Refactor arenas array (fixes deadlock).

Refactor the arenas array, which contains pointers to all extant arenas,
such that it starts out as a sparse array of maximum size, and use
double-checked atomics-based reads as the basis for fast and simple
arena_get().  Additionally, reduce arenas_lock's role such that it only
protects against arena initalization races.  These changes remove the
possibility for arena lookups to trigger locking, which resolves at
least one known (fork-related) deadlock.

This resolves #315.

include/jemalloc/internal/arena.h

@@ -290,14 +290,14 @@ struct arena_s {
 
 	/*
 	 * Number of threads currently assigned to this arena.  This field is
-	 * protected by arenas_lock.
+	 * synchronized via atomic operations.
 	 */
 	unsigned		nthreads;
 
 	/*
 	 * There are three classes of arena operations from a locking
 	 * perspective:
-	 * 1) Thread assignment (modifies nthreads) is protected by arenas_lock.
+	 * 1) Thread assignment (modifies nthreads) is synchronized via atomics.
 	 * 2) Bin-related operations are protected by bin locks.
 	 * 3) Chunk- and run-related operations are protected by this mutex.
 	 */
@@ -465,7 +465,6 @@ struct arena_s {
 
 /* Used in conjunction with tsd for fast arena-related context lookup. */
 struct arena_tdata_s {
-	arena_t			*arena;
 	ticker_t		decay_ticker;
 };
 #endif /* JEMALLOC_ARENA_STRUCTS_B */
@@ -578,6 +577,9 @@ void	arena_stats_merge(arena_t *arena, const char **dss,
     ssize_t *lg_dirty_mult, ssize_t *decay_time, size_t *nactive,
     size_t *ndirty, arena_stats_t *astats, malloc_bin_stats_t *bstats,
     malloc_large_stats_t *lstats, malloc_huge_stats_t *hstats);
+unsigned	arena_nthreads_get(arena_t *arena);
+void	arena_nthreads_inc(arena_t *arena);
+void	arena_nthreads_dec(arena_t *arena);
 arena_t	*arena_new(unsigned ind);
 bool	arena_boot(void);
 void	arena_prefork(arena_t *arena);

include/jemalloc/internal/atomic.h

@@ -28,8 +28,8 @@
  * callers.
  *
  *   <t> atomic_read_<t>(<t> *p) { return (*p); }
- *   <t> atomic_add_<t>(<t> *p, <t> x) { return (*p + x); }
+ *   <t> atomic_add_<t>(<t> *p, <t> x) { return (*p += x); }
- *   <t> atomic_sub_<t>(<t> *p, <t> x) { return (*p - x); }
+ *   <t> atomic_sub_<t>(<t> *p, <t> x) { return (*p -= x); }
  *   bool atomic_cas_<t>(<t> *p, <t> c, <t> s)
  *   {
  *     if (*p != c)

include/jemalloc/internal/jemalloc_internal.h.in

@@ -438,7 +438,13 @@ extern unsigned	opt_narenas;
 extern bool	in_valgrind;
 
 /* Number of CPUs. */
-extern unsigned		ncpus;
+extern unsigned	ncpus;
+
+/*
+ * Arenas that are used to service external requests.  Not all elements of the
+ * arenas array are necessarily used; arenas are created lazily as needed.
+ */
+extern arena_t	**arenas;
 
 /*
  * index2size_tab encodes the same information as could be computed (at
@@ -452,21 +458,17 @@ extern size_t const	index2size_tab[NSIZES+1];
  */
 extern uint8_t const	size2index_tab[];
 
