erts: Move all I/O polling to a seperate thread

10 files changed, 334 insertions, 476 deletions

diff --git a/erts/emulator/beam/erl_alloc.types b/erts/emulator/beam/erl_alloc.types
index 3ec1389d47..bf5487d31e 100644
--- a/erts/emulator/beam/erl_alloc.types
+++ b/erts/emulator/beam/erl_alloc.types
@@ -354,7 +354,6 @@ type	DRV_EV_STATE	LONG_LIVED	SYSTEM		driver_event_state
 type	DRV_SEL_D_STATE	FIXED_SIZE	SYSTEM		driver_select_data_state
 type	NIF_SEL_D_STATE	FIXED_SIZE	SYSTEM		enif_select_data_state
 type	FD_LIST		SHORT_LIVED	SYSTEM		fd_list
-type	ACTIVE_FD_ARR	SHORT_LIVED	SYSTEM		active_fd_array
 type	POLLSET		LONG_LIVED	SYSTEM		pollset
 type	POLLSET_UPDREQ	SHORT_LIVED	SYSTEM		pollset_update_req
 type	POLL_FDS	LONG_LIVED	SYSTEM		poll_fds
diff --git a/erts/emulator/beam/erl_bif_info.c b/erts/emulator/beam/erl_bif_info.c
index c782d7a8ce..d0c2accec6 100644
--- a/erts/emulator/beam/erl_bif_info.c
+++ b/erts/emulator/beam/erl_bif_info.c
@@ -98,9 +98,6 @@ static char erts_system_version[] = ("Erlang/OTP " ERLANG_OTP_RELEASE
 #ifdef HIPE
 				     " [hipe]"
 #endif	
-#ifdef ERTS_ENABLE_KERNEL_POLL
-				     " [kernel-poll:%s]"
-#endif	
 #ifdef ET_DEBUG
 #if ET_DEBUG
 				     " [type-assertions]"
@@ -372,9 +369,6 @@ erts_print_system_version(fmtfn_t to, void *arg, Process *c_p)
 		      , total, online
 		      , dirty_cpu, dirty_cpu_onln, dirty_io
 		      , erts_async_max_threads
-#ifdef ERTS_ENABLE_KERNEL_POLL
-		      , erts_use_kernel_poll ? "true" : "false"
-#endif
 	);
 }
 
@@ -2696,7 +2690,7 @@ BIF_RETTYPE system_info_1(BIF_ALIST_1)
 	BIF_RET(make_small(CONTEXT_REDS));
     } else if (ERTS_IS_ATOM_STR("kernel_poll", BIF_ARG_1)) {
 #ifdef ERTS_ENABLE_KERNEL_POLL
-	BIF_RET(erts_use_kernel_poll ? am_true : am_false);
+	BIF_RET(am_true);
 #else
 	BIF_RET(am_false);
 #endif    
diff --git a/erts/emulator/beam/erl_init.c b/erts/emulator/beam/erl_init.c
index 8c14c86219..b4d84f6b16 100644
--- a/erts/emulator/beam/erl_init.c
+++ b/erts/emulator/beam/erl_init.c
@@ -49,6 +49,7 @@
 #include "erl_bif_unique.h"
 #define ERTS_WANT_TIMER_WHEEL_API
 #include "erl_time.h"
+#include "erl_check_io.h"
 
 #ifdef HIPE
 #include "hipe_mode_switch.h"	/* for hipe_mode_switch_init() */
@@ -76,8 +77,10 @@ const char etp_otp_release[] = ERLANG_OTP_RELEASE;
 const char etp_compile_date[] = ERLANG_COMPILE_DATE;
 const char etp_arch[] = ERLANG_ARCHITECTURE;
 #ifdef ERTS_ENABLE_KERNEL_POLL
+const int erts_use_kernel_poll = 1;
 const int etp_kernel_poll_support = 1;
 #else
+const int erts_use_kernel_poll = 0;
 const int etp_kernel_poll_support = 0;
 #endif
 #if defined(ARCH_64)
@@ -203,16 +206,16 @@ int erts_no_line_info = 0;	/* -L: Don't load line information */
  */
 
 ErtsModifiedTimings erts_modified_timings[] = {
-    /* 0 */	{make_small(0), CONTEXT_REDS, INPUT_REDUCTIONS},
-    /* 1 */	{make_small(0), (3*CONTEXT_REDS)/4, 2*INPUT_REDUCTIONS},
-    /* 2 */	{make_small(0), CONTEXT_REDS/2, INPUT_REDUCTIONS/2},
-    /* 3 */	{make_small(0), (7*CONTEXT_REDS)/8, 3*INPUT_REDUCTIONS},
-    /* 4 */	{make_small(0), CONTEXT_REDS/3, 3*INPUT_REDUCTIONS},
-    /* 5 */	{make_small(0), (10*CONTEXT_REDS)/11, INPUT_REDUCTIONS/2},
-    /* 6 */	{make_small(1), CONTEXT_REDS/4, 2*INPUT_REDUCTIONS},
-    /* 7 */	{make_small(1), (5*CONTEXT_REDS)/7, INPUT_REDUCTIONS/3},
-    /* 8 */	{make_small(10), CONTEXT_REDS/5, 3*INPUT_REDUCTIONS},
-    /* 9 */	{make_small(10), (6*CONTEXT_REDS)/7, INPUT_REDUCTIONS/4}
+    /* 0 */	{make_small(0), CONTEXT_REDS},
+    /* 1 */	{make_small(0), (3*CONTEXT_REDS)/4},
+    /* 2 */	{make_small(0), CONTEXT_REDS/2},
+    /* 3 */	{make_small(0), (7*CONTEXT_REDS)/8},
+    /* 4 */	{make_small(0), CONTEXT_REDS/3},
+    /* 5 */	{make_small(0), (10*CONTEXT_REDS)/11},
+    /* 6 */	{make_small(1), CONTEXT_REDS/4},
+    /* 7 */	{make_small(1), (5*CONTEXT_REDS)/7},
+    /* 8 */	{make_small(10), CONTEXT_REDS/5},
+    /* 9 */	{make_small(10), (6*CONTEXT_REDS)/7}
 };
 
 #define ERTS_MODIFIED_TIMING_LEVELS \
@@ -310,8 +313,9 @@ erl_init(int ncpu,
     erts_init_sys_common_misc();
     erts_init_process(ncpu, proc_tab_sz, legacy_proc_tab);
     erts_init_scheduling(no_schedulers,
-			 no_schedulers_online
-			 , no_dirty_cpu_schedulers,
+			 no_schedulers_online,
+                         erts_no_poll_threads,
+			 no_dirty_cpu_schedulers,
 			 no_dirty_cpu_schedulers_online,
 			 no_dirty_io_schedulers
 			 );
@@ -558,9 +562,19 @@ void erts_usage(void)
     erts_fprintf(stderr, "-hmqd  val     set default message queue data flag for processes,\n");
     erts_fprintf(stderr, "               valid values are: off_heap | on_heap\n");
 
+    erts_fprintf(stderr, "-IOp number    set number of pollsets to be used to poll for I/O,\n");
+    erts_fprintf(stderr, "               This value has to be equal or smaller than the\n");
+    erts_fprintf(stderr, "               number of poll threads. If the current platform\n");
+    erts_fprintf(stderr, "               does not support concurrent update of pollsets\n");
+    erts_fprintf(stderr, "               this value is ignored.\n");
+    erts_fprintf(stderr, "-IOt number    set number of threads to be used to poll for I/O\n");
+    erts_fprintf(stderr, "-IOPp number   set number of pollsets as a percentage of the\n");
+    erts_fprintf(stderr, "               number of poll threads.");
+    erts_fprintf(stderr, "-IOPt number   set number of threads to be used to poll for I/O\n");
+    erts_fprintf(stderr, "               as a percentage of the number of schedulers.");
+
     /*    erts_fprintf(stderr, "-i module  set the boot module (default init)\n"); */
 
