Don't go to single-scheduler mode when managing breakpoints

Calls to erlang:set_trace_pattern/3 will no longer block all
other schedulers.

We will still go to single-scheduler mode when new code is loaded
for a module that is traced, or when loading code when there is a
default trace pattern set. That is not impossible to fix, but that
requires much closer cooperation between tracing BIFs and the loader
BIFs.

14 files changed, 693 insertions, 597 deletions

diff --git a/erts/emulator/beam/beam_bif_load.c b/erts/emulator/beam/beam_bif_load.c
index 9c65769b86..94f8edf165 100644
--- a/erts/emulator/beam/beam_bif_load.c
+++ b/erts/emulator/beam/beam_bif_load.c
@@ -672,11 +672,11 @@ set_default_trace_pattern(Eterm module)
     if (trace_pattern_is_on) {
 	Eterm mfa[1];
 	mfa[0] = module;
-	(void) erts_set_trace_pattern(mfa, 1,
+	(void) erts_set_trace_pattern(0, mfa, 1,
 				      match_spec,
 				      meta_match_spec,
 				      1, trace_pattern_flags,
-				      meta_tracer_pid);
+				      meta_tracer_pid, 1);
     }
 }
 
@@ -1006,12 +1006,11 @@ delete_code(Module* modp)
 		if (ep->code[3] == (BeamInstr) em_apply_bif) {
 		    continue;
 		}
-		else if (ep->code[3] == (BeamInstr) em_call_traced_function) {
+		else if (ep->code[3] ==
+			 (BeamInstr) BeamOp(op_i_generic_breakpoint)) {
 		    ERTS_SMP_LC_ASSERT(erts_smp_thr_progress_is_blocking());
 		    ASSERT(modp->curr.num_traced_exports > 0);
-		    --modp->curr.num_traced_exports;
-		    MatchSetUnref(ep->match_prog_set);
-		    ep->match_prog_set = NULL;
+		    erts_clear_export_break(modp, ep->code+3);
 		}
 		else ASSERT(ep->code[3] == (BeamInstr) em_call_error_handler
 			    || !erts_initialized);
@@ -1019,7 +1018,6 @@ delete_code(Module* modp)
 	    ep->addressv[code_ix] = ep->code+3;
 	    ep->code[3] = (BeamInstr) em_call_error_handler;
 	    ep->code[4] = 0;
-	    ASSERT(ep->match_prog_set == NULL);
 	}
     }
 
diff --git a/erts/emulator/beam/beam_bp.c b/erts/emulator/beam/beam_bp.c
index ebd7e1372f..50d18b0347 100644
--- a/erts/emulator/beam/beam_bp.c
+++ b/erts/emulator/beam/beam_bp.c
@@ -61,8 +61,9 @@
 #define ERTS_BPF_DEBUG             0x10
 #define ERTS_BPF_TIME_TRACE        0x20
 #define ERTS_BPF_TIME_TRACE_ACTIVE 0x40
+#define ERTS_BPF_GLOBAL_TRACE      0x80
 
-#define ERTS_BPF_ALL               0x7F
+#define ERTS_BPF_ALL               0xFF
 
 extern Eterm beam_return_to_trace[1];   /* OpCode(i_return_to_trace) */
 extern Eterm beam_return_trace[1];      /* OpCode(i_return_trace) */
@@ -80,7 +81,7 @@ erts_smp_atomic32_t erts_staging_bp_index;
 ** Helpers
 */
 static Eterm do_call_trace(Process* c_p, BeamInstr* I, Eterm* reg,
-			   Binary* ms, Eterm tracer_pid);
+			   int local, Binary* ms, Eterm tracer_pid);
 static void set_break(BpFunctions* f, Binary *match_spec, Uint break_flags,
 		      enum erts_break_op count_op, Eterm tracer_pid);
 static void set_function_break(BeamInstr *pc,
@@ -101,7 +102,7 @@ static void bp_time_diff(bp_data_time_item_t *item,
 static void bp_meta_unref(BpMetaPid* bmp);
 static void bp_count_unref(BpCount* bcp);
 static void bp_time_unref(BpDataTime* bdt);
-static void consolidate_bp_data(Module* modp, BeamInstr* pc);
+static void consolidate_bp_data(Module* modp, BeamInstr* pc, int local);
 static void uninstall_breakpoint(BeamInstr* pc);
 
 /* bp_hash */
@@ -201,30 +202,79 @@ erts_bp_match_functions(BpFunctions* f, Eterm mfa[3], int specified)
 }
 
 void
+erts_bp_match_export(BpFunctions* f, Eterm mfa[3], int specified)
+{
+    ErtsCodeIndex code_ix = erts_active_code_ix();
+    int i;
+    int num_exps = export_list_size(code_ix);
+    int ne;
+
+    f->matching = (BpFunction *) Alloc(num_exps*sizeof(BpFunction));
+    ne = 0;
+    for (i = 0; i < num_exps; i++) {
+	Export* ep = export_list(i, code_ix);
+	BeamInstr* pc;
+	int j;
+
+	for (j = 0; j < specified && mfa[j] == ep->code[j]; j++) {
+	    /* Empty loop body */
+	}
+	if (j < specified) {
+	    continue;
+	}
+	pc = ep->code+3;
+	if (ep->addressv[code_ix] == pc) {
+	    if ((*pc == (BeamInstr) em_apply_bif ||
+		 *pc == (BeamInstr) em_call_error_handler)) {
+		continue;
+	    }
+	    ASSERT(*pc == (BeamInstr) BeamOp(op_i_generic_breakpoint));
+	} else if (erts_is_native_break(ep->addressv[code_ix])) {
+	    continue;
+	}
+
+	f->matching[ne].pc = pc;
+	f->matching[ne].mod = erts_get_module(ep->code[0], code_ix);
+	ne++;
+
+    }
+    f->matched = ne;
+}
+
+void
 erts_bp_free_matched_functions(BpFunctions* f)
 {
     Free(f->matching);
 }
 
 void
-erts_consolidate_bp_data(BpFunctions* f)
+erts_consolidate_bp_data(BpFunctions* f, int local)
 {
     BpFunction* fs = f->matching;
     Uint i;
     Uint n = f->matched;
 
-    for (i = 0; i < BIF_SIZE; i++) {
-	Export *ep = bif_export[i];
-	consolidate_bp_data(0, ep->code+3);
-    }
+    ERTS_SMP_LC_ASSERT(erts_is_code_ix_locked());
 
     for (i = 0; i < n; i++) {
-	consolidate_bp_data(fs[i].mod, fs[i].pc);
+	consolidate_bp_data(fs[i].mod, fs[i].pc, local);
+    }
+}
+
+void
+erts_consolidate_bif_bp_data(void)
+{
+    int i;
+
+    ERTS_SMP_LC_ASSERT(erts_is_code_ix_locked());
+    for (i = 0; i < BIF_SIZE; i++) {
+	Export *ep = bif_export[i];
+	consolidate_bp_data(0, ep->code+3, 0);
     }
 }
 
 static void
-consolidate_bp_data(Module* modp, BeamInstr* pc)
+consolidate_bp_data(Module* modp, BeamInstr* pc, int local)
 {
     GenericBp* g = (GenericBp *) pc[-4];
     GenericBpData* src;
@@ -244,7 +294,7 @@ consolidate_bp_data(Module* modp, BeamInstr* pc)
      */
 
     flags = dst->flags;
-    if (flags & ERTS_BPF_LOCAL_TRACE) {
+    if (flags & (ERTS_BPF_LOCAL_TRACE|ERTS_BPF_GLOBAL_TRACE)) {
 	MatchSetUnref(dst->local_ms);
     }
     if (flags & ERTS_BPF_META_TRACE) {
@@ -264,13 +314,18 @@ consolidate_bp_data(Module* modp, BeamInstr* pc)
 
     flags = dst->flags = src->flags;
     if (flags == 0) {
-	pc[-4] = 0;
-	Free(g);
 	if (modp) {
-	    modp->curr.num_breakpoints--;
+	    if (local) {
+		modp->curr.num_breakpoints--;
+	    } else {
+		modp->curr.num_traced_exports--;
+	    }
 	    ASSERT(modp->curr.num_breakpoints >= 0);
+	    ASSERT(modp->curr.num_traced_exports >= 0);
 	    ASSERT(*pc != (BeamInstr) BeamOp(op_i_generic_breakpoint));
 	}
+	pc[-4] = 0;
+	Free(g);
 	return;
     }
 
@@ -279,7 +334,7 @@ consolidate_bp_data(Module* modp, BeamInstr* pc)
      * for the next time it will be used).
      */
 
-    if (flags & ERTS_BPF_LOCAL_TRACE) {
+    if (flags & (ERTS_BPF_LOCAL_TRACE|ERTS_BPF_GLOBAL_TRACE)) {
 	dst->local_ms = src->local_ms;
 	MatchSetRef(dst->local_ms);
     }
@@ -377,21 +432,26 @@ uninstall_breakpoint(BeamInstr* pc)
 void
 erts_set_trace_break(BpFunctions* f, Binary *match_spec)
 {
-    ERTS_SMP_LC_ASSERT(erts_smp_thr_progress_is_blocking());
     set_break(f, match_spec, ERTS_BPF_LOCAL_TRACE, 0, am_true);
 }
 
 void
 erts_set_mtrace_break(BpFunctions* f, Binary *match_spec, Eterm tracer_pid)
 {
-    ERTS_SMP_LC_ASSERT(erts_smp_thr_progress_is_blocking());
     set_break(f, match_spec, ERTS_BPF_META_TRACE, 0, tracer_pid);
 }
 
 void
+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);
+}
+
+void
 erts_set_mtrace_bif(BeamInstr *pc, Binary *match_spec, Eterm tracer_pid)
 {
-    ERTS_SMP_LC_ASSERT(erts_smp_thr_progress_is_blocking());
     set_function_break(pc, match_spec, ERTS_BPF_META_TRACE, 0, tracer_pid);
 }
 
