Merge ^/head r277327 through r277718.

1 files changed, 533 insertions, 520 deletions

diff --git a/sys/kern/kern_timeout.c b/sys/kern/kern_timeout.c
index 4336faa..13822fd 100644
--- a/sys/kern/kern_timeout.c
+++ b/sys/kern/kern_timeout.c
@@ -54,8 +54,6 @@ __FBSDID("$FreeBSD$");
 #include <sys/lock.h>
 #include <sys/malloc.h>
 #include <sys/mutex.h>
-#include <sys/rmlock.h>
-#include <sys/rwlock.h>
 #include <sys/proc.h>
 #include <sys/sdt.h>
 #include <sys/sleepqueue.h>
@@ -126,216 +124,37 @@ SYSCTL_INT(_kern, OID_AUTO, pin_pcpu_swi, CTLFLAG_RDTUN | CTLFLAG_NOFETCH, &pin_
  */
 u_int callwheelsize, callwheelmask;
 
-typedef void callout_mutex_op_t(struct lock_object *);
-typedef int callout_owned_op_t(struct lock_object *);
-
-struct callout_mutex_ops {
-	callout_mutex_op_t *lock;
-	callout_mutex_op_t *unlock;
-	callout_owned_op_t *owned;
-};
-
-enum {
-	CALLOUT_LC_UNUSED_0,
-	CALLOUT_LC_UNUSED_1,
-	CALLOUT_LC_UNUSED_2,
-	CALLOUT_LC_UNUSED_3,
-	CALLOUT_LC_SPIN,
-	CALLOUT_LC_MUTEX,
-	CALLOUT_LC_RW,
-	CALLOUT_LC_RM,
-};
-
-static void
-callout_mutex_op_none(struct lock_object *lock)
-{
-}
-
-static int
-callout_owned_op_none(struct lock_object *lock)
-{
-	return (0);
-}
-
-static void
-callout_mutex_lock(struct lock_object *lock)
-{
-	mtx_lock((struct mtx *)lock);
-}
-
-static void
-callout_mutex_unlock(struct lock_object *lock)
-{
-	mtx_unlock((struct mtx *)lock);
-}
-
-static void
-callout_mutex_lock_spin(struct lock_object *lock)
-{
-	mtx_lock_spin((struct mtx *)lock);
-}
-
-static void
-callout_mutex_unlock_spin(struct lock_object *lock)
-{
-	mtx_unlock_spin((struct mtx *)lock);
-}
-
-static int
-callout_mutex_owned(struct lock_object *lock)
-{
-	return (mtx_owned((struct mtx *)lock));
-}
-
-static void
-callout_rm_wlock(struct lock_object *lock)
-{
-	rm_wlock((struct rmlock *)lock);
-}
-
-static void
-callout_rm_wunlock(struct lock_object *lock)
-{
-	rm_wunlock((struct rmlock *)lock);
-}
-
-static int
-callout_rm_owned(struct lock_object *lock)
-{
-	return (rm_wowned((struct rmlock *)lock));
-}
-
-static void
-callout_rw_wlock(struct lock_object *lock)
-{
-	rw_wlock((struct rwlock *)lock);
-}
-
-static void
-callout_rw_wunlock(struct lock_object *lock)
-{
-	rw_wunlock((struct rwlock *)lock);
-}
-
-static int
-callout_rw_owned(struct lock_object *lock)
-{
-	return (rw_wowned((struct rwlock *)lock));
-}
-
-static const struct callout_mutex_ops callout_mutex_ops[8] = {
-	[CALLOUT_LC_UNUSED_0] = {
-		.lock = callout_mutex_op_none,
-		.unlock = callout_mutex_op_none,
-		.owned = callout_owned_op_none,
-	},
-	[CALLOUT_LC_UNUSED_1] = {
-		.lock = callout_mutex_op_none,
-		.unlock = callout_mutex_op_none,
-		.owned = callout_owned_op_none,
-	},
-	[CALLOUT_LC_UNUSED_2] = {
-		.lock = callout_mutex_op_none,
-		.unlock = callout_mutex_op_none,
-		.owned = callout_owned_op_none,
-	},
-	[CALLOUT_LC_UNUSED_3] = {
-		.lock = callout_mutex_op_none,
-		.unlock = callout_mutex_op_none,
-		.owned = callout_owned_op_none,
-	},
-	[CALLOUT_LC_SPIN] = {
-		.lock = callout_mutex_lock_spin,
-		.unlock = callout_mutex_unlock_spin,
-		.owned = callout_mutex_owned,
-	},
-	[CALLOUT_LC_MUTEX] = {
-		.lock = callout_mutex_lock,
-		.unlock = callout_mutex_unlock,
-		.owned = callout_mutex_owned,
-	},
-	[CALLOUT_LC_RW] = {
-		.lock = callout_rw_wlock,
-		.unlock = callout_rw_wunlock,
-		.owned = callout_rw_owned,
-	},
-	[CALLOUT_LC_RM] = {
-		.lock = callout_rm_wlock,
-		.unlock = callout_rm_wunlock,
-		.owned = callout_rm_owned,
-	},
-};
-
-static void
-callout_lock_client(int c_flags, struct lock_object *c_lock)
-{
-	callout_mutex_ops[CALLOUT_GET_LC(c_flags)].lock(c_lock);
-}
-
-static void
-callout_unlock_client(int c_flags, struct lock_object *c_lock)
-{
-	callout_mutex_ops[CALLOUT_GET_LC(c_flags)].unlock(c_lock);
-}
-
-#ifdef SMP
-static int
-callout_lock_owned_client(int c_flags, struct lock_object *c_lock)
-{
-	return (callout_mutex_ops[CALLOUT_GET_LC(c_flags)].owned(c_lock));
-}
-#endif
-
 /*
- * The callout CPU exec structure represent information necessary for
- * describing the state of callouts currently running on the CPU and
- * for handling deferred callout restarts.
- *
- * In particular, the first entry of the array cc_exec_entity holds
- * information for callouts running from the SWI thread context, while
- * the second one holds information for callouts running directly from
- * the hardware interrupt context.
+ * The callout cpu exec entities represent informations necessary for
+ * describing the state of callouts currently running on the CPU and the ones
+ * necessary for migrating callouts to the new callout cpu. In particular,
+ * the first entry of the array cc_exec_entity holds informations for callout
+ * running in SWI thread context, while the second one holds informations
+ * for callout running directly from hardware interrupt context.
+ * The cached informations are very important for deferring migration when
+ * the migrating callout is already running.
  */
 struct cc_exec {
-	/*
-	 * The "cc_curr" points to the currently executing callout and
-	 * is protected by the "cc_lock" spinlock. If no callback is
-	 * currently executing it is equal to "NULL".
-	 */
+	struct callout		*cc_next;
 	struct callout		*cc_curr;
-	/*
-	 * The "cc_restart_args" structure holds the argument for a
-	 * deferred callback restart and is protected by the "cc_lock"
-	 * spinlock. The structure is only valid if "cc_restart" is
-	 * "true". If "cc_restart" is "false" the information in the
-	 * "cc_restart_args" structure shall be ignored.
-	 */
-	struct callout_args	cc_restart_args;
-	bool			cc_restart;
-	/*
-	 * The "cc_cancel" variable allows the currently pending
-	 * callback to be atomically cancelled. This field is write
-	 * protected by the "cc_lock" spinlock.
-	 */
-	bool cc_cancel;
-	/*
-	 * The "cc_drain_fn" points to a function which shall be
-	 * called with the argument stored in "cc_drain_arg" when an
-	 * asynchronous drain is performed. This field is write
-	 * protected by the "cc_lock" spinlock.
-	 */
-	callout_func_t *cc_drain_fn;
-	void *cc_drain_arg;
+#ifdef SMP
+	void			(*ce_migration_func)(void *);
+	void			*ce_migration_arg;
+	int			ce_migration_cpu;
+	sbintime_t		ce_migration_time;
+	sbintime_t		ce_migration_prec;
+#endif
+	bool			cc_cancel;
+	bool			cc_waiting;
 };
 