-    erts_fprintf(stderr, "-K boolean     enable or disable kernel poll\n");
     erts_fprintf(stderr, "-n[s|a|d]      Control behavior of signals to ports\n");
     erts_fprintf(stderr, "               Note that this flag is deprecated!\n");
     erts_fprintf(stderr, "-M<X> <Y>      memory allocator switches,\n");
@@ -727,7 +741,6 @@ early_init(int *argc, char **argv) /*
     char envbuf[21]; /* enough for any 64-bit integer */
     size_t envbufsz;
 
-
     erts_save_emu_args(*argc, argv);
 
     erts_sched_compact_load = 1;
@@ -839,6 +852,7 @@ early_init(int *argc, char **argv) /*
 		    }
 		    break;
 		}
+
 		case 'S' :
 		    if (argv[i][2] == 'P') {
 			int ptot, ponln;
@@ -1100,6 +1114,9 @@ early_init(int *argc, char **argv) /*
     erts_alloc_init(argc, argv, &alloc_opts); /* Handles (and removes)
 						 -M flags. */
     /* Require allocators */
+
+    erts_init_check_io(argc, argv);
+
     /*
      * Thread progress management:
      *
@@ -1107,13 +1124,14 @@ early_init(int *argc, char **argv) /*
      * ** Scheduler threads (see erl_process.c)
      * ** Aux thread (see erl_process.c)
      * ** Sys message dispatcher thread (see erl_trace.c)
+     * ** IO Poll threads (see erl_check_io.c)
      *
      * * Unmanaged threads that need to register:
      * ** Async threads (see erl_async.c)
      * ** Dirty scheduler threads
      */
     erts_thr_progress_init(no_schedulers,
-			   no_schedulers+2,
+			   no_schedulers+2+erts_no_poll_threads,
 			   erts_async_max_threads +
 			   erts_no_dirty_cpu_schedulers +
 			   erts_no_dirty_io_schedulers
@@ -1561,16 +1579,6 @@ erl_start(int argc, char **argv)
 	    have_break_handler = 0;
 	  break;
 
-	case 'K':
-	    /* If kernel poll support is present,
-	       erl_sys_args() will remove the K parameter
-	       and value */
-	    get_arg(argv[i]+2, argv[i+1], &i);
-	    erts_fprintf(stderr,
-		       "kernel-poll not supported; \"K\" parameter ignored\n",
-		       arg);
-	    break;
-
 	case 'n':
 	    arg = get_arg(argv[i]+2, argv[i+1], &i);
 	    switch (arg[0]) {
diff --git a/erts/emulator/beam/erl_lock_count.c b/erts/emulator/beam/erl_lock_count.c
index 11890e44d1..1ae6076b12 100644
--- a/erts/emulator/beam/erl_lock_count.c
+++ b/erts/emulator/beam/erl_lock_count.c
@@ -554,16 +554,6 @@ erts_lock_flags_t erts_lcnt_get_category_mask() {
     return lcnt_category_mask__;
 }
 
-#ifdef ERTS_ENABLE_KERNEL_POLL
-/* erl_poll/erl_check_io only exports one of these variants at a time, and we
- * may need to use either one depending on emulator startup flags. */
-void erts_lcnt_update_pollset_locks_nkp(int);
-void erts_lcnt_update_pollset_locks_kp(int);
-
-void erts_lcnt_update_cio_locks_nkp(int);
-void erts_lcnt_update_cio_locks_kp(int);
-#endif
-
 void erts_lcnt_set_category_mask(erts_lock_flags_t mask) {
     erts_lock_flags_t changed_categories;
 
@@ -590,16 +580,6 @@ void erts_lcnt_set_category_mask(erts_lock_flags_t mask) {
     }
 
     if(changed_categories & ERTS_LOCK_FLAGS_CATEGORY_IO) {
-#ifdef ERTS_ENABLE_KERNEL_POLL
-        if(erts_use_kernel_poll) {
-            erts_lcnt_update_pollset_locks_kp(mask & ERTS_LOCK_FLAGS_CATEGORY_IO);
-        } else {
-            erts_lcnt_update_pollset_locks_nkp(mask & ERTS_LOCK_FLAGS_CATEGORY_IO);
-        }
-#else
-        erts_lcnt_update_pollset_locks(mask & ERTS_LOCK_FLAGS_CATEGORY_IO);
-#endif
-
         erts_lcnt_update_cio_locks(mask & ERTS_LOCK_FLAGS_CATEGORY_IO);
         erts_lcnt_update_driver_locks(mask & ERTS_LOCK_FLAGS_CATEGORY_IO);
         erts_lcnt_update_port_locks(mask & ERTS_LOCK_FLAGS_CATEGORY_IO);
diff --git a/erts/emulator/beam/erl_port_task.c b/erts/emulator/beam/erl_port_task.c
index 86df0f4cb8..a588477320 100644
--- a/erts/emulator/beam/erl_port_task.c
+++ b/erts/emulator/beam/erl_port_task.c
@@ -582,12 +582,10 @@ static ERTS_INLINE void
 reset_executed_io_task_handle(ErtsPortTask *ptp)
 {
     if (ptp->u.alive.handle) {
-	ErtsIoTask *itp = ErtsContainerStruct(ptp->u.alive.handle, ErtsIoTask, task);
-
 	ASSERT(ptp == handle2task(ptp->u.alive.handle));
-	erts_io_notify_port_task_executed(ptp->u.alive.handle);
-	reset_port_task_handle(ptp->u.alive.handle);
-	erts_check_io_interrupt(itp->pollset, 1);
+        /* The port task handle is reset inside task_executed */
+	erts_io_notify_port_task_executed(ptp->type, ptp->u.alive.handle,
+                                          reset_port_task_handle);
     }
 }
 
diff --git a/erts/emulator/beam/erl_process.c b/erts/emulator/beam/erl_process.c
index 66dcd85af6..ffb67aefaf 100644
--- a/erts/emulator/beam/erl_process.c
+++ b/erts/emulator/beam/erl_process.c
@@ -24,6 +24,8 @@
 #  include "config.h"
 #endif
 
+#define ERTS_WANT_BREAK_HANDLING
+
 #include <stddef.h> /* offsetof() */
 #include "sys.h"
 #include "erl_vm.h"
@@ -50,6 +52,7 @@
 #include "erl_time.h"
 #include "erl_nfunc_sched.h"
 #include "erl_check_io.h"
+#include "erl_poll.h"
 
 #define ERTS_CHECK_TIME_REDS CONTEXT_REDS
 #define ERTS_DELAYED_WAKEUP_INFINITY (~(Uint64) 0)
@@ -208,6 +211,7 @@ static struct ErtsSchedBusyWait_ {
 int erts_disable_proc_not_running_opt;
 
 static ErtsAuxWorkData *aux_thread_aux_work_data;
+static ErtsAuxWorkData *poll_thread_aux_work_data;
 
 #define ERTS_SCHDLR_SSPND_CHNG_NMSB		(((erts_aint32_t) 1) << 0)
 #define ERTS_SCHDLR_SSPND_CHNG_MSB		(((erts_aint32_t) 1) << 1)
@@ -432,6 +436,7 @@ typedef union {
 static ErtsAlignedSchedulerSleepInfo *aligned_sched_sleep_info;
 static ErtsAlignedSchedulerSleepInfo *aligned_dirty_cpu_sched_sleep_info;
 static ErtsAlignedSchedulerSleepInfo *aligned_dirty_io_sched_sleep_info;
+static ErtsAlignedSchedulerSleepInfo *aligned_poll_thread_sleep_info;
 
 static Uint last_reductions;
 static Uint last_exact_reductions;
@@ -499,9 +504,13 @@ ERTS_SCHED_PREF_QUICK_ALLOC_IMPL(proclist,
 				 200,
 				 ERTS_ALC_T_PROC_LIST)
 
