Merge branch 'maint'

* maint:
  Fix busy_port_SUITE:io_to_busy test-case
  Ensure "runnable port" trace messages are not sent out of order
  Ensure "runnable proc" trace messages are not sent out of order

7 files changed, 186 insertions, 60 deletions

diff --git a/erts/emulator/beam/erl_lock_check.c b/erts/emulator/beam/erl_lock_check.c
index c665aa51a2..b105ece6f1 100644
--- a/erts/emulator/beam/erl_lock_check.c
+++ b/erts/emulator/beam/erl_lock_check.c
@@ -139,7 +139,6 @@ static erts_lc_lock_order_t erts_lock_order[] = {
     {	"drv_tsd",				NULL			},
     {	"async_enq_mtx",			NULL			},
 #ifdef ERTS_SMP
-    {	"sys_msg_q", 				NULL			},
     {	"atom_tab",				NULL			},
     {	"make_ref",				NULL			},
     {	"misc_op_list_pre_alloc_lock",		"address"		},
@@ -148,6 +147,7 @@ static erts_lc_lock_order_t erts_lock_order[] = {
     {	"btm_pre_alloc_lock",			NULL,			},
     {	"dist_entry_out_queue",			"address"		},
     {	"port_sched_lock",			"port_id"		},
+    {	"sys_msg_q", 				NULL			},
     {   "port_table",                           NULL                    },
 #endif
     {	"mtrace_op",				NULL			},
diff --git a/erts/emulator/beam/erl_message.c b/erts/emulator/beam/erl_message.c
index 59a677a12c..8870fac7d9 100644
--- a/erts/emulator/beam/erl_message.c
+++ b/erts/emulator/beam/erl_message.c
@@ -415,7 +415,13 @@ erts_queue_dist_message(Process *rcvr,
 	if (!(*rcvr_locks & ERTS_PROC_LOCK_MSGQ))
 	    erts_smp_proc_unlock(rcvr, ERTS_PROC_LOCK_MSGQ);
 
-	erts_proc_notify_new_message(rcvr);
+	erts_proc_notify_new_message(rcvr,
+#ifdef ERTS_SMP
+				     *rcvr_locks
+#else
+				     0
+#endif
+	    );
     }
 }
 
@@ -542,7 +548,13 @@ queue_message(Process *c_p,
     if (locked_msgq)
 	erts_smp_proc_unlock(receiver, ERTS_PROC_LOCK_MSGQ);
 
-    erts_proc_notify_new_message(receiver);
+    erts_proc_notify_new_message(receiver,
+#ifdef ERTS_SMP
+				 *receiver_locks
+#else
+				 0
+#endif
+	);
 
 #ifndef ERTS_SMP
     ERTS_HOLE_CHECK(receiver);
diff --git a/erts/emulator/beam/erl_port_task.c b/erts/emulator/beam/erl_port_task.c
index 4103d1192a..682f6f8f4b 100644
--- a/erts/emulator/beam/erl_port_task.c
+++ b/erts/emulator/beam/erl_port_task.c
@@ -999,6 +999,7 @@ static ERTS_INLINE int
 finalize_exec(Port *pp, ErtsPortTask **execq, int processing_busy_q)
 {
     erts_aint32_t act;
+    unsigned int prof_runnable_ports;
 
     if (!processing_busy_q)
 	pp->sched.taskq.local.first = *execq;
@@ -1015,6 +1016,10 @@ finalize_exec(Port *pp, ErtsPortTask **execq, int processing_busy_q)
     if (act & ERTS_PTS_FLG_CHK_UNSET_BUSY_PORT_Q)
 	act = check_unset_busy_port_q(pp, act, pp->sched.taskq.bpq);
 
+    prof_runnable_ports = erts_system_profile_flags.runnable_ports;
+    if (prof_runnable_ports)
+	erts_port_task_sched_lock(&pp->sched);
+
     while (1) {
 	erts_aint32_t new, exp;
 
