Merge branch 'sverk/trapping-bin2term' into OTP_R16B03

* sverk/trapping-bin2term:
  erts: Adjust term_to_binary reduction factors
  erts: Yield after trapping term_to_binary if gc has been ordered
  erts: Let term_to_binary disable gc while trapping
  erts: Improve stress of binary_to_term in binary_SUITE
  erts: Fix bug in binary_to_term for compressed on halfword
  erts: Fix crash when binary_to_term throws badarg
  erts: Trapping memcpy in binary_to_term
  erts: Cleanup code for trapping binary_to_term
  erts: Add erlang wrappers to binary_to_term
  trapping uncompress
  trapping size calculation
  trapping binary_to_term/2
  trapping STRING_EXT
  trapping lists and tuples
  trapping binary_to_term passing binary_SUITE
  Parallel check_process_code when code_server purge a module
  Functionality for disabling garbage collection
  Use asynchronous check_process_code in code_parallel_SUITE
  Execution of system tasks in context of another process

Conflicts:
	erts/emulator/beam/external.c
	erts/emulator/beam/sys.h
	erts/emulator/test/binary_SUITE.erl
	erts/preloaded/ebin/erlang.beam
	erts/preloaded/ebin/erts_internal.beam

20 files changed, 2200 insertions, 408 deletions

diff --git a/erts/emulator/beam/atom.names b/erts/emulator/beam/atom.names
index c2f32ba089..71addedfdc 100644
--- a/erts/emulator/beam/atom.names
+++ b/erts/emulator/beam/atom.names
@@ -78,6 +78,7 @@ atom allocated_areas
 atom allocator
 atom allocator_sizes
 atom alloc_util_allocators
+atom allow_gc
 atom allow_passive_connect
 atom already_loaded
 atom amd64
@@ -114,6 +115,7 @@ atom binary_longest_prefix_trap
 atom binary_longest_suffix_trap
 atom binary_match_trap
 atom binary_matches_trap
+atom binary_to_term_trap
 atom block
 atom blocked
 atom bm
diff --git a/erts/emulator/beam/beam_bif_load.c b/erts/emulator/beam/beam_bif_load.c
index 649594a334..3f92c5b025 100644
--- a/erts/emulator/beam/beam_bif_load.c
+++ b/erts/emulator/beam/beam_bif_load.c
@@ -36,7 +36,7 @@
 #include "erl_thr_progress.h"
 
 static void set_default_trace_pattern(Eterm module);
-static Eterm check_process_code(Process* rp, Module* modp);
+static Eterm check_process_code(Process* rp, Module* modp, int allow_gc, int *redsp);
 static void delete_code(Module* modp);
 static void decrement_refc(BeamInstr* code);
 static int is_native(BeamInstr* code);
@@ -427,69 +427,82 @@ check_old_code_1(BIF_ALIST_1)
 }
 
 Eterm
-check_process_code_2(BIF_ALIST_2)
+erts_check_process_code(Process *c_p, Eterm module, int allow_gc, int *redsp)
 {
-    Process* rp;
     Module* modp;
+    Eterm res;
+    ErtsCodeIndex code_ix;
 
-    if (is_not_atom(BIF_ARG_2)) {
-	goto error;
-    }
-    if (is_internal_pid(BIF_ARG_1)) {
-	Eterm res;
-	ErtsCodeIndex code_ix;
-
-	code_ix = erts_active_code_ix();
-	modp = erts_get_module(BIF_ARG_2, code_ix);
-	if (modp == NULL) {		/* Doesn't exist. */
-	    return am_false;
-	}
-	erts_rlock_old_code(code_ix);
-	if (modp->old.code == NULL) { /* No old code. */
-	    erts_runlock_old_code(code_ix);
-	    return am_false;
-	}
-	erts_runlock_old_code(code_ix);
-	
-#ifdef ERTS_SMP
-	rp = erts_pid2proc_suspend(BIF_P, ERTS_PROC_LOCK_MAIN,
-				   BIF_ARG_1, ERTS_PROC_LOCK_MAIN);
-#else
-	rp = erts_pid2proc(BIF_P, 0, BIF_ARG_1, 0);
-#endif
-	if (!rp) {
-	    BIF_RET(am_false);
-	}
-	if (rp == ERTS_PROC_LOCK_BUSY) {
-	    ERTS_BIF_YIELD2(bif_export[BIF_check_process_code_2], BIF_P,
-			    BIF_ARG_1, BIF_ARG_2);
-	}
-	erts_rlock_old_code(code_ix);
-	if (modp->old.code != NULL) { /* must check again */
-	    res = check_process_code(rp, modp);
-	}
-	else {
-	    res = am_false;
-	}
-	erts_runlock_old_code(code_ix);
-#ifdef ERTS_SMP
-	if (BIF_P != rp) {
-	    erts_resume(rp, ERTS_PROC_LOCK_MAIN);
-	    erts_smp_proc_unlock(rp, ERTS_PROC_LOCK_MAIN);
+    (*redsp)++;
+
+    ASSERT(is_atom(module));
+
+    code_ix = erts_active_code_ix();
+    modp = erts_get_module(module, code_ix);
+    if (!modp)
+	return am_false;
+    erts_rlock_old_code(code_ix);
+    res = modp->old.code ? check_process_code(c_p, modp, allow_gc, redsp) : am_false;
+    erts_runlock_old_code(code_ix);
+
+    return res;
+}
+
+BIF_RETTYPE erts_internal_check_process_code_2(BIF_ALIST_2)
+{
+    int reds = 0;
+    Eterm res;
+    Eterm olist = BIF_ARG_2;
+    int allow_gc = 1;
+
+    if (is_not_atom(BIF_ARG_1))
+	goto badarg;
+
+    while (is_list(olist)) {
+	Eterm *lp = list_val(olist);
+	Eterm opt = CAR(lp);
+	if (is_tuple(opt)) {
+	    Eterm* tp = tuple_val(opt);
+	    switch (arityval(tp[0])) {
+	    case 2:
+		switch (tp[1]) {
+		case am_allow_gc:
+		    switch (tp[2]) {
+		    case am_false:
+			allow_gc = 0;
+			break;
+		    case am_true:
+			allow_gc = 1;
+			break;
+		    default:
+			goto badarg;
+		    }
+		    break;
+		default:
+		    goto badarg;
+		}
+		break;
+	    default:
+		goto badarg;
+	    }
 	}
-#endif
-	BIF_RET(res);
-    }
-    else if (is_external_pid(BIF_ARG_1)
-	     && external_pid_dist_entry(BIF_ARG_1) == erts_this_dist_entry) {
-	BIF_RET(am_false);
+	else
+	    goto badarg;
+	olist = CDR(lp);
     }
+    if (is_not_nil(olist))
+	goto badarg;
+
+    res = erts_check_process_code(BIF_P, BIF_ARG_1, allow_gc, &reds);
+
+    ASSERT(is_value(res));
+
+    BIF_RET2(res, reds);
 
- error:
+badarg:
     BIF_ERROR(BIF_P, BADARG);
 }
 
-
 BIF_RETTYPE delete_module_1(BIF_ALIST_1)
 {
     ErtsCodeIndex code_ix;
@@ -710,7 +723,7 @@ set_default_trace_pattern(Eterm module)
 }
 
 static Eterm
-check_process_code(Process* rp, Module* modp)
+check_process_code(Process* rp, Module* modp, int allow_gc, int *redsp)
 {
     BeamInstr* start;
     char* mod_start;
@@ -773,6 +786,16 @@ check_process_code(Process* rp, Module* modp)
 	}
     }
 
+    if (rp->flags & F_DISABLE_GC) {
+	/*
+	 * Cannot proceed. Process has disabled gc in order to
+	 * safely leave inconsistent data on the heap and/or
+	 * off heap lists. Need to wait for gc to be enabled
+	 * again.
+	 */ 
+	return THE_NON_VALUE;
+    }
+
     /*
      * See if there are funs that refer to the old version of the module.
      */
@@ -786,6 +809,8 @@ check_process_code(Process* rp, Module* modp)
 		if (done_gc) {
 		    return am_true;
 		} else {
+		    if (!allow_gc)
+			return am_aborted;
 		    /*
 		    * Try to get rid of this fun by garbage collecting.
 		    * Clear both fvalue and ftrace to make sure they
@@ -796,7 +821,7 @@ check_process_code(Process* rp, Module* modp)
 		    rp->ftrace = NIL;
 		    done_gc = 1;
 		    FLAGS(rp) |= F_NEED_FULLSWEEP;
-		    (void) erts_garbage_collect(rp, 0, rp->arg_reg, rp->arity);
+		    *redsp += erts_garbage_collect(rp, 0, rp->arg_reg, rp->arity);
 		    goto rescan;
 		}
 	    }
@@ -850,6 +875,9 @@ check_process_code(Process* rp, Module* modp)
 	    Uint lit_size;
 	    struct erl_off_heap_header* oh;
 
+	    if (!allow_gc)
+		return am_aborted;
+
 	    /*
 	     * Try to get rid of constants by by garbage collecting.
 	     * Clear both fvalue and ftrace.
@@ -859,11 +887,12 @@ check_process_code(Process* rp, Module* modp)
 	    rp->ftrace = NIL;
 	    done_gc = 1;
 	    FLAGS(rp) |= F_NEED_FULLSWEEP;
-	    (void) erts_garbage_collect(rp, 0, rp->arg_reg, rp->arity);
+	    *redsp += erts_garbage_collect(rp, 0, rp->arg_reg, rp->arity);
 	    literals = (Eterm *) modp->old.code[MI_LITERALS_START];
 	    lit_size = (Eterm *) modp->old.code[MI_LITERALS_END] - literals;
 	    oh = (struct erl_off_heap_header *)
 		modp->old.code[MI_LITERALS_OFF_HEAP];
+	    *redsp += lit_size / 10; /* Need, better value... */
 	    erts_garbage_collect_literals(rp, literals, lit_size, oh);
 	}
     }
diff --git a/erts/emulator/beam/bif.c b/erts/emulator/beam/bif.c
index 13d31285b2..96666d98ed 100644
--- a/erts/emulator/beam/bif.c
+++ b/erts/emulator/beam/bif.c
@@ -3764,45 +3764,6 @@ BIF_RETTYPE now_0(BIF_ALIST_0)
 
 /**********************************************************************/
 
-BIF_RETTYPE garbage_collect_1(BIF_ALIST_1)
-{
-    int reds;
-    Process *rp;
-
-    if (is_not_pid(BIF_ARG_1)) {
-	BIF_ERROR(BIF_P, BADARG);
-    }
-
-    if (BIF_P->common.id == BIF_ARG_1)
-	rp = BIF_P;
-    else {
-#ifdef ERTS_SMP
-	rp = erts_pid2proc_suspend(BIF_P, ERTS_PROC_LOCK_MAIN,
-				   BIF_ARG_1, ERTS_PROC_LOCK_MAIN);
-	if (rp == ERTS_PROC_LOCK_BUSY)
-	    ERTS_BIF_YIELD1(bif_export[BIF_garbage_collect_1], BIF_P, BIF_ARG_1);
-#else
-	rp = erts_proc_lookup(BIF_ARG_1);
-#endif
-	if (!rp)
-	    BIF_RET(am_false);
-    }
-
-    /* The GC cost is taken for the process executing this BIF. */
-
-    FLAGS(rp) |= F_NEED_FULLSWEEP;
-    reds = erts_garbage_collect(rp, 0, rp->arg_reg, rp->arity);
-
-#ifdef ERTS_SMP
-    if (BIF_P != rp) {
-	erts_resume(rp, ERTS_PROC_LOCK_MAIN);
-	erts_smp_proc_unlock(rp, ERTS_PROC_LOCK_MAIN);
-    }
-#endif
-
-    BIF_RET2(am_true, reds);
-}
-
 BIF_RETTYPE garbage_collect_0(BIF_ALIST_0)
 {
     int reds;
diff --git a/erts/emulator/beam/bif.tab b/erts/emulator/beam/bif.tab
index 6037c08dd8..6c5fd17902 100644
--- a/erts/emulator/beam/bif.tab
+++ b/erts/emulator/beam/bif.tab
@@ -45,8 +45,6 @@ bif erlang:apply/3
 bif erlang:atom_to_list/1
 bif erlang:binary_to_list/1
 bif erlang:binary_to_list/3
-bif erlang:binary_to_term/1	
-bif erlang:check_process_code/2
 bif erlang:crc32/1
 bif erlang:crc32/2
 bif erlang:crc32_combine/3
@@ -67,7 +65,6 @@ bif erlang:float_to_list/1
 bif erlang:float_to_list/2
 bif erlang:fun_info/2
 bif erlang:garbage_collect/0
-bif erlang:garbage_collect/1
 bif erlang:get/0
 bif erlang:get/1
 bif erlang:get_keys/1
@@ -154,6 +151,12 @@ bif erts_internal:port_command/3
 bif erts_internal:port_control/3
 bif erts_internal:port_close/1
 bif erts_internal:port_connect/2
+bif erts_internal:binary_to_term/1
+bif erts_internal:binary_to_term/2
+
+bif erts_internal:request_system_task/3
+bif erts_internal:check_process_code/2
+
 
 # inet_db support
 bif erlang:port_set_data/2
@@ -477,11 +480,6 @@ bif erlang:call_on_load_function/1
 bif erlang:finish_after_on_load/2
 
 #
-# New Bifs in R13B4
-#
-bif erlang:binary_to_term/2
-
-#
 # The binary match bifs (New in R14A - EEP9)
 #
 
diff --git a/erts/emulator/beam/break.c b/erts/emulator/beam/break.c
index b7e1092907..7d4f52ee23 100644
--- a/erts/emulator/beam/break.c
+++ b/erts/emulator/beam/break.c
@@ -112,10 +112,12 @@ process_killer(void)
 		    erts_smp_proc_lock(rp, rp_locks);
 		    state = erts_smp_atomic32_read_acqb(&rp->state);
 		    if (state & (ERTS_PSFLG_FREE
-				  | ERTS_PSFLG_EXITING
-				  | ERTS_PSFLG_ACTIVE
-				  | ERTS_PSFLG_IN_RUNQ
-				  | ERTS_PSFLG_RUNNING)) {
+				 | ERTS_PSFLG_EXITING
+				 | ERTS_PSFLG_ACTIVE
+				 | ERTS_PSFLG_ACTIVE_SYS
+				 | ERTS_PSFLG_IN_RUNQ
+				 | ERTS_PSFLG_RUNNING
+				 | ERTS_PSFLG_RUNNING_SYS)) {
 			erts_printf("Can only kill WAITING processes this way\n");
 		    }
 		    else {
diff --git a/erts/emulator/beam/erl_alloc.types b/erts/emulator/beam/erl_alloc.types
index bb5eba80be..32308fae9b 100644
--- a/erts/emulator/beam/erl_alloc.types
+++ b/erts/emulator/beam/erl_alloc.types
@@ -268,6 +268,8 @@ type	PROC_INTERVAL	LONG_LIVED	SYSTEM		process_interval
 type    BUSY_CALLER_TAB	SHORT_LIVED	SYSTEM		busy_caller_table
 type    BUSY_CALLER	SHORT_LIVED	SYSTEM		busy_caller
 type	PORT_DATA_HEAP	STANDARD	SYSTEM		port_data_heap
+type	PROC_SYS_TSK	SHORT_LIVED	PROCESSES	proc_sys_task
+type	PROC_SYS_TSK_QS	SHORT_LIVED	PROCESSES	proc_sys_task_queues
 
 +if threads_no_smp
 # Need thread safe allocs, but std_alloc and fix_alloc are not;
diff --git a/erts/emulator/beam/erl_bif_info.c b/erts/emulator/beam/erl_bif_info.c
index d7f1e2d971..1da9358334 100755
--- a/erts/emulator/beam/erl_bif_info.c
+++ b/erts/emulator/beam/erl_bif_info.c
@@ -3603,6 +3603,20 @@ BIF_RETTYPE erts_debug_set_internal_state_2(BIF_ALIST_2)
 		BIF_RET(am_true);
 	    }
 	}
+	else if (ERTS_IS_ATOM_STR("gc_state", BIF_ARG_1)) {
+	    /* Used by process_SUITE (emulator) */
+	    int res, enable;
+
+	    switch (BIF_ARG_2) {
+	    case am_true: enable = 1; break;
+	    case am_false: enable = 0; break;
+	    default: BIF_ERROR(BIF_P, BADARG); break;
+	    }
+ 
+            res = (BIF_P->flags & F_DISABLE_GC) ? am_false : am_true;
+	    erts_set_gc_state(BIF_P, enable);
+	    BIF_RET(res);
+	}
 	else if (ERTS_IS_ATOM_STR("send_fake_exit_signal", BIF_ARG_1)) {
 	    /* Used by signal_SUITE (emulator) */
 
diff --git a/erts/emulator/beam/erl_debug.c b/erts/emulator/beam/erl_debug.c
index b90d00f236..873a9860da 100644
--- a/erts/emulator/beam/erl_debug.c
+++ b/erts/emulator/beam/erl_debug.c
@@ -1,7 +1,7 @@
 /*
  * %CopyrightBegin%
  *
- * Copyright Ericsson AB 1998-2012. All Rights Reserved.
+ * Copyright Ericsson AB 1998-2013. All Rights Reserved.
  *
  * The contents of this file are subject to the Erlang Public License,
  * Version 1.1, (the "License"); you may not use this file except in
@@ -299,6 +299,9 @@ void erts_check_for_holes(Process* p)
     ErlHeapFragment* hf;
     Eterm* start;
 
+    if (p->flags & F_DISABLE_GC)
+	return;
+
     start = p->last_htop ? p->last_htop : HEAP_START(p);
     check_memory(start, HEAP_TOP(p));
     p->last_htop = HEAP_TOP(p);
diff --git a/erts/emulator/beam/erl_gc.c b/erts/emulator/beam/erl_gc.c
index e89725c190..c5585d39e8 100644
--- a/erts/emulator/beam/erl_gc.c
+++ b/erts/emulator/beam/erl_gc.c
@@ -400,10 +400,16 @@ erts_garbage_collect(Process* p, int need, Eterm* objv, int nobj)
     Uint reclaimed_now = 0;
     int done = 0;
     Uint ms1, s1, us1;
-    ErtsSchedulerData *esdp = erts_get_scheduler_data();
+    ErtsSchedulerData *esdp;
 #ifdef USE_VM_PROBES
     DTRACE_CHARBUF(pidbuf, DTRACE_TERM_BUF_SIZE);
 #endif
+
+    if (p->flags & F_DISABLE_GC)
+	return 1;
+
+    esdp = erts_get_scheduler_data();
+
     if (IS_TRACED_FL(p, F_TRACE_GC)) {
         trace_gc(p, am_gc_start);
     }
@@ -532,6 +538,9 @@ erts_garbage_collect_hibernate(Process* p)
     Uint area_size;
     Sint offs;
 
+    if (p->flags & F_DISABLE_GC)
+	ERTS_INTERNAL_ERROR("GC disabled");
+
     /*
      * Preliminaries.
      */
@@ -667,6 +676,8 @@ erts_garbage_collect_literals(Process* p, Eterm* literals,
     Uint n;
     struct erl_off_heap_header** prev;
 
+    if (p->flags & F_DISABLE_GC)
+	return;
     /*
      * Set GC state.
      */
diff --git a/erts/emulator/beam/erl_message.h b/erts/emulator/beam/erl_message.h
index 771eba431f..0f3bb8d281 100644
--- a/erts/emulator/beam/erl_message.h
+++ b/erts/emulator/beam/erl_message.h
@@ -46,6 +46,11 @@ typedef struct erl_off_heap {
     Uint64 overhead;     /* Administrative overhead (used to force GC). */
 } ErlOffHeap;
 
+#define ERTS_INIT_OFF_HEAP(OHP)			\
+    do {					\
+	(OHP)->first = NULL;			\
+	(OHP)->overhead = 0;			\
+    } while (0)
 #include "external.h"
 #include "erl_process.h"
 
diff --git a/erts/emulator/beam/erl_process.c b/erts/emulator/beam/erl_process.c
index 1efd070afd..21fd8dd50a 100644
--- a/erts/emulator/beam/erl_process.c
+++ b/erts/emulator/beam/erl_process.c
@@ -283,6 +283,40 @@ struct erts_system_profile_flags_t erts_system_profile_flags;
 #error "Need to store process_count in another type"
 #endif
 
