Perform potentially long GC on dirty schedulers if available

11 files changed, 563 insertions, 200 deletions

diff --git a/erts/emulator/beam/beam_bif_load.c b/erts/emulator/beam/beam_bif_load.c
index 36bbe81ed5..22576503ef 100644
--- a/erts/emulator/beam/beam_bif_load.c
+++ b/erts/emulator/beam/beam_bif_load.c
@@ -913,7 +913,7 @@ erts_proc_copy_literal_area(Process *c_p, int *redsp, int fcalls, int gc_allowed
 
     la = ERTS_COPY_LITERAL_AREA();
     if (!la)
-	return am_ok;
+        goto return_ok;
 
     oh = la->off_heap;
     literals = (char *) &la->start[0];
@@ -977,6 +977,11 @@ erts_proc_copy_literal_area(Process *c_p, int *redsp, int fcalls, int gc_allowed
 	 * this is not completely certain). We go for
 	 * the GC directly instead of scanning everything
 	 * one more time...
+	 *
+	 * Also note that calling functions expect a
+	 * major GC to be performed if gc_allowed is set
+	 * to true. If you change this, you need to fix
+	 * callers...
 	 */
 	goto literal_gc;
     }
@@ -1051,6 +1056,13 @@ erts_proc_copy_literal_area(Process *c_p, int *redsp, int fcalls, int gc_allowed
 	}
     }
 
+return_ok:
+
+#ifdef ERTS_DIRTY_SCHEDULERS
+    if (ERTS_SCHEDULER_IS_DIRTY(erts_proc_sched_data(c_p)))
+	c_p->flags &= ~F_DIRTY_CLA;
+#endif
+
     return am_ok;
 
 literal_gc:
@@ -1061,13 +1073,13 @@ literal_gc:
     if (c_p->flags & F_DISABLE_GC)
 	return THE_NON_VALUE;
 
-    FLAGS(c_p) |= F_NEED_FULLSWEEP;
-
-    *redsp += erts_garbage_collect_nobump(c_p, 0, c_p->arg_reg, c_p->arity, fcalls);
+    *redsp += erts_garbage_collect_literals(c_p, (Eterm *) literals, lit_bsize,
+					    oh, fcalls);
 
-    erts_garbage_collect_literals(c_p, (Eterm *) literals, lit_bsize, oh);
-
-    *redsp += lit_bsize / 64; /* Need, better value... */
+#ifdef ERTS_DIRTY_SCHEDULERS
+    if (c_p->flags & F_DIRTY_CLA)
+	return THE_NON_VALUE;
+#endif
 
     return am_ok;
 }
diff --git a/erts/emulator/beam/beam_emu.c b/erts/emulator/beam/beam_emu.c
index cfdf61eed0..85d92321b8 100644
--- a/erts/emulator/beam/beam_emu.c
+++ b/erts/emulator/beam/beam_emu.c
@@ -5327,10 +5327,25 @@ void erts_dirty_process_main(ErtsSchedulerData *esdp)
     ASSERT(!(c_p->flags & F_HIPE_MODE));
     ERTS_MSACC_UPDATE_CACHE_X();
 
-    reg = esdp->x_reg_array;
-    {
+    /*
+     * Set fcalls even though we ignore it, so we don't
+     * confuse code accessing it...
+     */
+    if (ERTS_PROC_GET_SAVED_CALLS_BUF(c_p))
+	c_p->fcalls = 0;
+    else
+	c_p->fcalls = CONTEXT_REDS;
+
+    if (erts_smp_atomic32_read_nob(&c_p->state) & ERTS_PSFLG_DIRTY_RUNNING_SYS) {
+	erts_execute_dirty_system_task(c_p);
+	goto do_dirty_schedule;
+    }
+    else {
+	ErtsCodeMFA *codemfa;
 	Eterm* argp;
-	int i;
+	int i, exiting;
+
+	reg = esdp->x_reg_array;
 
 	argp = c_p->arg_reg;
 	for (i = c_p->arity - 1; i >= 0; i--) {
@@ -5346,15 +5361,6 @@ void erts_dirty_process_main(ErtsSchedulerData *esdp)
 
 	I = c_p->i;
 
-	/*
-	 * Set fcalls even though we ignore it, so we don't
-	 * confuse code accessing it...
-	 */
-	if (ERTS_PROC_GET_SAVED_CALLS_BUF(c_p))
-	    c_p->fcalls = 0;
-	else
-	    c_p->fcalls = CONTEXT_REDS;
-
 	SWAPIN;
 
 #ifdef USE_VM_PROBES
@@ -5378,62 +5384,55 @@ void erts_dirty_process_main(ErtsSchedulerData *esdp)
             DTRACE2(process_scheduled, process_buf, fun_buf);
         }
 #endif
-    }
 
-    {
-	int exiting;
-
-	{
-	    /*
-	     * call_nif is always first instruction in function:
-	     *
-	     * I[-3]: Module
-	     * I[-2]: Function
-	     * I[-1]: Arity
-	     * I[0]: &&call_nif
-	     * I[1]: Function pointer to NIF function
-	     * I[2]: Pointer to erl_module_nif
-	     * I[3]: Function pointer to dirty NIF
-             *
-             * This layout is determined by the NifExport struct
-	     */
-            ErtsCodeMFA *codemfa;
+	/*
+	 * call_nif is always first instruction in function:
+	 *
+	 * I[-3]: Module
+	 * I[-2]: Function
+	 * I[-1]: Arity
+	 * I[0]: &&call_nif
+	 * I[1]: Function pointer to NIF function
+	 * I[2]: Pointer to erl_module_nif
+	 * I[3]: Function pointer to dirty NIF
+	 *
+	 * This layout is determined by the NifExport struct
+	 */
 
-	    ERTS_MSACC_SET_STATE_CACHED_M_X(ERTS_MSACC_STATE_NIF);
+	ERTS_MSACC_SET_STATE_CACHED_M_X(ERTS_MSACC_STATE_NIF);
 
-            codemfa = erts_code_to_codemfa(I);
+	codemfa = erts_code_to_codemfa(I);
 
-	    DTRACE_NIF_ENTRY(c_p, codemfa);
-	    c_p->current = codemfa;
-	    SWAPOUT;
-	    PROCESS_MAIN_CHK_LOCKS(c_p);
-	    ERTS_SMP_UNREQ_PROC_MAIN_LOCK(c_p);
+	DTRACE_NIF_ENTRY(c_p, codemfa);
+	c_p->current = codemfa;
+	SWAPOUT;
+	PROCESS_MAIN_CHK_LOCKS(c_p);
+	ERTS_SMP_UNREQ_PROC_MAIN_LOCK(c_p);
 
-	    ASSERT(!ERTS_PROC_IS_EXITING(c_p));
-	    if (em_apply_bif == (BeamInstr *) *I) {
-		exiting = erts_call_dirty_bif(esdp, c_p, I, reg);
-	    }
-	    else {
-		ASSERT(em_call_nif == (BeamInstr *) *I);
-		exiting = erts_call_dirty_nif(esdp, c_p, I, reg);
-	    }
+	ASSERT(!ERTS_PROC_IS_EXITING(c_p));
+	if (em_apply_bif == (BeamInstr *) *I) {
+	    exiting = erts_call_dirty_bif(esdp, c_p, I, reg);
+	}
+	else {
+	    ASSERT(em_call_nif == (BeamInstr *) *I);
+	    exiting = erts_call_dirty_nif(esdp, c_p, I, reg);
+	}
 
-	    ASSERT(!(c_p->flags & F_HIBERNATE_SCHED));
+	ASSERT(!(c_p->flags & F_HIBERNATE_SCHED));
 
