Merge branch 'rickard/dealloc/OTP-10162' into maint

* rickard/dealloc/OTP-10162:
  Improve the enqueue operation of delayed dealloc
  Implement delayed aux work wake up

5 files changed, 269 insertions, 108 deletions

diff --git a/erts/emulator/beam/erl_alloc_util.c b/erts/emulator/beam/erl_alloc_util.c
index a0abd1c405..97ba306a79 100644
--- a/erts/emulator/beam/erl_alloc_util.c
+++ b/erts/emulator/beam/erl_alloc_util.c
@@ -829,46 +829,83 @@ init_dd_queue(ErtsAllctrDDQueue_t *ddq)
     ddq->head.used_marker = 1;
 }
 
-static ERTS_INLINE erts_aint_t
-ddq_managed_thread_enqueue(ErtsAllctrDDQueue_t *ddq, void *ptr)
+static ERTS_INLINE int
+ddq_managed_thread_enqueue(ErtsAllctrDDQueue_t *ddq, void *ptr, int cinit)
 {
-    erts_aint_t ilast, itmp;
-    ErtsAllctrDDBlock_t *this = ptr;
+    erts_aint_t itmp;
+    ErtsAllctrDDBlock_t *enq, *this = ptr;
 
     erts_atomic_init_nob(&this->atmc_next, ERTS_AINT_NULL);
-
     /* Enqueue at end of list... */
 
-    ilast = erts_atomic_read_nob(&ddq->tail.data.last);
-    while (1) {
-	ErtsAllctrDDBlock_t *last = (ErtsAllctrDDBlock_t *) ilast;
-	itmp = erts_atomic_cmpxchg_mb(&last->atmc_next,
-				       (erts_aint_t) this,
-				       ERTS_AINT_NULL);
-	if (itmp == ERTS_AINT_NULL)
-	    break;
-	ilast = itmp;
+    enq = (ErtsAllctrDDBlock_t *) erts_atomic_read_nob(&ddq->tail.data.last);
+    itmp = erts_atomic_cmpxchg_relb(&enq->atmc_next,
+				    (erts_aint_t) this,
+				    ERTS_AINT_NULL);
+    if (itmp == ERTS_AINT_NULL) {
+	/* We are required to move last pointer */
+#ifdef DEBUG
+	ASSERT(ERTS_AINT_NULL == erts_atomic_read_nob(&this->atmc_next));
+	ASSERT(((erts_aint_t) enq)
+	       == erts_atomic_xchg_relb(&ddq->tail.data.last,
+					(erts_aint_t) this));
+#else
+	erts_atomic_set_relb(&ddq->tail.data.last, (erts_aint_t) this);
+#endif
+	return 1;
     }
+    else {
+	/*
+	 * We *need* to insert element somewhere in between the
+	 * last element we read earlier and the actual last element.
+	 */
+	int i = cinit;
 
-    /* Move last pointer forward... */
-    while (1) {
-	if (erts_atomic_read_rb(&this->atmc_next) != ERTS_AINT_NULL) {
-	    /* Someone else will move it forward */
-	    return erts_atomic_read_rb(&ddq->tail.data.last);
+	while (1) {
+	    erts_aint_t itmp2;
+	    erts_atomic_set_nob(&this->atmc_next, itmp);
+	    itmp2 = erts_atomic_cmpxchg_relb(&enq->atmc_next,
+					     (erts_aint_t) this,
+					     itmp);
+	    if (itmp == itmp2)
+		return 0; /* inserted this */
+	    if ((i & 1) == 0)
+		itmp = itmp2;
+	    else {
+		enq = (ErtsAllctrDDBlock_t *) itmp2;
+		itmp = erts_atomic_read_acqb(&enq->atmc_next);
+		ASSERT(itmp != ERTS_AINT_NULL);
+	    }
+	    i++;
 	}
-	itmp = erts_atomic_cmpxchg_mb(&ddq->tail.data.last,
-				      (erts_aint_t) this,
-				      ilast);
-	if (ilast == itmp)
-	    return (erts_aint_t) this;
-	ilast = itmp;
     }
 }
 
