Merge branch 'maint'

* maint:
  Fix faulty use of thread progress in handle_aux_work()

Conflicts:
	erts/emulator/beam/erl_process.c

Needed modifications:
       erts/emulator/beam/beam_bif_load.c

7 files changed, 71 insertions, 35 deletions

diff --git a/erts/emulator/beam/beam_bif_load.c b/erts/emulator/beam/beam_bif_load.c
index 4bf1c13421..9c65769b86 100644
--- a/erts/emulator/beam/beam_bif_load.c
+++ b/erts/emulator/beam/beam_bif_load.c
@@ -361,7 +361,7 @@ staging_epilogue(Process* c_p, int commit, Eterm res, int is_blocking,
 	 * without any memory barriers at all. 
 	 */
     
-	later = erts_thr_progress_later(); 
+	later = erts_thr_progress_later(c_p->scheduler_data);
 	erts_thr_progress_wakeup(c_p->scheduler_data, later);
 	erts_notify_code_ix_activation(c_p, later);
 	erts_suspend(c_p, ERTS_PROC_LOCK_MAIN, NULL);
diff --git a/erts/emulator/beam/erl_alloc_util.c b/erts/emulator/beam/erl_alloc_util.c
index 62225d3572..9a011e2adc 100644
--- a/erts/emulator/beam/erl_alloc_util.c
+++ b/erts/emulator/beam/erl_alloc_util.c
@@ -952,7 +952,7 @@ ddq_check_incoming(ErtsAllctrDDQueue_t *ddq)
 		ERTS_THR_MEMORY_BARRIER;
 	    else {
 		ddq->head.next.unref_end = (ErtsAllctrDDBlock_t *) ilast;
-		ddq->head.next.thr_progress = erts_thr_progress_later();
+		ddq->head.next.thr_progress = erts_thr_progress_later(NULL);
 		erts_atomic32_set_relb(&ddq->tail.data.um_refc_ix,
 				       um_refc_ix);
 		ddq->head.next.um_refc_ix = um_refc_ix == 0 ? 1 : 0;
diff --git a/erts/emulator/beam/erl_process.c b/erts/emulator/beam/erl_process.c
index bca8fac259..3d978f4994 100644
--- a/erts/emulator/beam/erl_process.c
+++ b/erts/emulator/beam/erl_process.c
@@ -1145,13 +1145,13 @@ unset_aux_work_flags(ErtsSchedulerSleepInfo *ssi, erts_aint32_t flgs)
 #ifdef ERTS_SMP
 
 static ERTS_INLINE void
-thr_prgr_current_reset(ErtsAuxWorkData *awdp)
+haw_thr_prgr_current_reset(ErtsAuxWorkData *awdp)
 {
     awdp->current_thr_prgr = ERTS_THR_PRGR_INVALID;
 }
 
 static ERTS_INLINE ErtsThrPrgrVal
-thr_prgr_current(ErtsAuxWorkData *awdp)
+haw_thr_prgr_current(ErtsAuxWorkData *awdp)
 {
     ErtsThrPrgrVal current = awdp->current_thr_prgr;
     if (current == ERTS_THR_PRGR_INVALID) {
@@ -1161,6 +1161,21 @@ thr_prgr_current(ErtsAuxWorkData *awdp)
     return current;
 }
 
+static ERTS_INLINE void
+haw_thr_prgr_current_check_progress(ErtsAuxWorkData *awdp)
+{
+    ErtsThrPrgrVal current = awdp->current_thr_prgr;
+    if (current != ERTS_THR_PRGR_INVALID
+	&& !erts_thr_progress_equal(current, erts_thr_progress_current())) {
+	/*
+	 * We have used a previouly read current value that isn't the
+	 * latest; need to poke ourselfs in order to guarantee no loss
+	 * of wakeups.
+	 */
+	erts_sched_poke(awdp->ssi);
+    }
+}
+
 #endif
 
 typedef struct erts_misc_aux_work_t_ erts_misc_aux_work_t;
@@ -1259,7 +1274,7 @@ static ERTS_INLINE erts_aint32_t
 handle_misc_aux_work_thr_prgr(ErtsAuxWorkData *awdp,
 			      erts_aint32_t aux_work)
 {
-    if (!erts_thr_progress_has_reached_this(thr_prgr_current(awdp),
+    if (!erts_thr_progress_has_reached_this(haw_thr_prgr_current(awdp),
 					    awdp->misc.thr_prgr))
 	return aux_work & ~ERTS_SSI_AUX_WORK_MISC_THR_PRGR;
 