-	    PROCESS_MAIN_CHK_LOCKS(c_p);
-	    ERTS_SMP_REQ_PROC_MAIN_LOCK(c_p);
-	    ERTS_VERIFY_UNUSED_TEMP_ALLOC(c_p);
-	    ERTS_MSACC_SET_STATE_CACHED_M_X(ERTS_MSACC_STATE_EMULATOR);
-	    if (exiting)
-		goto do_dirty_schedule;
-	    ASSERT(!ERTS_PROC_IS_EXITING(c_p));
+	PROCESS_MAIN_CHK_LOCKS(c_p);
+	ERTS_SMP_REQ_PROC_MAIN_LOCK(c_p);
+	ERTS_VERIFY_UNUSED_TEMP_ALLOC(c_p);
+	ERTS_MSACC_SET_STATE_CACHED_M_X(ERTS_MSACC_STATE_EMULATOR);
+	if (exiting)
+	    goto do_dirty_schedule;
+	ASSERT(!ERTS_PROC_IS_EXITING(c_p));
 
-	    DTRACE_NIF_RETURN(c_p, codemfa);
-	    ERTS_HOLE_CHECK(c_p);
-	    SWAPIN;
-	    I = c_p->i;
-	    goto context_switch;
-	}
+	DTRACE_NIF_RETURN(c_p, codemfa);
+	ERTS_HOLE_CHECK(c_p);
+	SWAPIN;
+	I = c_p->i;
+	goto context_switch;
     }
 #endif /* ERTS_DIRTY_SCHEDULERS */
 }
diff --git a/erts/emulator/beam/bif.c b/erts/emulator/beam/bif.c
index cae346267c..a5b66db35c 100644
--- a/erts/emulator/beam/bif.c
+++ b/erts/emulator/beam/bif.c
@@ -3865,13 +3865,6 @@ BIF_RETTYPE now_0(BIF_ALIST_0)
 
 /**********************************************************************/
 
-BIF_RETTYPE garbage_collect_0(BIF_ALIST_0)
-{
-    FLAGS(BIF_P) |= F_NEED_FULLSWEEP;
-    erts_garbage_collect(BIF_P, 0, NULL, 0);
-    return am_true;
-}
-
 /*
  * Pass atom 'minor' for relaxed generational GC run. This is only
  * recommendation, major run may still be chosen by VM.
@@ -4324,7 +4317,7 @@ BIF_RETTYPE group_leader_2(BIF_ALIST_2)
 		erts_smp_proc_unlock(new_member, ERTS_PROC_LOCK_STATUS);
 		if (new_member == BIF_P
 		    || !(erts_smp_atomic32_read_nob(&new_member->state)
-			 & (ERTS_PSFLG_DIRTY_RUNNING|ERTS_PSFLG_DIRTY_RUNNING_SYS))) {
+			 & ERTS_PSFLG_DIRTY_RUNNING)) {
 		    new_member->group_leader = STORE_NC_IN_PROC(new_member,
 								BIF_ARG_1);
 		}
diff --git a/erts/emulator/beam/bif.tab b/erts/emulator/beam/bif.tab
index f461da913e..47fdcfa7a4 100644
--- a/erts/emulator/beam/bif.tab
+++ b/erts/emulator/beam/bif.tab
@@ -68,7 +68,6 @@ gcbif erlang:float/1
 bif erlang:float_to_list/1
 bif erlang:float_to_list/2
 bif erlang:fun_info/2
-bif erlang:garbage_collect/0
 bif erts_internal:garbage_collect/1
 bif erlang:get/0
 bif erlang:get/1
diff --git a/erts/emulator/beam/erl_gc.c b/erts/emulator/beam/erl_gc.c
index 6093b0cf39..5b299d9acf 100644
--- a/erts/emulator/beam/erl_gc.c
+++ b/erts/emulator/beam/erl_gc.c
@@ -118,11 +118,14 @@ static Eterm *full_sweep_heaps(Process *p,
 			       char *oh, Uint oh_size,
 			       Eterm *objv, int nobj);
 static int garbage_collect(Process* p, ErlHeapFragment *live_hf_end,
-			   int need, Eterm* objv, int nobj, int fcalls);
+			   int need, Eterm* objv, int nobj, int fcalls,
+			   Uint max_young_gen_usage);
 static int major_collection(Process* p, ErlHeapFragment *live_hf_end,
-			    int need, Eterm* objv, int nobj, Uint *recl);
+			    int need, Eterm* objv, int nobj,
+			    Uint ygen_usage, Uint *recl);
 static int minor_collection(Process* p, ErlHeapFragment *live_hf_end,
-			    int need, Eterm* objv, int nobj, Uint *recl);
+			    int need, Eterm* objv, int nobj,
+			    Uint ygen_usage, Uint *recl);
 static void do_minor(Process *p, ErlHeapFragment *live_hf_end,
 		     char *mature, Uint mature_size,
 		     Uint new_sz, Eterm* objv, int nobj);
@@ -413,15 +416,15 @@ erts_gc_after_bif_call_lhf(Process* p, ErlHeapFragment *live_hf_end,
 		regs = erts_proc_sched_data(p)->x_reg_array;
 	    }
 #endif
-	    cost = garbage_collect(p, live_hf_end, 0, regs, p->arity, p->fcalls);
+	    cost = garbage_collect(p, live_hf_end, 0, regs, p->arity, p->fcalls, 0);
 	} else {
-	    cost = garbage_collect(p, live_hf_end, 0, regs, arity, p->fcalls);
+	    cost = garbage_collect(p, live_hf_end, 0, regs, arity, p->fcalls, 0);
 	}
     } else {
 	Eterm val[1];
 
 	val[0] = result;
-	cost = garbage_collect(p, live_hf_end, 0, val, 1, p->fcalls);
+	cost = garbage_collect(p, live_hf_end, 0, val, 1, p->fcalls, 0);
 	result = val[0];
     }
     BUMP_REDS(p, cost);
@@ -599,6 +602,32 @@ young_gen_usage(Process *p)
 	}							\
     } while (0)
 
+#ifdef ERTS_DIRTY_SCHEDULERS
+
+static ERTS_INLINE void
+check_for_possibly_long_gc(Process *p, Uint ygen_usage)
+{
+    int major;
+    Uint sz;
+
+    major = (p->flags & F_NEED_FULLSWEEP) || GEN_GCS(p) >= MAX_GEN_GCS(p);
+
+    sz = ygen_usage;
+    sz += p->hend - p->stop;
+    if (p->flags & F_ON_HEAP_MSGQ)
+        sz += p->msg.len;
+    if (major)
+	sz += p->old_htop - p->old_heap;
+
+    if (sz >= ERTS_POTENTIALLY_LONG_GC_HSIZE) {
+        ASSERT(!(p->flags & (F_DISABLE_GC|F_DELAY_GC)));
+	p->flags |= major ? F_DIRTY_MAJOR_GC : F_DIRTY_MINOR_GC;
+        erts_schedule_dirty_sys_execution(p);
+    }
+}
+
+#endif
+
 /*
  * Garbage collect a process.
  *
@@ -609,13 +638,15 @@ young_gen_usage(Process *p)
  */
 static int
 garbage_collect(Process* p, ErlHeapFragment *live_hf_end,
-		int need, Eterm* objv, int nobj, int fcalls)
+		int need, Eterm* objv, int nobj, int fcalls,
+		Uint max_young_gen_usage)
 {
     Uint reclaimed_now = 0;
+    Uint ygen_usage;
     Eterm gc_trace_end_tag;
     int reds;
     ErtsMonotonicTime start_time = 0; /* Shut up faulty warning... */
-    ErtsSchedulerData *esdp;
+    ErtsSchedulerData *esdp = erts_proc_sched_data(p);
     erts_aint32_t state;
     ERTS_MSACC_PUSH_STATE_M();
 #ifdef USE_VM_PROBES
@@ -624,13 +655,26 @@ garbage_collect(Process* p, ErlHeapFragment *live_hf_end,
 
     ERTS_CHK_MBUF_SZ(p);
 
