Merge branch 'sverk/cpool-search-improvement' into maint

OTP-12323

* sverk/cpool-search-improvement:
  erts: Add internal docs for super carrier and new cpool search.
  erts: Fix bug causing mbc to be deleted from cpool before it was inserted
  erts: Fix bug causing mbc removed from cpool to be used at pool entrance
  erts: Add pooled_list and traitor_list

4 files changed, 565 insertions, 54 deletions

diff --git a/erts/emulator/beam/erl_alloc_util.c b/erts/emulator/beam/erl_alloc_util.c
index 55052430e1..e3172dc4fb 100644
--- a/erts/emulator/beam/erl_alloc_util.c
+++ b/erts/emulator/beam/erl_alloc_util.c
@@ -205,7 +205,7 @@ MBC after deallocating first block:
      ASSERT(((UWord)(F) & (~FLG_MASK|THIS_FREE_BLK_HDR_FLG|PREV_FREE_BLK_HDR_FLG)) == THIS_FREE_BLK_HDR_FLG), \
      (B)->bhdr = ((Sz) | (F)), \
      (B)->u.carrier = (C))
-      
+
 #  define IS_MBC_FIRST_ABLK(AP,B) \
   ((((UWord)(B) & ~ERTS_SACRR_UNIT_MASK) == MBC_HEADER_SIZE(AP)) \
    && ((B)->bhdr & MBC_ABLK_OFFSET_MASK) == 0)
@@ -378,9 +378,8 @@ do {										\
 
 #ifdef ERTS_SMP
 #define SBC_HEADER_SIZE	   						\
-    (UNIT_CEILING(sizeof(Carrier_t)					\
-		  - sizeof(ErtsAlcCPoolData_t)				\
-		  + ABLK_HDR_SZ)					\
+    (UNIT_CEILING(offsetof(Carrier_t, cpool)                            \
+	          + ABLK_HDR_SZ)	                                \
      - ABLK_HDR_SZ)
 #else
 #define SBC_HEADER_SIZE	   						\
@@ -929,6 +928,88 @@ unlink_carrier(CarrierList_t *cl, Carrier_t *crr)
 
 #ifdef ERTS_SMP
 
+#ifdef DEBUG
+static int is_in_list(ErtsDoubleLink_t* sentinel, ErtsDoubleLink_t* node)
+{
+    ErtsDoubleLink_t* p;
+
+    ASSERT(node != sentinel);
+    for (p = sentinel->next; p != sentinel; p = p->next) {
+	if (p == node)
+	    return 1;
+    }
+    return 0;
+}
+#endif /* DEBUG */
+
+static ERTS_INLINE void
+link_edl_after(ErtsDoubleLink_t* after_me, ErtsDoubleLink_t* node)
+{
+    ErtsDoubleLink_t* before_me = after_me->next;
+    ASSERT(node != after_me && node != before_me);
+    node->next = before_me;
+    node->prev = after_me;
+    before_me->prev = node;
+    after_me->next = node;
+}
+
+static ERTS_INLINE void
+link_edl_before(ErtsDoubleLink_t* before_me, ErtsDoubleLink_t* node)
+{
+    ErtsDoubleLink_t* after_me = before_me->prev;
+    ASSERT(node != before_me && node != after_me);
+    node->next = before_me;
+    node->prev = after_me;
+    before_me->prev = node;
+    after_me->next = node;
+}
+
+static ERTS_INLINE void
+unlink_edl(ErtsDoubleLink_t* node)
+{
+    node->next->prev = node->prev;
+    node->prev->next = node->next;
+}
+
+static ERTS_INLINE void
+relink_edl_before(ErtsDoubleLink_t* before_me, ErtsDoubleLink_t* node)
+{
+    if (node != before_me && node != before_me->prev) {
+	unlink_edl(node);
+	link_edl_before(before_me, node);
+    }
+}
+
+static ERTS_INLINE int is_abandoned(Carrier_t *crr)
+{
+    return crr->cpool.abandoned.next != NULL;
+}
+
+static ERTS_INLINE void
+link_abandoned_carrier(ErtsDoubleLink_t* list, Carrier_t *crr)
+{
+    ASSERT(!is_abandoned(crr));
+
+    link_edl_after(list, &crr->cpool.abandoned);
+
+    ASSERT(crr->cpool.abandoned.next != &crr->cpool.abandoned);
+    ASSERT(crr->cpool.abandoned.prev != &crr->cpool.abandoned);
+}
+
+static ERTS_INLINE void
+unlink_abandoned_carrier(Carrier_t *crr)
+{
+    ASSERT(is_in_list(&crr->cpool.orig_allctr->cpool.pooled_list,
+		      &crr->cpool.abandoned) ||
+	   is_in_list(&crr->cpool.orig_allctr->cpool.traitor_list,
+		      &crr->cpool.abandoned));
+
+    unlink_edl(&crr->cpool.abandoned);
+
+    crr->cpool.abandoned.next = NULL;
+    crr->cpool.abandoned.prev = NULL;
+}
+
 static ERTS_INLINE void
 clear_busy_pool_carrier(Allctr_t *allctr, Carrier_t *crr)
 {
@@ -955,7 +1036,7 @@ clear_busy_pool_carrier(Allctr_t *allctr, Carrier_t *crr)
     }
 }
 