+#define ERTS_POLL_THREAD_SLEEP_INFO_IX(IX)                              \
+        (ASSERT(0 <= ((int) (IX))                                       \
+                && ((int) (IX)) < ((int) erts_no_poll_threads)),        \
+         &aligned_poll_thread_sleep_info[(IX)].ssi)
 #define ERTS_SCHED_SLEEP_INFO_IX(IX)					\
-    (ASSERT(-1 <= ((int) (IX))					        \
-		 && ((int) (IX)) < ((int) erts_no_schedulers)),		\
+    (ASSERT(((int)-1) <= ((int) (IX))                 \
+            && ((int) (IX)) < ((int) erts_no_schedulers)),		\
      &aligned_sched_sleep_info[(IX)].ssi)
 #define ERTS_DIRTY_CPU_SCHED_SLEEP_INFO_IX(IX)				\
     (ASSERT(0 <= ((int) (IX))					        \
@@ -1560,14 +1569,14 @@ erts_sched_finish_poke(ErtsSchedulerSleepInfo *ssi,
 {
     switch (flags & ERTS_SSI_FLGS_SLEEP_TYPE) {
     case ERTS_SSI_FLG_POLL_SLEEPING:
-	erts_check_io_interrupt(ssi->esdp->pollset, 1);
+	erts_check_io_interrupt(ssi->psi, 1);
 	break;
     case ERTS_SSI_FLG_POLL_SLEEPING|ERTS_SSI_FLG_TSE_SLEEPING:
 	/*
 	 * Thread progress blocking while poll sleeping; need
 	 * to signal on both...
 	 */
-	erts_check_io_interrupt(ssi->esdp->pollset, 1);
+	erts_check_io_interrupt(ssi->psi, 1);
 	/* fall through */
     case ERTS_SSI_FLG_TSE_SLEEPING:
 	erts_tse_set(ssi->event);
@@ -2827,36 +2836,6 @@ erts_set_aux_work_timeout(int ix, erts_aint32_t type, int enable)
     return old;
 }
 
-
-
-static ERTS_INLINE void
-sched_waiting_sys(Uint no, ErtsRunQueue *rq)
-{
-    ERTS_LC_ASSERT(erts_lc_runq_is_locked(rq));
-    ASSERT(rq->waiting >= 0);
-    (void) ERTS_RUNQ_FLGS_SET(rq, (ERTS_RUNQ_FLG_OUT_OF_WORK
-				   | ERTS_RUNQ_FLG_HALFTIME_OUT_OF_WORK));
-    rq->waiting++;
-    rq->waiting *= -1;
-    rq->woken = 0;
-    if (erts_system_profile_flags.scheduler)
-	profile_scheduler(make_small(no), am_inactive);
-}
-
-static ERTS_INLINE void
-sched_active_sys(Uint no, ErtsRunQueue *rq)
-{
-    ERTS_LC_ASSERT(erts_lc_runq_is_locked(rq));
-
-    ASSERT(!ERTS_RUNQ_IX_IS_DIRTY(rq->ix));
-
-    ASSERT(rq->waiting < 0);
-    rq->waiting *= -1;
-    rq->waiting--;
-    if (erts_system_profile_flags.scheduler)
-	profile_scheduler(make_small(no), am_active);
-}
-
 Uint
 erts_active_schedulers(void)
 {
@@ -2867,13 +2846,6 @@ erts_active_schedulers(void)
     return as;
 }
 
-static ERTS_INLINE int
-prepare_for_sys_schedule(int non_blocking)
-{
-    return 1;
-}
-
-
 static ERTS_INLINE void
 sched_waiting(Uint no, ErtsRunQueue *rq)
 {
@@ -3045,7 +3017,7 @@ sched_set_sleeptype(ErtsSchedulerSleepInfo *ssi, erts_aint32_t sleep_type)
 	erts_tse_reset(ssi->event);
     else {
 	ASSERT(sleep_type == ERTS_SSI_FLG_POLL_SLEEPING);
-	erts_check_io_interrupt(ssi->esdp->pollset, 0);
+        erts_check_io_interrupt(ssi->psi, 0);
     }
 
     while (1) {
@@ -3113,6 +3085,12 @@ thr_prgr_fin_wait(void *vssi)
 
 static void init_aux_work_data(ErtsAuxWorkData *awdp, ErtsSchedulerData *esdp, char *dawwp);
 
+void
+erts_aux_thread_poke()
+{
+    erts_sched_poke(ERTS_SCHED_SLEEP_INFO_IX(-1));
+}
+
 static void *
 aux_thread(void *unused)
 {
@@ -3121,6 +3099,7 @@ aux_thread(void *unused)
     erts_aint32_t aux_work;
     ErtsThrPrgrCallbacks callbacks;
     int thr_prgr_active = 1;
+    ERTS_MSACC_DECLARE_CACHE();
 
 #ifdef ERTS_ENABLE_LOCK_CHECK
     {
@@ -3144,9 +3123,14 @@ aux_thread(void *unused)
     init_aux_work_data(awdp, NULL, NULL);
     awdp->ssi = ssi;
 
+#if ERTS_POLL_USE_FALLBACK
+    ssi->psi = erts_create_pollset_thread(-1);
+#endif
 
     sched_prep_spin_wait(ssi);
 
+    ERTS_MSACC_SET_STATE_CACHED(ERTS_MSACC_STATE_OTHER);
+
     while (1) {
 	erts_aint32_t flgs;
 
@@ -3155,30 +3139,128 @@ aux_thread(void *unused)
 	    if (!thr_prgr_active)
 		erts_thr_progress_active(NULL, thr_prgr_active = 1);
 	    aux_work = handle_aux_work(awdp, aux_work, 1);
+            ERTS_MSACC_UPDATE_CACHE();
 	    if (aux_work && erts_thr_progress_update(NULL))
 		erts_thr_progress_leader_update(NULL);
 	}
 
 	if (!aux_work) {
+
+#ifdef ERTS_BREAK_REQUESTED
+            if (ERTS_BREAK_REQUESTED)
+                erts_do_break_handling();
+#endif
+
 	    if (thr_prgr_active)
 		erts_thr_progress_active(NULL, thr_prgr_active = 0);
-	    erts_thr_progress_prepare_wait(NULL);
+
+#if ERTS_POLL_USE_FALLBACK
 
 	    flgs = sched_spin_wait(ssi, 0);
 
