diff --git a/include/jemalloc/internal/arena.h b/include/jemalloc/internal/arena.h
index 681b5802..894ce9af 100644
--- a/include/jemalloc/internal/arena.h
+++ b/include/jemalloc/internal/arena.h
@@ -389,7 +389,7 @@ bool	arena_dss_prec_set(arena_t *arena, dss_prec_t dss_prec);
 void	arena_stats_merge(arena_t *arena, const char **dss, size_t *nactive,
     size_t *ndirty, arena_stats_t *astats, malloc_bin_stats_t *bstats,
     malloc_large_stats_t *lstats);
-bool	arena_new(arena_t *arena, unsigned ind);
+arena_t	*arena_new(unsigned ind);
 void	arena_boot(void);
 void	arena_prefork(arena_t *arena);
 void	arena_postfork_parent(arena_t *arena);
@@ -924,8 +924,10 @@ arena_malloc(tsd_t *tsd, arena_t *arena, size_t size, bool zero,
 		    true)) != NULL))
 			return (tcache_alloc_small(tcache, size, zero));
 		else {
-			return (arena_malloc_small(choose_arena(tsd, arena),
-			    size, zero));
+			arena = arena_choose(tsd, arena);
+			if (unlikely(arena == NULL))
+				return (NULL);
+			return (arena_malloc_small(arena, size, zero));
 		}
 	} else {
 		/*
@@ -936,8 +938,10 @@ arena_malloc(tsd_t *tsd, arena_t *arena, size_t size, bool zero,
 		    tcache_get(tsd, true)) != NULL))
 			return (tcache_alloc_large(tcache, size, zero));
 		else {
-			return (arena_malloc_large(choose_arena(tsd, arena),
-			    size, zero));
+			arena = arena_choose(tsd, arena);
+			if (unlikely(arena == NULL))
+				return (NULL);
+			return (arena_malloc_large(arena, size, zero));
 		}
 	}
 }
diff --git a/include/jemalloc/internal/jemalloc_internal.h.in b/include/jemalloc/internal/jemalloc_internal.h.in
index 8f0beb9e..c7a5fd8a 100644
--- a/include/jemalloc/internal/jemalloc_internal.h.in
+++ b/include/jemalloc/internal/jemalloc_internal.h.in
@@ -386,20 +386,6 @@ extern bool	in_valgrind;
 /* Number of CPUs. */
 extern unsigned		ncpus;
 
-/* Protects arenas initialization (arenas, arenas_total). */
-extern malloc_mutex_t	arenas_lock;
-/*
- * Arenas that are used to service external requests.  Not all elements of the
- * arenas array are necessarily used; arenas are created lazily as needed.
- *
- * arenas[0..narenas_auto) are used for automatic multiplexing of threads and
- * arenas.  arenas[narenas_auto..narenas_total) are only used if the application
- * takes some action to create them and allocate from them.
- */
-extern arena_t		**arenas;
-extern unsigned		narenas_total;
-extern unsigned		narenas_auto; /* Read-only after initialization. */
-
 /*
  * index2size_tab encodes the same information as could be computed (at
  * unacceptable cost in some code paths) by index2size_compute().
@@ -412,11 +398,23 @@ extern size_t const	index2size_tab[NSIZES];
  */
 extern uint8_t const	size2index_tab[];
 
+arena_t	*a0get(void);
+void	*a0malloc(size_t size);
+void	*a0calloc(size_t num, size_t size);
+void	a0free(void *ptr);
 arena_t	*arenas_extend(unsigned ind);
-arena_t	*choose_arena_hard(tsd_t *tsd);
+arena_t	*arena_init(unsigned ind);
+unsigned	narenas_total_get(void);
+arena_t	*arena_get_hard(tsd_t *tsd, unsigned ind, bool init_if_missing);
+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);
+void	arenas_cache_cleanup(tsd_t *tsd);
+void	narenas_cache_cleanup(tsd_t *tsd);
+void	arenas_cache_bypass_cleanup(tsd_t *tsd);
 void	jemalloc_prefork(void);
 void	jemalloc_postfork_parent(void);
 void	jemalloc_postfork_child(void);
@@ -475,8 +473,9 @@ size_t	s2u_compute(size_t size);
 size_t	s2u_lookup(size_t size);
 size_t	s2u(size_t size);
 size_t	sa2u(size_t size, size_t alignment);
-unsigned	narenas_total_get(void);
-arena_t	*choose_arena(tsd_t *tsd, arena_t *arena);
+arena_t	*arena_choose(tsd_t *tsd, arena_t *arena);
+arena_t	*arena_get(tsd_t *tsd, unsigned ind, bool init_if_missing,
+    bool refresh_if_missing);
 #endif
 
 #if (defined(JEMALLOC_ENABLE_INLINE) || defined(JEMALLOC_C_))
@@ -709,34 +708,51 @@ sa2u(size_t size, size_t alignment)
 	return (usize);
 }
 
-JEMALLOC_INLINE unsigned
-narenas_total_get(void)
-{
-	unsigned narenas;
-
-	malloc_mutex_lock(&arenas_lock);
-	narenas = narenas_total;
-	malloc_mutex_unlock(&arenas_lock);
-
-	return (narenas);
-}
-
 /* Choose an arena based on a per-thread value. */
 JEMALLOC_INLINE arena_t *
-choose_arena(tsd_t *tsd, arena_t *arena)
+arena_choose(tsd_t *tsd, arena_t *arena)
 {
 	arena_t *ret;
 
 	if (arena != NULL)
 		return (arena);
 
-	if (unlikely((ret = tsd_arena_get(tsd)) == NULL)) {
-		ret = choose_arena_hard(tsd);
-		assert(ret != NULL);
-	}
+	if (unlikely((ret = tsd_arena_get(tsd)) == NULL))
+		ret = arena_choose_hard(tsd);
 
 	return (ret);
 }
+
+JEMALLOC_INLINE arena_t *
+arena_get(tsd_t *tsd, unsigned ind, bool init_if_missing,
+    bool refresh_if_missing)
+{
+	arena_t *arena;
+	arena_t **arenas_cache = tsd_arenas_cache_get(tsd);
+
+	/* init_if_missing requires refresh_if_missing. */
+	assert(!init_if_missing || refresh_if_missing);
+
+	if (unlikely(arenas_cache == NULL)) {
+		/* arenas_cache hasn't been initialized yet. */
+		return (arena_get_hard(tsd, ind, init_if_missing));
+	}
+	if (unlikely(ind >= tsd_narenas_cache_get(tsd))) {
+		/*
+		 * ind is invalid, cache is old (too small), or arena to be
+		 * initialized.
+		 */
+		return (refresh_if_missing ?  arena_get_hard(tsd, ind,
+		    init_if_missing) : NULL);
+	}
+	arena = arenas_cache[ind];
+	if (likely(arena != NULL) || !refresh_if_missing)
+		return (arena);
+	if (init_if_missing)
+		return (arena_get_hard(tsd, ind, init_if_missing));
+	else
+		return (NULL);
+}
 #endif
 