-#endif
+#endif /* ERTS_SMP */
 
 #if 0
 #define ERTS_DBG_CHK_FIX_LIST(A, FIX, IX, B)			\
@@ -2575,10 +2656,9 @@ mbc_realloc(Allctr_t *allctr, void *p, Uint size, Uint32 alcu_flgs,
 #ifdef ERTS_SMP
 
 #define ERTS_ALC_MAX_DEALLOC_CARRIER		10
-#define ERTS_ALC_CPOOL_MAX_FETCH_INSPECT	10
+#define ERTS_ALC_CPOOL_MAX_FETCH_INSPECT	20
+#define ERTS_ALC_CPOOL_MAX_TRAITOR_INSPECT	10
 #define ERTS_ALC_CPOOL_CHECK_LIMIT_COUNT	100
-#define ERTS_ALC_CPOOL_MAX_NO_CARRIERS		5
-#define ERTS_ALC_CPOOL_INSERT_ALLOWED_OFFSET	100
 #define ERTS_ALC_CPOOL_MAX_FAILED_STAT_READS	3
 
 #define ERTS_ALC_CPOOL_PTR_MOD_MRK		(((erts_aint_t) 1) << 0)
@@ -2755,9 +2835,6 @@ cpool_insert(Allctr_t *allctr, Carrier_t *crr)
 			(erts_aint_t) CARRIER_SZ(crr));
     erts_atomic_inc_nob(&allctr->cpool.stat.no_carriers);
 
-    erts_smp_atomic_set_nob(&crr->allctr,
-			    ((erts_aint_t) allctr)|ERTS_CRR_ALCTR_FLG_IN_POOL);
-
     /*
      * We search in 'next' direction and begin by passing
      * one element before trying to insert. This in order to
@@ -2816,6 +2893,9 @@ cpool_insert(Allctr_t *allctr, Carrier_t *crr)
     cpool_set_mod_marked(&cpd2p->prev,
 			 (erts_aint_t) &crr->cpool,
 			 (erts_aint_t) cpd1p);
+
+    erts_smp_atomic_set_wb(&crr->allctr,
+			   ((erts_aint_t) allctr)|ERTS_CRR_ALCTR_FLG_IN_POOL);
 }
 
 static void
@@ -2916,59 +2996,163 @@ cpool_delete(Allctr_t *allctr, Allctr_t *prev_allctr, Carrier_t *crr)
 static Carrier_t *
 cpool_fetch(Allctr_t *allctr, UWord size)
 {
-    int i;
+    int i, i_stop, has_passed_sentinel;
     Carrier_t *crr;
     ErtsAlcCPoolData_t *cpdp;
-    ErtsAlcCPoolData_t *sentinel = &carrier_pool[allctr->alloc_no].sentinel;
+    ErtsAlcCPoolData_t *cpool_entrance;
+    ErtsAlcCPoolData_t *sentinel;
+    ErtsDoubleLink_t* dl;
+    ErtsDoubleLink_t* first_old_traitor;
 
     ERTS_ALC_CPOOL_ASSERT(allctr->alloc_no == ERTS_ALC_A_INVALID /* testcase */
 			  || erts_thr_progress_is_managed_thread());
 
-    i = 0;
+    i = ERTS_ALC_CPOOL_MAX_FETCH_INSPECT;
+    first_old_traitor = allctr->cpool.traitor_list.next;
+    cpool_entrance = NULL;
 