@@ -410,14 +470,12 @@ erts_clear_time_trace_bif(BeamInstr *pc) {
 
 void
 erts_set_debug_break(BpFunctions* f) {
-    ERTS_SMP_LC_ASSERT(erts_smp_thr_progress_is_blocking());
     set_break(f, NULL, ERTS_BPF_DEBUG, 0, NIL);
 }
 
 void
 erts_set_count_break(BpFunctions* f, enum erts_break_op count_op)
 {
-    ERTS_SMP_LC_ASSERT(erts_smp_thr_progress_is_blocking());
     set_break(f, 0, ERTS_BPF_COUNT|ERTS_BPF_COUNT_ACTIVE,
 	      count_op, NIL);
 }
@@ -425,7 +483,6 @@ erts_set_count_break(BpFunctions* f, enum erts_break_op count_op)
 void
 erts_set_time_break(BpFunctions* f, enum erts_break_op count_op)
 {
-    ERTS_SMP_LC_ASSERT(erts_smp_thr_progress_is_blocking());
     set_break(f, 0, ERTS_BPF_TIME_TRACE|ERTS_BPF_TIME_TRACE_ACTIVE,
 	      count_op, NIL);
 }
@@ -433,14 +490,25 @@ erts_set_time_break(BpFunctions* f, enum erts_break_op count_op)
 void
 erts_clear_trace_break(BpFunctions* f)
 {
-    ERTS_SMP_LC_ASSERT(erts_smp_thr_progress_is_blocking());
     clear_break(f, ERTS_BPF_LOCAL_TRACE);
 }
 
 void
+erts_clear_call_trace_bif(BeamInstr *pc, int local)
+{
+    GenericBp* g = (GenericBp *) pc[-4];
+
+    if (g) {
+	Uint flags = local ? ERTS_BPF_LOCAL_TRACE : ERTS_BPF_GLOBAL_TRACE;
+	if (g->data[erts_staging_bp_ix()].flags & flags) {
+	    clear_function_break(pc, flags);
+	}
+    }
+}
+
+void
 erts_clear_mtrace_break(BpFunctions* f)
 {
-    ERTS_SMP_LC_ASSERT(erts_smp_thr_progress_is_blocking());
     clear_break(f, ERTS_BPF_META_TRACE);
 }
 
@@ -460,21 +528,18 @@ erts_clear_debug_break(BpFunctions* f)
 void
 erts_clear_count_break(BpFunctions* f)
 {
-    ERTS_SMP_LC_ASSERT(erts_smp_thr_progress_is_blocking());
     clear_break(f, ERTS_BPF_COUNT|ERTS_BPF_COUNT_ACTIVE);
 }
 
 void
 erts_clear_time_break(BpFunctions* f)
 {
-    ERTS_SMP_LC_ASSERT(erts_smp_thr_progress_is_blocking());
     clear_break(f, ERTS_BPF_TIME_TRACE|ERTS_BPF_TIME_TRACE_ACTIVE);
 }
 
 void
 erts_clear_all_breaks(BpFunctions* f)
 {
-    ERTS_SMP_LC_ASSERT(erts_smp_thr_progress_is_blocking());
     clear_break(f, ERTS_BPF_ALL);
 }
 
@@ -511,12 +576,24 @@ erts_clear_module_break(Module *modp) {
 	    continue;
 	}
 	uninstall_breakpoint(pc);
-	consolidate_bp_data(modp, pc);
+	consolidate_bp_data(modp, pc, 1);
 	ASSERT(pc[-4] == 0);
     }
     return n;
 }
 
+void
+erts_clear_export_break(Module* modp, BeamInstr* pc)
+{
+    ERTS_SMP_LC_ASSERT(erts_smp_thr_progress_is_blocking());
+
+    clear_function_break(pc, ERTS_BPF_ALL);
+    erts_commit_staged_bp();
+    *pc = (BeamInstr) 0;
+    consolidate_bp_data(modp, pc, 0);
+    ASSERT(pc[-4] == 0);
+}
+
 BeamInstr
 erts_generic_breakpoint(Process* c_p, BeamInstr* I, Eterm* reg)
 {
@@ -529,9 +606,12 @@ erts_generic_breakpoint(Process* c_p, BeamInstr* I, Eterm* reg)
     bp = &g->data[ix];
     bp_flags = bp->flags;
     ASSERT((bp_flags & ~ERTS_BPF_ALL) == 0);
-    if (bp_flags & (ERTS_BPF_LOCAL_TRACE|ERTS_BPF_TIME_TRACE_ACTIVE) &&
+    if (bp_flags & (ERTS_BPF_LOCAL_TRACE|
+		    ERTS_BPF_GLOBAL_TRACE|
+		    ERTS_BPF_TIME_TRACE_ACTIVE) &&
 	!IS_TRACED_FL(c_p, F_TRACE_CALLS)) {
 	bp_flags &= ~(ERTS_BPF_LOCAL_TRACE|
+		      ERTS_BPF_GLOBAL_TRACE|
 		      ERTS_BPF_TIME_TRACE|
 		      ERTS_BPF_TIME_TRACE_ACTIVE);
 	if (bp_flags == 0) {	/* Quick exit */
@@ -540,7 +620,10 @@ erts_generic_breakpoint(Process* c_p, BeamInstr* I, Eterm* reg)
     }
 
     if (bp_flags & ERTS_BPF_LOCAL_TRACE) {
-	(void) do_call_trace(c_p, I, reg, bp->local_ms, am_true);
+	ASSERT((bp_flags & ERTS_BPF_GLOBAL_TRACE) == 0);
+	(void) do_call_trace(c_p, I, reg, 1, bp->local_ms, am_true);
+    } else if (bp_flags & ERTS_BPF_GLOBAL_TRACE) {
+	(void) do_call_trace(c_p, I, reg, 0, bp->local_ms, am_true);
     }
 
     if (bp_flags & ERTS_BPF_META_TRACE) {
@@ -548,7 +631,7 @@ erts_generic_breakpoint(Process* c_p, BeamInstr* I, Eterm* reg)
 	Eterm new_pid;
 
 	old_pid = (Eterm) erts_smp_atomic_read_nob(&bp->meta_pid->pid);
-	new_pid = do_call_trace(c_p, I, reg, bp->meta_ms, old_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);
 	}
@@ -590,9 +673,188 @@ erts_generic_breakpoint(Process* c_p, BeamInstr* I, Eterm* reg)
     }
 }
 
+/*
+ * Entry point called by the trace wrap functions in erl_bif_wrap.c
+ *
+ * The trace wrap functions are themselves called through the export
+ * entries instead of the original BIF functions.
+ */
+Eterm
+erts_bif_trace(int bif_index, Process* p, Eterm* args, BeamInstr* I)
+{
+    Eterm result;
+    Eterm (*func)(Process*, Eterm*, BeamInstr*);
+    Export* ep = bif_export[bif_index];
+    Uint32 flags = 0, flags_meta = 0;
+    Eterm meta_tracer_pid = NIL;
+    int applying = (I == &(ep->code[3])); /* Yup, the apply code for a bif
+					   * is actually in the
+					   * export entry */
+    BeamInstr *cp = p->cp;
+    GenericBp* g;
+    GenericBpData* bp;
+    Uint bp_flags = 0;
+
+    ERTS_SMP_CHK_HAVE_ONLY_MAIN_PROC_LOCK(p);
+
+    g = (GenericBp *) ep->fake_op_func_info_for_hipe[1];
+    if (g) {
+	bp = &g->data[erts_active_bp_ix()];
+	bp_flags = bp->flags;
+    }
+
+    /*
+     * Make continuation pointer OK, it is not during direct BIF calls,
+     * but it is correct during apply of bif.
+     */
+    if (!applying) {
+	p->cp = I;
+    }
+    if (bp_flags & (ERTS_BPF_LOCAL_TRACE|ERTS_BPF_GLOBAL_TRACE) &&
+	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, &p->tracer_proc);
+    }
+    if (bp_flags & ERTS_BPF_META_TRACE) {
+	Eterm tpid1, tpid2;
+
+	tpid1 = tpid2 =
+	    (Eterm) erts_smp_atomic_read_nob(&bp->meta_pid->pid);
+	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);
+	}
+    }
+    if (bp_flags & ERTS_BPF_TIME_TRACE_ACTIVE &&
+	IS_TRACED_FL(p, F_TRACE_CALLS)) {
+	BeamInstr *pc = (BeamInstr *)ep->code+3;
+	erts_trace_time_call(p, pc, bp->time);
+    }
+
+    /* Restore original continuation pointer (if changed). */
+    p->cp = cp;
+
+    func = bif_table[bif_index].f;
+
+    result = func(p, args, I);
+
+    if (applying && (flags & MATCH_SET_RETURN_TO_TRACE)) {
+	BeamInstr i_return_trace      = beam_return_trace[0];
+	BeamInstr i_return_to_trace   = beam_return_to_trace[0];
+	BeamInstr i_return_time_trace = beam_return_time_trace[0];
+	Eterm *cpp;
+	/* Maybe advance cp to skip trace stack frames */
+	for (cpp = p->stop;  ;  cp = cp_val(*cpp++)) {
+	    if (*cp == i_return_trace) {
+		/* Skip stack frame variables */
+		while (is_not_CP(*cpp)) cpp++;
+		cpp += 2; /* Skip return_trace parameters */
+	    } else if (*cp == i_return_time_trace) {
+		/* Skip stack frame variables */
+		while (is_not_CP(*cpp)) cpp++;
+		cpp += 1; /* Skip return_time_trace parameters */
+	    } else if (*cp == i_return_to_trace) {
+		/* A return_to trace message is going to be generated
+		 * by normal means, so we do not have to.
+		 */
+		cp = NULL;
+		break;
+	    } else break;
+	}
+    }
+
+    /* Try to get these in the order
+     * they usually appear in normal code... */
+    if (is_non_value(result)) {
+	Uint reason = p->freason;
+	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;
+		ASSERT(is_tuple(value));
+		tp = tuple_val(value);
+		value = tp[1];
+	    }
+	    if ((reason & EXF_THROWN) && (p->catches <= 0)) {
+		value = TUPLE2(nocatch, am_nocatch, value);
+		reason = EXC_ERROR;
+	    }
+	    /* Note: expand_error_value() could theoretically
+	     * allocate on the heap, but not for any error
+	     * returned by a BIF, and it would do no harm,
+	     * just be annoying.
+	     */
+	    value = expand_error_value(p, reason, value);
+	    class = exception_tag[GET_EXC_CLASS(reason)];
+
+	    if (flags_meta & MATCH_SET_EXCEPTION_TRACE) {
+		erts_trace_exception(p, ep->code, class, value,
+				     &meta_tracer_pid);
+	    }
+	    if (flags & MATCH_SET_EXCEPTION_TRACE) {
+		erts_trace_exception(p, ep->code, class, value,
+				     &p->tracer_proc);
+	    }
+	    if ((flags & MATCH_SET_RETURN_TO_TRACE) && p->catches > 0) {
+		/* can only happen if(local)*/
+		Eterm *ptr = p->stop;
+		ASSERT(is_CP(*ptr));
+		ASSERT(ptr <= STACK_START(p));
+		/* Search the nearest stack frame for a catch */
+		while (++ptr < STACK_START(p)) {
+		    if (is_CP(*ptr)) break;
+		    if (is_catch(*ptr)) {
+			if (applying) {
+			    /* Apply of BIF, cp is in calling function */
+			    if (cp) erts_trace_return_to(p, cp);
+			} else {
+			    /* Direct bif call, I points into
+			     * calling function */
+			    erts_trace_return_to(p, I);
+			}
+		    }
+		}
+	    }
+	    UnUseTmpHeapNoproc(3);
+	    if ((flags_meta|flags) & MATCH_SET_EXCEPTION_TRACE) {
+		erts_smp_proc_lock(p, ERTS_PROC_LOCKS_ALL_MINOR);
+		p->trace_flags |= F_EXCEPTION_TRACE;
+		erts_smp_proc_unlock(p, ERTS_PROC_LOCKS_ALL_MINOR);
+	    }
+	}
+    } else {
+	if (flags_meta & MATCH_SET_RX_TRACE) {
+	    erts_trace_return(p, ep->code, result, &meta_tracer_pid);
+	}
+	/* MATCH_SET_RETURN_TO_TRACE cannot occur if(meta) */
+	if (flags & MATCH_SET_RX_TRACE) {
+	    erts_trace_return(p, ep->code, result, &p->tracer_proc);
+	}
+	if (flags & MATCH_SET_RETURN_TO_TRACE) {
+	    /* can only happen if(local)*/
+	    if (applying) {
+		/* Apply of BIF, cp is in calling function */
+		if (cp) erts_trace_return_to(p, cp);
+	    } else {
+		/* Direct bif call, I points into calling function */
+		erts_trace_return_to(p, I);
+	    }
+	}
+    }
+    ERTS_SMP_CHK_HAVE_ONLY_MAIN_PROC_LOCK(p);
+    return result;
+}
+
 static Eterm
 do_call_trace(Process* c_p, BeamInstr* I, Eterm* reg,
-	      Binary* ms, Eterm tracer_pid)
+	      int local, Binary* ms, Eterm tracer_pid)
 {
     Eterm* cpp;
     int return_to_trace = 0;
@@ -632,7 +894,7 @@ 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, 1, &tracer_pid);
+    flags = erts_call_trace(c_p, I-3, ms, reg, local, &tracer_pid);
     ERTS_SMP_REQ_PROC_MAIN_LOCK(c_p);
     if (cpp) {
 	c_p->cp = cp_save;
@@ -821,44 +1083,11 @@ erts_trace_time_return(Process *p, BeamInstr *pc)
     }
 }
 