 #include "jemalloc/internal/bitmap.h"
@@ -833,8 +849,10 @@ ipalloct(tsd_t *tsd, size_t usize, size_t alignment, bool zero, bool try_tcache,
 		ret = arena_malloc(tsd, arena, usize, zero, try_tcache);
 	else {
 		if (usize <= arena_maxclass) {
-			ret = arena_palloc(choose_arena(tsd, arena), usize,
-			    alignment, zero);
+			arena = arena_choose(tsd, arena);
+			if (unlikely(arena == NULL))
+				return (NULL);
+			ret = arena_palloc(arena, usize, alignment, zero);
 		} else if (alignment <= chunksize)
 			ret = huge_malloc(tsd, arena, usize, zero);
 		else
diff --git a/include/jemalloc/internal/private_symbols.txt b/include/jemalloc/internal/private_symbols.txt
index 1a7fde4b..d5e6fdcf 100644
--- a/include/jemalloc/internal/private_symbols.txt
+++ b/include/jemalloc/internal/private_symbols.txt
@@ -1,11 +1,16 @@
 a0calloc
 a0free
+a0get
 a0malloc
+arena_get
+arena_get_hard
 arena_alloc_junk_small
 arena_bin_index
 arena_bin_info
 arena_bitselm_get
 arena_boot
+arena_choose
+arena_choose_hard
 arena_chunk_alloc_huge
 arena_chunk_dalloc_huge
 arena_cleanup
@@ -19,6 +24,7 @@ arena_dalloc_large_locked
 arena_dalloc_small
 arena_dss_prec_get
 arena_dss_prec_set
+arena_init
 arena_malloc
 arena_malloc_large
 arena_malloc_small
@@ -42,9 +48,11 @@ arena_mapbitsp_read
 arena_mapbitsp_write
 arena_maxclass
 arena_maxrun
+arena_migrate
 arena_miscelm_get
 arena_miscelm_to_pageind
 arena_miscelm_to_rpages
+arena_nbound
 arena_new
 arena_palloc
 arena_postfork_child
@@ -69,10 +77,8 @@ arena_salloc
 arena_sdalloc
 arena_stats_merge
 arena_tcache_fill_small
-arenas
-arenas_cleanup
-arenas_extend
-arenas_lock
+arenas_cache_bypass_cleanup
+arenas_cache_cleanup
 atomic_add_u
 atomic_add_uint32
 atomic_add_uint64
@@ -100,8 +106,6 @@ bitmap_size
 bitmap_unset
 bt_init
 buferror
-choose_arena
-choose_arena_hard
 chunk_alloc_arena
 chunk_alloc_base
 chunk_alloc_default
@@ -247,7 +251,8 @@ malloc_mutex_unlock
 malloc_printf
 malloc_snprintf
 malloc_strtoumax
-malloc_tsd_boot
+malloc_tsd_boot0
+malloc_tsd_boot1
 malloc_tsd_cleanup_register
 malloc_tsd_dalloc
 malloc_tsd_malloc
@@ -259,8 +264,7 @@ map_bias
 map_misc_offset
 mb_write
 mutex_boot
-narenas_auto
-narenas_total
+narenas_cache_cleanup
 narenas_total_get
 ncpus
 nhbins
@@ -363,6 +367,7 @@ tcache_alloc_small
 tcache_alloc_small_hard
 tcache_arena_associate
 tcache_arena_dissociate
+tcache_arena_reassociate
 tcache_bin_flush_large
 tcache_bin_flush_small
 tcache_bin_info
@@ -388,11 +393,14 @@ tsd_booted
 tsd_arena_get
 tsd_arena_set
 tsd_boot
+tsd_boot0
+tsd_boot1
 tsd_cleanup
 tsd_cleanup_wrapper
 tsd_fetch
 tsd_get
-tsd_get_wrapper
+tsd_wrapper_get
+tsd_wrapper_set
 tsd_initialized
 tsd_init_check_recursion
 tsd_init_finish
diff --git a/include/jemalloc/internal/tcache.h b/include/jemalloc/internal/tcache.h
index da8e4ef4..02eec5db 100644
--- a/include/jemalloc/internal/tcache.h
+++ b/include/jemalloc/internal/tcache.h
@@ -109,6 +109,7 @@ void	tcache_bin_flush_small(tcache_bin_t *tbin, index_t binind, unsigned rem,
 void	tcache_bin_flush_large(tcache_bin_t *tbin, index_t binind, unsigned rem,
     tcache_t *tcache);
 void	tcache_arena_associate(tcache_t *tcache, arena_t *arena);
+void	tcache_arena_reassociate(tcache_t *tcache, arena_t *arena);
 void	tcache_arena_dissociate(tcache_t *tcache);
 tcache_t *tcache_get_hard(tsd_t *tsd);
 tcache_t *tcache_create(arena_t *arena);
diff --git a/include/jemalloc/internal/tsd.h b/include/jemalloc/internal/tsd.h
index 25450391..b5658f8e 100644
--- a/include/jemalloc/internal/tsd.h
+++ b/include/jemalloc/internal/tsd.h
@@ -2,7 +2,7 @@
 #ifdef JEMALLOC_H_TYPES
 
 /* Maximum number of malloc_tsd users with cleanup functions. */
-#define	MALLOC_TSD_CLEANUPS_MAX	8
+#define	MALLOC_TSD_CLEANUPS_MAX	2
 
 typedef bool (*malloc_tsd_cleanup_t)(void);
 
@@ -23,7 +23,7 @@ typedef enum {
 
 /*
  * TLS/TSD-agnostic macro-based implementation of thread-specific data.  There
- * are four macros that support (at least) three use cases: file-private,
+ * are five macros that support (at least) three use cases: file-private,
  * library-private, and library-private inlined.  Following is an example
  * library-private tsd variable:
  *
@@ -33,18 +33,19 @@ typedef enum {
  *           int y;
  *   } example_t;
  *   #define EX_INITIALIZER JEMALLOC_CONCAT({0, 0})
- *   malloc_tsd_protos(, example_, example_t *)
- *   malloc_tsd_externs(example_, example_t *)
+ *   malloc_tsd_types(example_, example_t)
+ *   malloc_tsd_protos(, example_, example_t)
+ *   malloc_tsd_externs(example_, example_t)
  * In example.c:
- *   malloc_tsd_data(, example_, example_t *, EX_INITIALIZER)
- *   malloc_tsd_funcs(, example_, example_t *, EX_INITIALIZER,
+ *   malloc_tsd_data(, example_, example_t, EX_INITIALIZER)
+ *   malloc_tsd_funcs(, example_, example_t, EX_INITIALIZER,
  *       example_tsd_cleanup)
  *
  * The result is a set of generated functions, e.g.:
  *
  *   bool example_tsd_boot(void) {...}
- *   example_t **example_tsd_get() {...}
- *   void example_tsd_set(example_t **val) {...}
+ *   example_t *example_tsd_get() {...}
+ *   void example_tsd_set(example_t *val) {...}
  *
  * Note that all of the functions deal in terms of (a_type *) rather than
  * (a_type)  so that it is possible to support non-pointer types (unlike
@@ -70,9 +71,32 @@ typedef enum {
  * non-NULL.
  */
 
+/* malloc_tsd_types(). */
+#ifdef JEMALLOC_MALLOC_THREAD_CLEANUP
+#define	malloc_tsd_types(a_name, a_type)
+#elif (defined(JEMALLOC_TLS))
+#define	malloc_tsd_types(a_name, a_type)
+#elif (defined(_WIN32))
+#define	malloc_tsd_types(a_name, a_type)				\
+typedef struct {							\
+	bool	initialized;						\
+	a_type	val;							\
+} a_name##tsd_wrapper_t;
+#else
+#define	malloc_tsd_types(a_name, a_type)				\
+typedef struct {							\
+	bool	initialized;						\
+	a_type	val;							\
+} a_name##tsd_wrapper_t;
+#endif
+
 /* malloc_tsd_protos(). */
 #define	malloc_tsd_protos(a_attr, a_name, a_type)			\
 a_attr bool								\
+a_name##tsd_boot0(void);						\
+a_attr void								\
+a_name##tsd_boot1(void);						\
+a_attr bool								\
 a_name##tsd_boot(void);							\
 a_attr a_type *								\
 a_name##tsd_get(void);							\
@@ -93,11 +117,13 @@ extern bool		a_name##tsd_booted;
 #elif (defined(_WIN32))
 #define	malloc_tsd_externs(a_name, a_type)				\
 extern DWORD		a_name##tsd_tsd;				\
+extern a_name##tsd_wrapper_t	a_name##tsd_boot_wrapper;		\
 extern bool		a_name##tsd_booted;
 #else
 #define	malloc_tsd_externs(a_name, a_type)				\
 extern pthread_key_t	a_name##tsd_tsd;				\
 extern tsd_init_head_t	a_name##tsd_init_head;				\
+extern a_name##tsd_wrapper_t	a_name##tsd_boot_wrapper;		\
 extern bool		a_name##tsd_booted;
 #endif
 
@@ -118,6 +144,10 @@ a_attr bool		a_name##tsd_booted = false;
 #elif (defined(_WIN32))
 #define	malloc_tsd_data(a_attr, a_name, a_type, a_initializer)		\
 a_attr DWORD		a_name##tsd_tsd;				\
+a_attr a_name##tsd_wrapper_t a_name##tsd_boot_wrapper = {		\
+	false,								\
+	a_initializer							\
+};									\
 a_attr bool		a_name##tsd_booted = false;
 #else
 #define	malloc_tsd_data(a_attr, a_name, a_type, a_initializer)		\
@@ -126,6 +156,10 @@ a_attr tsd_init_head_t	a_name##tsd_init_head = {			\
 	ql_head_initializer(blocks),					\
 	MALLOC_MUTEX_INITIALIZER					\
 };									\
+a_attr a_name##tsd_wrapper_t a_name##tsd_boot_wrapper = {		\
+	false,								\
+	a_initializer							\
+};									\
 a_attr bool		a_name##tsd_booted = false;
 #endif
 
