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_sub_<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); }
  *   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,
-    bool refresh_if_missing);
+arena_t	*arena_get(unsigned ind, bool init_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,
-    bool refresh_if_missing)
+arena_get(unsigned ind, bool init_if_missing)
 {
-	arena_tdata_t *tdata;
+	arena_t *ret;
 
-	/* init_if_missing requires refresh_if_missing. */
-	assert(!init_if_missing || refresh_if_missing);
+	assert(ind <= MALLOCX_ARENA_MAX);
 
-	tdata = arena_tdata_get(tsd, ind, refresh_if_missing);
-	if (unlikely(tdata == NULL || tdata->arena == NULL))
-		return (arena_get_hard(tsd, ind, init_if_missing, tdata));
-
-	return (tdata->arena);
+	ret = arenas[ind];
+	if (unlikely(ret == NULL)) {
+		ret = atomic_read_p((void *)&arenas[ind]);
+		if (init_if_missing && unlikely(ret == NULL))
+			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