@@ -1026,12 +1031,24 @@ finalize_exec(Port *pp, ErtsPortTask **execq, int processing_busy_q)
 
 	act = erts_smp_atomic32_cmpxchg_relb(&pp->sched.flags, new, exp);
 
-	ASSERT(!(act & ERTS_PTS_FLG_IN_RUNQ));
+	ERTS_LC_ASSERT(!(act & ERTS_PTS_FLG_IN_RUNQ));
+	ERTS_LC_ASSERT(!(act & ERTS_PTS_FLG_EXEC_IMM));
 
 	if (exp == act)
 	    break;
     }
 
+    if (prof_runnable_ports | IS_TRACED_FL(pp, F_TRACE_SCHED_PORTS)) {
+	/* trace port scheduling, out */
+	if (IS_TRACED_FL(pp, F_TRACE_SCHED_PORTS))
+	    trace_sched_ports(pp, am_out);
+	if (prof_runnable_ports) {
+	    if (!(act & (ERTS_PTS_FLG_EXEC_IMM|ERTS_PTS_FLG_HAVE_TASKS)))
+		profile_runnable_port(pp, am_inactive);
+	    erts_port_task_sched_unlock(&pp->sched);
+	}
+    }
+
     return (act & ERTS_PTS_FLG_HAVE_TASKS) != 0;
 }
 
@@ -1377,6 +1394,7 @@ erts_port_task_schedule(Eterm id,
     Port *pp;
     ErtsPortTask *ptp = NULL;
     erts_aint32_t act, add_flags;
+    unsigned int prof_runnable_ports;
 
     if (pthp && erts_port_task_is_scheduled(pthp)) {
 	ASSERT(0);
@@ -1465,6 +1483,10 @@ erts_port_task_schedule(Eterm id,
     if (ns_pthlp)
 	add_flags |= ERTS_PTS_FLG_HAVE_NS_TASKS;
 
+    prof_runnable_ports = erts_system_profile_flags.runnable_ports;
+    if (prof_runnable_ports)
+	erts_port_task_sched_lock(&pp->sched);
+
     while (1) {
 	erts_aint32_t new, exp;
 
@@ -1489,6 +1511,13 @@ erts_port_task_schedule(Eterm id,
 	    goto done; /* Died after our task insert... */
     }
 
+    if (prof_runnable_ports) {
+	if (!(act & ERTS_PTS_FLG_EXEC_IMM))
+	    profile_runnable_port(pp, am_active);
+	erts_port_task_sched_unlock(&pp->sched);
+	prof_runnable_ports = 0;
+    }
+
     /* Enqueue port on run-queue */
 
     runq = erts_port_runq(pp);
@@ -1510,10 +1539,6 @@ erts_port_task_schedule(Eterm id,
 #endif
 
     enqueue_port(runq, pp);
-	    
-    if (erts_system_profile_flags.runnable_ports) {
-	profile_runnable_port(pp, am_active);
-    }
 
     erts_smp_runq_unlock(runq);
 
@@ -1521,6 +1546,9 @@ erts_port_task_schedule(Eterm id,
 
 done:
 
+    if (prof_runnable_ports)
+	erts_port_task_sched_unlock(&pp->sched);
+
 #ifdef ERTS_SMP
     if (dhndl != ERTS_THR_PRGR_DHANDLE_MANAGED)
 	erts_port_dec_refc(pp);
@@ -1777,10 +1805,6 @@ erts_port_task_execute(ErtsRunQueue *runq, Port **curr_port_pp)
 
     erts_unblock_fpe(fpe_was_unmasked);
 
-    /* trace port scheduling, out */
-    if (IS_TRACED_FL(pp, F_TRACE_SCHED_PORTS)) {
-    	trace_sched_ports(pp, am_out);
-    }
 
     if (io_tasks_executed) {
 	ASSERT(erts_smp_atomic_read_nob(&erts_port_task_outstanding_io_tasks)
@@ -1803,11 +1827,7 @@ erts_port_task_execute(ErtsRunQueue *runq, Port **curr_port_pp)
 
     erts_smp_runq_lock(runq);
  
