1 files changed, 472 insertions, 123 deletions

diff --git a/erts/emulator/beam/erl_alloc_util.c b/erts/emulator/beam/erl_alloc_util.c
index 236ee35d18..6fddba4b34 100644
--- a/erts/emulator/beam/erl_alloc_util.c
+++ b/erts/emulator/beam/erl_alloc_util.c
@@ -1,7 +1,7 @@
 /*
  * %CopyrightBegin%
  *
- * Copyright Ericsson AB 2002-2013. All Rights Reserved.
+ * Copyright Ericsson AB 2002-2017. All Rights Reserved.
  *
  * Licensed under the Apache License, Version 2.0 (the "License");
  * you may not use this file except in compliance with the License.
@@ -52,6 +52,7 @@
 #ifdef ERTS_ENABLE_LOCK_COUNT
 #include "erl_lock_count.h"
 #endif
+#include "lttng-wrapper.h"
 
 #if defined(ERTS_ALLOC_UTIL_HARD_DEBUG) && defined(__GNUC__)
 #warning "* * * * * * * * * *"
@@ -751,12 +752,77 @@ internal_free(void *ptr)
 
 #endif
 
+#ifdef ARCH_32
+
+/*
+ * Bit vector for the entire 32-bit virtual address space
+ * with one bit for each super aligned memory segment.
+ */
+
+#define VSPACE_MAP_BITS  (1 << (32 - ERTS_MMAP_SUPERALIGNED_BITS))
+#define VSPACE_MAP_SZ    (VSPACE_MAP_BITS / ERTS_VSPACE_WORD_BITS)
+
+static ERTS_INLINE void set_bit(UWord* map, Uint ix)
+{
+    ASSERT(ix / ERTS_VSPACE_WORD_BITS < VSPACE_MAP_SZ);
+    map[ix / ERTS_VSPACE_WORD_BITS]
+        |= ((UWord)1 << (ix % ERTS_VSPACE_WORD_BITS));
+}
+
+static ERTS_INLINE void clr_bit(UWord* map, Uint ix)
+{
+    ASSERT(ix / ERTS_VSPACE_WORD_BITS < VSPACE_MAP_SZ);
+    map[ix / ERTS_VSPACE_WORD_BITS]
+        &= ~((UWord)1 << (ix % ERTS_VSPACE_WORD_BITS));
+}
+
+static ERTS_INLINE int is_bit_set(UWord* map, Uint ix)
+{
+    ASSERT(ix / ERTS_VSPACE_WORD_BITS < VSPACE_MAP_SZ);
+    return map[ix / ERTS_VSPACE_WORD_BITS]
+        & ((UWord)1 << (ix % ERTS_VSPACE_WORD_BITS));
+}
+
+UWord erts_literal_vspace_map[VSPACE_MAP_SZ];
+
+static void set_literal_range(void* start, Uint size)
+{
+    Uint ix = (UWord)start >> ERTS_MMAP_SUPERALIGNED_BITS;
+    Uint n = size >> ERTS_MMAP_SUPERALIGNED_BITS;
+
+    ASSERT(!((UWord)start & ERTS_INV_SUPERALIGNED_MASK));
+    ASSERT(!((UWord)size & ERTS_INV_SUPERALIGNED_MASK));
+    ASSERT(n);
+    while (n--) {
+        ASSERT(!is_bit_set(erts_literal_vspace_map, ix));
+        set_bit(erts_literal_vspace_map, ix);
+        ix++;
+    }
+}
+
+static void clear_literal_range(void* start, Uint size)
+{
+    Uint ix = (UWord)start >> ERTS_MMAP_SUPERALIGNED_BITS;
+    Uint n = size >> ERTS_MMAP_SUPERALIGNED_BITS;
+
+    ASSERT(!((UWord)start & ERTS_INV_SUPERALIGNED_MASK));
+    ASSERT(!((UWord)size & ERTS_INV_SUPERALIGNED_MASK));
+    ASSERT(n);
+    while (n--) {
+        ASSERT(is_bit_set(erts_literal_vspace_map, ix));
+        clr_bit(erts_literal_vspace_map, ix);
+        ix++;
+    }
+}
+
+#endif /* ARCH_32 */
+
 /* mseg ... */
 
 #if HAVE_ERTS_MSEG
 
-static ERTS_INLINE void *
-alcu_mseg_alloc(Allctr_t *allctr, Uint *size_p, Uint flags)
+static void*
+erts_alcu_mseg_alloc(Allctr_t *allctr, Uint *size_p, Uint flags)
 {
     void *res;
     UWord size = (UWord) *size_p;
@@ -766,8 +832,9 @@ alcu_mseg_alloc(Allctr_t *allctr, Uint *size_p, Uint flags)
     return res;
 }
 
-static ERTS_INLINE void *
-alcu_mseg_realloc(Allctr_t *allctr, void *seg, Uint old_size, Uint *new_size_p)
+static void*
+erts_alcu_mseg_realloc(Allctr_t *allctr, void *seg,
+                       Uint old_size, Uint *new_size_p)
 {
     void *res;
     UWord new_size = (UWord) *new_size_p;
@@ -778,19 +845,162 @@ alcu_mseg_realloc(Allctr_t *allctr, void *seg, Uint old_size, Uint *new_size_p)
     return res;
 }
 
-static ERTS_INLINE void
-alcu_mseg_dealloc(Allctr_t *allctr, void *seg, Uint size, Uint flags)
+static void
+erts_alcu_mseg_dealloc(Allctr_t *allctr, void *seg, Uint size, Uint flags)
 {
     erts_mseg_dealloc_opt(allctr->alloc_no, seg, (UWord) size, flags, &allctr->mseg_opt);
     INC_CC(allctr->calls.mseg_dealloc);
 }
 