@@ -145,7 +179,7 @@ a_name##tsd_cleanup_wrapper(void)					\
 	return (a_name##tsd_initialized);				\
 }									\
 a_attr bool								\
-a_name##tsd_boot(void)							\
+a_name##tsd_boot0(void)							\
 {									\
 									\
 	if (a_cleanup != malloc_tsd_no_cleanup) {			\
@@ -155,6 +189,18 @@ a_name##tsd_boot(void)							\
 	a_name##tsd_booted = true;					\
 	return (false);							\
 }									\
+a_attr void								\
+a_name##tsd_boot1()							\
+{									\
+									\
+	/* Do nothing. */						\
+}									\
+a_attr bool								\
+a_name##tsd_boot(void)							\
+{									\
+									\
+	return (a_name##tsd_boot0());					\
+}									\
 /* Get/set. */								\
 a_attr a_type *								\
 a_name##tsd_get(void)							\
@@ -177,7 +223,7 @@ a_name##tsd_set(a_type *val)						\
     a_cleanup)								\
 /* Initialization/cleanup. */						\
 a_attr bool								\
-a_name##tsd_boot(void)							\
+a_name##tsd_boot0(void)							\
 {									\
 									\
 	if (a_cleanup != malloc_tsd_no_cleanup) {			\
@@ -188,6 +234,18 @@ a_name##tsd_boot(void)							\
 	a_name##tsd_booted = true;					\
 	return (false);							\
 }									\
+a_attr void								\
+a_name##tsd_boot1()							\
+{									\
+									\
+	/* Do nothing. */						\
+}									\
+a_attr bool								\
+a_name##tsd_boot(void)							\
+{									\
+									\
+	return (a_name##tsd_boot0());					\
+}									\
 /* Get/set. */								\
 a_attr a_type *								\
 a_name##tsd_get(void)							\
@@ -215,11 +273,6 @@ a_name##tsd_set(a_type *val)						\
 #elif (defined(_WIN32))
 #define	malloc_tsd_funcs(a_attr, a_name, a_type, a_initializer,		\
     a_cleanup)								\
-/* Data structure. */							\
-typedef struct {							\
-	bool	initialized;						\
-	a_type	val;							\
-} a_name##tsd_wrapper_t;						\
 /* Initialization/cleanup. */						\
 a_attr bool								\
 a_name##tsd_cleanup_wrapper(void)					\
@@ -241,23 +294,18 @@ a_name##tsd_cleanup_wrapper(void)					\
 	malloc_tsd_dalloc(wrapper);					\
 	return (false);							\
 }									\
-a_attr bool								\
-a_name##tsd_boot(void)							\
+a_attr void								\
+a_name##tsd_wrapper_set(a_name##tsd_wrapper_t *wrapper)			\
 {									\
 									\
-	a_name##tsd_tsd = TlsAlloc();					\
-	if (a_name##tsd_tsd == TLS_OUT_OF_INDEXES)			\
-		return (true);						\
-	if (a_cleanup != malloc_tsd_no_cleanup) {			\
-		malloc_tsd_cleanup_register(				\
-		    &a_name##tsd_cleanup_wrapper);			\
+	if (!TlsSetValue(a_name##tsd_tsd, (void *)wrapper)) {		\
+		malloc_write("<jemalloc>: Error setting"		\
+		    " TSD for "#a_name"\n");				\
+		abort();						\
 	}								\
-	a_name##tsd_booted = true;					\
-	return (false);							\
 }									\
-/* Get/set. */								\
 a_attr a_name##tsd_wrapper_t *						\
-a_name##tsd_get_wrapper(void)						\
+a_name##tsd_wrapper_get(void)						\
 {									\
 	a_name##tsd_wrapper_t *wrapper = (a_name##tsd_wrapper_t *)	\
 	    TlsGetValue(a_name##tsd_tsd);				\
@@ -273,21 +321,63 @@ a_name##tsd_get_wrapper(void)						\
 			wrapper->initialized = false;			\
 			wrapper->val = a_initializer;			\
 		}							\
-		if (!TlsSetValue(a_name##tsd_tsd, (void *)wrapper)) {	\
-			malloc_write("<jemalloc>: Error setting"	\
-			    " TSD for "#a_name"\n");			\
-			abort();					\
-		}							\
+		a_name##tsd_wrapper_set(wrapper);			\
 	}								\
 	return (wrapper);						\
 }									\
+a_attr bool								\
+a_name##tsd_boot0(void)							\
+{									\
+									\
+	a_name##tsd_tsd = TlsAlloc();					\
+	if (a_name##tsd_tsd == TLS_OUT_OF_INDEXES)			\
+		return (true);						\
+	if (a_cleanup != malloc_tsd_no_cleanup) {			\
+		malloc_tsd_cleanup_register(				\
+		    &a_name##tsd_cleanup_wrapper);			\
+	}								\
+	a_name##tsd_wrapper_set(&a_name##tsd_boot_wrapper);		\
+	a_name##tsd_booted = true;					\
+	return (false);							\
+}									\
+a_attr void								\
+a_name##tsd_boot1()							\
+{									\
+	a_name##tsd_wrapper_t *wrapper;					\
+	wrapper = (a_name##tsd_wrapper_t *)				\
+	    malloc_tsd_malloc(sizeof(a_name##tsd_wrapper_t));		\
+	if (wrapper == NULL) {						\
+		malloc_write("<jemalloc>: Error allocating"		\
+		    " TSD for "#a_name"\n");				\
+		abort();						\
+	}								\
+	memcpy(wrapper, &a_name##tsd_boot_wrapper,			\
+	    sizeof(a_name##tsd_wrapper_t));				\
+	a_name##tsd_wrapper_set(wrapper);				\
+}									\
+a_attr bool								\
+a_name##tsd_boot(void)							\
+{									\
+									\
+	if (a_name##tsd_boot0())					\
+		return (true);						\
+	a_name##tsd_boot1();						\
+	return (false);							\
+}									\
+a_attr bool								\
+a_name##tsd_boot(void)							\
+{									\
+									\
+	return (false);							\
+}									\
+/* Get/set. */								\
 a_attr a_type *								\
 a_name##tsd_get(void)							\
 {									\
 	a_name##tsd_wrapper_t *wrapper;					\
 									\
 	assert(a_name##tsd_booted);					\
-	wrapper = a_name##tsd_get_wrapper();				\
+	wrapper = a_name##tsd_wrapper_get();				\
 	return (&wrapper->val);						\
 }									\
 a_attr void								\
@@ -296,7 +386,7 @@ a_name##tsd_set(a_type *val)						\
 	a_name##tsd_wrapper_t *wrapper;					\
 									\
 	assert(a_name##tsd_booted);					\
-	wrapper = a_name##tsd_get_wrapper();				\
+	wrapper = a_name##tsd_wrapper_get();				\
 	wrapper->val = *(val);						\
 	if (a_cleanup != malloc_tsd_no_cleanup)				\
 		wrapper->initialized = true;				\
@@ -304,11 +394,6 @@ a_name##tsd_set(a_type *val)						\
 #else
 #define	malloc_tsd_funcs(a_attr, a_name, a_type, a_initializer,		\
     a_cleanup)								\
-/* Data structure. */							\
-typedef struct {							\
-	bool	initialized;						\
-	a_type	val;							\
-} a_name##tsd_wrapper_t;						\
 /* Initialization/cleanup. */						\
 a_attr void								\
 a_name##tsd_cleanup_wrapper(void *arg)					\
@@ -333,19 +418,19 @@ a_name##tsd_cleanup_wrapper(void *arg)					\
 	}								\
 	malloc_tsd_dalloc(wrapper);					\
 }									\
-a_attr bool								\
-a_name##tsd_boot(void)							\
+a_attr void								\
+a_name##tsd_wrapper_set(a_name##tsd_wrapper_t *wrapper)			\
 {									\
 									\
-	if (pthread_key_create(&a_name##tsd_tsd,			\
-	    a_name##tsd_cleanup_wrapper) != 0)				\
-		return (true);						\
-	a_name##tsd_booted = true;					\
-	return (false);							\
+	if (pthread_setspecific(a_name##tsd_tsd,			\
+	    (void *)wrapper)) {						\
+		malloc_write("<jemalloc>: Error setting"		\
+		    " TSD for "#a_name"\n");				\
+		abort();						\
+	}								\
 }									\
-/* Get/set. */								\
 a_attr a_name##tsd_wrapper_t *						\
-a_name##tsd_get_wrapper(void)						\
+a_name##tsd_wrapper_get(void)						\
 {									\
 	a_name##tsd_wrapper_t *wrapper = (a_name##tsd_wrapper_t *)	\
 	    pthread_getspecific(a_name##tsd_tsd);			\
@@ -367,23 +452,54 @@ a_name##tsd_get_wrapper(void)						\
 			wrapper->initialized = false;			\
 			wrapper->val = a_initializer;			\
 		}							\
-		if (pthread_setspecific(a_name##tsd_tsd,		\
-		    (void *)wrapper)) {					\
-			malloc_write("<jemalloc>: Error setting"	\
-			    " TSD for "#a_name"\n");			\
-			abort();					\
-		}							\
+		a_name##tsd_wrapper_set(wrapper);			\
 		tsd_init_finish(&a_name##tsd_init_head, &block);	\
 	}								\
 	return (wrapper);						\
 }									\
+a_attr bool								\
+a_name##tsd_boot0(void)							\
+{									\
+									\
+	if (pthread_key_create(&a_name##tsd_tsd,			\
+	    a_name##tsd_cleanup_wrapper) != 0)				\
+		return (true);						\
+	a_name##tsd_wrapper_set(&a_name##tsd_boot_wrapper);		\
+	a_name##tsd_booted = true;					\
+	return (false);							\
+}									\
+a_attr void								\
+a_name##tsd_boot1()							\
+{									\
+	a_name##tsd_wrapper_t *wrapper;					\
+	wrapper = (a_name##tsd_wrapper_t *)				\
+	    malloc_tsd_malloc(sizeof(a_name##tsd_wrapper_t));		\
+	if (wrapper == NULL) {						\
+		malloc_write("<jemalloc>: Error allocating"		\
+		    " TSD for "#a_name"\n");				\
+		abort();						\
+	}								\
+	memcpy(wrapper, &a_name##tsd_boot_wrapper,			\
+	    sizeof(a_name##tsd_wrapper_t));				\
+	a_name##tsd_wrapper_set(wrapper);				\
+}									\
+a_attr bool								\
+a_name##tsd_boot(void)							\
+{									\
+									\
+	if (a_name##tsd_boot0())					\
+		return (true);						\
+	a_name##tsd_boot1();						\
+	return (false);							\
+}									\
+/* Get/set. */								\
 a_attr a_type *								\
 a_name##tsd_get(void)							\
 {									\
 	a_name##tsd_wrapper_t *wrapper;					\
 									\
 	assert(a_name##tsd_booted);					\
-	wrapper = a_name##tsd_get_wrapper();				\
+	wrapper = a_name##tsd_wrapper_get();				\
 	return (&wrapper->val);						\
 }									\
 a_attr void								\
@@ -392,7 +508,7 @@ a_name##tsd_set(a_type *val)						\
 	a_name##tsd_wrapper_t *wrapper;					\
 									\
 	assert(a_name##tsd_booted);					\
-	wrapper = a_name##tsd_get_wrapper();				\
+	wrapper = a_name##tsd_wrapper_get();				\
 	wrapper->val = *(val);						\
 	if (a_cleanup != malloc_tsd_no_cleanup)				\
 		wrapper->initialized = true;				\
@@ -423,6 +539,9 @@ struct tsd_init_head_s {
     O(thread_deallocated,	uint64_t)				\
     O(prof_tdata,		prof_tdata_t *)				\
     O(arena,			arena_t *)				\
+    O(arenas_cache,		arena_t **)				\
+    O(narenas_cache,		unsigned)				\
+    O(arenas_cache_bypass,	bool)					\
     O(tcache_enabled,		tcache_enabled_t)			\
     O(quarantine,		quarantine_t *)				\
 