 /*
- * There is one "struct callout_cpu" per CPU, holding all relevant
+ * There is one struct callout_cpu per cpu, holding all relevant
  * state for the callout processing thread on the individual CPU.
  */
 struct callout_cpu {
 	struct mtx_padalign	cc_lock;
 	struct cc_exec 		cc_exec_entity[2];
-	struct callout		*cc_exec_next_dir;
 	struct callout		*cc_callout;
 	struct callout_list	*cc_callwheel;
 	struct callout_tailq	cc_expireq;
@@ -347,7 +166,27 @@ struct callout_cpu {
 	char			cc_ktr_event_name[20];
 };
 
+#define	cc_exec_curr		cc_exec_entity[0].cc_curr
+#define	cc_exec_next		cc_exec_entity[0].cc_next
+#define	cc_exec_cancel		cc_exec_entity[0].cc_cancel
+#define	cc_exec_waiting		cc_exec_entity[0].cc_waiting
+#define	cc_exec_curr_dir	cc_exec_entity[1].cc_curr
+#define	cc_exec_next_dir	cc_exec_entity[1].cc_next
+#define	cc_exec_cancel_dir	cc_exec_entity[1].cc_cancel
+#define	cc_exec_waiting_dir	cc_exec_entity[1].cc_waiting
+
 #ifdef SMP
+#define	cc_migration_func	cc_exec_entity[0].ce_migration_func
+#define	cc_migration_arg	cc_exec_entity[0].ce_migration_arg
+#define	cc_migration_cpu	cc_exec_entity[0].ce_migration_cpu
+#define	cc_migration_time	cc_exec_entity[0].ce_migration_time
+#define	cc_migration_prec	cc_exec_entity[0].ce_migration_prec
+#define	cc_migration_func_dir	cc_exec_entity[1].ce_migration_func
+#define	cc_migration_arg_dir	cc_exec_entity[1].ce_migration_arg
+#define	cc_migration_cpu_dir	cc_exec_entity[1].ce_migration_cpu
+#define	cc_migration_time_dir	cc_exec_entity[1].ce_migration_time
+#define	cc_migration_prec_dir	cc_exec_entity[1].ce_migration_prec
+
 struct callout_cpu cc_cpu[MAXCPU];
 #define	CPUBLOCK	MAXCPU
 #define	CC_CPU(cpu)	(&cc_cpu[(cpu)])
@@ -372,9 +211,60 @@ static void	softclock_call_cc(struct callout *c, struct callout_cpu *cc,
 
 static MALLOC_DEFINE(M_CALLOUT, "callout", "Callout datastructures");
 
