuple, 1);
+
+    ERL_MESSAGE_TOKEN(mp) = am_undefined;
+    erts_queue_message(rp, &rp_locks, mp, tuple);
     erts_smp_proc_unlock(rp, rp_locks);
     if (!scheduler)
 	erts_proc_dec_refc(rp);
@@ -3820,6 +3874,11 @@ terminate_port(Port *prt)
 	ASSERT(!prt->xports);
 #endif
     }
+
+    if (is_internal_port(send_closed_port_id)
+        && IS_TRACED_FL(prt, F_TRACE_SEND))
+        trace_port_send(prt, connected_id, am_closed, 1);
+
     if(drv->handle != NULL) {
 	erts_smp_rwmtx_rlock(&erts_driver_list_lock);
 	erts_ddll_decrement_port_count(drv->handle); 
@@ -3851,7 +3910,7 @@ terminate_port(Port *prt)
 	erts_flush_async_exit(erts_halt_code, "");
     }
     if (is_internal_port(send_closed_port_id))
-	deliver_result(send_closed_port_id, connected_id, am_closed);
+	deliver_result(NULL, send_closed_port_id, connected_id, am_closed);
 }
 
 void
@@ -3884,7 +3943,7 @@ static void sweep_one_monitor(ErtsMonitor *mon, void *vpsc)
 
 
 typedef struct {
-    Eterm port;
+    Port *port;
     Eterm reason;
 } SweepContext;
 
@@ -3893,10 +3952,13 @@ static void sweep_one_link(ErtsLink *lnk, void *vpsc)
     SweepContext *psc = vpsc;
     DistEntry *dep;
     Process *rp;
-    
+    Eterm port_id = psc->port->common.id;
 
     ASSERT(lnk->type == LINK_PID);
-    
+
+    if (IS_TRACED_FL(psc->port, F_TRACE_PORTS))
+        trace_port(psc->port, am_unlink, lnk->pid);
+
     if (is_external_pid(lnk->pid)) {
 	dep = external_pid_dist_entry(lnk->pid);
 	if(dep != erts_this_dist_entry) {
@@ -3909,9 +3971,9 @@ static void sweep_one_link(ErtsLink *lnk, void *vpsc)
 	    case ERTS_DSIG_PREP_NOT_CONNECTED:
 		break;
 	    case ERTS_DSIG_PREP_CONNECTED:
-		erts_remove_dist_link(&dld, psc->port, lnk->pid, dep);
+		erts_remove_dist_link(&dld, port_id, lnk->pid, dep);
 		erts_destroy_dist_link(&dld);
-		code = erts_dsig_send_exit(&dsd, psc->port, lnk->pid,
+		code = erts_dsig_send_exit(&dsd, port_id, lnk->pid,
 					   psc->reason);
 		ASSERT(code == ERTS_DSIG_SEND_OK);
 		break;
@@ -3925,23 +3987,25 @@ static void sweep_one_link(ErtsLink *lnk, void *vpsc)
 	ASSERT(is_internal_pid(lnk->pid));
 	rp = erts_pid2proc(NULL, 0, lnk->pid, rp_locks);
 	if (rp) {
-	    ErtsLink *rlnk = erts_remove_link(&ERTS_P_LINKS(rp), psc->port);
+	    ErtsLink *rlnk = erts_remove_link(&ERTS_P_LINKS(rp), port_id);
 
 	    if (rlnk) {
 		int xres = erts_send_exit_signal(NULL,
-						 psc->port,
+						 port_id,
 						 rp,
 						 &rp_locks, 
 						 psc->reason,
 						 NIL,
 						 NULL,
 						 0);
-		if (xres >= 0 && IS_TRACED_FL(rp, F_TRACE_PROCS)) {
+		if (xres >= 0) {
+                    if (rp_locks & ERTS_PROC_LOCKS_XSIG_SEND) {
+                        erts_smp_proc_unlock(rp, ERTS_PROC_LOCKS_XSIG_SEND);
+                        rp_locks &= ~ERTS_PROC_LOCKS_XSIG_SEND;
+                    }
 		    /* We didn't exit the process and it is traced */
-		    if (IS_TRACED_FL(rp, F_TRACE_PROCS)) {
-			trace_proc(NULL, rp, am_getting_unlinked,
-				   psc->port);
-		    }
+		    if (IS_TRACED_FL(rp, F_TRACE_PROCS))
+			trace_proc(NULL, 0, rp, am_getting_unlinked, port_id);
 		}
 		erts_destroy_link(rlnk);
 	    }
@@ -3963,7 +4027,8 @@ static void sweep_one_link(ErtsLink *lnk, void *vpsc)
  */
 