@@ -433,6 +552,9 @@ struct tsd_init_head_s {
     0,									\
     NULL,								\
     NULL,								\
+    NULL,								\
+    0,									\
+    false,								\
     tcache_enabled_default,						\
     NULL								\
 }
@@ -447,6 +569,8 @@ MALLOC_TSD
 
 static const tsd_t tsd_initializer = TSD_INITIALIZER;
 
+malloc_tsd_types(, tsd_t)
+
 #endif /* JEMALLOC_H_STRUCTS */
 /******************************************************************************/
 #ifdef JEMALLOC_H_EXTERNS
@@ -455,7 +579,8 @@ void	*malloc_tsd_malloc(size_t size);
 void	malloc_tsd_dalloc(void *wrapper);
 void	malloc_tsd_no_cleanup(void *arg);
 void	malloc_tsd_cleanup_register(bool (*f)(void));
-bool	malloc_tsd_boot(void);
+bool	malloc_tsd_boot0(void);
+void	malloc_tsd_boot1(void);
 #if (!defined(JEMALLOC_MALLOC_THREAD_CLEANUP) && !defined(JEMALLOC_TLS) && \
     !defined(_WIN32))
 void	*tsd_init_check_recursion(tsd_init_head_t *head,
diff --git a/src/arena.c b/src/arena.c
index 49a30572..86e54404 100644
--- a/src/arena.c
+++ b/src/arena.c
@@ -2192,27 +2192,37 @@ arena_stats_merge(arena_t *arena, const char **dss, size_t *nactive,
 	}
 }
 
-bool
-arena_new(arena_t *arena, unsigned ind)
+arena_t *
+arena_new(unsigned ind)
 {
+	arena_t *arena;
 	unsigned i;
 	arena_bin_t *bin;
 
+	/*
+	 * Allocate arena and arena->lstats contiguously, mainly because there
+	 * is no way to clean up if base_alloc() OOMs.
+	 */
+	if (config_stats) {
+		arena = (arena_t *)base_alloc(CACHELINE_CEILING(sizeof(arena_t))
+		    + nlclasses * sizeof(malloc_large_stats_t));
+	} else
+		arena = (arena_t *)base_alloc(sizeof(arena_t));
+	if (arena == NULL)
+		return (NULL);
+
 	arena->ind = ind;
 	arena->nthreads = 0;
 	arena->chunk_alloc = chunk_alloc_default;
 	arena->chunk_dalloc = chunk_dalloc_default;
 
 	if (malloc_mutex_init(&arena->lock))
-		return (true);
+		return (NULL);
 
 	if (config_stats) {
 		memset(&arena->stats, 0, sizeof(arena_stats_t));
-		arena->stats.lstats =
-		    (malloc_large_stats_t *)base_alloc(nlclasses *
-		    sizeof(malloc_large_stats_t));
-		if (arena->stats.lstats == NULL)
-			return (true);
+		arena->stats.lstats = (malloc_large_stats_t *)(((void *)arena) +
+		    CACHELINE_CEILING(sizeof(arena_t)));
 		memset(arena->stats.lstats, 0, nlclasses *
 		    sizeof(malloc_large_stats_t));
 		if (config_tcache)
@@ -2236,14 +2246,14 @@ arena_new(arena_t *arena, unsigned ind)
 	for (i = 0; i < NBINS; i++) {
 		bin = &arena->bins[i];
 		if (malloc_mutex_init(&bin->lock))
-			return (true);
+			return (NULL);
 		bin->runcur = NULL;
 		arena_run_tree_new(&bin->runs);
 		if (config_stats)
 			memset(&bin->stats, 0, sizeof(malloc_bin_stats_t));
 	}
 
-	return (false);
+	return (arena);
 }
 
 /*
diff --git a/src/chunk.c b/src/chunk.c
index 618aaca0..f65b67af 100644
--- a/src/chunk.c
+++ b/src/chunk.c
@@ -254,9 +254,17 @@ void *
 chunk_alloc_default(void *new_addr, size_t size, size_t alignment, bool *zero,
     unsigned arena_ind)
 {
+	arena_t *arena;
+
+	arena = arena_get(tsd_fetch(), arena_ind, false, true);
+	/*
+	 * The arena we're allocating on behalf of must have been initialized
+	 * already.
+	 */
+	assert(arena != NULL);
 
 	return (chunk_alloc_core(new_addr, size, alignment, false, zero,
-	    arenas[arena_ind]->dss_prec));
+	    arena->dss_prec));
 }
 
 static void
diff --git a/src/ctl.c b/src/ctl.c
index f1f3234b..37f8f42a 100644
--- a/src/ctl.c
+++ b/src/ctl.c
@@ -447,7 +447,7 @@ ctl_arena_init(ctl_arena_stats_t *astats)
 {
 
 	if (astats->lstats == NULL) {
-		astats->lstats = (malloc_large_stats_t *)base_alloc(nlclasses *
+		astats->lstats = (malloc_large_stats_t *)a0malloc(nlclasses *
 		    sizeof(malloc_large_stats_t));
 		if (astats->lstats == NULL)
 			return (true);
@@ -567,31 +567,24 @@ ctl_arena_refresh(arena_t *arena, unsigned i)
 static bool
 ctl_grow(void)
 {
-	tsd_t *tsd;
 	ctl_arena_stats_t *astats;
-	arena_t **tarenas;
 
-	tsd = tsd_fetch();
+	/* Initialize new arena. */
+	if (arena_init(ctl_stats.narenas) == NULL)
+		return (true);
 
-	/* Allocate extended arena stats and arenas arrays. */
-	astats = (ctl_arena_stats_t *)imalloc(tsd, (ctl_stats.narenas + 2) *
+	/* Allocate extended arena stats. */
+	astats = (ctl_arena_stats_t *)a0malloc((ctl_stats.narenas + 2) *
 	    sizeof(ctl_arena_stats_t));
 	if (astats == NULL)
 		return (true);
-	tarenas = (arena_t **)imalloc(tsd, (ctl_stats.narenas + 1) *
-	    sizeof(arena_t *));
-	if (tarenas == NULL) {
-		idalloc(tsd, astats);
-		return (true);
-	}
 