 	    if (flgs & ERTS_SSI_FLG_SLEEPING) {
 		ASSERT(flgs & ERTS_SSI_FLG_WAITING);
+		flgs = sched_set_sleeptype(ssi, ERTS_SSI_FLG_POLL_SLEEPING);
+		if (flgs & ERTS_SSI_FLG_SLEEPING) {
+		    ASSERT(flgs & ERTS_SSI_FLG_POLL_SLEEPING);
+		    ASSERT(flgs & ERTS_SSI_FLG_WAITING);
+                    erts_check_io(ssi->psi);
+		}
+            }
+#else
+            erts_thr_progress_prepare_wait(NULL);
+
+	    flgs = sched_spin_wait(ssi, 0);
+
+	    if (flgs & ERTS_SSI_FLG_SLEEPING) {
 		flgs = sched_set_sleeptype(ssi, ERTS_SSI_FLG_TSE_SLEEPING);
 		if (flgs & ERTS_SSI_FLG_SLEEPING) {
-		    int res;
+                    int res;
 		    ASSERT(flgs & ERTS_SSI_FLG_TSE_SLEEPING);
 		    ASSERT(flgs & ERTS_SSI_FLG_WAITING);
-		    do {
-			res = erts_tse_wait(ssi->event);
-		    } while (res == EINTR);
+                    ERTS_MSACC_SET_STATE_CACHED(ERTS_MSACC_STATE_SLEEP);
+                    do {
+                        res = erts_tse_wait(ssi->event);
+                    } while (res == EINTR);
+                    ERTS_MSACC_SET_STATE_CACHED(ERTS_MSACC_STATE_OTHER);
+		}
+            }
+            erts_thr_progress_finalize_wait(NULL);
+#endif
+	}
+
+	flgs = sched_prep_spin_wait(ssi);
+    }
+    return NULL;
+}
+
+static void *
+poll_thread(void *arg)
+{
+    int id = (int)(UWord)arg;
+    ErtsAuxWorkData *awdp = poll_thread_aux_work_data+id;
+    ErtsSchedulerSleepInfo *ssi = ERTS_POLL_THREAD_SLEEP_INFO_IX(id);
+    erts_aint32_t aux_work;
+    ErtsThrPrgrCallbacks callbacks;
+    int thr_prgr_active = 1;
+    struct erts_poll_thread *psi = erts_create_pollset_thread(id);
+    ERTS_MSACC_DECLARE_CACHE();
+
+#ifdef ERTS_ENABLE_LOCK_CHECK
+    {
+	char buf[] = "poll_thread";
+	erts_lc_set_thread_name(buf);
+    }
+#endif
+
+    erts_port_task_pre_alloc_init_thread();
+    ssi->event = erts_tse_fetch();
+
+    erts_msacc_init_thread("poll", id, 0);
+
+    callbacks.arg = (void *) ssi;
+    callbacks.wakeup = thr_prgr_wakeup;
+    callbacks.prepare_wait = thr_prgr_prep_wait;
+    callbacks.wait = thr_prgr_wait;
+    callbacks.finalize_wait = thr_prgr_fin_wait;
+
+    erts_thr_progress_register_managed_thread(NULL, &callbacks, 0);
+    init_aux_work_data(awdp, NULL, NULL);
+    awdp->ssi = ssi;
+    ssi->psi = psi;
+
+    sched_prep_spin_wait(ssi);
+
+    ERTS_MSACC_SET_STATE_CACHED(ERTS_MSACC_STATE_OTHER);
+
+    while (1) {
+	erts_aint32_t flgs;
+
+	aux_work = erts_atomic32_read_acqb(&ssi->aux_work);
+	if (aux_work) {
+	    if (!thr_prgr_active)
+		erts_thr_progress_active(NULL, thr_prgr_active = 1);
+	    aux_work = handle_aux_work(awdp, aux_work, 1);
+            ERTS_MSACC_UPDATE_CACHE();
+	    if (aux_work && erts_thr_progress_update(NULL))
+		erts_thr_progress_leader_update(NULL);
+	}
+
+	if (!aux_work) {
+	    if (thr_prgr_active)
+		erts_thr_progress_active(NULL, thr_prgr_active = 0);
+
+	    flgs = sched_spin_wait(ssi, 0);
+
+	    if (flgs & ERTS_SSI_FLG_SLEEPING) {
+		ASSERT(flgs & ERTS_SSI_FLG_WAITING);
+		flgs = sched_set_sleeptype(ssi, ERTS_SSI_FLG_POLL_SLEEPING);
+		if (flgs & ERTS_SSI_FLG_SLEEPING) {
+		    ASSERT(flgs & ERTS_SSI_FLG_POLL_SLEEPING);
+		    ASSERT(flgs & ERTS_SSI_FLG_WAITING);
+                    erts_check_io(psi);
 		}
 	    }
-	    erts_thr_progress_finalize_wait(NULL);
 	}
 
 	flgs = sched_prep_spin_wait(ssi);
@@ -3219,273 +3301,136 @@ scheduler_wait(int *fcalls, ErtsSchedulerData *esdp, ErtsRunQueue *rq)
         dirty_active(esdp, -1);
     }
 
-    /*
-     * If all schedulers are waiting, one of them *should*
-     * be waiting in erl_sys_schedule()
-     */
-
-    if (ERTS_SCHEDULER_IS_DIRTY(esdp) || !prepare_for_sys_schedule(0)) {
+    sched_waiting(esdp->no, rq);
 
-	sched_waiting(esdp->no, rq);
-
-	erts_runq_unlock(rq);
-
-	spincount = sched_busy_wait.tse;
-
-	if (ERTS_SCHEDULER_IS_DIRTY(esdp))
-	    dirty_sched_wall_time_change(esdp, working = 0);
-        else if (thr_prgr_active != working)
-	    sched_wall_time_change(esdp, working = thr_prgr_active);
-
-	while (1) {
-	    ErtsMonotonicTime current_time = 0;
+    erts_runq_unlock(rq);
 
-	    aux_work = erts_atomic32_read_acqb(&ssi->aux_work);
-	    if (aux_work && !ERTS_SCHEDULER_IS_DIRTY(esdp)) {
-		if (!thr_prgr_active) {
-		    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, 1);
-                ERTS_MSACC_UPDATE_CACHE();
-		if (aux_work && erts_thr_progress_update(esdp))
-		    erts_thr_progress_leader_update(esdp);
-	    }
+    spincount = sched_busy_wait.tse;
 
-	    if (aux_work) {
-		if (!ERTS_SCHEDULER_IS_DIRTY(esdp)) {
-		    flgs = erts_atomic32_read_acqb(&ssi->flags);
-		    current_time = erts_get_monotonic_time(esdp);
-		    if (current_time >= erts_next_timeout_time(esdp->next_tmo_ref)) {
-			if (!thr_prgr_active) {
-			    erts_thr_progress_active(esdp, thr_prgr_active = 1);
-			    sched_wall_time_change(esdp, 1);
-			}
-			erts_bump_timers(esdp->timer_wheel, current_time);
-		    }
-		}
-	    }
-	    else {
-		ErtsMonotonicTime timeout_time;
-		int do_timeout = 0;
-		if (!ERTS_SCHEDULER_IS_DIRTY(esdp)) {
-		    timeout_time = erts_check_next_timeout_time(esdp);
-		    current_time = erts_get_monotonic_time(esdp);
-		    do_timeout = (current_time >= timeout_time);
-		} else {
-		    current_time = 0;
-		    timeout_time = ERTS_MONOTONIC_TIME_MAX;
-		}
-		if (do_timeout) {
-		    if (!thr_prgr_active) {
-			erts_thr_progress_active(esdp, thr_prgr_active = 1);
-			sched_wall_time_change(esdp, 1);
-		    }
-		}
-		else {
-		    if (!ERTS_SCHEDULER_IS_DIRTY(esdp)) {
-			if (thr_prgr_active) {
-			    erts_thr_progress_active(esdp, thr_prgr_active = 0);
-			    sched_wall_time_change(esdp, 0);
-			}
-			erts_thr_progress_prepare_wait(esdp);
-		    }
+    if (ERTS_SCHEDULER_IS_DIRTY(esdp))
+        dirty_sched_wall_time_change(esdp, working = 0);
+    else if (thr_prgr_active != working)
+        sched_wall_time_change(esdp, working = thr_prgr_active);
 
-		    flgs = sched_spin_wait(ssi, spincount);
-		    if (flgs & ERTS_SSI_FLG_SLEEPING) {
-			ASSERT(flgs & ERTS_SSI_FLG_WAITING);
-			flgs = sched_set_sleeptype(ssi, ERTS_SSI_FLG_TSE_SLEEPING);
-			if (flgs & ERTS_SSI_FLG_SLEEPING) {
-			    int res;
-			    ASSERT(flgs & ERTS_SSI_FLG_TSE_SLEEPING);
-			    ASSERT(flgs & ERTS_SSI_FLG_WAITING);
-			    current_time = ERTS_SCHEDULER_IS_DIRTY(esdp) ? 0 :
-				erts_get_monotonic_time(esdp);
-			    do {
-				Sint64 timeout;
-				if (current_time >= timeout_time)
-				    break;
-				if (!ERTS_SCHEDULER_IS_DIRTY(esdp)) {
-				    timeout = ERTS_MONOTONIC_TO_NSEC(timeout_time
-								     - current_time
-								     - 1) + 1;
-				} else
-				    timeout = -1;
-                                ERTS_MSACC_SET_STATE_CACHED_M(ERTS_MSACC_STATE_SLEEP);
-				res = erts_tse_twait(ssi->event, timeout);
-                                ERTS_MSACC_POP_STATE_M();
-				current_time = ERTS_SCHEDULER_IS_DIRTY(esdp) ? 0 :
-				    erts_get_monotonic_time(esdp);
-			    } while (res == EINTR);
-			}
-		    }
-		    if (!ERTS_SCHEDULER_IS_DIRTY(esdp))
-			erts_thr_progress_finalize_wait(esdp);
-		}
-		if (!ERTS_SCHEDULER_IS_DIRTY(esdp) && current_time >= timeout_time)
-		    erts_bump_timers(esdp->timer_wheel, current_time);
-	    }
+    while (1) {
+        ErtsMonotonicTime current_time = 0;
 
