Merge branch 'rickard/dirty_scheduler_collapse/maint-21/OTP-15509' into maint-21

* rickard/dirty_scheduler_collapse/maint-21/OTP-15509:
  Fix bug causing dirty scheduler sleeper list inconsistency

2 files changed, 110 insertions, 50 deletions

diff --git a/erts/emulator/beam/erl_process.c b/erts/emulator/beam/erl_process.c
index 2427d87f66..19093ebfdd 100644
--- a/erts/emulator/beam/erl_process.c
+++ b/erts/emulator/beam/erl_process.c
@@ -92,10 +92,6 @@
 #undef HARDDEBUG
 #endif
 
-#ifdef HARDDEBUG
-#define HARDDEBUG_RUNQS
-#endif
-
 #ifdef HIPE
 #include "hipe_mode_switch.h"	/* for hipe_init_process() */
 #include "hipe_signal.h"	/* for hipe_thread_signal_init() */
@@ -3300,6 +3296,46 @@ check_io_timer(void *null)
 #define prepare_for_sys_schedule() 0
 #endif
 
+#ifdef HARDDEBUG
+#define ERTS_HDBG_CHK_SLEEP_LIST(SL, L, F, FN) \
+    check_sleepers_list((SL), (L), (F), (FN))
+static void check_sleepers_list(ErtsSchedulerSleepList *sl,
+                                int lock,
+                                ErtsSchedulerSleepInfo *find,
+                                ErtsSchedulerSleepInfo *find_not)
+{
+    ErtsSchedulerSleepInfo *last_out;
+    int found = 0;
+
+    if (lock)
+        erts_spin_lock(&sl->lock);
+
+    ERTS_ASSERT(!find_not || (!find_not->next && !find_not->prev));
+    
+    last_out = sl->list;
+    if (last_out) {
+        ErtsSchedulerSleepInfo *tmp = last_out;
+        do {
+            ERTS_ASSERT(tmp->next);
+            ERTS_ASSERT(tmp->prev);
+            ERTS_ASSERT(tmp->next->prev == tmp);
+            ERTS_ASSERT(tmp->prev->next == tmp);
+            ERTS_ASSERT(tmp != find_not);
+            if (tmp == find)
+                found = !0;
+            tmp = tmp->next;
+            
+        } while (tmp != last_out);
+    }
+    ERTS_ASSERT(!find || found);
+
+    if (lock)
+        erts_spin_unlock(&sl->lock);
+}
+#else
+#define ERTS_HDBG_CHK_SLEEP_LIST(SL, L, F, FN) ((void) 0)
+#endif
+
 static void
 scheduler_wait(int *fcalls, ErtsSchedulerData *esdp, ErtsRunQueue *rq)
 {
@@ -3313,23 +3349,29 @@ scheduler_wait(int *fcalls, ErtsSchedulerData *esdp, ErtsRunQueue *rq)
 
     ERTS_LC_ASSERT(erts_lc_runq_is_locked(rq));
 
-    if (ERTS_RUNQ_IX_IS_DIRTY(rq->ix))
-	erts_spin_lock(&rq->sleepers.lock);
     flgs = sched_prep_spin_wait(ssi);
     if (flgs & ERTS_SSI_FLG_SUSPENDED) {
 	/* Go suspend instead... */
-	if (ERTS_RUNQ_IX_IS_DIRTY(rq->ix))
-	    erts_spin_unlock(&rq->sleepers.lock);
 	return;
     }
 
