Merge branch 'maint'

* maint:
  Light weight statistics of run queue lengths

Conflicts:
	erts/preloaded/ebin/erlang.beam

4 files changed, 114 insertions, 40 deletions

diff --git a/erts/emulator/beam/atom.names b/erts/emulator/beam/atom.names
index 13c2a0f8f9..ae7f405cce 100644
--- a/erts/emulator/beam/atom.names
+++ b/erts/emulator/beam/atom.names
@@ -71,6 +71,7 @@ atom absoluteURI
 atom ac
 atom accessor
 atom active
+atom active_tasks
 atom all
 atom all_but_first
 atom all_names
@@ -509,6 +510,7 @@ atom return_from
 atom return_to
 atom return_trace
 atom run_queue
+atom run_queue_lengths
 atom runnable
 atom runnable_ports
 atom runnable_procs
@@ -576,7 +578,9 @@ atom timeout_value
 atom Times='*'
 atom timestamp
 atom total
+atom total_active_tasks
 atom total_heap_size
+atom total_run_queue_lengths
 atom tpkt
 atom trace trace_ts traced 
 atom trace_control_word
diff --git a/erts/emulator/beam/erl_bif_info.c b/erts/emulator/beam/erl_bif_info.c
index f952f937ce..ab3ebd6bac 100644
--- a/erts/emulator/beam/erl_bif_info.c
+++ b/erts/emulator/beam/erl_bif_info.c
@@ -3199,6 +3199,39 @@ BIF_RETTYPE statistics_1(BIF_ALIST_1)
 	if (is_non_value(res))
 	    BIF_RET(am_undefined);
 	BIF_TRAP1(gather_sched_wall_time_res_trap, BIF_P, res);
+    } else if (BIF_ARG_1 == am_total_active_tasks
+	       || BIF_ARG_1 == am_total_run_queue_lengths) {
+	Uint no = erts_run_queues_len(NULL, 0, BIF_ARG_1 == am_total_active_tasks);
+	if (IS_USMALL(0, no))
+	    res = make_small(no);
+	else {
+	    Eterm *hp = HAlloc(BIF_P, BIG_UINT_HEAP_SIZE);
+	    res = uint_to_big(no, hp);
+	}
+	BIF_RET(res);
+    } else if (BIF_ARG_1 == am_active_tasks
+	       || BIF_ARG_1 == am_run_queue_lengths) {
+	Eterm res, *hp, **hpp;
+	Uint sz, *szp;
+	int no_qs = erts_no_run_queues;
+	Uint *qszs = erts_alloc(ERTS_ALC_T_TMP,sizeof(Uint)*no_qs*2);
+	(void) erts_run_queues_len(qszs, 0, BIF_ARG_1 == am_active_tasks);
+	sz = 0;
+	szp = &sz;
+	hpp = NULL;
+	while (1) {
+	    int i;
+	    for (i = 0; i < no_qs; i++)
+		qszs[no_qs+i] = erts_bld_uint(hpp, szp, qszs[i]);
+	    res = erts_bld_list(hpp, szp, no_qs, &qszs[no_qs]);
+	    if (hpp) {
+		erts_free(ERTS_ALC_T_TMP, qszs);
+		BIF_RET(res);
+	    }
+	    hp = HAlloc(BIF_P, sz);
+	    szp = NULL;
+	    hpp = &hp;
+	}
     } else if (BIF_ARG_1 == am_context_switches) {
 	Eterm cs = erts_make_integer(erts_get_total_context_switches(), BIF_P);
 	hp = HAlloc(BIF_P, 3);
@@ -3247,7 +3280,7 @@ BIF_RETTYPE statistics_1(BIF_ALIST_1)
 	res = TUPLE2(hp, b1, b2);
 	BIF_RET(res);
     } else if (BIF_ARG_1 ==  am_run_queue) {
-	res = erts_run_queues_len(NULL);
+	res = erts_run_queues_len(NULL, 1, 0);
 	BIF_RET(make_small(res));
     } else if (BIF_ARG_1 == am_wall_clock) {
 	UWord w1, w2;
@@ -3267,7 +3300,7 @@ BIF_RETTYPE statistics_1(BIF_ALIST_1)
 	Uint sz, *szp;
 	int no_qs = erts_no_run_queues;
 	Uint *qszs = erts_alloc(ERTS_ALC_T_TMP,sizeof(Uint)*no_qs*2);
-	(void) erts_run_queues_len(qszs);
+	(void) erts_run_queues_len(qszs, 0, 0);
 	sz = 0;
 	szp = &sz;
 	hpp = NULL;
diff --git a/erts/emulator/beam/erl_process.c b/erts/emulator/beam/erl_process.c
index 2c62685f8c..9495436ba6 100644
--- a/erts/emulator/beam/erl_process.c
+++ b/erts/emulator/beam/erl_process.c
@@ -3060,7 +3060,7 @@ scheduler_wait(int *fcalls, ErtsSchedulerData *esdp, ErtsRunQueue *rq)
 	    }
 
