Merge branch 'lukas/erts/enif_trace_functions/OTP-13442'

* lukas/erts/enif_trace_functions/OTP-13442:
  erts: Fix bug in enif_send
  erts: Fix windows nif port tests
  erts: Fix bug in enif_term_to_binary
  erts: Remove printout when dec_term fails in DEBUG
  erts: Polish erl_nif docs
  erts: Improve enif_binary_to_term
  erts: Add enif_port_command
  erts: Add enif_term_to_binary and enif_binary_to_term
  erts: Add enif_is_process/port_alive
  erts: Add enif_cpu/now_time and enif_make_unique_integer

13 files changed, 512 insertions, 50 deletions

diff --git a/erts/emulator/beam/beam_load.c b/erts/emulator/beam/beam_load.c
index f115df935f..d3d278fb81 100644
--- a/erts/emulator/beam/beam_load.c
+++ b/erts/emulator/beam/beam_load.c
@@ -1548,7 +1548,7 @@ read_literal_table(LoaderState* stp)
             erts_factory_heap_frag_init(&factory,
 					new_literal_fragment(heap_size));
 	    factory.alloc_type = ERTS_ALC_T_PREPARED_CODE;
-            val = erts_decode_ext(&factory, &p);
+            val = erts_decode_ext(&factory, &p, 0);
 
             if (is_non_value(val)) {
                 LoadError1(stp, "literal %d: bad external format", i);
@@ -1559,7 +1559,7 @@ read_literal_table(LoaderState* stp)
         }
         else {
             erts_factory_dummy_init(&factory);
-            val = erts_decode_ext(&factory, &p);
+            val = erts_decode_ext(&factory, &p, 0);
             if (is_non_value(val)) {
                 LoadError1(stp, "literal %d: bad external format", i);
             }
@@ -5719,7 +5719,7 @@ attributes_for_module(Process* p, /* Process whose heap to use. */
     if (ext != NULL) {
 	ErtsHeapFactory factory;
 	erts_factory_proc_prealloc_init(&factory, p, code_hdr->attr_size_on_heap);
-	result = erts_decode_ext(&factory, &ext);
+	result = erts_decode_ext(&factory, &ext, 0);
 	if (is_value(result)) {
 	    erts_factory_close(&factory);
 	}
@@ -5742,7 +5742,7 @@ compilation_info_for_module(Process* p, /* Process whose heap to use. */
     if (ext != NULL) {
 	ErtsHeapFactory factory;
 	erts_factory_proc_prealloc_init(&factory, p, code_hdr->compile_size_on_heap);
-	result = erts_decode_ext(&factory, &ext);
+	result = erts_decode_ext(&factory, &ext, 0);
 	if (is_value(result)) {
 	    erts_factory_close(&factory);
 	}
diff --git a/erts/emulator/beam/erl_bif_unique.c b/erts/emulator/beam/erl_bif_unique.c
index c4a39b8897..c9c8561f64 100644
--- a/erts/emulator/beam/erl_bif_unique.c
+++ b/erts/emulator/beam/erl_bif_unique.c
@@ -266,17 +266,19 @@ static ERTS_INLINE Eterm unique_integer_bif(Process *c_p, int positive)
 }
 
 Uint
-erts_raw_unique_integer_heap_size(Uint64 val[ERTS_UNIQUE_INT_RAW_VALUES])
+erts_raw_unique_integer_heap_size(Uint64 val[ERTS_UNIQUE_INT_RAW_VALUES],
+                                  int positive)
 {
     Uint sz;
-    bld_unique_integer_term(NULL, &sz, val[0], val[1], 0);
+    bld_unique_integer_term(NULL, &sz, val[0], val[1], positive);
     return sz;
 }
 
 Eterm
-erts_raw_make_unique_integer(Eterm **hpp, Uint64 val[ERTS_UNIQUE_INT_RAW_VALUES])
+erts_raw_make_unique_integer(Eterm **hpp, Uint64 val[ERTS_UNIQUE_INT_RAW_VALUES],
+    int positive)
 {
-    return bld_unique_integer_term(hpp, NULL, val[0], val[1], 0);
+    return bld_unique_integer_term(hpp, NULL, val[0], val[1], positive);
 }
 
 void
@@ -426,16 +428,16 @@ erts_raw_get_unique_monotonic_integer(void)
 }
 
 Uint
-erts_raw_unique_monotonic_integer_heap_size(Sint64 raw)
+erts_raw_unique_monotonic_integer_heap_size(Sint64 raw, int positive)
 {
-    return get_unique_monotonic_integer_heap_size(raw, 0);
+    return get_unique_monotonic_integer_heap_size(raw, positive);
 }
 