-    /* First; check our own pending dealloc carrier list... */
-    crr = allctr->cpool.dc_list.last;
-    while (crr && i < ERTS_ALC_CPOOL_MAX_FETCH_INSPECT) {
-	if (erts_atomic_read_nob(&crr->cpool.max_size) >= size) {
-	    unlink_carrier(&allctr->cpool.dc_list, crr);
-#ifdef ERTS_ALC_CPOOL_DEBUG
-	    ERTS_ALC_CPOOL_ASSERT(erts_smp_atomic_xchg_nob(&crr->allctr,
-							   ((erts_aint_t) allctr))
-				  == (((erts_aint_t) allctr) & ~ERTS_CRR_ALCTR_FLG_MASK));
-#else
-	    erts_smp_atomic_set_nob(&crr->allctr, ((erts_aint_t) allctr));
-#endif
-	    return crr;
+    /*
+     * Search my own pooled_list,
+     * i.e my abandoned carriers that were in the pool last time I checked.
+     */
+
+    dl = allctr->cpool.pooled_list.next;
+    while(dl != &allctr->cpool.pooled_list) {
+	erts_aint_t exp, act;
+	crr = (Carrier_t *) (((char *) dl) - offsetof(Carrier_t, cpool.abandoned));
+
+	ASSERT(!is_in_list(&allctr->cpool.traitor_list, dl));
+	ASSERT(crr->cpool.orig_allctr == allctr);
+	dl = dl->next;
+	exp = erts_smp_atomic_read_rb(&crr->allctr);
+	if ((exp & ERTS_CRR_ALCTR_FLG_MASK) == ERTS_CRR_ALCTR_FLG_IN_POOL
+	    && erts_atomic_read_nob(&crr->cpool.max_size) >= size) {
+	    /* Try to fetch it... */
+	    act = erts_smp_atomic_cmpxchg_mb(&crr->allctr,
+					     (erts_aint_t) allctr,
+					     exp);
+	    if (act == exp) {
+		cpool_delete(allctr, ((Allctr_t *) (act & ~ERTS_CRR_ALCTR_FLG_MASK)), crr);
+		unlink_abandoned_carrier(crr);
+
+		/* Move sentinel to continue next search from here */
+		relink_edl_before(dl, &allctr->cpool.pooled_list);
+		return crr;
+	    }
+	    exp = act;
+	}
+	if (exp & ERTS_CRR_ALCTR_FLG_IN_POOL) {
+	    if (!cpool_entrance)
+		cpool_entrance = &crr->cpool;
+	}
+	else { /* Not in pool, move to traitor_list */
+	    unlink_abandoned_carrier(crr);
+	    link_abandoned_carrier(&allctr->cpool.traitor_list, crr);
+	}
+	if (--i <= 0) {
+	    /* Move sentinel to continue next search from here */
+	    relink_edl_before(dl, &allctr->cpool.pooled_list);
+	    return NULL;
 	}
-	crr = crr->prev;
-	i++;
     }
 
-    /* ... then the pool ... */
+    /* Now search traitor_list.
+     * i.e carriers employed by other allocators last time I checked.
+     * They might have been abandoned since then.
+     */
+
+    i_stop = (i < ERTS_ALC_CPOOL_MAX_TRAITOR_INSPECT ?
+	      0 : i - ERTS_ALC_CPOOL_MAX_TRAITOR_INSPECT);
+    dl = first_old_traitor;
+    while(dl != &allctr->cpool.traitor_list) {
+	erts_aint_t exp, act;
+	crr = (Carrier_t *) (((char *) dl) - offsetof(Carrier_t, cpool.abandoned));
+	ASSERT(dl != &allctr->cpool.pooled_list);
+	ASSERT(crr->cpool.orig_allctr == allctr);
+	dl = dl->next;
+	exp = erts_smp_atomic_read_rb(&crr->allctr);
+	if (exp & ERTS_CRR_ALCTR_FLG_IN_POOL) {
+	    if (!(exp & ERTS_CRR_ALCTR_FLG_BUSY)
+		&& erts_atomic_read_nob(&crr->cpool.max_size) >= size) {
+		/* Try to fetch it... */
+		act = erts_smp_atomic_cmpxchg_mb(&crr->allctr,
+						 (erts_aint_t) allctr,
+						 exp);
+		if (act == exp) {
+		    cpool_delete(allctr, ((Allctr_t *) (act & ~ERTS_CRR_ALCTR_FLG_MASK)), crr);
+		    unlink_abandoned_carrier(crr);
+
+		    /* Move sentinel to continue next search from here */
+		    relink_edl_before(dl, &allctr->cpool.traitor_list);
+		    return crr;
+		}
+		exp = act;
+	    }
+	    if (exp & ERTS_CRR_ALCTR_FLG_IN_POOL) {
+		if (!cpool_entrance)
+		    cpool_entrance = &crr->cpool;
+
+		/* Move to pooled_list */
+		unlink_abandoned_carrier(crr);
+		link_abandoned_carrier(&allctr->cpool.pooled_list, crr);
+	    }
+	}
+	if (--i <= i_stop) {
+	    /* Move sentinel to continue next search from here */
+	    relink_edl_before(dl, &allctr->cpool.traitor_list);
+	    if (i > 0)
+		break;
+	    else
+		return NULL;
+	}
+    }
 