 #ifndef ERTS_SMP
-	    if (rq->len != 0 || rq->misc.start)
+	    if (erts_smp_atomic32_read_dirty(&rq->len) != 0 || rq->misc.start)
 		goto sys_woken;
 #else
 	    flgs = erts_smp_atomic32_read_acqb(&ssi->flags);
@@ -3159,7 +3159,7 @@ scheduler_wait(int *fcalls, ErtsSchedulerData *esdp, ErtsRunQueue *rq)
 	}
 
 #ifndef ERTS_SMP
-	if (rq->len == 0 && !rq->misc.start)
+	if (erts_smp_atomic32_read_dirty(&rq->len) == 0 && !rq->misc.start)
 	    goto sys_aux_work;
     sys_woken:
 #else
@@ -4876,7 +4876,7 @@ erts_fprintf(stderr, "--------------------------------\n");
 	rq->out_of_work_count = 0;
 	(void) ERTS_RUNQ_FLGS_READ_BSET(rq, ERTS_RUNQ_FLGS_MIGRATION_INFO, flags);
 
-	rq->max_len = rq->len;
+	rq->max_len = erts_smp_atomic32_read_dirty(&rq->len);
 	for (pix = 0; pix < ERTS_NO_PRIO_LEVELS; pix++) {
 	    ErtsRunQueueInfo *rqi;
 	    rqi = (pix == ERTS_PORT_PRIO_LEVEL
@@ -5034,7 +5034,7 @@ wakeup_other_check(ErtsRunQueue *rq, Uint32 flags)
 {
     int wo_reds = rq->wakeup_other_reds;
     if (wo_reds) {
-	int left_len = rq->len - 1;
+	int left_len = erts_smp_atomic32_read_dirty(&rq->len) - 1;
 	if (left_len < 1) {
 	    int wo_reduce = wo_reds << wakeup_other.dec_shift;
 	    wo_reduce &= wakeup_other.dec_mask;
@@ -5107,7 +5107,7 @@ wakeup_other_check_legacy(ErtsRunQueue *rq, Uint32 flags)
 {
     int wo_reds = rq->wakeup_other_reds;
     if (wo_reds) {
-	erts_aint32_t len = rq->len;
+	erts_aint32_t len = erts_smp_atomic32_read_dirty(&rq->len);
 	if (len < 2) {
 	    rq->wakeup_other -= ERTS_WAKEUP_OTHER_DEC_LEGACY*wo_reds;
 	    if (rq->wakeup_other < 0)
@@ -5203,7 +5203,7 @@ runq_supervisor(void *unused)
 	    ErtsRunQueue *rq = ERTS_RUNQ_IX(ix);
 	    if (ERTS_RUNQ_FLGS_GET(rq) & ERTS_RUNQ_FLG_NONEMPTY) {
 		erts_smp_runq_lock(rq);
-		if (rq->len != 0)
+		if (erts_smp_atomic32_read_dirty(&rq->len) != 0)
 		    wake_scheduler_on_empty_runq(rq); /* forced wakeup... */
 		erts_smp_runq_unlock(rq);
 	    }
@@ -5544,7 +5544,7 @@ erts_init_scheduling(int no_schedulers, int no_schedulers_online
 	}
 	rq->out_of_work_count = 0;
 	rq->max_len = 0;
-	rq->len = 0;
+	erts_smp_atomic32_set_nob(&rq->len, 0);
 	rq->wakeup_other = 0;
 	rq->wakeup_other_reds = 0;
 	rq->halt_in_progress = 0;
@@ -7913,6 +7913,9 @@ sched_thread_func(void *vesdp)
 