 	/* Initialize the new astats element. */
 	memcpy(astats, ctl_stats.arenas, (ctl_stats.narenas + 1) *
 	    sizeof(ctl_arena_stats_t));
 	memset(&astats[ctl_stats.narenas + 1], 0, sizeof(ctl_arena_stats_t));
 	if (ctl_arena_init(&astats[ctl_stats.narenas + 1])) {
-		idalloc(tsd, tarenas);
-		idalloc(tsd, astats);
+		a0free(astats);
 		return (true);
 	}
 	/* Swap merged stats to their new location. */
@@ -604,32 +597,7 @@ ctl_grow(void)
 		memcpy(&astats[ctl_stats.narenas + 1], &tstats,
 		    sizeof(ctl_arena_stats_t));
 	}
-	/* Initialize the new arenas element. */
-	tarenas[ctl_stats.narenas] = NULL;
-	{
-		arena_t **arenas_old = arenas;
-		/*
-		 * Swap extended arenas array into place.  Although ctl_mtx
-		 * protects this function from other threads extending the
-		 * array, it does not protect from other threads mutating it
-		 * (i.e. initializing arenas and setting array elements to
-		 * point to them).  Therefore, array copying must happen under
-		 * the protection of arenas_lock.
-		 */
-		malloc_mutex_lock(&arenas_lock);
-		arenas = tarenas;
-		memcpy(arenas, arenas_old, ctl_stats.narenas *
-		    sizeof(arena_t *));
-		narenas_total++;
-		arenas_extend(narenas_total - 1);
-		malloc_mutex_unlock(&arenas_lock);
-		/*
-		 * Deallocate arenas_old only if it came from imalloc() (not
-		 * base_alloc()).
-		 */
-		if (ctl_stats.narenas != narenas_auto)
-			idalloc(tsd, arenas_old);
-	}
+	a0free(ctl_stats.arenas);
 	ctl_stats.arenas = astats;
 	ctl_stats.narenas++;
 
@@ -639,6 +607,7 @@ ctl_grow(void)
 static void
 ctl_refresh(void)
 {
+	tsd_t *tsd;
 	unsigned i;
 	VARIABLE_ARRAY(arena_t *, tarenas, ctl_stats.narenas);
 
@@ -657,15 +626,17 @@ ctl_refresh(void)
 	ctl_stats.arenas[ctl_stats.narenas].nthreads = 0;
 	ctl_arena_clear(&ctl_stats.arenas[ctl_stats.narenas]);
 
-	malloc_mutex_lock(&arenas_lock);
-	memcpy(tarenas, arenas, sizeof(arena_t *) * ctl_stats.narenas);
+	tsd = tsd_fetch();
+	for (i = 0; i < ctl_stats.narenas; i++)
+		tarenas[i] = arena_get(tsd, i, false, (i == 0));
+
 	for (i = 0; i < ctl_stats.narenas; i++) {
-		if (arenas[i] != NULL)
-			ctl_stats.arenas[i].nthreads = arenas[i]->nthreads;
+		if (tarenas[i] != NULL)
+			ctl_stats.arenas[i].nthreads = arena_nbound(i);
 		else
 			ctl_stats.arenas[i].nthreads = 0;
 	}
-	malloc_mutex_unlock(&arenas_lock);
+
 	for (i = 0; i < ctl_stats.narenas; i++) {
 		bool initialized = (tarenas[i] != NULL);
 
@@ -698,9 +669,8 @@ ctl_init(void)
 		 * Allocate space for one extra arena stats element, which
 		 * contains summed stats across all arenas.
 		 */
-		assert(narenas_auto == narenas_total_get());
-		ctl_stats.narenas = narenas_auto;
-		ctl_stats.arenas = (ctl_arena_stats_t *)base_alloc(
+		ctl_stats.narenas = narenas_total_get();
+		ctl_stats.arenas = (ctl_arena_stats_t *)a0malloc(
 		    (ctl_stats.narenas + 1) * sizeof(ctl_arena_stats_t));
 		if (ctl_stats.arenas == NULL) {
 			ret = true;
@@ -718,6 +688,13 @@ ctl_init(void)
 			unsigned i;
 			for (i = 0; i <= ctl_stats.narenas; i++) {
 				if (ctl_arena_init(&ctl_stats.arenas[i])) {
+					unsigned j;
+					for (j = 0; j < i; j++) {
+						a0free(
+						    ctl_stats.arenas[j].lstats);
+					}
+					a0free(ctl_stats.arenas);
+					ctl_stats.arenas = NULL;
 					ret = true;
 					goto label_return;
 				}
@@ -1231,17 +1208,19 @@ thread_arena_ctl(const size_t *mib, size_t miblen, void *oldp, size_t *oldlenp,
 {
 	int ret;
 	tsd_t *tsd;
+	arena_t *arena;
 	unsigned newind, oldind;
 
 	tsd = tsd_fetch();
+	arena = arena_choose(tsd, NULL);
+	if (arena == NULL)
+		return (EAGAIN);
 
 	malloc_mutex_lock(&ctl_mtx);
-	newind = oldind = choose_arena(tsd, NULL)->ind;
+	newind = oldind = arena->ind;
 	WRITE(newind, unsigned);
 	READ(oldind, unsigned);
 	if (newind != oldind) {
-		arena_t *arena;
-
 		if (newind >= ctl_stats.narenas) {
 			/* New arena index is out of range. */
 			ret = EFAULT;
@@ -1249,28 +1228,18 @@ thread_arena_ctl(const size_t *mib, size_t miblen, void *oldp, size_t *oldlenp,
 		}
 
 		/* Initialize arena if necessary. */
-		malloc_mutex_lock(&arenas_lock);
-		if ((arena = arenas[newind]) == NULL && (arena =
-		    arenas_extend(newind)) == NULL) {
-			malloc_mutex_unlock(&arenas_lock);
+		arena = arena_get(tsd, newind, true, true);
+		if (arena == NULL) {
 			ret = EAGAIN;
 			goto label_return;
 		}
-		assert(arena == arenas[newind]);
-		arenas[oldind]->nthreads--;
-		arenas[newind]->nthreads++;
-		malloc_mutex_unlock(&arenas_lock);
-
-		/* Set new arena association. */
+		/* Set new arena/tcache associations. */
+		arena_migrate(tsd, oldind, newind);
 		if (config_tcache) {
 			tcache_t *tcache = tsd_tcache_get(tsd);
-			if (tcache != NULL) {
-				tcache_arena_dissociate(tcache);
-				tcache_arena_associate(tcache, arena);
-			}
+			if (tcache != NULL)
+				tcache_arena_reassociate(tcache, arena);
 		}
-
-		tsd_arena_set(tsd, arena);
 	}
 
 	ret = 0;
@@ -1400,11 +1369,13 @@ label_return:
 static void
 arena_purge(unsigned arena_ind)
 {
+	tsd_t *tsd;
+	unsigned i;
 	VARIABLE_ARRAY(arena_t *, tarenas, ctl_stats.narenas);
 
-	malloc_mutex_lock(&arenas_lock);
-	memcpy(tarenas, arenas, sizeof(arena_t *) * ctl_stats.narenas);
-	malloc_mutex_unlock(&arenas_lock);
+	tsd = tsd_fetch();
+	for (i = 0; i < ctl_stats.narenas; i++)
+		tarenas[i] = arena_get(tsd, i, false, (i == 0));
 
 	if (arena_ind == ctl_stats.narenas) {
 		unsigned i;
@@ -1467,7 +1438,7 @@ 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 = arenas[arena_ind];
+		arena_t *arena = arena_get(tsd_fetch(), arena_ind, false, true);
 		if (arena == NULL || (dss_prec != dss_prec_limit &&
 		    arena_dss_prec_set(arena, dss_prec))) {
 			ret = EFAULT;
@@ -1501,7 +1472,8 @@ arena_i_chunk_alloc_ctl(const size_t *mib, size_t miblen, void *oldp,
 	arena_t *arena;
 
 	malloc_mutex_lock(&ctl_mtx);
-	if (arena_ind < narenas_total && (arena = arenas[arena_ind]) != NULL) {
+	if (arena_ind < narenas_total_get() && (arena = arena_get(tsd_fetch(),
+	    arena_ind, false, true)) != NULL) {
 		malloc_mutex_lock(&arena->lock);
 		READ(arena->chunk_alloc, chunk_alloc_t *);
 		WRITE(arena->chunk_alloc, chunk_alloc_t *);
@@ -1527,7 +1499,8 @@ arena_i_chunk_dalloc_ctl(const size_t *mib, size_t miblen, void *oldp,
 	arena_t *arena;
 
 	malloc_mutex_lock(&ctl_mtx);
-	if (arena_ind < narenas_total && (arena = arenas[arena_ind]) != NULL) {
+	if (arena_ind < narenas_total_get() && (arena = arena_get(tsd_fetch(),
+	    arena_ind, false, true)) != NULL) {
 		malloc_mutex_lock(&arena->lock);
 		READ(arena->chunk_dalloc, chunk_dalloc_t *);
 		WRITE(arena->chunk_dalloc, chunk_dalloc_t *);
diff --git a/src/huge.c b/src/huge.c
index ae416253..1376729a 100644
--- a/src/huge.c
+++ b/src/huge.c
@@ -50,7 +50,11 @@ huge_palloc(tsd_t *tsd, arena_t *arena, size_t usize, size_t alignment,
 	 * it is possible to make correct junk/zero fill decisions below.
 	 */
 	is_zeroed = zero;
-	arena = choose_arena(tsd, arena);
+	arena = arena_choose(tsd, arena);
+	if (unlikely(arena == NULL)) {
+		base_node_dalloc(node);
+		return (NULL);
+	}
 	ret = arena_chunk_alloc_huge(arena, NULL, csize, alignment, &is_zeroed);
 	if (ret == NULL) {
 		base_node_dalloc(node);
diff --git a/src/jemalloc.c b/src/jemalloc.c
index f3750b40..3c889e8a 100644
--- a/src/jemalloc.c
+++ b/src/jemalloc.c
@@ -4,8 +4,6 @@
 /******************************************************************************/
 /* Data. */
 