     /*
-     * We search in 'prev' direction and begin by passing
-     * one element before trying to fetch. This in order to
-     * avoid contention with threads inserting elements.
+     * Finally search the shared pool and try employ foreign carriers
      */
 
-    cpdp = cpool_aint2cpd(cpool_read(&sentinel->prev));
-    if (cpdp == sentinel)
-	return NULL;
+    sentinel = &carrier_pool[allctr->alloc_no].sentinel;
+    if (cpool_entrance) {
+	/* We saw a pooled carried above, use it as entrance into the pool
+	 */
+	cpdp = cpool_entrance;
+    }
+    else {
+	/* No pooled carried seen above. Start search at cpool sentinel,
+	 * but begin by passing one element before trying to fetch.
+	 * This in order to avoid contention with threads inserting elements.
+	 */
+	cpool_entrance = sentinel;
+	cpdp = cpool_aint2cpd(cpool_read(&cpool_entrance->prev));
+	if (cpdp == sentinel)
+	    return NULL;
+    }
 
-    while (i < ERTS_ALC_CPOOL_MAX_FETCH_INSPECT) {
+    has_passed_sentinel = 0;
+    while (1) {
 	erts_aint_t exp;
 	cpdp = cpool_aint2cpd(cpool_read(&cpdp->prev));
-	if (cpdp == sentinel) {
+	if (cpdp == cpool_entrance) {
+	    if (cpool_entrance == sentinel) {
+		cpdp = cpool_aint2cpd(cpool_read(&cpdp->prev));
+		if (cpdp == sentinel)
+		    return NULL;
+	    }
+	    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;
+	    }
 	    cpdp = cpool_aint2cpd(cpool_read(&cpdp->prev));
 	    if (cpdp == sentinel)
 		return NULL;
-	    i = ERTS_ALC_CPOOL_MAX_FETCH_INSPECT; /* Last one to inspect */
+	    has_passed_sentinel = 1;
 	}
-	crr = (Carrier_t *) (((char *) cpdp) - offsetof(Carrier_t, cpool));
+	crr = (Carrier_t *)(((char *)cpdp) - offsetof(Carrier_t, cpool));
 	exp = erts_smp_atomic_read_rb(&crr->allctr);
-	if (((exp & (ERTS_CRR_ALCTR_FLG_IN_POOL|ERTS_CRR_ALCTR_FLG_BUSY))
-	     == ERTS_CRR_ALCTR_FLG_IN_POOL)
+	if (((exp & (ERTS_CRR_ALCTR_FLG_MASK)) == ERTS_CRR_ALCTR_FLG_IN_POOL)
 	    && (erts_atomic_read_nob(&cpdp->max_size) >= size)) {
 	    erts_aint_t act;
 	    /* Try to fetch it... */
@@ -2977,11 +3161,35 @@ cpool_fetch(Allctr_t *allctr, UWord size)
 					     exp);
 	    if (act == exp) {
 		cpool_delete(allctr, ((Allctr_t *) (act & ~ERTS_CRR_ALCTR_FLG_MASK)), crr);
+		if (crr->cpool.orig_allctr == allctr) {
+		    unlink_abandoned_carrier(crr);
+		}
 		return crr;
 	    }
 	}
-	i++;
+	if (--i <= 0)
+	    return NULL;
     }
+
+    /* 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) {
+	    unlink_carrier(&allctr->cpool.dc_list, crr);
+#ifdef ERTS_ALC_CPOOL_DEBUG
+	    ERTS_ALC_CPOOL_ASSERT(erts_smp_atomic_xchg_nob(&crr->allctr,
+							   ((erts_aint_t) allctr))
+				  == (((erts_aint_t) allctr) & ~ERTS_CRR_ALCTR_FLG_MASK));
+#else
+	    erts_smp_atomic_set_nob(&crr->allctr, ((erts_aint_t) allctr));
+#endif
+	    return crr;
+	}
+	crr = crr->prev;
+	if (--i <= 0)
+	    return NULL;
+    }
+
     return NULL;
 }
 
@@ -3078,6 +3286,9 @@ schedule_dealloc_carrier(Allctr_t *allctr, Carrier_t *crr)
 	return;
     }
 
