Merge branch 'wpk-multiple-task-queues' into 'master'

Multiple task queues See merge request isc-projects/bind9!878
2025-08-30 14:07:59 +00:00 · 2018-11-06 04:44:34 -05:00
parent 151b64aa3e b673f509d2
commit 319c166a50
17 changed files with 459 additions and 225 deletions
--- a/3
+++ b/3
@@ -1,3 +1,6 @@
+5081.	[func]		Use per-worker queues in task manager, make task
+			runners CPU-affine. [GL #659]
+
 5080.	[func]		Improvements to "rndc nta" user interface:
 			- catch and report invalid command line options
 			- when removing an NTA from all views, do not
--- a/bin/named/server.c
+++ b/bin/named/server.c
@@ -9327,8 +9327,7 @@ load_zones(named_server_t *server, bool init, bool reconfig) {
 		 * the initial server setup; it isn't necessary during
 		 * a reload.)
 		 */
-		isc_taskmgr_setmode(named_g_taskmgr,
-				    isc_taskmgrmode_privileged);
+		isc_taskmgr_setprivilegedmode(named_g_taskmgr);
 	}

 	isc_task_endexclusive(server->task);
--- a/bin/tests/system/fetchlimit/tests.sh
+++ b/bin/tests/system/fetchlimit/tests.sh
@@ -166,7 +166,7 @@ touch ans4/norespond
 for try in 1 2 3 4 5; do
    burst b $try 400
    $DIG @10.53.0.3 -p ${PORT}  a ${try}.example > dig.out.ns3.$try
-    stat 360 || exceeded=`expr $exceeded + 1`
+    stat 370 || exceeded=`expr $exceeded + 1`
    grep "status: NOERROR" dig.out.ns3.$try > /dev/null 2>&1 && \
            success=`expr $success + 1`
    grep "status: SERVFAIL" dig.out.ns3.$try > /dev/null 2>&1 && \
@@ -177,7 +177,7 @@ echo_i "$success successful valid queries (expected 5)"
 [ "$success" -eq 5 ] || { echo_i "failed"; ret=1; }
 echo_i "$fail SERVFAIL responses (expected 0)"
 [ "$fail" -eq 0 ] || { echo_i "failed"; ret=1; }
-echo_i "clients count exceeded 360 on $exceeded trials (expected 0)"
+echo_i "clients count exceeded 370 on $exceeded trials (expected 0)"
 [ "$exceeded" -eq 0 ] || { echo_i "failed"; ret=1; }
 if [ $ret != 0 ]; then echo_i "failed"; fi
 status=`expr $status + $ret`
--- a/config.h.in
+++ b/config.h.in
@@ -75,6 +75,9 @@
 /* Define to 1 if you have the <cmocka.h> header file. */
 #undef HAVE_CMOCKA_H

+/* Define to 1 if you have the `cpuset_setaffinty' function. */
+#undef HAVE_CPUSET_SETAFFINITY
+
 /* Define to 1 if you have the `CRYPTO_zalloc' function. */
 #undef HAVE_CRYPTO_ZALLOC

@@ -285,6 +288,9 @@
 /* define if OpenSSL supports Ed25519 */
 #undef HAVE_OPENSSL_ED25519

+/* Define to 1 if you have the `processor_bind' function. */
+#undef HAVE_PROCESSOR_BIND
+
 /* Define if you have POSIX threads libraries and header files. */
 #undef HAVE_PTHREAD

@@ -303,6 +309,9 @@
 /* Have PTHREAD_PRIO_INHERIT. */
 #undef HAVE_PTHREAD_PRIO_INHERIT

+/* Define to 1 if you have the `pthread_setaffinity_np' function. */
+#undef HAVE_PTHREAD_SETAFFINITY_NP
+
 /* Define to 1 if you have the `pthread_setname_np' function. */
 #undef HAVE_PTHREAD_SETNAME_NP

@@ -333,6 +342,9 @@
 /* Define to 1 if you have the <sched.h> header file. */
 #undef HAVE_SCHED_H

+/* Define to 1 if you have the `sched_setaffinity' function. */
+#undef HAVE_SCHED_SETAFFINITY
+
 /* Define to 1 if you have the `sched_yield' function. */
 #undef HAVE_SCHED_YIELD

@@ -387,6 +399,9 @@
 /* Define to 1 if you have the <sys/capability.h> header file. */
 #undef HAVE_SYS_CAPABILITY_H

+/* Define to 1 if you have the <sys/cpuset.h> header file. */
+#undef HAVE_SYS_CPUSET_H
+
 /* Define to 1 if you have the <sys/devpoll.h> header file. */
 #undef HAVE_SYS_DEVPOLL_H

@@ -396,6 +411,12 @@
 /* Define to 1 if you have the <sys/param.h> header file. */
 #undef HAVE_SYS_PARAM_H

+/* Define to 1 if you have the <sys/prctl.h> header file. */
+#undef HAVE_SYS_PRCTL_H
+
+/* Define to 1 if you have the <sys/procset.h> header file. */
+#undef HAVE_SYS_PROCSET_H
+
 /* Define to 1 if you have the <sys/select.h> header file. */
 #undef HAVE_SYS_SELECT_H

--- a/51
+++ b/51
@@ -844,6 +844,7 @@ infodir
 docdir
 oldincludedir
 includedir
+runstatedir
 localstatedir
 sharedstatedir
 sysconfdir
@@ -1003,6 +1004,7 @@ datadir='${datarootdir}'
 sysconfdir='${prefix}/etc'
 sharedstatedir='${prefix}/com'
 localstatedir='${prefix}/var'
+runstatedir='${localstatedir}/run'
 includedir='${prefix}/include'
 oldincludedir='/usr/include'
 docdir='${datarootdir}/doc/${PACKAGE_TARNAME}'
@@ -1255,6 +1257,15 @@ do
  | -silent | --silent | --silen | --sile | --sil)
    silent=yes ;;

+  -runstatedir | --runstatedir | --runstatedi | --runstated \
+  | --runstate | --runstat | --runsta | --runst | --runs \
+  | --run | --ru | --r)
+    ac_prev=runstatedir ;;
+  -runstatedir=* | --runstatedir=* | --runstatedi=* | --runstated=* \
+  | --runstate=* | --runstat=* | --runsta=* | --runst=* | --runs=* \
+  | --run=* | --ru=* | --r=*)
+    runstatedir=$ac_optarg ;;
+
  -sbindir | --sbindir | --sbindi | --sbind | --sbin | --sbi | --sb)
    ac_prev=sbindir ;;
  -sbindir=* | --sbindir=* | --sbindi=* | --sbind=* | --sbin=* \
@@ -1392,7 +1403,7 @@ fi
 for ac_var in	exec_prefix prefix bindir sbindir libexecdir datarootdir \
 		datadir sysconfdir sharedstatedir localstatedir includedir \
 		oldincludedir docdir infodir htmldir dvidir pdfdir psdir \
-		libdir localedir mandir
+		libdir localedir mandir runstatedir
 do
  eval ac_val=\$$ac_var
  # Remove trailing slashes.
@@ -1545,6 +1556,7 @@ Fine tuning of the installation directories:
  --sysconfdir=DIR        read-only single-machine data [PREFIX/etc]
  --sharedstatedir=DIR    modifiable architecture-independent data [PREFIX/com]
  --localstatedir=DIR     modifiable single-machine data [PREFIX/var]
+  --runstatedir=DIR       modifiable per-process data [LOCALSTATEDIR/run]
  --libdir=DIR            object code libraries [EPREFIX/lib]
  --includedir=DIR        C header files [PREFIX/include]
  --oldincludedir=DIR     C header files for non-gcc [/usr/include]
@@ -15235,6 +15247,43 @@ fi
 done