+typedef enum {
+    ERTS_PSTT_GC,	/* Garbage Collect */
+    ERTS_PSTT_CPC	/* Check Process Code */
+} ErtsProcSysTaskType;
+
+#define ERTS_MAX_PROC_SYS_TASK_ARGS 2
+
+struct ErtsProcSysTask_ {
+    ErtsProcSysTask *next;
+    ErtsProcSysTask *prev;
+    ErtsProcSysTaskType type;
+    Eterm requester;
+    Eterm reply_tag;
+    Eterm req_id;
+    Uint req_id_sz;
+    Eterm arg[ERTS_MAX_PROC_SYS_TASK_ARGS];
+    ErlOffHeap off_heap;
+    Eterm heap[1];
+};
+
+#define ERTS_PROC_SYS_TASK_SIZE(HSz) \
+    (sizeof(ErtsProcSysTask) - sizeof(Eterm) + sizeof(Eterm)*(HSz))
+
+struct ErtsProcSysTaskQs_ {
+    int qmask;
+    int ncount;
+    ErtsProcSysTask *q[ERTS_NO_PROC_PRIO_LEVELS];
+};
+
+ERTS_SCHED_PREF_QUICK_ALLOC_IMPL(proc_sys_task_queues,
+				 ErtsProcSysTaskQs,
+				 50,
+				 ERTS_ALC_T_PROC_SYS_TSK_QS)
+
 ERTS_SCHED_PREF_QUICK_ALLOC_IMPL(misc_op_list,
 				 ErtsMiscOpList,
 				 10,
@@ -353,6 +387,14 @@ static void aux_work_timeout_early_init(int no_schedulers);
 static void aux_work_timeout_late_init(void);
 static void setup_aux_work_timer(void);
 
+static int execute_sys_tasks(Process *c_p,
+			     erts_aint32_t *statep,
+			     int in_reds);
+static int cleanup_sys_tasks(Process *c_p,
+			     erts_aint32_t in_state,
+			     int in_reds);
+
+
 #if defined(DEBUG) || 0
 #define ERTS_DBG_CHK_AUX_WORK_VAL(V) dbg_chk_aux_work_val((V))
 static void
@@ -471,6 +513,11 @@ erts_pre_init_process(void)
      erts_psd_required_locks[ERTS_PSD_CALL_TIME_BP].set_locks
 	 = ERTS_PSD_CALL_TIME_BP_SET_LOCKS;
 
+     erts_psd_required_locks[ERTS_PSD_DELAYED_GC_TASK_QS].get_locks
+	 = ERTS_PSD_DELAYED_GC_TASK_QS_GET_LOCKS;
+     erts_psd_required_locks[ERTS_PSD_DELAYED_GC_TASK_QS].set_locks
+	 = ERTS_PSD_DELAYED_GC_TASK_QS_SET_LOCKS;
+
      /* Check that we have locks for all entries */
      for (ix = 0; ix < ERTS_PSD_SIZE; ix++) {
 	 ERTS_SMP_LC_ASSERT(erts_psd_required_locks[ix].get_locks);
@@ -2848,7 +2895,7 @@ dequeue_process(ErtsRunQueue *runq, int prio_q, erts_aint32_t *statep)
     if (statep)
 	*statep = state;
 
-    prio = (int) (ERTS_PSFLG_PRIO_MASK & state);
+    prio = (int) ERTS_PSFLGS_GET_PRQ_PRIO(state);
 
     rqi = &runq->procs.prio_info[prio];
 
@@ -2997,7 +3044,7 @@ immigrate(ErtsRunQueue *c_rq, ErtsMigrationPath *mp)
 		    erts_aint32_t state;
 		    state = erts_smp_atomic32_read_acqb(&proc->state);
 		    if (!(ERTS_PSFLG_BOUND & state)
-			&& (prio == (int) (ERTS_PSFLG_PRIO_MASK & state))) {
+			&& (prio == (int) ERTS_PSFLGS_GET_PRQ_PRIO(state))) {
 			ErtsRunQueueInfo *rqi = &rq->procs.prio_info[prio];
 			unqueue_process(rq, rpq, rqi, prio, prev_proc, proc);
 			erts_smp_runq_unlock(rq);
@@ -3079,7 +3126,7 @@ schedule_bound_processes(ErtsRunQueue *rq,
     while (proc) {
 	erts_aint32_t state = erts_smp_atomic32_read_acqb(&proc->state);
 	next = proc->next;
-	enqueue_process(rq, (int) (ERTS_PSFLG_PRIO_MASK & state), proc);
+	enqueue_process(rq, (int) ERTS_PSFLGS_GET_PRQ_PRIO(state), proc);
 	proc = next;
     }
 }
@@ -3184,7 +3231,7 @@ evacuate_run_queue(ErtsRunQueue *rq,
 		sbpp->last = proc;
 	    }
 	    else {
-		int prio = (int) (ERTS_PSFLG_PRIO_MASK & state);
+		int prio = (int) ERTS_PSFLGS_GET_PRQ_PRIO(state);
 		erts_smp_runq_unlock(rq);
 
 		to_rq = mp->prio[prio].runq;
@@ -3257,7 +3304,7 @@ try_steal_task_from_victim(ErtsRunQueue *rq, int *rq_lockedp, ErtsRunQueue *vrq,
 	    erts_aint32_t state = erts_smp_atomic32_read_acqb(&proc->state);
 	    if (!(ERTS_PSFLG_BOUND & state)) {
 		/* Steal process */
-		int prio = (int) (ERTS_PSFLG_PRIO_MASK & state);
+		int prio = (int) ERTS_PSFLGS_GET_PRQ_PRIO(state);
 		ErtsRunQueueInfo *rqi = &vrq->procs.prio_info[prio];
 		unqueue_process(vrq, rpq, rqi, prio, prev_proc, proc);
 		erts_smp_runq_unlock(vrq);
@@ -4575,6 +4622,7 @@ erts_init_scheduling(int no_schedulers, int no_schedulers_online)
 #endif
 
     init_misc_op_list_alloc();
+    init_proc_sys_task_queues_alloc();
 
 #ifdef ERTS_SMP
     set_wakeup_other_data();
@@ -4858,46 +4906,166 @@ erts_get_scheduler_data(void)
 
 #endif
 
+static Process *
+make_proxy_proc(Process *prev_proxy, Process *proc, erts_aint32_t prio)
+{
+    erts_aint32_t state;
+    Process *proxy;
+#ifdef ERTS_SMP
+    ErtsRunQueue *rq = RUNQ_READ_RQ(&proc->run_queue);
+#endif
+
+    state = (ERTS_PSFLG_PROXY
+	     | ERTS_PSFLG_IN_RUNQ
+	     | (((erts_aint32_t) 1) << (prio + ERTS_PSFLGS_IN_PRQ_MASK_OFFSET))
+	     | (prio << ERTS_PSFLGS_PRQ_PRIO_OFFSET)
+	     | (prio << ERTS_PSFLGS_USR_PRIO_OFFSET)
+	     | (prio << ERTS_PSFLGS_ACT_PRIO_OFFSET));
+
+    if (prev_proxy) {
+	proxy = prev_proxy;
+	ASSERT(erts_smp_atomic32_read_nob(&proxy->state) & ERTS_PSFLG_PROXY);
+	erts_smp_atomic32_set_nob(&proxy->state, state);
+#ifdef ERTS_SMP
+	RUNQ_SET_RQ(&proc->run_queue, rq);
+#endif
+    }
+    else {
+	proxy = erts_alloc(ERTS_ALC_T_PROC, sizeof(Process));
+#ifdef DEBUG
+	{
+	    int i;
+	    Uint32 *ui32 = (Uint32 *) (char *) proxy;
+	    for (i = 0; i < sizeof(Process)/sizeof(Uint32); i++)
+		ui32[i] = (Uint32) 0xdeadbeef;
+	}
+#endif
+	erts_smp_atomic32_init_nob(&proxy->state, state);
+#ifdef ERTS_SMP
+	erts_smp_atomic_init_nob(&proxy->run_queue,
+				 erts_smp_atomic_read_nob(&proc->run_queue));
+#endif
+    }
+
+    proxy->common.id = proc->common.id;
+
+    return proxy;
+}
+
+static ERTS_INLINE void
+free_proxy_proc(Process *proxy)
+{
+    ASSERT(erts_smp_atomic32_read_nob(&proxy->state) & ERTS_PSFLG_PROXY);
+    erts_free(ERTS_ALC_T_PROC, proxy);
+}
+
+
+static ERTS_INLINE int
+check_enqueue_in_prio_queue(erts_aint32_t *prq_prio_p,
+			    erts_aint32_t *newp,
+			    erts_aint32_t actual)
+{
+    erts_aint32_t aprio, qbit, max_qbit;
+
+    aprio = (actual >> ERTS_PSFLGS_ACT_PRIO_OFFSET) & ERTS_PSFLGS_PRIO_MASK;
+    qbit = 1 << aprio;
+
+    *prq_prio_p = aprio;
+
+    max_qbit = (actual >> ERTS_PSFLGS_IN_PRQ_MASK_OFFSET) & ERTS_PSFLGS_QMASK;
+    max_qbit |= 1 << ERTS_PSFLGS_QMASK_BITS;
+    max_qbit &= -max_qbit;
+    /*
+     * max_qbit now either contain bit set for highest prio queue or a bit
+     * out of range (which will have a value larger than valid range).
+     */
+
+    if (qbit >= max_qbit)
+	return 0; /* Already queued in higher or equal prio */
+
+    /* Need to enqueue (if already enqueued, it is in lower prio) */
+    *newp |= qbit << ERTS_PSFLGS_IN_PRQ_MASK_OFFSET;
+
+    if ((actual & (ERTS_PSFLG_IN_RUNQ|ERTS_PSFLGS_USR_PRIO_MASK))
+	!= (aprio << ERTS_PSFLGS_USR_PRIO_OFFSET)) {
+	/*
+	 * Process struct already enqueued, or actual prio not
+	 * equal to user prio, i.e., enqueue using proxy.
+	 */
+	return -1;
+    }
+
+    /*
+     * Enqueue using process struct.
+     */
+    *newp &= ~ERTS_PSFLGS_PRQ_PRIO_MASK;
+    *newp |= ERTS_PSFLG_IN_RUNQ | (aprio << ERTS_PSFLGS_PRQ_PRIO_OFFSET);
+    return 1;
+}
+
 /*
  * scheduler_out_process() return with c_rq locked.
  */
 static ERTS_INLINE int
-schedule_out_process(ErtsRunQueue *c_rq, erts_aint32_t state, Process *p)
+schedule_out_process(ErtsRunQueue *c_rq, erts_aint32_t state, Process *p, Process *proxy)
 {
-    erts_aint32_t a, e, n;
+    erts_aint32_t a, e, n, enq_prio = -1;
     int res = 0;
+    int enqueue; /* < 0 -> use proxy */
 
     a = state;
 
     while (1) {
 	n = e = a;
 
-	ASSERT(a & ERTS_PSFLG_RUNNING);
-	ASSERT(!(a & ERTS_PSFLG_IN_RUNQ));
+	ASSERT(a & (ERTS_PSFLG_RUNNING|ERTS_PSFLG_RUNNING_SYS));
 
-	n &= ~ERTS_PSFLG_RUNNING;
-	if ((a & (ERTS_PSFLG_ACTIVE|ERTS_PSFLG_SUSPENDED)) == ERTS_PSFLG_ACTIVE)
-	    n |= ERTS_PSFLG_IN_RUNQ;
+	enqueue = 0;
+
+	n &= ~(ERTS_PSFLG_RUNNING|ERTS_PSFLG_RUNNING_SYS);
+	if (a & ERTS_PSFLG_ACTIVE_SYS
+	    || (a & (ERTS_PSFLG_ACTIVE|ERTS_PSFLG_SUSPENDED)) == ERTS_PSFLG_ACTIVE) {
+	    enqueue = check_enqueue_in_prio_queue(&enq_prio, &n, a);
+	}
 	a = erts_smp_atomic32_cmpxchg_mb(&p->state, n, e);
 	if (a == e)
 	    break;
     }
 
-    if (!(n & ERTS_PSFLG_IN_RUNQ)) {
-	if (erts_system_profile_flags.runnable_procs)
-	    profile_runnable_proc(p, am_inactive);
+    if (!enqueue) {
+
+	if (erts_system_profile_flags.runnable_procs) {
+
+	    if (!(a & ERTS_PSFLG_ACTIVE_SYS)
+		&& (!(a & ERTS_PSFLG_ACTIVE)
+		    || (a & ERTS_PSFLG_SUSPENDED))) {
+		/* Process inactive */
+		profile_runnable_proc(p, am_inactive);
+	    }
+	}
+
+	if (proxy)
+	    free_proxy_proc(proxy);
     }
     else {
-	int prio = (int) (ERTS_PSFLG_PRIO_MASK & n);
+	Process *sched_p;
 	ErtsRunQueue *runq = erts_get_runq_proc(p);
 
-	ASSERT(!(n & ERTS_PSFLG_SUSPENDED));
+	ASSERT(!(n & ERTS_PSFLG_SUSPENDED) || (n & ERTS_PSFLG_ACTIVE_SYS));
+
+	if (enqueue < 0)
+	    sched_p = make_proxy_proc(proxy, p, enq_prio);
+	else {
+	    sched_p = p;
+	    if (proxy)
+		free_proxy_proc(proxy);
+	}
 
 #ifdef ERTS_SMP
 	if (!(ERTS_PSFLG_BOUND & n)) {
-	    ErtsRunQueue *new_runq = erts_check_emigration_need(runq, prio);
+	    ErtsRunQueue *new_runq = erts_check_emigration_need(runq, enq_prio);
 	    if (new_runq) {
-		RUNQ_SET_RQ(&p->run_queue, new_runq);
+		RUNQ_SET_RQ(&sched_p->run_queue, new_runq);
 		runq = new_runq;
 	    }
 	}
@@ -4908,7 +5076,7 @@ schedule_out_process(ErtsRunQueue *c_rq, erts_aint32_t state, Process *p)
 	erts_smp_runq_lock(runq);
 
 	/* Enqueue the process */
-	enqueue_process(runq, prio, p);
+	enqueue_process(runq, (int) enq_prio, sched_p);
 
 	if (runq == c_rq)
 	    return res;
@@ -4920,14 +5088,13 @@ schedule_out_process(ErtsRunQueue *c_rq, erts_aint32_t state, Process *p)
 }
 
 static ERTS_INLINE void
-add2runq(Process *p, erts_aint32_t state)
+add2runq(Process *p, erts_aint32_t state, erts_aint32_t prio)
 {
-    int prio = (int) (ERTS_PSFLG_PRIO_MASK & state);
     ErtsRunQueue *runq = erts_get_runq_proc(p);
 
 #ifdef ERTS_SMP
     if (!(ERTS_PSFLG_BOUND & state)) {
-	ErtsRunQueue *new_runq = erts_check_emigration_need(runq, prio);
+	ErtsRunQueue *new_runq = erts_check_emigration_need(runq, (int) prio);
 	if (new_runq) {
 	    RUNQ_SET_RQ(&p->run_queue, new_runq);
 	    runq = new_runq;
@@ -4939,101 +5106,236 @@ add2runq(Process *p, erts_aint32_t state)
     erts_smp_runq_lock(runq);
 
     /* Enqueue the process */
-    enqueue_process(runq, prio, p);
+    enqueue_process(runq, (int) prio, p);
 
     erts_smp_runq_unlock(runq);
     smp_notify_inc_runq(runq);
 
 }
 
-static ERTS_INLINE void
-schedule_process(Process *p, erts_aint32_t state, int active_enq)
+static ERTS_INLINE int
+change_proc_schedule_state(Process *p,
+			   erts_aint32_t clear_state_flags,
+			   erts_aint32_t set_state_flags,
+			   erts_aint32_t *statep,
+			   erts_aint32_t *enq_prio_p)
 {
-    erts_aint32_t a = state, n;
+    /*
+     * NOTE: ERTS_PSFLG_RUNNING, ERTS_PSFLG_RUNNING_SYS and
+     *       ERTS_PSFLG_ACTIVE_SYS are not allowed to be
+     *       altered by this function!
+     */
+    erts_aint32_t a = *statep, n;
+    int enqueue; /* < 0 -> use proxy */
+
+    ASSERT(!(a & ERTS_PSFLG_PROXY));
+    ASSERT((clear_state_flags & (ERTS_PSFLG_RUNNING
+				 | ERTS_PSFLG_RUNNING_SYS
+				 | ERTS_PSFLG_ACTIVE_SYS)) == 0);
+    ASSERT((set_state_flags & (ERTS_PSFLG_RUNNING
+			       | ERTS_PSFLG_RUNNING_SYS
+			       | ERTS_PSFLG_ACTIVE_SYS)) == 0);
 
     while (1) {
 	erts_aint32_t e;
 	n = e = a;
 
+	enqueue = 0;
+
 	if (a & ERTS_PSFLG_FREE)
-	    return; /* We don't want to schedule free processes... */
-	n |= ERTS_PSFLG_ACTIVE;
-	if (!(a & (ERTS_PSFLG_SUSPENDED|ERTS_PSFLG_RUNNING)))
-	    n |= ERTS_PSFLG_IN_RUNQ;
+	    break; /* We don't want to schedule free processes... */
+
+	if (clear_state_flags)
+	    n &= ~clear_state_flags;
+
+	if (set_state_flags)
+	    n |= set_state_flags;
+
+	if ((n & (ERTS_PSFLG_SUSPENDED
+		  | ERTS_PSFLG_RUNNING
+		  | ERTS_PSFLG_RUNNING_SYS
+		  | ERTS_PSFLG_IN_RUNQ
+		  | ERTS_PSFLG_ACTIVE)) == ERTS_PSFLG_ACTIVE) {
+	    /*
+	     * Active and seemingly need to be enqueued, but
+	     * process may be in a run queue via proxy, need
+	     * further inspection...
+	     */
+	    enqueue = check_enqueue_in_prio_queue(enq_prio_p, &n, a);
+	}
+
 	a = erts_smp_atomic32_cmpxchg_mb(&p->state, n, e);
 	if (a == e)
 	    break;
-	if (!active_enq && (a & ERTS_PSFLG_ACTIVE))
-	    return; /* Someone else activated process ... */
+	if (enqueue == 0 && n == a)
+	    break;
     }
 
-    if (erts_system_profile_flags.runnable_procs
-	&& !(a & (ERTS_PSFLG_ACTIVE|ERTS_PSFLG_SUSPENDED))) {
-    	profile_runnable_proc(p, am_active);
+    if (erts_system_profile_flags.runnable_procs) {
+
+	if (((n & (ERTS_PSFLG_SUSPENDED
+		   | ERTS_PSFLG_ACTIVE)) == ERTS_PSFLG_ACTIVE)
+	    && (!(a & (ERTS_PSFLG_ACTIVE_SYS
+		       | ERTS_PSFLG_RUNNING
+		       | ERTS_PSFLG_RUNNING_SYS)
+		  && (!(a & ERTS_PSFLG_ACTIVE)
+		      || (a & ERTS_PSFLG_SUSPENDED))))) {
+	    /* We activated a prevously inactive process */
+	    profile_runnable_proc(p, am_active);
+	}
+
     }
 