+    if (is_abandoned(crr))
+	unlink_abandoned_carrier(crr);
+
     if (crr->cpool.thr_prgr == ERTS_THR_PRGR_INVALID
 	|| erts_thr_progress_has_reached(crr->cpool.thr_prgr)) {
 	dealloc_carrier(allctr, crr, 1);
@@ -3124,6 +3335,8 @@ cpool_init_carrier_data(Allctr_t *allctr, Carrier_t *crr)
 	    limit = (csz/100)*allctr->cpool.util_limit;
 	crr->cpool.abandon_limit = limit;
     }
+    crr->cpool.abandoned.next = NULL;
+    crr->cpool.abandoned.prev = NULL;
 }
 
 static void
@@ -3154,6 +3367,9 @@ abandon_carrier(Allctr_t *allctr, Carrier_t *crr)
     STAT_MBC_CPOOL_INSERT(allctr, crr);
 
     unlink_carrier(&allctr->mbc_list, crr);
+    if (crr->cpool.orig_allctr == allctr) {
+	link_abandoned_carrier(&allctr->cpool.pooled_list, crr);
+    }
 
     allctr->remove_mbc(allctr, crr);
 
@@ -3661,6 +3877,11 @@ destroy_carrier(Allctr_t *allctr, Block_t *blk, Carrier_t **busy_pcrr_pp)
 	if (busy_pcrr_pp && *busy_pcrr_pp) {
 	    ERTS_ALC_CPOOL_ASSERT(*busy_pcrr_pp == crr);
 	    *busy_pcrr_pp = NULL;
+	    ERTS_ALC_CPOOL_ASSERT(erts_smp_atomic_read_nob(&crr->allctr)
+				  == (((erts_aint_t) allctr)
+				      | ERTS_CRR_ALCTR_FLG_IN_POOL
+				      | ERTS_CRR_ALCTR_FLG_BUSY));
+	    erts_smp_atomic_set_nob(&crr->allctr, ((erts_aint_t) allctr));
 	    cpool_delete(allctr, allctr, crr);
 	}
 	else
@@ -5540,6 +5761,10 @@ erts_alcu_start(Allctr_t *allctr, AllctrInit_t *init)
 	    allctr->min_block_size = sz;
     }
 
+    allctr->cpool.pooled_list.next  = &allctr->cpool.pooled_list;
+    allctr->cpool.pooled_list.prev  = &allctr->cpool.pooled_list;
+    allctr->cpool.traitor_list.next = &allctr->cpool.traitor_list;
+    allctr->cpool.traitor_list.prev = &allctr->cpool.traitor_list;
     allctr->cpool.dc_list.first = NULL;
     allctr->cpool.dc_list.last = NULL;
     allctr->cpool.abandon_limit = 0;
diff --git a/erts/emulator/beam/erl_alloc_util.h b/erts/emulator/beam/erl_alloc_util.h
index 7be6b1ed9d..eee920e66c 100644
--- a/erts/emulator/beam/erl_alloc_util.h
+++ b/erts/emulator/beam/erl_alloc_util.h
@@ -268,6 +268,11 @@ typedef union {char c[ERTS_ALLOC_ALIGN_BYTES]; long l; double d;} Unit_t;
 
 #ifdef ERTS_SMP
 
+typedef struct ErtsDoubleLink_t_ {
+    struct ErtsDoubleLink_t_ *next;
+    struct ErtsDoubleLink_t_ *prev;
+}ErtsDoubleLink_t;
+
 typedef struct {
     erts_atomic_t next;
     erts_atomic_t prev;
@@ -277,6 +282,7 @@ typedef struct {
     UWord abandon_limit;
     UWord blocks;
     UWord blocks_size;
+    ErtsDoubleLink_t abandoned; /* node in pooled_list or traitor_list */
 } ErtsAlcCPoolData_t;
 
 #endif
@@ -500,7 +506,12 @@ struct Allctr_t_ {
     CarrierList_t	sbc_list;
 #ifdef ERTS_SMP
     struct {
-	CarrierList_t	dc_list;
+	/* pooled_list, traitor list and dc_list contain only
+           carriers _created_ by this allocator */
+	ErtsDoubleLink_t pooled_list;
+	ErtsDoubleLink_t traitor_list;
+	CarrierList_t	 dc_list;
+
 	UWord		abandon_limit;
 	int		disable_abandon;
 	int		check_limit_count;
diff --git a/erts/emulator/internal_doc/CarrierMigration.md b/erts/emulator/internal_doc/CarrierMigration.md
index b93c11c6ec..7afdb70aef 100644
--- a/erts/emulator/internal_doc/CarrierMigration.md
+++ b/erts/emulator/internal_doc/CarrierMigration.md
@@ -146,28 +146,53 @@ Since the carrier has been unlinked from the data structure of
 available free blocks, no more allocations will be made in the
 carrier. The allocator instance putting the carrier into the pool,
 however, still has the responsibility of performing deallocations in
-it while it remains in the pool.
+it while it remains in the pool. The allocator instance with this
+deallocation responsibility is here called the **employer**.
 