+static ERTS_INLINE erts_aint_t
+check_insert_marker(ErtsAllctrDDQueue_t *ddq, erts_aint_t ilast)
+{
+    if (!ddq->head.used_marker
+	&& ddq->head.unref_end == (ErtsAllctrDDBlock_t *) ilast) {
+	erts_aint_t itmp;
+	ErtsAllctrDDBlock_t *last = (ErtsAllctrDDBlock_t *) ilast;
+
+	erts_atomic_init_nob(&ddq->tail.data.marker.atmc_next, ERTS_AINT_NULL);
+	itmp = erts_atomic_cmpxchg_relb(&last->atmc_next,
+					(erts_aint_t) &ddq->tail.data.marker,
+					ERTS_AINT_NULL);
+	if (itmp == ERTS_AINT_NULL) {
+	    ilast = (erts_aint_t) &ddq->tail.data.marker;
+	    ddq->head.used_marker = !0;
+	    erts_atomic_set_relb(&ddq->tail.data.last, ilast);
+	}
+    }
+    return ilast;
+}
+
 static ERTS_INLINE int
-ddq_enqueue(ErtsAlcType_t type, ErtsAllctrDDQueue_t *ddq, void *ptr)
+ddq_enqueue(ErtsAlcType_t type, ErtsAllctrDDQueue_t *ddq, void *ptr, int cinit)
 {
-    erts_aint_t ilast;
+    int last_elem;
     int um_refc_ix = 0;
     int managed_thread = erts_thr_progress_is_managed_thread();
     if (!managed_thread) {
@@ -884,11 +921,11 @@ ddq_enqueue(ErtsAlcType_t type, ErtsAllctrDDQueue_t *ddq, void *ptr)
 	}
     }
 
-    ilast = ddq_managed_thread_enqueue(ddq, ptr);
+    last_elem = ddq_managed_thread_enqueue(ddq, ptr, cinit);
 
     if (!managed_thread)
 	erts_atomic_dec_relb(&ddq->tail.data.um_refc[um_refc_ix]);
-    return ilast == (erts_aint_t) ptr;
+    return last_elem;
 }
 
 static ERTS_INLINE void *
@@ -934,20 +971,16 @@ ddq_check_incoming(ErtsAllctrDDQueue_t *ddq)
 	int um_refc_ix;
 	ddq->head.next.thr_progress_reached = 1;
 	um_refc_ix = ddq->head.next.um_refc_ix;
-	if (erts_atomic_read_acqb(&ddq->tail.data.um_refc[um_refc_ix]) == 0) {
+	if (erts_atomic_read_nob(&ddq->tail.data.um_refc[um_refc_ix]) == 0) {
 	    /* Move unreferenced end pointer forward... */
 
+	    ETHR_MEMBAR(ETHR_LoadLoad|ETHR_LoadStore);
+
 	    ddq->head.unref_end = ddq->head.next.unref_end;
 
-	    if (!ddq->head.used_marker
-		&& ddq->head.unref_end == (ErtsAllctrDDBlock_t *) ilast) {
-		ddq->head.used_marker = 1;
-		ilast = ddq_managed_thread_enqueue(ddq, &ddq->tail.data.marker);
-	    }
+	    ilast = check_insert_marker(ddq, ilast);
 
-	    if (ddq->head.unref_end == (ErtsAllctrDDBlock_t *) ilast)
-		ERTS_THR_MEMORY_BARRIER;
-	    else {
+	    if (ddq->head.unref_end != (ErtsAllctrDDBlock_t *) ilast) {
 		ddq->head.next.unref_end = (ErtsAllctrDDBlock_t *) ilast;
 		ddq->head.next.thr_progress = erts_thr_progress_later(NULL);
 		erts_atomic32_set_relb(&ddq->tail.data.um_refc_ix,
@@ -1092,12 +1125,15 @@ handle_delayed_dealloc(Allctr_t *allctr,
 }
 
 static ERTS_INLINE void
-enqueue_dealloc_other_instance(ErtsAlcType_t type, Allctr_t *allctr, void *ptr)
+enqueue_dealloc_other_instance(ErtsAlcType_t type,
+			       Allctr_t *allctr,
+			       void *ptr,
+			       int cinit)
 {
     if (allctr->fix)
 	((UWord *) ptr)[ERTS_ALCU_DD_FIX_TYPE_OFFS] = (UWord) type;
 