-    if ((n & ERTS_PSFLG_IN_RUNQ) && !(a & ERTS_PSFLG_IN_RUNQ))
-	add2runq(p, n);
+    *statep = a;
+
+    return enqueue;
+}
+
+static ERTS_INLINE void
+schedule_process(Process *p, erts_aint32_t in_state)
+{
+    erts_aint32_t enq_prio  = -1;
+    erts_aint32_t state = in_state;
+    int enqueue = change_proc_schedule_state(p,
+					     0,
+					     ERTS_PSFLG_ACTIVE,
+					     &state,
+					     &enq_prio);
+    if (enqueue)
+	add2runq(enqueue > 0 ? p : make_proxy_proc(NULL, p, enq_prio),
+		 state,
+		 enq_prio);
 }
 
 void
 erts_schedule_process(Process *p, erts_aint32_t state)
 {
-    schedule_process(p, state, 0);
+    schedule_process(p, state);
+}
+
+static void
+schedule_process_sys_task(Process *p, erts_aint32_t state, Process *proxy)
+{
+    erts_aint32_t a = state, n, enq_prio = -1;
+    int enqueue; /* < 0 -> use proxy */
+
+    ASSERT(!(state & ERTS_PSFLG_PROXY));
+
+    while (1) {
+	erts_aint32_t e;
+	n = e = a;
+
+	if (a & ERTS_PSFLG_FREE)
+	    return; /* We don't want to schedule free processes... */
+
+	enqueue = 0;
+	n |= ERTS_PSFLG_ACTIVE_SYS;
+	if (!(a & (ERTS_PSFLG_RUNNING|ERTS_PSFLG_RUNNING_SYS)))
+	    enqueue = check_enqueue_in_prio_queue(&enq_prio, &n, a);
+	a = erts_smp_atomic32_cmpxchg_mb(&p->state, n, e);
+	if (a == e)
+	    break;
+	if (a == n && !enqueue)
+	    goto cleanup;
+    }
+
+    if (erts_system_profile_flags.runnable_procs) {
+
+	if (!(a & (ERTS_PSFLG_ACTIVE_SYS
+		   | ERTS_PSFLG_RUNNING
+		   | ERTS_PSFLG_RUNNING_SYS))
+	    && (!(a & ERTS_PSFLG_ACTIVE) || (a & ERTS_PSFLG_SUSPENDED))) {
+	    /* We activated a prevously inactive process */
+	    profile_runnable_proc(p, am_active);
+	}
+
+    }
+
+    if (enqueue) {
+	Process *sched_p;
+	if (enqueue > 0)
+	    sched_p = p;
+	else {
+	    sched_p = make_proxy_proc(proxy, p, enq_prio);
+	    proxy = NULL;
+	}
+	add2runq(sched_p, n, enq_prio);
+    }
+
+cleanup:
+    if (proxy)
+	free_proxy_proc(proxy);
 }
 
 static ERTS_INLINE int
 suspend_process(Process *c_p, Process *p)
 {
-    erts_aint32_t state = erts_smp_atomic32_read_acqb(&p->state);
+    erts_aint32_t state;
     int suspended = 0;
     ERTS_SMP_LC_ASSERT(ERTS_PROC_LOCK_STATUS & erts_proc_lc_my_proc_locks(p));
 
+    state = erts_smp_atomic32_read_acqb(&p->state);
+
     if ((state & ERTS_PSFLG_SUSPENDED))
 	suspended = -1;
     else {
 	if (c_p == p) {
 	    state = erts_smp_atomic32_read_bor_relb(&p->state,
 						    ERTS_PSFLG_SUSPENDED);
-	    state |= ERTS_PSFLG_SUSPENDED;
 	    ASSERT(state & ERTS_PSFLG_RUNNING);
-	    suspended = 1;
+	    suspended = (state & ERTS_PSFLG_SUSPENDED) ? -1: 1;
 	}
 	else {
 	    while (!(state & (ERTS_PSFLG_RUNNING|ERTS_PSFLG_EXITING))) {
-		erts_aint32_t e, n;
+		erts_aint32_t n, e;
+
 		n = e = state;
 		n |= ERTS_PSFLG_SUSPENDED;
 		state = erts_smp_atomic32_cmpxchg_relb(&p->state, n, e);
 		if (state == e) {
-		    state = n;
 		    suspended = 1;
 		    break;
 		}
+		if (state & ERTS_PSFLG_SUSPENDED) {
+		    suspended = -1;
+		    break;
+		}
 	    }
 	}
     }
 
-    if (state & ERTS_PSFLG_SUSPENDED) {
+    if (suspended) {
 
 	ASSERT(!(ERTS_PSFLG_RUNNING & state)
 	       || p == erts_get_current_process());
 
-	if (erts_system_profile_flags.runnable_procs
-	    && (p->rcount == 0)
-	    && (state & ERTS_PSFLG_ACTIVE)) {
-	    profile_runnable_proc(p, am_inactive);
+	if (suspended > 0 && erts_system_profile_flags.runnable_procs) {
+
+	    /* 'state' is before our change... */
+
+	    if ((state & (ERTS_PSFLG_ACTIVE
+			  | ERTS_PSFLG_ACTIVE_SYS
+			  | ERTS_PSFLG_RUNNING
+			  | ERTS_PSFLG_RUNNING_SYS
+			  | ERTS_PSFLG_SUSPENDED)) == ERTS_PSFLG_ACTIVE) {
+		/* We made process inactive */
+		profile_runnable_proc(p, am_inactive);
+	    }
+
 	}
 
 	p->rcount++;  /* count number of suspend */
     }
+
     return suspended;
 }
 
 static ERTS_INLINE void
 resume_process(Process *p)
 {
-    erts_aint32_t state;
+    erts_aint32_t state, enq_prio = -1;
+    int enqueue;
+
     ERTS_SMP_LC_ASSERT(ERTS_PROC_LOCK_STATUS & erts_proc_lc_my_proc_locks(p));
 
     ASSERT(p->rcount > 0);
@@ -5041,14 +5343,16 @@ resume_process(Process *p)
     if (--p->rcount > 0)  /* multiple suspend */
 	return;
 
-    state = erts_smp_atomic32_read_band_mb(&p->state, ~ERTS_PSFLG_SUSPENDED);
-    state &= ~ERTS_PSFLG_SUSPENDED;
-    if ((state & (ERTS_PSFLG_EXITING
-		  | ERTS_PSFLG_ACTIVE
-		  | ERTS_PSFLG_IN_RUNQ
-		  | ERTS_PSFLG_RUNNING)) == ERTS_PSFLG_ACTIVE) {
-	schedule_process(p, state, 1);
-    }
+    state = erts_smp_atomic32_read_nob(&p->state);
+    enqueue = change_proc_schedule_state(p,
+					 ERTS_PSFLG_SUSPENDED,
+					 0,
+					 &state,
+					 &enq_prio);
+    if (enqueue)
+	add2runq(enqueue > 0 ? p : make_proxy_proc(NULL, p, enq_prio),
+		 state,
+		 enq_prio);
 }
 
 int
@@ -6032,7 +6336,8 @@ pid2proc_not_running(Process *c_p, ErtsProcLocks c_p_locks,
 		goto done;
 	}
 	else {
-	    if (!(ERTS_PSFLG_RUNNING & erts_smp_atomic32_read_acqb(&rp->state)))
+	    if (!((ERTS_PSFLG_RUNNING|ERTS_PSFLG_RUNNING_SYS)
+		  & erts_smp_atomic32_read_acqb(&rp->state)))
 		goto done;
 
 	}
@@ -6695,7 +7000,7 @@ Eterm
 erts_get_process_priority(Process *p)
 {
     erts_aint32_t state = erts_smp_atomic32_read_nob(&p->state);
-    switch (state & ERTS_PSFLG_PRIO_MASK) {
+    switch (ERTS_PSFLGS_GET_USR_PRIO(state)) {
     case PRIORITY_MAX:		return am_max;
     case PRIORITY_HIGH:		return am_high;
     case PRIORITY_NORMAL:	return am_normal;
@@ -6718,18 +7023,68 @@ erts_set_process_priority(Process *p, Eterm value)
     }
 
     a = erts_smp_atomic32_read_nob(&p->state);
-    if (nprio == (a & ERTS_PSFLG_PRIO_MASK))
+    if (nprio == ERTS_PSFLGS_GET_USR_PRIO(a))
 	oprio = nprio;
     else {
-	erts_aint32_t e, n;
+	int slocked = 0;
+	erts_aint32_t e, n, aprio;
+
+	if (a & ERTS_PSFLG_ACTIVE_SYS) {
+	    erts_smp_proc_lock(p, ERTS_PROC_LOCK_STATUS);
+	    slocked = 1;
+	}
+
 	do {
-	    oprio = a & ERTS_PSFLG_PRIO_MASK;
+	    oprio = ERTS_PSFLGS_GET_USR_PRIO(a);
 	    n = e = a;
 
-	    ASSERT(!(a & ERTS_PSFLG_IN_RUNQ));
+	    if (!(a & (ERTS_PSFLG_ACTIVE_SYS|ERTS_PSFLG_DELAYED_SYS)))
+		aprio = nprio;
+	    else {
+		int max_qbit;
+
+		if (!slocked) {
+		    erts_smp_proc_lock(p, ERTS_PROC_LOCK_STATUS);
+		    slocked = 1;
+		}
+
+		max_qbit = 0;
+		if (a & ERTS_PSFLG_ACTIVE_SYS)
+		    max_qbit |= p->sys_task_qs->qmask;
+		if (a & ERTS_PSFLG_DELAYED_SYS) {
+		    ErtsProcSysTaskQs *qs;
+		    qs = ERTS_PROC_GET_DELAYED_GC_TASK_QS(p);
+		    ASSERT(qs);
+		    max_qbit |= qs->qmask;
+		}
+		max_qbit &= -max_qbit;
+		switch (max_qbit) {
+		case MAX_BIT:
+		    aprio = PRIORITY_MAX;
+		    break;
+		case HIGH_BIT:
+		    aprio = PRIORITY_HIGH;
+		    break;
+		case NORMAL_BIT:
+		    aprio = PRIORITY_NORMAL;
+		    break;
+		case LOW_BIT:
+		    aprio = PRIORITY_LOW;
+		    break;
+		default:
+		    ERTS_INTERNAL_ERROR("Invalid qmask");
+		    aprio = -1;
+		}
+
+		if (aprio > nprio) /* low value -> high prio */
+		    aprio = nprio;
+	    }
+
+	    n &= ~(ERTS_PSFLGS_USR_PRIO_MASK
+		   | ERTS_PSFLGS_ACT_PRIO_MASK);
+	    n |= ((nprio << ERTS_PSFLGS_USR_PRIO_OFFSET)
+		  | (aprio << ERTS_PSFLGS_ACT_PRIO_OFFSET));
 
-	    n &= ~ERTS_PSFLG_PRIO_MASK;
-	    n |= nprio;
 	    a = erts_smp_atomic32_cmpxchg_mb(&p->state, n, e);
 	} while (a != e);
     }
@@ -6763,6 +7118,7 @@ erts_set_process_priority(Process *p, Eterm value)
 
 Process *schedule(Process *p, int calls)
 {
+    Process *proxy_p = NULL;
     ErtsRunQueue *rq;
     erts_aint_t dt;
     ErtsSchedulerData *esdp;
@@ -6805,6 +7161,8 @@ Process *schedule(Process *p, int calls)
 	actual_reds = reds = 0;
 	erts_smp_runq_lock(rq);
     } else {
+    sched_out_proc:
+
 #ifdef ERTS_SMP
 	ERTS_SMP_CHK_HAVE_ONLY_MAIN_PROC_LOCK(p);
 	esdp = p->scheduler_data;
@@ -6858,10 +7216,11 @@ Process *schedule(Process *p, int calls)
 
 	esdp->reductions += reds;
 
-	schedule_out_process(rq, state, p); /* Returns with rq locked! */
+	schedule_out_process(rq, state, p, proxy_p); /* Returns with rq locked! */
+	proxy_p = NULL;
 
 	ERTS_PROC_REDUCTIONS_EXECUTED(rq,
-				      (int) (state & ERTS_PSFLG_PRIO_MASK),
+				      (int) ERTS_PSFLGS_GET_USR_PRIO(state),
 				      reds,
 				      actual_reds);
 
@@ -6870,18 +7229,19 @@ Process *schedule(Process *p, int calls)
 	p->scheduler_data = NULL;
 #endif
 
+	erts_smp_proc_unlock(p, ERTS_PROC_LOCK_MAIN|ERTS_PROC_LOCK_STATUS);
 
 	if (state & ERTS_PSFLG_FREE) {
 #ifdef ERTS_SMP
 	    ASSERT(esdp->free_process == p);
 	    esdp->free_process = NULL;
-#else	    
-	    erts_free_proc(p);
+#else
+	    state = erts_smp_atomic32_read_nob(&p->state);
+	    if (!(state & ERTS_PSFLG_IN_RUNQ))
+		erts_free_proc(p);
 #endif
 	}
 
-	erts_smp_proc_unlock(p, ERTS_PROC_LOCK_MAIN|ERTS_PROC_LOCK_STATUS);
-
 #ifdef ERTS_SMP
 	ASSERT(!esdp->free_process);
 #endif
@@ -7088,6 +7448,8 @@ Process *schedule(Process *p, int calls)
 	 * Find a new process to run.
 	 */
     pick_next_process: {
+	    erts_aint32_t psflg_band_mask;
+	    erts_aint32_t running_flag;
 	    int prio_q;
 	    int qmask;
 
@@ -7121,18 +7483,62 @@ Process *schedule(Process *p, int calls)
 
 	    ASSERT(p); /* Wrong qmask in rq->flags? */
 
+	    psflg_band_mask = ~(((erts_aint32_t) 1) << (ERTS_PSFLGS_GET_PRQ_PRIO(state)
+							+ ERTS_PSFLGS_IN_PRQ_MASK_OFFSET));
+
+	    if (!(state & ERTS_PSFLG_PROXY))
+		psflg_band_mask &= ~ERTS_PSFLG_IN_RUNQ;
+	    else {
+		proxy_p = p;
+		p = erts_proc_lookup_raw(proxy_p->common.id);
+		if (!p) {
+		    free_proxy_proc(proxy_p);
+		    proxy_p = NULL;
+		    goto pick_next_process;
+		}
+		state = erts_smp_atomic32_read_nob(&p->state);
+	    }
+
+
+	    if (state & ERTS_PSFLG_ACTIVE_SYS)
+		running_flag = ERTS_PSFLG_RUNNING_SYS;
+	    else
+		running_flag = ERTS_PSFLG_RUNNING;
+
 	    while (1) {
 		erts_aint32_t exp, new, tmp;
 		tmp = new = exp = state;
-		new &= ~ERTS_PSFLG_IN_RUNQ;
-		tmp = state & (ERTS_PSFLG_SUSPENDED|ERTS_PSFLG_PENDING_EXIT);
-		if (tmp != ERTS_PSFLG_SUSPENDED)
-		    new |= ERTS_PSFLG_RUNNING;
+		new &= psflg_band_mask;
+		if (!(state & (ERTS_PSFLG_RUNNING
+			       | ERTS_PSFLG_RUNNING_SYS))) {
+		    tmp = state & (ERTS_PSFLG_SUSPENDED
+				   | ERTS_PSFLG_PENDING_EXIT
+				   | ERTS_PSFLG_ACTIVE_SYS);
+		    if (tmp != ERTS_PSFLG_SUSPENDED)
+			new |= running_flag;
+		}
 		state = erts_smp_atomic32_cmpxchg_relb(&p->state, new, exp);
 		if (state == exp) {
-		    tmp = state & (ERTS_PSFLG_SUSPENDED|ERTS_PSFLG_PENDING_EXIT);
-		    if (tmp == ERTS_PSFLG_SUSPENDED)
+		    if ((state & (ERTS_PSFLG_RUNNING
+				  | ERTS_PSFLG_RUNNING_SYS
+				  | ERTS_PSFLG_FREE))
+			|| ((state & (ERTS_PSFLG_SUSPENDED
+				      | ERTS_PSFLG_PENDING_EXIT
+				      | ERTS_PSFLG_ACTIVE_SYS))
+			    == ERTS_PSFLG_SUSPENDED)) {
+			if (state & ERTS_PSFLG_FREE) {
+#ifdef ERTS_SMP
+			    erts_smp_proc_dec_refc(p);
+#else
+			    erts_free_proc(p);
+#endif
+			}
+			if (proxy_p) {
+			    free_proxy_proc(proxy_p);
+			    proxy_p = NULL;
+			}
 			goto pick_next_process;
+		    }
 		    state = new;
 		    break;
 		}
@@ -7162,7 +7568,7 @@ Process *schedule(Process *p, int calls)
 					  (UWord) esdp->no);
 	    int migrated = old && old != esdp->no;
 
-	    prio = (int) (state & ERTS_PSFLG_PRIO_MASK);
+	    prio = (int) ERTS_PSFLGS_GET_USR_PRIO(state);
 
 	    erts_smp_spin_lock(&erts_sched_stat.lock);
 	    erts_sched_stat.prio[prio].total_executed++;
@@ -7182,9 +7588,6 @@ Process *schedule(Process *p, int calls)
 	ASSERT(!p->scheduler_data);
 	p->scheduler_data = esdp;
 #endif
-	/* Never run a suspended process */
-	ASSERT(!(ERTS_PSFLG_SUSPENDED & erts_smp_atomic32_read_nob(&p->state)));
-
 	reds = context_reds;
 
 	if (IS_TRACED(p)) {
@@ -7210,21 +7613,780 @@ Process *schedule(Process *p, int calls)
 	    erts_check_my_tracer_proc(p);
 #endif
 
+	if (state & ERTS_PSFLG_RUNNING_SYS) {
+	    reds -= execute_sys_tasks(p, &state, reds);
+	    if (reds <= 0) {
+		p->fcalls = reds;
+		goto sched_out_proc;
+	    }
+
+	    ASSERT(state & ERTS_PSFLG_RUNNING_SYS);
+	    ASSERT(!(state & ERTS_PSFLG_RUNNING));
+
+	    while (1) {
+		erts_aint32_t n, e;
+
+		if (((state & (ERTS_PSFLG_SUSPENDED
+			       | ERTS_PSFLG_ACTIVE)) != ERTS_PSFLG_ACTIVE)
+		    && !(state & ERTS_PSFLG_EXITING))
+		    goto sched_out_proc;
+
+		n = e = state;
+		n &= ~ERTS_PSFLG_RUNNING_SYS;
+		n |= ERTS_PSFLG_RUNNING;
+
+		state = erts_smp_atomic32_cmpxchg_mb(&p->state, n, e);
+		if (state == e) {
+		    state = n;
+		    break;
+		}
+
+		ASSERT(state & ERTS_PSFLG_RUNNING_SYS);
+		ASSERT(!(state & ERTS_PSFLG_RUNNING));
+	    }
+	}
+
 	if (!(state & ERTS_PSFLG_EXITING)
 	    && ((FLAGS(p) & F_FORCE_GC)
 		|| (MSO(p).overhead > BIN_VHEAP_SZ(p)))) {
 	    reds -= erts_garbage_collect(p, 0, p->arg_reg, p->arity);
-	    if (reds < 0) {
-		reds = 1;
+	    if (reds <= 0) {
+		p->fcalls = reds;
+		goto sched_out_proc;
 	    }
 	}
+
+	if (proxy_p) {
+	    free_proxy_proc(proxy_p);
+	    proxy_p = NULL;
+	}
 	    
 	p->fcalls = reds;
 	ERTS_SMP_CHK_HAVE_ONLY_MAIN_PROC_LOCK(p);
+
+	/* Never run a suspended process */
+	ASSERT(!(ERTS_PSFLG_SUSPENDED & erts_smp_atomic32_read_nob(&p->state)));
+
 	return p;
     }
 }
 
