background_thread: add max thread count config

Looking at the thread counts in our services, jemalloc's background thread
is useful, but mostly idle.  Add a config option to tune down the number of threads.

doc/jemalloc.xml.in

@@ -761,6 +761,18 @@ mallctl("arena." STRINGIFY(MALLCTL_ARENAS_ALL) ".decay",
         selected pthread-based platforms.</para></listitem>
       </varlistentry>
 
+      <varlistentry id="max_background_threads">
+        <term>
+          <mallctl>max_background_threads</mallctl>
+          (<type>size_t</type>)
+          <literal>rw</literal>
+        </term>
+        <listitem><para>Maximum number of background worker threads that will
+        be created.  This value is capped at <link
+        linkend="opt.max_background_threads"><mallctl>opt.max_background_threads</mallctl></link> at
+        startup.</para></listitem>
+      </varlistentry>
+
       <varlistentry id="config.cache_oblivious">
         <term>
           <mallctl>config.cache_oblivious</mallctl>
@@ -1009,6 +1021,17 @@ mallctl("arena." STRINGIFY(MALLCTL_ARENAS_ALL) ".decay",
         default.</para></listitem>
       </varlistentry>
 
+      <varlistentry id="opt.max_background_threads">
+        <term>
+          <mallctl>opt.max_background_threads</mallctl>
+          (<type>const size_t</type>)
+          <literal>r-</literal>
+        </term>
+        <listitem><para>Maximum number of background threads that will be created
+        if <link linkend="background_thread">background_thread</link> is set.
+        Defaults to number of cpus.</para></listitem>
+      </varlistentry>
+
       <varlistentry id="opt.dirty_decay_ms">
         <term>
           <mallctl>opt.dirty_decay_ms</mallctl>

include/jemalloc/internal/background_thread_externs.h

@@ -2,9 +2,11 @@
 #define JEMALLOC_INTERNAL_BACKGROUND_THREAD_EXTERNS_H
 
 extern bool opt_background_thread;
+extern size_t opt_max_background_threads;
 extern malloc_mutex_t background_thread_lock;
 extern atomic_b_t background_thread_enabled_state;
 extern size_t n_background_threads;
+extern size_t max_background_threads;
 extern background_thread_info_t *background_thread_info;
 extern bool can_enable_background_thread;
 

include/jemalloc/internal/background_thread_structs.h

@@ -8,6 +8,7 @@
 #endif
 
 #define BACKGROUND_THREAD_INDEFINITE_SLEEP UINT64_MAX
+#define MAX_BACKGROUND_THREAD_LIMIT MALLOCX_ARENA_LIMIT
 
 typedef enum {
 	background_thread_stopped,

src/background_thread.c

@@ -11,12 +11,14 @@
 #define BACKGROUND_THREAD_DEFAULT false
 /* Read-only after initialization. */
 bool opt_background_thread = BACKGROUND_THREAD_DEFAULT;
+size_t opt_max_background_threads = MAX_BACKGROUND_THREAD_LIMIT;
 
 /* Used for thread creation, termination and stats. */
 malloc_mutex_t background_thread_lock;
 /* Indicates global state.  Atomic because decay reads this w/o locking. */
 atomic_b_t background_thread_enabled_state;
 size_t n_background_threads;
+size_t max_background_threads;
 /* Thread info per-index. */
 background_thread_info_t *background_thread_info;
 
@@ -287,7 +289,7 @@ background_work_sleep_once(tsdn_t *tsdn, background_thread_info_t *info, unsigne
 	uint64_t min_interval = BACKGROUND_THREAD_INDEFINITE_SLEEP;
 	unsigned narenas = narenas_total_get();
 