-Each carrier has a flag field containing information about allocator
-instance owning the carrier, a flag indicating if the carrier is in
+Each carrier has a flag field containing information about the
+employing allocator instance, a flag indicating if the carrier is in
 the pool or not, and a flag indicating if it is busy or not. When the
-carrier is in the pool, the owning allocator instance needs to mark it
+carrier is in the pool, the employing allocator instance needs to mark it
 as busy while operating on it. If another thread inspects it in order
-to try to fetch it from the pool, it will abort the fetch if it is
-busy. When fetching the carrier from the pool, ownership will changed
-and further deallocations in the carrier will be redirected to the new
-owner using the delayed dealloc functionality.
+to try to fetch it from the pool, it will skip it if it is busy. When
+fetching the carrier from the pool, employment will change and further
+deallocations in the carrier will be redirected to the new
+employer using the delayed dealloc functionality.
 
 If a carrier in the pool becomes empty, it will be withdrawn from the
 pool. All carriers that become empty are also always passed to its
-originating allocator instance for deallocation using the delayed
+**owning** allocator instance for deallocation using the delayed
 dealloc functionality. Since carriers this way always will be
-deallocated by the allocator instance that allocated the carrier the
+deallocated by the owner, that allocated the carrier, the
 underlying functionality of allocating and deallocating carriers can
 remain simple and doesn't have to bother about multiple threads. In a
 NUMA system we will also not mix carriers originating from multiple
 NUMA nodes.
 
+In short:
+
+* The allocator instance that created a carrier **owns** it.
+* An empty carrier is always deallocated by its **owner**.
+* **Ownership** never changes.
+* The allocator instance that uses a carrier **employs** it.
+* An **employer** can abandon a carrier into the pool.
+* Pooled carriers are not allocated from.
+* Deallocation in a pooled carrier is still performed by its **employer**.
+* **Employment** can only change when a carrier is fetched from the pool.
+
+### Searching the pool ###
+
+To harbor real time characteristics, searching the pool is
+limited. We only inspect a limited number of carriers. If none of
+those carriers had a free block large enough to satisfy the allocation
+request, the search will fail. A carrier in the pool can also be busy,
+if another thread is currently doing block deallocation work on the
+carrier. A busy carrier will also be skipped by the search as it can
+not satisfy the request. The pool is lock free and we do not want to
+block, waiting for the other thread to finish.
+
+#### Before OTP 17.4 ####
+
 When an allocator instance needs more carrier space, it always begins
 by inspecting its own carriers that are waiting for thread progress
 before they can be deallocated. If no such carrier could be found, it
@@ -176,6 +201,65 @@ it will allocate a new carrier. Regardless of where the allocator
 instance gets the carrier from it the just links in the carrier into
 its data structure of free blocks.
 