-    if (!active) {
-	if (erts_system_profile_flags.runnable_ports)
-	    profile_runnable_port(pp, am_inactive);
-    }
-    else {
+    if (active) {
 #ifdef ERTS_SMP
 	ErtsRunQueue *xrunq;
 #endif
diff --git a/erts/emulator/beam/erl_port_task.h b/erts/emulator/beam/erl_port_task.h
index 1d30465ec9..9ef0cfcedc 100644
--- a/erts/emulator/beam/erl_port_task.h
+++ b/erts/emulator/beam/erl_port_task.h
@@ -78,6 +78,7 @@ extern erts_smp_atomic_t erts_port_task_outstanding_io_tasks;
 #define ERTS_PTS_FLG_PARALLELISM		(((erts_aint32_t) 1) <<  9)
 #define ERTS_PTS_FLG_FORCE_SCHED		(((erts_aint32_t) 1) << 10)
 #define ERTS_PTS_FLG_EXITING			(((erts_aint32_t) 1) << 11)
+#define ERTS_PTS_FLG_EXEC_IMM			(((erts_aint32_t) 1) << 12)
 
 #define ERTS_PTS_FLGS_BUSY \
     (ERTS_PTS_FLG_BUSY_PORT | ERTS_PTS_FLG_BUSY_PORT_Q)
@@ -87,6 +88,7 @@ extern erts_smp_atomic_t erts_port_task_outstanding_io_tasks;
      | ERTS_PTS_FLG_HAVE_BUSY_TASKS		\
      | ERTS_PTS_FLG_HAVE_TASKS			\
      | ERTS_PTS_FLG_EXEC			\
+     | ERTS_PTS_FLG_EXEC_IMM			\
      | ERTS_PTS_FLG_FORCE_SCHED			\
      | ERTS_PTS_FLG_EXITING)
 
diff --git a/erts/emulator/beam/erl_process.c b/erts/emulator/beam/erl_process.c
index e96d5c4ccd..685004f267 100644
--- a/erts/emulator/beam/erl_process.c
+++ b/erts/emulator/beam/erl_process.c
@@ -5883,6 +5883,9 @@ schedule_out_process(ErtsRunQueue *c_rq, erts_aint32_t state, Process *p, Proces
     case ERTS_ENQUEUE_NOT:
 	if (erts_system_profile_flags.runnable_procs) {
 
+	    /* Status lock prevents out of order "runnable proc" trace msgs */
+	    ERTS_SMP_LC_ASSERT(ERTS_PROC_LOCK_STATUS & erts_proc_lc_my_proc_locks(p));
+
 	    if (!(a & ERTS_PSFLG_ACTIVE_SYS)
 		&& (!(a & ERTS_PSFLG_ACTIVE)
 		    || (a & ERTS_PSFLG_SUSPENDED))) {
@@ -5996,7 +5999,8 @@ 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 *enq_prio_p,
+			   ErtsProcLocks locks)
 {
     /*
      * NOTE: ERTS_PSFLG_RUNNING, ERTS_PSFLG_RUNNING_SYS and
@@ -6005,6 +6009,11 @@ change_proc_schedule_state(Process *p,
      */
     erts_aint32_t a = *statep, n;
     int enqueue; /* < 0 -> use proxy */
+    unsigned int prof_runnable_procs = erts_system_profile_flags.runnable_procs;
+    unsigned int lock_status = (prof_runnable_procs
+				&& !(locks & ERTS_PROC_LOCK_STATUS));
+
+    ERTS_SMP_LC_ASSERT(locks == erts_proc_lc_my_proc_locks(p));
 