+static int
+notify_sys_task_executed(Process *c_p, ErtsProcSysTask *st, Eterm st_result)
+{
+    Process *rp = erts_proc_lookup(st->requester);
+    if (rp) {
+	ErtsProcLocks rp_locks;
+	ErlOffHeap *ohp;
+	ErlHeapFragment* bp;
+	Eterm *hp, msg, req_id, result;
+	Uint st_result_sz, hsz;
+#ifdef DEBUG
+	Eterm *hp_start;
+#endif
+
+	rp_locks = (c_p == rp) ? ERTS_PROC_LOCK_MAIN : 0;
+
+	st_result_sz = is_immed(st_result) ? 0 : size_object(st_result);
+	hsz = st->req_id_sz + st_result_sz + 4 /* 3-tuple */;
+
+	hp = erts_alloc_message_heap(hsz,
+				     &bp,
+				     &ohp,
+				     rp,
+				     &rp_locks);
+
+#ifdef DEBUG
+	hp_start = hp;
+#endif
+
+	req_id = st->req_id_sz == 0 ? st->req_id : copy_struct(st->req_id,
+							       st->req_id_sz,
+							       &hp,
+							       ohp);
+
+	result = st_result_sz == 0 ? st_result :  copy_struct(st_result,
+							      st_result_sz,
+							      &hp,
+							      ohp);
+
+	ASSERT(is_immed(st->reply_tag));
+
+	msg = TUPLE3(hp, st->reply_tag, req_id, result);
+
+#ifdef DEBUG
+	hp += 4;
+	ASSERT(hp_start + hsz == hp);
+#endif
+
+	erts_queue_message(rp,
+			   &rp_locks,
+			   bp,
+			   msg,
+			   NIL
+#ifdef USE_VM_PROBES
+			   , NIL
+#endif
+	    );
+
+	if (c_p == rp)
+	    rp_locks &= ~ERTS_PROC_LOCK_MAIN;
+
+	if (rp_locks)
+	    erts_smp_proc_unlock(rp, rp_locks);
+    }
+
+    erts_cleanup_offheap(&st->off_heap);
+
+    erts_free(ERTS_ALC_T_PROC_SYS_TSK, st);
+
+    return rp ? 1 : 0;
+}
+
+static ERTS_INLINE ErtsProcSysTask *
+fetch_sys_task(Process *c_p, erts_aint32_t state, int *qmaskp, int *priop)
+{
+    ErtsProcSysTaskQs *unused_qs = NULL;
+    int qbit, qmask;
+    ErtsProcSysTask *st, **qp;
+
+    *priop = -1; /* Shut up annoying erroneous warning */
+
+    erts_smp_proc_lock(c_p, ERTS_PROC_LOCK_STATUS);
+
+    if (!c_p->sys_task_qs) {
+	qmask = 0;
+	st = NULL;
+	goto update_state;
+    }
+
+    qmask = c_p->sys_task_qs->qmask;
+
+    if ((state & (ERTS_PSFLG_ACTIVE
+		  | ERTS_PSFLG_EXITING
+		  | ERTS_PSFLG_SUSPENDED)) == ERTS_PSFLG_ACTIVE) {
+	/* No sys tasks if we got exclusively higher prio user work to do */
+	st = NULL;
+	switch (ERTS_PSFLGS_GET_USR_PRIO(state)) {
+	case PRIORITY_MAX:
+	    if (!(qmask & MAX_BIT))
+		goto done;
+	    break;
+	case PRIORITY_HIGH:
+	    if (!(qmask & (MAX_BIT|HIGH_BIT)))
+		goto done;
+	    break;
+	default:
+	    break;
+	}
+    }
+
+    qbit = qmask & -qmask;
+    switch (qbit) {
+    case MAX_BIT:
+	qp = &c_p->sys_task_qs->q[PRIORITY_MAX];
+	*priop = PRIORITY_MAX;
+	break;
+    case HIGH_BIT:
+	qp = &c_p->sys_task_qs->q[PRIORITY_HIGH];
+	*priop = PRIORITY_HIGH;
+	break;
+    case NORMAL_BIT:
+	if (!(qmask & PRIORITY_LOW)
+	    || ++c_p->sys_task_qs->ncount <= RESCHEDULE_LOW) {
+	    qp = &c_p->sys_task_qs->q[PRIORITY_NORMAL];
+	    *priop = PRIORITY_NORMAL;
+	    break;
+	}
+	c_p->sys_task_qs->ncount = 0;
+	/* Fall through */
+    case LOW_BIT:
+	qp = &c_p->sys_task_qs->q[PRIORITY_LOW];
+	*priop = PRIORITY_LOW;
+	break;
+    default:
+	ERTS_INTERNAL_ERROR("Invalid qmask");
+    }
+
+    st = *qp;
+    ASSERT(st);
+    if (st->next != st) {
+	*qp = st->next;
+	st->next->prev = st->prev;
+	st->prev->next = st->next;
+    }
+    else {
+	erts_aint32_t a, e, n, st_prio, qmask2;
+
+	*qp = NULL;
+	qmask &= ~qbit;
+	c_p->sys_task_qs->qmask = qmask;
+
+    update_state:
+
+	qmask2 = qmask;
+
+	if (state & ERTS_PSFLG_DELAYED_SYS) {
+	    ErtsProcSysTaskQs *qs = ERTS_PROC_GET_DELAYED_GC_TASK_QS(c_p);
+	    ASSERT(qs);
+	    qmask2 |= qs->qmask;
+	}
+
+	switch (qmask2 & -qmask2) {
+	case MAX_BIT:
+	    st_prio = PRIORITY_MAX;
+	    break;
+	case HIGH_BIT:
+	    st_prio = PRIORITY_HIGH;
+	    break;
+	case NORMAL_BIT:
+	    st_prio = PRIORITY_NORMAL;
+	    break;
+	case LOW_BIT:
+	case 0:
+	    st_prio = PRIORITY_LOW;
+	    break;
+	default:
+	    ERTS_INTERNAL_ERROR("Invalid qmask");
+	}
+
+	if (!qmask) {
+	    unused_qs = c_p->sys_task_qs;
+	    c_p->sys_task_qs = NULL;
+	}
+
+	a = state;
+	do {
+	    erts_aint32_t prio = ERTS_PSFLGS_GET_USR_PRIO(a);
+
+	    if (prio > st_prio)
+		prio = st_prio;
+
+	    n = e = a;
+
+	    n &= ~ERTS_PSFLGS_ACT_PRIO_MASK;
+	    n |= (prio << ERTS_PSFLGS_ACT_PRIO_OFFSET);
+
+	    if (!qmask)
+		n &= ~ERTS_PSFLG_ACTIVE_SYS;
+
+	    if (a == n)
+		break;
+	    a = erts_smp_atomic32_cmpxchg_nob(&c_p->state, n, e);
+	} while (a != e);
+    }
+
+done:
+
+    erts_smp_proc_unlock(c_p, ERTS_PROC_LOCK_STATUS);
+
+    if (unused_qs)
+	proc_sys_task_queues_free(unused_qs);
+
+    *qmaskp = qmask;
+
+    return st;
+}
+
+static void save_gc_task(Process *c_p, ErtsProcSysTask *st, int prio);
+
+static int
+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 qmask = 0;
+
+    ERTS_SMP_LC_ASSERT(erts_proc_lc_my_proc_locks(c_p) == ERTS_PROC_LOCK_MAIN);
+
+    do {
+	ErtsProcSysTask *st;
+	int st_prio;
+	Eterm st_res;
+
+	if (state & (ERTS_PSFLG_EXITING|ERTS_PSFLG_PENDING_EXIT)) {
+#ifdef ERTS_SMP
+	    if (state & ERTS_PSFLG_PENDING_EXIT)
+		erts_handle_pending_exit(c_p, ERTS_PROC_LOCK_MAIN);
+#endif
+	    ASSERT(ERTS_PROC_IS_EXITING(c_p));
+	    break;
+	}
+
+	st = fetch_sys_task(c_p, state, &qmask, &st_prio);
+	if (!st)
+	    break;
+
+	switch (st->type) {
+	case ERTS_PSTT_GC:
+	    if (c_p->flags & F_DISABLE_GC) {
+		save_gc_task(c_p, st, st_prio);
+		st = NULL;
+		reds++;
+	    }
+	    else {
+		if (!garbage_collected) {
+		    FLAGS(c_p) |= F_NEED_FULLSWEEP;
+		    reds += erts_garbage_collect(c_p,
+						 0,
+						 c_p->arg_reg,
+						 c_p->arity);
+		    garbage_collected = 1;
+		}
+		st_res = am_true;
+	    }
+	    break;
+	case ERTS_PSTT_CPC:
+	    st_res = erts_check_process_code(c_p,
+					     st->arg[0],
+					     st->arg[1] == am_true,
+					     &reds);
+	    if (is_non_value(st_res)) {
+		/* Needed gc, but gc was disabled */
+		save_gc_task(c_p, st, st_prio);
+		st = NULL;
+	    }
+	    break;
+	default:
+	    ERTS_INTERNAL_ERROR("Invalid process sys task type");
+	    st_res = am_false;
+	}
+
+	if (st)
+	    reds += notify_sys_task_executed(c_p, st, st_res);
+
+	state = erts_smp_atomic32_read_acqb(&c_p->state);
+    } while (qmask && reds < max_reds);
+
+    *statep = state;
+
+    return reds;
+}
+
+static int
+cleanup_sys_tasks(Process *c_p, erts_aint32_t in_state, int in_reds)
+{
+    erts_aint32_t state = in_state;
+    int max_reds = in_reds;
+    int reds = 0;
+    int qmask = 0;
+
+    ERTS_SMP_LC_ASSERT(erts_proc_lc_my_proc_locks(c_p) == ERTS_PROC_LOCK_MAIN);
+
+    do {
+	ErtsProcSysTask *st;
+	Eterm st_res;
+	int st_prio;
+
+	st = fetch_sys_task(c_p, state, &qmask, &st_prio);
+	if (!st)
+	    break;
+
+	switch (st->type) {
+	case ERTS_PSTT_GC:
+	    st_res = am_false;
+	    break;
+	case ERTS_PSTT_CPC:
+	    st_res = am_false;
+	    break;
+	default:
+	    ERTS_INTERNAL_ERROR("Invalid process sys task type");
+	    st_res = am_false;
+	    break;
+	}
+
+	reds += notify_sys_task_executed(c_p, st, st_res);
+
+	state = erts_smp_atomic32_read_acqb(&c_p->state);
+    } while (qmask && reds < max_reds);
+
+    return reds;
+}
+
+BIF_RETTYPE
+erts_internal_request_system_task_3(BIF_ALIST_3)
+{
+    Process *rp = erts_proc_lookup(BIF_ARG_1);
+    ErtsProcSysTaskQs *stqs, *free_stqs = NULL;
+    ErtsProcSysTask *st = NULL;
+    erts_aint32_t prio, rp_state;
+    int rp_locked;
+    Eterm noproc_res, req_type;
+
+    if (!rp && !is_internal_pid(BIF_ARG_1)) {
+	if (!is_external_pid(BIF_ARG_1))
+	    goto badarg;
+	if (external_pid_dist_entry(BIF_ARG_1) != erts_this_dist_entry)
+	    goto badarg;
+    }
+
+    switch (BIF_ARG_2) {
+    case am_max:	prio = PRIORITY_MAX;	break;
+    case am_high:	prio = PRIORITY_HIGH;	break;
+    case am_normal:	prio = PRIORITY_NORMAL;	break;
+    case am_low:	prio = PRIORITY_LOW;	break;
+    default: goto badarg;
+    }
+
+    if (is_not_tuple(BIF_ARG_3))
+	goto badarg;
+    else {
+	int i;
+	Eterm *tp = tuple_val(BIF_ARG_3);
+	Uint arity = arityval(*tp);
+	Eterm req_id;
+	Uint req_id_sz;
+	Eterm arg[ERTS_MAX_PROC_SYS_TASK_ARGS];
+	Uint arg_sz[ERTS_MAX_PROC_SYS_TASK_ARGS];
+	Uint tot_sz;
+	Eterm *hp;
+
+	if (arity < 2)
+	    goto badarg;
+	if (arity > 2 + ERTS_MAX_PROC_SYS_TASK_ARGS)
+	    goto badarg;
+	req_type = tp[1];
+	req_id = tp[2];
+	req_id_sz = is_immed(req_id) ? req_id : size_object(req_id);
+	tot_sz = req_id_sz;
+	for (i = 0; i < ERTS_MAX_PROC_SYS_TASK_ARGS; i++) {
+	    int tix = 3 + i;
+	    if (tix > arity) {
+		arg[i] = THE_NON_VALUE;
+		arg_sz[i] = 0;
+	    }
+	    else {
+		arg[i] = tp[tix];
+		if (is_immed(arg[i]))
+		    arg_sz[i] = 0;
+		else {
+		    arg_sz[i] = size_object(arg[i]);
+		    tot_sz += arg_sz[i];
+		}
+	    }
+	}
+	st = erts_alloc(ERTS_ALC_T_PROC_SYS_TSK,
+			ERTS_PROC_SYS_TASK_SIZE(tot_sz));
+	st->next = st->prev = st; /* Prep for empty prio queue */
+	ERTS_INIT_OFF_HEAP(&st->off_heap);
+	hp = &st->heap[0];
+
+	st->requester = BIF_P->common.id;
+	st->reply_tag = req_type;
+	st->req_id_sz = req_id_sz;
+	st->req_id = req_id_sz == 0 ? req_id : copy_struct(req_id,
+							   req_id_sz,
+							   &hp,
+							   &st->off_heap);
+
+	for (i = 0; i < ERTS_MAX_PROC_SYS_TASK_ARGS; i++)
+	    st->arg[i] = arg_sz[i] == 0 ? arg[i] : copy_struct(arg[i],
+							       arg_sz[i],
+							       &hp,
+							       &st->off_heap);
+	ASSERT(&st->heap[0] + tot_sz == hp);
+    }
+
+    switch (req_type) {
+
+    case am_garbage_collect:
+	st->type = ERTS_PSTT_GC;
+	noproc_res = am_false;
+	if (!rp)
+	    goto noproc;
+	break;
+
+    case am_check_process_code:
+	if (is_not_atom(st->arg[0]))
+	    goto badarg;
+	if (st->arg[1] != am_true && st->arg[1] != am_false)
+	    goto badarg;
+	noproc_res = am_false;
+	st->type = ERTS_PSTT_CPC;
+	if (!rp)
+	    goto noproc;
+	break;
+
+    default:
+	goto badarg;
+    }
+
+    rp_state = erts_smp_atomic32_read_nob(&rp->state);
+
+    rp_locked = 0;
+
+    free_stqs = NULL;
+    if (rp_state & ERTS_PSFLG_ACTIVE_SYS)
+	stqs = NULL;
+    else {
+    alloc_qs:
+	stqs = proc_sys_task_queues_alloc();
+	stqs->qmask = 1 << prio;
+	stqs->ncount = 0;
+	stqs->q[PRIORITY_MAX] = NULL;
+	stqs->q[PRIORITY_HIGH] = NULL;
+	stqs->q[PRIORITY_NORMAL] = NULL;
+	stqs->q[PRIORITY_LOW] = NULL;
+	stqs->q[prio] = st;
+    }
+
+    if (!rp_locked) {
+	rp_locked = 1;
+	erts_smp_proc_lock(rp, ERTS_PROC_LOCK_STATUS);
+
+	rp_state = erts_smp_atomic32_read_nob(&rp->state);
+	if (rp_state & ERTS_PSFLG_EXITING) {
+	    erts_smp_proc_unlock(rp, ERTS_PROC_LOCK_STATUS);
+	    rp = NULL;
+	    free_stqs = stqs;
+	    goto noproc;
+	}
+    }
+
+    if (!rp->sys_task_qs) {
+	if (stqs)
+	    rp->sys_task_qs = stqs;
+	else
+	    goto alloc_qs;
+    }
+    else {
+	if (stqs)
+	    free_stqs = stqs;
+	stqs = rp->sys_task_qs;
+	if (!stqs->q[prio]) {
+	    stqs->q[prio] = st;
+	    stqs->qmask |= 1 << prio;
+	}
+	else {
+	    st->next = stqs->q[prio];
+	    st->prev = stqs->q[prio]->prev;
+	    st->next->prev = st;
+	    st->prev->next = st;
+	    ASSERT(stqs->qmask & (1 << prio));
+	}
+    }
+
+    if (ERTS_PSFLGS_GET_ACT_PRIO(rp_state) > prio) {
+	erts_aint32_t n, a, e;
+	/* Need to elevate actual prio */
+
+	a = rp_state;
+	do {
+	    if (ERTS_PSFLGS_GET_ACT_PRIO(a) <= prio) {
+		n = a;
+		break;
+	    }
+	    n = e = a;
+	    n &= ~ERTS_PSFLGS_ACT_PRIO_MASK;
+	    n |= (prio << ERTS_PSFLGS_ACT_PRIO_OFFSET);
+	    a = erts_smp_atomic32_cmpxchg_nob(&rp->state, n, e);    
+	} while (a != e);
+	rp_state = n;
+    }
+
+    erts_smp_proc_unlock(rp, ERTS_PROC_LOCK_STATUS);
+
+    schedule_process_sys_task(rp, rp_state, NULL);
+
+    if (free_stqs)
+	proc_sys_task_queues_free(free_stqs);
+
+    BIF_RET(am_ok);
+
+noproc:
+
+    notify_sys_task_executed(BIF_P, st, noproc_res);
+    if (free_stqs)
+	proc_sys_task_queues_free(free_stqs);
+    BIF_RET(am_ok);
+
+badarg:
+
+    if (st) {
+	erts_cleanup_offheap(&st->off_heap);
+	erts_free(ERTS_ALC_T_PROC_SYS_TSK, st);
+    }
+    if (free_stqs)
+	proc_sys_task_queues_free(free_stqs);
+    BIF_ERROR(BIF_P, BADARG);
+}
+
+static void
+save_gc_task(Process *c_p, ErtsProcSysTask *st, int prio)
+{
+    erts_aint32_t state;
+    ErtsProcSysTaskQs *qs;
+
+    ERTS_SMP_LC_ASSERT(ERTS_PROC_LOCK_MAIN == erts_proc_lc_my_proc_locks(c_p));
+
+    qs = ERTS_PROC_GET_DELAYED_GC_TASK_QS(c_p);
+    if (!qs) {
+	st->next = st->prev = st;
+	qs = proc_sys_task_queues_alloc();
+	qs->qmask = 1 << prio;
+	qs->ncount = 0;
+	qs->q[PRIORITY_MAX] = NULL;
+	qs->q[PRIORITY_HIGH] = NULL;
+	qs->q[PRIORITY_NORMAL] = NULL;
+	qs->q[PRIORITY_LOW] = NULL;
+	qs->q[prio] = st;
+	(void) ERTS_PROC_SET_DELAYED_GC_TASK_QS(c_p, ERTS_PROC_LOCK_MAIN, qs);
+    }
+    else {
+	if (!qs->q[prio]) {
+	    st->next = st->prev = st;
+	    qs->q[prio] = st;
+	    qs->qmask |= 1 << prio;
+	}
+	else {
+	    st->next = qs->q[prio];
+	    st->prev = qs->q[prio]->prev;
+	    st->next->prev = st;
+	    st->prev->next = st;
+	    ASSERT(qs->qmask & (1 << prio));
+	}
+    }
+
+    state = erts_smp_atomic32_read_nob(&c_p->state);
+    ASSERT((ERTS_PSFLG_RUNNING|ERTS_PSFLG_RUNNING_SYS) & state);
+
+    while (!(state & ERTS_PSFLG_DELAYED_SYS)
+	   || prio < ERTS_PSFLGS_GET_ACT_PRIO(state)) {
+	erts_aint32_t n, e;
+
+	n = e = state;
+	n |= ERTS_PSFLG_DELAYED_SYS;
+	if (prio < ERTS_PSFLGS_GET_ACT_PRIO(state)) {
+	    n &= ~ERTS_PSFLGS_ACT_PRIO_MASK;
+	    n |= prio << ERTS_PSFLGS_ACT_PRIO_OFFSET;
+	}
+	state = erts_smp_atomic32_cmpxchg_relb(&c_p->state, n, e);
+	if (state == e)
+	    break;
+    }
+}
+
+int
+erts_set_gc_state(Process *c_p, int enable)
+{
+    ErtsProcSysTaskQs *dgc_tsk_qs;
+    ASSERT(c_p == erts_get_current_process());
+    ASSERT((ERTS_PSFLG_RUNNING|ERTS_PSFLG_RUNNING_SYS)
+	   & erts_smp_atomic32_read_nob(&c_p->state));
+    ERTS_SMP_LC_ASSERT(ERTS_PROC_LOCK_MAIN == erts_proc_lc_my_proc_locks(c_p));
+
+    if (!enable) {
+	c_p->flags |= F_DISABLE_GC;
+	return 0;
+    }
+
+    c_p->flags &= ~F_DISABLE_GC;
+
+    dgc_tsk_qs = ERTS_PROC_GET_DELAYED_GC_TASK_QS(c_p);
+    if (!dgc_tsk_qs)
+	return 0;
+
+    /* Move delayed gc tasks into sys tasks queues. */
+
+    erts_smp_proc_lock(c_p, ERTS_PROC_LOCK_STATUS);
+
+    if (!c_p->sys_task_qs) {
+	c_p->sys_task_qs = dgc_tsk_qs;
+	dgc_tsk_qs = NULL;
+    }
+    else {
+	ErtsProcSysTaskQs *stsk_qs;
+	int prio;
+
+	/*
+	 * We push delayed tasks to the front of the queue
+	 * since they have already made it to the front
+	 * once and then been delayed after that.
+	 */
+
+	stsk_qs = c_p->sys_task_qs;
+
+	while (dgc_tsk_qs->qmask) {
+	    int qbit = dgc_tsk_qs->qmask & -dgc_tsk_qs->qmask;
+	    dgc_tsk_qs->qmask &= ~qbit;
+	    switch (qbit) {
+	    case MAX_BIT:
+		prio = PRIORITY_MAX;
+		break;
+	    case HIGH_BIT:
+		prio = PRIORITY_HIGH;
+		break;
+	    case NORMAL_BIT:
+		prio = PRIORITY_NORMAL;
+		break;
+	    case LOW_BIT:
+		prio = PRIORITY_LOW;
+		break;
+	    default:
+		ERTS_INTERNAL_ERROR("Invalid qmask");
+		prio = -1;
+		break;
+	    }
+
+	    ASSERT(dgc_tsk_qs->q[prio]);
+
+	    if (!stsk_qs->q[prio]) {
+		stsk_qs->q[prio] = dgc_tsk_qs->q[prio];
+		stsk_qs->qmask |= 1 << prio;
+	    }
+	    else {
+		ErtsProcSysTask *first1, *last1, *first2, *last2;
+
+		ASSERT(stsk_qs->qmask & (1 << prio));
+		first1 = dgc_tsk_qs->q[prio];
+		last1 = first1->prev;
+		first2 = stsk_qs->q[prio];
+		last2 = first1->prev;
+
+		last1->next = first2;
+		first2->prev = last1;
+
+		first1->prev = last2;
+		last2->next = first1;
+
+		stsk_qs->q[prio] = first1;
+	    }
+
+	}
+    }
+
+#ifdef DEBUG
+    {
+	int qmask;
+	erts_aint32_t aprio, state =
+#endif
+
+	    erts_smp_atomic32_read_bset_nob(&c_p->state,
+					    (ERTS_PSFLG_DELAYED_SYS
+					     | ERTS_PSFLG_ACTIVE_SYS),
+					    ERTS_PSFLG_ACTIVE_SYS);
+
+#ifdef DEBUG
+	ASSERT(state & ERTS_PSFLG_DELAYED_SYS);
+	qmask = c_p->sys_task_qs->qmask;
+	aprio = ERTS_PSFLGS_GET_ACT_PRIO(state);
+	ASSERT(ERTS_PSFLGS_GET_USR_PRIO(state) >= aprio);
+	ASSERT((qmask & -qmask) >= (1 << aprio));
+    }
+#endif
+
+    erts_smp_proc_unlock(c_p, ERTS_PROC_LOCK_STATUS);
+
+    (void) ERTS_PROC_SET_DELAYED_GC_TASK_QS(c_p, ERTS_PROC_LOCK_MAIN, NULL);
+
+    if (dgc_tsk_qs)
+	proc_sys_task_queues_free(dgc_tsk_qs);
+
+    return 1;
+}
+
 void
 erts_sched_stat_modify(int what)
 {
@@ -7521,7 +8683,8 @@ erl_create_process(Process* parent, /* Parent of process (default group leader).
 	prio = (erts_aint32_t) so->priority;
     }
 