-/*
- * SMP NOTE: Process p may have become exiting on return!
- */
-Uint32
-erts_bif_mtrace(Process *p, BeamInstr *pc, Eterm *args, int local,
-		Eterm *tracer_pid)
-{
-    GenericBp* g;
-
-    ASSERT(tracer_pid);
-    g = (GenericBp *) pc[-4];
-    if (g) {
-	Eterm tpid1, tpid2;
-	Uint32 flags;
-	GenericBpData* bp;
-	ErtsBpIndex ix = erts_active_bp_ix();
-
-	bp = &g->data[ix];
-	tpid1 = tpid2 =
-	    (Eterm) erts_smp_atomic_read_nob(&bp->meta_pid->pid);
-	flags = erts_call_trace(p, pc-3/*mfa*/, bp->meta_ms, args,
-				local, &tpid2);
-	*tracer_pid = tpid2;
-	if (tpid1 != tpid2) {
-	    erts_smp_atomic_set_nob(&bp->meta_pid->pid, tpid2);
-	}
-	return flags;
-    }
-    *tracer_pid = NIL;
-    return 0;
-}
-
-
-
 int 
-erts_is_trace_break(BeamInstr *pc, Binary **match_spec_ret)
+erts_is_trace_break(BeamInstr *pc, Binary **match_spec_ret, int local)
 {
-    GenericBpData* bp = check_break(pc, ERTS_BPF_LOCAL_TRACE);
+    Uint flags = local ? ERTS_BPF_LOCAL_TRACE : ERTS_BPF_GLOBAL_TRACE;
+    GenericBpData* bp = check_break(pc, flags);
 
     if (bp) {
 	if (match_spec_ret) {
@@ -1220,6 +1449,7 @@ set_function_break(BeamInstr *pc, Binary *match_spec, Uint break_flags,
     Uint common;
     ErtsBpIndex ix = erts_staging_bp_ix();
 
+    ERTS_SMP_LC_ASSERT(erts_is_code_ix_locked());
     g = (GenericBp *) pc[-4];
     if (g == 0) {
 	int i;
@@ -1241,7 +1471,7 @@ set_function_break(BeamInstr *pc, Binary *match_spec, Uint break_flags,
      */
 
     common = break_flags & bp->flags;
-    if (common & ERTS_BPF_LOCAL_TRACE) {
+    if (common & (ERTS_BPF_LOCAL_TRACE|ERTS_BPF_GLOBAL_TRACE)) {
 	MatchSetUnref(bp->local_ms);
     } else if (common & ERTS_BPF_META_TRACE) {
 	MatchSetUnref(bp->meta_ms);
@@ -1276,7 +1506,7 @@ set_function_break(BeamInstr *pc, Binary *match_spec, Uint break_flags,
      * Initialize the new breakpoint data.
      */
 
-    if (break_flags & ERTS_BPF_LOCAL_TRACE) {
+    if (break_flags & (ERTS_BPF_LOCAL_TRACE|ERTS_BPF_GLOBAL_TRACE)) {
 	MatchSetRef(match_spec);
 	bp->local_ms = match_spec;
     } else if (break_flags & ERTS_BPF_META_TRACE) {
@@ -1335,6 +1565,8 @@ clear_function_break(BeamInstr *pc, Uint break_flags)
     Uint common;
     ErtsBpIndex ix = erts_staging_bp_ix();
 
+    ERTS_SMP_LC_ASSERT(erts_is_code_ix_locked());
+
     if ((g = (GenericBp *) pc[-4]) == 0) {
 	return 1;
     }
@@ -1343,7 +1575,7 @@ clear_function_break(BeamInstr *pc, Uint break_flags)
     ASSERT((bp->flags & ~ERTS_BPF_ALL) == 0);
     common = bp->flags & break_flags;
     bp->flags &= ~break_flags;
-    if (common & ERTS_BPF_LOCAL_TRACE) {
+    if (common & (ERTS_BPF_LOCAL_TRACE|ERTS_BPF_GLOBAL_TRACE)) {
 	MatchSetUnref(bp->local_ms);
     }
     if (common & ERTS_BPF_META_TRACE) {
@@ -1426,13 +1658,6 @@ get_time_break(BeamInstr *pc)
     return bp ? bp->time : 0;
 }
 
-BpDataTime*
-erts_get_active_time_break(BeamInstr *pc)
-{
-    GenericBpData* bp = check_break(pc, ERTS_BPF_TIME_TRACE_ACTIVE);
-    return bp ? bp->time : 0;
-}
-
 static GenericBpData*
 check_break(BeamInstr *pc, Uint break_flags)
 {
diff --git a/erts/emulator/beam/beam_bp.h b/erts/emulator/beam/beam_bp.h
index e854df2e19..28aaaa462a 100644
--- a/erts/emulator/beam/beam_bp.h
+++ b/erts/emulator/beam/beam_bp.h
@@ -119,20 +119,27 @@ ERTS_GLB_INLINE ErtsBpIndex erts_active_bp_ix(void);
 ERTS_GLB_INLINE ErtsBpIndex erts_staging_bp_ix(void);
 
 void erts_bp_match_functions(BpFunctions* f, Eterm mfa[3], int specified);
+void erts_bp_match_export(BpFunctions* f, Eterm mfa[3], int specified);
 void erts_bp_free_matched_functions(BpFunctions* f);
 
 void erts_install_breakpoints(BpFunctions* f);
 void erts_uninstall_breakpoints(BpFunctions* f);
-void erts_consolidate_bp_data(BpFunctions* f);
+void erts_consolidate_bp_data(BpFunctions* f, int local);
+void erts_consolidate_bif_bp_data(void);
 
 void erts_set_trace_break(BpFunctions *f, Binary *match_spec);
 void erts_clear_trace_break(BpFunctions *f);
+
+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);
 void erts_clear_mtrace_break(BpFunctions *f);
 void erts_set_mtrace_bif(BeamInstr *pc, Binary *match_spec,
 			 Eterm tracer_pid);
 void erts_clear_mtrace_bif(BeamInstr *pc);
+
 void erts_set_debug_break(BpFunctions *f);
 void erts_clear_debug_break(BpFunctions *f);
 void erts_set_count_break(BpFunctions *f, enum erts_break_op);
@@ -141,14 +148,13 @@ void erts_clear_count_break(BpFunctions *f);
 
 void erts_clear_all_breaks(BpFunctions* f);
 int erts_clear_module_break(Module *modp);
+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 erts_bif_mtrace(Process *p, BeamInstr *pc, Eterm *args,
-		       int local, Eterm *tracer_pid);
 
-int erts_is_trace_break(BeamInstr *pc, Binary **match_spec_ret);
+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);
 int erts_is_mtrace_bif(BeamInstr *pc, Binary **match_spec_ret,
@@ -166,7 +172,6 @@ void erts_clear_time_break(BpFunctions *f);
 int erts_is_time_trace_bif(Process *p, BeamInstr *pc, Eterm *call_time);
 void erts_set_time_trace_bif(BeamInstr *pc, enum erts_break_op);
 void erts_clear_time_trace_bif(BeamInstr *pc);
-BpDataTime* erts_get_active_time_break(BeamInstr *pc);
 
 BeamInstr *erts_find_local_func(Eterm mfa[3]);
 
diff --git a/erts/emulator/beam/beam_debug.c b/erts/emulator/beam/beam_debug.c
index 8d72627b38..a609ed8c71 100644
--- a/erts/emulator/beam/beam_debug.c
+++ b/erts/emulator/beam/beam_debug.c
@@ -132,7 +132,7 @@ erts_debug_breakpoint_2(BIF_ALIST_2)
 	erts_commit_staged_bp();
 	erts_uninstall_breakpoints(&f);
     }
-    erts_consolidate_bp_data(&f);
+    erts_consolidate_bp_data(&f, 1);
     res = make_small(f.matched);
     erts_bp_free_matched_functions(&f);
 
diff --git a/erts/emulator/beam/beam_emu.c b/erts/emulator/beam/beam_emu.c
index b5282830af..efce070061 100644
--- a/erts/emulator/beam/beam_emu.c
+++ b/erts/emulator/beam/beam_emu.c
@@ -217,7 +217,6 @@ BeamInstr beam_continue_exit[1];
 