     ASSERT(!(a & ERTS_PSFLG_PROXY));
     ASSERT((clear_state_flags & (ERTS_PSFLG_RUNNING
@@ -6014,6 +6023,9 @@ change_proc_schedule_state(Process *p,
 			       | ERTS_PSFLG_RUNNING_SYS
 			       | ERTS_PSFLG_ACTIVE_SYS)) == 0);
 
+    if (lock_status)
+	erts_smp_proc_lock(p, ERTS_PROC_LOCK_STATUS);
+
     while (1) {
 	erts_aint32_t e;
 	n = e = a;
@@ -6049,7 +6061,9 @@ change_proc_schedule_state(Process *p,
 	    break;
     }
 
-    if (erts_system_profile_flags.runnable_procs) {
+    if (prof_runnable_procs) {
+
+	/* Status lock prevents out of order "runnable proc" trace msgs */
 
 	if (((n & (ERTS_PSFLG_SUSPENDED
 		   | ERTS_PSFLG_ACTIVE)) == ERTS_PSFLG_ACTIVE)
@@ -6062,15 +6076,18 @@ change_proc_schedule_state(Process *p,
 	    profile_runnable_proc(p, am_active);
 	}
 
+	if (lock_status)
+	    erts_smp_proc_unlock(p, ERTS_PROC_LOCK_STATUS);
     }
 
+
     *statep = a;
 
     return enqueue;
 }
 
 static ERTS_INLINE void
-schedule_process(Process *p, erts_aint32_t in_state)
+schedule_process(Process *p, erts_aint32_t in_state, ErtsProcLocks locks)
 {
     erts_aint32_t enq_prio  = -1;
     erts_aint32_t state = in_state;
@@ -6078,7 +6095,8 @@ schedule_process(Process *p, erts_aint32_t in_state)
 					     0,
 					     ERTS_PSFLG_ACTIVE,
 					     &state,
-					     &enq_prio);
+					     &enq_prio,
+					     locks);
     if (enqueue != ERTS_ENQUEUE_NOT)
 	add2runq(enqueue > 0 ? p : make_proxy_proc(NULL, p, enq_prio),
 		 state,
@@ -6086,16 +6104,27 @@ schedule_process(Process *p, erts_aint32_t in_state)
 }
 
 void
-erts_schedule_process(Process *p, erts_aint32_t state)
+erts_schedule_process(Process *p, erts_aint32_t state, ErtsProcLocks locks)
 {
-    schedule_process(p, state);
+    schedule_process(p, state, locks);
 }
 
 static void
 schedule_process_sys_task(Process *p, erts_aint32_t state, Process *proxy)
 {
+    /*
+     * Expects status lock to be locked when called, and
+     * returns with status lock unlocked...
+     */
     erts_aint32_t a = state, n, enq_prio = -1;
     int enqueue; /* < 0 -> use proxy */
+    unsigned int prof_runnable_procs = erts_system_profile_flags.runnable_procs;
+
+    /* Status lock prevents out of order "runnable proc" trace msgs */
+    ERTS_SMP_LC_ASSERT(ERTS_PROC_LOCK_STATUS & erts_proc_lc_my_proc_locks(p));
+
+    if (!prof_runnable_procs)
+	erts_smp_proc_unlock(p, ERTS_PROC_LOCK_STATUS);
 
     ASSERT(!(state & ERTS_PSFLG_PROXY));
 
@@ -6104,7 +6133,7 @@ schedule_process_sys_task(Process *p, erts_aint32_t state, Process *proxy)
 	n = e = a;
 
 	if (a & ERTS_PSFLG_FREE)
-	    return; /* We don't want to schedule free processes... */
+	    goto cleanup; /* We don't want to schedule free processes... */
 
 	enqueue = ERTS_ENQUEUE_NOT;
 	n |= ERTS_PSFLG_ACTIVE_SYS;
@@ -6117,7 +6146,7 @@ schedule_process_sys_task(Process *p, erts_aint32_t state, Process *proxy)
 	    goto cleanup;
     }
 