     if (ERTS_RUNQ_IX_IS_DIRTY(rq->ix)) {
-	ssi->prev = NULL;
-	ssi->next = rq->sleepers.list;
-	if (rq->sleepers.list)
-	    rq->sleepers.list->prev = ssi;
-	rq->sleepers.list = ssi;
-	erts_spin_unlock(&rq->sleepers.lock);
+        erts_spin_lock(&rq->sleepers.lock);
+        ERTS_HDBG_CHK_SLEEP_LIST(&rq->sleepers, 0, NULL, ssi);
+        ASSERT(!ssi->next); /* Not in sleepers list */
+        ASSERT(!ssi->prev);
+        if (!rq->sleepers.list) {
+            ssi->next = ssi->prev = ssi;
+            rq->sleepers.list = ssi;
+        }
+        else {
+            ssi->prev = rq->sleepers.list;
+            ssi->next = rq->sleepers.list->next;
+            ssi->prev->next = ssi;
+            ssi->next->prev = ssi;
+        }
+        ERTS_HDBG_CHK_SLEEP_LIST(&rq->sleepers, 0, ssi, NULL);
+        erts_spin_unlock(&rq->sleepers.lock);
         dirty_active(esdp, -1);
     }
 
@@ -3470,8 +3512,28 @@ scheduler_wait(int *fcalls, ErtsSchedulerData *esdp, ErtsRunQueue *rq)
                                         (ERTS_SSI_FLG_SUSPENDED
                                          | ERTS_SSI_FLG_MSB_EXEC));
 
-    if (ERTS_SCHEDULER_IS_DIRTY(esdp))
+    if (ERTS_SCHEDULER_IS_DIRTY(esdp)) {
         dirty_sched_wall_time_change(esdp, working = 1);
+        erts_spin_lock(&rq->sleepers.lock);
+        ERTS_HDBG_CHK_SLEEP_LIST(&rq->sleepers, 0, ssi->next ? ssi : NULL, NULL);
+        if (ssi->next) { /* Still in list... */
+            if (ssi->next == ssi) {
+                ASSERT(rq->sleepers.list == ssi);
+                ASSERT(ssi->prev == ssi);
+                rq->sleepers.list = NULL;
+            }
+            else {
+                ASSERT(ssi->prev != ssi);
+                if (rq->sleepers.list == ssi)
+                    rq->sleepers.list = ssi->next;
+                ssi->prev->next = ssi->next;
+                ssi->next->prev = ssi->prev;
+            }
+            ssi->next = ssi->prev = NULL;
+        }
+        ERTS_HDBG_CHK_SLEEP_LIST(&rq->sleepers, 0, NULL, ssi);
+        erts_spin_unlock(&rq->sleepers.lock);
+    }
     else if (!thr_prgr_active) {
         erts_thr_progress_active(erts_thr_prgr_data(esdp), thr_prgr_active = 1);
         sched_wall_time_change(esdp, 1);
@@ -3559,56 +3621,44 @@ wake_scheduler(ErtsRunQueue *rq)
 }
 
 static void
-wake_dirty_schedulers(ErtsRunQueue *rq, int one)
+wake_dirty_scheduler(ErtsRunQueue *rq)
 {
-    ErtsSchedulerSleepInfo *ssi;
+    ErtsSchedulerSleepInfo *lo_ssi, *fo_ssi;
     ErtsSchedulerSleepList *sl;
 
     ASSERT(ERTS_RUNQ_IX_IS_DIRTY(rq->ix));
 
     sl = &rq->sleepers;
     erts_spin_lock(&sl->lock);
-    ssi = sl->list;
-    if (!ssi) {
+    ERTS_HDBG_CHK_SLEEP_LIST(&rq->sleepers, 0, NULL, NULL);
+    lo_ssi = sl->list;
+    if (!lo_ssi) {
 	erts_spin_unlock(&sl->lock);
-	if (one)
-	    wake_scheduler(rq);
-    } else if (one) {
+        wake_scheduler(rq);
+    }
+    else {
 	erts_aint32_t flgs;
-	if (ssi->prev)
-	    ssi->prev->next = ssi->next;
-	else {
-	    ASSERT(sl->list == ssi);
-	    sl->list = ssi->next;
+        fo_ssi = lo_ssi->next;
+        ASSERT(fo_ssi->prev == lo_ssi);
+        if (fo_ssi == lo_ssi) {
+	    ASSERT(lo_ssi->prev == lo_ssi);
+            sl->list = NULL;
+        }
+        else {
+            ASSERT(lo_ssi->prev != lo_ssi);
+	    lo_ssi->next = fo_ssi->next;
+	    fo_ssi->next->prev = fo_ssi->prev;
 	}
-	if (ssi->next)
-	    ssi->next->prev = ssi->prev;
-
-	erts_spin_unlock(&sl->lock);
-
-	ERTS_THR_MEMORY_BARRIER;
-	flgs = ssi_flags_set_wake(ssi);
-	erts_sched_finish_poke(ssi, flgs);
-    } else {
-	sl->list = NULL;
+        fo_ssi->next = fo_ssi->prev = NULL;
+        ERTS_HDBG_CHK_SLEEP_LIST(&rq->sleepers, 0, NULL, fo_ssi);
 	erts_spin_unlock(&sl->lock);
 