+for ac_header in sys/cpuset.h
+do :
+  ac_fn_c_check_header_mongrel "$LINENO" "sys/cpuset.h" "ac_cv_header_sys_cpuset_h" "$ac_includes_default"
+if test "x$ac_cv_header_sys_cpuset_h" = xyes; then :
+  cat >>confdefs.h <<_ACEOF
+#define HAVE_SYS_CPUSET_H 1
+_ACEOF
+
+fi
+
+done
+
+for ac_header in sys/procset.h
+do :
+  ac_fn_c_check_header_mongrel "$LINENO" "sys/procset.h" "ac_cv_header_sys_procset_h" "$ac_includes_default"
+if test "x$ac_cv_header_sys_procset_h" = xyes; then :
+  cat >>confdefs.h <<_ACEOF
+#define HAVE_SYS_PROCSET_H 1
+_ACEOF
+
+fi
+
+done
+
+for ac_func in pthread_setaffinity_np cpuset_setaffinity processor_bind sched_setaffinity
+do :
+  as_ac_var=`$as_echo "ac_cv_func_$ac_func" | $as_tr_sh`
+ac_fn_c_check_func "$LINENO" "$ac_func" "$as_ac_var"
+if eval test \"x\$"$as_ac_var"\" = x"yes"; then :
+  cat >>confdefs.h <<_ACEOF
+#define `$as_echo "HAVE_$ac_func" | $as_tr_cpp` 1
+_ACEOF
+
+fi
+done
+
+
 # Look for functions relating to thread naming
 for ac_func in pthread_setname_np pthread_set_name_np
 do :
--- a/configure.ac
+++ b/configure.ac
@@ -701,6 +701,10 @@ AC_CHECK_HEADERS([sched.h])
 AC_SEARCH_LIBS([sched_yield],[rt])
 AC_CHECK_FUNCS([sched_yield pthread_yield pthread_yield_np])

+AC_CHECK_HEADERS([sys/cpuset.h])
+AC_CHECK_HEADERS([sys/procset.h])
+AC_CHECK_FUNCS([pthread_setaffinity_np cpuset_setaffinity processor_bind sched_setaffinity])
+
 # Look for functions relating to thread naming
 AC_CHECK_FUNCS([pthread_setname_np pthread_set_name_np])
 AC_CHECK_HEADERS([pthread_np.h], [], [], [#include <pthread.h>])
--- a/lib/dns/client.c
+++ b/lib/dns/client.c
@@ -461,7 +461,7 @@ dns_client_createx(isc_mem_t *mctx, isc_appctx_t *actx,
 	client->timermgr = timermgr;

 	client->task = NULL;
-	result = isc_task_create(client->taskmgr, 0, &client->task);
+	result = isc_task_create(client->taskmgr, 50, &client->task);
 	if (result != ISC_R_SUCCESS)
 		goto cleanup;

--- a/lib/dns/dispatch.c
+++ b/lib/dns/dispatch.c
@@ -2503,7 +2503,7 @@ dns_dispatch_createtcp(dns_dispatchmgr_t *mgr, isc_socket_t *sock,

 	disp->ntasks = 1;
 	disp->task[0] = NULL;
-	result = isc_task_create(taskmgr, 0, &disp->task[0]);
+	result = isc_task_create(taskmgr, 50, &disp->task[0]);
 	if (result != ISC_R_SUCCESS)
 		goto kill_socket;

@@ -2941,7 +2941,7 @@ dispatch_createudp(dns_dispatchmgr_t *mgr, isc_socketmgr_t *sockmgr,
 		disp->ntasks = 1;
 	for (i = 0; i < disp->ntasks; i++) {
 		disp->task[i] = NULL;
-		result = isc_task_create(taskmgr, 0, &disp->task[i]);
+		result = isc_task_create(taskmgr, 50, &disp->task[i]);
 		if (result != ISC_R_SUCCESS) {
 			while (--i >= 0) {
 				isc_task_shutdown(disp->task[i]);
--- a/lib/isc/include/isc/task.h
+++ b/lib/isc/include/isc/task.h
@@ -208,8 +208,12 @@ isc_task_detach(isc_task_t **taskp);

 void
 isc_task_send(isc_task_t *task, isc_event_t **eventp);
+
+void
+isc_task_sendto(isc_task_t *task, isc_event_t **eventp, int c);
 /*%<
- * Send '*event' to 'task'.
+ * Send '*event' to 'task', if task is idle try starting it on cpu 'c'
+ * If 'c' is smaller than 0 then cpu is selected randomly.
 *
 * Requires:
 *
@@ -221,11 +225,15 @@ isc_task_send(isc_task_t *task, isc_event_t **eventp);
 *\li	*eventp == NULL.
 */

+void
+isc_task_sendtoanddetach(isc_task_t **taskp, isc_event_t **eventp, int c);
+
 void
 isc_task_sendanddetach(isc_task_t **taskp, isc_event_t **eventp);
 /*%<
 * Send '*event' to '*taskp' and then detach '*taskp' from its
- * task.
+ * task. If task is idle try starting it on cpu 'c'
+ * If 'c' is smaller than 0 then cpu is selected randomly.
 *
 * Requires:
 *
@@ -666,7 +674,7 @@ isc_taskmgr_create(isc_mem_t *mctx, unsigned int workers,
 */

 void
-isc_taskmgr_setmode(isc_taskmgr_t *manager, isc_taskmgrmode_t mode);
+isc_taskmgr_setprivilegedmode(isc_taskmgr_t *manager);

 isc_taskmgrmode_t
 isc_taskmgr_mode(isc_taskmgr_t *manager);
--- a/lib/isc/pthreads/include/isc/thread.h
+++ b/lib/isc/pthreads/include/isc/thread.h
@@ -44,6 +44,9 @@ isc_thread_yield(void);
 void
 isc_thread_setname(isc_thread_t thread, const char *name);

+isc_result_t
+isc_thread_setaffinity(int cpu);
+
 /* XXX We could do fancier error handling... */

 #define isc_thread_join(t, rp) \
--- a/lib/isc/pthreads/thread.c
+++ b/lib/isc/pthreads/thread.c
@@ -18,6 +18,17 @@
 #include <sched.h>
 #endif

+#if defined(HAVE_CPUSET_H)
+#include <sys/param.h>
+#include <sys/cpuset.h>
+#endif
+
+#if defined(HAVE_SYS_PROCESET_H)
+#include <sys/types.h>
+#include <sys/processor.h>
+#include <sys/procset.h>
+#endif
+
 #include <isc/thread.h>
 #include <isc/util.h>