-    ASSERT(CONTEXT_REDS - ERTS_REDS_LEFT(p, fcalls)
-	   >= erts_proc_sched_data(p)->virtual_reds);
+    ASSERT(CONTEXT_REDS - ERTS_REDS_LEFT(p, fcalls) >= esdp->virtual_reds);
 
     state = erts_smp_atomic32_read_nob(&p->state);
 
-    if (p->flags & (F_DISABLE_GC|F_DELAY_GC) || state & ERTS_PSFLG_EXITING)
+    if ((p->flags & (F_DISABLE_GC|F_DELAY_GC)) || state & ERTS_PSFLG_EXITING) {
+#ifdef ERTS_DIRTY_SCHEDULERS
+    delay_gc_before_start:
+#endif
 	return delay_garbage_collection(p, live_hf_end, need, fcalls);
+    }
+
+    ygen_usage = max_young_gen_usage ? max_young_gen_usage : young_gen_usage(p);
+
+#ifdef ERTS_DIRTY_SCHEDULERS
+    if (!ERTS_SCHEDULER_IS_DIRTY(esdp)) {
+	check_for_possibly_long_gc(p, ygen_usage);
+	if (p->flags & (F_DIRTY_MAJOR_GC|F_DIRTY_MINOR_GC))
+	    goto delay_gc_before_start;
+    }
+#endif
 
     if (p->abandoned_heap)
 	live_hf_end = ERTS_INVALID_HFRAG_PTR;
@@ -639,8 +683,6 @@ garbage_collect(Process* p, ErlHeapFragment *live_hf_end,
 
     ERTS_MSACC_SET_STATE_CACHED_M(ERTS_MSACC_STATE_GC);
 
-    esdp = erts_get_scheduler_data();
-
     erts_smp_atomic32_read_bor_nob(&p->state, ERTS_PSFLG_GC);
     if (erts_system_monitor_long_gc != 0)
 	start_time = erts_get_monotonic_time(esdp);
@@ -667,14 +709,25 @@ garbage_collect(Process* p, ErlHeapFragment *live_hf_end,
             trace_gc(p, am_gc_minor_start, need, THE_NON_VALUE);
         }
         DTRACE2(gc_minor_start, pidbuf, need);
-        reds = minor_collection(p, live_hf_end, need, objv, nobj, &reclaimed_now);
+        reds = minor_collection(p, live_hf_end, need, objv, nobj,
+				ygen_usage, &reclaimed_now);
         DTRACE2(gc_minor_end, pidbuf, reclaimed_now);
         if (reds == -1) {
             if (IS_TRACED_FL(p, F_TRACE_GC)) {
                 trace_gc(p, am_gc_minor_end, reclaimed_now, THE_NON_VALUE);
             }
+#ifdef ERTS_DIRTY_SCHEDULERS
+	    if (!ERTS_SCHEDULER_IS_DIRTY(esdp)) {
+		p->flags |= F_NEED_FULLSWEEP;
+		check_for_possibly_long_gc(p, ygen_usage);
+		if (p->flags & F_DIRTY_MAJOR_GC)
+		    goto delay_gc_after_start;
+	    }
+#endif
             goto do_major_collection;
         }
+        if (ERTS_SCHEDULER_IS_DIRTY(esdp))
+            p->flags &= ~F_DIRTY_MINOR_GC;
         gc_trace_end_tag = am_gc_minor_end;
     } else {
 do_major_collection:
@@ -683,7 +736,10 @@ do_major_collection:
             trace_gc(p, am_gc_major_start, need, THE_NON_VALUE);
         }
         DTRACE2(gc_major_start, pidbuf, need);
-        reds = major_collection(p, live_hf_end, need, objv, nobj, &reclaimed_now);
+        reds = major_collection(p, live_hf_end, need, objv, nobj,
+				ygen_usage, &reclaimed_now);
+        if (ERTS_SCHEDULER_IS_DIRTY(esdp))
+            p->flags &= ~(F_DIRTY_MAJOR_GC|F_DIRTY_MINOR_GC);
         DTRACE2(gc_major_end, pidbuf, reclaimed_now);
         gc_trace_end_tag = am_gc_major_end;
         ERTS_MSACC_SET_STATE_CACHED_M_X(ERTS_MSACC_STATE_GC);
@@ -713,6 +769,9 @@ do_major_collection:
                               am_kill, NIL, NULL, 0);
         erts_smp_proc_unlock(p, locks & ERTS_PROC_LOCKS_ALL_MINOR);
 
+#ifdef ERTS_DIRTY_SCHEDULERS
+    delay_gc_after_start:
+#endif
         /* erts_send_exit_signal looks for ERTS_PSFLG_GC, so
            we have to remove it after the signal is sent */
         erts_smp_atomic32_read_band_nob(&p->state, ~ERTS_PSFLG_GC);
@@ -795,7 +854,7 @@ do_major_collection:
 int
 erts_garbage_collect_nobump(Process* p, int need, Eterm* objv, int nobj, int fcalls)
 {
-    int reds = garbage_collect(p, ERTS_INVALID_HFRAG_PTR, need, objv, nobj, fcalls);
+    int reds = garbage_collect(p, ERTS_INVALID_HFRAG_PTR, need, objv, nobj, fcalls, 0);
     int reds_left = ERTS_REDS_LEFT(p, fcalls);
     if (reds > reds_left)
 	reds = reds_left;
@@ -806,7 +865,7 @@ erts_garbage_collect_nobump(Process* p, int need, Eterm* objv, int nobj, int fca
 void
 erts_garbage_collect(Process* p, int need, Eterm* objv, int nobj)
 {
-    int reds = garbage_collect(p, ERTS_INVALID_HFRAG_PTR, need, objv, nobj, p->fcalls);
+    int reds = garbage_collect(p, ERTS_INVALID_HFRAG_PTR, need, objv, nobj, p->fcalls, 0);
     BUMP_REDS(p, reds);
     ASSERT(CONTEXT_REDS - ERTS_BIF_REDS_LEFT(p)
 	   >= erts_proc_sched_data(p)->virtual_reds);
@@ -832,6 +891,20 @@ erts_garbage_collect_hibernate(Process* p)
     if (p->flags & F_DISABLE_GC)
 	ERTS_INTERNAL_ERROR("GC disabled");
 
+#ifdef ERTS_DIRTY_SCHEDULERS
+    if (ERTS_SCHEDULER_IS_DIRTY(erts_proc_sched_data(p)))
+        p->flags &= ~(F_DIRTY_GC_HIBERNATE|F_DIRTY_MAJOR_GC|F_DIRTY_MINOR_GC);
+    else {
+	Uint flags = p->flags;
+	p->flags |= F_NEED_FULLSWEEP;
+	check_for_possibly_long_gc(p, (p->htop - p->heap) + p->mbuf_sz);
+	if (p->flags & (F_DIRTY_MAJOR_GC|F_DIRTY_MINOR_GC)) {
+	    p->flags = flags|F_DIRTY_GC_HIBERNATE;
+	    return;
+	}
+	p->flags = flags;
+    }
+#endif
     /*
      * Preliminaries.
      */
@@ -843,7 +916,6 @@ erts_garbage_collect_hibernate(Process* p)
      * Do it.
      */
 
-
     heap_size = p->heap_sz + (p->old_htop - p->old_heap) + p->mbuf_sz;
 
     heap = (Eterm*) ERTS_HEAP_ALLOC(ERTS_ALC_T_TMP_HEAP,
@@ -984,10 +1056,11 @@ static ERTS_INLINE void offset_nstack(Process* p, Sint offs,
 