-	    if (!(flgs & ERTS_SSI_FLG_WAITING)) {
-		ASSERT(!(flgs & ERTS_SSI_FLG_SLEEPING));
-		break;
-	    }
+        aux_work = erts_atomic32_read_acqb(&ssi->aux_work);
+        if (aux_work && !ERTS_SCHEDULER_IS_DIRTY(esdp)) {
+            if (!thr_prgr_active) {
+                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, 1);
+            ERTS_MSACC_UPDATE_CACHE();
+            if (aux_work && erts_thr_progress_update(esdp))
+                erts_thr_progress_leader_update(esdp);
+        }
 
-	    flgs = sched_prep_cont_spin_wait(ssi);
-	    spincount = sched_busy_wait.aux_work;
+        if (aux_work) {
+            if (!ERTS_SCHEDULER_IS_DIRTY(esdp)) {
+                flgs = erts_atomic32_read_acqb(&ssi->flags);
+                current_time = erts_get_monotonic_time(esdp);
+                if (current_time >= erts_next_timeout_time(esdp->next_tmo_ref)) {
+                    if (!thr_prgr_active) {
+                        erts_thr_progress_active(esdp, thr_prgr_active = 1);
+                        sched_wall_time_change(esdp, 1);
+                    }
+                    erts_bump_timers(esdp->timer_wheel, current_time);
+                }
+            }
+        }
+        else {
+            ErtsMonotonicTime timeout_time;
+            int do_timeout = 0;
+            if (!ERTS_SCHEDULER_IS_DIRTY(esdp)) {
+                timeout_time = erts_check_next_timeout_time(esdp);
+                current_time = erts_get_monotonic_time(esdp);
+                do_timeout = (current_time >= timeout_time);
+            } else {
+                current_time = 0;
+                timeout_time = ERTS_MONOTONIC_TIME_MAX;
+            }
+            if (do_timeout) {
+                if (!thr_prgr_active) {
+                    erts_thr_progress_active(esdp, thr_prgr_active = 1);
+                    sched_wall_time_change(esdp, 1);
+                }
+            }
+            else {
+                if (!ERTS_SCHEDULER_IS_DIRTY(esdp)) {
+                    if (thr_prgr_active) {
+                        erts_thr_progress_active(esdp, thr_prgr_active = 0);
+                        sched_wall_time_change(esdp, 0);
+                    }
+                    erts_thr_progress_prepare_wait(esdp);
+                }
 
-	    if (!(flgs & ERTS_SSI_FLG_WAITING)) {
-		ASSERT(!(flgs & ERTS_SSI_FLG_SLEEPING));
-		break;
-	    }
+                flgs = sched_spin_wait(ssi, spincount);
+                if (flgs & ERTS_SSI_FLG_SLEEPING) {
+                    ASSERT(flgs & ERTS_SSI_FLG_WAITING);
+                    flgs = sched_set_sleeptype(ssi, ERTS_SSI_FLG_TSE_SLEEPING);
+                    if (flgs & ERTS_SSI_FLG_SLEEPING) {
+                        int res;
+                        ASSERT(flgs & ERTS_SSI_FLG_TSE_SLEEPING);
+                        ASSERT(flgs & ERTS_SSI_FLG_WAITING);
+                        current_time = ERTS_SCHEDULER_IS_DIRTY(esdp) ? 0 :
+                            erts_get_monotonic_time(esdp);
+                        do {
+                            Sint64 timeout;
+                            if (current_time >= timeout_time)
+                                break;
+                            if (!ERTS_SCHEDULER_IS_DIRTY(esdp)) {
+                                timeout = ERTS_MONOTONIC_TO_NSEC(timeout_time
+                                                                 - current_time
+                                                                 - 1) + 1;
+                            } else
+                                timeout = -1;
+                            ERTS_MSACC_SET_STATE_CACHED_M(ERTS_MSACC_STATE_SLEEP);
+                            res = erts_tse_twait(ssi->event, timeout);
+                            ERTS_MSACC_POP_STATE_M();
+                            current_time = ERTS_SCHEDULER_IS_DIRTY(esdp) ? 0 :
+                                erts_get_monotonic_time(esdp);
+                        } while (res == EINTR);
+                    }
+                }
+                if (!ERTS_SCHEDULER_IS_DIRTY(esdp))
+                    erts_thr_progress_finalize_wait(esdp);
+            }
+            if (!ERTS_SCHEDULER_IS_DIRTY(esdp) && current_time >= timeout_time)
+                erts_bump_timers(esdp->timer_wheel, current_time);
+        }
 
-	}
+        if (!(flgs & ERTS_SSI_FLG_WAITING)) {
+            ASSERT(!(flgs & ERTS_SSI_FLG_SLEEPING));
+            break;
+        }
 
-	if (flgs & ~(ERTS_SSI_FLG_SUSPENDED|ERTS_SSI_FLG_MSB_EXEC))
-	    erts_atomic32_read_band_nob(&ssi->flags,
-                                            (ERTS_SSI_FLG_SUSPENDED
-                                             | ERTS_SSI_FLG_MSB_EXEC));
+        flgs = sched_prep_cont_spin_wait(ssi);
+        spincount = sched_busy_wait.aux_work;
 
-	if (ERTS_SCHEDULER_IS_DIRTY(esdp))
-	    dirty_sched_wall_time_change(esdp, working = 1);
-        else if (!thr_prgr_active) {
-	    erts_thr_progress_active(esdp, thr_prgr_active = 1);
-	    sched_wall_time_change(esdp, 1);
+        if (!(flgs & ERTS_SSI_FLG_WAITING)) {
+            ASSERT(!(flgs & ERTS_SSI_FLG_SLEEPING));
+            break;
         }
 
-	erts_runq_lock(rq);
-	sched_active(esdp->no, rq);
-
     }
-    else
-    {
-
-	esdp->function_calls = 0;
-	*fcalls = 0;
-
-	ASSERT(!ERTS_SCHEDULER_IS_DIRTY(esdp));
-
-	sched_waiting_sys(esdp->no, rq);
-
-	erts_runq_unlock(rq);
-
-	ASSERT(working);
-	sched_wall_time_change(esdp, working = 0);
 
-	spincount = sched_busy_wait.sys_schedule;
-	if (spincount == 0)
-	    goto sys_aux_work;
+    if (flgs & ~(ERTS_SSI_FLG_SUSPENDED|ERTS_SSI_FLG_MSB_EXEC))
+        erts_atomic32_read_band_nob(&ssi->flags,
+                                        (ERTS_SSI_FLG_SUSPENDED
+                                         | ERTS_SSI_FLG_MSB_EXEC));
 