-	for (unsigned i = ind; i < narenas; i += ncpus) {
+	for (unsigned i = ind; i < narenas; i += max_background_threads) {
 		arena_t *arena = arena_get(tsdn, i, false);
 		if (!arena) {
 			continue;
@@ -390,7 +392,7 @@ check_background_thread_creation(tsd_t *tsd, unsigned *n_created,
 
 	tsdn_t *tsdn = tsd_tsdn(tsd);
 	malloc_mutex_unlock(tsdn, &background_thread_info[0].mtx);
-	for (unsigned i = 1; i < ncpus; i++) {
+	for (unsigned i = 1; i < max_background_threads; i++) {
 		if (created_threads[i]) {
 			continue;
 		}
@@ -430,9 +432,9 @@ check_background_thread_creation(tsd_t *tsd, unsigned *n_created,
 static void
 background_thread0_work(tsd_t *tsd) {
 	/* Thread0 is also responsible for launching / terminating threads. */
-	VARIABLE_ARRAY(bool, created_threads, ncpus);
+	VARIABLE_ARRAY(bool, created_threads, max_background_threads);
 	unsigned i;
-	for (i = 1; i < ncpus; i++) {
+	for (i = 1; i < max_background_threads; i++) {
 		created_threads[i] = false;
 	}
 	/* Start working, and create more threads when asked. */
@@ -455,7 +457,7 @@ background_thread0_work(tsd_t *tsd) {
 	 * the global background_thread mutex (and is waiting) for us.
 	 */
 	assert(!background_thread_enabled());
-	for (i = 1; i < ncpus; i++) {
+	for (i = 1; i < max_background_threads; i++) {
 		background_thread_info_t *info = &background_thread_info[i];
 		assert(info->state != background_thread_paused);
 		if (created_threads[i]) {
@@ -502,7 +504,7 @@ background_work(tsd_t *tsd, unsigned ind) {
 static void *
 background_thread_entry(void *ind_arg) {
 	unsigned thread_ind = (unsigned)(uintptr_t)ind_arg;
-	assert(thread_ind < ncpus);
+	assert(thread_ind < max_background_threads);
 #ifdef JEMALLOC_HAVE_PTHREAD_SETNAME_NP
 	pthread_setname_np(pthread_self(), "jemalloc_bg_thd");
 #endif
@@ -536,7 +538,7 @@ background_thread_create(tsd_t *tsd, unsigned arena_ind) {
 	malloc_mutex_assert_owner(tsd_tsdn(tsd), &background_thread_lock);
 
 	/* We create at most NCPUs threads. */
-	size_t thread_ind = arena_ind % ncpus;
+	size_t thread_ind = arena_ind % max_background_threads;
 	background_thread_info_t *info = &background_thread_info[thread_ind];
 
 	bool need_new_thread;
@@ -590,9 +592,9 @@ background_threads_enable(tsd_t *tsd) {
 	assert(background_thread_enabled());
 	malloc_mutex_assert_owner(tsd_tsdn(tsd), &background_thread_lock);
 
-	VARIABLE_ARRAY(bool, marked, ncpus);
+	VARIABLE_ARRAY(bool, marked, max_background_threads);
 	unsigned i, nmarked;
-	for (i = 0; i < ncpus; i++) {
+	for (i = 0; i < max_background_threads; i++) {
 		marked[i] = false;
 	}
 	nmarked = 0;
@@ -601,18 +603,18 @@ background_threads_enable(tsd_t *tsd) {
 	/* Mark the threads we need to create for thread 0. */
 	unsigned n = narenas_total_get();
 	for (i = 1; i < n; i++) {
-		if (marked[i % ncpus] ||
+		if (marked[i % max_background_threads] ||
 		    arena_get(tsd_tsdn(tsd), i, false) == NULL) {
 			continue;
 		}
 		background_thread_info_t *info = &background_thread_info[
-		    i % ncpus];
+		    i % max_background_threads];
 		malloc_mutex_lock(tsd_tsdn(tsd), &info->mtx);
 		assert(info->state == background_thread_stopped);
 		background_thread_init(tsd, info);
 		malloc_mutex_unlock(tsd_tsdn(tsd), &info->mtx);
-		marked[i % ncpus] = true;
+		marked[i % max_background_threads] = true;
-		if (++nmarked == ncpus) {
+		if (++nmarked == max_background_threads) {
 			break;
 		}
 	}
@@ -727,14 +729,14 @@ background_thread_prefork0(tsdn_t *tsdn) {
 