 Eterm
-erts_raw_make_unique_monotonic_integer_value(Eterm **hpp, Sint64 raw)
+erts_raw_make_unique_monotonic_integer_value(Eterm **hpp, Sint64 raw, int positive)
 {
-    Uint hsz = get_unique_monotonic_integer_heap_size(raw, 0);
-    Eterm res = make_unique_monotonic_integer_value(*hpp, hsz, raw, 0);
+    Uint hsz = get_unique_monotonic_integer_heap_size(raw, positive);
+    Eterm res = make_unique_monotonic_integer_value(*hpp, hsz, raw, positive);
     *hpp += hsz;
     return res;
 }
diff --git a/erts/emulator/beam/erl_bif_unique.h b/erts/emulator/beam/erl_bif_unique.h
index 37d5d91c39..94fd6163f2 100644
--- a/erts/emulator/beam/erl_bif_unique.h
+++ b/erts/emulator/beam/erl_bif_unique.h
@@ -41,8 +41,9 @@ void erts_make_ref_in_array(Uint32 ref[ERTS_MAX_REF_NUMBERS]);
  * not necessarily correspond to the end result.
  */
 Sint64 erts_raw_get_unique_monotonic_integer(void);
-Uint erts_raw_unique_monotonic_integer_heap_size(Sint64 raw);
-Eterm erts_raw_make_unique_monotonic_integer_value(Eterm **hpp, Sint64 raw);
+Uint erts_raw_unique_monotonic_integer_heap_size(Sint64 raw, int positive);
+Eterm erts_raw_make_unique_monotonic_integer_value(Eterm **hpp, Sint64 raw,
+                                                   int positive);
 
 Sint64 erts_get_min_unique_monotonic_integer(void);
 
@@ -53,8 +54,11 @@ Eterm erts_debug_get_unique_monotonic_integer_state(Process *c_p);
 #define ERTS_UNIQUE_INT_RAW_VALUES 2
 #define ERTS_MAX_UNIQUE_INT_HEAP_SIZE ERTS_UINT64_ARRAY_TO_BIG_MAX_HEAP_SZ(2)
 
-Uint erts_raw_unique_integer_heap_size(Uint64 val[ERTS_UNIQUE_INT_RAW_VALUES]);
-Eterm erts_raw_make_unique_integer(Eterm **hpp, Uint64 val[ERTS_UNIQUE_INT_RAW_VALUES]);
+Uint erts_raw_unique_integer_heap_size(Uint64 val[ERTS_UNIQUE_INT_RAW_VALUES],
+                                       int positive);
+Eterm erts_raw_make_unique_integer(Eterm **hpp,
+                                   Uint64 val[ERTS_UNIQUE_INT_RAW_VALUES],
+                                   int postive);
 void erts_raw_get_unique_integer(Uint64 val[ERTS_UNIQUE_INT_RAW_VALUES]);
 Sint64 erts_get_min_unique_integer(void);
 
diff --git a/erts/emulator/beam/erl_nif.c b/erts/emulator/beam/erl_nif.c
index 855d5deea1..d3030070b7 100644
--- a/erts/emulator/beam/erl_nif.c
+++ b/erts/emulator/beam/erl_nif.c
@@ -209,6 +209,7 @@ static void flush_env(ErlNifEnv* env)
 */
 static void cache_env(ErlNifEnv* env)
 {
+    env->heap_frag = MBUF(env->proc);
     if (env->heap_frag == NULL) {
 	ASSERT(env->hp_end == HEAP_LIMIT(env->proc));
 	ASSERT(env->hp <= HEAP_TOP(env->proc));
@@ -216,10 +217,6 @@ static void cache_env(ErlNifEnv* env)
 	env->hp = HEAP_TOP(env->proc);
     }
     else {
-	ASSERT(env->hp_end != HEAP_LIMIT(env->proc));
-	ASSERT(env->hp_end - env->hp <= env->heap_frag->alloc_size);       
-	env->heap_frag = MBUF(env->proc);
-	ASSERT(env->heap_frag != NULL);
 	env->hp = env->heap_frag->mem + env->heap_frag->used_size;
 	env->hp_end = env->heap_frag->mem + env->heap_frag->alloc_size;
     }
@@ -375,6 +372,29 @@ int enif_send(ErlNifEnv* env, const ErlNifPid* to_pid,
     return 1;
 }
 
+int
+enif_port_command(ErlNifEnv *env, const ErlNifPort* to_port,
+                  ErlNifEnv *msg_env, ERL_NIF_TERM msg)
+{
+
+    ErtsSchedulerData *esdp = erts_get_scheduler_data();
+    int scheduler = esdp ? esdp->no : 0;
+    Port *prt;
+
+    if (scheduler == 0 || !env)
+        return 0;
+
+    prt = erts_port_lookup(to_port->port_id,
+                           (erts_port_synchronous_ops
+                            ? ERTS_PORT_SFLGS_INVALID_DRIVER_LOOKUP
+                            : ERTS_PORT_SFLGS_INVALID_LOOKUP));
+
+    if (!prt)
+        return 0;
+
+    return erts_port_output_async(prt, env->proc->common.id, msg);
+}
+
 ERL_NIF_TERM enif_make_copy(ErlNifEnv* dst_env, ERL_NIF_TERM src_term)
 {
     Uint sz;
@@ -404,12 +424,28 @@ static int is_offheap(const ErlOffHeap* oh)
 
 ErlNifPid* enif_self(ErlNifEnv* caller_env, ErlNifPid* pid)
 {
+    if (caller_env->proc->common.id == ERTS_INVALID_PID)
+        return NULL;
     pid->pid = caller_env->proc->common.id;
     return pid;
 }
+
 int enif_get_local_pid(ErlNifEnv* env, ERL_NIF_TERM term, ErlNifPid* pid)
 {
-    return is_internal_pid(term) ? (pid->pid=term, 1) : 0;
+    if (is_internal_pid(term)) {
+        pid->pid=term;
+        return 1;
+    }
+    return 0;
+}
+
+int enif_get_local_port(ErlNifEnv* env, ERL_NIF_TERM term, ErlNifPort* port)
+{
+    if (is_internal_port(term)) {
+        port->port_id=term;
+        return 1;
+    }
+    return 0;
 }
 
 int enif_is_atom(ErlNifEnv* env, ERL_NIF_TERM term)
@@ -622,6 +658,68 @@ unsigned char* enif_make_new_binary(ErlNifEnv* env, size_t size,
     return binary_bytes(*termp);
 }
 