-	while (spincount-- > 0) {
-	    ErtsMonotonicTime current_time;
-
-	sys_poll_aux_work:
-
-	    if (working)
-		sched_wall_time_change(esdp, working = 0);
-
-	    ERTS_MSACC_SET_STATE_CACHED_M(ERTS_MSACC_STATE_CHECK_IO);
-
-            LTTNG2(scheduler_poll, esdp->no, 1);
-	    erl_sys_schedule(1); /* Might give us something to do */
-
-	    ERTS_MSACC_POP_STATE_M();
-
-	    if (!ERTS_SCHEDULER_IS_DIRTY(esdp)) {
-		current_time = erts_get_monotonic_time(esdp);
-		if (current_time >= erts_next_timeout_time(esdp->next_tmo_ref))
-		    erts_bump_timers(esdp->timer_wheel, current_time);
-	    }
-
-	sys_aux_work:
-
-	    aux_work = erts_atomic32_read_acqb(&ssi->aux_work);
-	    if (aux_work && !ERTS_SCHEDULER_IS_DIRTY(esdp)) {
-		if (!working)
-		    sched_wall_time_change(esdp, working = 1);
-		if (!thr_prgr_active)
-		    erts_thr_progress_active(esdp, thr_prgr_active = 1);
-		aux_work = handle_aux_work(&esdp->aux_work_data, aux_work, 1);
-                ERTS_MSACC_UPDATE_CACHE();
-		if (aux_work && erts_thr_progress_update(esdp))
-		    erts_thr_progress_leader_update(esdp);
-	    }
-
-	    flgs = erts_atomic32_read_acqb(&ssi->flags);
-	    if (!(flgs & ERTS_SSI_FLG_WAITING)) {
-		ASSERT(!(flgs & ERTS_SSI_FLG_SLEEPING));
-		goto sys_woken;
-	    }
-	}
-
-	erts_runq_lock(rq);
-
-	if (aux_work) {
-	    erts_runq_unlock(rq);
-	    goto sys_poll_aux_work;
-	}
-	flgs = sched_set_sleeptype(ssi, ERTS_SSI_FLG_POLL_SLEEPING);
-	if (!(flgs & ERTS_SSI_FLG_SLEEPING)) {
-	    if (!(flgs & ERTS_SSI_FLG_WAITING)) {
-		ASSERT(!(flgs & ERTS_SSI_FLG_SLEEPING));
-		goto sys_locked_woken;
-	    }
-	    erts_runq_unlock(rq);
-	    flgs = sched_prep_cont_spin_wait(ssi);
-	    if (!(flgs & ERTS_SSI_FLG_WAITING)) {
-		ASSERT(!(flgs & ERTS_SSI_FLG_SLEEPING));
-		goto sys_woken;
-	    }
-	    goto sys_poll_aux_work;
-	}
-
-	ASSERT(flgs & ERTS_SSI_FLG_POLL_SLEEPING);
-	ASSERT(flgs & ERTS_SSI_FLG_WAITING);
-
-	erts_runq_unlock(rq);
-
-	if (working)
-	    sched_wall_time_change(esdp, working = 0);
-
-	if (thr_prgr_active)
-	    erts_thr_progress_active(esdp, thr_prgr_active = 0);
-
-	ERTS_MSACC_SET_STATE_CACHED_M(ERTS_MSACC_STATE_CHECK_IO);
-        LTTNG2(scheduler_poll, esdp->no, 0);
-
-	erl_sys_schedule(0);
-
-	ERTS_MSACC_POP_STATE_M();
-
-	if (!ERTS_SCHEDULER_IS_DIRTY(esdp)) {
-	    ErtsMonotonicTime current_time = erts_get_monotonic_time(esdp);
-	    if (current_time >= erts_next_timeout_time(esdp->next_tmo_ref))
-		erts_bump_timers(esdp->timer_wheel, current_time);
-	}
-
-	flgs = sched_prep_cont_spin_wait(ssi);
-	if (flgs & ERTS_SSI_FLG_WAITING)
-	    goto sys_aux_work;
-
-    sys_woken:
-	if (!thr_prgr_active)
-	    erts_thr_progress_active(esdp, thr_prgr_active = 1);
-	erts_runq_lock(rq);
-    sys_locked_woken:
-	if (!thr_prgr_active) {
-	    erts_runq_unlock(rq);
-	    erts_thr_progress_active(esdp, thr_prgr_active = 1);
-	    erts_runq_lock(rq);
-	}
-	if (flgs & ~(ERTS_SSI_FLG_SUSPENDED|ERTS_SSI_FLG_MSB_EXEC))
-	    erts_atomic32_read_band_nob(&ssi->flags,
-                                            (ERTS_SSI_FLG_SUSPENDED
-                                             | ERTS_SSI_FLG_MSB_EXEC));
-	if (!working)
-	    sched_wall_time_change(esdp, working = 1);
-	sched_active_sys(esdp->no, rq);
+    if (ERTS_SCHEDULER_IS_DIRTY(esdp))
+        dirty_sched_wall_time_change(esdp, working = 1);
+    else if (!thr_prgr_active) {
+        erts_thr_progress_active(esdp, thr_prgr_active = 1);
+        sched_wall_time_change(esdp, 1);
     }
 
+    erts_runq_lock(rq);
+    sched_active(esdp->no, rq);
+
     if (ERTS_SCHEDULER_IS_DIRTY(esdp))
         dirty_active(esdp, 1);
 
@@ -5755,7 +5700,6 @@ init_scheduler_data(ErtsSchedulerData* esdp, int num,
     esdp->run_queue = runq;
     if (ERTS_RUNQ_IX_IS_DIRTY(runq->ix)) {
 	esdp->no = 0;
-        esdp->pollset = NULL;
         if (runq == ERTS_DIRTY_CPU_RUNQ)
             esdp->type = ERTS_SCHED_DIRTY_CPU;
         else {
@@ -5775,7 +5719,6 @@ init_scheduler_data(ErtsSchedulerData* esdp, int num,
         esdp->type = ERTS_SCHED_NORMAL;
 	esdp->no = (Uint) num;
 	esdp->dirty_no = 0;
-        esdp->pollset = erts_get_pollset(num);
         runq->scheduler = esdp;
     }
     esdp->dirty_shadow_process = shadow_proc;
@@ -5788,8 +5731,6 @@ init_scheduler_data(ErtsSchedulerData* esdp, int num,
 	shadow_proc->static_flags = ERTS_STC_FLG_SHADOW_PROC;
     }
 
-    esdp->function_calls = 0;
-
     ssi->esdp = esdp;
     esdp->ssi = ssi;
     esdp->current_process = NULL;
@@ -5821,8 +5762,8 @@ init_scheduler_data(ErtsSchedulerData* esdp, int num,
 }
 
 void
-erts_init_scheduling(int no_schedulers, int no_schedulers_online
-		     , int no_dirty_cpu_schedulers, int no_dirty_cpu_schedulers_online,
+erts_init_scheduling(int no_schedulers, int no_schedulers_online, int no_poll_threads,
+		     int no_dirty_cpu_schedulers, int no_dirty_cpu_schedulers_online,
 		     int no_dirty_io_schedulers
 		     )
 {
@@ -5849,6 +5790,7 @@ erts_init_scheduling(int no_schedulers, int no_schedulers_online
     ASSERT(no_dirty_cpu_schedulers >= 1);
     ASSERT(no_dirty_cpu_schedulers_online <= no_schedulers_online);
     ASSERT(no_dirty_cpu_schedulers_online >= 1);
+    ASSERT(erts_no_poll_threads == no_poll_threads);
 
     /* Create and initialize run queues */
 
@@ -5946,7 +5888,7 @@ erts_init_scheduling(int no_schedulers, int no_schedulers_online
     erts_no_total_schedulers += no_dirty_io_schedulers;
 
     /* Create and initialize scheduler sleep info */
-    no_ssi = n+1;
+    no_ssi = n + 1 /* aux thread */;
     aligned_sched_sleep_info =
 	erts_alloc_permanent_cache_aligned(
 	    ERTS_ALC_T_SCHDLR_SLP_INFO,
@@ -5957,12 +5899,13 @@ erts_init_scheduling(int no_schedulers, int no_schedulers_online
 	ssi->next = NULL;
 	ssi->prev = NULL;
 #endif
+        ssi->esdp = NULL;
 	erts_atomic32_init_nob(&ssi->flags, 0);
 	ssi->event = NULL; /* initialized in sched_thread_func */
 	erts_atomic32_init_nob(&ssi->aux_work, 0);
     }
 
-    aligned_sched_sleep_info++;
+    aligned_sched_sleep_info += 1 /* aux thread */;
 
     aligned_dirty_cpu_sched_sleep_info =
 	erts_alloc_permanent_cache_aligned(
@@ -5985,6 +5928,18 @@ erts_init_scheduling(int no_schedulers, int no_schedulers_online
 	erts_atomic32_init_nob(&ssi->aux_work, 0);
     }
 
+    aligned_poll_thread_sleep_info =
+        erts_alloc_permanent_cache_aligned(
+	    ERTS_ALC_T_SCHDLR_SLP_INFO,
+	    no_poll_threads*sizeof(ErtsAlignedSchedulerSleepInfo));
+    for (ix = 0; ix < no_poll_threads; ix++) {
+        ErtsSchedulerSleepInfo *ssi = &aligned_poll_thread_sleep_info[ix].ssi;
+        ssi->esdp = NULL;
+	erts_atomic32_init_nob(&ssi->flags, 0);
+	ssi->event = NULL; /* initialized in poll_thread */
+	erts_atomic32_init_nob(&ssi->aux_work, 0);
+    }
+
     /* Create and initialize scheduler specific data */
 
     daww_sz = ERTS_ALC_CACHE_LINE_ALIGN_SIZE((sizeof(ErtsDelayedAuxWorkWakeupJob)
@@ -6041,11 +5996,14 @@ erts_init_scheduling(int no_schedulers, int no_schedulers_online
     erts_atomic32_init_nob(&debug_wait_completed_count, 0); /* debug only */
     debug_wait_completed_flags = 0;
 