-arena_t	*a0get(void);
 void	*a0malloc(size_t size);
 void	a0dalloc(void *ptr);
 void	*bootstrap_malloc(size_t size);
 void	*bootstrap_calloc(size_t num, size_t size);
 void	bootstrap_free(void *ptr);
 arena_t	*arenas_extend(unsigned ind);
-arena_t	*arena_init(unsigned ind);
 unsigned	narenas_total_get(void);
+arena_t	*arena_init(unsigned ind);
 arena_tdata_t	*arena_tdata_get_hard(tsd_t *tsd, unsigned ind);
-arena_t	*arena_get_hard(tsd_t *tsd, unsigned ind, bool init_if_missing,
-    arena_tdata_t *tdata);
 arena_t	*arena_choose_hard(tsd_t *tsd);
 void	arena_migrate(tsd_t *tsd, unsigned oldind, unsigned newind);
-unsigned	arena_nbound(unsigned ind);
 void	thread_allocated_cleanup(tsd_t *tsd);
 void	thread_deallocated_cleanup(tsd_t *tsd);
 void	arena_cleanup(tsd_t *tsd);
@@ -543,8 +545,7 @@ size_t	sa2u(size_t size, size_t alignment);
 arena_t	*arena_choose(tsd_t *tsd, arena_t *arena);
 arena_tdata_t	*arena_tdata_get(tsd_t *tsd, unsigned ind,
     bool refresh_if_missing);
-arena_t	*arena_get(tsd_t *tsd, unsigned ind, bool init_if_missing,
+arena_t	*arena_get(unsigned ind, bool init_if_missing);
-    bool refresh_if_missing);
 ticker_t	*decay_ticker_get(tsd_t *tsd, unsigned ind);
 #endif
 
@@ -819,19 +820,19 @@ arena_tdata_get(tsd_t *tsd, unsigned ind, bool refresh_if_missing)
 }
 
 JEMALLOC_INLINE arena_t *
-arena_get(tsd_t *tsd, unsigned ind, bool init_if_missing,
+arena_get(unsigned ind, bool init_if_missing)
-    bool refresh_if_missing)
 {
-	arena_tdata_t *tdata;
+	arena_t *ret;
 
-	/* init_if_missing requires refresh_if_missing. */
+	assert(ind <= MALLOCX_ARENA_MAX);
-	assert(!init_if_missing || refresh_if_missing);
 
-	tdata = arena_tdata_get(tsd, ind, refresh_if_missing);
+	ret = arenas[ind];
-	if (unlikely(tdata == NULL || tdata->arena == NULL))
+	if (unlikely(ret == NULL)) {
-		return (arena_get_hard(tsd, ind, init_if_missing, tdata));
+		ret = atomic_read_p((void *)&arenas[ind]);
-
+		if (init_if_missing && unlikely(ret == NULL))
-	return (tdata->arena);
+			ret = arena_init(ind);
+	}
+	return (ret);
 }
 
 JEMALLOC_INLINE ticker_t *

include/jemalloc/internal/private_symbols.txt

@@ -1,5 +1,4 @@
 a0dalloc
-a0get
 a0malloc
 arena_aalloc
 arena_alloc_junk_small
@@ -34,7 +33,6 @@ arena_decay_ticks
 arena_dss_prec_get
 arena_dss_prec_set
 arena_get
-arena_get_hard
 arena_init
 arena_lg_dirty_mult_default_get
 arena_lg_dirty_mult_default_set
@@ -73,10 +71,12 @@ arena_migrate
 arena_miscelm_get
 arena_miscelm_to_pageind
 arena_miscelm_to_rpages
-arena_nbound
 arena_new
 arena_node_alloc
 arena_node_dalloc
+arena_nthreads_dec
+arena_nthreads_get
+arena_nthreads_inc
 arena_palloc
 arena_postfork_child
 arena_postfork_parent
@@ -106,6 +106,7 @@ arena_stats_merge
 arena_tcache_fill_small
 arena_tdata_get
 arena_tdata_get_hard
+arenas
 atomic_add_p
 atomic_add_u
 atomic_add_uint32

src/arena.c

@@ -3261,6 +3261,27 @@ arena_stats_merge(arena_t *arena, const char **dss, ssize_t *lg_dirty_mult,
 	}
 }
 