 int
-erts_deliver_port_exit(Port *p, Eterm from, Eterm reason, int send_closed)
+erts_deliver_port_exit(Port *p, Eterm from, Eterm reason, int send_closed,
+                       int drop_normal)
 {
    ErtsLink *lnk;
    Eterm rreason;
@@ -3993,8 +4058,10 @@ erts_deliver_port_exit(Port *p, Eterm from, Eterm reason, int send_closed)
 		| ERTS_PORT_SFLG_CLOSING))
        return 0;
 
-   if (reason == am_normal && from != ERTS_PORT_GET_CONNECTED(p) && from != p->common.id)
+   if (reason == am_normal && from != ERTS_PORT_GET_CONNECTED(p)
+       && from != p->common.id && drop_normal) {
        return 0;
+   }
 
    set_state_flags = ERTS_PORT_SFLG_EXITING;
    if (send_closed)
@@ -4005,9 +4072,8 @@ erts_deliver_port_exit(Port *p, Eterm from, Eterm reason, int send_closed)
    state = erts_atomic32_read_bor_mb(&p->state, set_state_flags);
    state |= set_state_flags;
 
-   if (IS_TRACED_FL(p, F_TRACE_PORTS)) {
+   if (IS_TRACED_FL(p, F_TRACE_PORTS))
    	trace_port(p, am_closed, reason);
-   }
 
    erts_trace_check_exiting(p->common.id);
 
@@ -4017,7 +4083,7 @@ erts_deliver_port_exit(Port *p, Eterm from, Eterm reason, int send_closed)
        (void) erts_unregister_name(NULL, 0, p, p->common.u.alive.reg->name);
 
    {
-       SweepContext sc = {p->common.id, rreason};
+       SweepContext sc = {p, rreason};
        lnk = ERTS_P_LINKS(p);
        ERTS_P_LINKS(p) = NULL;
        erts_sweep_links(lnk, &sweep_one_link, &sc);
@@ -4139,7 +4205,10 @@ call_driver_control(Eterm caller,
                 command, size);
     }
 #endif
-    
+
+    if (IS_TRACED_FL(prt, F_TRACE_RECEIVE))
+        trace_port_receive(prt, caller, am_control, command, bufp, size);
+
     ERTS_MSACC_SET_STATE_CACHED_M(ERTS_MSACC_STATE_PORT);
 
 #ifdef USE_LTTNG_VM_TRACEPOINTS
@@ -4166,6 +4235,9 @@ call_driver_control(Eterm caller,
     if (cres < 0)
 	return ERTS_PORT_OP_BADARG;
 
+    if (IS_TRACED_FL(prt, F_TRACE_SEND))
+        trace_port_send_binary(prt, caller, am_control, *resp_bufp, cres);
+
     *from_size = (ErlDrvSizeT) cres;
 
     return ERTS_PORT_OP_DONE;
@@ -4575,6 +4647,9 @@ call_driver_call(Eterm caller,
     }
 #endif
 
+    if (IS_TRACED_FL(prt, F_TRACE_RECEIVE))
+        trace_port_receive(prt, caller, am_call, command, bufp, size);
+
     ERTS_MSACC_SET_STATE_CACHED_M(ERTS_MSACC_STATE_PORT);
 
     prt->caller = caller;
@@ -4593,6 +4668,9 @@ call_driver_call(Eterm caller,
 	|| ((byte) (*resp_bufp)[0]) != VERSION_MAGIC)
 	return ERTS_PORT_OP_BADARG;
 
+    if (IS_TRACED_FL(prt, F_TRACE_SEND))
+        trace_port_send_binary(prt, caller, am_call, *resp_bufp, cres);
+
     *from_size = (ErlDrvSizeT) cres;
 
     return ERTS_PORT_OP_DONE;
@@ -5036,7 +5114,8 @@ reply_io_bytes(void *vreq)
 	eout = erts_bld_uint64(&hp, NULL, out);
 
 	msg = TUPLE4(hp, ref, make_small(sched_id), ein, eout);
-	erts_queue_message(rp, &rp_locks, mp, msg, NIL);
+
+	erts_queue_message(rp, &rp_locks, mp, msg);
 
 	if (req->sched_id == sched_id)
 	    rp_locks &= ~ERTS_PROC_LOCK_MAIN;
@@ -5507,6 +5586,7 @@ void driver_report_exit(ErlDrvPort ix, int status)
    ErtsProcLocks rp_locks = 0;
    int scheduler = erts_get_scheduler_id() != 0;
    Port* prt = erts_drvport2port(ix);
+   int trace_send = IS_TRACED_FL(prt, F_TRACE_SEND);
 
    if (prt == ERTS_INVALID_ERL_DRV_PORT)
        return;
@@ -5523,13 +5603,17 @@ void driver_report_exit(ErlDrvPort ix, int status)
    if (!rp)
        return;
 