-#endif
 
-static ERTS_INLINE void *
-alcu_sys_alloc(Allctr_t *allctr, Uint size, int superalign)
+#if defined(ARCH_32)
+
+void*
+erts_alcu_literal_32_mseg_alloc(Allctr_t *allctr, Uint *size_p, Uint flags)
+{
+    void* res;
+    Uint sz = ERTS_SUPERALIGNED_CEILING(*size_p);
+    ERTS_LC_ASSERT(allctr->alloc_no == ERTS_ALC_A_LITERAL &&
+                   allctr->t == 0);
+    ERTS_SMP_LC_ASSERT(allctr->thread_safe);
+
+    res = erts_alcu_mseg_alloc(allctr, &sz, flags);
+    if (res) {
+        set_literal_range(res, sz);
+        *size_p = sz;
+    }
+    return res;
+}
+
+void*
+erts_alcu_literal_32_mseg_realloc(Allctr_t *allctr, void *seg,
+                                  Uint old_size, Uint *new_size_p)
+{
+    void* res;
+    Uint new_sz = ERTS_SUPERALIGNED_CEILING(*new_size_p);
+    ERTS_LC_ASSERT(allctr->alloc_no == ERTS_ALC_A_LITERAL &&
+                   allctr->t == 0);
+    ERTS_SMP_LC_ASSERT(allctr->thread_safe);
+
+    if (seg && old_size)
+        clear_literal_range(seg, old_size);
+    res = erts_alcu_mseg_realloc(allctr, seg, old_size, &new_sz);
+    if (res) {
+        set_literal_range(res, new_sz);
+        *new_size_p = new_sz;
+    }
+    return res;
+}
+
+void
+erts_alcu_literal_32_mseg_dealloc(Allctr_t *allctr, void *seg, Uint size,
+                               Uint flags)
+{
+    ERTS_LC_ASSERT(allctr->alloc_no == ERTS_ALC_A_LITERAL &&
+                   allctr->t == 0);
+    ERTS_SMP_LC_ASSERT(allctr->thread_safe);
+
+    erts_alcu_mseg_dealloc(allctr, seg, size, flags);
+
+    clear_literal_range(seg, size);
+}
+
+#elif defined(ARCH_64) && defined(ERTS_HAVE_OS_PHYSICAL_MEMORY_RESERVATION)
+
+/* For allocators that have their own mmapper (super carrier),
+ * like literal_alloc and exec_alloc on amd64
+ */
+void*
+erts_alcu_mmapper_mseg_alloc(Allctr_t *allctr, Uint *size_p, Uint flags)
+{
+    void* res;
+    UWord size = (UWord) *size_p;
+    Uint32 mmap_flags = ERTS_MMAPFLG_SUPERCARRIER_ONLY;
+    if (flags & ERTS_MSEG_FLG_2POW)
+        mmap_flags |= ERTS_MMAPFLG_SUPERALIGNED;
+
+    res = erts_mmap(allctr->mseg_mmapper, mmap_flags, &size);
+    *size_p = (Uint)size;
+    INC_CC(allctr->calls.mseg_alloc);
+    return res;
+}
+
+void*
+erts_alcu_mmapper_mseg_realloc(Allctr_t *allctr, void *seg,
+                               Uint old_size, Uint *new_size_p)
 {
     void *res;
+    UWord new_size = (UWord) *new_size_p;
+    res = erts_mremap(allctr->mseg_mmapper, ERTS_MSEG_FLG_NONE, seg, old_size, &new_size);
+    *new_size_p = (Uint) new_size;
+    INC_CC(allctr->calls.mseg_realloc);
+    return res;
+}
+
+void
+erts_alcu_mmapper_mseg_dealloc(Allctr_t *allctr, void *seg, Uint size,
+                               Uint flags)
+{
+    Uint32 mmap_flags = ERTS_MMAPFLG_SUPERCARRIER_ONLY;
+    if (flags & ERTS_MSEG_FLG_2POW)
+        mmap_flags |= ERTS_MMAPFLG_SUPERALIGNED;
+
+    erts_munmap(allctr->mseg_mmapper, mmap_flags, seg, (UWord)size);
+    INC_CC(allctr->calls.mseg_dealloc);
+}
+#endif /* ARCH_64 && ERTS_HAVE_OS_PHYSICAL_MEMORY_RESERVATION */
+
+#if defined(ERTS_ALC_A_EXEC) && !defined(ERTS_HAVE_EXEC_MMAPPER)
+
+/*
+ * For exec_alloc on non-amd64 that just need memory with PROT_EXEC
+ */
+void*
+erts_alcu_exec_mseg_alloc(Allctr_t *allctr, Uint *size_p, Uint flags)
+{
+    void* res = erts_alcu_mseg_alloc(allctr, size_p, flags);
+
+    if (res) {
+        int r = mprotect(res, *size_p, PROT_EXEC | PROT_READ | PROT_WRITE);
+        ASSERT(r == 0); (void)r;
+    }
+    return res;
+}
+
+void*
+erts_alcu_exec_mseg_realloc(Allctr_t *allctr, void *seg,
+                            Uint old_size, Uint *new_size_p)
+{
+    void *res;
+
+    if (seg && old_size) {
+        int r = mprotect(seg, old_size, PROT_READ | PROT_WRITE);
+        ASSERT(r == 0); (void)r;
+    }
+    res = erts_alcu_mseg_realloc(allctr, seg, old_size, new_size_p);
+    if (res) {
+        int r = mprotect(res, *new_size_p, PROT_EXEC | PROT_READ | PROT_WRITE);
+        ASSERT(r == 0); (void)r;
+    }
+    return res;
+}
+
+void
+erts_alcu_exec_mseg_dealloc(Allctr_t *allctr, void *seg, Uint size, Uint flags)
+{
+    int r = mprotect(seg, size, PROT_READ | PROT_WRITE);
+    ASSERT(r == 0); (void)r;
+    erts_alcu_mseg_dealloc(allctr, seg, size, flags);
+}
+#endif /* ERTS_ALC_A_EXEC && !ERTS_HAVE_EXEC_MMAPPER */
+
+#endif /* HAVE_ERTS_MSEG */
+
+static void*
+erts_alcu_sys_alloc(Allctr_t *allctr, Uint* size_p, int superalign)
+{
+    void *res;
+    const Uint size = *size_p;
 #if ERTS_SA_MB_CARRIERS && ERTS_HAVE_ERTS_SYS_ALIGNED_ALLOC
     if (superalign)
 	res = erts_sys_aligned_alloc(ERTS_SACRR_UNIT_SZ, size);
@@ -803,10 +1013,11 @@ alcu_sys_alloc(Allctr_t *allctr, Uint size, int superalign)
     return res;
 }
 
-static ERTS_INLINE void *
-alcu_sys_realloc(Allctr_t *allctr, void *ptr, Uint size, Uint old_size, int superalign)
+static void*
+erts_alcu_sys_realloc(Allctr_t *allctr, void *ptr, Uint *size_p, Uint old_size, int superalign)
 {
     void *res;
+    const Uint size = *size_p;
 
 #if ERTS_SA_MB_CARRIERS && ERTS_HAVE_ERTS_SYS_ALIGNED_ALLOC
     if (superalign)
@@ -824,8 +1035,8 @@ alcu_sys_realloc(Allctr_t *allctr, void *ptr, Uint size, Uint old_size, int supe
     return res;
 }
 
-static ERTS_INLINE void
-alcu_sys_free(Allctr_t *allctr, void *ptr, int superalign)
+static void
+erts_alcu_sys_dealloc(Allctr_t *allctr, void *ptr, Uint size, int superalign)
 {
 #if ERTS_SA_MB_CARRIERS && ERTS_HAVE_ERTS_SYS_ALIGNED_ALLOC
     if (superalign)
@@ -838,6 +1049,59 @@ alcu_sys_free(Allctr_t *allctr, void *ptr, int superalign)
 	erts_mtrace_crr_free(allctr->alloc_no, ERTS_ALC_A_SYSTEM, ptr);
 }
 