+unsigned
+arena_nthreads_get(arena_t *arena)
+{
+
+	return (atomic_read_u(&arena->nthreads));
+}
+
+void
+arena_nthreads_inc(arena_t *arena)
+{
+
+	atomic_add_u(&arena->nthreads, 1);
+}
+
+void
+arena_nthreads_dec(arena_t *arena)
+{
+
+	atomic_sub_u(&arena->nthreads, 1);
+}
+
 arena_t *
 arena_new(unsigned ind)
 {

src/chunk.c

@@ -415,9 +415,7 @@ chunk_arena_get(unsigned arena_ind)
 {
 	arena_t *arena;
 
-	/* Dodge tsd for a0 in order to avoid bootstrapping issues. */
+	arena = arena_get(arena_ind, false);
-	arena = (arena_ind == 0) ? a0get() : arena_get(tsd_fetch(), arena_ind,
-	     false, true);
 	/*
 	 * The arena we're allocating on behalf of must have been initialized
 	 * already.

src/ctl.c

@@ -694,9 +694,7 @@ ctl_grow(void)
 static void
 ctl_refresh(void)
 {
-	tsd_t *tsd;
 	unsigned i;
-	bool refreshed;
 	VARIABLE_ARRAY(arena_t *, tarenas, ctl_stats.narenas);
 
 	/*
@@ -706,19 +704,14 @@ ctl_refresh(void)
 	ctl_stats.arenas[ctl_stats.narenas].nthreads = 0;
 	ctl_arena_clear(&ctl_stats.arenas[ctl_stats.narenas]);
 
-	tsd = tsd_fetch();
+	for (i = 0; i < ctl_stats.narenas; i++)
-	for (i = 0, refreshed = false; i < ctl_stats.narenas; i++) {
+		tarenas[i] = arena_get(i, false);
-		tarenas[i] = arena_get(tsd, i, false, false);
-		if (tarenas[i] == NULL && !refreshed) {
-			tarenas[i] = arena_get(tsd, i, false, true);
-			refreshed = true;
-		}
-	}
 
 	for (i = 0; i < ctl_stats.narenas; i++) {
-		if (tarenas[i] != NULL)
+		if (tarenas[i] != NULL) {
-			ctl_stats.arenas[i].nthreads = arena_nbound(i);
+			ctl_stats.arenas[i].nthreads =
-		else
+			    arena_nthreads_get(arena_get(i, false));
+		} else
 			ctl_stats.arenas[i].nthreads = 0;
 	}
 
@@ -1332,7 +1325,7 @@ thread_arena_ctl(const size_t *mib, size_t miblen, void *oldp, size_t *oldlenp,
 		}
 
 		/* Initialize arena if necessary. */
-		newarena = arena_get(tsd, newind, true, true);
+		newarena = arena_get(newind, true);
 		if (newarena == NULL) {
 			ret = EAGAIN;
 			goto label_return;
@@ -1560,22 +1553,14 @@ arena_i_purge(unsigned arena_ind, bool all)
 
 	malloc_mutex_lock(&ctl_mtx);
 	{
-		tsd_t *tsd = tsd_fetch();
 		unsigned narenas = ctl_stats.narenas;
 
 		if (arena_ind == narenas) {
 			unsigned i;
-			bool refreshed;
 			VARIABLE_ARRAY(arena_t *, tarenas, narenas);
 
-			for (i = 0, refreshed = false; i < narenas; i++) {
+			for (i = 0; i < narenas; i++)
-				tarenas[i] = arena_get(tsd, i, false, false);
+				tarenas[i] = arena_get(i, false);
-				if (tarenas[i] == NULL && !refreshed) {
-					tarenas[i] = arena_get(tsd, i, false,
-					    true);
-					refreshed = true;
-				}
-			}
 
 			/*
 			 * No further need to hold ctl_mtx, since narenas and
@@ -1592,7 +1577,7 @@ arena_i_purge(unsigned arena_ind, bool all)
 
 			assert(arena_ind < narenas);
 
-			tarena = arena_get(tsd, arena_ind, false, true);
+			tarena = arena_get(arena_ind, false);
 