+int enif_term_to_binary(ErlNifEnv *dst_env, ERL_NIF_TERM term,
+                        ErlNifBinary *bin)
+{
+    Sint size;
+    byte *bp;
+    Binary* refbin;
+
+    size = erts_encode_ext_size(term);
+    if (!enif_alloc_binary(size, bin))
+        return 0;
+
+    refbin = bin->ref_bin;
+
+    bp = bin->data;
+
+    erts_encode_ext(term, &bp);
+
+    bin->size = bp - bin->data;
+    refbin->orig_size = bin->size;
+
+    ASSERT(bin->data + bin->size == bp);
+
+    return 1;
+}
+
+size_t enif_binary_to_term(ErlNifEnv *dst_env,
+                           const unsigned char* data,
+                           size_t data_sz,
+                           ERL_NIF_TERM *term,
+                           ErlNifBinaryToTerm opts)
+{
+    Sint size;
+    ErtsHeapFactory factory;
+    byte *bp = (byte*) data;
+
+    ERTS_CT_ASSERT(ERL_NIF_BIN2TERM_SAFE == ERTS_DIST_EXT_BTT_SAFE);
+
+    if (opts & ~ERL_NIF_BIN2TERM_SAFE) {
+        return 0;
+    }
+    if ((size = erts_decode_ext_size(bp, data_sz)) < 0)
+        return 0;
+
+    if (size > 0) {
+        flush_env(dst_env);
+        erts_factory_proc_prealloc_init(&factory, dst_env->proc, size);
+    } else {
+        erts_factory_dummy_init(&factory);
+    }
+
+    *term = erts_decode_ext(&factory, &bp, (Uint32)opts);
+
+    if (is_non_value(*term)) {
+        return 0;
+    }
+    erts_factory_close(&factory);
+    cache_env(dst_env);
+
+    ASSERT(bp > data);
+    return bp - data;
+}
+
 int enif_is_identical(Eterm lhs, Eterm rhs)
 {
     return EQ(lhs,rhs);
@@ -1158,6 +1256,103 @@ int enif_make_reverse_list(ErlNifEnv* env, ERL_NIF_TERM term, ERL_NIF_TERM *list
     return 1;
 }
 
+int enif_is_process_alive(ErlNifEnv* env, ErlNifPid *proc)
+{
+    ErtsProcLocks rp_locks = 0; /* We don't need any locks,
+                                   just to check if it is alive */
+    Eterm target = proc->pid;
+    Process* rp;
+    Process* c_p;
+    int scheduler = erts_get_scheduler_id() != 0;
+
+    if (env != NULL) {
+	c_p = env->proc;
+	if (target == c_p->common.id) {
+            /* We are alive! */
+	    return 1;
+	}
+    }
+    else {
+#ifdef ERTS_SMP
+	c_p = NULL;
+#else
+	erts_exit(ERTS_ABORT_EXIT,"enif_is_process_alive: "
+                  "env==NULL on non-SMP VM");
+#endif
+    }
+
+    rp = (scheduler
+	  ? erts_proc_lookup(target)
+	  : erts_pid2proc_opt(c_p, ERTS_PROC_LOCK_MAIN,
+			      target, rp_locks, ERTS_P2P_FLG_INC_REFC));
+    if (rp == NULL) {
+        ASSERT(env == NULL || target != c_p->common.id);
+	return 0;
+    } else {
+        if (!scheduler)
+            erts_proc_dec_refc(rp);
+        return 1;
+    }
+}
+
+int enif_is_port_alive(ErlNifEnv *env, ErlNifPort *port)
+{
+    /* only allowed if called from scheduler */
+    if (erts_get_scheduler_id() == 0)
+        erts_exit(ERTS_ABORT_EXIT,"enif_is_port_alive: called from non-scheduler");
+
+    return erts_port_lookup(
+        port->port_id,
+        (erts_port_synchronous_ops
+         ? ERTS_PORT_SFLGS_INVALID_DRIVER_LOOKUP
+         : ERTS_PORT_SFLGS_INVALID_LOOKUP)) != NULL;
+}
+
+ERL_NIF_TERM
+enif_now_time(ErlNifEnv *env)
+{
+    Uint mega, sec, micro;
+    Eterm *hp;
+    get_now(&mega, &sec, &micro);
+    hp = alloc_heap(env, 4);
+    return TUPLE3(hp, make_small(mega), make_small(sec), make_small(micro));
+}
+
+ERL_NIF_TERM
+enif_cpu_time(ErlNifEnv *env)
+{
+#ifdef HAVE_ERTS_NOW_CPU
+    Uint mega, sec, micro;
+    Eterm *hp;
+    erts_get_now_cpu(&mega, &sec, &micro);
+    hp = alloc_heap(env, 4);
+    return TUPLE3(hp, make_small(mega), make_small(sec), make_small(micro));
+#else
+    return enif_make_badarg(env);
+#endif
+}
+
+ERL_NIF_TERM
+enif_make_unique_integer(ErlNifEnv *env, ErlNifUniqueInteger properties)
+{
+    int monotonic = properties & ERL_NIF_UNIQUE_MONOTONIC;
+    int positive = properties & ERL_NIF_UNIQUE_POSITIVE;
+    Eterm *hp;
+    Uint hsz;
+
+    if (monotonic) {
+        Sint64 raw_unique = erts_raw_get_unique_monotonic_integer();
+        hsz = erts_raw_unique_monotonic_integer_heap_size(raw_unique, positive);
+        hp = alloc_heap(env, hsz);
+        return erts_raw_make_unique_monotonic_integer_value(&hp, raw_unique, positive);
+    } else {
+        Uint64 raw_unique[ERTS_UNIQUE_INT_RAW_VALUES];
+        erts_raw_get_unique_integer(raw_unique);
+        hsz = erts_raw_unique_integer_heap_size(raw_unique, positive);
+        hp = alloc_heap(env, hsz);
+        return erts_raw_make_unique_integer(&hp, raw_unique, positive);
+    }
+}
 