@@ -91,3 +102,31 @@ isc_thread_yield(void) {
 	pthread_yield_np();
 #endif
 }
+
+isc_result_t
+isc_thread_setaffinity(int cpu) {
+#if defined(HAVE_CPUSET_SETAFFINITY)
+	cpuset_t cpuset;
+	CPU_ZERO(&cpuset);
+	CPU_SET(cpu, &cpuset);
+	if (cpuset_setaffinity(CPU_LEVEL_WHICH, CPU_WHICH_TID, -1,
+	    &cpuset, sizeof(cpuset)) != 0) {
+		return (ISC_R_FAILURE);
+	}
+#elif defined(HAVE_PTHREAD_SETAFFINITY_NP)
+	cpu_set_t set;
+	CPU_ZERO(&set);
+	CPU_SET(cpu, &set);
+	if (pthread_setaffinity_np(pthread_self(), sizeof(cpu_set_t),
+	    &set) != 0) {
+		return (ISC_R_FAILURE);
+	}
+#elif defined(HAVE_PROCESSOR_BIND)
+	if (processor_bind(P_LWPID, P_MYID, cpu, NULL) != 0) {
+		return (ISC_R_FAILURE);
+	}
+#else
+	UNUSED(cpu);
+#endif
+	return (ISC_R_SUCCESS);
+}
--- a/lib/isc/task.c
+++ b/lib/isc/task.c
@@ -21,6 +21,7 @@
 #include <stdbool.h>

 #include <isc/app.h>
+#include <isc/atomic.h>
 #include <isc/condition.h>
 #include <isc/event.h>
 #include <isc/json.h>
@@ -31,6 +32,7 @@
 #include <isc/platform.h>
 #include <isc/print.h>
 #include <isc/string.h>
+#include <isc/random.h>
 #include <isc/task.h>
 #include <isc/thread.h>
 #include <isc/time.h>
@@ -41,18 +43,14 @@
 #include <openssl/err.h>
 #endif

-/*%
- * For BIND9 internal applications:
- * when built with threads we use multiple worker threads shared by the whole
- * application.
- * when built without threads we share a single global task manager and use
- * an integrated event loop for socket, timer, and other generic task events.
- * For generic library:
- * we don't use either of them: an application can have multiple task managers
- * whether or not it's threaded, and if the application is threaded each thread
- * is expected to have a separate manager; no "worker threads" are shared by
- * the application threads.
+/*
+ * Task manager is built around 'as little locking as possible' concept.
+ * Each thread has his own queue of tasks to be run, if a task is in running
+ * state it will stay on the runner it's currently on, if a task is in idle
+ * state it can be woken up on a specific runner with isc_task_sendto - that
+ * helps with data locality on CPU.
 */
+
 #ifdef ISC_TASK_TRACE
 #define XTRACE(m)		fprintf(stderr, "task %p thread %lu: %s\n", \
 				       task, isc_thread_self(), (m))
@@ -86,6 +84,7 @@ static const char *statenames[] = {

 typedef struct isc__task isc__task_t;
 typedef struct isc__taskmgr isc__taskmgr_t;
+typedef struct isc__taskqueue isc__taskqueue_t;

 struct isc__task {
 	/* Not locked. */
@@ -104,6 +103,7 @@ struct isc__task {
 	isc_time_t			tnow;
 	char				name[16];
 	void *				tag;
+	unsigned int			threadid;
 	/* Locked by task manager lock. */
 	LINK(isc__task_t)		link;
 	LINK(isc__task_t)		ready_link;
@@ -121,28 +121,41 @@ struct isc__task {

 typedef ISC_LIST(isc__task_t)	isc__tasklist_t;

+struct isc__taskqueue {
+	/* Everything locked by lock */
+	isc_mutex_t			lock;
+	isc__tasklist_t			ready_tasks;
+	isc__tasklist_t			ready_priority_tasks;
+	isc_condition_t			work_available;
+	isc_thread_t			thread;
+	unsigned int			threadid;
+	isc__taskmgr_t			*manager;
+};
+
 struct isc__taskmgr {
 	/* Not locked. */
 	isc_taskmgr_t			common;
 	isc_mem_t *			mctx;
 	isc_mutex_t			lock;
+	isc_mutex_t			halt_lock;
+	isc_condition_t			halt_cond;
 	unsigned int			workers;
-	isc_thread_t *			threads;
+	atomic_uint_fast32_t		tasks_running;
+	atomic_uint_fast32_t		tasks_ready;
+	atomic_uint_fast32_t		curq;
+	isc__taskqueue_t		*queues;
+
 	/* Locked by task manager lock. */
 	unsigned int			default_quantum;
 	LIST(isc__task_t)		tasks;
-	isc__tasklist_t			ready_tasks;
-	isc__tasklist_t			ready_priority_tasks;
 	isc_taskmgrmode_t		mode;
-	isc_condition_t			work_available;
-	isc_condition_t			exclusive_granted;
-	isc_condition_t			paused;
-	unsigned int			tasks_running;
-	unsigned int			tasks_ready;
 	bool				pause_requested;
 	bool				exclusive_requested;
 	bool				exiting;

+	/* Locked by halt_lock */
+	unsigned int			halted;
+
 	/*
 	 * Multiple threads can read/write 'excl' at the same time, so we need
 	 * to protect the access.  We can't use 'lock' since isc_task_detach()
@@ -175,22 +188,33 @@ isc_taskmgr_setexcltask(isc_taskmgr_t *mgr0, isc_task_t *task0);
 isc_result_t
 isc_taskmgr_excltask(isc_taskmgr_t *mgr0, isc_task_t **taskp);
 static inline bool
-empty_readyq(isc__taskmgr_t *manager);
+empty_readyq(isc__taskmgr_t *manager, int c);

 static inline isc__task_t *
-pop_readyq(isc__taskmgr_t *manager);
+pop_readyq(isc__taskmgr_t *manager, int c);

 static inline void
-push_readyq(isc__taskmgr_t *manager, isc__task_t *task);
+push_readyq(isc__taskmgr_t *manager, isc__task_t *task, int c);
+
+static inline void
+wake_all_queues(isc__taskmgr_t *manager);

 /***
 *** Tasks.
 ***/

+static inline void
+wake_all_queues(isc__taskmgr_t *manager) {
+	for (unsigned i=0; i < manager->workers; i++) {
+		LOCK(&manager->queues[i].lock);
+		BROADCAST(&manager->queues[i].work_available);
+		UNLOCK(&manager->queues[i].lock);
+	}
+}
+
 static void
 task_finished(isc__task_t *task) {
 	isc__taskmgr_t *manager = task->manager;
-
 	REQUIRE(EMPTY(task->events));
 	REQUIRE(task->nevents == 0);
 	REQUIRE(EMPTY(task->on_shutdown));
@@ -201,6 +225,7 @@ task_finished(isc__task_t *task) {

 	LOCK(&manager->lock);
 	UNLINK(manager->tasks, task, link);
+	UNLOCK(&manager->lock);
 	if (FINISHED(manager)) {
 		/*
 		 * All tasks have completed and the
@@ -208,10 +233,8 @@ task_finished(isc__task_t *task) {
 		 * any idle worker threads so they
 		 * can exit.
 		 */
-		BROADCAST(&manager->work_available);
+		wake_all_queues(manager);
 	}
-	UNLOCK(&manager->lock);
-
 	DESTROYLOCK(&task->lock);
 	task->common.impmagic = 0;
 	task->common.magic = 0;
@@ -235,6 +258,9 @@ isc_task_create(isc_taskmgr_t *manager0, unsigned int quantum,
 		return (ISC_R_NOMEMORY);
 	XTRACE("isc_task_create");
 	task->manager = manager;
+	task->threadid = atomic_fetch_add_explicit(&manager->curq, 1,
+						   memory_order_relaxed)
+						   % manager->workers;
 	result = isc_mutex_init(&task->lock);
 	if (result != ISC_R_SUCCESS) {
 		isc_mem_put(manager->mctx, task, sizeof(*task));
@@ -351,12 +377,11 @@ task_ready(isc__task_t *task) {
 	REQUIRE(task->state == task_state_ready);