 void
 background_thread_prefork1(tsdn_t *tsdn) {
-	for (unsigned i = 0; i < ncpus; i++) {
+	for (unsigned i = 0; i < max_background_threads; i++) {
 		malloc_mutex_prefork(tsdn, &background_thread_info[i].mtx);
 	}
 }
 
 void
 background_thread_postfork_parent(tsdn_t *tsdn) {
-	for (unsigned i = 0; i < ncpus; i++) {
+	for (unsigned i = 0; i < max_background_threads; i++) {
 		malloc_mutex_postfork_parent(tsdn,
 		    &background_thread_info[i].mtx);
 	}
@@ -743,7 +745,7 @@ background_thread_postfork_parent(tsdn_t *tsdn) {
 
 void
 background_thread_postfork_child(tsdn_t *tsdn) {
-	for (unsigned i = 0; i < ncpus; i++) {
+	for (unsigned i = 0; i < max_background_threads; i++) {
 		malloc_mutex_postfork_child(tsdn,
 		    &background_thread_info[i].mtx);
 	}
@@ -756,7 +758,7 @@ background_thread_postfork_child(tsdn_t *tsdn) {
 	malloc_mutex_lock(tsdn, &background_thread_lock);
 	n_background_threads = 0;
 	background_thread_enabled_set(tsdn, false);
-	for (unsigned i = 0; i < ncpus; i++) {
+	for (unsigned i = 0; i < max_background_threads; i++) {
 		background_thread_info_t *info = &background_thread_info[i];
 		malloc_mutex_lock(tsdn, &info->mtx);
 		info->state = background_thread_stopped;
@@ -780,7 +782,7 @@ background_thread_stats_read(tsdn_t *tsdn, background_thread_stats_t *stats) {
 	stats->num_threads = n_background_threads;
 	uint64_t num_runs = 0;
 	nstime_init(&stats->run_interval, 0);
-	for (unsigned i = 0; i < ncpus; i++) {
+	for (unsigned i = 0; i < max_background_threads; i++) {
 		background_thread_info_t *info = &background_thread_info[i];
 		malloc_mutex_lock(tsdn, &info->mtx);
 		if (info->state != background_thread_stopped) {
@@ -848,6 +850,12 @@ background_thread_boot1(tsdn_t *tsdn) {
 	assert(have_background_thread);
 	assert(narenas_total_get() > 0);
 
+	if (opt_max_background_threads == MAX_BACKGROUND_THREAD_LIMIT &&
+	    ncpus < MAX_BACKGROUND_THREAD_LIMIT) {
+		opt_max_background_threads = ncpus;
+	}
+	max_background_threads = opt_max_background_threads;
+
 	background_thread_enabled_set(tsdn, opt_background_thread);
 	if (malloc_mutex_init(&background_thread_lock,
 	    "background_thread_global",
@@ -857,12 +865,13 @@ background_thread_boot1(tsdn_t *tsdn) {
 	}
 
 	background_thread_info = (background_thread_info_t *)base_alloc(tsdn,
-	    b0get(), ncpus * sizeof(background_thread_info_t), CACHELINE);
+	    b0get(), opt_max_background_threads *
+	    sizeof(background_thread_info_t), CACHELINE);
 	if (background_thread_info == NULL) {
 		return true;
 	}
 