 ErlNifMutex* enif_mutex_create(char *name) { return erl_drv_mutex_create(name); }
 void enif_mutex_destroy(ErlNifMutex *mtx) {  erl_drv_mutex_destroy(mtx); }
diff --git a/erts/emulator/beam/erl_nif.h b/erts/emulator/beam/erl_nif.h
index c9c8b1d0af..bc1f59bc90 100644
--- a/erts/emulator/beam/erl_nif.h
+++ b/erts/emulator/beam/erl_nif.h
@@ -150,7 +150,12 @@ typedef enum
 typedef struct
 {
     ERL_NIF_TERM pid;  /* internal, may change */
-}ErlNifPid;
+} ErlNifPid;
+
+typedef struct
+{
+    ERL_NIF_TERM port_id;  /* internal, may change */
+}ErlNifPort;
 
 typedef ErlDrvSysInfo ErlNifSysInfo;
 
@@ -197,6 +202,15 @@ typedef enum {
     ERL_NIF_MAP_ITERATOR_TAIL = ERL_NIF_MAP_ITERATOR_LAST
 } ErlNifMapIteratorEntry;
 
+typedef enum {
+    ERL_NIF_UNIQUE_POSITIVE = (1 << 0),
+    ERL_NIF_UNIQUE_MONOTONIC = (1 << 1)
+} ErlNifUniqueInteger;
+
+typedef enum {
+    ERL_NIF_BIN2TERM_SAFE = 0x20000000
+} ErlNifBinaryToTerm;
+
 #if (defined(__WIN32__) || defined(_WIN32) || defined(_WIN32_))
 #  define ERL_NIF_API_FUNC_DECL(RET_TYPE, NAME, ARGS) RET_TYPE (*NAME) ARGS
 typedef struct {
diff --git a/erts/emulator/beam/erl_nif_api_funcs.h b/erts/emulator/beam/erl_nif_api_funcs.h
index 1448a508a2..35058afe7c 100644
--- a/erts/emulator/beam/erl_nif_api_funcs.h
+++ b/erts/emulator/beam/erl_nif_api_funcs.h
@@ -163,6 +163,15 @@ ERL_NIF_API_FUNC_DECL(int,enif_getenv,(const char* key, char* value, size_t* val
 ERL_NIF_API_FUNC_DECL(ErlNifTime, enif_monotonic_time, (ErlNifTimeUnit));
 ERL_NIF_API_FUNC_DECL(ErlNifTime, enif_time_offset, (ErlNifTimeUnit));
 ERL_NIF_API_FUNC_DECL(ErlNifTime, enif_convert_time_unit, (ErlNifTime, ErlNifTimeUnit, ErlNifTimeUnit));
+ERL_NIF_API_FUNC_DECL(ERL_NIF_TERM, enif_now_time, (ErlNifEnv *env));
+ERL_NIF_API_FUNC_DECL(ERL_NIF_TERM, enif_cpu_time, (ErlNifEnv *env));
+ERL_NIF_API_FUNC_DECL(ERL_NIF_TERM, enif_make_unique_integer, (ErlNifEnv *env, ErlNifUniqueInteger properties));
+ERL_NIF_API_FUNC_DECL(int, enif_is_process_alive, (ErlNifEnv *env, ErlNifPid *pid));
+ERL_NIF_API_FUNC_DECL(int, enif_is_port_alive, (ErlNifEnv *env, ErlNifPort *port_id));
+ERL_NIF_API_FUNC_DECL(int, enif_get_local_port, (ErlNifEnv* env, ERL_NIF_TERM, ErlNifPort* port_id));
+ERL_NIF_API_FUNC_DECL(int, enif_term_to_binary, (ErlNifEnv *env, ERL_NIF_TERM term, ErlNifBinary *bin));
+ERL_NIF_API_FUNC_DECL(size_t, enif_binary_to_term, (ErlNifEnv *env, const unsigned char* data, size_t sz, ERL_NIF_TERM *term, unsigned int opts));
+ERL_NIF_API_FUNC_DECL(int, enif_port_command, (ErlNifEnv *env, const ErlNifPort* to_port, ErlNifEnv *msg_env, ERL_NIF_TERM msg));
 