 	XTRACE("task_ready");
-
-	LOCK(&manager->lock);
-	push_readyq(manager, task);
+	LOCK(&manager->queues[task->threadid].lock);
+	push_readyq(manager, task, task->threadid);
 	if (manager->mode == isc_taskmgrmode_normal || has_privilege)
-		SIGNAL(&manager->work_available);
-	UNLOCK(&manager->lock);
+		SIGNAL(&manager->queues[task->threadid].work_available);
+	UNLOCK(&manager->queues[task->threadid].lock);
 }

 static inline bool
@@ -414,7 +439,7 @@ isc_task_detach(isc_task_t **taskp) {
 }

 static inline bool
-task_send(isc__task_t *task, isc_event_t **eventp) {
+task_send(isc__task_t *task, isc_event_t **eventp, int c) {
 	bool was_idle = false;
 	isc_event_t *event;

@@ -433,6 +458,7 @@ task_send(isc__task_t *task, isc_event_t **eventp) {

 	if (task->state == task_state_idle) {
 		was_idle = true;
+		task->threadid = c;
 		INSIST(EMPTY(task->events));
 		task->state = task_state_ready;
 	}
@@ -447,6 +473,16 @@ task_send(isc__task_t *task, isc_event_t **eventp) {

 void
 isc_task_send(isc_task_t *task0, isc_event_t **eventp) {
+	isc_task_sendto(task0, eventp, -1);
+}
+
+void
+isc_task_sendanddetach(isc_task_t **taskp, isc_event_t **eventp) {
+	isc_task_sendtoanddetach(taskp, eventp, -1);
+}
+
+void
+isc_task_sendto(isc_task_t *task0, isc_event_t **eventp, int c) {
 	isc__task_t *task = (isc__task_t *)task0;
 	bool was_idle;

@@ -455,16 +491,21 @@ isc_task_send(isc_task_t *task0, isc_event_t **eventp) {
 	 */

 	REQUIRE(VALID_TASK(task));
-
 	XTRACE("isc_task_send");

+	if (c < 0) {
+		c = atomic_fetch_add_explicit(&task->manager->curq, 1,
+					      memory_order_relaxed);
+	}
+	c %= task->manager->workers;
+
 	/*
 	 * We're trying hard to hold locks for as short a time as possible.
 	 * We're also trying to hold as few locks as possible.  This is why
 	 * some processing is deferred until after the lock is released.
 	 */
 	LOCK(&task->lock);
-	was_idle = task_send(task, eventp);
+	was_idle = task_send(task, eventp, c);
 	UNLOCK(&task->lock);

 	if (was_idle) {
@@ -488,7 +529,7 @@ isc_task_send(isc_task_t *task0, isc_event_t **eventp) {
 }

 void
-isc_task_sendanddetach(isc_task_t **taskp, isc_event_t **eventp) {
+isc_task_sendtoanddetach(isc_task_t **taskp, isc_event_t **eventp, int c) {
 	bool idle1, idle2;
 	isc__task_t *task;

@@ -500,11 +541,16 @@ isc_task_sendanddetach(isc_task_t **taskp, isc_event_t **eventp) {
 	REQUIRE(taskp != NULL);
 	task = (isc__task_t *)*taskp;
 	REQUIRE(VALID_TASK(task));
-
 	XTRACE("isc_task_sendanddetach");

+	if (c < 0) {
+		c = atomic_fetch_add_explicit(&task->manager->curq, 1,
+					      memory_order_relaxed);
+	}
+	c %= task->manager->workers;
+
 	LOCK(&task->lock);
-	idle1 = task_send(task, eventp);
+	idle1 = task_send(task, eventp, c);
 	idle2 = task_detach(task);
 	UNLOCK(&task->lock);

@@ -829,13 +875,13 @@ isc_task_getcurrenttimex(isc_task_t *task0, isc_time_t *t) {
 * Caller must hold the task manager lock.
 */
 static inline bool
-empty_readyq(isc__taskmgr_t *manager) {
+empty_readyq(isc__taskmgr_t *manager, int c) {
 	isc__tasklist_t queue;

 	if (manager->mode == isc_taskmgrmode_normal)
-		queue = manager->ready_tasks;
+		queue = manager->queues[c].ready_tasks;
 	else
-		queue = manager->ready_priority_tasks;
+		queue = manager->queues[c].ready_priority_tasks;

 	return (EMPTY(queue));
 }
@@ -849,18 +895,18 @@ empty_readyq(isc__taskmgr_t *manager) {
 * Caller must hold the task manager lock.
 */
 static inline isc__task_t *
-pop_readyq(isc__taskmgr_t *manager) {
+pop_readyq(isc__taskmgr_t *manager, int c) {
 	isc__task_t *task;

 	if (manager->mode == isc_taskmgrmode_normal)
-		task = HEAD(manager->ready_tasks);
+		task = HEAD(manager->queues[c].ready_tasks);
 	else
-		task = HEAD(manager->ready_priority_tasks);
+		task = HEAD(manager->queues[c].ready_priority_tasks);

 	if (task != NULL) {
-		DEQUEUE(manager->ready_tasks, task, ready_link);
+		DEQUEUE(manager->queues[c].ready_tasks, task, ready_link);
 		if (ISC_LINK_LINKED(task, ready_priority_link))
-			DEQUEUE(manager->ready_priority_tasks, task,
+			DEQUEUE(manager->queues[c].ready_priority_tasks, task,
 				ready_priority_link);
 	}

@@ -874,20 +920,25 @@ pop_readyq(isc__taskmgr_t *manager) {
 * Caller must hold the task manager lock.
 */
 static inline void
-push_readyq(isc__taskmgr_t *manager, isc__task_t *task) {
-	ENQUEUE(manager->ready_tasks, task, ready_link);
+push_readyq(isc__taskmgr_t *manager, isc__task_t *task, int c) {
+	ENQUEUE(manager->queues[c].ready_tasks, task, ready_link);
 	if ((task->flags & TASK_F_PRIVILEGED) != 0)
-		ENQUEUE(manager->ready_priority_tasks, task,
+		ENQUEUE(manager->queues[c].ready_priority_tasks, task,
 			ready_priority_link);
-	manager->tasks_ready++;
+	atomic_fetch_add_explicit(&manager->tasks_ready, 1,
+				  memory_order_acquire);
 }

 static void
-dispatch(isc__taskmgr_t *manager) {
+dispatch(isc__taskmgr_t *manager, unsigned int threadid) {
 	isc__task_t *task;

 	REQUIRE(VALID_MANAGER(manager));