 			/* No further need to hold ctl_mtx. */
 			malloc_mutex_unlock(&ctl_mtx);
@@ -1664,7 +1649,7 @@ arena_i_dss_ctl(const size_t *mib, size_t miblen, void *oldp, size_t *oldlenp,
 	}
 
 	if (arena_ind < ctl_stats.narenas) {
-		arena_t *arena = arena_get(tsd_fetch(), arena_ind, false, true);
+		arena_t *arena = arena_get(arena_ind, false);
 		if (arena == NULL || (dss_prec != dss_prec_limit &&
 		    arena_dss_prec_set(arena, dss_prec))) {
 			ret = EFAULT;
@@ -1697,7 +1682,7 @@ arena_i_lg_dirty_mult_ctl(const size_t *mib, size_t miblen, void *oldp,
 	unsigned arena_ind = (unsigned)mib[1];
 	arena_t *arena;
 
-	arena = arena_get(tsd_fetch(), arena_ind, false, true);
+	arena = arena_get(arena_ind, false);
 	if (arena == NULL) {
 		ret = EFAULT;
 		goto label_return;
@@ -1731,7 +1716,7 @@ arena_i_decay_time_ctl(const size_t *mib, size_t miblen, void *oldp,
 	unsigned arena_ind = (unsigned)mib[1];
 	arena_t *arena;
 
-	arena = arena_get(tsd_fetch(), arena_ind, false, true);
+	arena = arena_get(arena_ind, false);
 	if (arena == NULL) {
 		ret = EFAULT;
 		goto label_return;
@@ -1767,7 +1752,7 @@ arena_i_chunk_hooks_ctl(const size_t *mib, size_t miblen, void *oldp,
 
 	malloc_mutex_lock(&ctl_mtx);
 	if (arena_ind < narenas_total_get() && (arena =
-	    arena_get(tsd_fetch(), arena_ind, false, true)) != NULL) {
+	    arena_get(arena_ind, false)) != NULL) {
 		if (newp != NULL) {
 			chunk_hooks_t old_chunk_hooks, new_chunk_hooks;
 			WRITE(new_chunk_hooks, chunk_hooks_t);

src/jemalloc.c

@@ -47,7 +47,7 @@ bool	in_valgrind;
 
 unsigned	ncpus;
 
-/* Protects arenas initialization (arenas, narenas_total). */
+/* Protects arenas initialization. */
 static malloc_mutex_t	arenas_lock;
 /*
  * Arenas that are used to service external requests.  Not all elements of the
@@ -57,8 +57,8 @@ static malloc_mutex_t	arenas_lock;
  * arenas.  arenas[narenas_auto..narenas_total) are only used if the application
  * takes some action to create them and allocate from them.
  */
-static arena_t		**arenas;
+arena_t			**arenas;
-static unsigned		narenas_total;
+static unsigned		narenas_total; /* Use narenas_total_*(). */
 static arena_t		*a0; /* arenas[0]; read-only after initialization. */
 static unsigned		narenas_auto; /* Read-only after initialization. */
 
@@ -311,14 +311,6 @@ malloc_init(void)
  * cannot tolerate TLS variable access.
  */
 
-arena_t *
-a0get(void)
-{
-
-	assert(a0 != NULL);
-	return (a0);
-}
-
 static void *
 a0ialloc(size_t size, bool zero, bool is_metadata)
 {
@@ -327,7 +319,7 @@ a0ialloc(size_t size, bool zero, bool is_metadata)
 		return (NULL);
 
 	return (iallocztm(NULL, size, size2index(size), zero, false,
-	    is_metadata, a0get(), true));
+	    is_metadata, arena_get(0, false), true));
 }
 
 static void
@@ -391,47 +383,59 @@ bootstrap_free(void *ptr)
 	a0idalloc(ptr, false);
 }
 