 /*
 ** ADD NEW ENTRIES HERE (before this comment) !!!
@@ -318,6 +327,15 @@ ERL_NIF_API_FUNC_DECL(int,enif_is_on_dirty_scheduler,(ErlNifEnv*));
 #  define enif_monotonic_time ERL_NIF_API_FUNC_MACRO(enif_monotonic_time)
 #  define enif_time_offset ERL_NIF_API_FUNC_MACRO(enif_time_offset)
 #  define enif_convert_time_unit ERL_NIF_API_FUNC_MACRO(enif_convert_time_unit)
+#  define enif_now_time ERL_NIF_API_FUNC_MACRO(enif_now_time)
+#  define enif_cpu_time ERL_NIF_API_FUNC_MACRO(enif_cpu_time)
+#  define enif_make_unique_integer ERL_NIF_API_FUNC_MACRO(enif_make_unique_integer)
+#  define enif_is_process_alive ERL_NIF_API_FUNC_MACRO(enif_is_process_alive)
+#  define enif_is_port_alive ERL_NIF_API_FUNC_MACRO(enif_is_port_alive)
+#  define enif_get_local_port ERL_NIF_API_FUNC_MACRO(enif_get_local_port)
+#  define enif_term_to_binary ERL_NIF_API_FUNC_MACRO(enif_term_to_binary)
+#  define enif_binary_to_term ERL_NIF_API_FUNC_MACRO(enif_binary_to_term)
+#  define enif_port_command ERL_NIF_API_FUNC_MACRO(enif_port_command)
 
 /*
 ** ADD NEW ENTRIES HERE (before this comment)
diff --git a/erts/emulator/beam/erl_port.h b/erts/emulator/beam/erl_port.h
index c25e08545e..5b06bc7fd1 100644
--- a/erts/emulator/beam/erl_port.h
+++ b/erts/emulator/beam/erl_port.h
@@ -978,6 +978,8 @@ ErtsPortOpResult erts_port_control(Process *, Port *, unsigned int, Eterm, Eterm
 ErtsPortOpResult erts_port_call(Process *, Port *, unsigned int, Eterm, Eterm *);
 ErtsPortOpResult erts_port_info(Process *, Port *, Eterm, Eterm *);
 
+int erts_port_output_async(Port *, Eterm, Eterm);
+
 /*
  * Signals from ports to ports. Used by sys drivers.
  */
diff --git a/erts/emulator/beam/erl_port_task.c b/erts/emulator/beam/erl_port_task.c
index e85395e062..b200344af5 100644
--- a/erts/emulator/beam/erl_port_task.c
+++ b/erts/emulator/beam/erl_port_task.c
@@ -180,10 +180,9 @@ p2p_sig_data_to_task(ErtsProc2PortSigData *sigdp)
     return ptp;
 }
 
-ErtsProc2PortSigData *
-erts_port_task_alloc_p2p_sig_data(void)
+static ERTS_INLINE ErtsProc2PortSigData *
+p2p_sig_data_init(ErtsPortTask *ptp)
 {
-    ErtsPortTask *ptp = port_task_alloc();
 
     ptp->type = ERTS_PORT_TASK_PROC_SIG;
     ptp->u.alive.flags = ERTS_PT_FLG_SIG_DEP;
@@ -194,6 +193,31 @@ erts_port_task_alloc_p2p_sig_data(void)
     return &ptp->u.alive.td.psig.data;
 }
 
+ErtsProc2PortSigData *
+erts_port_task_alloc_p2p_sig_data(void)
+{
+    ErtsPortTask *ptp = port_task_alloc();
+
+    return p2p_sig_data_init(ptp);
+}
+
+ErtsProc2PortSigData *
+erts_port_task_alloc_p2p_sig_data_extra(size_t extra, void **extra_ptr)
+{
+    ErtsPortTask *ptp = erts_alloc(ERTS_ALC_T_PORT_TASK,
+                                   sizeof(ErtsPortTask) + extra);
+
+    *extra_ptr = ptp+1;
+
+    return p2p_sig_data_init(ptp);
+}
+
+void
+erts_port_task_free_p2p_sig_data(ErtsProc2PortSigData *sigdp)
+{
+    schedule_port_task_free(p2p_sig_data_to_task(sigdp));
+}
+
 static ERTS_INLINE Eterm
 task_caller(ErtsPortTask *ptp)
 {
diff --git a/erts/emulator/beam/erl_port_task.h b/erts/emulator/beam/erl_port_task.h
index 335f7a77d5..56cef4e352 100644
--- a/erts/emulator/beam/erl_port_task.h
+++ b/erts/emulator/beam/erl_port_task.h
@@ -269,6 +269,8 @@ int erts_port_task_schedule(Eterm,
 void erts_port_task_free_port(Port *);
 int erts_port_is_scheduled(Port *);
 ErtsProc2PortSigData *erts_port_task_alloc_p2p_sig_data(void);
+ErtsProc2PortSigData *erts_port_task_alloc_p2p_sig_data_extra(size_t extra, void **extra_ptr);
+void erts_port_task_free_p2p_sig_data(ErtsProc2PortSigData *sigdp);
 