-    if (ddq_enqueue(type, &allctr->dd.q, ptr))
+    if (ddq_enqueue(type, &allctr->dd.q, ptr, cinit))
 	erts_alloc_notify_delayed_dealloc(allctr->ix);
 }
 
@@ -3613,7 +3649,11 @@ erts_alcu_free_thr_pref(ErtsAlcType_t type, void *extra, void *p)
 	get_pref_allctr(extra, &pref_allctr);
 	ptr = get_used_allctr(extra, p, &used_allctr, NULL);
 	if (pref_allctr != used_allctr)
-	    enqueue_dealloc_other_instance(type, used_allctr, ptr);
+	    enqueue_dealloc_other_instance(type,
+					   used_allctr,
+					   ptr,
+					   (used_allctr->dd.ix
+					    - pref_allctr->dd.ix));
 	else {
 	    if (used_allctr->thread_safe)
 		erts_mtx_lock(&used_allctr->mutex);
@@ -3988,7 +4028,11 @@ realloc_thr_pref(ErtsAlcType_t type, void *extra, void *p, Uint size,
 	    sys_memcpy(res, p, cpy_size);
 
 	    if (!force_move || used_allctr != pref_allctr)
-		enqueue_dealloc_other_instance(type, used_allctr, ptr);
+		enqueue_dealloc_other_instance(type,
+					       used_allctr,
+					       ptr,
+					       (used_allctr->dd.ix
+						- pref_allctr->dd.ix));
 	    else {
 		do_erts_alcu_free(type, used_allctr, ptr);
 		ASSERT(pref_allctr == used_allctr);
@@ -4179,6 +4223,7 @@ erts_alcu_start(Allctr_t *allctr, AllctrInit_t *init)
 
 	allctr->dd.use = 1;
 	init_dd_queue(&allctr->dd.q);
+	allctr->dd.ix = init->ix;
     }
     else
 #endif
diff --git a/erts/emulator/beam/erl_process.c b/erts/emulator/beam/erl_process.c
index d2fa111b80..0fa2def5af 100644
--- a/erts/emulator/beam/erl_process.c
+++ b/erts/emulator/beam/erl_process.c
@@ -359,6 +359,7 @@ dbg_chk_aux_work_val(erts_aint32_t value)
     valid |= ERTS_SSI_AUX_WORK_ASYNC_READY_CLEAN;
 #endif
 #ifdef ERTS_SMP
+    valid |= ERTS_SSI_AUX_WORK_DELAYED_AW_WAKEUP;
     valid |= ERTS_SSI_AUX_WORK_MISC_THR_PRGR;
     valid |= ERTS_SSI_AUX_WORK_DD;
     valid |= ERTS_SSI_AUX_WORK_DD_THR_PRGR;
@@ -930,6 +931,45 @@ haw_thr_prgr_current_check_progress(ErtsAuxWorkData *awdp)
     }
 }
 