-    state |= (prio & ERTS_PSFLG_PRIO_MASK);
+    state |= (((prio & ERTS_PSFLGS_PRIO_MASK) << ERTS_PSFLGS_ACT_PRIO_OFFSET)
+	      | ((prio & ERTS_PSFLGS_PRIO_MASK) << ERTS_PSFLGS_USR_PRIO_OFFSET));
 
     if (!rq)
 	rq = erts_get_runq_proc(parent);
@@ -7599,6 +8762,8 @@ erl_create_process(Process* parent, /* Parent of process (default group leader).
     p->bin_old_vheap    = 0;
     p->bin_vheap_mature = 0;
 
+    p->sys_task_qs = NULL;
+
     /* No need to initialize p->fcalls. */
 
     p->current = p->initial+INITIAL_MOD;
@@ -7764,7 +8929,7 @@ erl_create_process(Process* parent, /* Parent of process (default group leader).
      * Schedule process for execution.
      */
 
-    schedule_process(p, state, 0);
+    schedule_process(p, state);
 
     VERBOSE(DEBUG_PROCESSES, ("Created a new process: %T\n",p->common.id));
 
@@ -7815,6 +8980,7 @@ void erts_init_empty_process(Process *p)
     p->bin_vheap_sz = BIN_VH_MIN_SIZE;
     p->bin_old_vheap_sz = BIN_VH_MIN_SIZE;
     p->bin_old_vheap = 0;
+    p->sys_task_qs = NULL;
     p->bin_vheap_mature = 0;
 #ifdef ERTS_SMP
     p->common.u.alive.ptimer = NULL;
@@ -8070,33 +9236,21 @@ delete_process(Process* p)
     p->fvalue = NIL;
 }
 
-static ERTS_INLINE erts_aint32_t
-set_proc_exiting_state(Process *p, erts_aint32_t state)
-{
-    erts_aint32_t a, n, e;
-    a = state;
-    while (1) {
-	n = e = a;
-	n &= ~(ERTS_PSFLG_SUSPENDED|ERTS_PSFLG_PENDING_EXIT);
-	n |= ERTS_PSFLG_EXITING|ERTS_PSFLG_ACTIVE;
-	if (!(a & (ERTS_PSFLG_IN_RUNQ|ERTS_PSFLG_RUNNING)))
-	    n |= ERTS_PSFLG_IN_RUNQ;
-	a = erts_smp_atomic32_cmpxchg_relb(&p->state, n, e);
-	if (a == e)
-	    break;
-    }
-    return a;
-}
-
 static ERTS_INLINE void
 set_proc_exiting(Process *p,
-		 erts_aint32_t state,
+		 erts_aint32_t in_state,
 		 Eterm reason,
 		 ErlHeapFragment *bp)
 {
+    erts_aint32_t state = in_state, enq_prio = -1;
+    int enqueue;
     ERTS_SMP_LC_ASSERT(erts_proc_lc_my_proc_locks(p) == ERTS_PROC_LOCKS_ALL);
 
-    state = set_proc_exiting_state(p, state);
+    enqueue = change_proc_schedule_state(p,
+					 ERTS_PSFLG_SUSPENDED|ERTS_PSFLG_PENDING_EXIT,
+					 ERTS_PSFLG_EXITING|ERTS_PSFLG_ACTIVE,
+					 &state,
+					 &enq_prio);
 
     p->fvalue = reason;
     if (bp)
@@ -8111,15 +9265,37 @@ set_proc_exiting(Process *p,
     cancel_timer(p);
     p->i = (BeamInstr *) beam_exit;
 
-    if (erts_system_profile_flags.runnable_procs
-	&& !(state & (ERTS_PSFLG_ACTIVE|ERTS_PSFLG_SUSPENDED))) {
-    	profile_runnable_proc(p, am_active);
-    }
-
-    if (!(state & (ERTS_PSFLG_IN_RUNQ|ERTS_PSFLG_RUNNING)))
-	add2runq(p, state);
+    if (enqueue)
+	add2runq(enqueue > 0 ? p : make_proxy_proc(NULL, p, enq_prio),
+		 state,
+		 enq_prio);
 }
 
+static ERTS_INLINE erts_aint32_t
+set_proc_self_exiting(Process *c_p)
+{
+#ifdef DEBUG
+    int enqueue;
+#endif
+    erts_aint32_t state, enq_prio = -1;
+
+    ERTS_SMP_LC_ASSERT(erts_proc_lc_my_proc_locks(c_p) == ERTS_PROC_LOCKS_ALL);
+
+    state = erts_smp_atomic32_read_nob(&c_p->state);
+    ASSERT(state & (ERTS_PSFLG_RUNNING|ERTS_PSFLG_RUNNING_SYS));
+
+#ifdef DEBUG
+    enqueue =
+#endif
+	change_proc_schedule_state(c_p,
+				   ERTS_PSFLG_SUSPENDED|ERTS_PSFLG_PENDING_EXIT,
+				   ERTS_PSFLG_EXITING|ERTS_PSFLG_ACTIVE,
+				   &state,
+				   &enq_prio);
+
+    ASSERT(!enqueue);
+    return state;
+}
 
 #ifdef ERTS_SMP
 
@@ -8167,7 +9343,7 @@ handle_pending_exiters(ErtsProcList *pnd_xtrs)
 	if (p) {
 	    if (erts_proclist_same(plp, p)) {
 		erts_aint32_t state = erts_smp_atomic32_read_acqb(&p->state);
-		if (!(state & ERTS_PSFLG_RUNNING)) {
+		if (!(state & (ERTS_PSFLG_RUNNING|ERTS_PSFLG_RUNNING_SYS))) {
 		    ASSERT(state & ERTS_PSFLG_PENDING_EXIT);
 		    erts_handle_pending_exit(p, ERTS_PROC_LOCKS_ALL);
 		}
@@ -8388,7 +9564,7 @@ send_exit_signal(Process *c_p,		/* current process if and only
 		}
 		set_proc_exiting(c_p, state, rsn, NULL);
 	    }
-	    else if (!(state & ERTS_PSFLG_RUNNING)) {
+	    else if (!(state & (ERTS_PSFLG_RUNNING|ERTS_PSFLG_RUNNING_SYS))) {
 		/* Process not running ... */
 		ErtsProcLocks need_locks = ~(*rp_locks) & ERTS_PROC_LOCKS_ALL;
 		if (need_locks
@@ -8765,9 +9941,6 @@ resume_suspend_monitor(ErtsSuspendMonitor *smon, void *vc_p)
 void 
 erts_do_exit_process(Process* p, Eterm reason)
 {
-#ifdef ERTS_SMP
-    erts_aint32_t state;
-#endif
     p->arity = 0;		/* No live registers */
     p->fvalue = reason;
 
@@ -8792,10 +9965,9 @@ erts_do_exit_process(Process* p, Eterm reason)
 #endif
 
 #ifndef ERTS_SMP
-    set_proc_exiting_state(p, erts_smp_atomic32_read_nob(&p->state));
+    set_proc_self_exiting(p);
 #else
-    state = set_proc_exiting_state(p, erts_smp_atomic32_read_nob(&p->state));
-    if (state & ERTS_PSFLG_PENDING_EXIT) {
+    if (ERTS_PSFLG_PENDING_EXIT & set_proc_self_exiting(p)) {
 	/* Process exited before pending exit was received... */
 	p->pending_exit.reason = THE_NON_VALUE;
 	if (p->pending_exit.bp) {
@@ -8849,6 +10021,7 @@ erts_continue_exit_process(Process *p)
     DistEntry *dep;
     struct saved_calls *scb;
     process_breakpoint_time_t *pbt;
+    erts_aint32_t state;
 
 #ifdef DEBUG
     int yield_allowed = 1;
@@ -8885,6 +10058,13 @@ erts_continue_exit_process(Process *p)
 	p->flags &= ~F_USING_DB;
     }
 
+    erts_set_gc_state(p, 1);
+    state = erts_smp_atomic32_read_acqb(&p->state);
+    if (state & ERTS_PSFLG_ACTIVE_SYS) {
+	if (cleanup_sys_tasks(p, state, CONTEXT_REDS) >= CONTEXT_REDS/2)
+	    goto yield;
+    }
+
     if (p->flags & F_USING_DDLL) {
 	erts_ddll_proc_dead(p, ERTS_PROC_LOCK_MAIN);
 	p->flags &= ~F_USING_DDLL;
@@ -8962,17 +10142,31 @@ erts_continue_exit_process(Process *p)
     {
 	/* Inactivate and notify free */
 	erts_aint32_t n, e, a = erts_smp_atomic32_read_nob(&p->state);
+#ifdef ERTS_SMP
+	int refc_inced = 0;
+#endif
 	while (1) {
 	    n = e = a;
 	    ASSERT(a & ERTS_PSFLG_EXITING);
 	    n |= ERTS_PSFLG_FREE;
 	    n &= ~ERTS_PSFLG_ACTIVE;
+#ifdef ERTS_SMP
+	    if ((n & ERTS_PSFLG_IN_RUNQ) && !refc_inced) {
+		erts_smp_proc_inc_refc(p);
+		refc_inced = 1;
+	    }
+#endif
 	    a = erts_smp_atomic32_cmpxchg_mb(&p->state, n, e);
 	    if (a == e)
 		break;
 	}
-    }
 
+#ifdef ERTS_SMP
+	if (refc_inced && !(n & ERTS_PSFLG_IN_RUNQ))
+	    erts_smp_proc_dec_refc(p);
+#endif
+    }
+    
     dep = ((p->flags & F_DISTRIBUTION)
 	   ? ERTS_PROC_SET_DIST_ENTRY(p, ERTS_PROC_LOCKS_ALL, NULL)
 	   : NULL);
@@ -9075,7 +10269,7 @@ timeout_proc(Process* p)
 
     state = erts_smp_atomic32_read_acqb(&p->state);
     if (!(state & ERTS_PSFLG_ACTIVE))
-	schedule_process(p, state, 0);
+	schedule_process(p, state);
 }
 
 
@@ -9153,7 +10347,9 @@ erts_program_counter_info(int to, void *to_arg, Process *p)
     print_function_from_pc(to, to_arg, p->cp);
     erts_print(to, to_arg, ")\n");
     state = erts_smp_atomic32_read_acqb(&p->state);
-    if (!(state & (ERTS_PSFLG_RUNNING|ERTS_PSFLG_GC))) {
+    if (!(state & (ERTS_PSFLG_RUNNING
+		   | ERTS_PSFLG_RUNNING_SYS
+		   | ERTS_PSFLG_GC))) {
         erts_print(to, to_arg, "arity = %d\n",p->arity);
 	if (!ERTS_IS_CRASH_DUMPING) {
 	    /*
diff --git a/erts/emulator/beam/erl_process.h b/erts/emulator/beam/erl_process.h
index 8d136f6e8b..043621125c 100644
--- a/erts/emulator/beam/erl_process.h
+++ b/erts/emulator/beam/erl_process.h
@@ -631,8 +631,9 @@ erts_smp_reset_max_len(ErtsRunQueue *rq, ErtsRunQueueInfo *rqi)
 #define ERTS_PSD_SCHED_ID			2
 #define ERTS_PSD_DIST_ENTRY			3
 #define ERTS_PSD_CALL_TIME_BP			4
+#define ERTS_PSD_DELAYED_GC_TASK_QS		5
 
-#define ERTS_PSD_SIZE				5
+#define ERTS_PSD_SIZE				6
 
 typedef struct {
     void *data[ERTS_PSD_SIZE];
@@ -656,6 +657,9 @@ typedef struct {
 #define ERTS_PSD_CALL_TIME_BP_GET_LOCKS ERTS_PROC_LOCK_MAIN
 #define ERTS_PSD_CALL_TIME_BP_SET_LOCKS ERTS_PROC_LOCK_MAIN
 
+#define ERTS_PSD_DELAYED_GC_TASK_QS_GET_LOCKS ERTS_PROC_LOCK_MAIN
+#define ERTS_PSD_DELAYED_GC_TASK_QS_SET_LOCKS ERTS_PROC_LOCK_MAIN
+
 typedef struct {
     ErtsProcLocks get_locks;
     ErtsProcLocks set_locks;
@@ -688,6 +692,9 @@ typedef struct {
     ErlHeapFragment *bp;
 } ErtsPendExit;
 
+typedef struct ErtsProcSysTask_ ErtsProcSysTask;
+typedef struct ErtsProcSysTaskQs_ ErtsProcSysTaskQs;
+
 #ifdef ERTS_SMP
 
 typedef struct ErtsPendingSuspend_ ErtsPendingSuspend;
@@ -855,6 +862,8 @@ struct process {
     Uint64 bin_old_vheap_sz;	/* Virtual old heap block size for binaries */
     Uint64 bin_old_vheap;	/* Virtual old heap size for binaries */
 
+    ErtsProcSysTaskQs *sys_task_qs;
+
     erts_smp_atomic32_t state;  /* Process state flags (see ERTS_PSFLG_*) */
 
 #ifdef ERTS_SMP
@@ -924,24 +933,67 @@ void erts_check_for_holes(Process* p);
 #  error "Need to increase ERTS_PSFLG_PRIO_SHIFT"
 #endif
 
-#define ERTS_PSFLG_PRIO_SHIFT 2
+#define ERTS_PSFLGS_PRIO_BITS 2
+#define ERTS_PSFLGS_PRIO_MASK \
+    ((((erts_aint32_t) 1) << ERTS_PSFLGS_PRIO_BITS) - 1)
 
-#define ERTS_PSFLG_BIT(N) \
-    (((erts_aint32_t) 1) << (ERTS_PSFLG_PRIO_SHIFT + (N)))
+#define ERTS_PSFLGS_ACT_PRIO_OFFSET (0*ERTS_PSFLGS_PRIO_BITS)
+#define ERTS_PSFLGS_USR_PRIO_OFFSET (1*ERTS_PSFLGS_PRIO_BITS)
+#define ERTS_PSFLGS_PRQ_PRIO_OFFSET (2*ERTS_PSFLGS_PRIO_BITS)
+#define ERTS_PSFLGS_ZERO_BIT_OFFSET (3*ERTS_PSFLGS_PRIO_BITS)
 
-#define ERTS_PSFLG_PRIO_MASK 		(ERTS_PSFLG_BIT(0) - 1)
+#define ERTS_PSFLGS_QMASK_BITS 4
+#define ERTS_PSFLGS_QMASK \
+    ((((erts_aint32_t) 1) << ERTS_PSFLGS_QMASK_BITS) - 1)
+#define ERTS_PSFLGS_IN_PRQ_MASK_OFFSET \
+    ERTS_PSFLGS_ZERO_BIT_OFFSET
 
-#define ERTS_PSFLG_FREE			ERTS_PSFLG_BIT(0)
-#define ERTS_PSFLG_EXITING		ERTS_PSFLG_BIT(1)
-#define ERTS_PSFLG_PENDING_EXIT		ERTS_PSFLG_BIT(2)
-#define ERTS_PSFLG_ACTIVE		ERTS_PSFLG_BIT(3)
-#define ERTS_PSFLG_IN_RUNQ		ERTS_PSFLG_BIT(4)
-#define ERTS_PSFLG_RUNNING		ERTS_PSFLG_BIT(5)
-#define ERTS_PSFLG_SUSPENDED		ERTS_PSFLG_BIT(6)
-#define ERTS_PSFLG_GC			ERTS_PSFLG_BIT(7)
-#define ERTS_PSFLG_BOUND		ERTS_PSFLG_BIT(8)
-#define ERTS_PSFLG_TRAP_EXIT		ERTS_PSFLG_BIT(9)
+#define ERTS_PSFLG_BIT(N) \
+    (((erts_aint32_t) 1) << (ERTS_PSFLGS_ZERO_BIT_OFFSET + (N)))
 