 BeamInstr* em_call_error_handler;
 BeamInstr* em_apply_bif;
-BeamInstr* em_call_traced_function;
 
 
 /* NOTE These should be the only variables containing trace instructions.
@@ -4571,64 +4570,6 @@ void process_main(void)
      * Trace and debugging support.
      */
 
-    /*
-     * At this point, I points to the code[3] in the export entry for
-     * a trace-enabled function.
-     *
-     * code[0]: Module
-     * code[1]: Function
-     * code[2]: Arity
-     * code[3]: &&call_traced_function
-     * code[4]: Address of function.
-     */
- OpCase(call_traced_function): {
-     if (IS_TRACED_FL(c_p, F_TRACE_CALLS)) {
-	 unsigned offset = offsetof(Export, code) + 3*sizeof(BeamInstr);
-	 Export* ep = (Export *) (((char *)I)-offset);
-	 Uint32 flags;
-
-	 SWAPOUT;
-	 reg[0] = r(0);
-	 PROCESS_MAIN_CHK_LOCKS(c_p);
-	 ERTS_SMP_UNREQ_PROC_MAIN_LOCK(c_p);
-	 flags = erts_call_trace(c_p, ep->code, ep->match_prog_set, reg,
-				 0, &c_p->tracer_proc);
-	 ERTS_VERIFY_UNUSED_TEMP_ALLOC(c_p);
-	 ERTS_SMP_REQ_PROC_MAIN_LOCK(c_p);
-	 PROCESS_MAIN_CHK_LOCKS(c_p);
-	 ASSERT(!ERTS_PROC_IS_EXITING(c_p));
-	 SWAPIN;
-	 
-	 if (flags & MATCH_SET_RX_TRACE) {
-	     ASSERT(c_p->htop <= E && E <= c_p->hend);
-	     if (E - 3 < HTOP) {
-		 /* SWAPOUT, SWAPIN was done and r(0) was saved above */
-		 PROCESS_MAIN_CHK_LOCKS(c_p);
-		 FCALLS -= erts_garbage_collect(c_p, 3, reg, ep->code[2]);
-		 ERTS_VERIFY_UNUSED_TEMP_ALLOC(c_p);
-		 PROCESS_MAIN_CHK_LOCKS(c_p);
-		 r(0) = reg[0];
-		 SWAPIN;
-	     }
-	     E -= 3;
-	     ASSERT(c_p->htop <= E && E <= c_p->hend);
-	     ASSERT(is_CP((BeamInstr)(ep->code)));
-	     ASSERT(is_internal_pid(c_p->tracer_proc) || 
-		    is_internal_port(c_p->tracer_proc));
-	     E[2] = make_cp(c_p->cp); /* Code in lower range on halfword */
-	     E[1] = am_true; /* Process tracer */
-	     E[0] = make_cp(ep->code);
-	     c_p->cp = (flags & MATCH_SET_EXCEPTION_TRACE)
-		 ? beam_exception_trace : beam_return_trace;
-	     erts_smp_proc_lock(c_p, ERTS_PROC_LOCKS_ALL_MINOR);
-	     c_p->trace_flags |= F_EXCEPTION_TRACE;
-	     erts_smp_proc_unlock(c_p, ERTS_PROC_LOCKS_ALL_MINOR);
-	 }
-     }
-     SET_I((BeamInstr *)Arg(0));
-     Dispatch();
- }
-
  OpCase(return_trace): {
      BeamInstr* code = (BeamInstr *) (UWord) E[0];
      
@@ -5015,7 +4956,6 @@ void process_main(void)
 #endif /* NO_JUMP_TABLE */
      
      em_call_error_handler = OpCode(call_error_handler);
-     em_call_traced_function = OpCode(call_traced_function);
      em_apply_bif = OpCode(apply_bif);
 
      beam_apply[0]             = (BeamInstr) OpCode(i_apply);
diff --git a/erts/emulator/beam/beam_load.h b/erts/emulator/beam/beam_load.h
index f1506a8684..1048f258a5 100644
--- a/erts/emulator/beam/beam_load.h
+++ b/erts/emulator/beam/beam_load.h
@@ -49,7 +49,6 @@ extern void** beam_ops;
 extern BeamInstr beam_debug_apply[];
 extern BeamInstr* em_call_error_handler;
 extern BeamInstr* em_apply_bif;
-extern BeamInstr* em_call_traced_function;
 
 /*
  * The following variables keep a sorted list of address ranges for
diff --git a/erts/emulator/beam/erl_bif_trace.c b/erts/emulator/beam/erl_bif_trace.c
index 971e5ed650..e88fb8c9f4 100644
--- a/erts/emulator/beam/erl_bif_trace.c
+++ b/erts/emulator/beam/erl_bif_trace.c
@@ -42,12 +42,24 @@
 #define DECL_AM(S) Eterm AM_ ## S = am_atom_put(#S, sizeof(#S) - 1)
 
 const struct trace_pattern_flags   erts_trace_pattern_flags_off = {0, 0, 0, 0, 0};
+
+/*
+ * The following variables are protected by code write permission.
+ */
 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 struct {			/* Protected by code write permission */
+    int current;
+    int install;
+    int local;
+    BpFunctions f;		/* Local functions */
+    BpFunctions e;		/* Export entries */
+} finish_bp;
+
 static Eterm
 trace_pattern(Process* p, Eterm MFA, Eterm Pattern, Eterm flaglist);
 static BIF_RETTYPE
@@ -60,12 +72,11 @@ static Eterm trace_info_pid(Process* p, Eterm pid_spec, Eterm key);
 static Eterm trace_info_func(Process* p, Eterm pid_spec, Eterm key);
 static Eterm trace_info_on_load(Process* p, Eterm key);
 
-static int setup_func_trace(Export* ep, void* match_prog, ErtsCodeIndex);
-static int reset_func_trace(Export* ep, ErtsCodeIndex);
-static void reset_bif_trace(int bif_index);
-static void setup_bif_trace(int bif_index);
-static void set_trace_bif(int bif_index, void* match_prog);
-static void clear_trace_bif(int bif_index);
+static void reset_bif_trace(void);
+static void setup_bif_trace(void);
+static void install_exp_breakpoints(BpFunctions* f);
+static void uninstall_exp_breakpoints(BpFunctions* f);
+static void clean_export_entries(BpFunctions* f);
 
 void
 erts_bif_trace_init(void)
@@ -107,12 +118,12 @@ trace_pattern(Process* p, Eterm MFA, Eterm Pattern, Eterm flaglist)
     int is_global;
     Process *meta_tracer_proc = p;
     Eterm meta_tracer_pid = p->id;
+    int is_blocking = 0;
 
     if (!erts_try_seize_code_write_permission(p)) {
 	ERTS_BIF_YIELD3(bif_export[BIF_trace_pattern_3], p, MFA, Pattern, flaglist);
     }
-    erts_smp_proc_unlock(p, ERTS_PROC_LOCK_MAIN);
-    erts_smp_thr_progress_block();
+    finish_bp.current = -1;
 
     UseTmpHeap(3,p);
     /*
@@ -328,16 +339,24 @@ trace_pattern(Process* p, Eterm MFA, Eterm Pattern, Eterm flaglist)
 	    meta_tracer_proc->trace_flags |= F_TRACER;
 	}
 
-	matches = erts_set_trace_pattern(mfa, specified, 
+	matches = erts_set_trace_pattern(p, mfa, specified,
 					 match_prog_set, match_prog_set,
-					 on, flags, meta_tracer_pid);
+					 on, flags, meta_tracer_pid, 0);
     }
 
  error:
     MatchSetUnref(match_prog_set);
     UnUseTmpHeap(3,p);
-    erts_smp_thr_progress_unblock();
-    erts_smp_proc_lock(p, ERTS_PROC_LOCK_MAIN);
+
+#ifdef ERTS_SMP
+    if (finish_bp.current >= 0) {
+	ASSERT(matches >= 0);
+	erts_notify_finish_breakpointing(p);
+	erts_suspend(p, ERTS_PROC_LOCK_MAIN, NULL);
+	ERTS_BIF_YIELD_RETURN(p, make_small(matches));
+    }
+#endif
+
     erts_release_code_write_permission();
 
     if (matches >= 0) {
@@ -355,6 +374,8 @@ erts_get_default_trace_pattern(int *trace_pattern_is_on,
 			       struct trace_pattern_flags *trace_pattern_flags,
 			       Eterm *meta_tracer_pid)
 {
+    ERTS_SMP_LC_ASSERT(erts_is_code_ix_locked() ||
+		       erts_smp_thr_progress_is_blocking());
     if (trace_pattern_is_on)
 	*trace_pattern_is_on = erts_default_trace_pattern_is_on;
     if (match_spec)
@@ -369,6 +390,8 @@ erts_get_default_trace_pattern(int *trace_pattern_is_on,
 
 int erts_is_default_trace_enabled(void)
 {
+    ERTS_SMP_LC_ASSERT(erts_is_code_ix_locked() ||
+		       erts_smp_thr_progress_is_blocking());
     return erts_default_trace_pattern_is_on;
 }
 
@@ -842,6 +865,11 @@ Eterm trace_info_2(BIF_ALIST_2)
     Eterm What = BIF_ARG_1;
     Eterm Key = BIF_ARG_2;
     Eterm res;
+
+    if (!erts_try_seize_code_write_permission(p)) {
+	ERTS_BIF_YIELD2(bif_export[BIF_trace_info_2], p, What, Key);
+    }
+
     if (What == am_on_load) {
 	res = trace_info_on_load(p, Key);
     } else if (is_atom(What) || is_pid(What)) {
@@ -849,8 +877,10 @@ Eterm trace_info_2(BIF_ALIST_2)
     } else if (is_tuple(What)) {
 	res = trace_info_func(p, What, Key);
     } else {
+	erts_release_code_write_permission();
 	BIF_ERROR(p, BADARG);
     }
+    erts_release_code_write_permission();
     BIF_RET(res);
 }
 
@@ -983,59 +1013,49 @@ static int function_is_traced(Process *p,
 {
     Export e;
     Export* ep;
-    int i;
-    BeamInstr *code;
+    BeamInstr* pc;
 