+	/* Wait for everything to initialize */
+	LOCK(&manager->lock);
+	UNLOCK(&manager->lock);
+
 	/*
 	 * Again we're trying to hold the lock for as short a time as possible
 	 * and to do as little locking and unlocking as possible.
@@ -937,8 +988,7 @@ dispatch(isc__taskmgr_t *manager) {
 	 * For N iterations of the loop, this code does N+1 locks and N+1
 	 * unlocks.  The while expression is always protected by the lock.
 	 */
-
-	LOCK(&manager->lock);
+	LOCK(&manager->queues[threadid].lock);

 	while (!FINISHED(manager)) {
 		/*
@@ -951,13 +1001,19 @@ dispatch(isc__taskmgr_t *manager) {
 		 * If a pause has been requested, don't do any work
 		 * until it's been released.
 		 */
-		while ((empty_readyq(manager) || manager->pause_requested ||
-			manager->exclusive_requested) && !FINISHED(manager))
+		while ((empty_readyq(manager, threadid) && !manager->pause_requested &&
+			!manager->exclusive_requested) && !FINISHED(manager))
 		{
 			XTHREADTRACE(isc_msgcat_get(isc_msgcat,
 						    ISC_MSGSET_GENERAL,
 						    ISC_MSG_WAIT, "wait"));
-			WAIT(&manager->work_available, &manager->lock);
+			XTHREADTRACE(isc_msgcat_get(isc_msgcat,
+						    ISC_MSGSET_GENERAL,
+						    ISC_MSG_WAIT, manager->pause_requested ? "paused" : "notpaused"));
+			XTHREADTRACE(isc_msgcat_get(isc_msgcat,
+						    ISC_MSGSET_GENERAL,
+						    ISC_MSG_WAIT, manager->exclusive_requested ? "excreq" : "notexcreq"));
+			WAIT(&manager->queues[threadid].work_available, &manager->queues[threadid].lock);
 			XTHREADTRACE(isc_msgcat_get(isc_msgcat,
 						    ISC_MSGSET_TASK,
 						    ISC_MSG_AWAKE, "awake"));
@@ -965,7 +1021,37 @@ dispatch(isc__taskmgr_t *manager) {
 		XTHREADTRACE(isc_msgcat_get(isc_msgcat, ISC_MSGSET_TASK,
 					    ISC_MSG_WORKING, "working"));

-		task = pop_readyq(manager);
+		if (manager->pause_requested || manager->exclusive_requested) {
+			UNLOCK(&manager->queues[threadid].lock);
+			XTHREADTRACE(isc_msgcat_get(isc_msgcat, ISC_MSGSET_TASK,
+					    ISC_MSG_WORKING, "halting"));
+
+			/*
+			 * Switching to exclusive mode is done as a 2-phase-lock,
+			 * checking if we have to switch is done without any locks
+			 * on pause_requested and exclusive_requested to save time -
+			 * the worst thing that can happen is that we'll launch one task
+			 * more and exclusive task will be postponed a bit.
+			 *
+			 * Broadcasting on halt_cond seems suboptimal, but exclusive tasks
+			 * are rare enought that we don't care.
+			 */
+			LOCK(&manager->halt_lock);
+			manager->halted++;
+			BROADCAST(&manager->halt_cond);
+			while (manager->pause_requested || manager->exclusive_requested) {
+				WAIT(&manager->halt_cond, &manager->halt_lock);
+			}
+			manager->halted--;
+			SIGNAL(&manager->halt_cond);
+			UNLOCK(&manager->halt_lock);
+
+			LOCK(&manager->queues[threadid].lock);
+			/* Restart the loop after */
+			continue;
+		}
+
+		task = pop_readyq(manager, threadid);
 		if (task != NULL) {
 			unsigned int dispatch_count = 0;
 			bool done = false;
@@ -976,19 +1062,22 @@ dispatch(isc__taskmgr_t *manager) {
 			INSIST(VALID_TASK(task));

 			/*
-			 * Note we only unlock the manager lock if we actually
-			 * have a task to do.  We must reacquire the manager
+			 * Note we only unlock the queue lock if we actually
+			 * have a task to do.  We must reacquire the queue
 			 * lock before exiting the 'if (task != NULL)' block.
 			 */
-			manager->tasks_ready--;
-			manager->tasks_running++;
-			UNLOCK(&manager->lock);
+			UNLOCK(&manager->queues[threadid].lock);
+			RUNTIME_CHECK(atomic_fetch_sub_explicit(&manager->tasks_ready,
+						  1, memory_order_release) > 0);
+			atomic_fetch_add_explicit(&manager->tasks_running, 1,
+						  memory_order_acquire);

 			LOCK(&task->lock);
 			INSIST(task->state == task_state_ready);
 			task->state = task_state_running;
 			XTRACE(isc_msgcat_get(isc_msgcat, ISC_MSGSET_GENERAL,
 					      ISC_MSG_RUNNING, "running"));
+			XTRACE(task->name);
 			TIME_NOW(&task->tnow);
 			task->now = isc_time_seconds(&task->tnow);
 			do {
@@ -1004,6 +1093,7 @@ dispatch(isc__taskmgr_t *manager) {
 							    ISC_MSGSET_TASK,
 							    ISC_MSG_EXECUTE,
 							    "execute action"));
+					XTRACE(task->name);
 					if (event->ev_action != NULL) {
 						UNLOCK(&task->lock);
 						(event->ev_action)(
@@ -1094,15 +1184,9 @@ dispatch(isc__taskmgr_t *manager) {
 			if (finished)
 				task_finished(task);

-			LOCK(&manager->lock);
-			manager->tasks_running--;
-			if (manager->exclusive_requested &&
-			    manager->tasks_running == 1) {
-				SIGNAL(&manager->exclusive_granted);
-			} else if (manager->pause_requested &&
-				   manager->tasks_running == 0) {
-				SIGNAL(&manager->paused);
-			}
+			RUNTIME_CHECK(atomic_fetch_sub_explicit(&manager->tasks_running,
+						1, memory_order_release) > 0);
+			LOCK(&manager->queues[threadid].lock);
 			if (requeue) {
 				/*
 				 * We know we're awake, so we don't have
@@ -1123,7 +1207,7 @@ dispatch(isc__taskmgr_t *manager) {
 				 * were usually nonempty, the 'optimization'
 				 * might even hurt rather than help.
 				 */
-				push_readyq(manager, task);
+				push_readyq(manager, task, threadid);
 			}
 		}

@@ -1133,27 +1217,62 @@ dispatch(isc__taskmgr_t *manager) {
 		 * we're stuck.  Automatically drop privileges at that
 		 * point and continue with the regular ready queue.
 		 */
-		if (manager->tasks_running == 0 && empty_readyq(manager)) {
+		if (manager->mode != isc_taskmgrmode_normal &&
+		    atomic_load_explicit(&manager->tasks_running,
+					 memory_order_acquire) == 0)
+		{
+			UNLOCK(&manager->queues[threadid].lock);
+			LOCK(&manager->lock);
+			/*
+			 * Check once again, under lock. Mode can only
+			 * change from privileged to normal anyway, and
+			 * if we enter this loop twice at the same time
+			 * we'll end up in a deadlock over queue locks.
+			 *
+			 */
+			if (manager->mode != isc_taskmgrmode_normal &&
+			    atomic_load_explicit(&manager->tasks_running,
+						 memory_order_acquire) == 0)
+			{
+				bool empty = true;
+				unsigned int i;
+				for (i=0; i<manager->workers && empty; i++)
+				{
+					LOCK(&manager->queues[i].lock);
+					empty &= empty_readyq(manager, i);
+					UNLOCK(&manager->queues[i].lock);
+				}
+				if (empty) {
 					manager->mode = isc_taskmgrmode_normal;
-			if (!empty_readyq(manager))
-				BROADCAST(&manager->work_available);
+					wake_all_queues(manager);
 				}
 			}
-
 			UNLOCK(&manager->lock);
+			LOCK(&manager->queues[threadid].lock);
+		}
+	}
+	UNLOCK(&manager->queues[threadid].lock);
+	/*
+	 * There might be other dispatchers waiting on empty tasks,
+	 * wake them up.
+	 */
+	wake_all_queues(manager);
 }

 static isc_threadresult_t
 #ifdef _WIN32
 WINAPI
 #endif
-run(void *uap) {
-	isc__taskmgr_t *manager = uap;
+run(void *queuep) {
+	isc__taskqueue_t *tq = queuep;
+	isc__taskmgr_t *manager = tq->manager;
+	int threadid = tq->threadid;
+	isc_thread_setaffinity(threadid);