+/*
+ * ACT_PRIO -> Active prio, i.e., currently active prio. This
+ *             prio may be higher than user prio.
+ * USR_PRIO -> User prio. i.e., prio the user has set.
+ * PRQ_PRIO -> Prio queue prio, i.e., prio queue currently
+ *             enqueued in. 
+ */
+#define ERTS_PSFLGS_ACT_PRIO_MASK \
+    (ERTS_PSFLGS_PRIO_MASK << ERTS_PSFLGS_ACT_PRIO_OFFSET)
+#define ERTS_PSFLGS_USR_PRIO_MASK \
+    (ERTS_PSFLGS_PRIO_MASK << ERTS_PSFLGS_USR_PRIO_OFFSET)
+#define ERTS_PSFLGS_PRQ_PRIO_MASK \
+    (ERTS_PSFLGS_PRIO_MASK << ERTS_PSFLGS_PRQ_PRIO_OFFSET)
+#define ERTS_PSFLG_IN_PRQ_MAX 		ERTS_PSFLG_BIT(0)
+#define ERTS_PSFLG_IN_PRQ_HIGH		ERTS_PSFLG_BIT(1)
+#define ERTS_PSFLG_IN_PRQ_NORMAL	ERTS_PSFLG_BIT(2)
+#define ERTS_PSFLG_IN_PRQ_LOW 		ERTS_PSFLG_BIT(3)
+#define ERTS_PSFLG_FREE			ERTS_PSFLG_BIT(4)
+#define ERTS_PSFLG_EXITING		ERTS_PSFLG_BIT(5)
+#define ERTS_PSFLG_PENDING_EXIT		ERTS_PSFLG_BIT(6)
+#define ERTS_PSFLG_ACTIVE		ERTS_PSFLG_BIT(7)
+#define ERTS_PSFLG_IN_RUNQ		ERTS_PSFLG_BIT(8)
+#define ERTS_PSFLG_RUNNING		ERTS_PSFLG_BIT(9)
+#define ERTS_PSFLG_SUSPENDED		ERTS_PSFLG_BIT(10)
+#define ERTS_PSFLG_GC			ERTS_PSFLG_BIT(11)
+#define ERTS_PSFLG_BOUND		ERTS_PSFLG_BIT(12)
+#define ERTS_PSFLG_TRAP_EXIT		ERTS_PSFLG_BIT(13)
+#define ERTS_PSFLG_ACTIVE_SYS		ERTS_PSFLG_BIT(14)
+#define ERTS_PSFLG_RUNNING_SYS		ERTS_PSFLG_BIT(15)
+#define ERTS_PSFLG_PROXY		ERTS_PSFLG_BIT(16)
+#define ERTS_PSFLG_DELAYED_SYS		ERTS_PSFLG_BIT(17)
+
+#define ERTS_PSFLGS_IN_PRQ_MASK 	(ERTS_PSFLG_IN_PRQ_MAX		\
+					 | ERTS_PSFLG_IN_PRQ_HIGH	\
+					 | ERTS_PSFLG_IN_PRQ_NORMAL	\
+					 | ERTS_PSFLG_IN_PRQ_LOW)
+
+#define ERTS_PSFLGS_GET_ACT_PRIO(PSFLGS) \
+    (((PSFLGS) >> ERTS_PSFLGS_ACT_PRIO_OFFSET) & ERTS_PSFLGS_PRIO_MASK)
+#define ERTS_PSFLGS_GET_USR_PRIO(PSFLGS) \
+    (((PSFLGS) >> ERTS_PSFLGS_USR_PRIO_OFFSET) & ERTS_PSFLGS_PRIO_MASK)
+#define ERTS_PSFLGS_GET_PRQ_PRIO(PSFLGS) \
+    (((PSFLGS) >> ERTS_PSFLGS_USR_PRIO_OFFSET) & ERTS_PSFLGS_PRIO_MASK)
 
 /* The sequential tracing token is a tuple of size 5:
  *
@@ -1056,6 +1108,7 @@ extern struct erts_system_profile_flags_t erts_system_profile_flags;
 #define F_HAVE_BLCKD_MSCHED  (1 <<  8) /* Process has blocked multi-scheduling */
 #define F_P2PNR_RESCHED      (1 <<  9) /* Process has been rescheduled via erts_pid2proc_not_running() */
 #define F_FORCE_GC           (1 << 10) /* Force gc at process in-scheduling */
+#define F_DISABLE_GC         (1 << 11) /* Disable GC */
 
 /* process trace_flags */
 #define F_SENSITIVE          (1 << 0)
@@ -1146,6 +1199,7 @@ void erts_late_init_process(void);
 void erts_early_init_scheduling(int);
 void erts_init_scheduling(int, int);
 
+int erts_set_gc_state(Process *c_p, int enable);
 Eterm erts_sched_wall_time_request(Process *c_p, int set, int enable);
 Eterm erts_gc_info_request(Process *c_p);
 Uint64 erts_get_proc_interval(void);
@@ -1591,6 +1645,11 @@ erts_psd_set(Process *p, ErtsProcLocks plocks, int ix, void *data)
 #define ERTS_PROC_SET_CALL_TIME(P, L, PBT) \
   ((process_breakpoint_time_t *) erts_psd_set((P), (L), ERTS_PSD_CALL_TIME_BP, (void *) (PBT)))
 
+#define ERTS_PROC_GET_DELAYED_GC_TASK_QS(P) \
+    ((ErtsProcSysTaskQs *) erts_psd_get((P), ERTS_PSD_DELAYED_GC_TASK_QS))
+#define ERTS_PROC_SET_DELAYED_GC_TASK_QS(P, L, PBT) \
+    ((ErtsProcSysTaskQs *) erts_psd_set((P), (L), ERTS_PSD_DELAYED_GC_TASK_QS, (void *) (PBT)))
+
 
 ERTS_GLB_INLINE Eterm erts_proc_get_error_handler(Process *p);
 ERTS_GLB_INLINE Eterm erts_proc_set_error_handler(Process *p,
diff --git a/erts/emulator/beam/erl_zlib.c b/erts/emulator/beam/erl_zlib.c
index 47fd92988e..8e33144f96 100644
--- a/erts/emulator/beam/erl_zlib.c
+++ b/erts/emulator/beam/erl_zlib.c
@@ -87,6 +87,46 @@ int ZEXPORT erl_zlib_deflate_finish(z_stream *streamp)
     return deflateEnd(streamp);
 }
 
+int ZEXPORT erl_zlib_inflate_start(z_stream *streamp, const Bytef* source,
+                                   uLong sourceLen)
+{
+    streamp->next_in = (Bytef*)source;
+    streamp->avail_in = (uInt)sourceLen;
+    streamp->total_out = streamp->avail_out = 0;
+    streamp->next_out = NULL;
+    erl_zlib_alloc_init(streamp);
+    return inflateInit(streamp);
+}
+/*
+ * Inflate a chunk, The destination length is the limit.
+ * Returns Z_OK if more to process, Z_STREAM_END if we are done.
+ */
+int ZEXPORT erl_zlib_inflate_chunk(z_stream *streamp, Bytef* dest, uLongf* destLen)
+{
+    int err;
+    uLongf last_tot = streamp->total_out;
+
+    streamp->next_out = dest;
+    streamp->avail_out = (uInt)*destLen;
+
+    if ((uLong)streamp->avail_out != *destLen) return Z_BUF_ERROR;
+
+    err = inflate(streamp, Z_NO_FLUSH);
+    ASSERT(err != Z_STREAM_ERROR);
+    *destLen = streamp->total_out - last_tot;
+    return err;
+}
+
+/*
+ * When we are done, free up the inflate structure
+ * Retyurns Z_OK or Error
+ */
+int ZEXPORT erl_zlib_inflate_finish(z_stream *streamp)
+{
+    return inflateEnd(streamp);
+}
+
+
 int ZEXPORT erl_zlib_compress2 (Bytef* dest, uLongf* destLen,
 				const Bytef* source, uLong sourceLen,
 				int level)
diff --git a/erts/emulator/beam/erl_zlib.h b/erts/emulator/beam/erl_zlib.h
index 5ac849d21c..160166c66b 100644
--- a/erts/emulator/beam/erl_zlib.h
+++ b/erts/emulator/beam/erl_zlib.h
@@ -39,6 +39,12 @@ int ZEXPORT erl_zlib_deflate_start(z_stream *streamp, const Bytef* source,
 int ZEXPORT erl_zlib_deflate_chunk(z_stream *streamp, Bytef* dest, uLongf* destLen);
 int ZEXPORT erl_zlib_deflate_finish(z_stream *streamp);
 
+int ZEXPORT erl_zlib_inflate_start(z_stream *streamp, const Bytef* source,
+                                   uLong sourceLen);
+int ZEXPORT erl_zlib_inflate_chunk(z_stream *streamp, Bytef* dest, uLongf* destLen);
+int ZEXPORT erl_zlib_inflate_finish(z_stream *streamp);
+
+
 /* Use instead of compress
 */
 #define erl_zlib_compress(dest,destLen,source,sourceLen) \
diff --git a/erts/emulator/beam/external.c b/erts/emulator/beam/external.c
index 22b0a02937..2cb44a5b64 100644
--- a/erts/emulator/beam/external.c
+++ b/erts/emulator/beam/external.c
@@ -61,6 +61,9 @@
  */
 #    define ERTS_DEBUG_USE_DIST_SEP
 #  endif
+#  define IF_DEBUG(X) X
+#else
+#  define IF_DEBUG(X)
 #endif
 
 /* Does Sint fit in Sint32?
@@ -89,10 +92,11 @@ static int enc_term_int(Process *p,ErtsAtomCacheMap *acmp, Eterm obj, byte* ep,
 static Uint is_external_string(Eterm obj, int* p_is_string);
 static byte* enc_atom(ErtsAtomCacheMap *, Eterm, byte*, Uint32);
 static byte* enc_pid(ErtsAtomCacheMap *, Eterm, byte*, Uint32);
-static byte* dec_term(ErtsDistExternal *, Eterm**, byte*, ErlOffHeap*, Eterm*);
+struct B2TContext_t;
+static byte* dec_term(ErtsDistExternal *, Eterm**, byte*, ErlOffHeap*, Eterm*, struct B2TContext_t*);
 static byte* dec_atom(ErtsDistExternal *, byte*, Eterm*);
 static byte* dec_pid(ErtsDistExternal *, Eterm**, byte*, ErlOffHeap*, Eterm*);
-static Sint decoded_size(byte *ep, byte* endp, int internal_tags);
+static Sint decoded_size(byte *ep, byte* endp, int internal_tags, struct B2TContext_t*);
 static BIF_RETTYPE term_to_binary_trap_1(BIF_ALIST_1);
 
 static Eterm erts_term_to_binary_int(Process* p, Eterm Term, int level, Uint flags, 
@@ -102,11 +106,19 @@ static Uint encode_size_struct2(ErtsAtomCacheMap *, Eterm, unsigned);
 static int encode_size_struct_int(Process *p, ErtsAtomCacheMap *acmp, Eterm obj, 
 				  unsigned dflags, Sint *reds, Uint *res);
 
+static Export binary_to_term_trap_export;
+static BIF_RETTYPE binary_to_term_trap_1(BIF_ALIST_1);
+static Eterm binary_to_term_int(Process* p, Uint32 flags, Eterm bin, Binary* context_b);
+
 void erts_init_external(void) {
 #if 1 /* In R16 */
     erts_init_trap_export(&term_to_binary_trap_export,
 			  am_erlang, am_term_to_binary_trap, 1,
 			  &term_to_binary_trap_1);
+
+    erts_init_trap_export(&binary_to_term_trap_export,
+			  am_erlang, am_binary_to_term_trap, 1,
+			  &binary_to_term_trap_1);
 #else
     sys_memset((void *) &term_to_binary_trap_export, 0, sizeof(Export));
     term_to_binary_trap_export.address = &term_to_binary_trap_export.code[3];
@@ -877,7 +889,7 @@ erts_decode_dist_ext_size(ErtsDistExternal *edep)
 	    goto fail;
 	ep = edep->extp+1;
     }
-    res = decoded_size(ep, edep->ext_endp, 0);
+    res = decoded_size(ep, edep->ext_endp, 0, NULL);
     if (res >= 0)
 	return res;
  fail:
@@ -889,12 +901,12 @@ Sint erts_decode_ext_size(byte *ext, Uint size)
 {
     if (size == 0 || *ext != VERSION_MAGIC)
 	return -1;
-    return decoded_size(ext+1, ext+size, 0);
+    return decoded_size(ext+1, ext+size, 0, NULL);
 }
 
 Sint erts_decode_ext_size_ets(byte *ext, Uint size)
 {
-    Sint sz = decoded_size(ext, ext+size, 1);
+    Sint sz = decoded_size(ext, ext+size, 1, NULL);
     ASSERT(sz >= 0);
     return sz;
 }
@@ -927,7 +939,7 @@ erts_decode_dist_ext(Eterm** hpp,
 	    goto error;
 	ep++;
     }
-    ep = dec_term(edep, hpp, ep, off_heap, &obj);
+    ep = dec_term(edep, hpp, ep, off_heap, &obj, NULL);
     if (!ep)
 	goto error;
 
@@ -948,7 +960,7 @@ Eterm erts_decode_ext(Eterm **hpp, ErlOffHeap *off_heap, byte **ext)
     byte *ep = *ext;
     if (*ep++ != VERSION_MAGIC)
 	return THE_NON_VALUE;
-    ep = dec_term(NULL, hpp, ep, off_heap, &obj);
+    ep = dec_term(NULL, hpp, ep, off_heap, &obj, NULL);
     if (!ep) {
 #ifdef DEBUG
 	bin_write(ERTS_PRINT_STDERR,NULL,*ext,500);
@@ -962,7 +974,7 @@ Eterm erts_decode_ext(Eterm **hpp, ErlOffHeap *off_heap, byte **ext)
 Eterm erts_decode_ext_ets(Eterm **hpp, ErlOffHeap *off_heap, byte *ext)
 {
     Eterm obj;
-    ext = dec_term(NULL, hpp, ext, off_heap, &obj);
+    ext = dec_term(NULL, hpp, ext, off_heap, &obj, NULL);
     ASSERT(ext);
     return obj;
 }
@@ -1043,9 +1055,14 @@ static BIF_RETTYPE term_to_binary_trap_1(BIF_ALIST_1)
     Binary *bin = ((ProcBin *) binary_val(bt))->val;
     Eterm res = erts_term_to_binary_int(BIF_P, Term, 0, 0,bin);
     if (is_tuple(res)) {
+	ASSERT(BIF_P->flags & F_DISABLE_GC);
 	BIF_TRAP1(&term_to_binary_trap_export,BIF_P,res);
     } else {
-	BIF_RET(res);
+        if (erts_set_gc_state(BIF_P, 1)
+            || MSO(BIF_P).overhead > BIN_VHEAP_SZ(BIF_P))
+            ERTS_BIF_YIELD_RETURN(BIF_P, res);
+        else
+            BIF_RET(res);
     }
 }
     
@@ -1053,8 +1070,10 @@ BIF_RETTYPE term_to_binary_1(BIF_ALIST_1)
 {
     Eterm res = erts_term_to_binary_int(BIF_P, BIF_ARG_1, 0, TERM_TO_BINARY_DFLAGS, NULL);
     if (is_tuple(res)) {
+	erts_set_gc_state(BIF_P, 0);
 	BIF_TRAP1(&term_to_binary_trap_export,BIF_P,res);
     } else {
+	ASSERT(!(BIF_P->flags & F_DISABLE_GC));
 	BIF_RET(res);
     }
 }
@@ -1106,12 +1125,72 @@ BIF_RETTYPE term_to_binary_2(BIF_ALIST_2)
 
     res = erts_term_to_binary_int(p, Term, level, flags, bin);
     if (is_tuple(res)) {
+	erts_set_gc_state(p, 0);
 	BIF_TRAP1(&term_to_binary_trap_export,BIF_P,res);
     } else {
+	ASSERT(!(BIF_P->flags & F_DISABLE_GC));
 	BIF_RET(res);
     }
 }
 
+
+enum B2TState { /* order is somewhat significant */
+    B2TPrepare,
+    B2TUncompressChunk,
+    B2TSizeInit,
+    B2TSize,
+    B2TDecodeInit,
+    B2TDecode,
+    B2TDecodeList,
+    B2TDecodeTuple,
+    B2TDecodeString,
+    B2TDecodeBinary,
+
+    B2TDone,
+    B2TDecodeFail,
+    B2TBadArg
+};
+
+typedef struct {
+    int heap_size;
+    int terms;
+    byte* ep;
+    int atom_extra_skip;
+} B2TSizeContext;
+
+typedef struct {
+    byte*  ep;
+    Eterm  res;
+    Eterm* next;
+    Eterm* hp_start;
+    Eterm* hp;
+    Eterm* hp_end;
+    int remaining_n;
+    char* remaining_bytes;
+} B2TDecodeContext;
+
+typedef struct {
+    z_stream stream;
+    byte* dbytes;
+    Uint dleft;
+} B2TUncompressContext;
+
+typedef struct B2TContext_t {
+    Sint heap_size;
+    byte* aligned_alloc;
+    ErtsBinary2TermState b2ts;
+    Uint32 flags;
+    SWord reds;
+    Eterm trap_bin;
+    enum B2TState state;
+    union {
+	B2TSizeContext sc;
+	B2TDecodeContext dc;
+	B2TUncompressContext uc;
+    } u;
+} B2TContext;
+
+
 static uLongf binary2term_uncomp_size(byte* data, Sint size)
 {
     z_stream stream;
@@ -1141,48 +1220,62 @@ static uLongf binary2term_uncomp_size(byte* data, Sint size)
     return err == Z_STREAM_END ? uncomp_size : 0;
 }
 
-static ERTS_INLINE Sint
-binary2term_prepare(ErtsBinary2TermState *state, byte *data, Sint data_size)
+static ERTS_INLINE int
+binary2term_prepare(ErtsBinary2TermState *state, byte *data, Sint data_size,
+		    B2TContext* ctx)
 {
-    Sint res;
     byte *bytes = data;
     Sint size = data_size;
 
     state->exttmp = 0;
 
     if (size < 1 || *bytes != VERSION_MAGIC) {
-    error:
-	if (state->exttmp)
-	    erts_free(ERTS_ALC_T_TMP, state->extp);
-	state->extp = NULL;
-	state->exttmp = 0;
 	return -1;
     }
     bytes++;
     size--;
     if (size < 5 || *bytes != COMPRESSED) {
 	state->extp = bytes;
+        if (ctx)
+	    ctx->state = B2TSizeInit;
     }
     else  {
 	uLongf dest_len = (Uint32) get_int32(bytes+1);
 	bytes += 5;
 	size -= 5;	
 	if (dest_len > 32*1024*1024
-	    || (state->extp = erts_alloc_fnf(ERTS_ALC_T_TMP, dest_len)) == NULL) {
+	    || (state->extp = erts_alloc_fnf(ERTS_ALC_T_EXT_TERM_DATA, dest_len)) == NULL) {
+            /*
+             * Try avoid out-of-memory crash due to corrupted 'dest_len'
+             * by checking the actual length of the uncompressed data.
+             * The only way to do that is to uncompress it. Sad but true.
+             */
 	    if (dest_len != binary2term_uncomp_size(bytes, size)) {
-		goto error;
+                return -1;
 	    }
-	    state->extp = erts_alloc(ERTS_ALC_T_TMP, dest_len);
+	    state->extp = erts_alloc(ERTS_ALC_T_EXT_TERM_DATA, dest_len);
+            ctx->reds -= dest_len;
 	}
 	state->exttmp = 1;
-	if (erl_zlib_uncompress(state->extp, &dest_len, bytes, size) != Z_OK)
-	    goto error;
+        if (ctx) {
+	    if (erl_zlib_inflate_start(&ctx->u.uc.stream, bytes, size) != Z_OK)
+		return -1;
+
+	    ctx->u.uc.dbytes = state->extp;
+	    ctx->u.uc.dleft = dest_len;
+	    ctx->state = B2TUncompressChunk;
+        }
+	else {
+	    uLongf dlen = dest_len;
+	    if (erl_zlib_uncompress(state->extp, &dlen, bytes, size) != Z_OK
+		|| dlen != dest_len) {
+		return -1;
+	    }
+        }
 	size = (Sint) dest_len;
     }
-    res = decoded_size(state->extp, state->extp + size, 0);
-    if (res < 0)
-	goto error;
-    return res;
+    state->extsize = size;
+    return 0;
 }
 
 static ERTS_INLINE void
@@ -1190,7 +1283,7 @@ binary2term_abort(ErtsBinary2TermState *state)
 {
     if (state->exttmp) {
 	state->exttmp = 0;
-	erts_free(ERTS_ALC_T_TMP, state->extp);
+	erts_free(ERTS_ALC_T_EXT_TERM_DATA, state->extp);
     }
 }
 
@@ -1198,11 +1291,11 @@ static ERTS_INLINE Eterm
 binary2term_create(ErtsDistExternal *edep, ErtsBinary2TermState *state, Eterm **hpp, ErlOffHeap *ohp)
 {
     Eterm res;
-    if (!dec_term(edep, hpp, state->extp, ohp, &res))
+    if (!dec_term(edep, hpp, state->extp, ohp, &res, NULL))
 	res = THE_NON_VALUE;
     if (state->exttmp) {
 	state->exttmp = 0;
-	erts_free(ERTS_ALC_T_TMP, state->extp);
+	erts_free(ERTS_ALC_T_EXT_TERM_DATA, state->extp);
     }
     return res;
 }
@@ -1210,7 +1303,18 @@ binary2term_create(ErtsDistExternal *edep, ErtsBinary2TermState *state, Eterm **
 Sint
 erts_binary2term_prepare(ErtsBinary2TermState *state, byte *data, Sint data_size)
 {
-    return binary2term_prepare(state, data, data_size);
+    Sint res;
+
+    if (binary2term_prepare(state, data, data_size, NULL) < 0 ||
+        (res=decoded_size(state->extp, state->extp + state->extsize, 0, NULL)) < 0) {
+
+        if (state->exttmp)
+            erts_free(ERTS_ALC_T_EXT_TERM_DATA, state->extp);
+        state->extp = NULL;
+	state->exttmp = 0;
+	return -1;
+    }
+    return res;
 }
 
 void
@@ -1225,68 +1329,233 @@ erts_binary2term_create(ErtsBinary2TermState *state, Eterm **hpp, ErlOffHeap *oh
     return binary2term_create(NULL,state, hpp, ohp);
 }
 