-   mp = erts_alloc_message_heap(rp, &rp_locks, 3+3, &hp, &ohp);
+   mp = erts_alloc_message_heap(trace_send ? NULL : rp, &rp_locks, 3+3, &hp, &ohp);
 
    tuple = TUPLE2(hp, am_exit_status, make_small(status));
    hp += 3;
    tuple = TUPLE2(hp, prt->common.id, tuple);
 
-   erts_queue_message(rp, &rp_locks, mp, tuple, am_undefined);
+    if (IS_TRACED_FL(prt, F_TRACE_SEND))
+        trace_port_send(prt, pid, tuple, 1);
+
+   ERL_MESSAGE_TOKEN(mp) = am_undefined;
+   erts_queue_message(rp, &rp_locks, mp, tuple);
 
    erts_smp_proc_unlock(rp, rp_locks);
    if (!scheduler)
@@ -5623,13 +5707,13 @@ cleanup_b2t_states(struct b2t_states__ *b2tsp)
  */
 
 static int
-driver_deliver_term(Eterm to, ErlDrvTermData* data, int len)
+driver_deliver_term(Port *prt, Eterm to, ErlDrvTermData* data, int len)
 {
 #define HEAP_EXTRA 200
 #define ERTS_DDT_FAIL do { res = -1; goto done; } while (0)
     Uint need = 0;
     int depth = 0;
-    int res;
+    int res = 0;
     ErlDrvTermData* ptr;
     ErlDrvTermData* ptr_end;
     DECLARE_ESTACK(stack); 
@@ -5848,11 +5932,26 @@ driver_deliver_term(Eterm to, ErlDrvTermData* data, int len)
 	  ? erts_proc_lookup(to)
 	  : erts_pid2proc_opt(NULL, 0, to, 0, ERTS_P2P_FLG_INC_REFC));
     if (!rp) {
-	res = 0;
-	goto done;
-    }
+        if (!prt || !IS_TRACED_FL(prt, F_TRACE_SEND))
+            goto done;
+        if (!erts_is_tracer_proc_enabled(NULL, 0, &prt->common, am_send))
+            goto done;
 
-    (void) erts_factory_message_create(&factory, rp, &rp_locks, need);
+	res = -2;
+
+        /* We allocate a temporary heap to be used to create
+           the message that may be sent using tracing */
+        erts_factory_tmp_init(&factory, erts_alloc(ERTS_ALC_T_DRIVER, need*sizeof(Eterm)),
+                              need, ERTS_ALC_T_DRIVER);
+
+    } else {
+        /* We force the creation of a heap fragment (rp == NULL) when send
+           tracing so that we don't have the main lock of the process while
+           tracing */
+        Process *trace_rp = prt && IS_TRACED_FL(prt, F_TRACE_SEND) ? NULL : rp;
+        (void) erts_factory_message_create(&factory, trace_rp, &rp_locks, need);
+        res = 1;
+    }
 
     /*
      * Interpret the instructions and build the term.
@@ -6115,22 +6214,40 @@ driver_deliver_term(Eterm to, ErlDrvTermData* data, int len)
 	ESTACK_PUSH(stack, mess);
     }
 
-    res = 1;
-
  done:
 