     /* First look for an export entry */
     e.code[0] = mfa[0];
     e.code[1] = mfa[1];
     e.code[2] = mfa[2];
     if ((ep = export_get(&e)) != NULL) {
-	if (ep->addressv[erts_active_code_ix()] == ep->code+3 &&
-	    ep->code[3] != (BeamInstr) em_call_error_handler) {
-	    if (ep->code[3] == (BeamInstr) em_call_traced_function) {
-		*ms = ep->match_prog_set;
+	pc = ep->code+3;
+	if (ep->addressv[erts_active_code_ix()] == pc &&
+	    *pc != (BeamInstr) em_call_error_handler) {
+
+	    int r = 0;
+
+	    ASSERT(*pc == (BeamInstr) em_apply_bif ||
+		   *pc == (BeamInstr) BeamOp(op_i_generic_breakpoint));
+
+	    if (erts_is_trace_break(pc, ms, 0)) {
 		return FUNC_TRACE_GLOBAL_TRACE;
 	    }
-	    if (ep->code[3] == (BeamInstr) em_apply_bif) {
-		for (i = 0; i < BIF_SIZE; ++i) {
-		    if (bif_export[i] == ep) {
-			int r = 0;
-			
-			if (erts_bif_trace_flags[i] & BIF_TRACE_AS_GLOBAL) {
-			    *ms = ep->match_prog_set;
-			    return FUNC_TRACE_GLOBAL_TRACE;
-			} else {
-			    if (erts_bif_trace_flags[i] & BIF_TRACE_AS_LOCAL) {
-				r |= FUNC_TRACE_LOCAL_TRACE;
-				*ms = ep->match_prog_set;
-			    }
-			    if (erts_is_mtrace_break(ep->code+3, ms_meta,
-						   tracer_pid_meta)) {
-				r |= FUNC_TRACE_META_TRACE;
-			    }
-			    if (erts_is_time_break(p, ep->code+3, call_time)) {
-				r |= FUNC_TRACE_TIME_TRACE;
-			    }
-			}
-			return r ? r : FUNC_TRACE_UNTRACED;
-		    }
-		}
-		erl_exit(1,"Impossible ghost bif encountered in trace_info.");
+
+	    if (erts_is_trace_break(pc, ms, 1)) {
+		r |= FUNC_TRACE_LOCAL_TRACE;
+	    }
+	    if (erts_is_mtrace_break(pc, ms_meta, tracer_pid_meta)) {
+		r |= FUNC_TRACE_META_TRACE;
 	    }
+	    if (erts_is_time_break(p, pc, call_time)) {
+		r |= FUNC_TRACE_TIME_TRACE;
+	    }
+	    return r ? r : FUNC_TRACE_UNTRACED;
 	}
     }
     
     /* OK, now look for breakpoint tracing */
-    if ((code = erts_find_local_func(mfa)) != NULL) {
+    if ((pc = erts_find_local_func(mfa)) != NULL) {
 	int r = 
-	    (erts_is_trace_break(code, ms)
+	    (erts_is_trace_break(pc, ms, 1)
 	     ? FUNC_TRACE_LOCAL_TRACE : 0) 
-	    | (erts_is_mtrace_break(code, ms_meta, tracer_pid_meta)
+	    | (erts_is_mtrace_break(pc, ms_meta, tracer_pid_meta)
 	       ? FUNC_TRACE_META_TRACE : 0)
-	    | (erts_is_count_break(code, count)
+	    | (erts_is_count_break(pc, count)
 	       ? FUNC_TRACE_COUNT_TRACE : 0)
-	    | (erts_is_time_break(p, code, call_time)
+	    | (erts_is_time_break(p, pc, call_time)
 	       ? FUNC_TRACE_TIME_TRACE : 0);
 	
 	return r ? r : FUNC_TRACE_UNTRACED;
@@ -1327,40 +1347,46 @@ trace_info_on_load(Process* p, Eterm key)
 #undef FUNC_TRACE_LOCAL_TRACE
 
 int
-erts_set_trace_pattern(Eterm* mfa, int specified, 
+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)
+		       Eterm meta_tracer_pid, int is_blocking)
 {
     const ErtsCodeIndex code_ix = erts_active_code_ix();
-    BpFunctions f;
     int matches = 0;
     int i;
+    int n;
+    BpFunction* fp;
 
     /*
      * First work on normal functions (not real BIFs).
      */
-    
-    for (i = 0; i < export_list_size(code_ix); i++) {
-	Export* ep = export_list(i, code_ix);
-	int j;
-	
-	if (ExportIsBuiltIn(ep)) {
-	    continue;
-	}
-	
-	for (j = 0; j < specified && mfa[j] == ep->code[j]; j++) {
-	    /* Empty loop body */
-	}
 
-	if (j == specified) {
-	    if (on) {
-		if (! flags.breakpoint)
-		    matches += setup_func_trace(ep, match_prog_set, code_ix);
-		else
-		    reset_func_trace(ep, code_ix);
-	    } else if (! flags.breakpoint) {
-		matches += reset_func_trace(ep, code_ix);
+    erts_bp_match_export(&finish_bp.e, mfa, specified);
+    fp = finish_bp.e.matching;
+    n = finish_bp.e.matched;
+
+    for (i = 0; i < n; i++) {
+	BeamInstr* pc = fp[i].pc;
+	Export* ep = (Export *)(((char *)(pc-3)) - offsetof(Export, code));
+
+	if (!on || flags.breakpoint) {
+	    erts_clear_call_trace_bif(pc, 0);
+	    if (pc[0] == (BeamInstr) BeamOp(op_i_generic_breakpoint)) {
+		pc[0] = (BeamInstr) BeamOp(op_jump_f);
+	    }
+	} else {
+	    if (ep->addressv[code_ix] != pc) {
+		fp[i].mod->curr.num_traced_exports++;
+#ifdef DEBUG
+		pc[-5] = (BeamInstr) BeamOp(op_i_func_info_IaaI);
+#endif
+		pc[0] = (BeamInstr) BeamOp(op_jump_f);
+		pc[1] = (BeamInstr) ep->addressv[code_ix];
+	    }
+	    erts_set_call_trace_bif(pc, match_prog_set, 0);
+	    if (ep->addressv[code_ix] != pc) {
+		pc[0] = (BeamInstr) BeamOp(op_i_generic_breakpoint);
 	    }
 	}
     }
@@ -1385,26 +1411,15 @@ erts_set_trace_pattern(Eterm* mfa, int specified,
 	    /* Empty loop body */
 	}
 	if (j == specified) {
+	    BeamInstr* pc = (BeamInstr *)bif_export[i]->code + 3;
+
 	    if (! flags.breakpoint) { /* Export entry call trace */
 		if (on) {
-		    if (erts_bif_trace_flags[i] & BIF_TRACE_AS_META) {
-			ASSERT(ExportIsBuiltIn(bif_export[i]));
-			erts_clear_mtrace_bif
-			    ((BeamInstr *)bif_export[i]->code + 3);
-			erts_bif_trace_flags[i] &= ~BIF_TRACE_AS_META;
-		    }
-		    set_trace_bif(i, match_prog_set);
-		    erts_bif_trace_flags[i] &= ~BIF_TRACE_AS_LOCAL;
-		    erts_bif_trace_flags[i] |= BIF_TRACE_AS_GLOBAL;
-		    setup_bif_trace(i);
+		    erts_clear_call_trace_bif(pc, 1);
+		    erts_clear_mtrace_bif(pc);
+		    erts_set_call_trace_bif(pc, match_prog_set, 0);
 		} else { /* off */
-		    if (erts_bif_trace_flags[i] & BIF_TRACE_AS_GLOBAL) {
-			clear_trace_bif(i);
-			erts_bif_trace_flags[i] &= ~BIF_TRACE_AS_GLOBAL;
-		    }
-		    if (! erts_bif_trace_flags[i]) {
-			reset_bif_trace(i);
-		    }
+		    erts_clear_call_trace_bif(pc, 0);
 		}
 		matches++;
 	    } else { /* Breakpoint call trace */
@@ -1412,52 +1427,33 @@ erts_set_trace_pattern(Eterm* mfa, int specified,
 		
 		if (on) {
 		    if (flags.local) {
-			set_trace_bif(i, match_prog_set);
-			erts_bif_trace_flags[i] |= BIF_TRACE_AS_LOCAL;
-			erts_bif_trace_flags[i] &= ~BIF_TRACE_AS_GLOBAL;
+			erts_clear_call_trace_bif(pc, 0);
+			erts_set_call_trace_bif(pc, match_prog_set, 1);
 			m = 1;
 		    }
 		    if (flags.meta) {
-			erts_set_mtrace_bif
-			    ((BeamInstr *)bif_export[i]->code + 3,
-			     meta_match_prog_set, meta_tracer_pid);
-			erts_bif_trace_flags[i] |= BIF_TRACE_AS_META;
-			erts_bif_trace_flags[i] &= ~BIF_TRACE_AS_GLOBAL;
+			erts_set_mtrace_bif(pc, meta_match_prog_set,
+					    meta_tracer_pid);
 			m = 1;
 		    }
 		    if (flags.call_time) {
-			erts_set_time_trace_bif(bif_export[i]->code + 3, on);
+			erts_set_time_trace_bif(pc, on);
 			/* I don't want to remove any other tracers */
-			erts_bif_trace_flags[i] |= BIF_TRACE_AS_CALL_TIME;
 			m = 1;
 		    }
-		    if (erts_bif_trace_flags[i]) {
-			setup_bif_trace(i);
-		    }
 		} else { /* off */
 		    if (flags.local) {
-			if (erts_bif_trace_flags[i] & BIF_TRACE_AS_LOCAL) {
-			    clear_trace_bif(i);
-			    erts_bif_trace_flags[i] &= ~BIF_TRACE_AS_LOCAL;
-			}
+			erts_clear_call_trace_bif(pc, 1);
 			m = 1;
 		    }
 		    if (flags.meta) {
-			if (erts_bif_trace_flags[i] & BIF_TRACE_AS_META) {
-			    erts_clear_mtrace_bif
-				((BeamInstr *)bif_export[i]->code + 3);
-			    erts_bif_trace_flags[i] &= ~BIF_TRACE_AS_META;
-			}
+			erts_clear_mtrace_bif(pc);
 			m = 1;
 		    }
 		    if (flags.call_time) {
-			erts_clear_time_trace_bif(bif_export[i]->code + 3);
-			erts_bif_trace_flags[i] &= ~BIF_TRACE_AS_CALL_TIME;
+			erts_clear_time_trace_bif(pc);
 			m = 1;
 		    }
-		    if (! erts_bif_trace_flags[i]) {
-			reset_bif_trace(i);
-		    }
 		}
 		matches += m;
 	    }
@@ -1467,184 +1463,242 @@ erts_set_trace_pattern(Eterm* mfa, int specified,
     /*
     ** So, now for breakpoint tracing
     */
-    erts_bp_match_functions(&f, mfa, specified);
+    erts_bp_match_functions(&finish_bp.f, mfa, specified);
     if (on) {
 	if (! flags.breakpoint) {
-	    erts_clear_all_breaks(&f);
-	    erts_commit_staged_bp();
-	    erts_uninstall_breakpoints(&f);
+	    erts_clear_all_breaks(&finish_bp.f);
 	} else {
 	    if (flags.local) {
-		erts_set_trace_break(&f, match_prog_set);
+		erts_set_trace_break(&finish_bp.f, match_prog_set);
 	    }
 	    if (flags.meta) {
-		erts_set_mtrace_break(&f, meta_match_prog_set,
+		erts_set_mtrace_break(&finish_bp.f, meta_match_prog_set,
 				      meta_tracer_pid);
 	    }
 	    if (flags.call_count) {
-		erts_set_count_break(&f, on);
+		erts_set_count_break(&finish_bp.f, on);
 	    }
 	    if (flags.call_time) {
-		erts_set_time_break(&f, on);
+		erts_set_time_break(&finish_bp.f, on);
 	    }
-	    erts_install_breakpoints(&f);
-	    erts_commit_staged_bp();
 	}
     } else {
 	if (flags.local) {
-	    erts_clear_trace_break(&f);
+	    erts_clear_trace_break(&finish_bp.f);
 	}
 	if (flags.meta) {
-	    erts_clear_mtrace_break(&f);
+	    erts_clear_mtrace_break(&finish_bp.f);
 	}
 	if (flags.call_count) {
-	    erts_clear_count_break(&f);
+	    erts_clear_count_break(&finish_bp.f);
 	}
 	if (flags.call_time) {
-	    erts_clear_time_break(&f);
+	    erts_clear_time_break(&finish_bp.f);
 	}
-	erts_commit_staged_bp();
-	erts_uninstall_breakpoints(&f);
     }
-    erts_consolidate_bp_data(&f);
+
+    finish_bp.current = 0;
+    finish_bp.install = on;
+    finish_bp.local = flags.breakpoint;
+
+#ifdef ERTS_SMP
+    if (is_blocking) {
+	ERTS_SMP_LC_ASSERT(erts_smp_thr_progress_is_blocking());
+#endif
+	while (erts_finish_breakpointing()) {
+	    /* Empty loop body */
+	}
+#ifdef ERTS_SMP
+	finish_bp.current = -1;
+    }
+#endif
+
     if (flags.breakpoint) {
-	matches += f.matched;
+	matches += finish_bp.f.matched;
+    } else {
+	matches += finish_bp.e.matched;
     }
-    erts_bp_free_matched_functions(&f);
     return matches;
 }
 