     erts_sched_init_time_sup(esdp);
 
+    (void) ERTS_RUNQ_FLGS_SET_NOB(esdp->run_queue,
+				  ERTS_RUNQ_FLG_EXEC);
+
 #ifdef ERTS_SMP
     tse = erts_tse_fetch();
     erts_tse_prepare_timed(tse);
@@ -8910,24 +8913,39 @@ resume_process_1(BIF_ALIST_1)
 }
 
 Uint
-erts_run_queues_len(Uint *qlen)
+erts_run_queues_len(Uint *qlen, int atomic_queues_read, int incl_active_sched)
 {
     int i = 0;
     Uint len = 0;
-    ERTS_ATOMIC_FOREACH_RUNQ(rq,
-    {
-	Sint pqlen = 0;
-	int pix;
-	for (pix = 0; pix < ERTS_NO_PROC_PRIO_LEVELS; pix++)
-	    pqlen += RUNQ_READ_LEN(&rq->procs.prio_info[pix].len);
+    if (atomic_queues_read)
+	ERTS_ATOMIC_FOREACH_RUNQ(rq,
+	 {
+	     Sint rq_len = (Sint) erts_smp_atomic32_read_dirty(&rq->len);
+	     ASSERT(rq_len >= 0);
+	     if (incl_active_sched
+		 && (ERTS_RUNQ_FLGS_GET_NOB(rq) & ERTS_RUNQ_FLG_EXEC)) {
+		 rq_len++;
+	     }
+	     if (qlen)
+		 qlen[i++] = rq_len;
+	     len += (Uint) rq_len;
+	 }
+	    );
+    else {
+	for (i = 0; i < erts_no_run_queues; i++) {
+	    ErtsRunQueue *rq = ERTS_RUNQ_IX(i);
+	    Sint rq_len = (Sint) erts_smp_atomic32_read_nob(&rq->len);
+	    ASSERT(rq_len >= 0);
+	     if (incl_active_sched
+		 && (ERTS_RUNQ_FLGS_GET_NOB(rq) & ERTS_RUNQ_FLG_EXEC)) {
+		 rq_len++;
+	     }
+	    if (qlen)
+		qlen[i] = rq_len;
+	    len += (Uint) rq_len;
+	}
 
-	if (pqlen < 0)
-	    pqlen = 0;
- 	if (qlen)
-	    qlen[i++] = pqlen;
-	len += pqlen;
     }
-	);
     return len;
 }
 
@@ -9356,8 +9374,10 @@ Process *schedule(Process *p, int calls)
 
 	if (flags & (ERTS_RUNQ_FLG_CHK_CPU_BIND|ERTS_RUNQ_FLG_SUSPENDED)) {
 	    if (flags & ERTS_RUNQ_FLG_SUSPENDED) {
+		(void) ERTS_RUNQ_FLGS_UNSET_NOB(rq, ERTS_RUNQ_FLG_EXEC);
 		suspend_scheduler(esdp);
-		flags = ERTS_RUNQ_FLGS_GET_NOB(rq);
+		flags = ERTS_RUNQ_FLGS_SET_NOB(rq, ERTS_RUNQ_FLG_EXEC);
+		flags |= ERTS_RUNQ_FLG_EXEC;
 	    }
 	    if (flags & ERTS_RUNQ_FLG_CHK_CPU_BIND) {
 		flags = ERTS_RUNQ_FLGS_UNSET(rq, ERTS_RUNQ_FLG_CHK_CPU_BIND);
@@ -9446,7 +9466,10 @@ Process *schedule(Process *p, int calls)
 	    }
 #endif
 