+static ERTS_INLINE erts_aint32_t
+handle_delayed_aux_work_wakeup(ErtsAuxWorkData *awdp, erts_aint32_t aux_work)
+{
+    int jix, max_jix;
+    unset_aux_work_flags(awdp->ssi, ERTS_SSI_AUX_WORK_DELAYED_AW_WAKEUP);
+
+    ERTS_THR_MEMORY_BARRIER;
+
+    max_jix = awdp->delayed_wakeup.jix;
+    awdp->delayed_wakeup.jix = -1;
+    for (jix = 0; jix <= max_jix; jix++) {
+	int sched = awdp->delayed_wakeup.job[jix].sched;
+	erts_aint32_t aux_work = awdp->delayed_wakeup.job[jix].aux_work;
+
+	ASSERT(awdp->delayed_wakeup.sched2jix[sched] == jix);
+	awdp->delayed_wakeup.sched2jix[sched] = -1;
+	set_aux_work_flags_wakeup_nob(ERTS_SCHED_SLEEP_INFO_IX(sched-1),
+				      aux_work);
+    }
+    return aux_work & ~ERTS_SSI_AUX_WORK_DELAYED_AW_WAKEUP;
+}
+
+static ERTS_INLINE void
+schedule_aux_work_wakeup(ErtsAuxWorkData *awdp, int sched, erts_aint32_t aux_work)
+{
+    int jix = awdp->delayed_wakeup.sched2jix[sched];
+    if (jix >= 0) {
+	ASSERT(awdp->delayed_wakeup.job[jix].sched == sched);
+	awdp->delayed_wakeup.job[jix].aux_work |= aux_work;
+    }
+    else {
+	jix = ++awdp->delayed_wakeup.jix;
+	awdp->delayed_wakeup.sched2jix[sched] = jix;
+	awdp->delayed_wakeup.job[jix].sched = sched;
+	awdp->delayed_wakeup.job[jix].aux_work = aux_work;
+    }
+    set_aux_work_flags_wakeup_nob(awdp->ssi, ERTS_SSI_AUX_WORK_DELAYED_AW_WAKEUP);
+}
+
 #endif
 
 typedef struct erts_misc_aux_work_t_ erts_misc_aux_work_t;
@@ -1186,8 +1226,14 @@ handle_fix_alloc(ErtsAuxWorkData *awdp, erts_aint32_t aux_work)
 void
 erts_alloc_notify_delayed_dealloc(int ix)
 {
-    set_aux_work_flags_wakeup_nob(ERTS_SCHED_SLEEP_INFO_IX(ix-1),
-				  ERTS_SSI_AUX_WORK_DD);
+    ErtsSchedulerData *esdp = erts_get_scheduler_data();
+    if (esdp)
+	schedule_aux_work_wakeup(&esdp->aux_work_data,
+				 ix,
+				 ERTS_SSI_AUX_WORK_DD);
+    else
+	set_aux_work_flags_wakeup_relb(ERTS_SCHED_SLEEP_INFO_IX(ix-1),
+				       ERTS_SSI_AUX_WORK_DD);
 }
 
 static ERTS_INLINE erts_aint32_t
@@ -1485,6 +1531,8 @@ handle_aux_work(ErtsAuxWorkData *awdp, erts_aint32_t orig_aux_work, int waiting)
      * eachother. Most frequent first.
      */
 #ifdef ERTS_SMP
+    HANDLE_AUX_WORK(ERTS_SSI_AUX_WORK_DELAYED_AW_WAKEUP,
+		    handle_delayed_aux_work_wakeup);
     HANDLE_AUX_WORK(ERTS_SSI_AUX_WORK_DD,
 		    handle_delayed_dealloc);
     /* DD must be before DD_THR_PRGR */
@@ -1963,7 +2011,7 @@ thr_prgr_fin_wait(void *vssi)
 				      | ERTS_SSI_FLG_TSE_SLEEPING));
 }
 
-static void init_aux_work_data(ErtsAuxWorkData *awdp, ErtsSchedulerData *esdp);
+static void init_aux_work_data(ErtsAuxWorkData *awdp, ErtsSchedulerData *esdp, char *dawwp);
 
 static void *
 aux_thread(void *unused)
@@ -1983,7 +2031,7 @@ aux_thread(void *unused)
     callbacks.finalize_wait = thr_prgr_fin_wait;
 
     erts_thr_progress_register_managed_thread(NULL, &callbacks, 1);
-    init_aux_work_data(awdp, NULL);
+    init_aux_work_data(awdp, NULL, NULL);
     awdp->ssi = ssi;
 