     if (res > 0) {
 	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);
+
 	/* send message */
-	erts_queue_message(rp, &rp_locks, factory.message, mess, am_undefined);
+        ERL_MESSAGE_TOKEN(factory.message) = am_undefined;
+	erts_queue_message(rp, &rp_locks, factory.message, mess);
+    }
+    else if (res == -2) {
+        /* this clause only happens when we were requested to
+           generate a send trace, but the process to send to
+           did not exist any more */
+        mess = ESTACK_POP(stack);  /* get resulting value */
+
+        trace_port_send(prt, to, mess, 0);
+
+	erts_factory_trim_and_close(&factory, &mess, 1);
+        erts_free(ERTS_ALC_T_DRIVER, factory.hp_start);
+        res = 0;
     }
     else {
 	if (b2t.ix > b2t.used)
 	    b2t.used = b2t.ix;
 	for (b2t.ix = 0; b2t.ix < b2t.used; b2t.ix++)
 	    erts_binary2term_abort(&b2t.state[b2t.ix]);
-	erts_factory_undo(&factory);
+        if (factory.mode != FACTORY_CLOSED) {
+            ERL_MESSAGE_TERM(factory.message) = am_undefined;
+            erts_factory_undo(&factory);
+        }
     }
     if (rp) {
 	if (rp_locks)
@@ -6146,7 +6263,8 @@ driver_deliver_term(Eterm to, ErlDrvTermData* data, int len)
 }
 
 static ERTS_INLINE int
-deliver_term_check_port(ErlDrvTermData port_id, Eterm *connected_p)
+deliver_term_check_port(ErlDrvTermData port_id, Eterm *connected_p,
+                        Port **trace_prt)
 {
 #ifdef ERTS_SMP
     ErtsThrPrgrDelayHandle dhndl = erts_thr_progress_unmanaged_delay();
@@ -6174,8 +6292,11 @@ deliver_term_check_port(ErlDrvTermData port_id, Eterm *connected_p)
     if (dhndl != ERTS_THR_PRGR_DHANDLE_MANAGED) {
 	erts_thr_progress_unmanaged_continue(dhndl);
 	ETHR_MEMBAR(ETHR_LoadLoad|ETHR_LoadStore);
-    }
+    } else
 #endif
+    {
+        *trace_prt = prt;
+    }
     ERTS_SMP_LC_ASSERT(dhndl == ERTS_THR_PRGR_DHANDLE_MANAGED
 		       ? erts_lc_is_port_locked(prt)
 		       : !erts_lc_is_port_locked(prt));
@@ -6186,10 +6307,11 @@ int erl_drv_output_term(ErlDrvTermData port_id, ErlDrvTermData* data, int len)
 {
     /* May be called from arbitrary thread */
     Eterm connected;
-    int res = deliver_term_check_port(port_id, &connected);
+    Port *prt = NULL;
+    int res = deliver_term_check_port(port_id, &connected, &prt);
     if (res <= 0)
 	return res;
-    return driver_deliver_term(connected, data, len);
+    return driver_deliver_term(prt, connected, data, len);
 }
 
 /*
@@ -6213,7 +6335,7 @@ driver_output_term(ErlDrvPort drvport, ErlDrvTermData* data, int len)
     if (state & ERTS_PORT_SFLG_CLOSING)
 	return 0;
 
-    return driver_deliver_term(ERTS_PORT_GET_CONNECTED(prt), data, len);
+    return driver_deliver_term(prt, ERTS_PORT_GET_CONNECTED(prt), data, len);
 }
 
 int erl_drv_send_term(ErlDrvTermData port_id,
@@ -6222,10 +6344,11 @@ int erl_drv_send_term(ErlDrvTermData port_id,
 		      int len)
 {
     /* May be called from arbitrary thread */
-    int res = deliver_term_check_port(port_id, NULL);
+    Port *prt = NULL;
+    int res = deliver_term_check_port(port_id, NULL, &prt);
     if (res <= 0)
 	return res;
-    return driver_deliver_term(to, data, len);
+    return driver_deliver_term(prt, to, data, len);
 }
 
 /*
@@ -6244,20 +6367,21 @@ driver_send_term(ErlDrvPort drvport,
      * to make this access safe without using a less efficient
      * internal data representation for ErlDrvPort.
      */
+    Port* prt = NULL;
     ERTS_SMP_CHK_NO_PROC_LOCKS;
 #ifdef ERTS_SMP
     if (erts_thr_progress_is_managed_thread())
 #endif
     {
 	erts_aint32_t state;
-	Port* prt = erts_drvport2port_state(drvport, &state);
+	prt = erts_drvport2port_state(drvport, &state);
 	if (prt == ERTS_INVALID_ERL_DRV_PORT)
 	    return -1; /* invalid (dead) */
 	ERTS_SMP_LC_ASSERT(erts_lc_is_port_locked(prt));
 	if (state & ERTS_PORT_SFLG_CLOSING)
 	    return 0;
     }
-    return driver_deliver_term(to, data, len);
+    return driver_deliver_term(prt, to, data, len);
 }
 