@@ -1364,7 +1379,7 @@ handle_async_ready_clean(ErtsAuxWorkData *awdp,
 
 #ifdef ERTS_SMP
     if (awdp->async_ready.need_thr_prgr
-	&& !erts_thr_progress_has_reached_this(thr_prgr_current(awdp),
+	&& !erts_thr_progress_has_reached_this(haw_thr_prgr_current(awdp),
 					       awdp->async_ready.thr_prgr)) {
 	return aux_work & ~ERTS_SSI_AUX_WORK_ASYNC_READY_CLEAN;
     }
@@ -1483,7 +1498,7 @@ handle_delayed_dealloc(ErtsAuxWorkData *awdp, erts_aint32_t aux_work)
 
     if (need_thr_progress) {
 	if (wakeup == ERTS_THR_PRGR_INVALID)
-	    wakeup = erts_thr_progress_later_than(thr_prgr_current(awdp));
+	    wakeup = erts_thr_progress_later(awdp->esdp);
 	awdp->dd.thr_prgr = wakeup;
 	set_aux_work_flags(ssi, ERTS_SSI_AUX_WORK_DD_THR_PRGR);
 	awdp->dd.thr_prgr = wakeup;
@@ -1504,7 +1519,7 @@ handle_delayed_dealloc_thr_prgr(ErtsAuxWorkData *awdp, erts_aint32_t aux_work)
     int need_thr_progress;
     int more_work;
     ErtsThrPrgrVal wakeup = ERTS_THR_PRGR_INVALID;
-    ErtsThrPrgrVal current = thr_prgr_current(awdp);
+    ErtsThrPrgrVal current = haw_thr_prgr_current(awdp);
 
     if (!erts_thr_progress_has_reached_this(current, awdp->dd.thr_prgr))
 	return aux_work & ~ERTS_SSI_AUX_WORK_DD_THR_PRGR;
@@ -1526,7 +1541,7 @@ handle_delayed_dealloc_thr_prgr(ErtsAuxWorkData *awdp, erts_aint32_t aux_work)
 
     if (need_thr_progress) {
 	if (wakeup == ERTS_THR_PRGR_INVALID)
-	    wakeup = erts_thr_progress_later_than(current);
+	    wakeup = erts_thr_progress_later(awdp->esdp);
 	awdp->dd.thr_prgr = wakeup;
 	erts_thr_progress_wakeup(awdp->esdp, wakeup);
     }
@@ -1715,7 +1730,7 @@ handle_setup_aux_work_timer(ErtsAuxWorkData *awdp, erts_aint32_t aux_work)
 }
 
 static erts_aint32_t
-handle_aux_work(ErtsAuxWorkData *awdp, erts_aint32_t orig_aux_work)
+handle_aux_work(ErtsAuxWorkData *awdp, erts_aint32_t orig_aux_work, int waiting)
 {
 #undef HANDLE_AUX_WORK
 #define HANDLE_AUX_WORK(FLG, HNDLR) \
@@ -1733,7 +1748,7 @@ handle_aux_work(ErtsAuxWorkData *awdp, erts_aint32_t orig_aux_work)
     erts_aint32_t ignore = 0;
 
 #ifdef ERTS_SMP
-    thr_prgr_current_reset(awdp);
+    haw_thr_prgr_current_reset(awdp);
 #endif
 
     ERTS_DBG_CHK_AUX_WORK_VAL(aux_work);
@@ -1804,6 +1819,11 @@ handle_aux_work(ErtsAuxWorkData *awdp, erts_aint32_t orig_aux_work)
 
     ERTS_DBG_CHK_AUX_WORK_VAL(aux_work);
 
+#ifdef ERTS_SMP
+    if (waiting && !aux_work)
+	haw_thr_prgr_current_check_progress(awdp);
+#endif
+
     return aux_work;
 
 #undef HANDLE_AUX_WORK
@@ -2262,7 +2282,7 @@ aux_thread(void *unused)
 	if (aux_work) {
 	    if (!thr_prgr_active)
 		erts_thr_progress_active(NULL, thr_prgr_active = 1);
-	    aux_work = handle_aux_work(awdp, aux_work);
+	    aux_work = handle_aux_work(awdp, aux_work, 1);
 	    if (aux_work && erts_thr_progress_update(NULL))
 		erts_thr_progress_leader_update(NULL);
 	}
@@ -2341,7 +2361,7 @@ scheduler_wait(int *fcalls, ErtsSchedulerData *esdp, ErtsRunQueue *rq)
 		    erts_thr_progress_active(esdp, thr_prgr_active = 1);
 		    sched_wall_time_change(esdp, 1);
 		}
-		aux_work = handle_aux_work(&esdp->aux_work_data, aux_work);
+		aux_work = handle_aux_work(&esdp->aux_work_data, aux_work, 1);
 		if (aux_work && erts_thr_progress_update(esdp))
 		    erts_thr_progress_leader_update(esdp);
 	    }
@@ -2445,7 +2465,7 @@ scheduler_wait(int *fcalls, ErtsSchedulerData *esdp, ErtsRunQueue *rq)
 		if (!thr_prgr_active)
 		    erts_thr_progress_active(esdp, thr_prgr_active = 1);
 #endif
-		aux_work = handle_aux_work(&esdp->aux_work_data, aux_work);
+		aux_work = handle_aux_work(&esdp->aux_work_data, aux_work, 1);
 #ifdef ERTS_SMP
 		if (aux_work && erts_thr_progress_update(esdp))
 		    erts_thr_progress_leader_update(esdp);
@@ -3570,7 +3590,7 @@ retire_mpaths(ErtsMigrationPaths *mps)
     if (!mpaths.retired.first)
 	mpaths.retired.last = NULL;
 