-malloc_tsd_data(, arenas, arena_t *, NULL)
-
 /* Runtime configuration options. */
 const char	*je_malloc_conf JEMALLOC_ATTR(weak);
 bool	opt_abort =
@@ -34,10 +32,20 @@ bool	in_valgrind;
 
 unsigned	ncpus;
 
-malloc_mutex_t		arenas_lock;
-arena_t			**arenas;
-unsigned		narenas_total;
-unsigned		narenas_auto;
+/* Protects arenas initialization (arenas, narenas_total). */
+static malloc_mutex_t	arenas_lock;
+/*
+ * Arenas that are used to service external requests.  Not all elements of the
+ * arenas array are necessarily used; arenas are created lazily as needed.
+ *
+ * arenas[0..narenas_auto) are used for automatic multiplexing of threads and
+ * 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;
+static unsigned		narenas_total;
+static arena_t		*a0; /* arenas[0]; read-only after initialization. */
+static unsigned		narenas_auto; /* Read-only after initialization. */
 
 /* Set to true once the allocator has been initialized. */
 static bool		malloc_initialized = false;
@@ -144,35 +152,288 @@ static bool	malloc_init_hard(void);
  * Begin miscellaneous support functions.
  */
 
-/* Create a new arena and insert it into the arenas array at index ind. */
-arena_t *
-arenas_extend(unsigned ind)
+JEMALLOC_ALWAYS_INLINE_C void
+malloc_thread_init(void)
 {
-	arena_t *ret;
-
-	ret = (arena_t *)base_alloc(sizeof(arena_t));
-	if (ret != NULL && !arena_new(ret, ind)) {
-		arenas[ind] = ret;
-		return (ret);
-	}
-	/* Only reached if there is an OOM error. */
 
 	/*
-	 * OOM here is quite inconvenient to propagate, since dealing with it
-	 * would require a check for failure in the fast path.  Instead, punt
-	 * by using arenas[0].  In practice, this is an extremely unlikely
-	 * failure.
+	 * TSD initialization can't be safely done as a side effect of
+	 * deallocation, because it is possible for a thread to do nothing but
+	 * deallocate its TLS data via free(), in which case writing to TLS
+	 * would cause write-after-free memory corruption.  The quarantine
+	 * facility *only* gets used as a side effect of deallocation, so make
+	 * a best effort attempt at initializing its TSD by hooking all
+	 * allocation events.
 	 */
-	malloc_write("<jemalloc>: Error initializing arena\n");
-	if (opt_abort)
-		abort();
-
-	return (arenas[0]);
+	if (config_fill && unlikely(opt_quarantine))
+		quarantine_alloc_hook();
 }
 
-/* Slow path, called only by choose_arena(). */
+JEMALLOC_ALWAYS_INLINE_C bool
+malloc_init(void)
+{
+
+	if (unlikely(!malloc_initialized) && malloc_init_hard())
+		return (true);
+	malloc_thread_init();
+
+	return (false);
+}
+
+/*
+ * The a0*() functions are used instead of i[mcd]alloc() in bootstrap-sensitive
+ * situations that cannot tolerate TLS variable access.  These functions are
+ * also exposed for use in static binaries on FreeBSD, hence the old-style
+ * malloc() API.
+ */
+
 arena_t *
-choose_arena_hard(tsd_t *tsd)
+a0get(void)
+{
+
+	assert(a0 != NULL);
+	return (a0);
+}
+
+static void *
+a0alloc(size_t size, bool zero)
+{
+	void *ret;
+
+	if (unlikely(malloc_init()))
+		return (NULL);
+
+	if (size == 0)
+		size = 1;
+
+	if (size <= arena_maxclass)
+		ret = arena_malloc(NULL, a0get(), size, zero, false);
+	else
+		ret = huge_malloc(NULL, a0get(), size, zero);
+
+	return (ret);
+}
+
+void *
+a0malloc(size_t size)
+{
+
+	return (a0alloc(size, false));
+}
+
+void *
+a0calloc(size_t num, size_t size)
+{
+
+	return (a0alloc(num * size, true));
+}
+
+void
+a0free(void *ptr)
+{
+	arena_chunk_t *chunk;
+
+	if (ptr == NULL)
+		return;
+
+	chunk = (arena_chunk_t *)CHUNK_ADDR2BASE(ptr);
+	if (chunk != ptr)
+		arena_dalloc(NULL, chunk, ptr, false);
+	else
+		huge_dalloc(ptr);
+}
+
+/* Create a new arena and insert it into the arenas array at index ind. */
+arena_t *
+arena_init(unsigned ind)
+{
+	arena_t *arena;
+
+	malloc_mutex_lock(&arenas_lock);
+
+	/* Expand arenas if necessary. */
+	assert(ind <= narenas_total);
+	if (ind == narenas_total) {
+		unsigned narenas_new = narenas_total + 1;
+		arena_t **arenas_new =
+		    (arena_t **)a0malloc(CACHELINE_CEILING(narenas_new *
+		    sizeof(arena_t *)));
+		if (arenas_new == NULL) {
+			arena = NULL;
+			goto label_return;
+		}
+		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)
+			a0free(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];
+	if (arena != NULL) {
+		assert(ind < narenas_auto);
+		goto label_return;
+	}
+
+	/* Actually initialize the arena. */
+	arena = arenas[ind] = arena_new(ind);
+label_return:
+	malloc_mutex_unlock(&arenas_lock);
+	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)
+{
+
+	malloc_mutex_lock(&arenas_lock);
+	arena_bind_locked(tsd, ind);
+	malloc_mutex_unlock(&arenas_lock);
+}
+
+void
+arena_migrate(tsd_t *tsd, unsigned oldind, unsigned newind)
+{
+	arena_t *oldarena, *newarena;
+
+	malloc_mutex_lock(&arenas_lock);
+	oldarena = arenas[oldind];
+	newarena = arenas[newind];
+	oldarena->nthreads--;
+	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 = arenas[ind];
+	arena->nthreads--;
+	malloc_mutex_unlock(&arenas_lock);
+	tsd_arena_set(tsd, NULL);
+}
+
+arena_t *
+arena_get_hard(tsd_t *tsd, unsigned ind, bool init_if_missing)
+{
+	arena_t *arena;
+	arena_t **arenas_cache = tsd_arenas_cache_get(tsd);
+	unsigned narenas_cache = tsd_narenas_cache_get(tsd);
+	unsigned narenas_actual = narenas_total_get();
+
+	/* Deallocate old cache if it's too small. */
+	if (arenas_cache != NULL && narenas_cache < narenas_actual) {
+		a0free(arenas_cache);
+		arenas_cache = NULL;
+		narenas_cache = 0;
+		tsd_arenas_cache_set(tsd, arenas_cache);
+		tsd_narenas_cache_set(tsd, narenas_cache);
+	}
+
+	/* Allocate cache if it's missing. */
+	if (arenas_cache == NULL) {
+		bool *arenas_cache_bypassp = tsd_arenas_cache_bypassp_get(tsd);
+		assert(ind < narenas_actual || !init_if_missing);
+		narenas_cache = (ind < narenas_actual) ? narenas_actual : ind+1;
+
+		if (!*arenas_cache_bypassp) {
+			*arenas_cache_bypassp = true;
+			arenas_cache = (arena_t **)a0malloc(sizeof(arena_t *) *
+			    narenas_cache);
+			*arenas_cache_bypassp = false;
+		} else
+			arenas_cache = NULL;
+		if (arenas_cache == NULL) {
+			/*
+			 * This function must always tell the truth, even if
+			 * it's slow, so don't let OOM or recursive allocation
+			 * avoidance (note arenas_cache_bypass check) get in the
+			 * way.
+			 */
+			if (ind >= narenas_actual)
+				return (NULL);
+			malloc_mutex_lock(&arenas_lock);
+			arena = arenas[ind];
+			malloc_mutex_unlock(&arenas_lock);
+			return (arena);
+		}
+		tsd_arenas_cache_set(tsd, arenas_cache);
+		tsd_narenas_cache_set(tsd, narenas_cache);
+	}
+
+	/*
+	 * Copy to cache.  It's possible that the actual number of arenas has
+	 * increased since narenas_total_get() was called above, but that causes
+	 * no correctness issues unless two threads concurrently execute the
+	 * arenas.extend mallctl, which we trust mallctl synchronization to
+	 * prevent.
+	 */
+	malloc_mutex_lock(&arenas_lock);
+	memcpy(arenas_cache, arenas, sizeof(arena_t *) * narenas_actual);
+	malloc_mutex_unlock(&arenas_lock);
+	if (narenas_cache > narenas_actual) {
+		memset(&arenas_cache[narenas_actual], 0, sizeof(arena_t *) *
+		    (narenas_cache - narenas_actual));
+	}
+
+	/* Read the refreshed cache, and init the arena if necessary. */
+	arena = arenas_cache[ind];
+	if (init_if_missing && arena == NULL)
+		arena = arenas_cache[ind] = arena_init(ind);
+	return (arena);
+}
+
+/* Slow path, called only by arena_choose(). */
+arena_t *
+arena_choose_hard(tsd_t *tsd)
 {
 	arena_t *ret;
 