 #endif /* HIPE */
 
-void
+int
 erts_garbage_collect_literals(Process* p, Eterm* literals,
 			      Uint byte_lit_size,
-			      struct erl_off_heap_header* oh)
+			      struct erl_off_heap_header* oh,
+			      int fcalls)
 {
     Uint lit_size = byte_lit_size / sizeof(Eterm);
     Uint old_heap_size;
@@ -999,20 +1072,49 @@ erts_garbage_collect_literals(Process* p, Eterm* literals,
     Uint area_size;
     Eterm* old_htop;
     Uint n;
+    Uint ygen_usage = 0;
     struct erl_off_heap_header** prev = NULL;
+    Sint64 reds;
+
+    if (p->flags & (F_DISABLE_GC|F_DELAY_GC))
+	ERTS_INTERNAL_ERROR("GC disabled");
+
+    /*
+     * First an ordinary major collection...
+     */
+
+    p->flags |= F_NEED_FULLSWEEP;
+
+#ifdef ERTS_DIRTY_SCHEDULERS
+    if (ERTS_SCHEDULER_IS_DIRTY(erts_proc_sched_data(p)))
+	p->flags &= ~F_DIRTY_CLA;
+    else {
+	ygen_usage = young_gen_usage(p);
+	check_for_possibly_long_gc(p,
+				   (byte_lit_size/sizeof(Uint)
+				    + 2*ygen_usage));
+	if (p->flags & F_DIRTY_MAJOR_GC) {
+	    p->flags |= F_DIRTY_CLA;
+	    return 10;
+	}
+    }
+#endif
+
+    reds = (Sint64) garbage_collect(p, ERTS_INVALID_HFRAG_PTR, 0,
+				    p->arg_reg, p->arity, fcalls,
+				    ygen_usage);
+
+    ASSERT(!(p->flags & (F_DIRTY_MAJOR_GC|F_DIRTY_MINOR_GC)));
 
-    if (p->flags & F_DISABLE_GC)
-	return;
     /*
      * Set GC state.
      */
     erts_smp_atomic32_read_bor_nob(&p->state, ERTS_PSFLG_GC);
 
     /*
-     * We assume that the caller has already done a major collection
-     * (which has discarded the old heap), so that we don't have to cope
-     * with pointer to literals on the old heap. We will now allocate
-     * an old heap to contain the literals.
+     * Just did a major collection (which has discarded the old heap),
+     * so that we don't have to cope with pointer to literals on the
+     * old heap. We will now allocate an old heap to contain the literals.
      */
     
     ASSERT(p->old_heap == 0);	/* Must NOT have an old heap yet. */
@@ -1155,15 +1257,21 @@ erts_garbage_collect_literals(Process* p, Eterm* literals,
      * Restore status.
      */
     erts_smp_atomic32_read_band_nob(&p->state, ~ERTS_PSFLG_GC);
+
+    reds += (Sint64) gc_cost((p->htop - p->heap) + byte_lit_size/sizeof(Uint), 0);
+    if (reds > INT_MAX)
+	return INT_MAX;
+    return (int) reds;
 }
 
 static int
 minor_collection(Process* p, ErlHeapFragment *live_hf_end,
-		 int need, Eterm* objv, int nobj, Uint *recl)
+		 int need, Eterm* objv, int nobj,
+		 Uint ygen_usage, Uint *recl)
 {
     Eterm *mature = p->abandoned_heap ? p->abandoned_heap : p->heap;
     Uint mature_size = p->high_water - mature;
-    Uint size_before = young_gen_usage(p);
+    Uint size_before = ygen_usage;
 
     /*
      * Check if we have gone past the max heap size limit
@@ -1534,7 +1642,8 @@ do_minor(Process *p, ErlHeapFragment *live_hf_end,
 
 static int
 major_collection(Process* p, ErlHeapFragment *live_hf_end,
-		 int need, Eterm* objv, int nobj, Uint *recl)
+		 int need, Eterm* objv, int nobj,
+		 Uint ygen_usage, Uint *recl)
 {
     Uint size_before, size_after, stack_size;
     Eterm* n_heap;
@@ -1552,7 +1661,7 @@ major_collection(Process* p, ErlHeapFragment *live_hf_end,
      * to receive all live data.
      */
 
-    size_before = young_gen_usage(p);
+    size_before = ygen_usage;
     size_before += p->old_htop - p->old_heap;
     stack_size = p->hend - p->stop;
 
diff --git a/erts/emulator/beam/erl_gc.h b/erts/emulator/beam/erl_gc.h
index 54ea9ca3c0..0dfdf52028 100644
--- a/erts/emulator/beam/erl_gc.h
+++ b/erts/emulator/beam/erl_gc.h
@@ -25,6 +25,8 @@
 
 /* GC declarations shared by beam/erl_gc.c and hipe/hipe_gc.c */
 
+#define ERTS_POTENTIALLY_LONG_GC_HSIZE (128*1024) /* Words */
+
 #include "erl_map.h"
 
 #if defined(DEBUG) && !ERTS_GLB_INLINE_INCL_FUNC_DEF
@@ -145,9 +147,10 @@ void erts_garbage_collect_hibernate(struct process* p);
 Eterm erts_gc_after_bif_call_lhf(struct process* p, ErlHeapFragment *live_hf_end,
 				 Eterm result, Eterm* regs, Uint arity);
 Eterm erts_gc_after_bif_call(struct process* p, Eterm result, Eterm* regs, Uint arity);
-void erts_garbage_collect_literals(struct process* p, Eterm* literals,
-				   Uint lit_size,
-				   struct erl_off_heap_header* oh);
+int erts_garbage_collect_literals(struct process* p, Eterm* literals,
+				  Uint lit_size,
+				  struct erl_off_heap_header* oh,
+				  int fcalls);
 Uint erts_next_heap_size(Uint, Uint);
 Eterm erts_heap_sizes(struct process* p);
 
diff --git a/erts/emulator/beam/erl_nfunc_sched.h b/erts/emulator/beam/erl_nfunc_sched.h
index 5c98957cb7..55a3a6dbf6 100644
--- a/erts/emulator/beam/erl_nfunc_sched.h
+++ b/erts/emulator/beam/erl_nfunc_sched.h
@@ -54,7 +54,6 @@ typedef struct {
     /* --- The following is only used on error --- */
     BeamInstr *pc;	/* Program counter */
     BeamInstr *cp;	/* Continuation pointer */
-    void *nif;		/* Original NIF/BIF call */
     ErtsCodeMFA *mfa;	/* MFA of original call */
     int argc;		/* Number of arguments in original call */
     int argv_size;	/* Allocated size of argv */
diff --git a/erts/emulator/beam/erl_process.c b/erts/emulator/beam/erl_process.c
index 955b98b35b..da4b468d66 100644
--- a/erts/emulator/beam/erl_process.c
+++ b/erts/emulator/beam/erl_process.c
@@ -6439,10 +6439,30 @@ schedule_out_process(ErtsRunQueue *c_rq, erts_aint32_t state, Process *p,
     int enqueue; /* < 0 -> use proxy */
     ErtsRunQueue* runq;
 
-    if (is_normal_sched)
-	running_flgs = ERTS_PSFLG_RUNNING|ERTS_PSFLG_RUNNING_SYS;
-    else
+    if (!is_normal_sched)
 	running_flgs = ERTS_PSFLG_DIRTY_RUNNING|ERTS_PSFLG_DIRTY_RUNNING_SYS;
+    else {
+	running_flgs = ERTS_PSFLG_RUNNING|ERTS_PSFLG_RUNNING_SYS;
+#ifdef ERTS_DIRTY_SCHEDULERS
+        if (state & ERTS_PSFLG_DIRTY_ACTIVE_SYS
+            && (p->flags & (F_DELAY_GC|F_DISABLE_GC))) {
+            /*
+             * Delay dirty GC; will be enabled automatically
+             * again by next GC...
+             */
+
+            /*
+             * No normal execution until dirty CLA or hibernat has
+             * been handled...
+             */
+            ASSERT(!(p->flags & (F_DIRTY_CLA | F_DIRTY_GC_HIBERNATE)));
+
+            state = erts_smp_atomic32_read_band_nob(&p->state,
+                                                    ~ERTS_PSFLG_DIRTY_ACTIVE_SYS);
+            state &= ~ERTS_PSFLG_DIRTY_ACTIVE_SYS;
+        }
+#endif
+    }
 
     a = state;
 