-    mps->thr_prgr = erts_thr_progress_later_than(current);
+    mps->thr_prgr = erts_thr_progress_later(NULL);
     mps->next = NULL;
 
     if (mpaths.retired.last)
@@ -5167,7 +5187,9 @@ suspend_scheduler(ErtsSchedulerData *esdp)
 			sched_wall_time_change(esdp, 1);
 		    }
 		    if (aux_work)
-			aux_work = handle_aux_work(&esdp->aux_work_data, aux_work);
+			aux_work = handle_aux_work(&esdp->aux_work_data,
+						   aux_work,
+						   1);
 		    if (aux_work && erts_thr_progress_update(esdp))
 			erts_thr_progress_leader_update(esdp);
 		    if (qmask) {
@@ -6839,7 +6861,7 @@ Process *schedule(Process *p, int calls)
 		if (leader_update)
 		    erts_thr_progress_leader_update(esdp);
 		if (aux_work)
-		    handle_aux_work(&esdp->aux_work_data, aux_work);
+		    handle_aux_work(&esdp->aux_work_data, aux_work, 0);
 		erts_smp_runq_lock(rq);
 	    }
 	}
@@ -6852,7 +6874,7 @@ Process *schedule(Process *p, int calls)
 	    erts_aint32_t aux_work;
 	    aux_work = erts_atomic32_read_acqb(&esdp->ssi->aux_work);
 	    if (aux_work)
-		handle_aux_work(&esdp->aux_work_data, aux_work);
+		handle_aux_work(&esdp->aux_work_data, aux_work, 0);
 	}
 #endif /* ERTS_SMP */
 
diff --git a/erts/emulator/beam/erl_sched_spec_pre_alloc.c b/erts/emulator/beam/erl_sched_spec_pre_alloc.c
index bff9d246a3..37b186abd9 100644
--- a/erts/emulator/beam/erl_sched_spec_pre_alloc.c
+++ b/erts/emulator/beam/erl_sched_spec_pre_alloc.c
@@ -227,7 +227,7 @@ fetch_remote(erts_sspa_chunk_header_t *chdr, int max)
 		    ERTS_THR_MEMORY_BARRIER;
 		else {
 		    chdr->head.next.unref_end = (erts_sspa_blk_t *) ilast;
-		    chdr->head.next.thr_progress = erts_thr_progress_later();
+		    chdr->head.next.thr_progress = erts_thr_progress_later(NULL);
 		    erts_atomic32_set_relb(&chdr->tail.data.um_refc_ix,
 					   um_refc_ix);
 		    chdr->head.next.um_refc_ix = um_refc_ix == 0 ? 1 : 0;
diff --git a/erts/emulator/beam/erl_thr_progress.c b/erts/emulator/beam/erl_thr_progress.c
index 9ef83746c5..88524bdd4c 100644
--- a/erts/emulator/beam/erl_thr_progress.c
+++ b/erts/emulator/beam/erl_thr_progress.c
@@ -891,16 +891,16 @@ has_reached_wakeup(ErtsThrPrgrVal wakeup)
 	ErtsThrPrgrVal limit;
 	/*
 	 * erts_thr_progress_later() returns values which are
-	 * equal to 'current + 2'. That is, users should never
-	 * get a hold of values larger than that.
+	 * equal to 'current + 2', or 'current + 3'. That is, users
+	 * should never get a hold of values larger than that.
 	 *
-	 * That is, valid values are values less than 'current + 3'.
+	 * That is, valid values are values less than 'current + 4'.
 	 *
 	 * Values larger than this won't work with the wakeup
 	 * algorithm.
 	 */
 