@@ -182,7 +443,7 @@ choose_arena_hard(tsd_t *tsd)
 		choose = 0;
 		first_null = narenas_auto;
 		malloc_mutex_lock(&arenas_lock);
-		assert(arenas[0] != NULL);
+		assert(a0get() != NULL);
 		for (i = 1; i < narenas_auto; i++) {
 			if (arenas[i] != NULL) {
 				/*
@@ -215,20 +476,20 @@ choose_arena_hard(tsd_t *tsd)
 			ret = arenas[choose];
 		} else {
 			/* Initialize a new arena. */
-			ret = arenas_extend(first_null);
+			choose = first_null;
+			ret = arena_init(choose);
+			if (ret == NULL) {
+				malloc_mutex_unlock(&arenas_lock);
+				return (NULL);
+			}
 		}
-		ret->nthreads++;
+		arena_bind_locked(tsd, choose);
 		malloc_mutex_unlock(&arenas_lock);
 	} else {
-		ret = arenas[0];
-		malloc_mutex_lock(&arenas_lock);
-		ret->nthreads++;
-		malloc_mutex_unlock(&arenas_lock);
+		ret = a0get();
+		arena_bind(tsd, 0);
 	}
 
-	if (tsd_nominal(tsd))
-		tsd_arena_set(tsd, ret);
-
 	return (ret);
 }
 
@@ -248,6 +509,33 @@ thread_deallocated_cleanup(tsd_t *tsd)
 
 void
 arena_cleanup(tsd_t *tsd)
+{
+	arena_t *arena;
+
+	arena = tsd_arena_get(tsd);
+	if (arena != NULL)
+		arena_unbind(tsd, arena->ind);
+}
+
+void
+arenas_cache_cleanup(tsd_t *tsd)
+{
+	arena_t **arenas_cache;
+
+	arenas_cache = tsd_arenas_cache_get(tsd);
+	if (arenas != NULL)
+		a0free(arenas_cache);
+}
+
+void
+narenas_cache_cleanup(tsd_t *tsd)
+{
+
+	/* Do nothing. */
+}
+
+void
+arenas_cache_bypass_cleanup(tsd_t *tsd)
 {
 
 	/* Do nothing. */
@@ -312,44 +600,6 @@ malloc_ncpus(void)
 	return ((result == -1) ? 1 : (unsigned)result);
 }
 
-void
-arenas_cleanup(void *arg)
-{
-	arena_t *arena = *(arena_t **)arg;
-
-	malloc_mutex_lock(&arenas_lock);
-	arena->nthreads--;
-	malloc_mutex_unlock(&arenas_lock);
-}
-
-JEMALLOC_ALWAYS_INLINE_C void
-malloc_thread_init(void)
-{
-
-	/*
-	 * TSD initialization can't be safely done as a side effect of
-	 * deallocation, because it is possible for a thread to do nothing but
-	 * deallocate its TLS data via free(), in which case writing to TLS
-	 * would cause write-after-free memory corruption.  The quarantine
-	 * facility *only* gets used as a side effect of deallocation, so make
-	 * a best effort attempt at initializing its TSD by hooking all
-	 * allocation events.
-	 */
-	if (config_fill && unlikely(opt_quarantine))
-		quarantine_alloc_hook();
-}
-
-JEMALLOC_ALWAYS_INLINE_C bool
-malloc_init(void)
-{
-
-	if (unlikely(!malloc_initialized) && malloc_init_hard())
-		return (true);
-	malloc_thread_init();
-
-	return (false);
-}
-
 static bool
 malloc_conf_next(char const **opts_p, char const **k_p, size_t *klen_p,
     char const **v_p, size_t *vlen_p)
@@ -745,7 +995,7 @@ malloc_init_hard(void)
 #endif
 	malloc_initializer = INITIALIZER;
 
-	if (malloc_tsd_boot()) {
+	if (malloc_tsd_boot0()) {
 		malloc_mutex_unlock(&init_lock);
 		return (true);
 	}
@@ -809,10 +1059,10 @@ malloc_init_hard(void)
 
 	/*
 	 * Initialize one arena here.  The rest are lazily created in
-	 * choose_arena_hard().
+	 * arena_choose_hard().
 	 */
-	arenas_extend(0);
-	if (arenas[0] == NULL) {
+	a0 = arena_init(0);
+	if (a0 == NULL) {
 		malloc_mutex_unlock(&init_lock);
 		return (true);
 	}
@@ -887,6 +1137,7 @@ malloc_init_hard(void)
 
 	malloc_initialized = true;
 	malloc_mutex_unlock(&init_lock);
+	malloc_tsd_boot1();
 
 	return (false);
 }
@@ -1428,8 +1679,8 @@ JEMALLOC_EXPORT void *(*__memalign_hook)(size_t alignment, size_t size) =
  * Begin non-standard functions.
  */
 
-JEMALLOC_ALWAYS_INLINE_C void
-imallocx_flags_decode_hard(size_t size, int flags, size_t *usize,
+JEMALLOC_ALWAYS_INLINE_C bool
+imallocx_flags_decode_hard(tsd_t *tsd, size_t size, int flags, size_t *usize,
     size_t *alignment, bool *zero, bool *try_tcache, arena_t **arena)
 {
 
@@ -1444,16 +1695,19 @@ imallocx_flags_decode_hard(size_t size, int flags, size_t *usize,
 	if ((flags & MALLOCX_ARENA_MASK) != 0) {
 		unsigned arena_ind = MALLOCX_ARENA_GET(flags);
 		*try_tcache = false;
-		*arena = arenas[arena_ind];
+		*arena = arena_get(tsd, arena_ind, true, true);
+		if (unlikely(*arena == NULL))
+			return (true);
 	} else {
 		*try_tcache = true;
 		*arena = NULL;
 	}
+	return (false);
 }
 
-JEMALLOC_ALWAYS_INLINE_C void
-imallocx_flags_decode(size_t size, int flags, size_t *usize, size_t *alignment,
-    bool *zero, bool *try_tcache, arena_t **arena)
+JEMALLOC_ALWAYS_INLINE_C bool
+imallocx_flags_decode(tsd_t *tsd, size_t size, int flags, size_t *usize,
+    size_t *alignment, bool *zero, bool *try_tcache, arena_t **arena)
 {
 
 	if (likely(flags == 0)) {
@@ -1463,9 +1717,10 @@ imallocx_flags_decode(size_t size, int flags, size_t *usize, size_t *alignment,
 		*zero = false;
 		*try_tcache = true;
 		*arena = NULL;
+		return (false);
 	} else {
-		imallocx_flags_decode_hard(size, flags, usize, alignment, zero,
-		    try_tcache, arena);
+		return (imallocx_flags_decode_hard(tsd, size, flags, usize,
+		    alignment, zero, try_tcache, arena));
 	}
 }
 
@@ -1524,8 +1779,9 @@ imallocx_prof(tsd_t *tsd, size_t size, int flags, size_t *usize)
 	arena_t *arena;
 	prof_tctx_t *tctx;
 