@@ -8642,8 +8662,7 @@ pid2proc_not_running(Process *c_p, ErtsProcLocks c_p_locks,
 		erts_aint32_t state;
 		state = erts_smp_atomic32_read_nob(&rp->state);
 		ASSERT((state & ERTS_PSFLG_PENDING_EXIT)
-		       || !(state & (ERTS_PSFLG_RUNNING
-				     | ERTS_PSFLG_DIRTY_RUNNING_SYS)));
+		       || !(state & ERTS_PSFLG_RUNNING));
 	    }
 #endif
 
@@ -9981,6 +10000,7 @@ Process *erts_schedule(ErtsSchedulerData *esdp, Process *p, int calls)
 
 					      | ERTS_PSFLG_FREE
 					      | ERTS_PSFLG_RUNNING_SYS
+					      | ERTS_PSFLG_DIRTY_RUNNING_SYS
 					      | ERTS_PSFLG_EXITING))
 				    == ERTS_PSFLG_EXITING)
 				   & (!!is_normal_sched))
@@ -9992,7 +10012,14 @@ Process *erts_schedule(ErtsSchedulerData *esdp, Process *p, int calls)
 					    | ERTS_PSFLG_PENDING_EXIT
 					    | ERTS_PSFLG_ACTIVE_SYS
 					    | ERTS_PSFLG_DIRTY_ACTIVE_SYS))
-				  != ERTS_PSFLG_SUSPENDED));
+				  != ERTS_PSFLG_SUSPENDED)
+#ifdef ERTS_DIRTY_SCHEDULERS
+                               & (!(state & (ERTS_PSFLG_EXITING
+                                             | ERTS_PSFLG_PENDING_EXIT))
+                                  | (!!is_normal_sched))
+#endif
+                    );
+
 		if (run_process) {
 		    if (state & (ERTS_PSFLG_ACTIVE_SYS
 				 | ERTS_PSFLG_DIRTY_ACTIVE_SYS))
@@ -10133,65 +10160,70 @@ Process *erts_schedule(ErtsSchedulerData *esdp, Process *p, int calls)
 	    }
 	}
 
-	if (state & (ERTS_PSFLG_RUNNING_SYS
-		     | ERTS_PSFLG_DIRTY_RUNNING_SYS)) {
-	    /*
-	     * GC is normally never delayed when a process
-	     * is scheduled out, but might be when executing
-	     * hand written beam assembly in
-	     * prim_eval:'receive'. If GC is delayed we are
-	     * not allowed to execute system tasks.
-	     */
-	    if (!(p->flags & F_DELAY_GC)) {
-		int cost = execute_sys_tasks(p, &state, reds);
-		calls += cost;
-		reds -= cost;
-		if (reds <= 0
+        if (is_normal_sched) {
+
+            if (state & ERTS_PSFLG_RUNNING_SYS) {
+                /*
+                 * GC is normally never delayed when a process
+                 * is scheduled out, but might be when executing
+                 * hand written beam assembly in
+                 * prim_eval:'receive'. If GC is delayed we are
+                 * not allowed to execute system tasks.
+                 */
+                if (!(p->flags & F_DELAY_GC)) {
+                    int cost = execute_sys_tasks(p, &state, reds);
+                    calls += cost;
+                    reds -= cost;
+                    if (reds <= 0)
+                        goto sched_out_proc;
 #ifdef ERTS_DIRTY_SCHEDULERS
-		    || !is_normal_sched
-		    || (state & ERTS_PSFLGS_DIRTY_WORK)
+                    if (state & ERTS_PSFLGS_DIRTY_WORK)
+                        goto sched_out_proc;
 #endif
-		    ) {
-		    goto sched_out_proc;
-		}
-	    }
+                }
 
-	    ASSERT(state & psflg_running_sys);
-	    ASSERT(!(state & psflg_running));
+                ASSERT(state & psflg_running_sys);
+                ASSERT(!(state & psflg_running));
 
-	    while (1) {
-		erts_aint32_t n, e;
+                while (1) {
+                    erts_aint32_t n, e;
 
-		if (((state & (ERTS_PSFLG_SUSPENDED
-			       | ERTS_PSFLG_ACTIVE)) != ERTS_PSFLG_ACTIVE)
-		    && !(state & ERTS_PSFLG_EXITING)) {
-		    goto sched_out_proc;
-		}
+                    if (((state & (ERTS_PSFLG_SUSPENDED
+                                   | ERTS_PSFLG_ACTIVE)) != ERTS_PSFLG_ACTIVE)
+                        && !(state & ERTS_PSFLG_EXITING)) {
+                        goto sched_out_proc;
+                    }
 
-		n = e = state;
-		n &= ~psflg_running_sys;
-		n |= psflg_running;
+                    n = e = state;
+                    n &= ~psflg_running_sys;
+                    n |= psflg_running;
 
-		state = erts_smp_atomic32_cmpxchg_mb(&p->state, n, e);
-		if (state == e) {
-		    state = n;
-		    break;
-		}
+                    state = erts_smp_atomic32_cmpxchg_mb(&p->state, n, e);
+                    if (state == e) {
+                        state = n;
+                        break;
+                    }
 
-		ASSERT(state & psflg_running_sys);
-		ASSERT(!(state & psflg_running));
-	    }
-	}
+                    ASSERT(state & psflg_running_sys);
+                    ASSERT(!(state & psflg_running));
+                }
+            }
 
-	if (ERTS_IS_GC_DESIRED(p) && !ERTS_SCHEDULER_IS_DIRTY_IO(esdp)) {
-	    if (!(state & ERTS_PSFLG_EXITING) && !(p->flags & (F_DELAY_GC|F_DISABLE_GC))) {
-		int cost = scheduler_gc_proc(p, reds);
-		calls += cost;
-		reds -= cost;
-		if (reds <= 0)
-		    goto sched_out_proc;
-	    }
-	}
+            if (ERTS_IS_GC_DESIRED(p)) {
+                if (!(state & ERTS_PSFLG_EXITING)
+                    && !(p->flags & (F_DELAY_GC|F_DISABLE_GC))) {
+                    int cost = scheduler_gc_proc(p, reds);
+                    calls += cost;
+                    reds -= cost;
+                    if (reds <= 0)
+                        goto sched_out_proc;
+#ifdef ERTS_DIRTY_SCHEDULERS
+                    if (p->flags & (F_DIRTY_MAJOR_GC|F_DIRTY_MINOR_GC))
+                        goto sched_out_proc;
+#endif
+                }
+            }
+        }
 
 	if (proxy_p) {
 	    free_proxy_proc(proxy_p);
@@ -10203,7 +10235,13 @@ Process *erts_schedule(ErtsSchedulerData *esdp, Process *p, int calls)
 	ERTS_SMP_CHK_HAVE_ONLY_MAIN_PROC_LOCK(p);
 
 	/* Never run a suspended process */
-	ASSERT(!(ERTS_PSFLG_SUSPENDED & erts_smp_atomic32_read_nob(&p->state)));
+#ifdef DEBUG
+        {
+            erts_aint32_t dstate = erts_smp_atomic32_read_nob(&p->state);
+            ASSERT(!(ERTS_PSFLG_SUSPENDED & dstate)
+                   || (ERTS_PSFLG_DIRTY_RUNNING_SYS & dstate));
+        }
+#endif
 
 	ASSERT(erts_proc_read_refc(p) > 0);
 