 	XTHREADTRACE(isc_msgcat_get(isc_msgcat, ISC_MSGSET_GENERAL,
 				    ISC_MSG_STARTING, "starting"));

-	dispatch(manager);
+	dispatch(manager, threadid);

 	XTHREADTRACE(isc_msgcat_get(isc_msgcat, ISC_MSGSET_GENERAL,
 				    ISC_MSG_EXITING, "exiting"));
@@ -1167,27 +1286,23 @@ run(void *uap) {

 static void
 manager_free(isc__taskmgr_t *manager) {
-	isc_mem_t *mctx;
-
-	(void)isc_condition_destroy(&manager->exclusive_granted);
-	(void)isc_condition_destroy(&manager->work_available);
-	(void)isc_condition_destroy(&manager->paused);
-	isc_mem_free(manager->mctx, manager->threads);
+	for (unsigned int i=0; i < manager->workers; i++) {
+		DESTROYLOCK(&manager->queues[i].lock);
+	}
 	DESTROYLOCK(&manager->lock);
-	DESTROYLOCK(&manager->excl_lock);
+	DESTROYLOCK(&manager->halt_lock);
+	isc_mem_put(manager->mctx, manager->queues,
+		    manager->workers * sizeof(isc__taskqueue_t));
 	manager->common.impmagic = 0;
 	manager->common.magic = 0;
-	mctx = manager->mctx;
-	isc_mem_put(mctx, manager, sizeof(*manager));
-	isc_mem_detach(&mctx);
+	isc_mem_putanddetach(&manager->mctx, manager, sizeof(*manager));
 }

 isc_result_t
 isc_taskmgr_create(isc_mem_t *mctx, unsigned int workers,
 		    unsigned int default_quantum, isc_taskmgr_t **managerp)
 {
-	isc_result_t result;
-	unsigned int i, started = 0;
+	unsigned int i;
 	isc__taskmgr_t *manager;

 	/*
@@ -1198,64 +1313,37 @@ isc_taskmgr_create(isc_mem_t *mctx, unsigned int workers,
 	REQUIRE(managerp != NULL && *managerp == NULL);

 	manager = isc_mem_get(mctx, sizeof(*manager));
-	if (manager == NULL)
-		return (ISC_R_NOMEMORY);
+	RUNTIME_CHECK(manager != NULL);
 	manager->common.impmagic = TASK_MANAGER_MAGIC;
 	manager->common.magic = ISCAPI_TASKMGR_MAGIC;
 	manager->mode = isc_taskmgrmode_normal;
 	manager->mctx = NULL;
-	result = isc_mutex_init(&manager->lock);
-	if (result != ISC_R_SUCCESS)
-		goto cleanup_mgr;
-	result = isc_mutex_init(&manager->excl_lock);
-	if (result != ISC_R_SUCCESS) {
-		DESTROYLOCK(&manager->lock);
-		goto cleanup_mgr;
-	}
+	RUNTIME_CHECK(isc_mutex_init(&manager->lock) == ISC_R_SUCCESS);
+	RUNTIME_CHECK(isc_mutex_init(&manager->excl_lock) == ISC_R_SUCCESS);
+
+	RUNTIME_CHECK(isc_mutex_init(&manager->halt_lock)
+		      == ISC_R_SUCCESS);
+	RUNTIME_CHECK(isc_condition_init(&manager->halt_cond)
+		      == ISC_R_SUCCESS);
+
+	manager->workers = workers;

-	manager->workers = 0;
-	manager->threads = isc_mem_allocate(mctx,
-					    workers * sizeof(isc_thread_t));
-	if (manager->threads == NULL) {
-		result = ISC_R_NOMEMORY;
-		goto cleanup_lock;
-	}
-	if (isc_condition_init(&manager->work_available) != ISC_R_SUCCESS) {
-		UNEXPECTED_ERROR(__FILE__, __LINE__,
-				 "isc_condition_init() %s",
-				 isc_msgcat_get(isc_msgcat, ISC_MSGSET_GENERAL,
-						ISC_MSG_FAILED, "failed"));
-		result = ISC_R_UNEXPECTED;
-		goto cleanup_threads;
-	}
-	if (isc_condition_init(&manager->exclusive_granted) != ISC_R_SUCCESS) {
-		UNEXPECTED_ERROR(__FILE__, __LINE__,
-				 "isc_condition_init() %s",
-				 isc_msgcat_get(isc_msgcat, ISC_MSGSET_GENERAL,
-						ISC_MSG_FAILED, "failed"));
-		result = ISC_R_UNEXPECTED;
-		goto cleanup_workavailable;
-	}
-	if (isc_condition_init(&manager->paused) != ISC_R_SUCCESS) {
-		UNEXPECTED_ERROR(__FILE__, __LINE__,
-				 "isc_condition_init() %s",
-				 isc_msgcat_get(isc_msgcat, ISC_MSGSET_GENERAL,
-						ISC_MSG_FAILED, "failed"));
-		result = ISC_R_UNEXPECTED;
-		goto cleanup_exclusivegranted;
-	}
 	if (default_quantum == 0)
 		default_quantum = DEFAULT_DEFAULT_QUANTUM;
 	manager->default_quantum = default_quantum;
 	INIT_LIST(manager->tasks);
-	INIT_LIST(manager->ready_tasks);
-	INIT_LIST(manager->ready_priority_tasks);
+	manager->queues = isc_mem_get(mctx, workers *
+				      sizeof(isc__taskqueue_t));
+	RUNTIME_CHECK(manager->queues != NULL);
+
 	manager->tasks_running = 0;
 	manager->tasks_ready = 0;
-	manager->exclusive_requested = false;
-	manager->pause_requested = false;
+	manager->curq = 0;
 	manager->exiting = false;
 	manager->excl = NULL;
+	manager->halted = 0;
+	manager->exclusive_requested = false;
+	manager->pause_requested = false;

 	isc_mem_attach(mctx, &manager->mctx);

