Index: sys/kern/kern_thread.c
===================================================================
--- sys/kern/kern_thread.c
+++ sys/kern/kern_thread.c
@@ -52,6 +52,7 @@
 #include <sys/syscallsubr.h>
 #include <sys/sysent.h>
 #include <sys/turnstile.h>
+#include <sys/kthread.h>
 #include <sys/ktr.h>
 #include <sys/rwlock.h>
 #include <sys/umtx.h>
@@ -128,7 +129,22 @@
  */
 static uma_zone_t thread_zone;
 
-static __exclusive_cache_line struct thread *thread_zombies;
+static void thread_reap_worker(void *arg);
+
+struct thread_domain_data {
+	struct thread	*tdd_zombies;
+} __aligned(CACHE_LINE_SIZE);
+
+static struct thread_domain_data thread_domain_data[MAXMEMDOM];
+
+static int
+thread_domain(struct thread *td)
+{
+
+	return (uma_item_domain(td));
+}
+
+static void thread_reap_all(void);
 
 static void thread_zombie(struct thread *);
 static int thread_unsuspend_one(struct thread *td, struct proc *p,
@@ -159,30 +175,45 @@
 EVENTHANDLER_LIST_DEFINE(thread_fini);
 
 static bool
-thread_count_inc(void)
+thread_count_inc_try(void)
 {
-	static struct timeval lastfail;
-	static int curfail;
 	int nthreads_new;
 
-	thread_reap();
-
 	nthreads_new = atomic_fetchadd_int(&nthreads, 1) + 1;
 	if (nthreads_new >= maxthread - 100) {
 		if (priv_check_cred(curthread->td_ucred, PRIV_MAXPROC) != 0 ||
 		    nthreads_new >= maxthread) {
 			atomic_subtract_int(&nthreads, 1);
-			if (ppsratecheck(&lastfail, &curfail, 1)) {
-				printf("maxthread limit exceeded by uid %u "
-				"(pid %d); consider increasing kern.maxthread\n",
-				curthread->td_ucred->cr_ruid, curproc->p_pid);
-			}
 			return (false);
 		}
 	}
 	return (true);
 }
 
+static bool
+thread_count_inc(void)
+{
+	static struct timeval lastfail;
+	static int curfail;
+
+	thread_reap();
+	if (thread_count_inc_try()) {
+		return (true);
+	}
+
+	thread_reap_all();
+	if (thread_count_inc_try()) {
+		return (true);
+	}
+
+	if (ppsratecheck(&lastfail, &curfail, 1)) {
+		printf("maxthread limit exceeded by uid %u "
+		    "(pid %d); consider increasing kern.maxthread\n",
+		    curthread->td_ucred->cr_ruid, curproc->p_pid);
+	}
+	return (false);
+}
+
 static void
 thread_count_sub(int n)
 {
@@ -452,18 +483,40 @@
 		rw_init(&tidhashtbl_lock[i], "tidhash");
 }
 
+/*
+ * Spawn per-domain reaper threads.
+ *
+ * threadinit is called before we can use kthread_add, hence additional routine.
+ */
+static void
+threadinit_reaper(void *arg __unused)
+{
+	int i, error;
+
+	for (i = 0; i < vm_ndomains; i++) {
+		error = kthread_add(thread_reap_worker,
+		    (void *)(uintptr_t)i, NULL, NULL, 0, 0, "tdreap%d", i);
+		if (error != 0)
+			panic("%s: could not start reaper thread %d: %d\n",
+			    __func__, i, error);
+	}
+}
+SYSINIT(threadreapper, SI_SUB_SMP, SI_ORDER_ANY, threadinit_reaper, NULL);
+
 /*
  * Place an unused thread on the zombie list.
  */
 void
 thread_zombie(struct thread *td)
 {
+	struct thread_domain_data *tdd;
 	struct thread *ztd;
 
-	ztd = atomic_load_ptr(&thread_zombies);
+	tdd = &thread_domain_data[thread_domain(td)];
+	ztd = atomic_load_ptr(&tdd->tdd_zombies);
 	for (;;) {
 		td->td_zombie = ztd;
-		if (atomic_fcmpset_rel_ptr((uintptr_t *)&thread_zombies,
+		if (atomic_fcmpset_rel_ptr((uintptr_t *)&tdd->tdd_zombies,
 		    (uintptr_t *)&ztd, (uintptr_t)td))
 			break;
 		continue;
@@ -481,10 +534,10 @@
 }
 
 /*
- * Reap zombie threads.
+ * Reap zombies from passed domain.
  */
-void
-thread_reap(void)
+static void
+thread_reap_domain(struct thread_domain_data *tdd)
 {
 	struct thread *itd, *ntd;
 	lwpid_t tidbatch[16];
@@ -494,11 +547,14 @@
 	 * Reading upfront is pessimal if followed by concurrent atomic_swap,
 	 * but most of the time the list is empty.
 	 */
-	if (thread_zombies == NULL)
+	if (tdd->tdd_zombies == NULL)
 		return;
 
-	itd = (struct thread *)atomic_swap_ptr((uintptr_t *)&thread_zombies,
+	itd = (struct thread *)atomic_swap_ptr((uintptr_t *)&tdd->tdd_zombies,
 	    (uintptr_t)NULL);
+	if (itd == NULL)
+		return;
+
 	tidbatchn = 0;
 	while (itd != NULL) {
 		ntd = itd->td_zombie;
@@ -520,6 +576,56 @@
 	}
 }
 
+/*
+ * Reap zombies from all domains.
+ */
+static void
+thread_reap_all(void)
+{
+	struct thread_domain_data *tdd;
+	int i, domain;
+
+	domain = PCPU_GET(domain);
+	for (i = 0; i < vm_ndomains; i++) {
+		tdd = &thread_domain_data[(i + domain) % vm_ndomains];
+		thread_reap_domain(tdd);
+	}
+}
+
+/*
+ * Reap zombies from local domain.
+ */
+void
+thread_reap(void)
+{
+	struct thread_domain_data *tdd;
+	int domain;
+
+	domain = PCPU_GET(domain);
+	tdd = &thread_domain_data[domain];
+
+	thread_reap_domain(tdd);
+}
+
+/*
+ * Per-domain workers.
+ */
+static void
+thread_reap_worker(void *arg)
+{
+	struct thread_domain_data *tdd;
+	int domain;
+
+	domain = (uintptr_t)arg;
+	tdd = &thread_domain_data[domain];
+
+	for (;;) {
+		kthread_suspend_check();
+		thread_reap_domain(tdd);
+		pause("threadreap", hz * 5);
+	}
+}
+
 /*
  * Allocate a thread.
  */
Index: sys/vm/uma_core.c
===================================================================
--- sys/vm/uma_core.c
+++ sys/vm/uma_core.c
@@ -3259,6 +3259,16 @@
 	    ("%s: unknown domain for item %p", __func__, item));
 	return (domain);
 }
+
+int
+uma_item_domain(void *item)
+{
+
+	/*
+	 * XXX assert passed object was allocated by UMA.
+	 */
+	return (item_domain(item));
+}
 #endif
 
 #if defined(INVARIANTS) || defined(DEBUG_MEMGUARD) || defined(WITNESS)