+/**
+ * Locked by cc_lock:
+ *   cc_curr         - If a callout is in progress, it is cc_curr.
+ *                     If cc_curr is non-NULL, threads waiting in
+ *                     callout_drain() will be woken up as soon as the
+ *                     relevant callout completes.
+ *   cc_cancel       - Changing to 1 with both callout_lock and cc_lock held
+ *                     guarantees that the current callout will not run.
+ *                     The softclock() function sets this to 0 before it
+ *                     drops callout_lock to acquire c_lock, and it calls
+ *                     the handler only if curr_cancelled is still 0 after
+ *                     cc_lock is successfully acquired.
+ *   cc_waiting      - If a thread is waiting in callout_drain(), then
+ *                     callout_wait is nonzero.  Set only when
+ *                     cc_curr is non-NULL.
+ */
+
+/*
+ * Resets the execution entity tied to a specific callout cpu.
+ */
+static void
+cc_cce_cleanup(struct callout_cpu *cc, int direct)
+{
+
+	cc->cc_exec_entity[direct].cc_curr = NULL;
+	cc->cc_exec_entity[direct].cc_next = NULL;
+	cc->cc_exec_entity[direct].cc_cancel = false;
+	cc->cc_exec_entity[direct].cc_waiting = false;
+#ifdef SMP
+	cc->cc_exec_entity[direct].ce_migration_cpu = CPUBLOCK;
+	cc->cc_exec_entity[direct].ce_migration_time = 0;
+	cc->cc_exec_entity[direct].ce_migration_prec = 0;
+	cc->cc_exec_entity[direct].ce_migration_func = NULL;
+	cc->cc_exec_entity[direct].ce_migration_arg = NULL;
+#endif
+}
+
+/*
+ * Checks if migration is requested by a specific callout cpu.
+ */
+static int
+cc_cce_migrating(struct callout_cpu *cc, int direct)
+{
+
+#ifdef SMP
+	return (cc->cc_exec_entity[direct].ce_migration_cpu != CPUBLOCK);
+#else
+	return (0);
+#endif
+}
+
 /*
- * Kernel low level callwheel initialization called from cpu0 during
- * kernel startup:
+ * Kernel low level callwheel initialization
+ * called on cpu0 during kernel startup.
  */
 static void
 callout_callwheel_init(void *dummy)
@@ -434,6 +324,8 @@ callout_cpu_init(struct callout_cpu *cc, int cpu)
 		LIST_INIT(&cc->cc_callwheel[i]);
 	TAILQ_INIT(&cc->cc_expireq);
 	cc->cc_firstevent = SBT_MAX;
+	for (i = 0; i < 2; i++)
+		cc_cce_cleanup(cc, i);
 	snprintf(cc->cc_ktr_event_name, sizeof(cc->cc_ktr_event_name),
 	    "callwheel cpu %d", cpu);
 	if (cc->cc_callout == NULL)	/* Only cpu0 handles timeout(9) */
@@ -441,11 +333,41 @@ callout_cpu_init(struct callout_cpu *cc, int cpu)
 	for (i = 0; i < ncallout; i++) {
 		c = &cc->cc_callout[i];
 		callout_init(c, 0);
-		c->c_flags |= CALLOUT_LOCAL_ALLOC;
+		c->c_flags = CALLOUT_LOCAL_ALLOC;
 		SLIST_INSERT_HEAD(&cc->cc_callfree, c, c_links.sle);
 	}
 }
 
+#ifdef SMP
+/*
+ * Switches the cpu tied to a specific callout.
+ * The function expects a locked incoming callout cpu and returns with
+ * locked outcoming callout cpu.
+ */
+static struct callout_cpu *
+callout_cpu_switch(struct callout *c, struct callout_cpu *cc, int new_cpu)
+{
+	struct callout_cpu *new_cc;
+
+	MPASS(c != NULL && cc != NULL);
+	CC_LOCK_ASSERT(cc);
+
+	/*
+	 * Avoid interrupts and preemption firing after the callout cpu
+	 * is blocked in order to avoid deadlocks as the new thread
+	 * may be willing to acquire the callout cpu lock.
+	 */
+	c->c_cpu = CPUBLOCK;
+	spinlock_enter();
+	CC_UNLOCK(cc);
+	new_cc = CC_CPU(new_cpu);
+	CC_LOCK(new_cc);
+	spinlock_exit();
+	c->c_cpu = new_cpu;
+	return (new_cc);
+}
+#endif
+
 /*
  * Start standard softclock thread.
  */
@@ -522,8 +444,9 @@ callout_process(sbintime_t now)
 #ifdef CALLOUT_PROFILING
 	int depth_dir = 0, mpcalls_dir = 0, lockcalls_dir = 0;
 #endif
+
 	cc = CC_SELF();
-	CC_LOCK(cc);
+	mtx_lock_spin_flags(&cc->cc_lock, MTX_QUIET);
 