-	limit = current + 3;
+	limit = current + 4;
 	if (limit == ERTS_THR_PRGR_VAL_WAITING)
 	    limit = 0;
 	else if (limit < current) /* Wrapped */
diff --git a/erts/emulator/beam/erl_thr_progress.h b/erts/emulator/beam/erl_thr_progress.h
index a71724b813..89486b065b 100644
--- a/erts/emulator/beam/erl_thr_progress.h
+++ b/erts/emulator/beam/erl_thr_progress.h
@@ -139,11 +139,12 @@ ERTS_GLB_INLINE ErtsThrPrgrVal erts_thr_prgr_read_mb__(ERTS_THR_PRGR_ATOMIC *atm
 
 ERTS_GLB_INLINE int erts_thr_progress_is_managed_thread(void);
 ERTS_GLB_INLINE ErtsThrPrgrVal erts_thr_progress_current_to_later__(ErtsThrPrgrVal val);
-ERTS_GLB_INLINE ErtsThrPrgrVal erts_thr_progress_later_than(ErtsThrPrgrVal val);
-ERTS_GLB_INLINE ErtsThrPrgrVal erts_thr_progress_later(void);
+ERTS_GLB_INLINE ErtsThrPrgrVal erts_thr_progress_later(ErtsSchedulerData *);
 ERTS_GLB_INLINE ErtsThrPrgrVal erts_thr_progress_current(void);
 ERTS_GLB_INLINE int erts_thr_progress_has_passed__(ErtsThrPrgrVal val1, ErtsThrPrgrVal val2);
 ERTS_GLB_INLINE int erts_thr_progress_has_reached_this(ErtsThrPrgrVal this, ErtsThrPrgrVal val);
+ERTS_GLB_INLINE int erts_thr_progress_equal(ErtsThrPrgrVal val1,
+					    ErtsThrPrgrVal val2);
 ERTS_GLB_INLINE int erts_thr_progress_cmp(ErtsThrPrgrVal val1, ErtsThrPrgrVal val2);
 ERTS_GLB_INLINE int erts_thr_progress_has_reached(ErtsThrPrgrVal val);
 
@@ -230,16 +231,23 @@ erts_thr_progress_current_to_later__(ErtsThrPrgrVal val)
 }
 
 ERTS_GLB_INLINE ErtsThrPrgrVal
-erts_thr_progress_later_than(ErtsThrPrgrVal val)
+erts_thr_progress_later(ErtsSchedulerData *esdp)
 {
-    ERTS_THR_MEMORY_BARRIER;
-    return erts_thr_progress_current_to_later__(val);
-}
-
-ERTS_GLB_INLINE ErtsThrPrgrVal
-erts_thr_progress_later(void)
-{
-    ErtsThrPrgrVal val = erts_thr_prgr_read_mb__(&erts_thr_prgr__.current);
+    ErtsThrPrgrData *tpd;
+    ErtsThrPrgrVal val;
+    if (esdp) {
+	tpd = &esdp->thr_progress_data;
+    managed_thread:
+	val = tpd->previous.local;
+	ERTS_THR_MEMORY_BARRIER;
+    }
+    else {
+	tpd = erts_tsd_get(erts_thr_prgr_data_key__);
+	if (tpd && tpd->is_managed)
+	    goto managed_thread;
+	val = erts_thr_prgr_read_mb__(&erts_thr_prgr__.current);
+    }
+    ASSERT(val != ERTS_THR_PRGR_VAL_WAITING);
     return erts_thr_progress_current_to_later__(val);
 }
 
@@ -279,6 +287,12 @@ erts_thr_progress_has_reached_this(ErtsThrPrgrVal this, ErtsThrPrgrVal val)
 }
 
 ERTS_GLB_INLINE int
+erts_thr_progress_equal(ErtsThrPrgrVal val1, ErtsThrPrgrVal val2)
+{
+    return val1 == val2 && val1 != ERTS_THR_PRGR_INVALID;
+}
+
+ERTS_GLB_INLINE int
 erts_thr_progress_cmp(ErtsThrPrgrVal val1, ErtsThrPrgrVal val2)
 {
     if (val1 == val2)
diff --git a/erts/emulator/beam/erl_thr_queue.c b/erts/emulator/beam/erl_thr_queue.c
index 70949ece76..f07964a265 100644
--- a/erts/emulator/beam/erl_thr_queue.c
+++ b/erts/emulator/beam/erl_thr_queue.c
@@ -422,7 +422,7 @@ clean(ErtsThrQ_t *q, int max_ops, int do_notify)
 	    else {
 		q->head.next.unref_end = (ErtsThrQElement_t *) ilast;
 #ifdef ERTS_SMP
-		q->head.next.thr_progress = erts_thr_progress_later();
+		q->head.next.thr_progress = erts_thr_progress_later(NULL);
 #endif
 		erts_atomic32_set_relb(&q->tail.data.um_refc_ix,
 				       um_refc_ix);