-	imallocx_flags_decode(size, flags, usize, &alignment, &zero,
-	    &try_tcache, &arena);
+	if (unlikely(imallocx_flags_decode(tsd, size, flags, usize, &alignment,
+	    &zero, &try_tcache, &arena)))
+		return (NULL);
 	tctx = prof_alloc_prep(tsd, *usize, true);
 	if (likely((uintptr_t)tctx == (uintptr_t)1U)) {
 		p = imallocx_maybe_flags(tsd, size, flags, *usize, alignment,
@@ -1558,8 +1814,9 @@ imallocx_no_prof(tsd_t *tsd, size_t size, int flags, size_t *usize)
 		return (imalloc(tsd, size));
 	}
 
-	imallocx_flags_decode_hard(size, flags, usize, &alignment, &zero,
-	    &try_tcache, &arena);
+	if (unlikely(imallocx_flags_decode_hard(tsd, size, flags, usize,
+	    &alignment, &zero, &try_tcache, &arena)))
+		return (NULL);
 	return (imallocx_flags(tsd, *usize, alignment, zero, try_tcache,
 	    arena));
 }
@@ -1685,9 +1942,10 @@ je_rallocx(void *ptr, size_t size, int flags)
 		arena_chunk_t *chunk;
 		try_tcache_alloc = false;
 		chunk = (arena_chunk_t *)CHUNK_ADDR2BASE(ptr);
-		try_tcache_dalloc = (chunk == ptr || chunk->arena !=
-		    arenas[arena_ind]);
-		arena = arenas[arena_ind];
+		arena = arena_get(tsd, arena_ind, true, true);
+		if (unlikely(arena == NULL))
+			goto label_oom;
+		try_tcache_dalloc = (chunk == ptr || chunk->arena != arena);
 	} else {
 		try_tcache_alloc = true;
 		try_tcache_dalloc = true;
@@ -1825,6 +2083,7 @@ je_xallocx(void *ptr, size_t size, size_t extra, int flags)
 
 	if (unlikely((flags & MALLOCX_ARENA_MASK) != 0)) {
 		unsigned arena_ind = MALLOCX_ARENA_GET(flags);
+		// XX Dangerous arenas read.
 		arena = arenas[arena_ind];
 	} else
 		arena = NULL;
@@ -1875,16 +2134,24 @@ je_sallocx(const void *ptr, int flags)
 void
 je_dallocx(void *ptr, int flags)
 {
+	tsd_t *tsd;
 	bool try_tcache;
 
 	assert(ptr != NULL);
 	assert(malloc_initialized || IS_INITIALIZER);
 
+	tsd = tsd_fetch();
 	if (unlikely((flags & MALLOCX_ARENA_MASK) != 0)) {
 		unsigned arena_ind = MALLOCX_ARENA_GET(flags);
 		arena_chunk_t *chunk = (arena_chunk_t *)CHUNK_ADDR2BASE(ptr);
-		try_tcache = (chunk == ptr || chunk->arena !=
-		    arenas[arena_ind]);
+		arena_t *arena = arena_get(tsd, arena_ind, true, true);
+		/*
+		 * If arena is NULL, the application passed an arena that has
+		 * never been used before, which is unsupported during
+		 * deallocation.
+		 */
+		assert(arena != NULL);
+		try_tcache = (chunk == ptr || chunk->arena != arena);
 	} else
 		try_tcache = true;
 
@@ -1908,6 +2175,7 @@ inallocx(size_t size, int flags)
 void
 je_sdallocx(void *ptr, size_t size, int flags)
 {
+	tsd_t *tsd;
 	bool try_tcache;
 	size_t usize;
 
@@ -1916,16 +2184,22 @@ je_sdallocx(void *ptr, size_t size, int flags)
 	usize = inallocx(size, flags);
 	assert(usize == isalloc(ptr, config_prof));
 
+	tsd = tsd_fetch();
 	if (unlikely((flags & MALLOCX_ARENA_MASK) != 0)) {
 		unsigned arena_ind = MALLOCX_ARENA_GET(flags);
 		arena_chunk_t *chunk = (arena_chunk_t *)CHUNK_ADDR2BASE(ptr);
-		try_tcache = (chunk == ptr || chunk->arena !=
-		    arenas[arena_ind]);
+		arena_t *arena = arena_get(tsd, arena_ind, true, true);
+		/*
+		 * If arena is NULL, the application passed an arena that has
+		 * never been used before, which is unsupported during
+		 * deallocation.
+		 */
+		try_tcache = (chunk == ptr || chunk->arena != arena);
 	} else
 		try_tcache = true;
 
 	UTRACE(ptr, 0, 0);
-	isfree(tsd_fetch(), ptr, usize, try_tcache);
+	isfree(tsd, ptr, usize, try_tcache);
 }
 
 size_t
@@ -2105,55 +2379,3 @@ jemalloc_postfork_child(void)
 }
 
 /******************************************************************************/
-/*
- * The following functions are used for TLS allocation/deallocation in static
- * binaries on FreeBSD.  The primary difference between these and i[mcd]alloc()
- * is that these avoid accessing TLS variables.
- */
-
-static void *
-a0alloc(size_t size, bool zero)
-{
-
-	if (unlikely(malloc_init()))
-		return (NULL);
-
-	if (size == 0)
-		size = 1;
-
-	if (size <= arena_maxclass)
-		return (arena_malloc(NULL, arenas[0], size, zero, false));
-	else
-		return (huge_malloc(NULL, arenas[0], size, zero));
-}
-
-void *
-a0malloc(size_t size)
-{
-
-	return (a0alloc(size, false));
-}
-
-void *
-a0calloc(size_t num, size_t size)
-{
-
-	return (a0alloc(num * size, true));
-}
-
-void
-a0free(void *ptr)
-{
-	arena_chunk_t *chunk;
-
-	if (ptr == NULL)
-		return;
-
-	chunk = (arena_chunk_t *)CHUNK_ADDR2BASE(ptr);
-	if (chunk != ptr)
-		arena_dalloc(NULL, chunk, ptr, false);
-	else
-		huge_dalloc(ptr);
-}
-
-/******************************************************************************/
diff --git a/src/tcache.c b/src/tcache.c
index 2c968c68..1bf70269 100644
--- a/src/tcache.c
+++ b/src/tcache.c
@@ -245,6 +245,14 @@ tcache_arena_associate(tcache_t *tcache, arena_t *arena)
 	tcache->arena = arena;
 }
 
+void
+tcache_arena_reassociate(tcache_t *tcache, arena_t *arena)
+{
+
+	tcache_arena_dissociate(tcache);
+	tcache_arena_associate(tcache, arena);
+}
+
 void
 tcache_arena_dissociate(tcache_t *tcache)
 {
@@ -261,13 +269,17 @@ tcache_arena_dissociate(tcache_t *tcache)
 tcache_t *
 tcache_get_hard(tsd_t *tsd)
 {
+	arena_t *arena;
 
 	if (!tcache_enabled_get()) {
 		if (tsd_nominal(tsd))
 			tcache_enabled_set(false); /* Memoize. */
 		return (NULL);
 	}
-	return (tcache_create(choose_arena(tsd, NULL)));
+	arena = arena_choose(tsd, NULL);
+	if (unlikely(arena == NULL))
+		return (NULL);
+	return (tcache_create(arena));
 }
 
 tcache_t *
diff --git a/src/tsd.c b/src/tsd.c
index cbc64e44..59253fe3 100644
--- a/src/tsd.c
+++ b/src/tsd.c
@@ -15,16 +15,14 @@ void *
 malloc_tsd_malloc(size_t size)
 {
 
-	/* Avoid choose_arena() in order to dodge bootstrapping issues. */
-	return (arena_malloc(NULL, arenas[0], CACHELINE_CEILING(size), false,
-	    false));
+	return (a0malloc(CACHELINE_CEILING(size)));
 }
 
 void
 malloc_tsd_dalloc(void *wrapper)
 {
 
-	idalloct(NULL, wrapper, false);
+	a0free(wrapper);
 }
 
 void
@@ -106,15 +104,24 @@ MALLOC_TSD
 }
 
 bool
-malloc_tsd_boot(void)
+malloc_tsd_boot0(void)
 {
 
 	ncleanups = 0;
-	if (tsd_boot())
+	if (tsd_boot0())
 		return (true);
+	*tsd_arenas_cache_bypassp_get(tsd_fetch()) = true;
 	return (false);
 }
 
+void
+malloc_tsd_boot1(void)
+{
+
+	tsd_boot1();
+	*tsd_arenas_cache_bypassp_get(tsd_fetch()) = false;
+}
+
 #ifdef _WIN32
 static BOOL WINAPI
 _tls_callback(HINSTANCE hinstDLL, DWORD fdwReason, LPVOID lpvReserved)
diff --git a/test/unit/tsd.c b/test/unit/tsd.c
index eb1c5976..b031c484 100644
--- a/test/unit/tsd.c
+++ b/test/unit/tsd.c
@@ -6,6 +6,7 @@ typedef unsigned int data_t;
 
 static bool data_cleanup_executed;
 
+malloc_tsd_types(data_, data_t)
 malloc_tsd_protos(, data_, data_t)
 
 void