 	/* Compute the buckets of the last scan and present times. */
 	firstb = callout_hash(cc->cc_lastscan);
@@ -626,7 +549,7 @@ next:
 	avg_mpcalls_dir += (mpcalls_dir * 1000 - avg_mpcalls_dir) >> 8;
 	avg_lockcalls_dir += (lockcalls_dir * 1000 - avg_lockcalls_dir) >> 8;
 #endif
-	CC_UNLOCK(cc);
+	mtx_unlock_spin_flags(&cc->cc_lock, MTX_QUIET);
 	/*
 	 * swi_sched acquires the thread lock, so we don't want to call it
 	 * with cc_lock held; incorrect locking order.
@@ -639,55 +562,49 @@ static struct callout_cpu *
 callout_lock(struct callout *c)
 {
 	struct callout_cpu *cc;
-	cc = CC_CPU(c->c_cpu);
-	CC_LOCK(cc);
+	int cpu;
+
+	for (;;) {
+		cpu = c->c_cpu;
+#ifdef SMP
+		if (cpu == CPUBLOCK) {
+			while (c->c_cpu == CPUBLOCK)
+				cpu_spinwait();
+			continue;
+		}
+#endif
+		cc = CC_CPU(cpu);
+		CC_LOCK(cc);
+		if (cpu == c->c_cpu)
+			break;
+		CC_UNLOCK(cc);
+	}
 	return (cc);
 }
 
-static struct callout_cpu *
-callout_cc_add_locked(struct callout *c, struct callout_cpu *cc,
-    struct callout_args *coa, bool can_swap_cpu)
+static void
+callout_cc_add(struct callout *c, struct callout_cpu *cc,
+    sbintime_t sbt, sbintime_t precision, void (*func)(void *),
+    void *arg, int cpu, int flags)
 {
-#ifndef NO_EVENTTIMERS
-	sbintime_t sbt;
-#endif
 	int bucket;
 
 	CC_LOCK_ASSERT(cc);
-
-	/* update flags before swapping locks, if any */
-	c->c_flags &= ~(CALLOUT_PROCESSED | CALLOUT_DIRECT | CALLOUT_DEFRESTART);
-	if (coa->flags & C_DIRECT_EXEC)
-		c->c_flags |= (CALLOUT_ACTIVE | CALLOUT_PENDING | CALLOUT_DIRECT);
-	else
-		c->c_flags |= (CALLOUT_ACTIVE | CALLOUT_PENDING);
-
-#ifdef SMP
-	/*
-	 * Check if we are changing the CPU on which the callback
-	 * should be executed and if we have a lock protecting us:
-	 */
-	if (can_swap_cpu != false && coa->cpu != c->c_cpu &&
-	    callout_lock_owned_client(c->c_flags, c->c_lock) != 0) {
-		CC_UNLOCK(cc);
-		c->c_cpu = coa->cpu;
-		cc = callout_lock(c);
-	}
-#endif
-	if (coa->time < cc->cc_lastscan)
-		coa->time = cc->cc_lastscan;
-	c->c_arg = coa->arg;
-	c->c_func = coa->func;
-	c->c_time = coa->time;
-	c->c_precision = coa->precision;
-
+	if (sbt < cc->cc_lastscan)
+		sbt = cc->cc_lastscan;
+	c->c_arg = arg;
+	c->c_flags |= (CALLOUT_ACTIVE | CALLOUT_PENDING);
+	if (flags & C_DIRECT_EXEC)
+		c->c_flags |= CALLOUT_DIRECT;
+	c->c_flags &= ~CALLOUT_PROCESSED;
+	c->c_func = func;
+	c->c_time = sbt;
+	c->c_precision = precision;
 	bucket = callout_get_bucket(c->c_time);
 	CTR3(KTR_CALLOUT, "precision set for %p: %d.%08x",
 	    c, (int)(c->c_precision >> 32),
 	    (u_int)(c->c_precision & 0xffffffff));
 	LIST_INSERT_HEAD(&cc->cc_callwheel[bucket], c, c_links.le);
-
-	/* Ensure we are first to be scanned, if called via a callback */
 	if (cc->cc_bucket == bucket)
 		cc->cc_exec_next_dir = c;
 #ifndef NO_EVENTTIMERS
@@ -700,16 +617,17 @@ callout_cc_add_locked(struct callout *c, struct callout_cpu *cc,
 	sbt = c->c_time + c->c_precision;
 	if (sbt < cc->cc_firstevent) {
 		cc->cc_firstevent = sbt;
-		cpu_new_callout(coa->cpu, sbt, c->c_time);
+		cpu_new_callout(cpu, sbt, c->c_time);
 	}
 #endif
-	return (cc);
 }
 