+#### After OTP 17.4 ####
+
+The old search algorithm had a problem as the search always started at
+the same position in the pool, the sentinel. This could lead to
+contention from concurrent searching processes. But even worse, it
+could lead to a "bad" state when searches fail with a high rate
+leading to new carriers instead being allocated. These new carriers
+may later be inserted into the pool due to bad utilization. If the
+frequency of insertions into the pool is higher than successful
+fetching from the pool, memory will eventually get exhausted.
+
+This "bad" state, consist of a cluster of small and/or highly
+fragmented carriers located at the sentinel in the pool. The largest free
+block in such a "bad" carrier is rather small, making it not able to satisfy
+most allocation requests. As the search always started at the
+sentinel, any such "bad" carriers that had been left in the pool would
+eventually cluster together at the sentinel. All searches first
+have to skip past this cluster of "bad" carriers to reach a "good"
+carrier. When the cluster gets to the same size as the search limit,
+all searches will essentially fail.
+
+To counter the "bad cluster" problem and also ease the contention, the
+search will now always start by first looking at the allocators **own**
+carriers. That is, carriers that were initially created by the
+allocator itself and later had been abandoned to the pool. If none of
+our own abandoned carrier would do, then the search continues into the
+pool, as before, to look for carriers created by other
+allocators. However, if we have at least one abandoned carrier of our
+own, that could not satisfy the request, we can use that as entry point
+into the pool.
+
+The result is that we prefer carriers created by the thread itself,
+which is good for NUMA performance. And we get more entry points when
+searching the pool, which will ease contention and clustering.
+
+To do the first search among own carriers, every allocator instance
+has two new lists; `pooled_list` and `traitor_list`. These lists are only
+accessed by the allocator itself and they only contain the allocators
+own carriers. When an owned carrier is abandoned and put in the
+pool, it is also linked into `pooled_list`. When we search our
+`pooled_list` and find a carrier that is no longer in the pool, we
+move that carrier from `pooled_list` to `traitor_list` as it is now
+employed by another allocator. If searching `pooled_list` fails, we
+also do a limited search of `traitor_list`. When finding an abandoned
+carrier in `traitor_list` it is either employed, or moved back to
+`pooled_list` if it could not satisfy the allocation request.
+
+When searching `pooled_list` and `traitor_list` we always start at the
+point where the last search ended. This to avoid clustering
+problems and increase the probability to find a "good" carrier. As
+`pooled_list` and `traitor_list` are only accessed by the owning
+allocator instance, they need no thread synchronization at all.
+
+Furthermore, the search for own carriers that are scheduled
+for deallocation is now done as the last search option. The idea is
+that it is better to reuse a poorly utilized carrier, than to
+resurrect an empty carrier that was just about to be released back to
+the OS.
+
 ### Result ###
 