+static void
+arena_set(unsigned ind, arena_t *arena)
+{
+
+	atomic_write_p((void **)&arenas[ind], arena);
+}
+
+static void
+narenas_total_set(unsigned narenas)
+{
+
+	atomic_write_u(&narenas_total, narenas);
+}
+
+static void
+narenas_total_inc(void)
+{
+
+	atomic_add_u(&narenas_total, 1);
+}
+
+unsigned
+narenas_total_get(void)
+{
+
+	return (atomic_read_u(&narenas_total));
+}
+
 /* Create a new arena and insert it into the arenas array at index ind. */
 static arena_t *
 arena_init_locked(unsigned ind)
 {
 	arena_t *arena;
 
-	/* Expand arenas if necessary. */
+	assert(ind <= narenas_total_get());
-	assert(ind <= narenas_total);
 	if (ind > MALLOCX_ARENA_MAX)
 		return (NULL);
-	if (ind == narenas_total) {
+	if (ind == narenas_total_get())
-		unsigned narenas_new = narenas_total + 1;
+		narenas_total_inc();
-		arena_t **arenas_new =
-		    (arena_t **)a0malloc(CACHELINE_CEILING(narenas_new *
-		    sizeof(arena_t *)));
-		if (arenas_new == NULL)
-			return (NULL);
-		memcpy(arenas_new, arenas, narenas_total * sizeof(arena_t *));
-		arenas_new[ind] = NULL;
-		/*
-		 * Deallocate only if arenas came from a0malloc() (not
-		 * base_alloc()).
-		 */
-		if (narenas_total != narenas_auto)
-			a0dalloc(arenas);
-		arenas = arenas_new;
-		narenas_total = narenas_new;
-	}
 
 	/*
 	 * Another thread may have already initialized arenas[ind] if it's an
 	 * auto arena.
 	 */
-	arena = arenas[ind];
+	arena = arena_get(ind, false);
 	if (arena != NULL) {
 		assert(ind < narenas_auto);
 		return (arena);
 	}
 
 	/* Actually initialize the arena. */
-	arena = arenas[ind] = arena_new(ind);
+	arena = arena_new(ind);
+	arena_set(ind, arena);
 	return (arena);
 }
 
@@ -446,37 +450,16 @@ arena_init(unsigned ind)
 	return (arena);
 }
 
-unsigned
-narenas_total_get(void)
-{
-	unsigned narenas;
-
-	malloc_mutex_lock(&arenas_lock);
-	narenas = narenas_total;
-	malloc_mutex_unlock(&arenas_lock);
-
-	return (narenas);
-}
-
-static void
-arena_bind_locked(tsd_t *tsd, unsigned ind)
-{
-	arena_t *arena;
-
-	arena = arenas[ind];
-	arena->nthreads++;
-
-	if (tsd_nominal(tsd))
-		tsd_arena_set(tsd, arena);
-}
-
 static void
 arena_bind(tsd_t *tsd, unsigned ind)
 {
+	arena_t *arena;
 
-	malloc_mutex_lock(&arenas_lock);
+	arena = arena_get(ind, false);
-	arena_bind_locked(tsd, ind);
+	arena_nthreads_inc(arena);
-	malloc_mutex_unlock(&arenas_lock);
+
+	if (tsd_nominal(tsd))
+		tsd_arena_set(tsd, arena);
 }
 
 void
@@ -484,35 +467,20 @@ arena_migrate(tsd_t *tsd, unsigned oldind, unsigned newind)
 {
 	arena_t *oldarena, *newarena;
 
-	malloc_mutex_lock(&arenas_lock);
+	oldarena = arena_get(oldind, false);
-	oldarena = arenas[oldind];
+	newarena = arena_get(newind, false);
-	newarena = arenas[newind];
+	arena_nthreads_dec(oldarena);
-	oldarena->nthreads--;
+	arena_nthreads_inc(newarena);
-	newarena->nthreads++;
-	malloc_mutex_unlock(&arenas_lock);
 	tsd_arena_set(tsd, newarena);
 }
 