-/*
- * Setup function tracing for the given exported function.
- *
- * Return Value: 1 if entry refers to a BIF or loaded function,
- * 0 if the entry refers to a function not loaded.
- */
-
-static int
-setup_func_trace(Export* ep, void* match_prog, ErtsCodeIndex code_ix)
+int
+erts_finish_breakpointing(void)
 {
-    Module* modp;
-
-    if (ep->addressv[code_ix] == ep->code+3) {
-	if (ep->code[3] == (BeamInstr) em_call_error_handler) {
-	    return 0;
-	} else if (ep->code[3] == (BeamInstr) em_call_traced_function) {
-	    MatchSetUnref(ep->match_prog_set);
-	    ep->match_prog_set = match_prog;
-	    MatchSetRef(ep->match_prog_set);
-	    return 1;
-	} else {
-	    /*
-	     * We ignore apply/3 and anything else.
-	     */
-	    return 0;
-	}
-    }
-    
+    ERTS_SMP_LC_ASSERT(erts_is_code_ix_locked());
+
     /*
-     * Currently no trace support for native code.
+     * Memory barriers will be issued for all processes *before*
+     * each of the stages below. (Unless the other schedulers
+     * are blocked, in which case memory barriers will be issued
+     * when they are awaken.)
      */
-    if (erts_is_native_break(ep->addressv[code_ix])) {
+
+    switch (finish_bp.current++) {
+    case 0:
+	/*
+	 * At this point, in all functions that are to be breakpointed,
+	 * a pointer to a GenericBp struct has already been added,
+	 *
+	 * Insert the new breakpoints (if any) into the
+	 * code. Different schedulers may see breakpoint instruction
+	 * at different times, but it does not matter since the newly
+	 * added breakpoints are disabled.
+	 */
+	if (finish_bp.install) {
+	    if (finish_bp.local) {
+		erts_install_breakpoints(&finish_bp.f);
+	    } else {
+		install_exp_breakpoints(&finish_bp.e);
+	    }
+	}
+	setup_bif_trace();
+	return 1;
+    case 1:
+	/*
+	 * Switch index for the breakpoint data, activating the staged
+	 * data. (Depending on the changes in the breakpoint data,
+	 * that could either activate breakpoints or disable
+	 * breakpoints.)
+	 */
+	erts_commit_staged_bp();
+	return 1;
+    case 2:
+	/*
+	 * Remove breakpoints instructions for disabled breakpoints
+	 * (if any).
+	 */
+	if (finish_bp.install) {
+	    if (finish_bp.local) {
+		uninstall_exp_breakpoints(&finish_bp.e);
+	    } else {
+		erts_uninstall_breakpoints(&finish_bp.f);
+	    }
+	} else {
+	    if (finish_bp.local) {
+		erts_uninstall_breakpoints(&finish_bp.f);
+	    } else {
+		uninstall_exp_breakpoints(&finish_bp.e);
+	    }
+	}
+	reset_bif_trace();
+	return 1;
+    case 3:
+	/*
+	 * Now all breakpoints have either been inserted or removed.
+	 * For all updated breakpoints, copy the active breakpoint
+	 * data to the staged breakpoint data to make them equal
+	 * (simplifying for the next time breakpoints are to be
+	 * updated).  If any breakpoints have been totally disabled,
+	 * deallocate the GenericBp structs for them.
+	 */
+	erts_consolidate_bif_bp_data();
+	clean_export_entries(&finish_bp.e);
+	erts_consolidate_bp_data(&finish_bp.e, 0);
+	erts_consolidate_bp_data(&finish_bp.f, 1);
+	erts_bp_free_matched_functions(&finish_bp.e);
+	erts_bp_free_matched_functions(&finish_bp.f);
 	return 0;
+    default:
+	ASSERT(0);
     }
+    return 0;
+}
 
-    ep->code[3] = (BeamInstr) em_call_traced_function;
-    ep->code[4] = (BeamInstr) ep->addressv[code_ix];
-    ep->addressv[code_ix] = ep->code+3;
-    ep->match_prog_set = match_prog;
-    MatchSetRef(ep->match_prog_set);
+static void
+install_exp_breakpoints(BpFunctions* f)
+{
+    const ErtsCodeIndex code_ix = erts_active_code_ix();
+    BpFunction* fp = f->matching;
+    Uint ne = f->matched;
+    Uint i;
+    Uint offset = offsetof(Export, code) + 3*sizeof(BeamInstr);
 
-    modp = erts_get_module(ep->code[0], code_ix);
-    ASSERT(modp);
-    modp->curr.num_traced_exports++;
-    return 1;
-}
+    for (i = 0; i < ne; i++) {
+	BeamInstr* pc = fp[i].pc;
+	Export* ep = (Export *) (((char *)pc)-offset);
 
-static void setup_bif_trace(int bif_index) {
-    Export *ep = bif_export[bif_index];
-    
-    ASSERT(ExportIsBuiltIn(ep));
-    ASSERT(ep->code[4]);
-    ep->code[4] = (BeamInstr) bif_table[bif_index].traced;
+	ep->addressv[code_ix] = pc;
+    }
 }
 
-static void set_trace_bif(int bif_index, void* match_prog) {
-    Export *ep = bif_export[bif_index];
-    
-#ifdef HARDDEBUG
-    erts_fprintf(stderr, "set_trace_bif: %T:%T/%bpu\n",
-		 ep->code[0], ep->code[1], ep->code[2]);
-#endif
-    ASSERT(ExportIsBuiltIn(ep));
-    MatchSetUnref(ep->match_prog_set);
-    ep->match_prog_set = match_prog;
-    MatchSetRef(ep->match_prog_set);
-}
+static void
+uninstall_exp_breakpoints(BpFunctions* f)
+{
+    const ErtsCodeIndex code_ix = erts_active_code_ix();
+    BpFunction* fp = f->matching;
+    Uint ne = f->matched;
+    Uint i;
+    Uint offset = offsetof(Export, code) + 3*sizeof(BeamInstr);
 
-/*
- * Reset function tracing for the given exported function.
- *
- * Return Value: 1 if entry refers to a BIF or loaded function,
- * 0 if the entry refers to a function not loaded.
- */
+    for (i = 0; i < ne; i++) {
+	BeamInstr* pc = fp[i].pc;
+	Export* ep = (Export *) (((char *)pc)-offset);
 
-static int
-reset_func_trace(Export* ep, ErtsCodeIndex code_ix)
-{
-    if (ep->addressv[code_ix] == ep->code+3) {
-	if (ep->code[3] == (BeamInstr) em_call_error_handler) {
-	    return 0;
-	} else if (ep->code[3] == (BeamInstr) em_call_traced_function) {
-	    Module* modp = erts_get_module(ep->code[0], code_ix);
-	    ASSERT(modp);
-	    modp->curr.num_traced_exports--;
-
-	    ep->addressv[code_ix] = (Uint *) ep->code[4];
-	    MatchSetUnref(ep->match_prog_set);
-	    ep->match_prog_set = NULL;
-	    return 1;
-	} else {
-	    /*
-	     * We ignore apply/3 and anything else.
-	     */
-	    return 0;
+	if (ep->addressv[code_ix] != pc) {
+	    continue;
 	}
+	ASSERT(*pc == (BeamInstr) BeamOp(op_jump_f));
+	ep->addressv[code_ix] = (BeamInstr *) ep->code[4];
     }
-    
-    /*
-     * Currently no trace support for native code.
-     */
-    if (erts_is_native_break(ep->addressv[code_ix])) {
-	return 0;
-    }
-    
-    /*
-     * Nothing to do, but the export entry matches.
-     */
+}
+
+static void
+clean_export_entries(BpFunctions* f)
+{
+    const ErtsCodeIndex code_ix = erts_active_code_ix();
+    BpFunction* fp = f->matching;
+    Uint ne = f->matched;
+    Uint i;
+    Uint offset = offsetof(Export, code) + 3*sizeof(BeamInstr);
+
+    for (i = 0; i < ne; i++) {
+	BeamInstr* pc = fp[i].pc;
+	Export* ep = (Export *) (((char *)pc)-offset);
 