-    if (erts_system_profile_flags.runnable_procs) {
+    if (prof_runnable_procs) {
 
 	if (!(a & (ERTS_PSFLG_ACTIVE_SYS
 		   | ERTS_PSFLG_RUNNING
@@ -6127,6 +6156,8 @@ schedule_process_sys_task(Process *p, erts_aint32_t state, Process *proxy)
 	    profile_runnable_proc(p, am_active);
 	}
 
+	erts_smp_proc_unlock(p, ERTS_PROC_LOCK_STATUS);
+	prof_runnable_procs = 0;
     }
 
     if (enqueue != ERTS_ENQUEUE_NOT) {
@@ -6141,8 +6172,14 @@ schedule_process_sys_task(Process *p, erts_aint32_t state, Process *proxy)
     }
 
 cleanup:
+
+    if (prof_runnable_procs)
+	erts_smp_proc_unlock(p, ERTS_PROC_LOCK_STATUS);
+
     if (proxy)
 	free_proxy_proc(proxy);
+
+    ERTS_SMP_LC_ASSERT(!(ERTS_PROC_LOCK_STATUS & erts_proc_lc_my_proc_locks(p)));
 }
 
 static ERTS_INLINE int
@@ -6209,7 +6246,7 @@ suspend_process(Process *c_p, Process *p)
 }
 
 static ERTS_INLINE void
-resume_process(Process *p)
+resume_process(Process *p, ErtsProcLocks locks)
 {
     erts_aint32_t state, enq_prio = -1;
     int enqueue;
@@ -6226,7 +6263,8 @@ resume_process(Process *p)
 					 ERTS_PSFLG_SUSPENDED,
 					 0,
 					 &state,
-					 &enq_prio);
+					 &enq_prio,
+					 locks);
     if (enqueue)
 	add2runq(enqueue > 0 ? p : make_proxy_proc(NULL, p, enq_prio),
 		 state,
@@ -8042,7 +8080,8 @@ handle_pend_sync_suspend(Process *suspendee,
 	}
 	/* suspender is suspended waiting for suspendee to suspend;
 	   resume suspender */
-	resume_process(suspender);
+	ASSERT(suspendee != suspender);
+	resume_process(suspender, ERTS_PROC_LOCK_STATUS);
 	erts_smp_proc_unlock(suspender, ERTS_PROC_LOCK_STATUS);
     }
 }
@@ -8077,7 +8116,7 @@ pid2proc_not_running(Process *c_p, ErtsProcLocks c_p_locks,
 	ASSERT(c_p->flags & F_P2PNR_RESCHED);
 	c_p->flags &= ~F_P2PNR_RESCHED;
 	if (!suspend && rp)
-	    resume_process(rp);
+	    resume_process(rp, rp_locks);
     }
     else {
 
@@ -8235,7 +8274,8 @@ handle_pend_bif_sync_suspend(Process *suspendee,
 	}
 	/* suspender is suspended waiting for suspendee to suspend;
 	   resume suspender */
-	resume_process(suspender);
+	ASSERT(suspender != suspendee);
+	resume_process(suspender, ERTS_PROC_LOCK_LINK|ERTS_PROC_LOCK_STATUS);
 	erts_smp_proc_unlock(suspender,
 			     ERTS_PROC_LOCK_LINK|ERTS_PROC_LOCK_STATUS);
     }
@@ -8595,7 +8635,8 @@ resume_process_1(BIF_ALIST_1)
 
