Drop high rank locks when creating threads.

Avoid holding arenas_lock and background_thread_lock when creating background
threads, because pthread_create may take internal locks, and potentially cause
deadlock with jemalloc internal locks.

Author: interwq

include/jemalloc/internal/arena_externs.h

@@ -15,6 +15,7 @@ extern percpu_arena_mode_t opt_percpu_arena;
 extern const char *percpu_arena_mode_names[];
 
 extern const uint64_t h_steps[SMOOTHSTEP_NSTEPS];
+extern malloc_mutex_t arenas_lock;
 
 void arena_stats_large_nrequests_add(tsdn_t *tsdn, arena_stats_t *arena_stats,
     szind_t szind, uint64_t nrequests);

src/arena.c

@@ -2050,17 +2050,6 @@ arena_new(tsdn_t *tsdn, unsigned ind, extent_hooks_t *extent_hooks) {
 			hooks_arena_new_hook();
 		}
 		post_reentrancy(tsdn_tsd(tsdn));
-
-		/* background_thread_create() handles reentrancy internally. */
-		if (have_background_thread) {
-			bool err;
-			malloc_mutex_lock(tsdn, &background_thread_lock);
-			err = background_thread_create(tsdn_tsd(tsdn), ind);
-			malloc_mutex_unlock(tsdn, &background_thread_lock);
-			if (err) {
-				goto label_error;
-			}
-		}
 	}
 
 	return arena;

src/background_thread.c

@@ -352,12 +352,15 @@ background_thread_create(tsd_t *tsd, unsigned arena_ind) {
 	}
 
 	pre_reentrancy(tsd);
+	malloc_mutex_unlock(tsd_tsdn(tsd), &background_thread_lock);
 	/*
 	 * To avoid complications (besides reentrancy), create internal
-	 * background threads with the underlying pthread_create.
+	 * background threads with the underlying pthread_create, and drop
+	 * background_thread_lock (pthread_create may take internal locks).
 	 */
 	int err = pthread_create_wrapper(&info->thread, NULL,
 	    background_thread_entry, (void *)thread_ind);
+	malloc_mutex_lock(tsd_tsdn(tsd), &background_thread_lock);
 	post_reentrancy(tsd);
 
 	if (err != 0) {

src/ctl.c

@@ -1501,6 +1501,7 @@ background_thread_ctl(tsd_t *tsd, const size_t *mib, size_t miblen,
 	}
 	background_thread_ctl_init(tsd_tsdn(tsd));
 
+	malloc_mutex_lock(tsd_tsdn(tsd), &ctl_mtx);
 	malloc_mutex_lock(tsd_tsdn(tsd), &background_thread_lock);
 	if (newp == NULL) {
 		oldval = background_thread_enabled();
@@ -1535,6 +1536,8 @@ background_thread_ctl(tsd_t *tsd, const size_t *mib, size_t miblen,
 	ret = 0;
 label_return:
 	malloc_mutex_unlock(tsd_tsdn(tsd), &background_thread_lock);
+	malloc_mutex_unlock(tsd_tsdn(tsd), &ctl_mtx);
+
 	return ret;
 }
 

src/jemalloc.c

@@ -70,7 +70,7 @@ unsigned	opt_narenas = 0;
 unsigned	ncpus;
 
 /* Protects arenas initialization. */
-static malloc_mutex_t	arenas_lock;
+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.
@@ -335,13 +335,35 @@ arena_init_locked(tsdn_t *tsdn, unsigned ind, extent_hooks_t *extent_hooks) {
 	return arena;
 }
 
+static void
+arena_new_create_background_thread(tsdn_t *tsdn, unsigned ind) {
+	if (ind == 0) {
+		return;
+	}
+	/* background_thread_create() handles reentrancy internally. */
+	if (have_background_thread) {
+		bool err;
+		malloc_mutex_lock(tsdn, &background_thread_lock);
+		err = background_thread_create(tsdn_tsd(tsdn), ind);
+		malloc_mutex_unlock(tsdn, &background_thread_lock);
+		if (err) {
+			malloc_printf("<jemalloc>: error in background thread "
+				      "creation for arena %u. Abort.\n", ind);
+			abort();
+		}
+	}
+}
+
 arena_t *
 arena_init(tsdn_t *tsdn, unsigned ind, extent_hooks_t *extent_hooks) {
 	arena_t *arena;
 
 	malloc_mutex_lock(tsdn, &arenas_lock);
 	arena = arena_init_locked(tsdn, ind, extent_hooks);
 	malloc_mutex_unlock(tsdn, &arenas_lock);
+
+	arena_new_create_background_thread(tsdn, ind);
+
 	return arena;
 }
 
@@ -475,6 +497,7 @@ arena_choose_hard(tsd_t *tsd, bool internal) {
 
 	if (narenas_auto > 1) {
 		unsigned i, j, choose[2], first_null;
+		bool is_new_arena[2];
 
 		/*
 		 * Determine binding for both non-internal and internal
@@ -486,6 +509,7 @@ arena_choose_hard(tsd_t *tsd, bool internal) {
 
 		for (j = 0; j < 2; j++) {
 			choose[j] = 0;
+			is_new_arena[j] = false;
 		}
 
 		first_null = narenas_auto;
@@ -545,13 +569,23 @@ arena_choose_hard(tsd_t *tsd, bool internal) {
 					    &arenas_lock);
 					return NULL;
 				}
+				is_new_arena[j] = true;
 				if (!!j == internal) {
 					ret = arena;
 				}
 			}
 			arena_bind(tsd, choose[j], !!j);
 		}
 		malloc_mutex_unlock(tsd_tsdn(tsd), &arenas_lock);
+
+		for (j = 0; j < 2; j++) {
+			if (is_new_arena[j]) {
+				assert(choose[j] > 0);
+				arena_new_create_background_thread(
+				    tsd_tsdn(tsd), choose[j]);
+			}
+		}
+
 	} else {
 		ret = arena_get(tsd_tsdn(tsd), 0, false);
 		arena_bind(tsd, 0, false);