+#ifdef ARCH_32
+
+void*
+erts_alcu_literal_32_sys_alloc(Allctr_t *allctr, Uint* size_p, int superalign)
+{
+    void* res;
+    Uint size = ERTS_SUPERALIGNED_CEILING(*size_p);
+    ERTS_LC_ASSERT(allctr->alloc_no == ERTS_ALC_A_LITERAL &&
+                   allctr->t == 0);
+    ERTS_SMP_LC_ASSERT(allctr->thread_safe);
+
+    res = erts_alcu_sys_alloc(allctr, &size, 1);
+    if (res) {
+        set_literal_range(res, size);
+        *size_p = size;
+    }
+    return res;
+}
+
+void*
+erts_alcu_literal_32_sys_realloc(Allctr_t *allctr, void *ptr, Uint* size_p, Uint old_size, int superalign)
+{
+    void* res;
+    Uint size = ERTS_SUPERALIGNED_CEILING(*size_p);
+
+    ERTS_LC_ASSERT(allctr->alloc_no == ERTS_ALC_A_LITERAL &&
+                   allctr->t == 0);
+    ERTS_SMP_LC_ASSERT(allctr->thread_safe);
+
+    if (ptr && old_size)
+        clear_literal_range(ptr, old_size);
+    res = erts_alcu_sys_realloc(allctr, ptr, &size, old_size, 1);
+    if (res) {
+        set_literal_range(res, size);
+        *size_p = size;
+    }
+    return res;
+}
+
+void
+erts_alcu_literal_32_sys_dealloc(Allctr_t *allctr, void *ptr, Uint size, int superalign)
+{
+    ERTS_LC_ASSERT(allctr->alloc_no == ERTS_ALC_A_LITERAL &&
+                   allctr->t == 0);
+    ERTS_SMP_LC_ASSERT(allctr->thread_safe);
+
+    erts_alcu_sys_dealloc(allctr, ptr, size, 1);
+
+    clear_literal_range(ptr, size);
+}
+
+#endif /* ARCH_32 */
+
 static Uint
 get_next_mbc_size(Allctr_t *allctr)
 {
@@ -1143,6 +1407,7 @@ fix_cpool_alloc(Allctr_t *allctr, ErtsAlcType_t type, Uint size)
 	   && type <= ERTS_ALC_N_MAX_A_FIXED_SIZE);
 
     fix = &allctr->fix[type - ERTS_ALC_N_MIN_A_FIXED_SIZE];
+    ASSERT(size == fix->type_size);
 
     res = fix->list;
     if (res) {
@@ -1154,8 +1419,6 @@ fix_cpool_alloc(Allctr_t *allctr, ErtsAlcType_t type, Uint size)
 	fix_cpool_check_shrink(allctr, type, fix, NULL);
 	return res;
     }
-    if (size < 2*sizeof(UWord))
-	size += sizeof(UWord);
     if (size >= allctr->sbc_threshold) {
 	Block_t *blk;
 	blk = create_carrier(allctr, size, CFLG_SBC);
@@ -1275,6 +1538,7 @@ fix_nocpool_alloc(Allctr_t *allctr, ErtsAlcType_t type, Uint size)
 	   && type <= ERTS_ALC_N_MAX_A_FIXED_SIZE);
 
     fix = &allctr->fix[type - ERTS_ALC_N_MIN_A_FIXED_SIZE];
+    ASSERT(size == fix->type_size);
 
     ERTS_DBG_CHK_FIX_LIST(allctr, fix, ix, 1);
     fix->u.nocpool.used++;
@@ -1297,8 +1561,6 @@ fix_nocpool_alloc(Allctr_t *allctr, ErtsAlcType_t type, Uint size)
 	ERTS_DBG_CHK_FIX_LIST(allctr, fix, ix, 0);
 	return res;
     }
-    if (size < 2*sizeof(UWord))
-	size += sizeof(UWord);
     if (fix->u.nocpool.limit < fix->u.nocpool.used)
 	fix->u.nocpool.limit = fix->u.nocpool.used;
     if (fix->u.nocpool.max_used < fix->u.nocpool.used)
@@ -1437,6 +1699,16 @@ erts_alcu_fix_alloc_shrink(Allctr_t *allctr, erts_aint32_t flgs)
 
 static void dealloc_carrier(Allctr_t *allctr, Carrier_t *crr, int superaligned);
 
+static ERTS_INLINE void
+dealloc_mbc(Allctr_t *allctr, Carrier_t *crr)
+{
+    ASSERT(IS_MB_CARRIER(crr));
+    if (allctr->destroying_mbc)
+        allctr->destroying_mbc(allctr, crr);
+
+    dealloc_carrier(allctr, crr, 1);
+}
+
 #ifdef ERTS_SMP
 
 static ERTS_INLINE Allctr_t*
@@ -2074,7 +2346,7 @@ mbc_alloc_block(Allctr_t *allctr, Uint size, Uint *blk_szp)
 
     if (!blk) {
 	blk = create_carrier(allctr, get_blk_sz, CFLG_MBC);
-#if !HALFWORD_HEAP && !ERTS_SUPER_ALIGNED_MSEG_ONLY
+#if !ERTS_SUPER_ALIGNED_MSEG_ONLY
 	if (!blk) {
 	    /* Emergency! We couldn't create the carrier as we wanted.
 	       Try to place it in a sys_alloced sbc. */
@@ -2920,6 +3192,7 @@ cpool_insert(Allctr_t *allctr, Carrier_t *crr)
 
     erts_smp_atomic_set_wb(&crr->allctr,
 			   ((erts_aint_t) allctr)|ERTS_CRR_ALCTR_FLG_IN_POOL);
+    LTTNG3(carrier_pool_put, ERTS_ALC_A2AD(allctr->alloc_no), allctr->ix, CARRIER_SZ(crr));
 }
 
 static void
@@ -3035,6 +3308,7 @@ cpool_fetch(Allctr_t *allctr, UWord size)
     first_old_traitor = allctr->cpool.traitor_list.next;
     cpool_entrance = NULL;
 
+    LTTNG3(carrier_pool_get, ERTS_ALC_A2AD(allctr->alloc_no), allctr->ix, (unsigned long)size);
     /*
      * Search my own pooled_list,
      * i.e my abandoned carriers that were in the pool last time I checked.
@@ -3149,7 +3423,7 @@ cpool_fetch(Allctr_t *allctr, UWord size)
 	cpool_entrance = sentinel;
 	cpdp = cpool_aint2cpd(cpool_read(&cpool_entrance->prev));
 	if (cpdp == sentinel)
-	    return NULL;
+	    goto check_dc_list;
     }
 
     has_passed_sentinel = 0;
@@ -3160,18 +3434,18 @@ cpool_fetch(Allctr_t *allctr, UWord size)
 	    if (cpool_entrance == sentinel) {
 		cpdp = cpool_aint2cpd(cpool_read(&cpdp->prev));
 		if (cpdp == sentinel)
-		    return NULL;
+		    break;
 	    }
 	    i = 0; /* Last one to inspect */
 	}
 	else if (cpdp == sentinel) {
 	    if (has_passed_sentinel) {
 		/* We been here before. cpool_entrance must have been removed */
-		return NULL;
+		break;
 	    }
 	    cpdp = cpool_aint2cpd(cpool_read(&cpdp->prev));
 	    if (cpdp == sentinel)
-		return NULL;
+                break;
 	    has_passed_sentinel = 1;
 	}
 	crr = (Carrier_t *)(((char *)cpdp) - offsetof(Carrier_t, cpool));
@@ -3195,10 +3469,12 @@ cpool_fetch(Allctr_t *allctr, UWord size)
 	    return NULL;
     }
 