-unsigned
-arena_nbound(unsigned ind)
-{
-	unsigned nthreads;
-
-	malloc_mutex_lock(&arenas_lock);
-	nthreads = arenas[ind]->nthreads;
-	malloc_mutex_unlock(&arenas_lock);
-	return (nthreads);
-}
-
 static void
 arena_unbind(tsd_t *tsd, unsigned ind)
 {
 	arena_t *arena;
 
-	malloc_mutex_lock(&arenas_lock);
+	arena = arena_get(ind, false);
-	arena = arenas[ind];
+	arena_nthreads_dec(arena);
-	arena->nthreads--;
-	malloc_mutex_unlock(&arenas_lock);
 	tsd_arena_set(tsd, NULL);
 }
 
@@ -568,14 +536,6 @@ arena_tdata_get_hard(tsd_t *tsd, unsigned ind)
 	 * the arenas.extend mallctl, which we trust mallctl synchronization to
 	 * prevent.
 	 */
-	malloc_mutex_lock(&arenas_lock);
-	for (i = 0; i < narenas_actual; i++)
-		arenas_tdata[i].arena = arenas[i];
-	malloc_mutex_unlock(&arenas_lock);
-	if (narenas_tdata > narenas_actual) {
-		memset(&arenas_tdata[narenas_actual], 0, sizeof(arena_tdata_t)
-		    * (narenas_tdata - narenas_actual));
-	}
 
 	/* Copy/initialize tickers. */
 	for (i = 0; i < narenas_actual; i++) {
@@ -587,6 +547,10 @@ arena_tdata_get_hard(tsd_t *tsd, unsigned ind)
 			    DECAY_NTICKS_PER_UPDATE);
 		}
 	}
+	if (narenas_tdata > narenas_actual) {
+		memset(&arenas_tdata[narenas_actual], 0, sizeof(arena_tdata_t)
+		    * (narenas_tdata - narenas_actual));
+	}
 
 	/* Read the refreshed tdata array. */
 	tdata = &arenas_tdata[ind];
@@ -596,33 +560,6 @@ label_return:
 	return (tdata);
 }
 
-arena_t *
-arena_get_hard(tsd_t *tsd, unsigned ind, bool init_if_missing,
-    arena_tdata_t *tdata)
-{
-	arena_t *arena;
-	unsigned narenas_actual;
-
-	if (init_if_missing && tdata != NULL) {
-		tdata->arena = arena_init(ind);
-		if (tdata->arena != NULL)
-			return (tdata->arena);
-	}
-
-	/*
-	 * This function must always tell the truth, even if it's slow, so don't
-	 * let OOM, thread cleanup (note tsd_nominal check), nor recursive
-	 * allocation avoidance (note arenas_tdata_bypass check) get in the way.
-	 */
-	narenas_actual = narenas_total_get();
-	if (ind >= narenas_actual)
-		return (NULL);
-	malloc_mutex_lock(&arenas_lock);
-	arena = arenas[ind];
-	malloc_mutex_unlock(&arenas_lock);
-	return (arena);
-}
-
 /* Slow path, called only by arena_choose(). */
 arena_t *
 arena_choose_hard(tsd_t *tsd)
@@ -635,15 +572,16 @@ arena_choose_hard(tsd_t *tsd)
 		choose = 0;
 		first_null = narenas_auto;
 		malloc_mutex_lock(&arenas_lock);
-		assert(a0get() != NULL);
+		assert(arena_get(0, false) != NULL);
 		for (i = 1; i < narenas_auto; i++) {
-			if (arenas[i] != NULL) {
+			if (arena_get(i, false) != NULL) {
 				/*
 				 * Choose the first arena that has the lowest
 				 * number of threads assigned to it.
 				 */
-				if (arenas[i]->nthreads <
+				if (arena_nthreads_get(arena_get(i, false)) <
-				    arenas[choose]->nthreads)
+				    arena_nthreads_get(arena_get(choose,
+				    false)))
 					choose = i;
 			} else if (first_null == narenas_auto) {
 				/*
@@ -659,13 +597,13 @@ arena_choose_hard(tsd_t *tsd)
 			}
 		}
 