 static void
 callout_cc_del(struct callout *c, struct callout_cpu *cc)
 {
 
+	if ((c->c_flags & CALLOUT_LOCAL_ALLOC) == 0)
+		return;
 	c->c_func = NULL;
 	SLIST_INSERT_HEAD(&cc->cc_callfree, c, c_links.sle);
 }
@@ -721,10 +639,20 @@ softclock_call_cc(struct callout *c, struct callout_cpu *cc,
 #endif
     int direct)
 {
-	callout_func_t *c_func;
+	struct rm_priotracker tracker;
+	void (*c_func)(void *);
 	void *c_arg;
+	struct lock_class *class;
 	struct lock_object *c_lock;
+	uintptr_t lock_status;
 	int c_flags;
+#ifdef SMP
+	struct callout_cpu *new_cc;
+	void (*new_func)(void *);
+	void *new_arg;
+	int flags, new_cpu;
+	sbintime_t new_prec, new_time;
+#endif
 #if defined(DIAGNOSTIC) || defined(CALLOUT_PROFILING) 
 	sbintime_t sbt1, sbt2;
 	struct timespec ts2;
@@ -735,39 +663,37 @@ softclock_call_cc(struct callout *c, struct callout_cpu *cc,
 	KASSERT((c->c_flags & (CALLOUT_PENDING | CALLOUT_ACTIVE)) ==
 	    (CALLOUT_PENDING | CALLOUT_ACTIVE),
 	    ("softclock_call_cc: pend|act %p %x", c, c->c_flags));
+	class = (c->c_lock != NULL) ? LOCK_CLASS(c->c_lock) : NULL;
+	lock_status = 0;
+	if (c->c_flags & CALLOUT_SHAREDLOCK) {
+		if (class == &lock_class_rm)
+			lock_status = (uintptr_t)&tracker;
+		else
+			lock_status = 1;
+	}
 	c_lock = c->c_lock;
 	c_func = c->c_func;
 	c_arg = c->c_arg;
 	c_flags = c->c_flags;
-
-	/* remove pending bit */
-	c->c_flags &= ~CALLOUT_PENDING;
-
-	/* reset our local state */
+	if (c->c_flags & CALLOUT_LOCAL_ALLOC)
+		c->c_flags = CALLOUT_LOCAL_ALLOC;
+	else
+		c->c_flags &= ~CALLOUT_PENDING;
 	cc->cc_exec_entity[direct].cc_curr = c;
-	cc->cc_exec_entity[direct].cc_restart = false;
-	cc->cc_exec_entity[direct].cc_drain_fn = NULL;
-	cc->cc_exec_entity[direct].cc_drain_arg = NULL;
-
+	cc->cc_exec_entity[direct].cc_cancel = false;
+	CC_UNLOCK(cc);
 	if (c_lock != NULL) {
-		cc->cc_exec_entity[direct].cc_cancel = false;
-		CC_UNLOCK(cc);
-
-		/* unlocked region for switching locks */
-
-		callout_lock_client(c_flags, c_lock);
-
+		class->lc_lock(c_lock, lock_status);
 		/*
-		 * Check if the callout may have been cancelled while
-		 * we were switching locks. Even though the callout is
-		 * specifying a lock, it might not be certain this
-		 * lock is locked when starting and stopping callouts.
+		 * The callout may have been cancelled
+		 * while we switched locks.
 		 */
-		CC_LOCK(cc);
 		if (cc->cc_exec_entity[direct].cc_cancel) {
-			callout_unlock_client(c_flags, c_lock);
-			goto skip_cc_locked;
+			class->lc_unlock(c_lock);
+			goto skip;
 		}
+		/* The callout cannot be stopped now. */
+		cc->cc_exec_entity[direct].cc_cancel = true;
 		if (c_lock == &Giant.lock_object) {
 #ifdef CALLOUT_PROFILING
 			(*gcalls)++;
@@ -788,11 +714,6 @@ softclock_call_cc(struct callout *c, struct callout_cpu *cc,
 		CTR3(KTR_CALLOUT, "callout %p func %p arg %p",
 		    c, c_func, c_arg);
 	}
-	/* The callout cannot be stopped now! */
-	cc->cc_exec_entity[direct].cc_cancel = true;
-	CC_UNLOCK(cc);
-
-	/* unlocked region */
 	KTR_STATE3(KTR_SCHED, "callout", cc->cc_ktr_event_name, "running",
 	    "func:%p", c_func, "arg:%p", c_arg, "direct:%d", direct);
 #if defined(DIAGNOSTIC) || defined(CALLOUT_PROFILING)
@@ -819,40 +740,85 @@ softclock_call_cc(struct callout *c, struct callout_cpu *cc,
 #endif
 	KTR_STATE0(KTR_SCHED, "callout", cc->cc_ktr_event_name, "idle");
 	CTR1(KTR_CALLOUT, "callout %p finished", c);
-
-	/*
-	 * At this point the callback structure might have been freed,
-	 * so we need to check the previously copied value of
-	 * "c->c_flags":
-	 */
 	if ((c_flags & CALLOUT_RETURNUNLOCKED) == 0)
-		callout_unlock_client(c_flags, c_lock);
-
+		class->lc_unlock(c_lock);
+skip:
 	CC_LOCK(cc);
-
-skip_cc_locked:
 	KASSERT(cc->cc_exec_entity[direct].cc_curr == c, ("mishandled cc_curr"));
 	cc->cc_exec_entity[direct].cc_curr = NULL;
-
-	/* Check if there is anything which needs draining */
-	if (cc->cc_exec_entity[direct].cc_drain_fn != NULL) {
+	if (cc->cc_exec_entity[direct].cc_waiting) {
 		/*
-		 * Unlock the CPU callout last, so that any use of
-		 * structures belonging to the callout are complete:
+		 * There is someone waiting for the
+		 * callout to complete.
+		 * If the callout was scheduled for
+		 * migration just cancel it.
 		 */
+		if (cc_cce_migrating(cc, direct)) {
+			cc_cce_cleanup(cc, direct);
+
+			/*
+			 * It should be assert here that the callout is not
+			 * destroyed but that is not easy.
+			 */
+			c->c_flags &= ~CALLOUT_DFRMIGRATION;
+		}
+		cc->cc_exec_entity[direct].cc_waiting = false;
 		CC_UNLOCK(cc);
-		/* call drain function unlocked */
-		cc->cc_exec_entity[direct].cc_drain_fn(
-		    cc->cc_exec_entity[direct].cc_drain_arg);
+		wakeup(&cc->cc_exec_entity[direct].cc_waiting);
 		CC_LOCK(cc);
-	} else if (c_flags & CALLOUT_LOCAL_ALLOC) {
-		/* return callout back to freelist */
-		callout_cc_del(c, cc);
-	} else if (cc->cc_exec_entity[direct].cc_restart) {
-		/* [re-]schedule callout, if any */
-		cc = callout_cc_add_locked(c, cc,
-		    &cc->cc_exec_entity[direct].cc_restart_args, false);
+	} else if (cc_cce_migrating(cc, direct)) {
+		KASSERT((c_flags & CALLOUT_LOCAL_ALLOC) == 0,
+		    ("Migrating legacy callout %p", c));
+#ifdef SMP
+		/*
+		 * If the callout was scheduled for
+		 * migration just perform it now.
+		 */
+		new_cpu = cc->cc_exec_entity[direct].ce_migration_cpu;
+		new_time = cc->cc_exec_entity[direct].ce_migration_time;
+		new_prec = cc->cc_exec_entity[direct].ce_migration_prec;
+		new_func = cc->cc_exec_entity[direct].ce_migration_func;
+		new_arg = cc->cc_exec_entity[direct].ce_migration_arg;
+		cc_cce_cleanup(cc, direct);
+
+		/*
+		 * It should be assert here that the callout is not destroyed
+		 * but that is not easy.
+		 *
+		 * As first thing, handle deferred callout stops.
+		 */
+		if ((c->c_flags & CALLOUT_DFRMIGRATION) == 0) {
+			CTR3(KTR_CALLOUT,
+			     "deferred cancelled %p func %p arg %p",
+			     c, new_func, new_arg);
+			callout_cc_del(c, cc);
+			return;
+		}
+		c->c_flags &= ~CALLOUT_DFRMIGRATION;
+
+		new_cc = callout_cpu_switch(c, cc, new_cpu);
+		flags = (direct) ? C_DIRECT_EXEC : 0;
+		callout_cc_add(c, new_cc, new_time, new_prec, new_func,
+		    new_arg, new_cpu, flags);
+		CC_UNLOCK(new_cc);
+		CC_LOCK(cc);
+#else
+		panic("migration should not happen");
+#endif
 	}
+	/*
+	 * If the current callout is locally allocated (from
+	 * timeout(9)) then put it on the freelist.
+	 *
+	 * Note: we need to check the cached copy of c_flags because
+	 * if it was not local, then it's not safe to deref the
+	 * callout pointer.
+	 */
+	KASSERT((c_flags & CALLOUT_LOCAL_ALLOC) == 0 ||
+	    c->c_flags == CALLOUT_LOCAL_ALLOC,
+	    ("corrupted callout"));
+	if (c_flags & CALLOUT_LOCAL_ALLOC)
+		callout_cc_del(c, cc);
 }
 