 The use of this strategy of abandoning carriers with poor utilization
diff --git a/erts/emulator/internal_doc/SuperCarrier.md b/erts/emulator/internal_doc/SuperCarrier.md
new file mode 100644
index 0000000000..0ad6af41de
--- /dev/null
+++ b/erts/emulator/internal_doc/SuperCarrier.md
@@ -0,0 +1,191 @@
+Super Carrier
+=============
+
+A super carrier is large memory area, allocated at VM start, which can
+be used during runtime to allocate normal carriers from.
+
+The super carrier feature was introduced in OTP R16B03. It is
+enabled with command line option +MMscs <size in Mb>
+and can be configured with other options.
+
+Problem
+-------
+
+The initial motivation for this feature was customers asking for a way
+to pre-allocate physcial memory at VM start for it to use.
+
+Other problems were different experienced limitations of the OS
+implementation of mmap:
+
+* Increasingly bad performance of mmap/munmap as the number of mmap'ed areas grow.
+* Fragmentation problem between mmap'ed areas.
+
+A third problem was management of low memory in the halfword
+emulator. The implementation used a naive linear search structure to
+hold free segments which would lead to poor performance when
+fragmentation increased.
+
+
+Solution
+--------
+
+Allocate one large continious area of address space at VM start and
+then use that area to satisfy our dynamic memory need during
+runtime. In other words: implement our own mmap.
+
+### Use cases ###
+
+If command line option +MMscrpm (Reserve Physical Memory) is set to
+false, only virtual space is allocated for the super carrier from
+start. The super carrier then acts as an "alternative mmap" implementation
+without changing the consumption of physical memory pages. Physical
+pages will be reserved on demand when an allocation is done from the super
+carrier and be unreserved when the memory is released back to the
+super carrier.
+
+If +MMscrpm is set to true, which is default, the initial allocation
+will reserve physical memory for the entire super carrier. This can be
+used by users that want to ensure a certain *minimum* amount of
+physical memory for the VM.
+
+However, what reservation of physical memory actually means highly
+depends on the operating system, and how it is configured. For
+example, different memory overcommit settings on Linux drastically
+change the behaviour.
+
+A third feature is to have the super carrier limit the *maximum*
+amount of memory used by the VM. If +MMsco (Super Carrier Only) is set
+to true, which is default, allocations will only be done from the
+super carrier. When the super carrier gets full, the VM will fail due
+to out of memory.
+If +MMsco is false, allocations will use mmap directly if the super
+carrier is full.
+
+
+
+### Implementation ###
+
+The entire super carrier implementation is kept in erl_mmap.c. The
+name suggest that it can be viewed as our own mmap implementation.
+
+A super carrier needs to satisfy two slightly different kinds of
+allocation requests; multi block carriers (MBC) and single block
+carriers (SBC). They are both rather large blocks of continious
+memory, but MBCs and SBCs have different demands on alignment and
+size.
+
+SBCs can have arbitrary size and do only need minimum 8-byte
+alignment.
+
+MBCs are more restricted. They can only have a number of fixed
+sizes that are powers of 2. The start address need to have a very
+large aligment (currently 256 kb, called "super alignment"). This is a
+design choice that allows very low overhead per allocated block in the
+MBC.
+
+To reduce fragmentation within the super carrier, it is good to keep SBCs
+and MBCs apart. MBCs with their uniform alignment and sizes can be
+packed very efficiently together. SBCs without demand for aligment can
+also be allocated quite efficiently together. But mixing them can lead
+to a lot of memory wasted when we need to create large holes of
+padding to the next alignment limit.
+
+The super carrier thus contains two areas. One area for MBCs growing from
+the bottom and up. And one area for SBCs growing from the top and
+down. Like a process with a heap and a stack growing towards each
+other.
+
+
+### Data structures ###
+
+The MBC area is called **sa** as in super aligned and the SBC area is
+called **sua** as in super un-aligned.
+
+Note that the "super" in super alignment and the "super" in super
+carrier has nothing to do with each other. We could have choosen
+another naming to avoid confusion, such as "meta" carrier or "giant"
+aligment.
+
+	+-------+ <---- sua.top
+	|  sua  |
+	|       |
+	|-------| <---- sua.bot
+	|       |
+	|       |
+	|       |
+	|-------| <---- sa.top
+	|       |
+	|  sa   |
+	|       |
+	+-------+ <---- sa.bot
+
+
+When a carrier is deallocated a free memory segment will be created
+inside the corresponding area, unless the carrier was at the very top
+(in `sa`) or bottom (in `sua`) in which case the area will just shrink
+down or up.
+
+We need to keep track of all the free segments in order to reuse them
+for new carrier allocations. One initial idea was to use the same
+mechanism that is used to keep track of free blocks within MBCs
+(alloc_util and the different strategies). However, that would not be
+as straight forward as one can think and can also waste quite a lot of
+memory as it uses prepended block headers. The granularity of the
+super carrier is one memory page (usually 4kb). We want to allocate
+and free entire pages and we don't want to waste an entire page just
+to hold the block header of the following pages.
+
+Instead we store the meta information about all the free segments in a
+dedicated area apart from the `sa` and `sua` areas. Every free segment is
+represented by a descriptor struct (`ErtsFreeSegDesc`).
+
+    typedef struct {
+        RBTNode snode;      /* node in 'stree' */
+        RBTNode anode;      /* node in 'atree' */
+        char* start;
+        char* end;
+    }ErtsFreeSegDesc;
+
+To find the smallest free segment that will satisfy a carrier allocation
+(best fit), the free segments are organized in a tree sorted by
+size (`stree`). We search in this tree at allocation. If no free segment of
+sufficient size was found, the area (`sa` or `sua`) is instead expanded.
+If two or more free segments with equal size exist, the one at lowest
+address is choosen for `sa` and highest address for `sua`.
+
+At carrier deallocation, we want to coalesce with any adjacent free
+segments, to form one large free segment. To do that, all free
+segments are also organized in a tree sorted in address order (`atree`).
+
+So, in total we keep four trees of free descriptors for the super
+carrier; two for `sa` and two for `sua`. They all use the same
+red-black-tree implementation that support the different sorting
+orders used.
+
+When allocating a new MBC we first search after a free segment in `sa`,
+then try to raise `sa.top`, and then as a fallback try to search after a
+free segment in `sua`. When an MBC is allocated in `sua`, a larger segment
+is allocated which is then trimmed to obtain the right
+alignment. Allocation search for an SBC is done in reverse order. When
+an SBC is allocated in `sa`, the size is aligned up to super aligned
+size.
+
+### The free descriptor area ###
+
+As mentioned above, the descriptors for the free segments are
+allocated in a separate area. This area has a constant configurable
+size (+MMscrfsd) that defaults to 65536 descriptors. This should be
+more than enough in most cases. If the descriptors area should fill up,
+new descriptor areas will be allocated first directly from the OS, and
+then from `sua` and `sa` in the super carrier, and lastly from the memory
+segment itself which is being deallocated. Allocating free descriptor
+areas from the super carrier is only a last resort, and should be
+avoided, as it creates fragmentation.
+
+### Halfword emulator ###
+
+The halfword emulator uses the super carrier implementation to manage
+its low memory mappings thar are needed for all term storage. The
+super carrier can here not be configured by command line options. One
+could imagine a second configurable instance of the super carrier used
+by high memory allocation, but that has not been implemented.