-
     aux_thread_aux_work_data =
 	erts_alloc_permanent_cache_aligned(ERTS_ALC_T_SCHDLR_DATA,
 					   sizeof(ErtsAuxWorkData));
 
+    poll_thread_aux_work_data =
+	erts_alloc_permanent_cache_aligned(ERTS_ALC_T_SCHDLR_DATA,
+					   no_poll_threads * sizeof(ErtsAuxWorkData));
+
     init_no_runqs(no_schedulers_online, no_schedulers_online);
     balance_info.last_active_runqs = no_schedulers;
     erts_mtx_init(&balance_info.update_mtx, "migration_info_update", NIL,
@@ -7034,12 +6992,8 @@ sched_set_suspended_sleeptype(ErtsSchedulerSleepInfo *ssi,
 			   | ERTS_SSI_FLG_WAITING
 			   | ERTS_SSI_FLG_SUSPENDED);
 
-    if (sleep_type == ERTS_SSI_FLG_TSE_SLEEPING)
-	erts_tse_reset(ssi->event);
-    else {
-	ASSERT(sleep_type == ERTS_SSI_FLG_POLL_SLEEPING);
-	erts_check_io_interrupt(ssi->esdp->pollset, 0);
-    }
+    ASSERT(sleep_type == ERTS_SSI_FLG_TSE_SLEEPING);
+    erts_tse_reset(ssi->event);
 
     while (1) {
 	oflgs = erts_atomic32_cmpxchg_acqb(&ssi->flags, nflgs, xflgs);
@@ -7291,17 +7245,6 @@ msb_scheduler_type_switch(ErtsSchedType sched_type,
     if (exec_type != ERTS_SCHED_NORMAL)
         schdlr_sspnd.last_msb_dirty_type = exec_type;
     else {
-        erts_aint32_t calls;
-        /*
-         * Going back to normal scheduler after
-         * dirty execution; make sure it will check
-         * for I/O...
-         */
-        if (ERTS_USE_MODIFIED_TIMING())
-            calls = ERTS_MODIFIED_TIMING_INPUT_REDS + 1;
-        else
-            calls = INPUT_REDUCTIONS + 1;
-        esdp->function_calls = calls;
 
         if ((nrml_prio == ERTS_MSB_NONE_PRIO_BIT)
             & ((dcpu_prio != ERTS_MSB_NONE_PRIO_BIT)
@@ -7376,7 +7319,7 @@ msb_scheduler_type_switch(ErtsSchedType sched_type,
 }
 
 static ERTS_INLINE void
-suspend_dirty_scheduler_sleep(ErtsSchedulerData *esdp)
+suspend_normal_scheduler_sleep(ErtsSchedulerData *esdp)
 {
     ErtsSchedulerSleepInfo *ssi = esdp->ssi;
     erts_aint32_t flgs = sched_spin_suspended(ssi,
@@ -7399,36 +7342,9 @@ suspend_dirty_scheduler_sleep(ErtsSchedulerData *esdp)
 }
 
 static ERTS_INLINE void
-suspend_normal_scheduler_sleep(ErtsSchedulerData *esdp)
+suspend_dirty_scheduler_sleep(ErtsSchedulerData *esdp)
 {
-    ErtsSchedulerSleepInfo *ssi = esdp->ssi;
-    erts_aint32_t flgs;
-    int spincount;
-
-    spincount = sched_busy_wait.sys_schedule;
-
-    while (spincount-- > 0) {
-
-        erl_sys_schedule(1);
-
-        flgs = erts_smp_atomic32_read_acqb(&ssi->flags);
-        if (flgs != (ERTS_SSI_FLG_SLEEPING
-                     | ERTS_SSI_FLG_WAITING
-                     | ERTS_SSI_FLG_SUSPENDED)) {
-            return;
-        }
-    }
-
-    flgs = sched_set_suspended_sleeptype(ssi, ERTS_SSI_FLG_POLL_SLEEPING);
-    if (flgs != (ERTS_SSI_FLG_SLEEPING
-                 | ERTS_SSI_FLG_POLL_SLEEPING
-                 | ERTS_SSI_FLG_WAITING
-                 | ERTS_SSI_FLG_SUSPENDED)) {
-        return;
-    }
-
-    /* Sleep on pollset... */
-    erl_sys_schedule(0);
+    suspend_normal_scheduler_sleep(esdp);
 }
 
 static void
@@ -7662,8 +7578,10 @@ suspend_scheduler(ErtsSchedulerData *esdp)
                             erts_thr_progress_active(esdp, thr_prgr_active = 0);
                             sched_wall_time_change(esdp, 0);
                         }
+                        erts_thr_progress_prepare_wait(NULL);
                         suspend_normal_scheduler_sleep(esdp);
-                        current_time = erts_get_monotonic_time(esdp);;
+                        erts_thr_progress_finalize_wait(NULL);
+                        current_time = erts_get_monotonic_time(esdp);
                     }
 
                     if (current_time >= timeout_time) {
@@ -8493,17 +8411,17 @@ sched_dirty_io_thread_func(void *vesdp)
     return NULL;
 }
 
-static ethr_tid aux_tid;
-
 void
 erts_start_schedulers(void)
 {
+    ethr_tid tid;
     int res = 0;
     Uint actual;
     Uint wanted = erts_no_schedulers;
     Uint wanted_no_schedulers = erts_no_schedulers;
     char name[16];
     ethr_thr_opts opts = ETHR_THR_OPTS_DEFAULT_INITER;
+    int ix;
 
     opts.detached = 1;
 
@@ -8548,7 +8466,6 @@ erts_start_schedulers(void)
     erts_no_schedulers = actual;
 
     {
-	int ix;
 	for (ix = 0; ix < erts_no_dirty_cpu_schedulers; ix++) {
 	    ErtsSchedulerData *esdp = ERTS_DIRTY_CPU_SCHEDULER_IX(ix);
 	    erts_snprintf(opts.name, 16, "%d_dirty_cpu_scheduler", ix + 1);
@@ -8571,10 +8488,18 @@ erts_start_schedulers(void)
 
     erts_snprintf(opts.name, 16, "aux");
 
-    res = ethr_thr_create(&aux_tid, aux_thread, NULL, &opts);
+    res = ethr_thr_create(&tid, aux_thread, NULL, &opts);
     if (res != 0)
 	erts_exit(ERTS_ERROR_EXIT, "Failed to create aux thread\n");
 
+    for (ix = 0; ix < erts_no_poll_threads; ix++) {
+        erts_snprintf(opts.name, 16, "%d_poller", ix);
+
+        res = ethr_thr_create(&tid, poll_thread, (void*)(UWord)ix, &opts);
+        if (res != 0)
+            erts_exit(ERTS_ERROR_EXIT, "Failed to create poll thread\n");
+    }
+
     if (actual < 1)
 	erts_exit(ERTS_ERROR_EXIT,
 		 "Failed to create any scheduler-threads: %s (%d)\n",
@@ -9641,7 +9566,6 @@ Process *erts_schedule(ErtsSchedulerData *esdp, Process *p, int calls)
     ErtsRunQueue *rq;
     int context_reds;
     int fcalls;
-    int input_reductions;
     int actual_reds;
     int reds;
     Uint32 flags;
@@ -9661,11 +9585,9 @@ Process *erts_schedule(ErtsSchedulerData *esdp, Process *p, int calls)
 
     if (ERTS_USE_MODIFIED_TIMING()) {
 	context_reds = ERTS_MODIFIED_TIMING_CONTEXT_REDS;
-	input_reductions = ERTS_MODIFIED_TIMING_INPUT_REDS;
     }
     else {
 	context_reds = CONTEXT_REDS;
-	input_reductions = INPUT_REDUCTIONS;
     }
 