-		if (arenas[choose]->nthreads == 0
+		if (arena_nthreads_get(arena_get(choose, false)) == 0
 		    || first_null == narenas_auto) {
 			/*
 			 * Use an unloaded arena, or the least loaded arena if
 			 * all arenas are already initialized.
 			 */
-			ret = arenas[choose];
+			ret = arena_get(choose, false);
 		} else {
 			/* Initialize a new arena. */
 			choose = first_null;
@@ -675,10 +613,10 @@ arena_choose_hard(tsd_t *tsd)
 				return (NULL);
 			}
 		}
-		arena_bind_locked(tsd, choose);
+		arena_bind(tsd, choose);
 		malloc_mutex_unlock(&arenas_lock);
 	} else {
-		ret = a0get();
+		ret = arena_get(0, false);
 		arena_bind(tsd, 0);
 	}
 
@@ -750,7 +688,7 @@ stats_print_atexit(void)
 		 * continue to allocate.
 		 */
 		for (i = 0, narenas = narenas_total_get(); i < narenas; i++) {
-			arena_t *arena = arenas[i];
+			arena_t *arena = arena_get(i, false);
 			if (arena != NULL) {
 				tcache_t *tcache;
 
@@ -1309,7 +1247,8 @@ malloc_init_hard_a0_locked(void)
 	 * Create enough scaffolding to allow recursive allocation in
 	 * malloc_ncpus().
 	 */
-	narenas_total = narenas_auto = 1;
+	narenas_auto = 1;
+	narenas_total_set(narenas_auto);
 	arenas = &a0;
 	memset(arenas, 0, sizeof(arena_t *) * narenas_auto);
 	/*
@@ -1391,28 +1330,22 @@ malloc_init_hard_finish(void)
 	}
 	narenas_auto = opt_narenas;
 	/*
-	 * Make sure that the arenas array can be allocated.  In practice, this
+	 * Limit the number of arenas to the indexing range of MALLOCX_ARENA().
-	 * limit is enough to allow the allocator to function, but the ctl
-	 * machinery will fail to allocate memory at far lower limits.
 	 */
-	if (narenas_auto > chunksize / sizeof(arena_t *)) {
+	if (narenas_auto > MALLOCX_ARENA_MAX) {
-		narenas_auto = (unsigned)(chunksize / sizeof(arena_t *));
+		narenas_auto = MALLOCX_ARENA_MAX;
 		malloc_printf("<jemalloc>: Reducing narenas to limit (%d)\n",
 		    narenas_auto);
 	}
-	narenas_total = narenas_auto;
+	narenas_total_set(narenas_auto);
 
 	/* Allocate and initialize arenas. */
-	arenas = (arena_t **)base_alloc(sizeof(arena_t *) * narenas_total);
+	arenas = (arena_t **)base_alloc(sizeof(arena_t *) *
+	    (MALLOCX_ARENA_MAX+1));
 	if (arenas == NULL)
 		return (true);
-	/*
-	 * Zero the array.  In practice, this should always be pre-zeroed,
-	 * since it was just mmap()ed, but let's be sure.
-	 */
-	memset(arenas, 0, sizeof(arena_t *) * narenas_total);
 	/* Copy the pointer to the one arena that was already initialized. */
-	arenas[0] = a0;
+	arena_set(0, a0);
 
 	malloc_init_state = malloc_init_initialized;
 	malloc_slow_flag_init();
@@ -2084,7 +2017,7 @@ imallocx_flags_decode_hard(tsd_t *tsd, size_t size, int flags, size_t *usize,
 		*tcache = tcache_get(tsd, true);
 	if ((flags & MALLOCX_ARENA_MASK) != 0) {
 		unsigned arena_ind = MALLOCX_ARENA_GET(flags);
-		*arena = arena_get(tsd, arena_ind, true, true);
+		*arena = arena_get(arena_ind, true);
 		if (unlikely(*arena == NULL))
 			return (true);
 	} else
@@ -2325,7 +2258,7 @@ je_rallocx(void *ptr, size_t size, int flags)
 