 #ifdef ERTS_SMP
 void erts_enqueue_port(ErtsRunQueue *rq, Port *pp);
diff --git a/erts/emulator/beam/erl_trace.c b/erts/emulator/beam/erl_trace.c
index 41876b8c62..9033c1b75c 100644
--- a/erts/emulator/beam/erl_trace.c
+++ b/erts/emulator/beam/erl_trace.c
@@ -176,7 +176,7 @@ take_timestamp(ErtsTraceTimeStamp *tsp, int ts_type)
 	    hsz += 3; /* 2-tuple */
 	    raw_unique = erts_raw_get_unique_monotonic_integer();
 	    tsp->u.monotonic.raw_unique = raw_unique;
-	    hsz += erts_raw_unique_monotonic_integer_heap_size(raw_unique);
+	    hsz += erts_raw_unique_monotonic_integer_heap_size(raw_unique, 0);
 	}
 	return hsz;
     }
@@ -216,8 +216,7 @@ write_timestamp(ErtsTraceTimeStamp *tsp, Eterm **hpp)
 	    return emtime;
 
 	raw = tsp->u.monotonic.raw_unique;
-	unique = erts_raw_make_unique_monotonic_integer_value(hpp,
-							      raw);
+	unique = erts_raw_make_unique_monotonic_integer_value(hpp, raw, 0);
 	res = TUPLE2(*hpp, emtime, unique);
 	*hpp += 3;
 	return res;
diff --git a/erts/emulator/beam/external.c b/erts/emulator/beam/external.c
index aac1490f0c..10f03636ec 100644
--- a/erts/emulator/beam/external.c
+++ b/erts/emulator/beam/external.c
@@ -967,19 +967,24 @@ erts_decode_dist_ext(ErtsHeapFactory* factory,
     return THE_NON_VALUE;
 }
 
-Eterm erts_decode_ext(ErtsHeapFactory* factory, byte **ext)
+Eterm erts_decode_ext(ErtsHeapFactory* factory, byte **ext, Uint32 flags)
 {
+    ErtsDistExternal ede, *edep;
     Eterm obj;
     byte *ep = *ext;
     if (*ep++ != VERSION_MAGIC) {
         erts_factory_undo(factory);
 	return THE_NON_VALUE;
     }
-    ep = dec_term(NULL, factory, ep, &obj, NULL);
+    if (flags) {
+        ASSERT(flags == ERTS_DIST_EXT_BTT_SAFE);
+        ede.flags = flags; /* a dummy struct just for the flags */
+        edep = &ede;
+    } else {
+        edep = NULL;
+    }
+    ep = dec_term(edep, factory, ep, &obj, NULL);
     if (!ep) {
-#ifdef DEBUG
-	bin_write(ERTS_PRINT_STDERR,NULL,*ext,500);
-#endif
 	return THE_NON_VALUE;
     }
     *ext = ep;
diff --git a/erts/emulator/beam/external.h b/erts/emulator/beam/external.h
index d12051c6b4..87eff2fe9f 100644
--- a/erts/emulator/beam/external.h
+++ b/erts/emulator/beam/external.h
@@ -191,7 +191,7 @@ Eterm erts_decode_dist_ext(ErtsHeapFactory* factory, ErtsDistExternal *);
 
 Sint erts_decode_ext_size(byte*, Uint);
 Sint erts_decode_ext_size_ets(byte*, Uint);
-Eterm erts_decode_ext(ErtsHeapFactory*, byte**);
+Eterm erts_decode_ext(ErtsHeapFactory*, byte**, Uint32 flags);
 Eterm erts_decode_ext_ets(ErtsHeapFactory*, byte*);
 
 Eterm erts_term_to_binary(Process* p, Eterm Term, int level, Uint flags);
diff --git a/erts/emulator/beam/io.c b/erts/emulator/beam/io.c
index b2fd8984fd..29f28cc9dc 100644
--- a/erts/emulator/beam/io.c
+++ b/erts/emulator/beam/io.c
@@ -1760,7 +1760,6 @@ cleanup_scheduled_outputv(ErlIOVec *ev, ErlDrvBinary *cbinp)
 	    driver_free_binary(ev->binv[i]);
     if (cbinp)
 	driver_free_binary(cbinp);
-    erts_free(ERTS_ALC_T_DRV_CMD_DATA, ev);
 }
 
 static int
@@ -1898,6 +1897,188 @@ port_sig_output(Port *prt, erts_aint32_t state, int op, ErtsProc2PortSigData *si
     return ERTS_PORT_REDS_CMD_OUTPUT;
 }
 