+check_dc_list:
     /* Last; check our own pending dealloc carrier list... */
     crr = allctr->cpool.dc_list.last;
     while (crr) {
 	if (erts_atomic_read_nob(&crr->cpool.max_size) >= size) {
+	    Block_t* blk;
 	    unlink_carrier(&allctr->cpool.dc_list, crr);
 #ifdef ERTS_ALC_CPOOL_DEBUG
 	    ERTS_ALC_CPOOL_ASSERT(erts_smp_atomic_xchg_nob(&crr->allctr,
@@ -3207,6 +3483,9 @@ cpool_fetch(Allctr_t *allctr, UWord size)
 #else
 	    erts_smp_atomic_set_nob(&crr->allctr, ((erts_aint_t) allctr));
 #endif
+	    blk = MBC_TO_FIRST_BLK(allctr, crr);
+	    ASSERT(FBLK_TO_MBC(blk) == crr);
+	    allctr->link_free_block(allctr, blk);
 	    return crr;
 	}
 	crr = crr->prev;
@@ -3237,7 +3516,7 @@ check_pending_dealloc_carrier(Allctr_t *allctr,
 
 	    dcrr = crr;
 	    crr = crr->next;
-	    dealloc_carrier(allctr, dcrr, 1);
+	    dealloc_mbc(allctr, dcrr);
 	    i++;
 	} while (crr && i < ERTS_ALC_MAX_DEALLOC_CARRIER);
 
@@ -3268,18 +3547,20 @@ static void
 schedule_dealloc_carrier(Allctr_t *allctr, Carrier_t *crr)
 {
     Allctr_t *orig_allctr;
+    Block_t *blk;
     int check_pending_dealloc;
     erts_aint_t max_size;
 
+    ASSERT(IS_MB_CARRIER(crr));
+
     if (!ERTS_ALC_IS_CPOOL_ENABLED(allctr)) {
-	dealloc_carrier(allctr, crr, 1);
+	dealloc_mbc(allctr, crr);
 	return;
     }
 
     orig_allctr = crr->cpool.orig_allctr;
 
     if (allctr != orig_allctr) {
-	Block_t *blk = MBC_TO_FIRST_BLK(allctr, crr);
 	int cinit = orig_allctr->dd.ix - allctr->dd.ix;
 
 	/*
@@ -3296,6 +3577,7 @@ schedule_dealloc_carrier(Allctr_t *allctr, Carrier_t *crr)
 	 * since the block is an mbc block that is free and last
 	 * in the carrier.
 	 */
+	blk = MBC_TO_FIRST_BLK(allctr, crr);
 	ERTS_ALC_CPOOL_ASSERT(IS_FREE_LAST_MBC_BLK(blk));
 
 	ERTS_ALC_CPOOL_ASSERT(IS_MBC_FIRST_ABLK(allctr, blk));
@@ -3315,11 +3597,13 @@ schedule_dealloc_carrier(Allctr_t *allctr, Carrier_t *crr)
 
     if (crr->cpool.thr_prgr == ERTS_THR_PRGR_INVALID
 	|| erts_thr_progress_has_reached(crr->cpool.thr_prgr)) {
-	dealloc_carrier(allctr, crr, 1);
+	dealloc_mbc(allctr, crr);
 	return;
     }
 
-    max_size = (erts_aint_t) allctr->largest_fblk_in_mbc(allctr, crr);
+    blk = MBC_TO_FIRST_BLK(allctr, crr);
+    ASSERT(IS_FREE_LAST_MBC_BLK(blk));
+    max_size = (erts_aint_t) MBC_FBLK_SZ(blk);
     erts_atomic_set_nob(&crr->cpool.max_size, max_size);
 
     crr->next = NULL;
@@ -3511,8 +3795,7 @@ create_carrier(Allctr_t *allctr, Uint umem_sz, UWord flags)
     int is_mseg = 0;
 #endif
 
-    if (HALFWORD_HEAP
-	|| (ERTS_SUPER_ALIGNED_MSEG_ONLY && (flags & CFLG_MBC))
+    if ((ERTS_SUPER_ALIGNED_MSEG_ONLY && (flags & CFLG_MBC))
 	|| !allow_sys_alloc_carriers) {
 	flags |= CFLG_FORCE_MSEG;
 	flags &= ~CFLG_FORCE_SYS_ALLOC;
@@ -3520,6 +3803,8 @@ create_carrier(Allctr_t *allctr, Uint umem_sz, UWord flags)
 	return NULL;
 #endif
     }
+    flags |= allctr->crr_set_flgs;
+    flags &= ~allctr->crr_clr_flgs;
 
     ASSERT((flags & CFLG_SBC && !(flags & CFLG_MBC))
 	   || (flags & CFLG_MBC && !(flags & CFLG_SBC)));
@@ -3535,7 +3820,21 @@ create_carrier(Allctr_t *allctr, Uint umem_sz, UWord flags)
 	return NULL;
     }
 
-    blk_sz = UMEMSZ2BLKSZ(allctr, umem_sz);
+    if (flags & CFLG_MAIN_CARRIER) {
+        ASSERT(flags & CFLG_MBC);
+        ASSERT(flags & CFLG_NO_CPOOL);
+        ASSERT(umem_sz == allctr->main_carrier_size);
+        ERTS_UNDEF(blk_sz, 0);
+
+        if (allctr->main_carrier_size < allctr->min_mbc_size)
+            allctr->main_carrier_size = allctr->min_mbc_size;
+        crr_sz = bcrr_sz = allctr->main_carrier_size;
+    }
+    else {
+        ERTS_UNDEF(bcrr_sz, 0);
+	ERTS_UNDEF(crr_sz, 0);
+        blk_sz = UMEMSZ2BLKSZ(allctr, umem_sz);
+    }
 
 #ifdef ERTS_SMP
     allctr->cpool.disable_abandon = ERTS_ALC_CPOOL_MAX_DISABLE_ABANDON;
@@ -3581,13 +3880,15 @@ create_carrier(Allctr_t *allctr, Uint umem_sz, UWord flags)
 	mseg_flags = ERTS_MSEG_FLG_NONE;
     }
     else {
-	crr_sz = (*allctr->get_next_mbc_size)(allctr);
-	if (crr_sz < MBC_HEADER_SIZE(allctr) + blk_sz)
-	    crr_sz = MBC_HEADER_SIZE(allctr) + blk_sz;
-	mseg_flags = ERTS_MSEG_FLG_2POW;
+        if (!(flags & CFLG_MAIN_CARRIER)) {
+            crr_sz = (*allctr->get_next_mbc_size)(allctr);
+            if (crr_sz < MBC_HEADER_SIZE(allctr) + blk_sz)
+                crr_sz = MBC_HEADER_SIZE(allctr) + blk_sz;
+        }
+        mseg_flags = ERTS_MSEG_FLG_2POW;
     }
 