 	if (unlikely((flags & MALLOCX_ARENA_MASK) != 0)) {
 		unsigned arena_ind = MALLOCX_ARENA_GET(flags);
-		arena = arena_get(tsd, arena_ind, true, true);
+		arena = arena_get(arena_ind, true);
 		if (unlikely(arena == NULL))
 			goto label_oom;
 	} else
@@ -2677,7 +2610,7 @@ JEMALLOC_EXPORT void
 _malloc_prefork(void)
 #endif
 {
-	unsigned i;
+	unsigned i, narenas;
 
 #ifdef JEMALLOC_MUTEX_INIT_CB
 	if (!malloc_initialized())
@@ -2689,9 +2622,11 @@ _malloc_prefork(void)
 	ctl_prefork();
 	prof_prefork();
 	malloc_mutex_prefork(&arenas_lock);
-	for (i = 0; i < narenas_total; i++) {
+	for (i = 0, narenas = narenas_total_get(); i < narenas; i++) {
-		if (arenas[i] != NULL)
+		arena_t *arena;
-			arena_prefork(arenas[i]);
+
+		if ((arena = arena_get(i, false)) != NULL)
+			arena_prefork(arena);
 	}
 	chunk_prefork();
 	base_prefork();
@@ -2705,7 +2640,7 @@ JEMALLOC_EXPORT void
 _malloc_postfork(void)
 #endif
 {
-	unsigned i;
+	unsigned i, narenas;
 
 #ifdef JEMALLOC_MUTEX_INIT_CB
 	if (!malloc_initialized())
@@ -2716,9 +2651,11 @@ _malloc_postfork(void)
 	/* Release all mutexes, now that fork() has completed. */
 	base_postfork_parent();
 	chunk_postfork_parent();
-	for (i = 0; i < narenas_total; i++) {
+	for (i = 0, narenas = narenas_total_get(); i < narenas; i++) {
-		if (arenas[i] != NULL)
+		arena_t *arena;
-			arena_postfork_parent(arenas[i]);
+
+		if ((arena = arena_get(i, false)) != NULL)
+			arena_postfork_parent(arena);
 	}
 	malloc_mutex_postfork_parent(&arenas_lock);
 	prof_postfork_parent();
@@ -2728,16 +2665,18 @@ _malloc_postfork(void)
 void
 jemalloc_postfork_child(void)
 {
-	unsigned i;
+	unsigned i, narenas;
 
 	assert(malloc_initialized());
 
 	/* Release all mutexes, now that fork() has completed. */
 	base_postfork_child();
 	chunk_postfork_child();
-	for (i = 0; i < narenas_total; i++) {
+	for (i = 0, narenas = narenas_total_get(); i < narenas; i++) {
-		if (arenas[i] != NULL)
+		arena_t *arena;
-			arena_postfork_child(arenas[i]);
+
+		if ((arena = arena_get(i, false)) != NULL)
+			arena_postfork_child(arena);
 	}
 	malloc_mutex_postfork_child(&arenas_lock);
 	prof_postfork_child();

src/tcache.c

@@ -325,7 +325,8 @@ tcache_create(tsd_t *tsd, arena_t *arena)
 	/* Avoid false cacheline sharing. */
 	size = sa2u(size, CACHELINE);
 
-	tcache = ipallocztm(tsd, size, CACHELINE, true, false, true, a0get());
+	tcache = ipallocztm(tsd, size, CACHELINE, true, false, true,
+	    arena_get(0, false));
 	if (tcache == NULL)
 		return (NULL);
 
@@ -453,7 +454,7 @@ tcaches_create(tsd_t *tsd, unsigned *r_ind)
 
 	if (tcaches_avail == NULL && tcaches_past > MALLOCX_TCACHE_MAX)
 		return (true);
-	tcache = tcache_create(tsd, a0get());
+	tcache = tcache_create(tsd, arena_get(0, false));
 	if (tcache == NULL)
 		return (true);