     sched_prep_spin_wait(ssi);
@@ -3850,7 +3898,7 @@ erts_sched_set_busy_wait_threshold(char *str)
     return 0;
 }
 static void
-init_aux_work_data(ErtsAuxWorkData *awdp, ErtsSchedulerData *esdp)
+init_aux_work_data(ErtsAuxWorkData *awdp, ErtsSchedulerData *esdp, char *dawwp)
 {
     awdp->sched_id = esdp ? (int) esdp->no : 0;
     awdp->esdp = esdp;
@@ -3868,12 +3916,32 @@ init_aux_work_data(ErtsAuxWorkData *awdp, ErtsSchedulerData *esdp)
 #endif
     awdp->async_ready.queue = NULL;
 #endif
+#ifdef ERTS_SMP
+    if (!dawwp) {
+	awdp->delayed_wakeup.job = NULL;
+	awdp->delayed_wakeup.sched2jix = NULL;
+	awdp->delayed_wakeup.jix = -1;
+    }
+    else {
+	int i;
+	awdp->delayed_wakeup.job = (ErtsDelayedAuxWorkWakeupJob *) dawwp;
+	dawwp += sizeof(ErtsDelayedAuxWorkWakeupJob)*(erts_no_schedulers+1);
+	awdp->delayed_wakeup.sched2jix = (int *) dawwp;
+	awdp->delayed_wakeup.jix = -1;
+	for (i = 0; i <= erts_no_schedulers; i++)
+	    awdp->delayed_wakeup.sched2jix[i] = -1;
+    }
+#endif
 }
 
 void
 erts_init_scheduling(int no_schedulers, int no_schedulers_online)
 {
     int ix, n, no_ssi;
+    char *daww_ptr;
+#ifdef ERTS_SMP
+    size_t daww_sz;
+#endif
 
     init_misc_op_list_alloc();
 
@@ -4006,6 +4074,15 @@ erts_init_scheduling(int no_schedulers, int no_schedulers_online)
 
     /* Create and initialize scheduler specific data */
 
+#ifdef ERTS_SMP
+    daww_sz = ERTS_ALC_CACHE_LINE_ALIGN_SIZE((sizeof(ErtsDelayedAuxWorkWakeupJob)
+					      + sizeof(int))*(n+1));
+    daww_ptr = erts_alloc_permanent_cache_aligned(ERTS_ALC_T_SCHDLR_DATA,
+						  daww_sz*n);
+#else
+    daww_ptr = NULL;
+#endif
+
     erts_aligned_scheduler_data = 
 	erts_alloc_permanent_cache_aligned(ERTS_ALC_T_SCHDLR_DATA,
 					   n*sizeof(ErtsAlignedSchedulerData));					   
@@ -4040,7 +4117,10 @@ erts_init_scheduling(int no_schedulers, int no_schedulers_online)
 	esdp->run_queue = ERTS_RUNQ_IX(ix);
 	esdp->run_queue->scheduler = esdp;
 
-	init_aux_work_data(&esdp->aux_work_data, esdp);
+	init_aux_work_data(&esdp->aux_work_data, esdp, daww_ptr);
+#ifdef ERTS_SMP
+	daww_ptr += daww_sz;
+#endif
 	init_sched_wall_time(&esdp->sched_wall_time);
     }
 
diff --git a/erts/emulator/beam/erl_process.h b/erts/emulator/beam/erl_process.h
index 5b79c40d93..9e7a5a5c74 100644
--- a/erts/emulator/beam/erl_process.h
+++ b/erts/emulator/beam/erl_process.h
@@ -253,18 +253,19 @@ typedef enum {
  * eachother. Most frequent - lowest bit number.
  */
 