 /*
@@ -933,11 +899,10 @@ timeout(timeout_t *ftn, void *arg, int to_ticks)
 		/* XXX Attempt to malloc first */
 		panic("timeout table full");
 	SLIST_REMOVE_HEAD(&cc->cc_callfree, c_links.sle);
+	callout_reset(new, to_ticks, ftn, arg);
 	handle.callout = new;
 	CC_UNLOCK(cc);
 
-	callout_reset(new, to_ticks, ftn, arg);
-
 	return (handle);
 }
 
@@ -945,7 +910,6 @@ void
 untimeout(timeout_t *ftn, void *arg, struct callout_handle handle)
 {
 	struct callout_cpu *cc;
-	bool match;
 
 	/*
 	 * Check for a handle that was initialized
@@ -956,11 +920,9 @@ untimeout(timeout_t *ftn, void *arg, struct callout_handle handle)
 		return;
 
 	cc = callout_lock(handle.callout);
-	match = (handle.callout->c_func == ftn && handle.callout->c_arg == arg);
-	CC_UNLOCK(cc);
-
-	if (match)
+	if (handle.callout->c_func == ftn && handle.callout->c_arg == arg)
 		callout_stop(handle.callout);
+	CC_UNLOCK(cc);
 }
 
 void
@@ -969,119 +931,6 @@ callout_handle_init(struct callout_handle *handle)
 	handle->callout = NULL;
 }
 