@@ -1264,40 +1352,29 @@ isc_taskmgr_create(isc_mem_t *mctx, unsigned int workers,
 	 * Start workers.
 	 */
 	for (i = 0; i < workers; i++) {
-		if (isc_thread_create(run, manager,
-				      &manager->threads[manager->workers]) ==
-		    ISC_R_SUCCESS) {
-			char name[16];	/* thread name limit on Linux */
+		INIT_LIST(manager->queues[i].ready_tasks);
+		INIT_LIST(manager->queues[i].ready_priority_tasks);
+		RUNTIME_CHECK(isc_mutex_init(&manager->queues[i].lock)
+			      == ISC_R_SUCCESS);
+		RUNTIME_CHECK(isc_condition_init(
+					   &manager->queues[i].work_available)
+			      == ISC_R_SUCCESS);
+		manager->queues[i].manager = manager;
+		manager->queues[i].threadid = i;
+		RUNTIME_CHECK(isc_thread_create(run, &manager->queues[i],
+						&manager->queues[i].thread)
+			      == ISC_R_SUCCESS);
+		char name[16];
 		snprintf(name, sizeof(name), "isc-worker%04u", i);
-			isc_thread_setname(manager->threads[manager->workers],
-					   name);
-			manager->workers++;
-			started++;
-		}
+		isc_thread_setname(manager->queues[i].thread, name);
 	}
 	UNLOCK(&manager->lock);

-	if (started == 0) {
-		manager_free(manager);
-		return (ISC_R_NOTHREADS);
-	}
 	isc_thread_setconcurrency(workers);

 	*managerp = (isc_taskmgr_t *)manager;

 	return (ISC_R_SUCCESS);
-
- cleanup_exclusivegranted:
-	(void)isc_condition_destroy(&manager->exclusive_granted);
- cleanup_workavailable:
-	(void)isc_condition_destroy(&manager->work_available);
- cleanup_threads:
-	isc_mem_free(mctx, manager->threads);
- cleanup_lock:
-	DESTROYLOCK(&manager->lock);
- cleanup_mgr:
-	isc_mem_put(mctx, manager, sizeof(*manager));
-	return (result);
 }

 void
@@ -1355,29 +1432,34 @@ isc_taskmgr_destroy(isc_taskmgr_t **managerp) {

 	/*
 	 * Post shutdown event(s) to every task (if they haven't already been
-	 * posted).
+	 * posted). To make things easier post idle tasks to worker 0.
 	 */
+	LOCK(&manager->queues[0].lock);
 	for (task = HEAD(manager->tasks);
 	     task != NULL;
 	     task = NEXT(task, link)) {
 		LOCK(&task->lock);
-		if (task_shutdown(task))
-			push_readyq(manager, task);
+		if (task_shutdown(task)) {
+			task->threadid = 0;
+			push_readyq(manager, task, 0);
+		}
 		UNLOCK(&task->lock);
 	}
+	UNLOCK(&manager->queues[0].lock);
+
 	/*
 	 * Wake up any sleeping workers.  This ensures we get work done if
 	 * there's work left to do, and if there are already no tasks left
 	 * it will cause the workers to see manager->exiting.
 	 */
-	BROADCAST(&manager->work_available);
+	wake_all_queues(manager);
 	UNLOCK(&manager->lock);

 	/*
 	 * Wait for all the worker threads to exit.
 	 */
 	for (i = 0; i < manager->workers; i++)
-		(void)isc_thread_join(manager->threads[i], NULL);
+		(void)isc_thread_join(manager->queues[i].thread, NULL);

 	manager_free(manager);

@@ -1385,11 +1467,11 @@ isc_taskmgr_destroy(isc_taskmgr_t **managerp) {
 }

 void
-isc_taskmgr_setmode(isc_taskmgr_t *manager0, isc_taskmgrmode_t mode) {
+isc_taskmgr_setprivilegedmode(isc_taskmgr_t *manager0) {
 	isc__taskmgr_t *manager = (isc__taskmgr_t *)manager0;

 	LOCK(&manager->lock);
-	manager->mode = mode;
+	manager->mode = isc_taskmgrmode_privileged;
 	UNLOCK(&manager->lock);
 }

@@ -1406,24 +1488,35 @@ isc_taskmgr_mode(isc_taskmgr_t *manager0) {
 void
 isc__taskmgr_pause(isc_taskmgr_t *manager0) {
 	isc__taskmgr_t *manager = (isc__taskmgr_t *)manager0;
-	manager->pause_requested = true;
-	LOCK(&manager->lock);
-	while (manager->tasks_running > 0) {
-		WAIT(&manager->paused, &manager->lock);
+
+	LOCK(&manager->halt_lock);
+	while (manager->exclusive_requested || manager->pause_requested) {
+		UNLOCK(&manager->halt_lock);
+		/* This is ugly but pause is used EXCLUSIVELY in tests */
+		isc_thread_yield();
+		LOCK(&manager->halt_lock);
 	}
-	UNLOCK(&manager->lock);
+
+	manager->pause_requested = true;
+	while (manager->halted < manager->workers) {
+		wake_all_queues(manager);
+		WAIT(&manager->halt_cond, &manager->halt_lock);
+	}
+	UNLOCK(&manager->halt_lock);
 }

 void
 isc__taskmgr_resume(isc_taskmgr_t *manager0) {
 	isc__taskmgr_t *manager = (isc__taskmgr_t *)manager0;
-
-	LOCK(&manager->lock);
+	LOCK(&manager->halt_lock);
 	if (manager->pause_requested) {
 		manager->pause_requested = false;
-		BROADCAST(&manager->work_available);
+		while (manager->halted > 0) {
+			BROADCAST(&manager->halt_cond);
+			WAIT(&manager->halt_cond, &manager->halt_lock);
 		}
-	UNLOCK(&manager->lock);
+	}
+	UNLOCK(&manager->halt_lock);
 }

 void
@@ -1462,24 +1555,25 @@ isc_result_t
 isc_task_beginexclusive(isc_task_t *task0) {
 	isc__task_t *task = (isc__task_t *)task0;
 	isc__taskmgr_t *manager = task->manager;
+
 	REQUIRE(VALID_TASK(task));

 	REQUIRE(task->state == task_state_running);
-/*
- *  TODO REQUIRE(task == task->manager->excl);
- *  it should be here, it fails on shutdown server->task
- */
+	REQUIRE(task == task->manager->excl ||
+		(task->manager->exiting && task->manager->excl == NULL));

-	LOCK(&manager->lock);
-	if (manager->exclusive_requested) {
-		UNLOCK(&manager->lock);
+	if (manager->exclusive_requested || manager->pause_requested) {
 		return (ISC_R_LOCKBUSY);
 	}
+
+	LOCK(&manager->halt_lock);
+	INSIST(!manager->exclusive_requested && !manager->pause_requested);
 	manager->exclusive_requested = true;
-	while (manager->tasks_running > 1) {
-		WAIT(&manager->exclusive_granted, &manager->lock);
+	while (manager->halted + 1 < manager->workers) {
+		wake_all_queues(manager);
+		WAIT(&manager->halt_cond, &manager->halt_lock);
 	}
-	UNLOCK(&manager->lock);
+	UNLOCK(&manager->halt_lock);
 	return (ISC_R_SUCCESS);
 }

@@ -1490,11 +1584,14 @@ isc_task_endexclusive(isc_task_t *task0) {

 	REQUIRE(VALID_TASK(task));
 	REQUIRE(task->state == task_state_running);
-	LOCK(&manager->lock);
+	LOCK(&manager->halt_lock);
 	REQUIRE(manager->exclusive_requested);
 	manager->exclusive_requested = false;
-	BROADCAST(&manager->work_available);
-	UNLOCK(&manager->lock);
+	while (manager->halted > 0) {
+		BROADCAST(&manager->halt_cond);
+		WAIT(&manager->halt_cond, &manager->halt_lock);
+	}
+	UNLOCK(&manager->halt_lock);
 }

 void
@@ -1515,14 +1612,14 @@ isc_task_setprivilege(isc_task_t *task0, bool priv) {
 	if (priv == oldpriv)
 		return;