-    crr = (Carrier_t *) alcu_mseg_alloc(allctr, &crr_sz, mseg_flags);
+    crr = (Carrier_t *) allctr->mseg_alloc(allctr, &crr_sz, mseg_flags);
     if (!crr) {
 	have_tried_mseg = 1;
 	if (!(have_tried_sys_alloc || flags & CFLG_FORCE_MSEG))
@@ -3619,23 +3920,22 @@ create_carrier(Allctr_t *allctr, Uint umem_sz, UWord flags)
     if (flags & CFLG_SBC) {
 	bcrr_sz = blk_sz + SBC_HEADER_SIZE;
     }
-    else {
+    else if (!(flags & CFLG_MAIN_CARRIER)) {
 	bcrr_sz = MBC_HEADER_SIZE(allctr) + blk_sz;
-	if (!(flags & CFLG_MAIN_CARRIER)
-	    && bcrr_sz < allctr->smallest_mbc_size)
-	    bcrr_sz = allctr->smallest_mbc_size;
+	if (bcrr_sz < allctr->smallest_mbc_size)
+            bcrr_sz = allctr->smallest_mbc_size;
     }
 
     crr_sz = (flags & CFLG_FORCE_SIZE
 	      ? UNIT_CEILING(bcrr_sz)
 	      : SYS_ALLOC_CARRIER_CEILING(bcrr_sz));
 
-    crr = (Carrier_t *) alcu_sys_alloc(allctr, crr_sz, flags & CFLG_MBC);
+    crr = (Carrier_t *) allctr->sys_alloc(allctr, &crr_sz, flags & CFLG_MBC);
 	
     if (!crr) {
 	if (crr_sz > UNIT_CEILING(bcrr_sz)) {
 	    crr_sz = UNIT_CEILING(bcrr_sz);
-	    crr = (Carrier_t *) alcu_sys_alloc(allctr, crr_sz, flags & CFLG_MBC);
+	    crr = (Carrier_t *) allctr->sys_alloc(allctr, &crr_sz, flags & CFLG_MBC);
 	}
 	if (!crr) {
 #if HAVE_ERTS_MSEG
@@ -3694,6 +3994,21 @@ create_carrier(Allctr_t *allctr, Uint umem_sz, UWord flags)
 
     }
 
+#ifdef USE_LTTNG_VM_TRACEPOINTS
+    if (LTTNG_ENABLED(carrier_create)) {
+        lttng_decl_carrier_stats(mbc_stats);
+        lttng_decl_carrier_stats(sbc_stats);
+        LTTNG_CARRIER_STATS_TO_LTTNG_STATS(&(allctr->mbcs), mbc_stats);
+        LTTNG_CARRIER_STATS_TO_LTTNG_STATS(&(allctr->sbcs), sbc_stats);
+        LTTNG5(carrier_create,
+                ERTS_ALC_A2AD(allctr->alloc_no),
+                allctr->ix,
+                crr_sz,
+                mbc_stats,
+                sbc_stats);
+    }
+#endif
+
     DEBUG_SAVE_ALIGNMENT(crr);
     return blk;
 }
@@ -3734,7 +4049,7 @@ resize_carrier(Allctr_t *allctr, Block_t *old_blk, Uint umem_sz, UWord flags)
 
 	    new_crr_sz = new_blk_sz + SBC_HEADER_SIZE;
 	    new_crr_sz = ERTS_SACRR_UNIT_CEILING(new_crr_sz);
-	    new_crr = (Carrier_t *) alcu_mseg_realloc(allctr,
+	    new_crr = (Carrier_t *) allctr->mseg_realloc(allctr,
 						      old_crr,
 						      old_crr_sz,
 						      &new_crr_sz);
@@ -3749,11 +4064,6 @@ resize_carrier(Allctr_t *allctr, Block_t *old_blk, Uint umem_sz, UWord flags)
 		DEBUG_SAVE_ALIGNMENT(new_crr);
 		return new_blk;
 	    }
-#if HALFWORD_HEAP
-	    /* Old carrier unchanged; restore stat */
-	    STAT_MSEG_SBC_ALLOC(allctr, old_crr_sz, old_blk_sz);
-	    return NULL;
-#endif	    
 	    create_flags |= CFLG_FORCE_SYS_ALLOC; /* since mseg_realloc()
 						     failed */
 	}
@@ -3764,7 +4074,7 @@ resize_carrier(Allctr_t *allctr, Block_t *old_blk, Uint umem_sz, UWord flags)
 		       (void *) BLK2UMEM(old_blk),
 		       MIN(new_blk_sz, old_blk_sz) - ABLK_HDR_SZ);
 	    unlink_carrier(&allctr->sbc_list, old_crr);
-	    alcu_mseg_dealloc(allctr, old_crr, old_crr_sz, ERTS_MSEG_FLG_NONE);
+            allctr->mseg_dealloc(allctr, old_crr, old_crr_sz, ERTS_MSEG_FLG_NONE);
 	}
 	else {
 	    /* Old carrier unchanged; restore stat */
@@ -3781,9 +4091,9 @@ resize_carrier(Allctr_t *allctr, Block_t *old_blk, Uint umem_sz, UWord flags)
 			  ? UNIT_CEILING(new_bcrr_sz)
 			  : SYS_ALLOC_CARRIER_CEILING(new_bcrr_sz));
 
-	    new_crr = (Carrier_t *) alcu_sys_realloc(allctr,
+	    new_crr = (Carrier_t *) allctr->sys_realloc(allctr,
 						     (void *) old_crr,
-						     new_crr_sz,
+						     &new_crr_sz,
 						     old_crr_sz,
 						     0);
 	    if (new_crr) {
@@ -3802,9 +4112,9 @@ resize_carrier(Allctr_t *allctr, Block_t *old_blk, Uint umem_sz, UWord flags)
 	    else if (new_crr_sz > UNIT_CEILING(new_bcrr_sz)) {
 		new_crr_sz = new_blk_sz + SBC_HEADER_SIZE;
 		new_crr_sz = UNIT_CEILING(new_crr_sz);
-		new_crr = (Carrier_t *) alcu_sys_realloc(allctr,
+		new_crr = (Carrier_t *) allctr->sys_realloc(allctr,
 							 (void *) old_crr,
-							 new_crr_sz,
+							 &new_crr_sz,
 							 old_crr_sz,
 							 0);
 		if (new_crr)
@@ -3825,7 +4135,7 @@ resize_carrier(Allctr_t *allctr, Block_t *old_blk, Uint umem_sz, UWord flags)
 		       (void *) BLK2UMEM(old_blk),
 		       MIN(new_blk_sz, old_blk_sz) - ABLK_HDR_SZ);
 	    unlink_carrier(&allctr->sbc_list, old_crr);
-	    alcu_sys_free(allctr, old_crr, 0);
+	    allctr->sys_dealloc(allctr, old_crr, CARRIER_SZ(old_crr), 0);
 	}
 	else {
 	    /* Old carrier unchanged; restore... */
@@ -3841,13 +4151,13 @@ dealloc_carrier(Allctr_t *allctr, Carrier_t *crr, int superaligned)
 {
 #if HAVE_ERTS_MSEG
     if (IS_MSEG_CARRIER(crr))
-	alcu_mseg_dealloc(allctr, crr, CARRIER_SZ(crr),
+	allctr->mseg_dealloc(allctr, crr, CARRIER_SZ(crr),
 			  (superaligned
 			   ? ERTS_MSEG_FLG_2POW
 			   : ERTS_MSEG_FLG_NONE));
     else
 #endif
-	alcu_sys_free(allctr, crr, superaligned);
+	allctr->sys_dealloc(allctr, crr, CARRIER_SZ(crr), superaligned);
 }
 
 static void
@@ -3887,16 +4197,13 @@ destroy_carrier(Allctr_t *allctr, Block_t *blk, Carrier_t **busy_pcrr_pp)
 	    ASSERT(IS_LAST_BLK(blk));
 