-BIF_RETTYPE binary_to_term_1(BIF_ALIST_1)
+static void b2t_destroy_context(B2TContext* context)
 {
-    Sint heap_size;
-    Eterm res;
+    erts_free_aligned_binary_bytes_extra(context->aligned_alloc,
+                                         ERTS_ALC_T_EXT_TERM_DATA);
+    context->aligned_alloc = NULL;
+    binary2term_abort(&context->b2ts);
+    if (context->state == B2TUncompressChunk) {
+	erl_zlib_inflate_finish(&context->u.uc.stream);
+    }
+}
+
+static void b2t_context_destructor(Binary *context_bin)
+{
+    B2TContext* ctx = (B2TContext*) ERTS_MAGIC_BIN_DATA(context_bin);
+    ASSERT(ERTS_MAGIC_BIN_DESTRUCTOR(context_bin) == b2t_context_destructor);
+
+    b2t_destroy_context(ctx);
+}
+
+static BIF_RETTYPE binary_to_term_trap_1(BIF_ALIST_1)
+{
+    Binary *context_bin = ((ProcBin *) binary_val(BIF_ARG_1))->val;
+    ASSERT(ERTS_MAGIC_BIN_DESTRUCTOR(context_bin) == b2t_context_destructor);
+
+    return binary_to_term_int(BIF_P, 0, THE_NON_VALUE, context_bin);
+}
+
+
+#define B2T_BYTES_PER_REDUCTION 128
+#define B2T_MEMCPY_FACTOR 8
+
+/* Define for testing */
+/*#define EXTREME_B2T_TRAPPING 1*/
+
+#ifdef EXTREME_B2T_TRAPPING
+static unsigned b2t_rand(void)
+{
+    static unsigned prev = 17;
+    prev = (prev * 214013 + 2531011);
+    return prev;
+}
+#endif
+
+
+static B2TContext* b2t_export_context(Process* p, B2TContext* src)
+{
+    Binary* context_b = erts_create_magic_binary(sizeof(B2TContext),
+                                                 b2t_context_destructor);
+    B2TContext* ctx = ERTS_MAGIC_BIN_DATA(context_b);
     Eterm* hp;
-    Eterm* endp;
-    Sint size;
-    byte* bytes;
-    byte* temp_alloc = NULL;
-    ErtsBinary2TermState b2ts;
+    sys_memcpy(ctx, src, sizeof(B2TContext));
+    if (ctx->state >= B2TDecode && ctx->u.dc.next == &src->u.dc.res) {
+        ctx->u.dc.next = &ctx->u.dc.res;
+    }
+    hp = HAlloc(p, PROC_BIN_SIZE);
+    ctx->trap_bin = erts_mk_magic_binary_term(&hp, &MSO(p), context_b);
+    return ctx;
+}
 
-    if ((bytes = erts_get_aligned_binary_bytes(BIF_ARG_1, &temp_alloc)) == NULL) {
-    error:
-	erts_free_aligned_binary_bytes(temp_alloc);
-	BIF_ERROR(BIF_P, BADARG);
+static Eterm binary_to_term_int(Process* p, Uint32 flags, Eterm bin, Binary* context_b)
+{
+#ifdef EXTREME_B2T_TRAPPING
+    SWord initial_reds = 1 + b2t_rand() % 4;
+#else
+    SWord initial_reds = (Uint)(ERTS_BIF_REDS_LEFT(p) * B2T_BYTES_PER_REDUCTION);
+#endif
+    B2TContext c_buff;
+    B2TContext *ctx;
+    int is_first_call;
+
+    if (context_b == NULL) {
+	/* Setup enough to get started */
+        is_first_call = 1;
+        ctx = &c_buff;
+	ctx->state = B2TPrepare;
+        ctx->aligned_alloc = NULL;
+        ctx->flags = flags;
+        IF_DEBUG(ctx->trap_bin = THE_NON_VALUE;)
+    } else {
+        is_first_call = 0;
+	ctx = ERTS_MAGIC_BIN_DATA(context_b);
+        ASSERT(ctx->state != B2TPrepare);
     }
-    size = binary_size(BIF_ARG_1);
+    ctx->reds = initial_reds;
+
+    do {
+        switch (ctx->state) {
+        case B2TPrepare: {
+	    byte* bytes;
+            Uint bin_size;
+            bytes = erts_get_aligned_binary_bytes_extra(bin,
+                                                        &ctx->aligned_alloc,
+                                                        ERTS_ALC_T_EXT_TERM_DATA,
+                                                        0);
+            if (bytes == NULL) {
+                ctx->b2ts.exttmp = 0;
+                ctx->state = B2TBadArg;
+                break;
+            }
+            bin_size = binary_size(bin);
+            if (ctx->aligned_alloc) {
+                ctx->reds -= bin_size / 8;
+            }
+            if (binary2term_prepare(&ctx->b2ts, bytes, bin_size, ctx) < 0) {
+		ctx->state = B2TBadArg;
+	    }
+            break;
+        }
+	case B2TUncompressChunk: {
+            uLongf chunk = ctx->reds;
+            int zret;
+
+            if (chunk > ctx->u.uc.dleft)
+                chunk = ctx->u.uc.dleft;
+            zret = erl_zlib_inflate_chunk(&ctx->u.uc.stream,
+                                          ctx->u.uc.dbytes, &chunk);
+            ctx->u.uc.dbytes += chunk;
+            ctx->u.uc.dleft  -= chunk;
+            if (zret == Z_OK && ctx->u.uc.dleft > 0) {
+                ctx->reds = 0;
+            }
+            else if (erl_zlib_inflate_finish(&ctx->u.uc.stream) == Z_OK
+                     && zret == Z_STREAM_END
+                     && ctx->u.uc.dleft == 0) {
+                ctx->reds -= chunk;
+                ctx->state = B2TSizeInit;
+            }
+            else {
+                ctx->state = B2TBadArg;
+            }
+            break;
+        }
+	case B2TSizeInit:
+	    ctx->u.sc.ep = NULL;
+	    ctx->state = B2TSize;
+	    /*fall through*/
+        case B2TSize:
+            ctx->heap_size = decoded_size(ctx->b2ts.extp,
+					  ctx->b2ts.extp + ctx->b2ts.extsize,
+                                          0, ctx);
+            break;
+
+        case B2TDecodeInit:
+            if (ctx == &c_buff && ctx->b2ts.extsize > ctx->reds) {
+                /* dec_term will maybe trap, allocate space for magic bin
+                   before result term to make it easy to trim with HRelease.
+                 */
+                ctx = b2t_export_context(p, &c_buff);
+            }
+            ctx->u.dc.ep = ctx->b2ts.extp;
+            ctx->u.dc.res = (Eterm) (UWord) NULL;
+            ctx->u.dc.next = &ctx->u.dc.res;
+            ctx->u.dc.hp_start = HAlloc(p, ctx->heap_size);
+            ctx->u.dc.hp       = ctx->u.dc.hp_start;
+            ctx->u.dc.hp_end   = ctx->u.dc.hp_start + ctx->heap_size;
+            ctx->state = B2TDecode;
+            /*fall through*/
+	case B2TDecode:
+        case B2TDecodeList:
+        case B2TDecodeTuple:
+        case B2TDecodeString:
+        case B2TDecodeBinary: {
+	    ErtsDistExternal fakedep;
+            fakedep.flags = ctx->flags;
+            dec_term(&fakedep, NULL, NULL, &MSO(p), NULL, ctx);
+            break;
+	}
+        case B2TDecodeFail:
+            HRelease(p, ctx->u.dc.hp_end, ctx->u.dc.hp_start);
+            /*fall through*/
+        case B2TBadArg:
+            b2t_destroy_context(ctx);
+            if (!is_first_call) {
+                erts_set_gc_state(p, 1);
+            }
+            BUMP_REDS(p, (initial_reds - ctx->reds) / B2T_BYTES_PER_REDUCTION);
+            BIF_ERROR(p, BADARG & ~EXF_SAVETRACE);
 
-    heap_size = binary2term_prepare(&b2ts, bytes, size);
-    if (heap_size < 0)
-	goto error;
+        case B2TDone:
+            b2t_destroy_context(ctx);
 
-    hp = HAlloc(BIF_P, heap_size);
-    endp = hp + heap_size;
+            if (ctx->u.dc.hp > ctx->u.dc.hp_end) {
+                erl_exit(1, ":%s, line %d: heap overrun by %d words(s)\n",
+                         __FILE__, __LINE__, ctx->u.dc.hp - ctx->u.dc.hp_end);
+            }
+            HRelease(p, ctx->u.dc.hp_end, ctx->u.dc.hp);
 
-    res = binary2term_create(NULL, &b2ts, &hp, &MSO(BIF_P));
+            if (!is_first_call) {
+                erts_set_gc_state(p, 1);
+            }
+            BUMP_REDS(p, (initial_reds - ctx->reds) / B2T_BYTES_PER_REDUCTION);
+            return ctx->u.dc.res;
 
-    erts_free_aligned_binary_bytes(temp_alloc);
+        default:
+            ASSERT(!"Unknown state in binary_to_term");
+        }
+    }while (ctx->reds > 0 || ctx->state >= B2TDone);
 
-    if (hp > endp) {
-	erl_exit(1, ":%s, line %d: heap overrun by %d words(s)\n",
-		 __FILE__, __LINE__, hp-endp);
+    if (ctx == &c_buff) {
+        ASSERT(ctx->trap_bin == THE_NON_VALUE);
+        ctx = b2t_export_context(p, &c_buff);
     }
+    ASSERT(ctx->trap_bin != THE_NON_VALUE);
 
-    HRelease(BIF_P, endp, hp);
-
-    if (res == THE_NON_VALUE)
-	goto error;
+    if (is_first_call) {
+        erts_set_gc_state(p, 0);
+    }
+    BUMP_ALL_REDS(p);
+    BIF_TRAP1(&binary_to_term_trap_export, p, ctx->trap_bin);
+}
 
-    return res;
+BIF_RETTYPE erts_internal_binary_to_term_1(BIF_ALIST_1)
+{
+    return binary_to_term_int(BIF_P, 0, BIF_ARG_1, NULL);
 }
 
-BIF_RETTYPE binary_to_term_2(BIF_ALIST_2)
+BIF_RETTYPE erts_internal_binary_to_term_2(BIF_ALIST_2)
 {
-    Sint heap_size;
-    Eterm res;
     Eterm opts;
     Eterm opt;
-    Eterm* hp;
-    Eterm* endp;
-    Sint size;
-    byte* bytes;
-    byte* temp_alloc = NULL;
-    ErtsBinary2TermState b2ts;
-    ErtsDistExternal fakedep;
+    Uint32 flags = 0;
 
-    fakedep.flags = 0;
     opts = BIF_ARG_2;
     while (is_list(opts)) {
         opt = CAR(list_val(opts));
         if (opt == am_safe) {
-	    fakedep.flags |= ERTS_DIST_EXT_BTT_SAFE;
+            flags |= ERTS_DIST_EXT_BTT_SAFE;
         }
 	else {
             goto error;
@@ -1297,35 +1566,10 @@ BIF_RETTYPE binary_to_term_2(BIF_ALIST_2)
     if (is_not_nil(opts))
         goto error;
 
-    if ((bytes = erts_get_aligned_binary_bytes(BIF_ARG_1, &temp_alloc)) == NULL) {
-    error:
-	erts_free_aligned_binary_bytes(temp_alloc);
-	BIF_ERROR(BIF_P, BADARG);
-    }
-    size = binary_size(BIF_ARG_1);
-
-    heap_size = binary2term_prepare(&b2ts, bytes, size);
-    if (heap_size < 0)
-	goto error;
-
-    hp = HAlloc(BIF_P, heap_size);
-    endp = hp + heap_size;
-
-    res = binary2term_create(&fakedep, &b2ts, &hp, &MSO(BIF_P));
-
-    erts_free_aligned_binary_bytes(temp_alloc);
-
-    if (hp > endp) {
-	erl_exit(1, ":%s, line %d: heap overrun by %d words(s)\n",
-		 __FILE__, __LINE__, hp-endp);
-    }
-
-    HRelease(BIF_P, endp, hp);
-
-    if (res == THE_NON_VALUE)
-	goto error;
+    return binary_to_term_int(BIF_P, flags, BIF_ARG_1, NULL);
 
-    return res;
+error:
+    BIF_ERROR(BIF_P, BADARG);
 }
 
 Eterm
@@ -1473,12 +1717,10 @@ erts_term_to_binary(Process* p, Eterm Term, int level, Uint flags) {
 /* #define EXTREME_TTB_TRAPPING 1 */
 
 #ifndef EXTREME_TTB_TRAPPING
-#define TERM_TO_BINARY_LOOP_FACTOR 500
-#define TERM_TO_BINARY_SIZE_FACTOR 500000
-#define TERM_TO_BINARY_COMPRESS_CHUNK 500000
+#define TERM_TO_BINARY_LOOP_FACTOR 32
+#define TERM_TO_BINARY_COMPRESS_CHUNK (1 << 18)
 #else
 #define TERM_TO_BINARY_LOOP_FACTOR 1
-#define TERM_TO_BINARY_SIZE_FACTOR 10
 #define TERM_TO_BINARY_COMPRESS_CHUNK 10
 #endif
 
@@ -1514,7 +1756,7 @@ typedef struct {
     } s;
 } TTBContext;
 
-static void context_destructor(Binary *context_bin) 
+static void ttb_context_destructor(Binary *context_bin)
 {
     TTBContext *context = ERTS_MAGIC_BIN_DATA(context_bin);
     if (context->alive) {
@@ -1567,7 +1809,7 @@ static Eterm erts_term_to_binary_int(Process* p, Eterm Term, int level, Uint fla
     do {								\
 	if (context_b == NULL) {					\
 	    context_b = erts_create_magic_binary(sizeof(TTBContext),    \
-                                                 context_destructor);   \
+                                                 ttb_context_destructor);   \
 	    context =  ERTS_MAGIC_BIN_DATA(context_b);			\
 	    memcpy(context,&c_buff,sizeof(TTBContext));			\
 	}								\
@@ -1615,7 +1857,7 @@ static Eterm erts_term_to_binary_int(Process* p, Eterm Term, int level, Uint fla
 		    /* Finish in one go */
 		    res = erts_term_to_binary_simple(p, Term, size, 
 						     level, flags);
-		    BUMP_REDS(p, size / TERM_TO_BINARY_SIZE_FACTOR);
+		    BUMP_REDS(p, 1);
 		    return res;
 		}
 
@@ -2604,21 +2846,112 @@ undo_offheap_in_area(ErlOffHeap* off_heap, Eterm* start, Eterm* end)
 #endif /* DEBUG */
 }
 
+
 /* Decode term from external format into *objp.
 ** On failure return NULL and (R13B04) *hpp will be unchanged.
 */
 static byte*
-dec_term(ErtsDistExternal *edep, Eterm** hpp, byte* ep, ErlOffHeap* off_heap, Eterm* objp)
+dec_term(ErtsDistExternal *edep, Eterm** hpp, byte* ep, ErlOffHeap* off_heap,
+         Eterm* objp, B2TContext* ctx)
 {
-    Eterm* hp_saved = *hpp;
+    Eterm* hp_saved;
     int n;
     ErtsAtomEncoding char_enc;
-    register Eterm* hp = *hpp;	/* Please don't take the address of hp */
-    Eterm* next = objp;
+    register Eterm* hp;        /* Please don't take the address of hp */
+    Eterm* next;
+    SWord reds;
+
+    if (ctx) {
+        hp_saved = ctx->u.dc.hp_start;
+        reds     = ctx->reds;
+        next     = ctx->u.dc.next;
+        ep       = ctx->u.dc.ep;
+        hpp      = &ctx->u.dc.hp;
+
+        if (ctx->state != B2TDecode) {
+            int n_limit = reds;
+
+	    n = ctx->u.dc.remaining_n;
+            if (ctx->state == B2TDecodeBinary) {
+                n_limit *= B2T_MEMCPY_FACTOR;
+                ASSERT(n_limit >= reds);
+		reds -= n / B2T_MEMCPY_FACTOR;
+            }
+	    else
+		reds -= n;
 
-    *next = (Eterm) (UWord) NULL;
+            if (n > n_limit) {
+                ctx->u.dc.remaining_n -= n_limit;
+                n = n_limit;
+                reds = 0;
+            }
+            else {
+                ctx->u.dc.remaining_n = 0;
+            }
+
+            switch (ctx->state) {
+            case B2TDecodeList:
+                objp = next - 2;
+                while (n > 0) {
+                    objp[0] = (Eterm) COMPRESS_POINTER(next);
+                    objp[1] = make_list(next);
+                    next = objp;
+                    objp -= 2;
+                    n--;
+                }
+                break;
+
+            case B2TDecodeTuple:
+                objp = next - 1;
+                while (n-- > 0) {
+                    objp[0] = (Eterm) COMPRESS_POINTER(next);
+                    next = objp;
+                    objp--;
+                }
+                break;
+
+            case B2TDecodeString:
+                hp = *hpp;
+                hp[-1] = make_list(hp);  /* overwrite the premature NIL */
+                while (n-- > 0) {
+                    hp[0] = make_small(*ep++);
+                    hp[1] = make_list(hp+2);
+                    hp += 2;
+                }
+                hp[-1] = NIL;
+                *hpp = hp;
+                break;
+
+            case B2TDecodeBinary:
+                sys_memcpy(ctx->u.dc.remaining_bytes, ep, n);
+                ctx->u.dc.remaining_bytes += n;
+                ep += n;
+                break;
+
+            default:
+                ASSERT(!"Unknown state");
+            }
+            if (!ctx->u.dc.remaining_n) {
+                ctx->state = B2TDecode;
+            }
+            if (reds <= 0) {
+                ctx->u.dc.next = next;
+                ctx->u.dc.ep = ep;
+                ctx->reds = 0;
+                return NULL;
+            }
+        }
+    }
+    else {
+        hp_saved = *hpp;
+        reds = ERTS_SWORD_MAX;
+        next = objp;
+        *next = (Eterm) (UWord) NULL;
+    }
+    hp = *hpp;
 
     while (next != NULL) {
+
 	objp = next;
 	next = (Eterm *) EXPAND_POINTER(*objp);
 