@@ -10224,9 +10262,18 @@ Process *erts_schedule(ErtsSchedulerData *esdp, Process *p, int calls)
 }
 
 static int
-notify_sys_task_executed(Process *c_p, ErtsProcSysTask *st, Eterm st_result)
+notify_sys_task_executed(Process *c_p, ErtsProcSysTask *st,
+			 Eterm st_result, int normal_sched)
 {
-    Process *rp = erts_proc_lookup(st->requester);
+    Process *rp;
+#ifdef ERTS_DIRTY_SCHEDULERS
+    if (!normal_sched)
+	rp = erts_pid2proc_opt(c_p, ERTS_PROC_LOCK_MAIN,
+			       st->requester, 0,
+			       ERTS_P2P_FLG_INC_REFC);
+    else
+#endif
+	rp = erts_proc_lookup(st->requester);
     if (rp) {
 	ErtsProcLocks rp_locks;
 	ErlOffHeap *ohp;
@@ -10274,6 +10321,11 @@ notify_sys_task_executed(Process *c_p, ErtsProcSysTask *st, Eterm st_result)
 
 	if (rp_locks)
 	    erts_smp_proc_unlock(rp, rp_locks);
+
+#ifdef ERTS_DIRTY_SCHEDULERS
+	if (!normal_sched)
+	    erts_proc_dec_refc(rp);
+#endif
     }
 
     erts_cleanup_offheap(&st->off_heap);
@@ -10429,18 +10481,23 @@ done:
 }
 
 static void save_gc_task(Process *c_p, ErtsProcSysTask *st, int prio);
+#ifdef ERTS_DIRTY_SCHEDULERS
+static void save_dirty_task(Process *c_p, ErtsProcSysTask *st);
+#endif
 
 static int
 execute_sys_tasks(Process *c_p, erts_aint32_t *statep, int in_reds)
 {
-    int garbage_collected = 0;
+    int minor_gc = 0, major_gc = 0;
     erts_aint32_t state = *statep;
     int reds = in_reds;
     int qmask = 0;
 
+    ASSERT(!ERTS_SCHEDULER_IS_DIRTY(erts_proc_sched_data(c_p)));
     ERTS_SMP_LC_ASSERT(erts_proc_lc_my_proc_locks(c_p) == ERTS_PROC_LOCK_MAIN);
 
     do {
+	ErtsProcSysTaskType type;
 	ErtsProcSysTask *st;
 	int st_prio;
 	Eterm st_res;
@@ -10458,7 +10515,9 @@ execute_sys_tasks(Process *c_p, erts_aint32_t *statep, int in_reds)
 	if (!st)
 	    break;
 
-	switch (st->type) {
+	type = st->type;
+
+	switch (type) {
         case ERTS_PSTT_GC_MAJOR:
         case ERTS_PSTT_GC_MINOR:
 	    if (c_p->flags & F_DISABLE_GC) {
@@ -10467,12 +10526,23 @@ execute_sys_tasks(Process *c_p, erts_aint32_t *statep, int in_reds)
 		reds--;
 	    }
 	    else {
-		if (!garbage_collected) {
-                    if (st->type == ERTS_PSTT_GC_MAJOR) {
+		if (!minor_gc
+		    || (!major_gc && type == ERTS_PSTT_GC_MAJOR)) {
+                    if (type == ERTS_PSTT_GC_MAJOR) {
                         FLAGS(c_p) |= F_NEED_FULLSWEEP;
                     }
 		    reds -= scheduler_gc_proc(c_p, reds);
-		    garbage_collected = 1;
+#ifdef ERTS_DIRTY_SCHEDULERS
+		    if (c_p->flags & (F_DIRTY_MAJOR_GC|F_DIRTY_MINOR_GC)) {
+			save_dirty_task(c_p, st);
+			st = NULL;
+			break;
+		    }
+#endif
+		    if (type == ERTS_PSTT_GC_MAJOR)
+			minor_gc = major_gc = 1;
+		    else
+			minor_gc = 1;
 		}
 		st_res = am_true;
 	    }
@@ -10499,20 +10569,31 @@ execute_sys_tasks(Process *c_p, erts_aint32_t *statep, int in_reds)
 	case ERTS_PSTT_CLA: {
 	    int fcalls;
             int cla_reds = 0;
+	    int do_gc;
+
 	    if (!ERTS_PROC_GET_SAVED_CALLS_BUF(c_p))
 		fcalls = reds;
 	    else
 		fcalls = reds - CONTEXT_REDS;
-	    st_res = erts_proc_copy_literal_area(c_p,
-						 &cla_reds,
-						 fcalls,
-						 st->arg[0] == am_true);
+	    do_gc = st->arg[0] == am_true;
+	    st_res = erts_proc_copy_literal_area(c_p, &cla_reds,
+						 fcalls, do_gc);
             reds -= cla_reds;
 	    if (is_non_value(st_res)) {
+#ifdef ERTS_DIRTY_SCHEDULERS
+		if (c_p->flags & F_DIRTY_CLA) {
+		    save_dirty_task(c_p, st);
+		    st = NULL;
+		    break;
+		}
+#endif
 		/* Needed gc, but gc was disabled */
 		save_gc_task(c_p, st, st_prio);
 		st = NULL;
+		break;
 	    }
+	    if (do_gc) /* We did a major gc */
+		minor_gc = major_gc = 1;
 	    break;
         }
 	case ERTS_PSTT_COHMQ:
@@ -10531,7 +10612,7 @@ execute_sys_tasks(Process *c_p, erts_aint32_t *statep, int in_reds)
 	}
 
 	if (st)
-	    reds += notify_sys_task_executed(c_p, st, st_res);
+	    reds += notify_sys_task_executed(c_p, st, st_res, 1);
 
 	state = erts_smp_atomic32_read_acqb(&c_p->state);
     } while (qmask && reds > 0);
@@ -10559,9 +10640,18 @@ cleanup_sys_tasks(Process *c_p, erts_aint32_t in_state, int in_reds)
 	Eterm st_res;
 	int st_prio;
 
-	st = fetch_sys_task(c_p, state, &qmask, &st_prio);
-	if (!st)
-	    break;
+#ifdef ERTS_DIRTY_SCHEDULERS
+	if (c_p->dirty_sys_tasks) {
+	    st = c_p->dirty_sys_tasks;
+	    c_p->dirty_sys_tasks = st->next;
+	}
+	else
+#endif
+	{
+	    st = fetch_sys_task(c_p, state, &qmask, &st_prio);
+	    if (!st)
+		break;
+	}
 
 	switch (st->type) {
         case ERTS_PSTT_GC_MAJOR:
@@ -10585,7 +10675,7 @@ cleanup_sys_tasks(Process *c_p, erts_aint32_t in_state, int in_reds)
 	    break;
 	}
 
-	reds += notify_sys_task_executed(c_p, st, st_res);
+	reds += notify_sys_task_executed(c_p, st, st_res, 1);
 
 	state = erts_smp_atomic32_read_acqb(&c_p->state);
     } while (qmask && reds < max_reds);
@@ -10595,6 +10685,92 @@ cleanup_sys_tasks(Process *c_p, erts_aint32_t in_state, int in_reds)
 