-#define ERTS_SSI_AUX_WORK_DD			(((erts_aint32_t) 1) << 0)
-#define ERTS_SSI_AUX_WORK_DD_THR_PRGR		(((erts_aint32_t) 1) << 1)
-#define ERTS_SSI_AUX_WORK_FIX_ALLOC_DEALLOC	(((erts_aint32_t) 1) << 2)
-#define ERTS_SSI_AUX_WORK_FIX_ALLOC_LOWER_LIM	(((erts_aint32_t) 1) << 3)
-#define ERTS_SSI_AUX_WORK_ASYNC_READY		(((erts_aint32_t) 1) << 4)
-#define ERTS_SSI_AUX_WORK_ASYNC_READY_CLEAN	(((erts_aint32_t) 1) << 5)
-#define ERTS_SSI_AUX_WORK_MISC_THR_PRGR		(((erts_aint32_t) 1) << 6)
-#define ERTS_SSI_AUX_WORK_MISC			(((erts_aint32_t) 1) << 7)
-#define ERTS_SSI_AUX_WORK_CHECK_CHILDREN	(((erts_aint32_t) 1) << 8)
-#define ERTS_SSI_AUX_WORK_SET_TMO		(((erts_aint32_t) 1) << 9)
-#define ERTS_SSI_AUX_WORK_MSEG_CACHE_CHECK	(((erts_aint32_t) 1) << 10)
-#define ERTS_SSI_AUX_WORK_REAP_PORTS		(((erts_aint32_t) 1) << 11)
+#define ERTS_SSI_AUX_WORK_DELAYED_AW_WAKEUP	(((erts_aint32_t) 1) << 0)
+#define ERTS_SSI_AUX_WORK_DD			(((erts_aint32_t) 1) << 1)
+#define ERTS_SSI_AUX_WORK_DD_THR_PRGR		(((erts_aint32_t) 1) << 2)
+#define ERTS_SSI_AUX_WORK_FIX_ALLOC_DEALLOC	(((erts_aint32_t) 1) << 3)
+#define ERTS_SSI_AUX_WORK_FIX_ALLOC_LOWER_LIM	(((erts_aint32_t) 1) << 4)
+#define ERTS_SSI_AUX_WORK_ASYNC_READY		(((erts_aint32_t) 1) << 5)
+#define ERTS_SSI_AUX_WORK_ASYNC_READY_CLEAN	(((erts_aint32_t) 1) << 6)
+#define ERTS_SSI_AUX_WORK_MISC_THR_PRGR		(((erts_aint32_t) 1) << 7)
+#define ERTS_SSI_AUX_WORK_MISC			(((erts_aint32_t) 1) << 8)
+#define ERTS_SSI_AUX_WORK_CHECK_CHILDREN	(((erts_aint32_t) 1) << 9)
+#define ERTS_SSI_AUX_WORK_SET_TMO		(((erts_aint32_t) 1) << 10)
+#define ERTS_SSI_AUX_WORK_MSEG_CACHE_CHECK	(((erts_aint32_t) 1) << 11)
+#define ERTS_SSI_AUX_WORK_REAP_PORTS		(((erts_aint32_t) 1) << 12)
 
 typedef struct ErtsSchedulerSleepInfo_ ErtsSchedulerSleepInfo;
 
@@ -403,6 +404,11 @@ typedef struct {
 } ErtsSchedWallTime;
 
 typedef struct {
+    int sched;
+    erts_aint32_t aux_work;
+} ErtsDelayedAuxWorkWakeupJob;
+
+typedef struct {
     int sched_id;
     ErtsSchedulerData *esdp;
     ErtsSchedulerSleepInfo *ssi;
@@ -431,6 +437,13 @@ typedef struct {
 	void *queue;
     } async_ready;
 #endif
+#ifdef ERTS_SMP
+    struct {
+	int *sched2jix;
+	int jix;
+	ErtsDelayedAuxWorkWakeupJob *job;
+    } delayed_wakeup;
+#endif
 } ErtsAuxWorkData;
 
 struct ErtsSchedulerData_ {
@@ -464,7 +477,6 @@ struct ErtsSchedulerData_ {
     int virtual_reds;
     int cpu_id;			/* >= 0 when bound */
     ErtsAuxWorkData aux_work_data;
-
     ErtsAtomCacheMap atom_cache_map;
 
     ErtsSchedAllocData alloc_data;
diff --git a/erts/emulator/beam/erl_sched_spec_pre_alloc.c b/erts/emulator/beam/erl_sched_spec_pre_alloc.c
index 37b186abd9..a490aec734 100644
--- a/erts/emulator/beam/erl_sched_spec_pre_alloc.c
+++ b/erts/emulator/beam/erl_sched_spec_pre_alloc.c
@@ -116,54 +116,84 @@ erts_sspa_create(size_t blk_sz, int pa_size)
     return data;
 }
 