 #ifdef ERTS_ALLOC_UTIL_HARD_DEBUG
-	    (*allctr->link_free_block)(allctr, blk, 0);
+	    (*allctr->link_free_block)(allctr, blk);
 	    HARD_CHECK_BLK_CARRIER(allctr, blk);
-	    (*allctr->unlink_free_block)(allctr, blk, 0);
+	    (*allctr->unlink_free_block)(allctr, blk);
 #endif
 	}
 #endif
 
-	if (allctr->destroying_mbc)
-	    (*allctr->destroying_mbc)(allctr, crr);
-
 #ifdef ERTS_SMP
 	if (busy_pcrr_pp && *busy_pcrr_pp) {
 	    ERTS_ALC_CPOOL_ASSERT(*busy_pcrr_pp == crr);
@@ -3920,12 +4227,30 @@ destroy_carrier(Allctr_t *allctr, Block_t *blk, Carrier_t **busy_pcrr_pp)
 	    else
 #endif
 		STAT_SYS_ALLOC_MBC_FREE(allctr, crr_sz);
+
+            if (allctr->remove_mbc)
+                allctr->remove_mbc(allctr, crr);
 	}
 
+#ifdef USE_LTTNG_VM_TRACEPOINTS
+        if (LTTNG_ENABLED(carrier_destroy)) {
+            lttng_decl_carrier_stats(mbc_stats);
+            lttng_decl_carrier_stats(sbc_stats);
+            LTTNG_CARRIER_STATS_TO_LTTNG_STATS(&(allctr->mbcs), mbc_stats);
+            LTTNG_CARRIER_STATS_TO_LTTNG_STATS(&(allctr->sbcs), sbc_stats);
+            LTTNG5(carrier_destroy,
+                ERTS_ALC_A2AD(allctr->alloc_no),
+                allctr->ix,
+                crr_sz,
+                mbc_stats,
+                sbc_stats);
+        }
+#endif
+
 #ifdef ERTS_SMP
 	schedule_dealloc_carrier(allctr, crr);
 #else
-	dealloc_carrier(allctr, crr, 1);
+	dealloc_mbc(allctr, crr);
 #endif
     }
 }
@@ -3942,9 +4267,6 @@ static struct {
     Eterm e;
     Eterm t;
     Eterm ramv;
-#if HALFWORD_HEAP
-    Eterm low;
-#endif
     Eterm sbct;
 #if HAVE_ERTS_MSEG
     Eterm asbcst;
@@ -4035,9 +4357,6 @@ init_atoms(Allctr_t *allctr)
 	AM_INIT(e);
 	AM_INIT(t);
 	AM_INIT(ramv);
-#if HALFWORD_HEAP
-	AM_INIT(low);
-#endif
 	AM_INIT(sbct);
 #if HAVE_ERTS_MSEG
 	AM_INIT(asbcst);
@@ -4198,7 +4517,7 @@ add_fix_types(Allctr_t *allctr, int internal, Uint **hpp, Uint *szp,
 static Eterm
 sz_info_fix(Allctr_t *allctr,
 	    int internal,
-	    int *print_to_p,
+	    fmtfn_t *print_to_p,
 	    void *print_to_arg,
 	    Uint **hpp,
 	    Uint *szp)
@@ -4219,7 +4538,7 @@ sz_info_fix(Allctr_t *allctr,
 		UWord used = fix->type_size * fix->u.cpool.used;
 
 		if (print_to_p) {
-		    int to = *print_to_p;
+		    fmtfn_t to = *print_to_p;
 		    void *arg = print_to_arg;
 		    erts_print(to,
 			       arg,
@@ -4247,7 +4566,7 @@ sz_info_fix(Allctr_t *allctr,
 	    UWord used = fix->type_size*fix->u.nocpool.used;
 
 	    if (print_to_p) {
-		int to = *print_to_p;
+		fmtfn_t to = *print_to_p;
 		void *arg = print_to_arg;
 		erts_print(to,
 			   arg,
@@ -4273,7 +4592,7 @@ static Eterm
 sz_info_carriers(Allctr_t *allctr,
 		 CarriersStats_t *cs,
 		 char *prefix,
-		 int *print_to_p,
+		 fmtfn_t *print_to_p,
 		 void *print_to_arg,
 		 Uint **hpp,
 		 Uint *szp)
@@ -4282,7 +4601,7 @@ sz_info_carriers(Allctr_t *allctr,
     UWord curr_size = cs->curr.norm.mseg.size + cs->curr.norm.sys_alloc.size;
 
     if (print_to_p) {
-	int to = *print_to_p;
+	fmtfn_t to = *print_to_p;
 	void *arg = print_to_arg;
 	erts_print(to,
 		   arg,
@@ -4323,7 +4642,7 @@ static Eterm
 info_cpool(Allctr_t *allctr,
 	   int sz_only,
 	   char *prefix,
-	   int *print_to_p,
+	   fmtfn_t *print_to_p,
 	   void *print_to_arg,
 	   Uint **hpp,
 	   Uint *szp)
@@ -4340,7 +4659,7 @@ info_cpool(Allctr_t *allctr,
     }
 
     if (print_to_p) {
-	int to = *print_to_p;
+	fmtfn_t to = *print_to_p;
 	void *arg = print_to_arg;
 	if (!sz_only)
 	    erts_print(to, arg, "%sblocks: %bpu\n", prefix, nob);
@@ -4377,7 +4696,7 @@ static Eterm
 info_carriers(Allctr_t *allctr,
 	      CarriersStats_t *cs,
 	      char *prefix,
-	      int *print_to_p,
+	      fmtfn_t *print_to_p,
 	      void *print_to_arg,
 	      Uint **hpp,
 	      Uint *szp)
@@ -4389,7 +4708,7 @@ info_carriers(Allctr_t *allctr,
     curr_size = cs->curr.norm.mseg.size + cs->curr.norm.sys_alloc.size;
 
     if (print_to_p) {
-	int to = *print_to_p;
+	fmtfn_t to = *print_to_p;
 	void *arg = print_to_arg;
 	erts_print(to,
 		   arg,
@@ -4498,7 +4817,7 @@ make_name_atoms(Allctr_t *allctr)
     size_t prefix_len = strlen(allctr->name_prefix);
 
     if (prefix_len > MAX_ATOM_CHARACTERS + sizeof(realloc) - 1)
-	erl_exit(1,"Too long allocator name: %salloc\n",allctr->name_prefix);
+	erts_exit(ERTS_ERROR_EXIT,"Too long allocator name: %salloc\n",allctr->name_prefix);
 
     memcpy((void *) buf, (void *) allctr->name_prefix, prefix_len);
 
@@ -4515,7 +4834,7 @@ make_name_atoms(Allctr_t *allctr)
 
 static Eterm
 info_calls(Allctr_t *allctr,
-	   int *print_to_p,
+	   fmtfn_t *print_to_p,
 	   void *print_to_arg,
 	   Uint **hpp,
 	   Uint *szp)
@@ -4532,7 +4851,7 @@ info_calls(Allctr_t *allctr,
 	erts_print(TO, TOA, "%s%s calls: %b64u\n",PRFX,NAME,CC)
 
 	char *prefix = allctr->name_prefix;
-	int to = *print_to_p;
+	fmtfn_t to = *print_to_p;
 	void *arg = print_to_arg;
 