-	LOCK(&manager->lock);
+	LOCK(&manager->queues[task->threadid].lock);
 	if (priv && ISC_LINK_LINKED(task, ready_link))
-		ENQUEUE(manager->ready_priority_tasks, task,
-			ready_priority_link);
+		ENQUEUE(manager->queues[task->threadid].ready_priority_tasks,
+			task, ready_priority_link);
 	else if (!priv && ISC_LINK_LINKED(task, ready_priority_link))
-		DEQUEUE(manager->ready_priority_tasks, task,
-			ready_priority_link);
-	UNLOCK(&manager->lock);
+		DEQUEUE(manager->queues[task->threadid].ready_priority_tasks,
+			task, ready_priority_link);
+	UNLOCK(&manager->queues[task->threadid].lock);
 }

 bool
@@ -1575,11 +1672,13 @@ isc_taskmgr_renderxml(isc_taskmgr_t *mgr0, xmlTextWriterPtr writer) {
 	TRY0(xmlTextWriterEndElement(writer)); /* default-quantum */

 	TRY0(xmlTextWriterStartElement(writer, ISC_XMLCHAR "tasks-running"));
-	TRY0(xmlTextWriterWriteFormatString(writer, "%d", mgr->tasks_running));
+	TRY0(xmlTextWriterWriteFormatString(writer, "%d",
+					    (int) mgr->tasks_running));
 	TRY0(xmlTextWriterEndElement(writer)); /* tasks-running */

 	TRY0(xmlTextWriterStartElement(writer, ISC_XMLCHAR "tasks-ready"));
-	TRY0(xmlTextWriterWriteFormatString(writer, "%d", mgr->tasks_ready));
+	TRY0(xmlTextWriterWriteFormatString(writer, "%d",
+					    (int) mgr->tasks_ready));
 	TRY0(xmlTextWriterEndElement(writer)); /* tasks-ready */

 	TRY0(xmlTextWriterEndElement(writer)); /* thread-model */
--- a/lib/isc/tests/task_test.c
+++ b/lib/isc/tests/task_test.c
@@ -67,7 +67,6 @@ set_and_drop(isc_task_t *task, isc_event_t *event) {
 	*value = (int) isc_taskmgr_mode(taskmgr);
 	counter++;
 	UNLOCK(&lock);
-	isc_taskmgr_setmode(taskmgr, isc_taskmgrmode_normal);
 }

 /*
@@ -230,7 +229,7 @@ ATF_TC_BODY(privileged_events, tc) {
 	isc_task_send(task2, &event);

 	ATF_CHECK_EQ(isc_taskmgr_mode(taskmgr), isc_taskmgrmode_normal);
-	isc_taskmgr_setmode(taskmgr, isc_taskmgrmode_privileged);
+	isc_taskmgr_setprivilegedmode(taskmgr);
 	ATF_CHECK_EQ(isc_taskmgr_mode(taskmgr), isc_taskmgrmode_privileged);

 	isc__taskmgr_resume(taskmgr);
@@ -351,7 +350,7 @@ ATF_TC_BODY(privilege_drop, tc) {
 	isc_task_send(task2, &event);

 	ATF_CHECK_EQ(isc_taskmgr_mode(taskmgr), isc_taskmgrmode_normal);
-	isc_taskmgr_setmode(taskmgr, isc_taskmgrmode_privileged);
+	isc_taskmgr_setprivilegedmode(taskmgr);
 	ATF_CHECK_EQ(isc_taskmgr_mode(taskmgr), isc_taskmgrmode_privileged);

 	isc__taskmgr_resume(taskmgr);
@@ -363,14 +362,12 @@ ATF_TC_BODY(privilege_drop, tc) {
 	}

 	/*
-	 * We can't guarantee what order the events fire, but
-	 * we do know *exactly one* of the privileged tasks will
-	 * have run in privileged mode...
+	 * We need to check that all privilege mode events were fired
+	 * in privileged mode, and non privileged in non-privileged.
 	 */
-	ATF_CHECK(a == isc_taskmgrmode_privileged ||
-		  c == isc_taskmgrmode_privileged ||
+	ATF_CHECK(a == isc_taskmgrmode_privileged &&
+		  c == isc_taskmgrmode_privileged &&
 		  d == isc_taskmgrmode_privileged);
-	ATF_CHECK(a + c + d == isc_taskmgrmode_privileged);

 	/* ...and neither of the non-privileged tasks did... */
 	ATF_CHECK(b == isc_taskmgrmode_normal || e == isc_taskmgrmode_normal);
--- a/lib/isc/win32/include/isc/thread.h
+++ b/lib/isc/win32/include/isc/thread.h
@@ -79,6 +79,9 @@ isc_thread_setconcurrency(unsigned int level);
 void
 isc_thread_setname(isc_thread_t, const char *);

+isc_result_t
+isc_thread_setaffinity(int cpu);
+
 int
 isc_thread_key_create(isc_thread_key_t *key, void (*func)(void *));

--- a/lib/isc/win32/libisc.def.in
+++ b/lib/isc/win32/libisc.def.in
@@ -580,6 +580,8 @@ isc_task_purgeevent
 isc_task_purgerange
 isc_task_send
 isc_task_sendanddetach
+isc_task_sendto
+isc_task_sendtoanddetach
 isc_task_setname
 isc_task_setprivilege
 isc_task_shutdown
@@ -596,7 +598,7 @@ isc_taskmgr_renderjson
 isc_taskmgr_renderxml
@END LIBXML2
 isc_taskmgr_setexcltask
-isc_taskmgr_setmode
+isc_taskmgr_setprivilegedmode
 isc_taskpool_create
 isc_taskpool_destroy
 isc_taskpool_expand
@@ -609,6 +611,7 @@ isc_thread_key_create
 isc_thread_key_delete
 isc_thread_key_getspecific
 isc_thread_key_setspecific
+isc_thread_setaffinity
 isc_thread_setconcurrency
 isc_thread_setname
 isc_time_add
--- a/lib/isc/win32/thread.c
+++ b/lib/isc/win32/thread.c
@@ -66,6 +66,12 @@ isc_thread_setname(isc_thread_t thread, const char *name) {
 	UNUSED(name);
 }

+isc_result_t
+isc_thread_setaffinity(int cpu) {
+	/* no-op on Windows for now */
+	return (ISC_R_SUCCESS);
+}
+
 void *
 isc_thread_key_getspecific(isc_thread_key_t key) {
 	return(TlsGetValue(key));
--- a/lib/ns/client.c
+++ b/lib/ns/client.c
@@ -2960,7 +2960,7 @@ client_create(ns_clientmgr_t *manager, ns_client_t **clientp) {
 	ns_server_attach(manager->sctx, &client->sctx);

 	client->task = NULL;
-	result = isc_task_create(manager->taskmgr, 0, &client->task);
+	result = isc_task_create(manager->taskmgr, 50, &client->task);
 	if (result != ISC_R_SUCCESS)
 		goto cleanup_client;
 	isc_task_setname(client->task, "client", client);