+
+/*
+ * This erts_port_output will always create a port task.
+ * The call is treated as a port_command call, i.e. no
+ * badsig i generated if the input in invalid. However
+ * an error_logger message is generated.
+ */
+int
+erts_port_output_async(Port *prt, Eterm from, Eterm list)
+{
+
+    ErtsPortOpResult res;
+    ErtsProc2PortSigData *sigdp;
+    erts_driver_t *drv = prt->drv_ptr;
+    size_t size;
+    int task_flags;
+    ErtsProc2PortSigCallback port_sig_callback;
+    ErlDrvBinary *cbin = NULL;
+    ErlIOVec *evp = NULL;
+    char *buf = NULL;
+    ErtsPortTaskHandle *ns_pthp;
+
+    if (drv->outputv) {
+        ErlIOVec ev;
+	SysIOVec* ivp;
+	ErlDrvBinary**  bvp;
+        int vsize;
+	Uint csize;
+	Uint pvsize;
+	Uint pcsize;
+        size_t iov_offset, binv_offset, alloc_size;
+        Uint blimit = 0;
+        char *ptr;
+        int i;
+
+        Eterm* bptr = NULL;
+        Uint offset;
+
+        if (is_binary(list)) {
+            /* We optimize for when we get a procbin without offset */
+            Eterm real_bin;
+            int bitoffs;
+            int bitsize;
+            ERTS_GET_REAL_BIN(list, real_bin, offset, bitoffs, bitsize);
+            bptr = binary_val(real_bin);
+            if (*bptr == HEADER_PROC_BIN && bitoffs == 0) {
+                size = binary_size(list);
+                vsize = 1;
+            } else
+                bptr = NULL;
+        }
+
+        if (!bptr) {
+            if (io_list_vec_len(list, &vsize, &csize, &pvsize, &pcsize, &size))
+                goto bad_value;
+
+            /* To pack or not to pack (small binaries) ...? */
+            if (vsize >= SMALL_WRITE_VEC) {
+                /* Do pack */
+                vsize = pvsize + 1;
+                csize = pcsize;
+                blimit = ERL_SMALL_IO_BIN_LIMIT;
+            }
+            cbin = driver_alloc_binary(csize);
+            if (!cbin)
+                erts_alloc_enomem(ERTS_ALC_T_DRV_BINARY, ERTS_SIZEOF_Binary(csize));
+        }
+
+
+        iov_offset = ERTS_ALC_DATA_ALIGN_SIZE(sizeof(ErlIOVec));
+	binv_offset = iov_offset;
+        binv_offset += ERTS_ALC_DATA_ALIGN_SIZE((vsize+1)*sizeof(SysIOVec));
+        alloc_size = binv_offset;
+	alloc_size += (vsize+1)*sizeof(ErlDrvBinary *);
+
+        sigdp = erts_port_task_alloc_p2p_sig_data_extra(alloc_size, (void**)&ptr);
+
+        evp = (ErlIOVec *) ptr;
+        ivp = evp->iov = (SysIOVec *) (ptr + iov_offset);
+        bvp = evp->binv = (ErlDrvBinary **) (ptr + binv_offset);
+
+        ivp[0].iov_base = NULL;
+	ivp[0].iov_len = 0;
+	bvp[0] = NULL;
+
+        if (bptr) {
+            ProcBin* pb = (ProcBin *) bptr;
+
+            ivp[1].iov_base = pb->bytes+offset;
+            ivp[1].iov_len = size;
+            bvp[1] = Binary2ErlDrvBinary(pb->val);
+
+            evp->vsize = 1;
+        } else {
+
+            evp->vsize = io_list_to_vec(list, ivp+1, bvp+1, cbin, blimit);
+            if (evp->vsize < 0) {
+                if (evp != &ev)
+                    erts_free(ERTS_ALC_T_DRV_CMD_DATA, evp);
+                driver_free_binary(cbin);
+                goto bad_value;
+            }
+        }
+#if 0
+	/* This assertion may say something useful, but it can
+        be falsified during the emulator test suites. */
+	ASSERT(evp->vsize == vsize);
+#endif
+	evp->vsize++;
+	evp->size = size;  /* total size */
+
+        /* Need to increase refc on all binaries */
+        for (i = 1; i < evp->vsize; i++)
+            if (bvp[i])
+                driver_binary_inc_refc(bvp[i]);
+
+	sigdp->flags = ERTS_P2P_SIG_TYPE_OUTPUTV;
+	sigdp->u.outputv.from = from;
+	sigdp->u.outputv.evp = evp;
+	sigdp->u.outputv.cbinp = cbin;
+	port_sig_callback = port_sig_outputv;
+    } else {
+        ErlDrvSizeT ERTS_DECLARE_DUMMY(r);
+
+	/*
+	 * Apperently there exist code that write 1 byte to
+	 * much in buffer. Where it resides I don't know, but
+	 * we can live with one byte extra allocated...
+	 */
+
+        if (erts_iolist_size(list, &size))
+            goto bad_value;
+
+        buf = erts_alloc(ERTS_ALC_T_DRV_CMD_DATA, size + 1);
+
+        r = erts_iolist_to_buf(list, buf, size);
+        ASSERT(ERTS_IOLIST_TO_BUF_SUCCEEDED(r));
+
+	sigdp = erts_port_task_alloc_p2p_sig_data();
+	sigdp->flags = ERTS_P2P_SIG_TYPE_OUTPUT;
+	sigdp->u.output.from = from;
+	sigdp->u.output.bufp = buf;
+	sigdp->u.output.size = size;
+	port_sig_callback = port_sig_output;
+    }
+    sigdp->flags = 0;
+    ns_pthp = NULL;
+    task_flags = 0;
+
+    res = erts_schedule_proc2port_signal(NULL,
+					 prt,
+					 ERTS_INVALID_PID,
+					 NULL,
+					 sigdp,
+					 task_flags,
+					 ns_pthp,
+					 port_sig_callback);
+
+    if (res != ERTS_PORT_OP_SCHEDULED) {
+	if (drv->outputv)
+	    cleanup_scheduled_outputv(evp, cbin);
+	else
+	    cleanup_scheduled_output(buf);
+	return 1;
+    }
+    return 1;
+
+bad_value:
+
+    /*
+     * We call badsig directly here as this function is called with
+     * the main lock of the calling process still held.
+     * At the moment this operation is always not a bang_op, so
+     * only an error_logger message should be generated, no badsig.
+    */
+
+    badsig_received(0, prt, erts_atomic32_read_nob(&prt->state), 1);
+
+    return 0;
+
+}
+
 ErtsPortOpResult
 erts_port_output(Process *c_p,
 		 int flags,
@@ -1907,7 +2088,7 @@ erts_port_output(Process *c_p,
 		 Eterm *refp)
 {
     ErtsPortOpResult res;
-    ErtsProc2PortSigData *sigdp;
+    ErtsProc2PortSigData *sigdp = NULL;
     erts_driver_t *drv = prt->drv_ptr;
     size_t size;
     int try_call;
@@ -1989,10 +2170,13 @@ erts_port_output(Process *c_p,
 	    evp = &ev;
 	}
 	else {
-	    char *ptr = erts_alloc((try_call
-				    ? ERTS_ALC_T_TMP
-				    : ERTS_ALC_T_DRV_CMD_DATA), alloc_size);
-
+            char *ptr;
+            if (try_call) {
+                ptr = erts_alloc(ERTS_ALC_T_TMP, alloc_size);
+            } else {
+                sigdp = erts_port_task_alloc_p2p_sig_data_extra(
+                    alloc_size, (void**)&ptr);
+            }
 	    evp = (ErlIOVec *) ptr;
 	    ivp = evp->iov = (SysIOVec *) (ptr + iov_offset);
 	    bvp = evp->binv = (ErlDrvBinary **) (ptr + binv_offset);
@@ -2021,9 +2205,12 @@ erts_port_output(Process *c_p,
 	bvp[0] = NULL;
 	evp->vsize = io_list_to_vec(list, ivp+1, bvp+1, cbin, blimit);
 	if (evp->vsize < 0) {
-	    if (evp != &ev)
-		erts_free(try_call ? ERTS_ALC_T_TMP : ERTS_ALC_T_DRV_CMD_DATA,
-			  evp);
+            if (evp != &ev) {
+                if (try_call)
+                    erts_free(ERTS_ALC_T_TMP, evp);
+                else
+                    erts_port_task_free_p2p_sig_data(sigdp);
+            }
 	    driver_free_binary(cbin);
 	    goto bad_value;
 	}
@@ -2075,8 +2262,10 @@ erts_port_output(Process *c_p,
 		/* Fall through... */
 	    case ERTS_TRY_IMM_DRV_CALL_INVALID_PORT:
 		driver_free_binary(cbin);
-		if (evp != &ev)
+		if (evp != &ev) {
+                    ASSERT(!sigdp);
 		    erts_free(ERTS_ALC_T_TMP, evp);
+                }
 		if (try_call_res != ERTS_TRY_IMM_DRV_CALL_OK)
 		    return ERTS_PORT_OP_DROPPED;
 		if (c_p)
@@ -2087,8 +2276,10 @@ erts_port_output(Process *c_p,
 		if (async_nosuspend
 		    && (sched_flags & (busy_flgs|ERTS_PTS_FLG_EXIT))) {
 		    driver_free_binary(cbin);
-		    if (evp != &ev)
+                    if (evp != &ev) {
+                        ASSERT(!sigdp);
 			erts_free(ERTS_ALC_T_TMP, evp);
+                    }
 		    return ((sched_flags & ERTS_PTS_FLG_EXIT)
 			    ? ERTS_PORT_OP_DROPPED
 			    : ERTS_PORT_OP_BUSY);
@@ -2103,9 +2294,16 @@ erts_port_output(Process *c_p,
 		if (bvp[i])
 		    driver_binary_inc_refc(bvp[i]);
 