 
@@ -7433,7 +7557,7 @@ driver_failure_term(ErlDrvPort ix, Eterm term, int eof)
     if (state & ERTS_PORT_SFLG_CLOSING) {
 	terminate_port(prt);
     } else if (eof && (state & ERTS_PORT_SFLG_SOFT_EOF)) {
-	deliver_result(prt->common.id, ERTS_PORT_GET_CONNECTED(prt), am_eof);
+	deliver_result(prt, prt->common.id, ERTS_PORT_GET_CONNECTED(prt), am_eof);
     } else {
 	/* XXX UGLY WORK AROUND, Let erts_deliver_port_exit() terminate the port */
 	if (prt->port_data_lock)
@@ -7441,7 +7565,7 @@ driver_failure_term(ErlDrvPort ix, Eterm term, int eof)
 	prt->ioq.size = 0;
 	if (prt->port_data_lock)
 	    driver_pdl_unlock(prt->port_data_lock);
-	erts_deliver_port_exit(prt, prt->common.id, eof ? am_normal : term, 0);
+	erts_deliver_port_exit(prt, prt->common.id, eof ? am_normal : term, 0, 0);
     }
     return 0;
 }
diff --git a/erts/emulator/beam/register.c b/erts/emulator/beam/register.c
index 071f6aee08..77f79fcea4 100644
--- a/erts/emulator/beam/register.c
+++ b/erts/emulator/beam/register.c
@@ -226,7 +226,8 @@ int erts_register_name(Process *c_p, Eterm name, Eterm id)
     rp = (RegProc*) hash_put(&process_reg, (void*) &r);
     if (proc && rp->p == proc) {
 	if (IS_TRACED_FL(proc, F_TRACE_PROCS)) {
-	    trace_proc(c_p, proc, am_register, name);
+	    trace_proc(proc, ERTS_PROC_LOCK_MAIN,
+                       proc, am_register, name);
 	}
 	proc->common.u.alive.reg = rp;
     }
@@ -470,8 +471,8 @@ int erts_unregister_name(Process *c_p,
     int res = 0;
     RegProc r, *rp;
     Port *port = c_prt;
+    ErtsProcLocks current_c_p_locks = 0;
 #ifdef ERTS_SMP
-    ErtsProcLocks current_c_p_locks;
 
     /*
      * SMP note: If 'c_prt != NULL' and 'c_prt->reg->name == name',
@@ -537,8 +538,12 @@ int erts_unregister_name(Process *c_p,
 	    ERTS_SMP_LC_ASSERT(erts_lc_is_port_locked(port));
 
 	    rp->pt->common.u.alive.reg = NULL;
-	    
+
 	    if (IS_TRACED_FL(port, F_TRACE_PORTS)) {
+                if (current_c_p_locks) {
+                    erts_smp_proc_unlock(c_p, current_c_p_locks);
+                    current_c_p_locks = 0;
+                }
 		trace_port(port, am_unregister, r.name);
 	    }
 
@@ -555,7 +560,8 @@ int erts_unregister_name(Process *c_p,
 #endif
 	    rp->p->common.u.alive.reg = NULL;
 	    if (IS_TRACED_FL(rp->p, F_TRACE_PROCS)) {
-		trace_proc(c_p, rp->p, am_unregister, r.name);
+                trace_proc(rp->p, (c_p == rp->p) ? c_p_locks : ERTS_PROC_LOCK_MAIN,
+                           rp->p, am_unregister, r.name);
 	    }
 #ifdef ERTS_SMP
 	    if (rp->p != c_p) {
diff --git a/erts/emulator/beam/sys.h b/erts/emulator/beam/sys.h
index 8b3eb2db1d..748fba15c7 100644
--- a/erts/emulator/beam/sys.h
+++ b/erts/emulator/beam/sys.h
@@ -644,6 +644,15 @@ typedef struct preload {
     unsigned char* code;	/* Code pointer */
 } Preload;
 
+/*
+ * ErtsTracer is either NIL, 'true' or [Mod | State]
+ *
+ * If set to NIL, it means no tracer.
+ * If set to 'true' it means the current process' tracer.
+ * If set to [Mod | State], there is a tracer.
+ *  See erts_tracer_update for more details
+ */
+typedef Eterm ErtsTracer;
 
 /*
  * This structure contains options to all built in drivers.
diff --git a/erts/emulator/beam/utils.c b/erts/emulator/beam/utils.c
index 384b340a1c..b280995488 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, NIL);
+	erts_queue_message(p, NULL /* only used for smp build */, mp, message);
     }
 #endif
 }