-	for (unsigned i = 0; i < ncpus; i++) {
+	for (unsigned i = 0; i < max_background_threads; i++) {
 		background_thread_info_t *info = &background_thread_info[i];
 		/* Thread mutex is rank_inclusive because of thread0. */
 		if (malloc_mutex_init(&info->mtx, "background_thread",

src/ctl.c

@@ -57,6 +57,7 @@ static const ctl_named_node_t	*n##_index(tsdn_t *tsdn,		\
 CTL_PROTO(version)
 CTL_PROTO(epoch)
 CTL_PROTO(background_thread)
+CTL_PROTO(max_background_threads)
 CTL_PROTO(thread_tcache_enabled)
 CTL_PROTO(thread_tcache_flush)
 CTL_PROTO(thread_prof_name)
@@ -85,6 +86,7 @@ CTL_PROTO(opt_dss)
 CTL_PROTO(opt_narenas)
 CTL_PROTO(opt_percpu_arena)
 CTL_PROTO(opt_background_thread)
+CTL_PROTO(opt_max_background_threads)
 CTL_PROTO(opt_dirty_decay_ms)
 CTL_PROTO(opt_muzzy_decay_ms)
 CTL_PROTO(opt_stats_print)
@@ -284,6 +286,7 @@ static const ctl_named_node_t opt_node[] = {
 	{NAME("narenas"),	CTL(opt_narenas)},
 	{NAME("percpu_arena"),	CTL(opt_percpu_arena)},
 	{NAME("background_thread"),	CTL(opt_background_thread)},
+	{NAME("max_background_threads"),	CTL(opt_max_background_threads)},
 	{NAME("dirty_decay_ms"), CTL(opt_dirty_decay_ms)},
 	{NAME("muzzy_decay_ms"), CTL(opt_muzzy_decay_ms)},
 	{NAME("stats_print"),	CTL(opt_stats_print)},
@@ -535,6 +538,7 @@ static const ctl_named_node_t	root_node[] = {
 	{NAME("version"),	CTL(version)},
 	{NAME("epoch"),		CTL(epoch)},
 	{NAME("background_thread"),	CTL(background_thread)},
+	{NAME("max_background_threads"),	CTL(max_background_threads)},
 	{NAME("thread"),	CHILD(named, thread)},
 	{NAME("config"),	CHILD(named, config)},
 	{NAME("opt"),		CHILD(named, opt)},
@@ -1564,6 +1568,71 @@ label_return:
 	return ret;
 }
 
+static int
+max_background_threads_ctl(tsd_t *tsd, const size_t *mib, size_t miblen,
+    void *oldp, size_t *oldlenp, void *newp, size_t newlen) {
+	int ret;
+	size_t oldval;
+
+	if (!have_background_thread) {
+		return ENOENT;
+	}
+	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 = max_background_threads;
+		READ(oldval, size_t);
+	} else {
+		if (newlen != sizeof(size_t)) {
+			ret = EINVAL;
+			goto label_return;
+		}
+		oldval = max_background_threads;
+		READ(oldval, size_t);
+
+		size_t newval = *(size_t *)newp;
+		if (newval == oldval) {
+			ret = 0;
+			goto label_return;
+		}
+		if (newval > opt_max_background_threads) {
+			ret = EINVAL;
+			goto label_return;
+		}
+
+		if (background_thread_enabled()) {
+			if (!can_enable_background_thread) {
+				malloc_printf("<jemalloc>: Error in dlsym("
+			            "RTLD_NEXT, \"pthread_create\"). Cannot "
+				    "enable background_thread\n");
+				ret = EFAULT;
+				goto label_return;
+			}
+			background_thread_enabled_set(tsd_tsdn(tsd), false);
+			if (background_threads_disable(tsd)) {
+				ret = EFAULT;
+				goto label_return;
+			}
+			max_background_threads = newval;
+			background_thread_enabled_set(tsd_tsdn(tsd), true);
+			if (background_threads_enable(tsd)) {
+				ret = EFAULT;
+				goto label_return;
+			}
+		} else {
+			max_background_threads = newval;
+		}
+	}
+	ret = 0;
+label_return:
+	malloc_mutex_unlock(tsd_tsdn(tsd), &background_thread_lock);
+	malloc_mutex_unlock(tsd_tsdn(tsd), &ctl_mtx);
+
+	return ret;
+}
+
 /******************************************************************************/
 