 	PRINT_CC_5(to, arg, prefix, "alloc",        allctr->calls.this_alloc);
@@ -4608,7 +4927,7 @@ info_calls(Allctr_t *allctr,
 
 static Eterm
 info_options(Allctr_t *allctr,
-             int *print_to_p,
+             fmtfn_t *print_to_p,
 	     void *print_to_arg,
 	     Uint **hpp,
 	     Uint *szp)
@@ -4643,9 +4962,6 @@ info_options(Allctr_t *allctr,
 		   "option e: true\n"
 		   "option t: %s\n"
 		   "option ramv: %s\n"
-#if HALFWORD_HEAP
-		   "option low: %s\n"
-#endif
 		   "option sbct: %beu\n"
 #if HAVE_ERTS_MSEG
 		   "option asbcst: %bpu\n"
@@ -4664,9 +4980,6 @@ info_options(Allctr_t *allctr,
 		   "option acul: %d\n",
 		   topt,
 		   allctr->ramv ? "true" : "false",
-#if HALFWORD_HEAP
-		   allctr->mseg_opt.low_mem ? "true" : "false",
-#endif
 		   allctr->sbc_threshold,
 #if HAVE_ERTS_MSEG
 		   allctr->mseg_opt.abs_shrink_th,
@@ -4729,9 +5042,6 @@ info_options(Allctr_t *allctr,
 	add_2tup(hpp, szp, &res,
 		 am.sbct,
 		 bld_uint(hpp, szp, allctr->sbc_threshold));
-#if HALFWORD_HEAP
-	add_2tup(hpp, szp, &res, am.low, allctr->mseg_opt.low_mem ? am_true : am_false);
-#endif
 	add_2tup(hpp, szp, &res, am.ramv, allctr->ramv ? am_true : am_false);
 	add_2tup(hpp, szp, &res, am.t, (allctr->t ? am_true : am_false));
 	add_2tup(hpp, szp, &res, am.e, am_true);
@@ -4769,7 +5079,7 @@ reset_max_values(CarriersStats_t *cs)
 \*                                                                         */
 
 Eterm
-erts_alcu_au_info_options(int *print_to_p, void *print_to_arg,
+erts_alcu_au_info_options(fmtfn_t *print_to_p, void *print_to_arg,
 			  Uint **hpp, Uint *szp)
 {
     Eterm res = THE_NON_VALUE;    
@@ -4812,7 +5122,7 @@ erts_alcu_au_info_options(int *print_to_p, void *print_to_arg,
 
 Eterm
 erts_alcu_info_options(Allctr_t *allctr,
-		       int *print_to_p,
+		       fmtfn_t *print_to_p,
 		       void *print_to_arg,
 		       Uint **hpp,
 		       Uint *szp)
@@ -4844,7 +5154,7 @@ Eterm
 erts_alcu_sz_info(Allctr_t *allctr,
 		  int internal,
 		  int begin_max_period,
-		  int *print_to_p,
+		  fmtfn_t *print_to_p,
 		  void *print_to_arg,
 		  Uint **hpp,
 		  Uint *szp)
@@ -4876,7 +5186,7 @@ erts_alcu_sz_info(Allctr_t *allctr,
 
     ERTS_ALCU_DBG_CHK_THR_ACCESS(allctr);
 
-    /* Update sbc values not continously updated */
+    /* Update sbc values not continuously updated */
     allctr->sbcs.blocks.curr.no
 	= allctr->sbcs.curr.norm.mseg.no + allctr->sbcs.curr.norm.sys_alloc.no;
     allctr->sbcs.blocks.max.no = allctr->sbcs.max.no;
@@ -4930,7 +5240,7 @@ Eterm
 erts_alcu_info(Allctr_t *allctr,
 	       int internal,
 	       int begin_max_period,
-	       int *print_to_p,
+	       fmtfn_t *print_to_p,
 	       void *print_to_arg,
 	       Uint **hpp,
 	       Uint *szp)
@@ -4962,7 +5272,7 @@ erts_alcu_info(Allctr_t *allctr,
 
     ERTS_ALCU_DBG_CHK_THR_ACCESS(allctr);
 
-    /* Update sbc values not continously updated */
+    /* Update sbc values not continuously updated */
     allctr->sbcs.blocks.curr.no
 	= allctr->sbcs.curr.norm.mseg.no + allctr->sbcs.curr.norm.sys_alloc.no;
     allctr->sbcs.blocks.max.no = allctr->sbcs.max.no;
@@ -5126,6 +5436,11 @@ do_erts_alcu_alloc(ErtsAlcType_t type, void *extra, Uint size)
 void *erts_alcu_alloc(ErtsAlcType_t type, void *extra, Uint size)
 {
     void *res;
+#ifdef ERTS_SMP
+    ASSERT(!"This is not thread safe");
+#elif defined(USE_THREADS)
+    ASSERT(erts_equal_tids(erts_main_thread, erts_thr_self()));
+#endif
     res = do_erts_alcu_alloc(type, extra, size);
     DEBUG_CHECK_ALIGNMENT(res);
     return res;
@@ -5416,11 +5731,7 @@ do_erts_alcu_realloc(ErtsAlcType_t type,
 	Block_t *new_blk;
 	if(IS_SBC_BLK(blk)) {
 	do_carrier_resize:
-#if HALFWORD_HEAP
-	    new_blk = resize_carrier(allctr, blk, size, CFLG_SBC | CFLG_FORCE_MSEG);
-#else
 	    new_blk = resize_carrier(allctr, blk, size, CFLG_SBC);
-#endif
 	    res = new_blk ? BLK2UMEM(new_blk) : NULL;
 	}
 	else if (alcu_flgs & ERTS_ALCU_FLG_FAIL_REALLOC_MOVE)
@@ -5700,7 +6011,7 @@ erts_alcu_start(Allctr_t *allctr, AllctrInit_t *init)
     /* erts_alcu_start assumes that allctr has been zeroed */
 
     if (((UWord)allctr & ERTS_CRR_ALCTR_FLG_MASK) != 0) {
-        erl_exit(ERTS_ABORT_EXIT, "%s:%d:erts_alcu_start: Alignment error\n",
+        erts_exit(ERTS_ABORT_EXIT, "%s:%d:erts_alcu_start: Alignment error\n",
                  __FILE__, __LINE__);
     }
 