 	ASSERT(ERTS_PSFLG_SUSPENDED
 	       & erts_smp_atomic32_read_nob(&suspendee->state));
-	resume_process(suspendee);
+	ASSERT(BIF_P != suspendee);
+	resume_process(suspendee, ERTS_PROC_LOCK_STATUS);
 
 	erts_smp_proc_unlock(suspendee, ERTS_PROC_LOCK_STATUS);
     }
@@ -8725,7 +8766,7 @@ erts_resume(Process* process, ErtsProcLocks process_locks)
     ERTS_SMP_LC_ASSERT(process_locks == erts_proc_lc_my_proc_locks(process));
     if (!(process_locks & ERTS_PROC_LOCK_STATUS))
 	erts_smp_proc_lock(process, ERTS_PROC_LOCK_STATUS);
-    resume_process(process);
+    resume_process(process, process_locks|ERTS_PROC_LOCK_STATUS);
     if (!(process_locks & ERTS_PROC_LOCK_STATUS))
 	erts_smp_proc_unlock(process, ERTS_PROC_LOCK_STATUS);
 }
@@ -8744,7 +8785,7 @@ erts_resume_processes(ErtsProcList *list)
 	proc = erts_pid2proc(NULL, 0, plp->pid, ERTS_PROC_LOCK_STATUS);
 	if (proc) {
 	    if (erts_proclist_same(plp, proc)) {
-		resume_process(proc);
+		resume_process(proc, ERTS_PROC_LOCK_STATUS);
 		nresumed++;
 	    }
 	    erts_smp_proc_unlock(proc, ERTS_PROC_LOCK_STATUS);
@@ -9980,8 +10021,10 @@ erts_internal_request_system_task_3(BIF_ALIST_3)
 	rp_state = n;
     }
 
-    erts_smp_proc_unlock(rp, ERTS_PROC_LOCK_STATUS);
-
+    /*
+     * schedule_process_sys_task() unlocks status
+     * lock on process.
+     */
     schedule_process_sys_task(rp, rp_state, NULL);
 
     if (free_stqs)
@@ -10726,7 +10769,7 @@ erl_create_process(Process* parent, /* Parent of process (default group leader).
      * Schedule process for execution.
      */
 
-    schedule_process(p, state);
+    schedule_process(p, state, 0);
 
     VERBOSE(DEBUG_PROCESSES, ("Created a new process: %T\n",p->common.id));
 
@@ -11047,7 +11090,8 @@ set_proc_exiting(Process *p,
 					 ERTS_PSFLG_SUSPENDED|ERTS_PSFLG_PENDING_EXIT,
 					 ERTS_PSFLG_EXITING|ERTS_PSFLG_ACTIVE,
 					 &state,
-					 &enq_prio);
+					 &enq_prio,
+					 ERTS_PROC_LOCKS_ALL);
 
     p->fvalue = reason;
     if (bp)
@@ -11088,7 +11132,8 @@ set_proc_self_exiting(Process *c_p)
 				   ERTS_PSFLG_SUSPENDED|ERTS_PSFLG_PENDING_EXIT,
 				   ERTS_PSFLG_EXITING|ERTS_PSFLG_ACTIVE,
 				   &state,
-				   &enq_prio);
+				   &enq_prio,
+				   ERTS_PROC_LOCKS_ALL);
 
     ASSERT(!enqueue);
     return state;
@@ -11733,8 +11778,9 @@ resume_suspend_monitor(ErtsSuspendMonitor *smon, void *vc_p)
     Process *suspendee = erts_pid2proc((Process *) vc_p, ERTS_PROC_LOCK_MAIN,
 				       smon->pid, ERTS_PROC_LOCK_STATUS);
     if (suspendee) {
+	ASSERT(suspendee != vc_p);
 	if (smon->active)
-	    resume_process(suspendee);
+	    resume_process(suspendee, ERTS_PROC_LOCK_STATUS);
 	erts_smp_proc_unlock(suspendee, ERTS_PROC_LOCK_STATUS);
     }
     erts_destroy_suspend_monitor(smon);
@@ -12067,7 +12113,7 @@ timeout_proc(Process* p)
 
     state = erts_smp_atomic32_read_acqb(&p->state);
     if (!(state & ERTS_PSFLG_ACTIVE))
-	schedule_process(p, state);
+	schedule_process(p, state, ERTS_PROC_LOCK_MAIN|ERTS_PROC_LOCK_STATUS);
 }
 
 
diff --git a/erts/emulator/beam/erl_process.h b/erts/emulator/beam/erl_process.h
index 31f4a09c94..9b740f049e 100644
--- a/erts/emulator/beam/erl_process.h
+++ b/erts/emulator/beam/erl_process.h
@@ -1700,17 +1700,17 @@ ErtsSchedulerData *erts_get_scheduler_data(void)
 #endif
 #endif
 