-static ERTS_INLINE erts_aint_t
+static ERTS_INLINE void
 enqueue_remote_managed_thread(erts_sspa_chunk_header_t *chdr,
 			      erts_sspa_blk_t *this,
-			      int want_last)
+			      int cinit)
 {
-    erts_aint_t ilast, itmp;
+    erts_aint_t itmp;
+    erts_sspa_blk_t *enq;
 
     erts_atomic_init_nob(&this->next_atmc, ERTS_AINT_NULL);
-
     /* Enqueue at end of list... */
 
-    ilast = erts_atomic_read_nob(&chdr->tail.data.last);
-    while (1) {
-	erts_sspa_blk_t *last = (erts_sspa_blk_t *) ilast;
-	itmp = erts_atomic_cmpxchg_mb(&last->next_atmc,
-				       (erts_aint_t) this,
-				       ERTS_AINT_NULL);
-	if (itmp == ERTS_AINT_NULL)
-	    break;
-	ilast = itmp;
+    enq = (erts_sspa_blk_t *) erts_atomic_read_nob(&chdr->tail.data.last);
+    itmp = erts_atomic_cmpxchg_relb(&enq->next_atmc,
+				    (erts_aint_t) this,
+				    ERTS_AINT_NULL);
+    if (itmp == ERTS_AINT_NULL) {
+	/* We are required to move last pointer */
+#ifdef DEBUG
+	ASSERT(ERTS_AINT_NULL == erts_atomic_read_nob(&this->next_atmc));
+	ASSERT(((erts_aint_t) enq)
+	       == erts_atomic_xchg_relb(&chdr->tail.data.last,
+				      (erts_aint_t) this));
+#else
+	erts_atomic_set_relb(&chdr->tail.data.last, (erts_aint_t) this);
+#endif
     }
+    else {
+	/*
+	 * We *need* to insert element somewhere in between the
+	 * last element we read earlier and the actual last element.
+	 */
+	int i = cinit;
 
-    /* Move last pointer forward... */
-    while (1) {
-	erts_aint_t itmp;
-	if (want_last) {
-	    if (erts_atomic_read_rb(&this->next_atmc) != ERTS_AINT_NULL) {
-		/* Someone else will move it forward */
-		return erts_atomic_read_nob(&chdr->tail.data.last);
+	while (1) {
+	    erts_aint_t itmp2;
+	    erts_atomic_set_nob(&this->next_atmc, itmp);
+	    itmp2 = erts_atomic_cmpxchg_relb(&enq->next_atmc,
+					     (erts_aint_t) this,
+					     itmp);
+	    if (itmp == itmp2)
+		break; /* inserted this */
+	    if ((i & 1) == 0)
+		itmp = itmp2;
+	    else {
+		enq = (erts_sspa_blk_t *) itmp;
+		itmp = erts_atomic_read_acqb(&enq->next_atmc);
+		ASSERT(itmp != ERTS_AINT_NULL);
 	    }
+	    i++;
 	}
-	else {
-	    if (erts_atomic_read_nob(&this->next_atmc) != ERTS_AINT_NULL) {
-		/* Someone else will move it forward */
-		return ERTS_AINT_NULL;
-	    }
+    }
+}
+
+static ERTS_INLINE erts_aint_t
+check_insert_marker(erts_sspa_chunk_header_t *chdr, erts_aint_t ilast)
+{
+    if (!chdr->head.used_marker
+	&& chdr->head.unref_end == (erts_sspa_blk_t *) ilast) {
+	erts_aint_t itmp;
+	erts_sspa_blk_t *last = (erts_sspa_blk_t *) ilast;
+
+	erts_atomic_init_nob(&chdr->tail.data.marker.next_atmc, ERTS_AINT_NULL);
+	itmp = erts_atomic_cmpxchg_relb(&last->next_atmc,
+					(erts_aint_t) &chdr->tail.data.marker,
+					ERTS_AINT_NULL);
+	if (itmp == ERTS_AINT_NULL) {
+	    ilast = (erts_aint_t) &chdr->tail.data.marker;
+	    chdr->head.used_marker = !0;
+	    erts_atomic_set_relb(&chdr->tail.data.last, ilast);
 	}
-	itmp = erts_atomic_cmpxchg_mb(&chdr->tail.data.last,
-				      (erts_aint_t) this,
-				      ilast);
-	if (ilast == itmp)
-	    return want_last ? (erts_aint_t) this : ERTS_AINT_NULL;
-	ilast = itmp;
     }
+    return ilast;
 }
 