+	    (void) ERTS_RUNQ_FLGS_UNSET(rq, ERTS_RUNQ_FLG_EXEC);
 	    scheduler_wait(&fcalls, esdp, rq);
+	    flags = ERTS_RUNQ_FLGS_SET_NOB(rq, ERTS_RUNQ_FLG_EXEC);
+	    flags |= ERTS_RUNQ_FLG_EXEC;
 
 #ifdef ERTS_SMP
 	    non_empty_runq(rq);
diff --git a/erts/emulator/beam/erl_process.h b/erts/emulator/beam/erl_process.h
index 4bc879eacb..71818b2330 100644
--- a/erts/emulator/beam/erl_process.h
+++ b/erts/emulator/beam/erl_process.h
@@ -170,8 +170,10 @@ extern int erts_sched_thread_suggested_stack_size;
   (((Uint32) 1) << (ERTS_RUNQ_FLG_BASE2 + 5))
 #define ERTS_RUNQ_FLG_PROTECTED \
   (((Uint32) 1) << (ERTS_RUNQ_FLG_BASE2 + 6))
+#define ERTS_RUNQ_FLG_EXEC \
+  (((Uint32) 1) << (ERTS_RUNQ_FLG_BASE2 + 7))
 
-#define ERTS_RUNQ_FLG_MAX (ERTS_RUNQ_FLG_BASE2 + 7)
+#define ERTS_RUNQ_FLG_MAX (ERTS_RUNQ_FLG_BASE2 + 8)
 
 #define ERTS_RUNQ_FLGS_MIGRATION_QMASKS	\
   (ERTS_RUNQ_FLGS_EMIGRATE_QMASK	\
@@ -215,6 +217,9 @@ extern int erts_sched_thread_suggested_stack_size;
 #define ERTS_RUNQ_FLGS_SET(RQ, FLGS)					\
     ((Uint32) erts_smp_atomic32_read_bor_relb(&(RQ)->flags,		\
 					      (erts_aint32_t) (FLGS)))
+#define ERTS_RUNQ_FLGS_SET_NOB(RQ, FLGS)				\
+    ((Uint32) erts_smp_atomic32_read_bor_nob(&(RQ)->flags,		\
+					     (erts_aint32_t) (FLGS)))
 #define ERTS_RUNQ_FLGS_BSET(RQ, MSK, FLGS)				\
     ((Uint32) erts_smp_atomic32_read_bset_relb(&(RQ)->flags,		\
 					       (erts_aint32_t) (MSK),	\
@@ -222,6 +227,9 @@ extern int erts_sched_thread_suggested_stack_size;
 #define ERTS_RUNQ_FLGS_UNSET(RQ, FLGS)					\
     ((Uint32) erts_smp_atomic32_read_band_relb(&(RQ)->flags,		\
 					       (erts_aint32_t) ~(FLGS)))
+#define ERTS_RUNQ_FLGS_UNSET_NOB(RQ, FLGS)					\
+    ((Uint32) erts_smp_atomic32_read_band_nob(&(RQ)->flags,		\
+					      (erts_aint32_t) ~(FLGS)))
 #define ERTS_RUNQ_FLGS_GET(RQ)						\
     ((Uint32) erts_smp_atomic32_read_acqb(&(RQ)->flags))
 #define ERTS_RUNQ_FLGS_GET_NOB(RQ)					\
@@ -464,7 +472,7 @@ struct ErtsRunQueue_ {
     int full_reds_history[ERTS_FULL_REDS_HISTORY_SIZE];
     int out_of_work_count;
     erts_aint32_t max_len;
-    erts_aint32_t len;
+    erts_smp_atomic32_t len;
     int wakeup_other;
     int wakeup_other_reds;
     int halt_in_progress;
@@ -712,7 +720,19 @@ erts_smp_inc_runq_len(ErtsRunQueue *rq, ErtsRunQueueInfo *rqi, int prio)
 