-    return 1;
+	if (ep->addressv[code_ix] == pc) {
+	    continue;
+	}
+	if (*pc == (BeamInstr) BeamOp(op_jump_f)) {
+	    ep->code[3] = (BeamInstr) 0;
+	    ep->code[4] = (BeamInstr) 0;
+	}
+    }
 }
 
-static void reset_bif_trace(int bif_index) {
-    Export *ep = bif_export[bif_index];
-    
-    ASSERT(ExportIsBuiltIn(ep));
-    ASSERT(ep->code[4]);
-    ASSERT(! ep->match_prog_set);
-    ep->code[4] = (BeamInstr) bif_table[bif_index].f;
+static void
+setup_bif_trace(void)
+{
+    int i;
+
+    for (i = 0; i < BIF_SIZE; ++i) {
+	Export *ep = bif_export[i];
+	GenericBp* g = (GenericBp *) ep->fake_op_func_info_for_hipe[1];
+	if (g) {
+	    if (ExportIsBuiltIn(ep)) {
+		ASSERT(ep->code[4]);
+		ep->code[4] = (BeamInstr) bif_table[i].traced;
+	    }
+	}
+    }
 }
 
-static void clear_trace_bif(int bif_index) {
-    Export *ep = bif_export[bif_index];
-    
-#ifdef HARDDEBUG
-    erts_fprintf(stderr, "clear_trace_bif: %T:%T/%bpu\n",
-		 ep->code[0], ep->code[1], ep->code[2]);
-#endif
-    ASSERT(ExportIsBuiltIn(ep));
-    MatchSetUnref(ep->match_prog_set);
-    ep->match_prog_set = NULL;
+static void
+reset_bif_trace(void)
+{
+    int i;
+    ErtsBpIndex active = erts_active_bp_ix();
+
+    for (i = 0; i < BIF_SIZE; ++i) {
+	Export *ep = bif_export[i];
+	BeamInstr* pc = ep->code+3;
+	GenericBp* g = (GenericBp *) pc[-4];
+	if (g && g->data[active].flags == 0) {
+	    if (ExportIsBuiltIn(ep)) {
+		ASSERT(ep->code[4]);
+		ep->code[4] = (BeamInstr) bif_table[i].f;
+	    }
+	}
+    }
 }
 
 /*
diff --git a/erts/emulator/beam/erl_process.c b/erts/emulator/beam/erl_process.c
index 3d978f4994..e07c9ae2b0 100644
--- a/erts/emulator/beam/erl_process.c
+++ b/erts/emulator/beam/erl_process.c
@@ -530,6 +530,7 @@ dbg_chk_aux_work_val(erts_aint32_t value)
 #endif
 #ifdef ERTS_SMP
     valid |= ERTS_SSI_AUX_WORK_CODE_IX_ACTIVATION;
+    valid |= ERTS_SSI_AUX_WORK_FINISH_BP;
 #endif
 #ifdef ERTS_SSI_AUX_WORK_REAP_PORTS
     valid |= ERTS_SSI_AUX_WORK_REAP_PORTS;
@@ -1446,6 +1447,55 @@ handle_code_ix_activation(ErtsAuxWorkData *awdp, erts_aint32_t aux_work)
 }
 #endif /* ERTS_SMP */
 
+#ifdef ERTS_SMP
+void
+erts_notify_finish_breakpointing(Process* p)
+{
+    ErtsAuxWorkData* awdp = &p->scheduler_data->aux_work_data;
+
+    ASSERT(awdp->bp_ix_activation.stager == NULL);
+    awdp->bp_ix_activation.stager = p;
+    awdp->bp_ix_activation.thr_prgr = erts_thr_progress_later(awdp->esdp);
+    erts_thr_progress_wakeup(awdp->esdp, awdp->bp_ix_activation.thr_prgr);
+    erts_smp_proc_inc_refc(p);
+    set_aux_work_flags_wakeup_relb(p->scheduler_data->ssi,
+				   ERTS_SSI_AUX_WORK_FINISH_BP);
+}
+
+static erts_aint32_t
+handle_finish_bp(ErtsAuxWorkData* awdp, erts_aint32_t aux_work)
+{
+    ErtsThrPrgrVal current = haw_thr_prgr_current(awdp);
+
+    if (!erts_thr_progress_has_reached_this(current,
+					    awdp->bp_ix_activation.thr_prgr)) {
+	return aux_work & ~ERTS_SSI_AUX_WORK_FINISH_BP;
+    }
+    if (erts_finish_breakpointing()) { /* Not done */
+	/* Arrange for being called again */
+	awdp->bp_ix_activation.thr_prgr =
+	    erts_thr_progress_later(awdp->esdp);
+	erts_thr_progress_wakeup(awdp->esdp, awdp->bp_ix_activation.thr_prgr);
+    } else {			/* Done */
+	Process* p;
+
+	unset_aux_work_flags(awdp->ssi, ERTS_SSI_AUX_WORK_FINISH_BP);
+	p = awdp->bp_ix_activation.stager;
+#ifdef DEBUG
+	awdp->bp_ix_activation.stager = NULL;
+#endif
+	erts_smp_proc_lock(p, ERTS_PROC_LOCK_STATUS);
+	if (!ERTS_PROC_IS_EXITING(p)) {
+	    erts_resume(p, ERTS_PROC_LOCK_STATUS);
+	}
+	erts_smp_proc_unlock(p, ERTS_PROC_LOCK_STATUS);
+	erts_smp_proc_dec_refc(p);
+	erts_release_code_write_permission();
+    }
+    return aux_work & ~ERTS_SSI_AUX_WORK_FINISH_BP;
+}
+#endif /* ERTS_SMP */
+
 static ERTS_INLINE erts_aint32_t
 handle_fix_alloc(ErtsAuxWorkData *awdp, erts_aint32_t aux_work)
 {
@@ -1817,6 +1867,11 @@ handle_aux_work(ErtsAuxWorkData *awdp, erts_aint32_t orig_aux_work, int waiting)
     HANDLE_AUX_WORK(ERTS_SSI_AUX_WORK_REAP_PORTS,
 		    handle_reap_ports);
 
+#ifdef ERTS_SMP
+    HANDLE_AUX_WORK(ERTS_SSI_AUX_WORK_FINISH_BP,
+		    handle_finish_bp);
+#endif
+
     ERTS_DBG_CHK_AUX_WORK_VAL(aux_work);
 
 #ifdef ERTS_SMP
diff --git a/erts/emulator/beam/erl_process.h b/erts/emulator/beam/erl_process.h
index e9e5a3365f..93e71681da 100644
--- a/erts/emulator/beam/erl_process.h
+++ b/erts/emulator/beam/erl_process.h
@@ -283,6 +283,7 @@ typedef enum {
 #define ERTS_SSI_AUX_WORK_MSEG_CACHE_CHECK	(((erts_aint32_t) 1) << 10)
 #define ERTS_SSI_AUX_WORK_CODE_IX_ACTIVATION	(((erts_aint32_t) 1) << 11)
 #define ERTS_SSI_AUX_WORK_REAP_PORTS		(((erts_aint32_t) 1) << 12)
+#define ERTS_SSI_AUX_WORK_FINISH_BP	        (((erts_aint32_t) 1) << 13)
 
 typedef struct ErtsSchedulerSleepInfo_ ErtsSchedulerSleepInfo;
 
@@ -472,6 +473,12 @@ typedef struct {
 	ErtsThrPrgrVal thr_prgr;
     } code_ix_activation;
 #endif
+#ifdef ERTS_SMP
+    struct {
+	Process* stager;
+	ErtsThrPrgrVal thr_prgr;
+    } bp_ix_activation;
+#endif
 } ErtsAuxWorkData;
 
 struct ErtsSchedulerData_ {
@@ -1188,6 +1195,7 @@ void erts_notify_check_async_ready_queue(void *);
 #endif
 #ifdef ERTS_SMP
 void erts_notify_code_ix_activation(Process* p, ErtsThrPrgrVal later);
+void erts_notify_finish_breakpointing(Process* p);
 #endif
 void erts_schedule_misc_aux_work(int sched_id,
 				 void (*func)(void *),
diff --git a/erts/emulator/beam/erl_trace.c b/erts/emulator/beam/erl_trace.c
index c64c9d8eee..d04a91f18c 100644
--- a/erts/emulator/beam/erl_trace.c
+++ b/erts/emulator/beam/erl_trace.c
@@ -2134,185 +2134,6 @@ void save_calls(Process *p, Export *e)
     }
 }
 