 #ifdef ERTS_DIRTY_SCHEDULERS
 
+void
+erts_execute_dirty_system_task(Process *c_p)
+{
+    Eterm cla_res = THE_NON_VALUE;
+    ErtsProcSysTask *stasks;
+
+    /*
+     * If multiple operations, perform them in the following
+     * order (in order to avoid unnecessary GC):
+     *  1. Copy Literal Area (implies major GC).
+     *  2. GC Hibernate (implies major GC if not woken).
+     *  3. Major GC (implies minor GC).
+     *  4. Minor GC.
+     *
+     * System task requests are handled after the actual
+     * operations have been performed...
+     */
+
+    ASSERT(!(c_p->flags & (F_DELAY_GC|F_DISABLE_GC)));
+
+    if (c_p->flags & F_DIRTY_CLA) {
+	int cla_reds = 0;
+	cla_res = erts_proc_copy_literal_area(c_p, &cla_reds, c_p->fcalls, 1);
+	ASSERT(is_value(cla_res));
+    }
+
+    if (c_p->flags & F_DIRTY_GC_HIBERNATE) {
+	erts_smp_proc_lock(c_p, ERTS_PROC_LOCK_MSGQ|ERTS_PROC_LOCK_STATUS);
+	ERTS_SMP_MSGQ_MV_INQ2PRIVQ(c_p);
+	if (c_p->msg.len)
+            c_p->flags &= ~F_DIRTY_GC_HIBERNATE; /* operation aborted... */
+        else {
+	    erts_smp_proc_unlock(c_p, ERTS_PROC_LOCK_MSGQ|ERTS_PROC_LOCK_STATUS);
+	    c_p->fvalue = NIL;
+	    erts_garbage_collect_hibernate(c_p);
+	    ERTS_VERIFY_UNUSED_TEMP_ALLOC(c_p);
+	    erts_smp_proc_lock(c_p, ERTS_PROC_LOCK_MSGQ|ERTS_PROC_LOCK_STATUS);
+	    ASSERT(!ERTS_PROC_IS_EXITING(c_p));
+	}
+	erts_smp_proc_unlock(c_p, ERTS_PROC_LOCK_MSGQ|ERTS_PROC_LOCK_STATUS);
+    }
+
+    if (c_p->flags & (F_DIRTY_MAJOR_GC|F_DIRTY_MINOR_GC)) {
+        if (c_p->flags & F_DIRTY_MAJOR_GC)
+            c_p->flags |= F_NEED_FULLSWEEP;
+	(void) erts_garbage_collect_nobump(c_p, 0, c_p->arg_reg,
+					   c_p->arity, c_p->fcalls);
+    }
+
+    ASSERT(!(c_p->flags & (F_DIRTY_CLA
+			   | F_DIRTY_GC_HIBERNATE
+			   | F_DIRTY_MAJOR_GC
+			   | F_DIRTY_MINOR_GC)));
+
+    stasks = c_p->dirty_sys_tasks;
+    c_p->dirty_sys_tasks = NULL;
+    
+    while (stasks) {
+	Eterm st_res;
+	ErtsProcSysTask *st = stasks;
+	stasks = st->next;
+
+	switch (st->type) {
+	case ERTS_PSTT_CLA:
+	    ASSERT(is_value(st_res));
+	    st_res = cla_res;
+	    break;
+	case ERTS_PSTT_GC_MAJOR:
+	    st_res = am_true;
+	    break;
+	case ERTS_PSTT_GC_MINOR:
+	    st_res = am_true;
+	    break;
+
+	default:
+	    ERTS_INTERNAL_ERROR("Not supported dirty system task");
+	    break;
+	}
+
+	(void) notify_sys_task_executed(c_p, st, st_res, 0);
+
+    }
+
+    erts_smp_atomic32_read_band_relb(&c_p->state, ~ERTS_PSFLG_DIRTY_ACTIVE_SYS);
+}
+
 static BIF_RETTYPE
 dispatch_system_task(Process *c_p, erts_aint_t fail_state,
 		     ErtsProcSysTask *st, Eterm target,
@@ -10797,7 +10973,7 @@ request_system_task(Process *c_p, Eterm requester, Eterm target,
 	    ERTS_INTERNAL_ERROR("Unknown failure schedule_process_sys_task()");
 	    failure = am_internal_error;
 	}
-	notify_sys_task_executed(c_p, st, failure);
+	notify_sys_task_executed(c_p, st, failure, 1);
     }
 
     ERTS_BIF_PREP_RET(ret, am_ok);
@@ -11012,6 +11188,15 @@ save_gc_task(Process *c_p, ErtsProcSysTask *st, int prio)
     }
 }
 
+#ifdef ERTS_DIRTY_SCHEDULERS
+static void
+save_dirty_task(Process *c_p, ErtsProcSysTask *st)
+{
+    st->next = c_p->dirty_sys_tasks;
+    c_p->dirty_sys_tasks = st;
+}
+#endif
+
 int
 erts_set_gc_state(Process *c_p, int enable)
 {
@@ -11348,6 +11533,7 @@ static void early_init_process_struct(void *varg, Eterm data)
     proc->common.id = make_internal_pid(data);
 #ifdef ERTS_DIRTY_SCHEDULERS
     erts_smp_atomic32_init_nob(&proc->dirty_state, 0);
+    proc->dirty_sys_tasks = NULL;
 #endif
     erts_smp_atomic32_init_relb(&proc->state, arg->state);
 
@@ -11856,6 +12042,7 @@ void erts_init_empty_process(Process *p)
 
 #ifdef ERTS_DIRTY_SCHEDULERS
     erts_smp_atomic32_init_nob(&p->dirty_state, 0);
+    p->dirty_sys_tasks = NULL;
 #endif
     erts_smp_atomic32_init_nob(&p->state, (erts_aint32_t) PRIORITY_NORMAL);
 
@@ -12443,7 +12630,9 @@ send_exit_signal(Process *c_p,		/* current process if and only
 		}
 		set_proc_exiting(c_p, state, rsn, NULL);
 	    }
-	    else if (!(state & (ERTS_PSFLG_RUNNING|ERTS_PSFLG_RUNNING_SYS))) {
+	    else if (!(state & (ERTS_PSFLG_RUNNING
+                                | ERTS_PSFLG_RUNNING_SYS
+                                | ERTS_PSFLG_DIRTY_RUNNING_SYS))) {
 		/* Process not running ... */
 		ErtsProcLocks need_locks = ~(*rp_locks) & ERTS_PROC_LOCKS_ALL;
 		ErlHeapFragment *bp = NULL;
@@ -12470,8 +12659,7 @@ send_exit_signal(Process *c_p,		/* current process if and only
 		    ErlOffHeap *ohp;
 		    Uint rsn_sz = size_object(rsn);
 #ifdef ERTS_DIRTY_SCHEDULERS
-		    if (state & (ERTS_PSFLG_DIRTY_RUNNING
-				 | ERTS_PSFLG_DIRTY_RUNNING_SYS)) {
+		    if (state & ERTS_PSFLG_DIRTY_RUNNING) {
 			bp = new_message_buffer(rsn_sz);
 			ohp = &bp->off_heap;
 			hp = &bp->mem[0];
@@ -12488,7 +12676,7 @@ send_exit_signal(Process *c_p,		/* current process if and only
 		set_proc_exiting(rp, state, rsn_cpy, bp);
 	    }
 	    else { /* Process running... */
-
+ 
 		/*
 		 * The pending exit will be discovered when the process
 		 * is scheduled out if not discovered earlier.
@@ -12510,8 +12698,35 @@ send_exit_signal(Process *c_p,		/* current process if and only
 							  &bp->off_heap);
 		    rp->pending_exit.bp = bp;
 		}
-		erts_smp_atomic32_read_bor_relb(&rp->state,
-						ERTS_PSFLG_PENDING_EXIT);
+
+		/*
+		 * If no dirty work has been scheduled, pending exit will
+                 * be discovered when the process is scheduled. If dirty work
+                 * has been scheduled, we may need to add it to a normal run
+                 * queue...
+		 */
+#ifndef ERTS_DIRTY_SCHEDULERS
+                (void) erts_smp_atomic32_read_bor_relb(&rp->state,
+                                                       ERTS_PSFLG_PENDING_EXIT);
+#else
+                {
+                    erts_aint32_t a = erts_smp_atomic32_read_nob(&rp->state);
+                    while (1) {
+                        erts_aint32_t n, e;
+                        int dwork;
+                        n = e = a;
+                        n |= ERTS_PSFLG_PENDING_EXIT;
+                        dwork = !!(n & ERTS_PSFLGS_DIRTY_WORK);
+                        n &= ~ERTS_PSFLGS_DIRTY_WORK;
+                        a = erts_smp_atomic32_cmpxchg_mb(&rp->state, n, e);
+                        if (a == e) {
+                            if (dwork)
+                                erts_schedule_process(rp, n, *rp_locks);
+                            break;
+                        }
+                    }
+                }
+#endif
 	    }
 	}
 	/* else:
diff --git a/erts/emulator/beam/erl_process.h b/erts/emulator/beam/erl_process.h
index 58a50f8cf4..3938c4c913 100644
--- a/erts/emulator/beam/erl_process.h
+++ b/erts/emulator/beam/erl_process.h
@@ -1060,6 +1060,9 @@ struct process {
     Uint64 bin_old_vheap;	/* Virtual old heap size for binaries */
 