 	ERTS_THR_MEMORY_BARRIER;
-	do {
-	    ErtsSchedulerSleepInfo *wake_ssi = ssi;
-	    ssi = ssi->next;
-	    erts_sched_finish_poke(wake_ssi, ssi_flags_set_wake(wake_ssi));
-	} while (ssi);
+	flgs = ssi_flags_set_wake(fo_ssi);
+	erts_sched_finish_poke(fo_ssi, flgs);
     }
 }
 
-static void
-wake_dirty_scheduler(ErtsRunQueue *rq)
-{
-    wake_dirty_schedulers(rq, 1);
-}
-
-
 #define ERTS_NO_USED_RUNQS_SHIFT 16
 #define ERTS_NO_RUNQS_MASK 0xffffU
 
@@ -5988,6 +6038,8 @@ erts_init_scheduling(int no_schedulers, int no_schedulers_online, int no_poll_th
     for (ix = 0; ix < no_dirty_cpu_schedulers; ix++) {
 	ErtsSchedulerSleepInfo *ssi = &aligned_dirty_cpu_sched_sleep_info[ix].ssi;
 	erts_atomic32_init_nob(&ssi->flags, 0);
+        ssi->next = NULL;
+        ssi->prev = NULL;
 	ssi->event = NULL; /* initialized in sched_dirty_cpu_thread_func */
 	erts_atomic32_init_nob(&ssi->aux_work, 0);
     }
@@ -5998,6 +6050,8 @@ erts_init_scheduling(int no_schedulers, int no_schedulers_online, int no_poll_th
     for (ix = 0; ix < no_dirty_io_schedulers; ix++) {
 	ErtsSchedulerSleepInfo *ssi = &aligned_dirty_io_sched_sleep_info[ix].ssi;
 	erts_atomic32_init_nob(&ssi->flags, 0);
+        ssi->next = NULL;
+        ssi->prev = NULL;
 	ssi->event = NULL; /* initialized in sched_dirty_io_thread_func */
 	erts_atomic32_init_nob(&ssi->aux_work, 0);
     }
@@ -7517,7 +7571,13 @@ suspend_scheduler(ErtsSchedulerData *esdp)
         return;
     }
 
+#ifdef HARDDEBUG
+    if (sched_type != ERTS_SCHED_NORMAL)
+        ERTS_HDBG_CHK_SLEEP_LIST(&esdp->run_queue->sleepers, !0, NULL, ssi);
+#endif
+
     if (erts_atomic32_read_nob(&ssi->flags) & ERTS_SSI_FLG_MSB_EXEC) {
+
         ASSERT(no == 1);
         if (!msb_scheduler_type_switch(sched_type, esdp, no))
             return;
diff --git a/erts/emulator/beam/erl_process.h b/erts/emulator/beam/erl_process.h
index a1b029adbe..0aa19e7bde 100644
--- a/erts/emulator/beam/erl_process.h
+++ b/erts/emulator/beam/erl_process.h
@@ -359,7 +359,7 @@ typedef struct ErtsSchedulerSleepInfo_ ErtsSchedulerSleepInfo;
 
 typedef struct {
     erts_spinlock_t lock;
-    ErtsSchedulerSleepInfo *list;
+    ErtsSchedulerSleepInfo *list; /* circular lifo list; points to last out */
 } ErtsSchedulerSleepList;
 
 struct ErtsSchedulerSleepInfo_ {