-static int
-callout_restart_async(struct callout *c, struct callout_args *coa,
-    callout_func_t *drain_fn, void *drain_arg)
-{
-	struct callout_cpu *cc;
-	int cancelled;
-	int direct;
-
-	cc = callout_lock(c);
-
-	/* Figure out if the callout is direct or not */
-	direct = ((c->c_flags & CALLOUT_DIRECT) != 0);
-
-	/*
-	 * Check if the callback is currently scheduled for
-	 * completion:
-	 */
-	if (cc->cc_exec_entity[direct].cc_curr == c) {
-		/*
-		 * Try to prevent the callback from running by setting
-		 * the "cc_cancel" variable to "true". Also check if
-		 * the callout was previously subject to a deferred
-		 * callout restart:
-		 */
-		if (cc->cc_exec_entity[direct].cc_cancel == false ||
-		    (c->c_flags & CALLOUT_DEFRESTART) != 0) {
-			cc->cc_exec_entity[direct].cc_cancel = true;
-			cancelled = 1;
-		} else {
-			cancelled = 0;
-		}
-
-		/*
-		 * Prevent callback restart if "callout_drain_xxx()"
-		 * is being called or we are stopping the callout or
-		 * the callback was preallocated by us:
-		 */
-		if (cc->cc_exec_entity[direct].cc_drain_fn != NULL ||
-		    coa == NULL || (c->c_flags & CALLOUT_LOCAL_ALLOC) != 0) {
-			CTR4(KTR_CALLOUT, "%s %p func %p arg %p",
-			    cancelled ? "cancelled and draining" : "draining",
-			    c, c->c_func, c->c_arg);
-
-			/* clear old flags, if any */
-			c->c_flags &= ~(CALLOUT_ACTIVE | CALLOUT_PENDING |
-			    CALLOUT_DEFRESTART | CALLOUT_PROCESSED);
-
-			/* clear restart flag, if any */
-			cc->cc_exec_entity[direct].cc_restart = false;
-
-			/* set drain function, if any */
-			if (drain_fn != NULL) {
-				cc->cc_exec_entity[direct].cc_drain_fn = drain_fn;
-				cc->cc_exec_entity[direct].cc_drain_arg = drain_arg;
-				cancelled |= 2;		/* XXX define the value */
-			}
-		} else {
-			CTR4(KTR_CALLOUT, "%s %p func %p arg %p",
-			    cancelled ? "cancelled and restarting" : "restarting",
-			    c, c->c_func, c->c_arg);
-
-			/* get us back into the game */
-			c->c_flags |= (CALLOUT_ACTIVE | CALLOUT_PENDING |
-			    CALLOUT_DEFRESTART);
-			c->c_flags &= ~CALLOUT_PROCESSED;
-
-			/* enable deferred restart */
-			cc->cc_exec_entity[direct].cc_restart = true;
-
-			/* store arguments for the deferred restart, if any */
-			cc->cc_exec_entity[direct].cc_restart_args = *coa;
-		}
-	} else {
-		/* stop callout */
-		if (c->c_flags & CALLOUT_PENDING) {
-			/*
-			 * The callback has not yet been executed, and
-			 * we simply just need to unlink it:
-			 */
-			if ((c->c_flags & CALLOUT_PROCESSED) == 0) {
-				if (cc->cc_exec_next_dir == c)
-					cc->cc_exec_next_dir = LIST_NEXT(c, c_links.le);
-				LIST_REMOVE(c, c_links.le);
-			} else {
-				TAILQ_REMOVE(&cc->cc_expireq, c, c_links.tqe);
-			}
-			cancelled = 1;
-		} else {
-			cancelled = 0;
-		}
-
-		CTR4(KTR_CALLOUT, "%s %p func %p arg %p",
-		    cancelled ? "rescheduled" : "scheduled",
-		    c, c->c_func, c->c_arg);
-
-		/* [re-]schedule callout, if any */
-		if (coa != NULL) {
-			cc = callout_cc_add_locked(c, cc, coa, true);
-		} else {
-			/* clear old flags, if any */
-			c->c_flags &= ~(CALLOUT_ACTIVE | CALLOUT_PENDING |
-			    CALLOUT_DEFRESTART | CALLOUT_PROCESSED);
-
-			/* return callback to pre-allocated list, if any */
-			if ((c->c_flags & CALLOUT_LOCAL_ALLOC) && cancelled != 0) {
-				callout_cc_del(c, cc);
-			}
-		}
-	}
-	CC_UNLOCK(cc);
-	return (cancelled);
-}
-
 /*
  * New interface; clients allocate their own callout structures.
  *
@@ -1100,32 +949,25 @@ callout_restart_async(struct callout *c, struct callout_args *coa,
  */
 int
 callout_reset_sbt_on(struct callout *c, sbintime_t sbt, sbintime_t precision,
-    callout_func_t *ftn, void *arg, int cpu, int flags)
+    void (*ftn)(void *), void *arg, int cpu, int flags)
 {
-	struct callout_args coa;
-
-	/* store arguments for callout add function */
-	coa.func = ftn;
-	coa.arg = arg;
-	coa.precision = precision;
-	coa.flags = flags;
-	coa.cpu = cpu;
-
-	/* compute the rest of the arguments needed */
-	if (coa.flags & C_ABSOLUTE) {
-		coa.time = sbt;
-	} else {
-		sbintime_t pr;
+	sbintime_t to_sbt, pr;
+	struct callout_cpu *cc;
+	int cancelled, direct;
 
-		if ((coa.flags & C_HARDCLOCK) && (sbt < tick_sbt))
+	cancelled = 0;
+	if (flags & C_ABSOLUTE) {
+		to_sbt = sbt;
+	} else {
+		if ((flags & C_HARDCLOCK) && (sbt < tick_sbt))
 			sbt = tick_sbt;
-		if ((coa.flags & C_HARDCLOCK) ||
+		if ((flags & C_HARDCLOCK) ||
 #ifdef NO_EVENTTIMERS
 		    sbt >= sbt_timethreshold) {
-			coa.time = getsbinuptime();
+			to_sbt = getsbinuptime();
 