     ErtsProcSysTaskQs *sys_task_qs;
+#ifdef ERTS_DIRTY_SCHEDULERS
+    ErtsProcSysTask *dirty_sys_tasks;
+#endif
 
     erts_smp_atomic32_t state;  /* Process state flags (see ERTS_PSFLG_*) */
 #ifdef ERTS_DIRTY_SCHEDULERS
@@ -1384,14 +1387,18 @@ extern int erts_system_profile_ts_type;
 #define F_FORCE_GC           (1 << 10) /* Force gc at process in-scheduling */
 #define F_DISABLE_GC         (1 << 11) /* Disable GC (see below) */
 #define F_OFF_HEAP_MSGQ      (1 << 12) /* Off heap msg queue */
-#define F_ON_HEAP_MSGQ       (1 << 13) /* Off heap msg queue */
+#define F_ON_HEAP_MSGQ       (1 << 13) /* On heap msg queue */
 #define F_OFF_HEAP_MSGQ_CHNG (1 << 14) /* Off heap msg queue changing */
 #define F_ABANDONED_HEAP_USE (1 << 15) /* Have usage of abandoned heap */
 #define F_DELAY_GC           (1 << 16) /* Similar to disable GC (see below) */
 #define F_SCHDLR_ONLN_WAITQ  (1 << 17) /* Process enqueued waiting to change schedulers online */
 #define F_HAVE_BLCKD_NMSCHED (1 << 18) /* Process has blocked normal multi-scheduling */
-#define F_HIPE_MODE          (1 << 19)
+#define F_HIPE_MODE          (1 << 19) /* Process is executing in HiPE mode */
 #define F_DELAYED_DEL_PROC   (1 << 20) /* Delay delete process (dirty proc exit case) */
+#define F_DIRTY_CLA          (1 << 21) /* Dirty copy literal area scheduled */
+#define F_DIRTY_GC_HIBERNATE (1 << 22) /* Dirty GC hibernate scheduled */
+#define F_DIRTY_MAJOR_GC     (1 << 23) /* Dirty major GC scheduled */
+#define F_DIRTY_MINOR_GC     (1 << 24) /* Dirty minor GC scheduled */
 
 /*
  * F_DISABLE_GC and F_DELAY_GC are similar. Both will prevent
@@ -1573,7 +1580,9 @@ void erts_init_scheduling(int, int
 			  , int, int, int
 #endif
 			  );
-
+#ifdef ERTS_DIRTY_SCHEDULERS
+void erts_execute_dirty_system_task(Process *c_p);
+#endif
 int erts_set_gc_state(Process *c_p, int enable);
 Eterm erts_sched_wall_time_request(Process *c_p, int set, int enable);
 Eterm erts_system_check_request(Process *c_p);
@@ -1918,6 +1927,8 @@ ErtsSchedulerData *erts_get_scheduler_data(void)
 void erts_schedule_process(Process *, erts_aint32_t, ErtsProcLocks);
 
 ERTS_GLB_INLINE void erts_proc_notify_new_message(Process *p, ErtsProcLocks locks);
+ERTS_GLB_INLINE void erts_schedule_dirty_sys_execution(Process *c_p);
+
 #if ERTS_GLB_INLINE_INCL_FUNC_DEF
 ERTS_GLB_INLINE void
 erts_proc_notify_new_message(Process *p, ErtsProcLocks locks)
@@ -1927,6 +1938,34 @@ erts_proc_notify_new_message(Process *p, ErtsProcLocks locks)
     if (!(state & ERTS_PSFLG_ACTIVE))
 	erts_schedule_process(p, state, locks);
 }
+
+ERTS_GLB_INLINE void
+erts_schedule_dirty_sys_execution(Process *c_p)
+{
+    erts_aint32_t a, n, e;
+
+    a = erts_smp_atomic32_read_nob(&c_p->state);
+
+    /*
+     * Only a currently executing process schedules
+     * itself for dirty-sys execution...
+     */
+
+    ASSERT(a & (ERTS_PSFLG_RUNNING|ERTS_PSFLG_RUNNING_SYS));
+
+    /* Don't set dirty-active-sys if we are about to exit... */
+
+    while (!(a & (ERTS_PSFLG_DIRTY_ACTIVE_SYS
+                  | ERTS_PSFLG_EXITING
+                  | ERTS_PSFLG_PENDING_EXIT))) {
+        e = a;
+        n = a | ERTS_PSFLG_DIRTY_ACTIVE_SYS;
+        a = erts_smp_atomic32_cmpxchg_mb(&c_p->state, n, e);
+        if (a == e)
+            break; /* dirty-active-sys set */
+    }
+}
+
 #endif
 
 #if defined(ERTS_SMP) && defined(ERTS_ENABLE_LOCK_CHECK)
diff --git a/erts/emulator/beam/ops.tab b/erts/emulator/beam/ops.tab
index 4ef04d020a..ec36b23059 100644
--- a/erts/emulator/beam/ops.tab
+++ b/erts/emulator/beam/ops.tab
@@ -802,15 +802,10 @@ call_ext_last u==1 Bif=u$bif:erts_internal:check_process_code/1 D => i_call_ext_
 call_ext_only u==1 Bif=u$bif:erts_internal:check_process_code/1 => i_call_ext_only Bif
 
 #
-# The BIFs erlang:garbage_collect/0 must be called like a function,
+# The BIFs erts_internal:garbage_collect/1 must be called like a function,
 # to allow them to invoke the garbage collector. (The stack pointer must
 # be saved and p->arity must be zeroed, which is not done on ordinary BIF calls.)
 #
-
-call_ext u==0 Bif=u$bif:erlang:garbage_collect/0 => i_call_ext Bif
-call_ext_last u==0 Bif=u$bif:erlang:garbage_collect/0 D => i_call_ext_last Bif D
-call_ext_only u==0 Bif=u$bif:erlang:garbage_collect/0 => i_call_ext_only Bif
-
 call_ext u==1 Bif=u$bif:erts_internal:garbage_collect/1 => i_call_ext Bif
 call_ext_last u==1 Bif=u$bif:erts_internal:garbage_collect/1 D => i_call_ext_last Bif D
 call_ext_only u==1 Bif=u$bif:erts_internal:garbage_collect/1 => i_call_ext_only Bif
diff --git a/erts/emulator/hipe/hipe_bif_list.m4 b/erts/emulator/hipe/hipe_bif_list.m4
index ec0d0f5a0d..f034c4700c 100644
--- a/erts/emulator/hipe/hipe_bif_list.m4
+++ b/erts/emulator/hipe/hipe_bif_list.m4
@@ -183,7 +183,7 @@ standard_bif_interface_0(nbif_ports_0, ports_0)
  */
 gc_bif_interface_1(nbif_erts_internal_check_process_code_1, hipe_erts_internal_check_process_code_1)
 gc_bif_interface_1(nbif_erase_1, erase_1)
-gc_bif_interface_0(nbif_garbage_collect_0, garbage_collect_0)
+gc_bif_interface_1(nbif_erts_internal_garbage_collect_1, erts_internal_garbage_collect_1)
 gc_nofail_primop_interface_1(nbif_gc_1, hipe_gc)
 gc_bif_interface_2(nbif_put_2, put_2)