@@ -5714,11 +6025,8 @@ erts_alcu_start(Allctr_t *allctr, AllctrInit_t *init)
 #ifdef ERTS_SMP
     if (init->tspec || init->tpref)
 	allctr->mseg_opt.sched_spec = 1;
-#endif
-# if HALFWORD_HEAP
-    allctr->mseg_opt.low_mem = init->low_mem;
-# endif
-#endif
+#endif /* ERTS_SMP */
+#endif /* HAVE_ERTS_MSEG */
 
     allctr->name_prefix			= init->name_prefix;
     if (!allctr->name_prefix)
@@ -5775,7 +6083,7 @@ erts_alcu_start(Allctr_t *allctr, AllctrInit_t *init)
 	goto error;
     allctr->min_block_size		= UNIT_CEILING(allctr->min_block_size
 						       + sizeof(FreeBlkFtr_t));
-#if ERTS_SMP
+#ifdef ERTS_SMP
     if (init->tpref) {
 	Uint sz = ABLK_HDR_SZ;
 	sz += (init->fix ? 
@@ -5826,18 +6134,18 @@ erts_alcu_start(Allctr_t *allctr, AllctrInit_t *init)
 #ifdef USE_THREADS
     if (init->ts) {
 	allctr->thread_safe = 1;
-	
+
 #ifdef ERTS_ENABLE_LOCK_COUNT
 	erts_mtx_init_x_opt(&allctr->mutex,
 			    "alcu_allocator",
 			    make_small(allctr->alloc_no),
-			    ERTS_LCNT_LT_ALLOC,1);
+			    ERTS_LCNT_LT_ALLOC);
 #else
 	erts_mtx_init_x(&allctr->mutex,
 			"alcu_allocator",
-			make_small(allctr->alloc_no),1);
+			make_small(allctr->alloc_no));
 #endif /*ERTS_ENABLE_LOCK_COUNT*/
-	
+
 #ifdef DEBUG
 	allctr->debug.saved_tid = 0;
 #endif
@@ -5867,24 +6175,60 @@ erts_alcu_start(Allctr_t *allctr, AllctrInit_t *init)
 					    + ABLK_HDR_SZ)
 			       - ABLK_HDR_SZ);
 
+    if (init->sys_alloc) {
+        ASSERT(init->sys_realloc && init->sys_dealloc);
+        allctr->sys_alloc   = init->sys_alloc;
+        allctr->sys_realloc = init->sys_realloc;
+        allctr->sys_dealloc = init->sys_dealloc;
+    }
+    else {
+        ASSERT(!init->sys_realloc && !init->sys_dealloc);
+        allctr->sys_alloc   = &erts_alcu_sys_alloc;
+        allctr->sys_realloc = &erts_alcu_sys_realloc;
+        allctr->sys_dealloc = &erts_alcu_sys_dealloc;
+    }
+#if HAVE_ERTS_MSEG
+    if (init->mseg_alloc) {
+        ASSERT(init->mseg_realloc && init->mseg_dealloc);
+        allctr->mseg_alloc   = init->mseg_alloc;
+        allctr->mseg_realloc = init->mseg_realloc;
+        allctr->mseg_dealloc = init->mseg_dealloc;
+        allctr->mseg_mmapper = init->mseg_mmapper;
+    }
+    else {
+        ASSERT(!init->mseg_realloc && !init->mseg_dealloc);
+        allctr->mseg_alloc   = &erts_alcu_mseg_alloc;
+        allctr->mseg_realloc = &erts_alcu_mseg_realloc;
+        allctr->mseg_dealloc = &erts_alcu_mseg_dealloc;
+    }
+    /* If a custom carrier alloc function is specified, make sure it's used */
+    if (init->mseg_alloc && !init->sys_alloc) {
+        allctr->crr_set_flgs = CFLG_FORCE_MSEG;
+        allctr->crr_clr_flgs = CFLG_FORCE_SYS_ALLOC;
+    }
+    else if (!init->mseg_alloc && init->sys_alloc) {
+        allctr->crr_set_flgs = CFLG_FORCE_SYS_ALLOC;
+        allctr->crr_clr_flgs = CFLG_FORCE_MSEG;
+    }
+#endif
+
     if (allctr->main_carrier_size) {
 	Block_t *blk;
 
 	blk = create_carrier(allctr,
 			     allctr->main_carrier_size,
-			     CFLG_MBC
+                             (ERTS_SUPER_ALIGNED_MSEG_ONLY
+                              ? CFLG_FORCE_MSEG : CFLG_FORCE_SYS_ALLOC)
+                             | CFLG_MBC
 			     | CFLG_FORCE_SIZE
 			     | CFLG_NO_CPOOL
-#if !HALFWORD_HEAP && !ERTS_SUPER_ALIGNED_MSEG_ONLY
-			     | CFLG_FORCE_SYS_ALLOC
-#endif
 			     | CFLG_MAIN_CARRIER);
 	if (!blk) {
 #ifdef USE_THREADS
 	  if (allctr->thread_safe)
 	    erts_mtx_destroy(&allctr->mutex);
 #endif
-	  erl_exit(ERTS_ABORT_EXIT,
+	  erts_exit(ERTS_ABORT_EXIT,
 	    "Failed to create main carrier for %salloc\n",
 	    init->name_prefix);
 	}
@@ -6029,7 +6373,7 @@ erts_alcu_test(UWord op, UWord a1, UWord a2)
     case 0x019:	return (UWord) PREV_BLK((Block_t *) a1);
     case 0x01a: return (UWord) IS_MBC_FIRST_BLK((Allctr_t*)a1, (Block_t *) a2);
     case 0x01b: return (UWord) sizeof(Unit_t);
-    case 0x01c: return (unsigned long) BLK_TO_MBC((Block_t*) a1);
+    case 0x01c: return (UWord) BLK_TO_MBC((Block_t*) a1);
     case 0x01d: ((Allctr_t*) a1)->add_mbc((Allctr_t*)a1, (Carrier_t*)a2); break;
     case 0x01e: ((Allctr_t*) a1)->remove_mbc((Allctr_t*)a1, (Carrier_t*)a2); break;
 #ifdef ERTS_SMP
@@ -6054,6 +6398,16 @@ erts_alcu_test(UWord op, UWord a1, UWord a2)
     case 0x023: return (UWord) 0;
     case 0x024: return (UWord) 0;
 #endif
+    case 0x025: /* UMEM2BLK_TEST*/
+#ifdef DEBUG
+# ifdef HARD_DEBUG
+	return (UWord)UMEM2BLK(a1-3*sizeof(UWord));
+# else
+	return (UWord)UMEM2BLK(a1-2*sizeof(UWord));
+# endif
+#else
+	return (UWord)UMEM2BLK(a1);
+#endif
 
     default:	ASSERT(0); return ~((UWord) 0);
     }
@@ -6090,7 +6444,7 @@ erts_alcu_verify_unused(Allctr_t *allctr)
     if (no) {
 	UWord sz = allctr->sbcs.blocks.curr.size;
 	sz += allctr->mbcs.blocks.curr.size;
-	erl_exit(ERTS_ABORT_EXIT,
+	erts_exit(ERTS_ABORT_EXIT,
 		 "%salloc() used when expected to be unused!\n"
 		 "Total amount of blocks allocated: %bpu\n"
 		 "Total amount of bytes allocated: %bpu\n",
@@ -6130,11 +6484,6 @@ check_blk_carrier(Allctr_t *allctr, Block_t *iblk)
 
 	ASSERT(SBC2BLK(allctr, sbc) == iblk);
 	ASSERT(CARRIER_SZ(sbc) - SBC_HEADER_SIZE >= SBC_BLK_SZ(iblk));
-#if HAVE_ERTS_MSEG
-	if (IS_MSEG_CARRIER(sbc)) {
-	    ASSERT(CARRIER_SZ(sbc) % ERTS_SACRR_UNIT_SZ == 0);
-	}
-#endif
 	crr = sbc;
 	cl = &allctr->sbc_list;
     }