 CTL_RO_CONFIG_GEN(config_cache_oblivious, bool)
@@ -1590,6 +1659,7 @@ CTL_RO_NL_GEN(opt_narenas, opt_narenas, unsigned)
 CTL_RO_NL_GEN(opt_percpu_arena, percpu_arena_mode_names[opt_percpu_arena],
     const char *)
 CTL_RO_NL_GEN(opt_background_thread, opt_background_thread, bool)
+CTL_RO_NL_GEN(opt_max_background_threads, opt_max_background_threads, size_t)
 CTL_RO_NL_GEN(opt_dirty_decay_ms, opt_dirty_decay_ms, ssize_t)
 CTL_RO_NL_GEN(opt_muzzy_decay_ms, opt_muzzy_decay_ms, ssize_t)
 CTL_RO_NL_GEN(opt_stats_print, opt_stats_print, bool)

src/jemalloc.c

@@ -1175,6 +1175,10 @@ malloc_conf_init(void) {
 			}
 			CONF_HANDLE_BOOL(opt_background_thread,
 			    "background_thread");
+			CONF_HANDLE_SIZE_T(opt_max_background_threads,
+					   "max_background_threads", 1,
+					   opt_max_background_threads, yes, yes,
+					   true);
 			if (config_prof) {
 				CONF_HANDLE_BOOL(opt_prof, "prof")
 				CONF_HANDLE_CHAR_P(opt_prof_prefix,

test/unit/background_thread_enable.c

@@ -1,6 +1,6 @@
 #include "test/jemalloc_test.h"
 
-const char *malloc_conf = "background_thread:false,narenas:1";
+const char *malloc_conf = "background_thread:false,narenas:1,max_background_threads:20";
 
 TEST_BEGIN(test_deferred) {
 	test_skip_if(!have_background_thread);
@@ -30,8 +30,54 @@ TEST_BEGIN(test_deferred) {
 }
 TEST_END
 
+TEST_BEGIN(test_max_background_threads) {
+	test_skip_if(!have_background_thread);
+
+	size_t maxt;
+	size_t opt_maxt;
+	size_t sz_m = sizeof(maxt);
+	assert_d_eq(mallctl("opt.max_background_threads",
+			    &opt_maxt, &sz_m, NULL, 0), 0,
+			    "Failed to get opt.max_background_threads");
+	assert_d_eq(mallctl("max_background_threads", &maxt, &sz_m, NULL, 0), 0,
+		    "Failed to get max background threads");
+	assert_zu_eq(20, maxt, "should be ncpus");
+	assert_zu_eq(opt_maxt, maxt,
+		     "max_background_threads and "
+		     "opt.max_background_threads should match");
+	assert_d_eq(mallctl("max_background_threads", NULL, NULL, &maxt, sz_m),
+		    0, "Failed to set max background threads");
+
+	unsigned id;
+	size_t sz_u = sizeof(unsigned);
+
+	for (unsigned i = 0; i < 10 * ncpus; i++) {
+		assert_d_eq(mallctl("arenas.create", &id, &sz_u, NULL, 0), 0,
+		    "Failed to create arena");
+	}
+
+	bool enable = true;
+	size_t sz_b = sizeof(bool);
+	assert_d_eq(mallctl("background_thread", NULL, NULL, &enable, sz_b), 0,
+	    "Failed to enable background threads");
+	assert_zu_eq(n_background_threads, maxt,
+		     "Number of background threads should be 3.\n");
+	maxt = 10;
+	assert_d_eq(mallctl("max_background_threads", NULL, NULL, &maxt, sz_m),
+		    0, "Failed to set max background threads");
+	assert_zu_eq(n_background_threads, maxt,
+		     "Number of background threads should be 10.\n");
+	maxt = 3;
+	assert_d_eq(mallctl("max_background_threads", NULL, NULL, &maxt, sz_m),
+		    0, "Failed to set max background threads");
+	assert_zu_eq(n_background_threads, maxt,
+		     "Number of background threads should be 3.\n");
+}
+TEST_END
+
 int
 main(void) {
 	return test_no_reentrancy(
-	    test_deferred);
+		test_deferred,
+		test_max_background_threads);
 }