     ERTS_SMP_LC_ASSERT(erts_smp_lc_runq_is_locked(rq));
 
-    len = erts_smp_atomic32_read_nob(&rqi->len);
+    len = erts_smp_atomic32_read_dirty(&rq->len);
+
+#ifdef ERTS_SMP
+    if (len == 0)
+	erts_non_empty_runq(rq);
+#endif
+    len++;
+    if (rq->max_len < len)
+	rq->max_len = len;
+    ASSERT(len > 0);
+    erts_smp_atomic32_set_nob(&rq->len, len);
+
+    len = erts_smp_atomic32_read_dirty(&rqi->len);
     ASSERT(len >= 0);
     if (len == 0) {
 	ASSERT((erts_smp_atomic32_read_nob(&rq->flags)
@@ -725,15 +745,6 @@ erts_smp_inc_runq_len(ErtsRunQueue *rq, ErtsRunQueueInfo *rqi, int prio)
 	rqi->max_len = len;
 
     erts_smp_atomic32_set_relb(&rqi->len, len);
-
-#ifdef ERTS_SMP
-    if (rq->len == 0)
-	erts_non_empty_runq(rq);
-#endif
-    rq->len++;
-    if (rq->max_len < rq->len)
-	rq->max_len = len;
-    ASSERT(rq->len > 0);
 }
 
 ERTS_GLB_INLINE void
@@ -743,7 +754,12 @@ erts_smp_dec_runq_len(ErtsRunQueue *rq, ErtsRunQueueInfo *rqi, int prio)
 
     ERTS_SMP_LC_ASSERT(erts_smp_lc_runq_is_locked(rq));
 
-    len = erts_smp_atomic32_read_nob(&rqi->len);
+    len = erts_smp_atomic32_read_dirty(&rq->len);
+    len--;
+    ASSERT(len >= 0);
+    erts_smp_atomic32_set_nob(&rq->len, len);
+
+    len = erts_smp_atomic32_read_dirty(&rqi->len);
     len--;
     ASSERT(len >= 0);
     if (len == 0) {
@@ -754,8 +770,6 @@ erts_smp_dec_runq_len(ErtsRunQueue *rq, ErtsRunQueueInfo *rqi, int prio)
     }
     erts_smp_atomic32_set_relb(&rqi->len, len);
 
-    rq->len--;
-    ASSERT(rq->len >= 0);
 }
 
 ERTS_GLB_INLINE void
@@ -765,7 +779,7 @@ erts_smp_reset_max_len(ErtsRunQueue *rq, ErtsRunQueueInfo *rqi)
 
     ERTS_SMP_LC_ASSERT(erts_smp_lc_runq_is_locked(rq));
 
-    len = erts_smp_atomic32_read_nob(&rqi->len);
+    len = erts_smp_atomic32_read_dirty(&rqi->len);
     ASSERT(rqi->max_len >= len);
     rqi->max_len = len;
 }
@@ -1684,7 +1698,7 @@ void erts_sched_notify_check_cpu_bind(void);
 Uint erts_active_schedulers(void);
 void erts_init_process(int, int, int);
 Eterm erts_process_status(Process *, ErtsProcLocks, Process *, Eterm);
-Uint erts_run_queues_len(Uint *);
+Uint erts_run_queues_len(Uint *, int, int);
 void erts_add_to_runq(Process *);
 Eterm erts_bound_schedulers_term(Process *c_p);
 Eterm erts_get_cpu_topology_term(Process *c_p, Eterm which);