 void
-erts_sspa_remote_free(erts_sspa_chunk_header_t *chdr, erts_sspa_blk_t *blk)
+erts_sspa_remote_free(erts_sspa_chunk_header_t *chdr,
+		      erts_sspa_blk_t *blk,
+		      int cinit)
 {
     int um_refc_ix = 0;
     int managed_thread = erts_thr_progress_is_managed_thread();
@@ -180,7 +210,7 @@ erts_sspa_remote_free(erts_sspa_chunk_header_t *chdr, erts_sspa_blk_t *blk)
 	}
     }
 
-    (void) enqueue_remote_managed_thread(chdr, blk, 0);
+    enqueue_remote_managed_thread(chdr, blk, cinit);
 
     if (!managed_thread)
 	erts_atomic_dec_relb(&chdr->tail.data.um_refc[um_refc_ix]);
@@ -208,24 +238,17 @@ fetch_remote(erts_sspa_chunk_header_t *chdr, int max)
 	    int um_refc_ix;
 	    chdr->head.next.thr_progress_reached = 1;
 	    um_refc_ix = chdr->head.next.um_refc_ix;
-	    if (erts_atomic_read_acqb(&chdr->tail.data.um_refc[um_refc_ix]) == 0) {
+	    if (erts_atomic_read_nob(&chdr->tail.data.um_refc[um_refc_ix]) == 0) {
+
+		ETHR_MEMBAR(ETHR_LoadLoad|ETHR_LoadStore);
 
 		/* Move unreferenced end pointer forward... */
 
 		chdr->head.unref_end = chdr->head.next.unref_end;
 
-		if (!chdr->head.used_marker
-		    && chdr->head.unref_end == (erts_sspa_blk_t *) ilast) {
-		    /* Need to equeue marker */
-		    chdr->head.used_marker = 1;
-		    ilast = enqueue_remote_managed_thread(chdr,
-							  &chdr->tail.data.marker,
-							  1);
-		}
+		ilast = check_insert_marker(chdr, ilast);
 
-		if (chdr->head.unref_end == (erts_sspa_blk_t *) ilast)
-		    ERTS_THR_MEMORY_BARRIER;
-		else {
+		if (chdr->head.unref_end != (erts_sspa_blk_t *) ilast) {
 		    chdr->head.next.unref_end = (erts_sspa_blk_t *) ilast;
 		    chdr->head.next.thr_progress = erts_thr_progress_later(NULL);
 		    erts_atomic32_set_relb(&chdr->tail.data.um_refc_ix,
diff --git a/erts/emulator/beam/erl_sched_spec_pre_alloc.h b/erts/emulator/beam/erl_sched_spec_pre_alloc.h
index d36066c399..bccb1aba7a 100644
--- a/erts/emulator/beam/erl_sched_spec_pre_alloc.h
+++ b/erts/emulator/beam/erl_sched_spec_pre_alloc.h
@@ -142,7 +142,8 @@ check_local_list(erts_sspa_chunk_header_t *chdr)
 erts_sspa_data_t *erts_sspa_create(size_t blk_sz,
 				   int pa_size);
 void erts_sspa_remote_free(erts_sspa_chunk_header_t *chdr,
-			   erts_sspa_blk_t *blk);
+			   erts_sspa_blk_t *blk,
+			   int cinit);
 erts_sspa_blk_t *erts_sspa_process_remote_frees(erts_sspa_chunk_header_t *chdr,
 						erts_sspa_blk_t *old_res);
 
@@ -216,7 +217,7 @@ erts_sspa_free(erts_sspa_data_t *data, int cix, char *cblk)
     chdr = &chnk->aligned.header;
     if (chnk_cix != cix) {
 	/* Remote chunk */
-	erts_sspa_remote_free(chdr, blk);
+	erts_sspa_remote_free(chdr, blk, chnk_cix - cix);
     }
     else {
 	/* Local chunk */