-void erts_schedule_process(Process *, erts_aint32_t);
+void erts_schedule_process(Process *, erts_aint32_t, ErtsProcLocks);
 
-ERTS_GLB_INLINE void erts_proc_notify_new_message(Process *p);
+ERTS_GLB_INLINE void erts_proc_notify_new_message(Process *p, ErtsProcLocks locks);
 #if ERTS_GLB_INLINE_INCL_FUNC_DEF
 ERTS_GLB_INLINE void
-erts_proc_notify_new_message(Process *p)
+erts_proc_notify_new_message(Process *p, ErtsProcLocks locks)
 {
     /* No barrier needed, due to msg lock */
     erts_aint32_t state = erts_smp_atomic32_read_nob(&p->state);
     if (!(state & ERTS_PSFLG_ACTIVE))
-	erts_schedule_process(p, state);
+	erts_schedule_process(p, state, locks);
 }
 #endif
 
diff --git a/erts/emulator/beam/io.c b/erts/emulator/beam/io.c
index 0f86d8e41d..ae053fc191 100644
--- a/erts/emulator/beam/io.c
+++ b/erts/emulator/beam/io.c
@@ -1218,9 +1218,10 @@ typedef struct {
 static ERTS_INLINE ErtsTryImmDrvCallResult
 try_imm_drv_call(ErtsTryImmDrvCallState *sp)
 {
+    unsigned int prof_runnable_ports;
     ErtsTryImmDrvCallResult res;
     int reds_left_in;
-    erts_aint32_t invalid_state, invalid_sched_flags;
+    erts_aint32_t act, exp, invalid_state, invalid_sched_flags;
     Port *prt = sp->port;
     Process *c_p = sp->c_p;
 
@@ -1247,18 +1248,39 @@ try_imm_drv_call(ErtsTryImmDrvCallState *sp)
 	goto locked_fail;
     }
 
-    sp->sched_flags = erts_smp_atomic32_read_nob(&prt->sched.flags);
-    if (sp->sched_flags & invalid_sched_flags) {
-	res = ERTS_TRY_IMM_DRV_CALL_INVALID_SCHED_FLAGS;
-	goto locked_fail;
-    }
+    prof_runnable_ports = erts_system_profile_flags.runnable_ports;
+    if (prof_runnable_ports)
+	erts_port_task_sched_lock(&prt->sched);
 
+    act = erts_smp_atomic32_read_nob(&prt->sched.flags);
+
+    do {
+	erts_aint32_t new;
+	
+	if (act & invalid_sched_flags) {
+	    res = ERTS_TRY_IMM_DRV_CALL_INVALID_SCHED_FLAGS;
+	    sp->sched_flags = act;
+	    goto locked_fail;
+	}
+	exp = act;
+	new = act | ERTS_PTS_FLG_EXEC_IMM;
+	act = erts_smp_atomic32_cmpxchg_mb(&prt->sched.flags, new, exp);
+    } while (act != exp);
+    
+    sp->sched_flags = act;
 
     if (!c_p)
 	reds_left_in = CONTEXT_REDS/10;
     else {
 	if (IS_TRACED_FL(c_p, F_TRACE_SCHED_PROCS))
 	    trace_virtual_sched(c_p, am_out);
+	/*
+	 * No status lock held while sending runnable
+	 * proc trace messages. It is however not needed
+	 * in this case, since only this thread can send
+	 * such messages for this process until the process
+	 * has been scheduled out.
+	 */
 	if (erts_system_profile_flags.runnable_procs
 	    && erts_system_profile_flags.exclusive)
 	    profile_runnable_proc(c_p, am_inactive);
@@ -1273,11 +1295,14 @@ try_imm_drv_call(ErtsTryImmDrvCallState *sp)
 