-	    new_evp = erts_alloc(ERTS_ALC_T_DRV_CMD_DATA, alloc_size);
+            /* The port task and iovec is allocated in the
+               same structure as an optimization. This
+               is especially important in erts_port_output_async
+               of when !try_call */
+            ASSERT(sigdp == NULL);
+            sigdp = erts_port_task_alloc_p2p_sig_data_extra(
+                alloc_size, (void**)&new_evp);
 
 	    if (evp != &ev) {
+                /* Copy from TMP alloc to port task */
 		sys_memcpy((void *) new_evp, (void *) evp, alloc_size);
 		new_evp->iov = (SysIOVec *) (((char *) new_evp)
 					     + iov_offset);
@@ -2153,7 +2351,6 @@ erts_port_output(Process *c_p,
 	    evp = new_evp;
 	}
 
-	sigdp = erts_port_task_alloc_p2p_sig_data();
 	sigdp->flags = ERTS_P2P_SIG_TYPE_OUTPUTV;
 	sigdp->u.outputv.from = from;
 	sigdp->u.outputv.evp = evp;
@@ -4386,7 +4583,7 @@ port_sig_call(Port *prt,
 
                 (void) erts_factory_message_create(&factory, rp, &rp_locks, hsz);
 		endp = (byte *) resp_bufp;
-		msg = erts_decode_ext(&factory, &endp);
+		msg = erts_decode_ext(&factory, &endp, 0);
 		if (is_value(msg)) {
                     hp = erts_produce_heap(&factory,
                                            3,
@@ -4505,7 +4702,7 @@ erts_port_call(Process* c_p,
 	    hsz += 3;
             erts_factory_proc_prealloc_init(&factory, c_p, hsz);
 	    endp = (byte *) resp_bufp;
-	    term = erts_decode_ext(&factory, &endp);
+	    term = erts_decode_ext(&factory, &endp, 0);
 	    if (term == THE_NON_VALUE)
 		return ERTS_PORT_OP_BADARG;
             hp = erts_produce_heap(&factory,3,0);