-/* 
- * Entry point called by the trace wrap functions in erl_bif_wrap.c
- *
- * The trace wrap functions are themselves called through the export
- * entries instead of the original BIF functions.
- */
-Eterm
-erts_bif_trace(int bif_index, Process* p, Eterm* args, BeamInstr* I)
-{
-    Eterm result;
-    int meta = !!(erts_bif_trace_flags[bif_index] & BIF_TRACE_AS_META);
-
-    ERTS_SMP_CHK_HAVE_ONLY_MAIN_PROC_LOCK(p);
-
-    if (!ARE_TRACE_FLAGS_ON(p, F_TRACE_CALLS) && (! meta)) {
-	/* Warning! This is an Optimization. 
-	 *
-	 * If neither meta trace is active nor process trace flags then 
-	 * no tracing will occur. Doing the whole else branch will 
-	 * also do nothing, only slower.
-	 */
-	Eterm (*func)(Process*, Eterm*, BeamInstr*) = bif_table[bif_index].f;
-	result = func(p, args, I);
-    } else {
-	Eterm (*func)(Process*, Eterm*, BeamInstr*);
-	Export* ep = bif_export[bif_index];
-	Uint32 flags = 0, flags_meta = 0;
-	int global = !!(erts_bif_trace_flags[bif_index] & BIF_TRACE_AS_GLOBAL);
-	int local  = !!(erts_bif_trace_flags[bif_index] & BIF_TRACE_AS_LOCAL);
-	int time   = !!(erts_bif_trace_flags[bif_index] & BIF_TRACE_AS_CALL_TIME);
-	Eterm meta_tracer_pid = NIL;
-	int applying = (I == &(ep->code[3])); /* Yup, the apply code for a bif
-					       * is actually in the 
-					       * export entry */
-	BeamInstr *cp = p->cp;
-	
-	/* 
-	 * Make continuation pointer OK, it is not during direct BIF calls,
-	 * but it is correct during apply of bif.
-	 */
-	if (!applying) { 
-	    p->cp = I;
-	}
-	if (global || local) {
-	    flags = erts_call_trace(p, ep->code, ep->match_prog_set, args, 
-				       local, &p->tracer_proc);
-	}
-	if (meta) {
-	    flags_meta = erts_bif_mtrace(p, ep->code+3, args, local, 
-					 &meta_tracer_pid);
-	}
-	if (time) {
-	    BpDataTime *bdt;
-	    BeamInstr *pc = (BeamInstr *)ep->code+3;
-
-	    bdt = erts_get_active_time_break(pc);
-	    if (bdt) {
-		erts_trace_time_call(p, pc, bdt);
-	    }
-	}
-	/* Restore original continuation pointer (if changed). */
-	p->cp = cp;
-	
-	func = bif_table[bif_index].f;
-
-	result = func(p, args, I);
-	
-	if (applying && (flags & MATCH_SET_RETURN_TO_TRACE)) {
-	    BeamInstr i_return_trace      = beam_return_trace[0];
-	    BeamInstr i_return_to_trace   = beam_return_to_trace[0];
-	    BeamInstr i_return_time_trace = beam_return_time_trace[0];
-	    Eterm *cpp;
-	    /* Maybe advance cp to skip trace stack frames */
-	    for (cpp = p->stop;  ;  cp = cp_val(*cpp++)) {
-		if (*cp == i_return_trace) {
-		    /* Skip stack frame variables */
-		    while (is_not_CP(*cpp)) cpp++;
-		    cpp += 2; /* Skip return_trace parameters */
-		} else if (*cp == i_return_time_trace) {
-		    /* Skip stack frame variables */
-		    while (is_not_CP(*cpp)) cpp++;
-		    cpp += 1; /* Skip return_time_trace parameters */
-		} else if (*cp == i_return_to_trace) {
-		    /* A return_to trace message is going to be generated
-		     * by normal means, so we do not have to.
-		     */
-		    cp = NULL;
-		    break;
-		} else break;
-	    }
-	}
-	
-	/* Try to get these in the order 
-	 * they usually appear in normal code... */
-	if (is_non_value(result)) {
-	    Uint reason = p->freason;
-	    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;
-		    ASSERT(is_tuple(value));
-		    tp = tuple_val(value);
-		    value = tp[1];
-		}
-		if ((reason & EXF_THROWN) && (p->catches <= 0)) {
-		    value = TUPLE2(nocatch, am_nocatch, value);
-		    reason = EXC_ERROR;
-		}
-		/* Note: expand_error_value() could theoretically 
-		 * allocate on the heap, but not for any error
-		 * returned by a BIF, and it would do no harm,
-		 * just be annoying.
-		 */
-		value = expand_error_value(p, reason, value);
-		class = exception_tag[GET_EXC_CLASS(reason)];
-		
-		if (flags_meta & MATCH_SET_EXCEPTION_TRACE) {
-		    erts_trace_exception(p, ep->code, class, value,
-					 &meta_tracer_pid);
-		}
-		if (flags & MATCH_SET_EXCEPTION_TRACE) {
-		    erts_trace_exception(p, ep->code, class, value,
-					 &p->tracer_proc);
-		}
-		if ((flags & MATCH_SET_RETURN_TO_TRACE) && p->catches > 0) {
-		    /* can only happen if(local)*/
-		    Eterm *ptr = p->stop;
-		    ASSERT(is_CP(*ptr));
-		    ASSERT(ptr <= STACK_START(p));
-		    /* Search the nearest stack frame for a catch */
-		    while (++ptr < STACK_START(p)) {
-			if (is_CP(*ptr)) break;
-			if (is_catch(*ptr)) {
-			    if (applying) {
-				/* Apply of BIF, cp is in calling function */
-				if (cp) erts_trace_return_to(p, cp);
-			    } else {
-				/* Direct bif call, I points into 
-				 * calling function */
-				erts_trace_return_to(p, I);
-			    }
-			}
-		    }
-		}
-		UnUseTmpHeapNoproc(3);
-		if ((flags_meta|flags) & MATCH_SET_EXCEPTION_TRACE) {
-		    erts_smp_proc_lock(p, ERTS_PROC_LOCKS_ALL_MINOR);
-		    p->trace_flags |= F_EXCEPTION_TRACE;
-		    erts_smp_proc_unlock(p, ERTS_PROC_LOCKS_ALL_MINOR);
-		}
-	    }
-	} else {
-	    if (flags_meta & MATCH_SET_RX_TRACE) {
-		erts_trace_return(p, ep->code, result, &meta_tracer_pid);
-	    }
-	    /* MATCH_SET_RETURN_TO_TRACE cannot occur if(meta) */
-	    if (flags & MATCH_SET_RX_TRACE) {
-		erts_trace_return(p, ep->code, result, &p->tracer_proc);
-	    }
-	    if (flags & MATCH_SET_RETURN_TO_TRACE) { 
-		/* can only happen if(local)*/
-		if (applying) {
-		    /* Apply of BIF, cp is in calling function */
-		    if (cp) erts_trace_return_to(p, cp);
-		} else {
-		    /* Direct bif call, I points into calling function */
-		    erts_trace_return_to(p, I);
-		}
-	    }
-	}
-    }
-    ERTS_SMP_CHK_HAVE_ONLY_MAIN_PROC_LOCK(p);
-    return result;
-}
-
 /* Sends trace message:
  *    {trace_ts, Pid, What, Msg, Timestamp}
  * or {trace, Pid, What, Msg}
diff --git a/erts/emulator/beam/export.c b/erts/emulator/beam/export.c
index 229641cb32..6b5121f917 100644
--- a/erts/emulator/beam/export.c
+++ b/erts/emulator/beam/export.c
@@ -137,7 +137,6 @@ export_alloc(struct export_entry* tmpl_e)
 	obj->code[2] = tmpl->code[2];
 	obj->code[3] = (BeamInstr) em_call_error_handler;
 	obj->code[4] = 0;
-	obj->match_prog_set = NULL;
 
 	for (ix=0; ix<ERTS_NUM_CODE_IX; ix++) {
 	    obj->addressv[ix] = obj->code+3;
@@ -260,8 +259,9 @@ erts_find_function(Eterm m, Eterm f, unsigned int a, ErtsCodeIndex code_ix)
     struct export_entry* ee;
 
     ee = hash_get(&export_tables[code_ix].htable, init_template(&templ, m, f, a));
-    if (ee == NULL || (ee->ep->addressv[code_ix] == ee->ep->code+3 &&
-		       ee->ep->code[3] != (BeamInstr) em_call_traced_function)) {
+    if (ee == NULL ||
+	(ee->ep->addressv[code_ix] == ee->ep->code+3 &&
+	 ee->ep->code[3] != (BeamInstr) BeamOp(op_i_generic_breakpoint))) {
 	return NULL;
     }
     return ee->ep;
diff --git a/erts/emulator/beam/export.h b/erts/emulator/beam/export.h
index ec9fcb26f2..ee06e69aff 100644
--- a/erts/emulator/beam/export.h
+++ b/erts/emulator/beam/export.h
@@ -37,7 +37,6 @@
 typedef struct export
 {
     void* addressv[ERTS_NUM_CODE_IX];  /* Pointer to code for function. */
-    struct binary* match_prog_set; /* Match program for tracing. */
 
     BeamInstr fake_op_func_info_for_hipe[2]; /* MUST be just before code[] */
     /*
@@ -46,12 +45,12 @@ typedef struct export
      * code[2]: Arity (untagged integer).
      * code[3]: This entry is 0 unless the 'address' field points to it.
      *          Threaded code instruction to load function
-     *		(em_call_error_handler), execute BIF (em_apply_bif,
-     *		em_apply_apply), or call a traced function
-     *		(em_call_traced_function).
-     * code[4]: Function pointer to BIF function (for BIFs only)
+     *		(em_call_error_handler), execute BIF (em_apply_bif),
+     *		or a breakpoint instruction (op_i_generic_breakpoint).
+     * code[4]: Function pointer to BIF function (for BIFs only),
      *		or pointer to threaded code if the module has an
-     *		on_load function that has not been run yet.
+     *		on_load function that has not been run yet, or pointer
+     *          to code for function code[3] is a breakpont instruction.
      *		Otherwise: 0.
      */
     BeamInstr code[5];
diff --git a/erts/emulator/beam/global.h b/erts/emulator/beam/global.h
index 9e387c550f..c9be20322d 100755
--- a/erts/emulator/beam/global.h
+++ b/erts/emulator/beam/global.h
@@ -1682,10 +1682,10 @@ struct trace_pattern_flags {
 };
 extern const struct trace_pattern_flags erts_trace_pattern_flags_off;
 extern int erts_call_time_breakpoint_tracing;
-int erts_set_trace_pattern(Eterm* mfa, int specified, 
+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,
-			   Eterm meta_tracer_pid);
+			   Eterm meta_tracer_pid, int is_blocking);
 void
 erts_get_default_trace_pattern(int *trace_pattern_is_on,
 			       Binary **match_spec,
@@ -1694,6 +1694,7 @@ erts_get_default_trace_pattern(int *trace_pattern_is_on,
 			       Eterm *meta_tracer_pid);
 int erts_is_default_trace_enabled(void);
 void erts_bif_trace_init(void);
+int erts_finish_breakpointing(void);
 
 /*
 ** Call_trace uses this API for the parameter matching functions
@@ -1739,14 +1740,6 @@ extern void erts_match_prog_foreach_offheap(Binary *b,
 					   breakpoint functions */
 #define MATCH_SET_EXCEPTION_TRACE (0x4) /* exception trace requested */
 #define MATCH_SET_RX_TRACE (MATCH_SET_RETURN_TRACE|MATCH_SET_EXCEPTION_TRACE)
-/*
- * Flag values when tracing bif
- * Future note: flag field is 8 bits
- */
-#define BIF_TRACE_AS_LOCAL      (0x1)
-#define BIF_TRACE_AS_GLOBAL     (0x2)
-#define BIF_TRACE_AS_META       (0x4)
-#define BIF_TRACE_AS_CALL_TIME  (0x8)
 
 extern erts_driver_t vanilla_driver;
 extern erts_driver_t spawn_driver;
diff --git a/erts/emulator/beam/ops.tab b/erts/emulator/beam/ops.tab
index 8cb57a9c60..6764e88c81 100644
--- a/erts/emulator/beam/ops.tab
+++ b/erts/emulator/beam/ops.tab
@@ -519,7 +519,6 @@ apply_bif
 call_nif
 call_error_handler
 error_action_code
-call_traced_function
 return_trace
 
 #