@@ -2738,7 +3071,16 @@ dec_term_atom_common:
 	    *objp = make_tuple(hp);
 	    *hp++ = make_arityval(n);
 	    hp += n;
-	    objp = hp - 1;
+            objp = hp - 1;
+            if (ctx) {
+                if (reds < n) {
+                    ASSERT(reds > 0);
+                    ctx->state = B2TDecodeTuple;
+                    ctx->u.dc.remaining_n = n - reds;
+                    n = reds;
+                }
+		reds -= n;
+	    }
 	    while (n-- > 0) {
 		objp[0] = (Eterm) COMPRESS_POINTER(next);
 		next = objp;
@@ -2756,17 +3098,27 @@ dec_term_atom_common:
 		break;
 	    }
 	    *objp = make_list(hp);
-	    hp += 2*n;
+            hp += 2 * n;
 	    objp = hp - 2;
 	    objp[0] = (Eterm) COMPRESS_POINTER((objp+1));
 	    objp[1] = (Eterm) COMPRESS_POINTER(next);
 	    next = objp;
 	    objp -= 2;
-	    while (--n > 0) {
+            n--;
+	    if (ctx) {
+                if (reds < n) {
+		    ctx->state = B2TDecodeList;
+		    ctx->u.dc.remaining_n = n - reds;
+		    n = reds;
+		}
+		reds -= n;
+	    }
+            while (n > 0) {
 		objp[0] = (Eterm) COMPRESS_POINTER(next);
-		objp[1] = make_list(objp + 2);
+		objp[1] = make_list(next);
 		next = objp;
 		objp -= 2;
+                n--;
 	    }
 	    break;
 	case STRING_EXT:
@@ -2777,6 +3129,14 @@ dec_term_atom_common:
 		break;
 	    }
 	    *objp = make_list(hp);
+            if (ctx) {
+                if (reds < n) {
+                    ctx->state = B2TDecodeString;
+                    ctx->u.dc.remaining_n = n - reds;
+                    n = reds;
+		}
+                reds -= n;
+            }
 	    while (n-- > 0) {
 		hp[0] = make_small(*ep++);
 		hp[1] = make_list(hp+2);
@@ -2984,7 +3344,6 @@ dec_term_atom_common:
 		    dbin->flags = 0;
 		    dbin->orig_size = n;
 		    erts_refc_init(&dbin->refc, 1);
-		    sys_memcpy(dbin->orig_bytes, ep, n);
 		    pb = (ProcBin *) hp;
 		    hp += PROC_BIN_SIZE;
 		    pb->thing_word = HEADER_PROC_BIN;
@@ -2995,7 +3354,20 @@ dec_term_atom_common:
 		    pb->bytes = (byte*) dbin->orig_bytes;
 		    pb->flags = 0;
 		    *objp = make_binary(pb);
-		}
+                    if (ctx) {
+                        int n_limit = reds * B2T_MEMCPY_FACTOR;
+                        if (n > n_limit) {
+                            ctx->state = B2TDecodeBinary;
+                            ctx->u.dc.remaining_n = n - n_limit;
+                            ctx->u.dc.remaining_bytes = dbin->orig_bytes + n_limit;
+                            n = n_limit;
+                            reds = 0;
+                        }
+                        else
+                            reds -= n / B2T_MEMCPY_FACTOR;
+                    }
+                    sys_memcpy(dbin->orig_bytes, ep, n);
+                }
 		ep += n;
 		break;
 	    }
@@ -3018,13 +3390,14 @@ dec_term_atom_common:
 		    sys_memcpy(hb->data, ep, n);
 		    bin = make_binary(hb);
 		    hp += heap_bin_size(n);
+                    ep += n;
 		} else {
 		    Binary* dbin = erts_bin_nrml_alloc(n);
 		    ProcBin* pb;
+
 		    dbin->flags = 0;
 		    dbin->orig_size = n;
 		    erts_refc_init(&dbin->refc, 1);
-		    sys_memcpy(dbin->orig_bytes, ep, n);
 		    pb = (ProcBin *) hp;
 		    pb->thing_word = HEADER_PROC_BIN;
 		    pb->size = n;
@@ -3035,8 +3408,23 @@ dec_term_atom_common:
 		    pb->flags = 0;
 		    bin = make_binary(pb);
 		    hp += PROC_BIN_SIZE;
-		}
-		ep += n;
+                    if (ctx) {
+                        int n_limit = reds * B2T_MEMCPY_FACTOR;
+                        if (n > n_limit) {
+                            ctx->state = B2TDecodeBinary;
+                            ctx->u.dc.remaining_n = n - n_limit;
+                            ctx->u.dc.remaining_bytes = dbin->orig_bytes + n_limit;
+                            n = n_limit;
+                            reds = 0;
+                        }
+                        else
+                            reds -= n / B2T_MEMCPY_FACTOR;
+                    }
+                    sys_memcpy(dbin->orig_bytes, ep, n);
+                    ep += n;
+                    n = pb->size;
+                }
+
 		if (bitsize == 8 || n == 0) {
 		    *objp = bin;
 		} else {
@@ -3067,7 +3455,7 @@ dec_term_atom_common:
 		    goto error;
 		}
 		*hpp = hp;
-		ep = dec_term(edep, hpp, ep, off_heap, &temp);
+		ep = dec_term(edep, hpp, ep, off_heap, &temp, NULL);
 		hp = *hpp;
 		if (ep == NULL) {
 		    goto error;
@@ -3127,7 +3515,7 @@ dec_term_atom_common:
 		}
 		*hpp = hp;
 		/* Index */
-		if ((ep = dec_term(edep, hpp, ep, off_heap, &temp)) == NULL) {
+		if ((ep = dec_term(edep, hpp, ep, off_heap, &temp, NULL)) == NULL) {
 		    goto error;
 		}
 		if (!is_small(temp)) {
@@ -3136,7 +3524,7 @@ dec_term_atom_common:
 		old_index = unsigned_val(temp);
 
 		/* Uniq */
-		if ((ep = dec_term(edep, hpp, ep, off_heap, &temp)) == NULL) {
+		if ((ep = dec_term(edep, hpp, ep, off_heap, &temp, NULL)) == NULL) {
 		    goto error;
 		}
 		if (!is_small(temp)) {
@@ -3204,7 +3592,7 @@ dec_term_atom_common:
 		}
 
 		/* Index */
-		if ((ep = dec_term(edep, hpp, ep, off_heap, &temp)) == NULL) {
+		if ((ep = dec_term(edep, hpp, ep, off_heap, &temp, NULL)) == NULL) {
 		    goto error;
 		}
 		if (!is_small(temp)) {
@@ -3213,7 +3601,7 @@ dec_term_atom_common:
 		old_index = unsigned_val(temp);
 
 		/* Uniq */
-		if ((ep = dec_term(edep, hpp, ep, off_heap, &temp)) == NULL) {
+		if ((ep = dec_term(edep, hpp, ep, off_heap, &temp, NULL)) == NULL) {
 		    goto error;
 		}
 		if (!is_small(temp)) {
@@ -3313,8 +3701,31 @@ dec_term_atom_common:
 	    }
 	    undo_offheap_in_area(off_heap, hp_saved, hp);
 	    *hpp = hp_saved;
-	    return NULL;
+            if (ctx) {
+                ctx->state = B2TDecodeFail;
+		ctx->reds = reds;
+            }
+            return NULL;
 	}
+
+        if (--reds <= 0) {
+            if (ctx) {
+                if (next || ctx->state != B2TDecode) {
+                    ctx->u.dc.ep = ep;
+                    ctx->u.dc.next = next;
+                    ctx->u.dc.hp = hp;
+                    ctx->reds = 0;
+                    return NULL;
+                }
+            }
+            else {
+                reds = ERTS_SWORD_MAX;
+            }
+        }
+    }
+    if (ctx) {
+        ctx->state = B2TDone;
+	ctx->reds = reds;
     }
     *hpp = hp;
     return ep;
@@ -3602,18 +4013,37 @@ encode_size_struct_int(Process *p, ErtsAtomCacheMap *acmp, Eterm obj,
 }
 
 static Sint
-decoded_size(byte *ep, byte* endp, int internal_tags)
+decoded_size(byte *ep, byte* endp, int internal_tags, B2TContext* ctx)
 {
-    int heap_size = 0;
+    int heap_size;
     int terms;
-    int atom_extra_skip = 0;
+    int atom_extra_skip;
     Uint n;
+    SWord reds;
+
+    if (ctx) {
+        reds = ctx->reds;
+        if (ctx->u.sc.ep) {
+            heap_size = ctx->u.sc.heap_size;
+            terms = ctx->u.sc.terms;
+            ep = ctx->u.sc.ep;
+            atom_extra_skip = ctx->u.sc.atom_extra_skip;
+            goto init_done;
+        }
+    }
+    else
+        reds = 0; /* not used but compiler warns anyway */
+
+    heap_size = 0;
+    terms = 1;
+    atom_extra_skip = 0;
+init_done:
 
 #define SKIP(sz)				\
     do {					\
 	if ((sz) <= endp-ep) {			\
 	    ep += (sz);				\
-        } else { return -1; };			\
+        } else { goto error; };			\
     } while (0)
 
 #define SKIP2(sz1, sz2)				\
@@ -3621,25 +4051,24 @@ decoded_size(byte *ep, byte* endp, int internal_tags)
 	Uint sz = (sz1) + (sz2);		\
 	if (sz1 < sz && (sz) <= endp-ep) {	\
 	    ep += (sz);				\
-        } else { return -1; }			\
+        } else { goto error; }			\
     } while (0)
 
 #define CHKSIZE(sz)				\
     do {					\
-	 if ((sz) > endp-ep) { return -1; }	\
+	 if ((sz) > endp-ep) { goto error; }	\
     } while (0)
 
 #define ADDTERMS(n)				\
     do {					\
         int before = terms;		        \
 	terms += (n);                           \
-	if (terms < before) return -1;     	\
+	if (terms < before) goto error;     	\
     } while (0)
 
-
-    for (terms=1; terms > 0; terms--) {
-	int tag;
-
+    ASSERT(terms > 0);
+    do {
+        int tag;
 	CHKSIZE(1);
 	tag = ep++[0];
 	switch (tag) {
@@ -3660,7 +4089,7 @@ decoded_size(byte *ep, byte* endp, int internal_tags)
 	    CHKSIZE(4);
 	    n = get_int32(ep);
 	    if (n > BIG_ARITY_MAX*sizeof(ErtsDigit)) {
-		return -1;
+		goto error;
 	    }
 	    SKIP2(n,4+1);		/* skip, size,sign,digits */
 	    heap_size += 1+1+(n+sizeof(Eterm)-1)/sizeof(Eterm); /* XXX: 1 too much? */
@@ -3669,7 +4098,7 @@ decoded_size(byte *ep, byte* endp, int internal_tags)
 	    CHKSIZE(2);
 	    n = get_int16(ep);
 	    if (n > MAX_ATOM_CHARACTERS) {
-		return -1;
+		goto error;
 	    }
 	    SKIP(n+2+atom_extra_skip);
 	    atom_extra_skip = 0;
@@ -3679,7 +4108,7 @@ decoded_size(byte *ep, byte* endp, int internal_tags)
 	    n = get_int16(ep);
 	    ep += 2;
 	    if (n > MAX_ATOM_SZ_LIMIT) {
-		return -1;
+		goto error;
 	    }
 	    SKIP(n+atom_extra_skip);
 	    atom_extra_skip = 0;
@@ -3688,7 +4117,7 @@ decoded_size(byte *ep, byte* endp, int internal_tags)
 	    CHKSIZE(1);
 	    n = get_int8(ep);
 	    if (n > MAX_ATOM_CHARACTERS) {
-		return -1;
+		goto error;
 	    }
 	    SKIP(n+1+atom_extra_skip);
 	    atom_extra_skip = 0;
@@ -3698,7 +4127,7 @@ decoded_size(byte *ep, byte* endp, int internal_tags)
 	    n = get_int8(ep);
 	    ep++;
 	    if (n > MAX_ATOM_SZ_LIMIT) {
-		return -1;
+		goto error;
 	    }
 	    SKIP(n+atom_extra_skip);
 	    atom_extra_skip = 0;
@@ -3727,7 +4156,7 @@ decoded_size(byte *ep, byte* endp, int internal_tags)
 		id_words = get_int16(ep);
 		    
 		if (id_words > ERTS_MAX_REF_NUMBERS)
-		    return -1;
+		    goto error;
 
 		ep += 2;
 		atom_extra_skip = 1 + 4*id_words;
@@ -3829,7 +4258,7 @@ decoded_size(byte *ep, byte* endp, int internal_tags)
 		num_free = get_int32(ep);
 		ep += 4;
 		if (num_free > MAX_ARG) {
-		    return -1;
+		    goto error;
 		}
 		terms += 4 + num_free;
 		heap_size += ERL_FUN_SIZE + num_free;
@@ -3846,24 +4275,47 @@ decoded_size(byte *ep, byte* endp, int internal_tags)
 
 	case BINARY_INTERNAL_REF:
 	    if (!internal_tags) {
-		return -1;
+		goto error;
 	    }
 	    SKIP(sizeof(ProcBin));
 	    heap_size += PROC_BIN_SIZE;
 	    break;
 	case BIT_BINARY_INTERNAL_REF:
 	    if (!internal_tags) {
-		return -1;
+		goto error;
 	    }
 	    SKIP(2+sizeof(ProcBin));
 	    heap_size += PROC_BIN_SIZE + ERL_SUB_BIN_SIZE;
 	    break;
 	default:
-	    return -1;
+	    goto error;
 	}
-    }
+        terms--;
+
+        if (ctx && --reds <= 0 && terms > 0) {
+            ctx->u.sc.heap_size = heap_size;
+            ctx->u.sc.terms = terms;
+            ctx->u.sc.ep = ep;
+            ctx->u.sc.atom_extra_skip = atom_extra_skip;
+            ctx->reds = 0;
+            return 0;
+        }
+    }while (terms > 0);
+
     /* 'terms' may be non-zero if it has wrapped around */
-    return terms==0 ? heap_size : -1;
+    if (terms == 0) {
+        if (ctx) {
+            ctx->state = B2TDecodeInit;
+            ctx->reds = reds;
+        }
+        return heap_size;
+    }
+
+error:
+    if (ctx) {
+        ctx->state = B2TBadArg;
+    }
+    return -1;
 #undef SKIP
 #undef SKIP2
 #undef CHKSIZE
diff --git a/erts/emulator/beam/external.h b/erts/emulator/beam/external.h
index ff29e84972..83001b2c7e 100644
--- a/erts/emulator/beam/external.h
+++ b/erts/emulator/beam/external.h
@@ -146,6 +146,7 @@ typedef struct {
 typedef struct {
     byte *extp;
     int exttmp;
+    Uint extsize;
 } ErtsBinary2TermState;
 
 /* -------------------------------------------------------------------------- */
diff --git a/erts/emulator/beam/global.h b/erts/emulator/beam/global.h
index 063d16c0c7..bfab25d2bd 100755
--- a/erts/emulator/beam/global.h
+++ b/erts/emulator/beam/global.h
@@ -655,6 +655,10 @@ Eterm erl_send(Process *p, Eterm to, Eterm msg);
 
 Eterm erl_is_function(Process* p, Eterm arg1, Eterm arg2);
 
+/* beam_bif_load.c */
+Eterm erts_check_process_code(Process *c_p, Eterm module, int allow_gc, int *redsp);
+
+
 /* beam_load.c */
 typedef struct {
     BeamInstr* current;		/* Pointer to: Mod, Name, Arity */
@@ -1110,7 +1114,12 @@ erts_alloc_message_heap_state(Uint size,
 	if (statep)
 	    *statep = state;
 	if ((state & (ERTS_PSFLG_EXITING|ERTS_PSFLG_PENDING_EXIT))
+	    || (receiver->flags & F_DISABLE_GC)
 	    || HEAP_LIMIT(receiver) - HEAP_TOP(receiver) <= size) {
+	    /*
+	     * The heap is either potentially in an inconsistent
+	     * state, or not large enough.
+	     */
 #ifdef ERTS_SMP
 	    if (locked_main) {
 		*receiver_locks &= ~ERTS_PROC_LOCK_MAIN;
diff --git a/erts/emulator/beam/ops.tab b/erts/emulator/beam/ops.tab
index 1e5ae46bfa..c29f3f9b1b 100644
--- a/erts/emulator/beam/ops.tab
+++ b/erts/emulator/beam/ops.tab
@@ -763,17 +763,17 @@ allocate_init t I y
 #################################################################
 
 #
-# The BIFs erlang:check_process_code/2 must be called like a function,
+# The BIFs erts_internal:check_process_code/2 must be called like a function,
 # to ensure that c_p->i (program counter) is set correctly (an ordinary
 # BIF call doesn't set it).
 #
 
-call_ext u==2 Bif=u$bif:erlang:check_process_code/2 => i_call_ext Bif
-call_ext_last u==2 Bif=u$bif:erlang:check_process_code/2 D => i_call_ext_last Bif D
-call_ext_only u==2 Bif=u$bif:erlang:check_process_code/2 => i_call_ext_only Bif
+call_ext u==2 Bif=u$bif:erts_internal:check_process_code/2 => i_call_ext Bif
+call_ext_last u==2 Bif=u$bif:erts_internal:check_process_code/2 D => i_call_ext_last Bif D
+call_ext_only u==2 Bif=u$bif:erts_internal:check_process_code/2 => i_call_ext_only Bif
 
 #
-# The BIFs erlang:garbage_collect/0,1 must be called like functions,
+# The BIFs erlang:garbage_collect/0 must be called like a function,
 # to allow them to invoke the garbage collector. (The stack pointer must
 # be saved and p->arity must be zeroed, which is not done on ordinary BIF calls.)
 #
@@ -782,10 +782,6 @@ call_ext u==0 Bif=u$bif:erlang:garbage_collect/0 => i_call_ext Bif
 call_ext_last u==0 Bif=u$bif:erlang:garbage_collect/0 D => i_call_ext_last Bif D
 call_ext_only u==0 Bif=u$bif:erlang:garbage_collect/0 => i_call_ext_only Bif
 
-call_ext u==1 Bif=u$bif:erlang:garbage_collect/1 => i_call_ext Bif
-call_ext_last u==1 Bif=u$bif:erlang:garbage_collect/1 D => i_call_ext_last Bif D
-call_ext_only u==1 Bif=u$bif:erlang:garbage_collect/1 => i_call_ext_only Bif
-
 #
 # put/2 and erase/1 must be able to do garbage collection, so we must call
 # them like functions.
diff --git a/erts/emulator/beam/sys.h b/erts/emulator/beam/sys.h
index 31252ed78f..d0ccf8a2fc 100644
--- a/erts/emulator/beam/sys.h
+++ b/erts/emulator/beam/sys.h
@@ -279,18 +279,21 @@ typedef unsigned long UWord;
 typedef long          SWord;
 #define SWORD_CONSTANT(Const) Const##L
 #define UWORD_CONSTANT(Const) Const##UL
+#define ERTS_UWORD_MAX ULONG_MAX
 #define ERTS_SWORD_MAX LONG_MAX
 #elif SIZEOF_VOID_P == SIZEOF_INT
 typedef unsigned int UWord;
 typedef int          SWord;
 #define SWORD_CONSTANT(Const) Const
 #define UWORD_CONSTANT(Const) Const##U
+#define ERTS_UWORD_MAX UINT_MAX
 #define ERTS_SWORD_MAX INT_MAX
 #elif SIZEOF_VOID_P == SIZEOF_LONG_LONG
 typedef unsigned long long UWord;
 typedef long long          SWord;
 #define SWORD_CONSTANT(Const) Const##LL
 #define UWORD_CONSTANT(Const) Const##ULL
+#define ERTS_UWORD_MAX ULLONG_MAX
 #define ERTS_SWORD_MAX LLONG_MAX
 #else
 #error Found no appropriate type to use for 'Eterm', 'Uint' and 'Sint'
@@ -304,6 +307,7 @@ typedef unsigned long Uint;
 typedef long          Sint;
 #define SWORD_CONSTANT(Const) Const##L
 #define UWORD_CONSTANT(Const) Const##UL
+#define ERTS_UWORD_MAX ULONG_MAX
 #define ERTS_SWORD_MAX LONG_MAX
 #define ERTS_SIZEOF_ETERM SIZEOF_LONG
 #define ErtsStrToSint strtol
@@ -313,6 +317,7 @@ typedef unsigned int Uint;
 typedef int          Sint;
 #define SWORD_CONSTANT(Const) Const
 #define UWORD_CONSTANT(Const) Const##U
+#define ERTS_UWORD_MAX UINT_MAX
 #define ERTS_SWORD_MAX INT_MAX
 #define ERTS_SIZEOF_ETERM SIZEOF_INT
 #define ErtsStrToSint strtol
@@ -322,6 +327,7 @@ typedef unsigned long long Uint;
 typedef long long          Sint;
 #define SWORD_CONSTANT(Const) Const##LL
 #define UWORD_CONSTANT(Const) Const##ULL
+#define ERTS_UWORD_MAX ULLONG_MAX
 #define ERTS_SWORD_MAX LLONG_MAX
 #define ERTS_SIZEOF_ETERM SIZEOF_LONG_LONG
 #if defined(__WIN32__)
diff --git a/erts/emulator/beam/utils.c b/erts/emulator/beam/utils.c
index 0d75bbcc77..5a889f7fce 100644
--- a/erts/emulator/beam/utils.c
+++ b/erts/emulator/beam/utils.c
@@ -1675,7 +1675,7 @@ static int do_send_to_logger(Eterm tag, Eterm gleader, char *buf, int len)
 	p = erts_whereis_process(NULL, 0, am_error_logger, 0, 0);
 	if (p) {
 	    erts_aint32_t state = erts_smp_atomic32_read_acqb(&p->state);
-	    if (state & ERTS_PSFLG_RUNNING)
+	    if (state & (ERTS_PSFLG_RUNNING|ERTS_PSFLG_RUNNING_SYS))
 		p = NULL;
 	}
     }