     ERTS_LC_ASSERT(ERTS_SCHEDULER_IS_DIRTY(erts_get_scheduler_data())
@@ -9685,7 +9607,6 @@ Process *erts_schedule(ErtsSchedulerData *esdp, Process *p, int calls)
 	}
 	rq = erts_get_runq_current(esdp);
 	ASSERT(esdp);
-	fcalls = esdp->function_calls;
 	actual_reds = reds = 0;
 	erts_runq_lock(rq);
     } else {
@@ -9711,8 +9632,6 @@ Process *erts_schedule(ErtsSchedulerData *esdp, Process *p, int calls)
 	    reds = ERTS_PROC_MIN_CONTEXT_SWITCH_REDS_COST;
 	esdp->virtual_reds = 0;
 
-        esdp->function_calls += reds;
-	fcalls = esdp->function_calls;
 	ASSERT(esdp && esdp == erts_get_scheduler_data());
 
 	rq = erts_get_runq_current(esdp);
@@ -9802,16 +9721,6 @@ Process *erts_schedule(ErtsSchedulerData *esdp, Process *p, int calls)
 
 	ERTS_CHK_NO_PROC_LOCKS;
 
-	if (is_normal_sched) {
-	    if (esdp->check_time_reds >= ERTS_CHECK_TIME_REDS)
-		(void) erts_get_monotonic_time(esdp);
-
-	    if (esdp->last_monotonic_time >= erts_next_timeout_time(esdp->next_tmo_ref)) {
-		erts_runq_unlock(rq);
-		erts_bump_timers(esdp->timer_wheel, esdp->last_monotonic_time);
-		erts_runq_lock(rq);
-	    }
-	}
     }
 
     ERTS_LC_ASSERT(!is_normal_sched || !erts_thr_progress_is_blocking());
@@ -9822,6 +9731,16 @@ Process *erts_schedule(ErtsSchedulerData *esdp, Process *p, int calls)
 	ErtsMigrationPath *mp;
 
 	if (is_normal_sched) {
+
+	    if (esdp->check_time_reds >= ERTS_CHECK_TIME_REDS)
+		(void) erts_get_monotonic_time(esdp);
+
+	    if (esdp->last_monotonic_time >= erts_next_timeout_time(esdp->next_tmo_ref)) {
+		erts_runq_unlock(rq);
+		erts_bump_timers(esdp->timer_wheel, esdp->last_monotonic_time);
+		erts_runq_lock(rq);
+	    }
+
 	    if (rq->check_balance_reds <= 0)
 		check_balance(rq);
 
@@ -9951,37 +9870,6 @@ Process *erts_schedule(ErtsSchedulerData *esdp, Process *p, int calls)
 
 	    goto check_activities_to_run;
 	}
-	else if (is_normal_sched
-		 && (fcalls > input_reductions
-		     && prepare_for_sys_schedule(!0))) {
-	    ErtsMonotonicTime current_time;
-	    /*
-	     * Schedule system-level activities.
-	     */
-
-	    ERTS_MSACC_PUSH_STATE_CACHED_M();
-
-	    esdp->function_calls = 0;
-	    fcalls = 0;
-
-#if 0 /* Not needed since we wont wait in sys schedule */
-	    erts_check_io_interrupt(esdp->pollset, 0);
-#endif
-	    erts_runq_unlock(rq);
-
-            ERTS_MSACC_SET_STATE_CACHED_M(ERTS_MSACC_STATE_CHECK_IO);
-            LTTNG2(scheduler_poll, esdp->no, 1);
-
-	    erl_sys_schedule(1);
-	    ERTS_MSACC_POP_STATE_M();
-
-	    current_time = erts_get_monotonic_time(esdp);
-	    if (current_time >= erts_next_timeout_time(esdp->next_tmo_ref))
-		erts_bump_timers(esdp->timer_wheel, current_time);
-
-	    erts_runq_lock(rq);
-	    goto continue_check_activities_to_run;
-	}
 
         if (flags & ERTS_RUNQ_FLG_MISC_OP)
 	    exec_misc_ops(rq);
diff --git a/erts/emulator/beam/erl_process.h b/erts/emulator/beam/erl_process.h
index b372146846..b4c06adc95 100644
--- a/erts/emulator/beam/erl_process.h
+++ b/erts/emulator/beam/erl_process.h
@@ -369,6 +369,7 @@ struct ErtsSchedulerSleepInfo_ {
     ErtsSchedulerSleepInfo *prev;
     erts_atomic32_t flags;
     erts_tse_t *event;
+    struct erts_poll_thread *psi;
     erts_atomic32_t aux_work;
 };
 
@@ -632,8 +633,6 @@ struct ErtsSchedulerData_ {
     void *match_pseudo_process; /* erl_db_util.c:db_prog_match() */
     Process *free_process;
     ErtsThrPrgrData thr_progress_data;
-    struct pollset_info* pollset;
-    int function_calls;
     ErtsSchedulerSleepInfo *ssi;
     Process *current_process;
     ErtsSchedType type;
@@ -1534,9 +1533,7 @@ extern int erts_system_profile_ts_type;
 void erts_pre_init_process(void);
 void erts_late_init_process(void);
 void erts_early_init_scheduling(int);
-void erts_init_scheduling(int, int
-			  , int, int, int
-			  );
+void erts_init_scheduling(int, int, int, int, int, int);
 void erts_execute_dirty_system_task(Process *c_p);
 int erts_set_gc_state(Process *c_p, int enable);
 Eterm erts_sched_wall_time_request(Process *c_p, int set, int enable,
@@ -2473,6 +2470,7 @@ void erts_notify_inc_runq(ErtsRunQueue *runq);
 
 void erts_sched_finish_poke(ErtsSchedulerSleepInfo *, erts_aint32_t);
 ERTS_GLB_INLINE void erts_sched_poke(ErtsSchedulerSleepInfo *ssi);
+void erts_aux_thread_poke(void);
 
 #if ERTS_GLB_INLINE_INCL_FUNC_DEF
 
diff --git a/erts/emulator/beam/erl_vm.h b/erts/emulator/beam/erl_vm.h
index 82977e62ea..295130f60c 100644
--- a/erts/emulator/beam/erl_vm.h
+++ b/erts/emulator/beam/erl_vm.h
@@ -46,8 +46,6 @@
  */
 #define ERTS_X_REGS_ALLOCATED (MAX_REG+3)
 
-#define INPUT_REDUCTIONS (CONTEXT_REDS / 4)
-
 #define H_DEFAULT_SIZE  233        /* default (heap + stack) min size */
 #define VH_DEFAULT_SIZE  32768     /* default virtual (bin) heap min size (words) */
 #define H_DEFAULT_MAX_SIZE 0       /* default max heap size is off */
diff --git a/erts/emulator/beam/global.h b/erts/emulator/beam/global.h
index c3f8e92f13..09aeba00fa 100644
--- a/erts/emulator/beam/global.h
+++ b/erts/emulator/beam/global.h
@@ -1101,7 +1101,6 @@ void erts_save_stacktrace(Process* p, struct StackTrace* s, int depth);
 typedef struct {
     Eterm delay_time;
     int context_reds;
-    int input_reds;
 } ErtsModifiedTimings;
 
 extern Export *erts_delay_trap;
diff --git a/erts/emulator/beam/sys.h b/erts/emulator/beam/sys.h
index 0bf2ecfc8b..9106091ca6 100644
--- a/erts/emulator/beam/sys.h
+++ b/erts/emulator/beam/sys.h
@@ -992,10 +992,6 @@ erts_refc_read(erts_refc_t *refcp, erts_aint_t min_val)
 
 #endif /* #if ERTS_GLB_INLINE_INCL_FUNC_DEF */
 
-#ifdef ERTS_ENABLE_KERNEL_POLL
-extern int erts_use_kernel_poll;
-#endif
-
 #define sys_memcpy(s1,s2,n)  memcpy(s1,s2,n)
 #define sys_memmove(s1,s2,n) memmove(s1,s2,n)
 #define sys_memcmp(s1,s2,n)  memcmp(s1,s2,n)