     ERTS_SMP_CHK_NO_PROC_LOCKS;
 
-    if (IS_TRACED_FL(prt, F_TRACE_SCHED_PORTS))
-	trace_sched_ports_where(prt, am_in, sp->port_op);
-    if (erts_system_profile_flags.runnable_ports
-	&& !erts_port_is_scheduled(prt))
-    	profile_runnable_port(prt, am_active);
+    if (prof_runnable_ports | IS_TRACED_FL(prt, F_TRACE_SCHED_PORTS)) {
+	if (prof_runnable_ports && !(act & (ERTS_PTS_FLG_IN_RUNQ|ERTS_PTS_FLG_EXEC)))
+	    profile_runnable_port(prt, am_active);
+	if (IS_TRACED_FL(prt, F_TRACE_SCHED_PORTS))
+	    trace_sched_ports_where(prt, am_in, sp->port_op);
+	if (prof_runnable_ports)
+	    erts_port_task_sched_unlock(&prt->sched);
+    }
 
     sp->fpe_was_unmasked = erts_block_fpe();
 
@@ -1294,17 +1319,31 @@ finalize_imm_drv_call(ErtsTryImmDrvCallState *sp)
     int reds;
     Port *prt = sp->port;
     Process *c_p = sp->c_p;
+    erts_aint32_t act;
+    unsigned int prof_runnable_ports;
 
     reds = prt->reds;
     reds += erts_port_driver_callback_epilogue(prt, NULL);
 
     erts_unblock_fpe(sp->fpe_was_unmasked);
 
-    if (IS_TRACED_FL(prt, F_TRACE_SCHED_PORTS))
-	trace_sched_ports_where(prt, am_out, sp->port_op);
-    if (erts_system_profile_flags.runnable_ports
-	&& !erts_port_is_scheduled(prt))
-    	profile_runnable_port(prt, am_inactive);
+    prof_runnable_ports = erts_system_profile_flags.runnable_ports;
+    if (prof_runnable_ports)
+	erts_port_task_sched_lock(&prt->sched);
+
+    act = erts_smp_atomic32_read_band_mb(&prt->sched.flags,
+					 ~ERTS_PTS_FLG_EXEC_IMM);
+    ERTS_SMP_LC_ASSERT(act & ERTS_PTS_FLG_EXEC_IMM);
+
+    if (prof_runnable_ports | IS_TRACED_FL(prt, F_TRACE_SCHED_PORTS)) {
+	if (IS_TRACED_FL(prt, F_TRACE_SCHED_PORTS))
+	    trace_sched_ports_where(prt, am_out, sp->port_op);
+	if (prof_runnable_ports) {
+	    if (!(act & (ERTS_PTS_FLG_IN_RUNQ|ERTS_PTS_FLG_EXEC)))
+		profile_runnable_port(prt, am_inactive);
+	    erts_port_task_sched_unlock(&prt->sched);
+	}
+    }
 
     erts_port_release(prt);
 
@@ -1319,6 +1358,13 @@ finalize_imm_drv_call(ErtsTryImmDrvCallState *sp)
 
 	if (IS_TRACED_FL(c_p, F_TRACE_SCHED_PROCS))
 	    trace_virtual_sched(c_p, am_in);
+	/*
+	 * No status lock held while sending runnable
+	 * proc trace messages. It is however not needed
+	 * in this case, since only this thread can send
+	 * such messages for this process until the process
+	 * has been scheduled out.
+	 */
 	if (erts_system_profile_flags.runnable_procs
 	    && erts_system_profile_flags.exclusive)
 	    profile_runnable_proc(c_p, am_active);