 			/* Add safety belt for the case of hz > 1000. */
-			coa.time += tc_tick_sbt - tick_sbt;
+			to_sbt += tc_tick_sbt - tick_sbt;
 #else
 		    sbt >= sbt_tickthreshold) {
 			/*
@@ -1135,29 +977,101 @@ callout_reset_sbt_on(struct callout *c, sbintime_t sbt, sbintime_t precision,
 			 * active ones.
 			 */
 #ifdef __LP64__
-			coa.time = DPCPU_GET(hardclocktime);
+			to_sbt = DPCPU_GET(hardclocktime);
 #else
 			spinlock_enter();
-			coa.time = DPCPU_GET(hardclocktime);
+			to_sbt = DPCPU_GET(hardclocktime);
 			spinlock_exit();
 #endif
 #endif
-			if ((coa.flags & C_HARDCLOCK) == 0)
-				coa.time += tick_sbt;
+			if ((flags & C_HARDCLOCK) == 0)
+				to_sbt += tick_sbt;
 		} else
-			coa.time = sbinuptime();
-		if (SBT_MAX - coa.time < sbt)
-			coa.time = SBT_MAX;
+			to_sbt = sbinuptime();
+		if (SBT_MAX - to_sbt < sbt)
+			to_sbt = SBT_MAX;
 		else
-			coa.time += sbt;
-		pr = ((C_PRELGET(coa.flags) < 0) ? sbt >> tc_precexp :
-		    sbt >> C_PRELGET(coa.flags));
-		if (pr > coa.precision)
-			coa.precision = pr;
+			to_sbt += sbt;
+		pr = ((C_PRELGET(flags) < 0) ? sbt >> tc_precexp :
+		    sbt >> C_PRELGET(flags));
+		if (pr > precision)
+			precision = pr;
+	}
+	/*
+	 * Don't allow migration of pre-allocated callouts lest they
+	 * become unbalanced.
+	 */
+	if (c->c_flags & CALLOUT_LOCAL_ALLOC)
+		cpu = c->c_cpu;
+	direct = (c->c_flags & CALLOUT_DIRECT) != 0;
+	KASSERT(!direct || c->c_lock == NULL,
+	    ("%s: direct callout %p has lock", __func__, c));
+	cc = callout_lock(c);
+	if (cc->cc_exec_entity[direct].cc_curr == c) {
+		/*
+		 * We're being asked to reschedule a callout which is
+		 * currently in progress.  If there is a lock then we
+		 * can cancel the callout if it has not really started.
+		 */
+		if (c->c_lock != NULL && !cc->cc_exec_entity[direct].cc_cancel)
+			cancelled = cc->cc_exec_entity[direct].cc_cancel = true;
+		if (cc->cc_exec_entity[direct].cc_waiting) {
+			/*
+			 * Someone has called callout_drain to kill this
+			 * callout.  Don't reschedule.
+			 */
+			CTR4(KTR_CALLOUT, "%s %p func %p arg %p",
+			    cancelled ? "cancelled" : "failed to cancel",
+			    c, c->c_func, c->c_arg);
+			CC_UNLOCK(cc);
+			return (cancelled);
+		}
+	}
+	if (c->c_flags & CALLOUT_PENDING) {
+		if ((c->c_flags & CALLOUT_PROCESSED) == 0) {
+			if (cc->cc_exec_next_dir == c)
+				cc->cc_exec_next_dir = LIST_NEXT(c, c_links.le);
+			LIST_REMOVE(c, c_links.le);
+		} else
+			TAILQ_REMOVE(&cc->cc_expireq, c, c_links.tqe);
+		cancelled = 1;
+		c->c_flags &= ~(CALLOUT_ACTIVE | CALLOUT_PENDING);
+	}
+
+#ifdef SMP
+	/*
+	 * If the callout must migrate try to perform it immediately.
+	 * If the callout is currently running, just defer the migration
+	 * to a more appropriate moment.
+	 */
+	if (c->c_cpu != cpu) {
+		if (cc->cc_exec_entity[direct].cc_curr == c) {
+			cc->cc_exec_entity[direct].ce_migration_cpu = cpu;
+			cc->cc_exec_entity[direct].ce_migration_time
+			    = to_sbt;
+			cc->cc_exec_entity[direct].ce_migration_prec 
+			    = precision;
+			cc->cc_exec_entity[direct].ce_migration_func = ftn;
+			cc->cc_exec_entity[direct].ce_migration_arg = arg;
+			c->c_flags |= CALLOUT_DFRMIGRATION;
+			CTR6(KTR_CALLOUT,
+		    "migration of %p func %p arg %p in %d.%08x to %u deferred",
+			    c, c->c_func, c->c_arg, (int)(to_sbt >> 32),
+			    (u_int)(to_sbt & 0xffffffff), cpu);
+			CC_UNLOCK(cc);
+			return (cancelled);
+		}
+		cc = callout_cpu_switch(c, cc, cpu);
 	}
+#endif
 
-	/* get callback started, if any */
-	return (callout_restart_async(c, &coa, NULL, NULL));
+	callout_cc_add(c, cc, to_sbt, precision, ftn, arg, cpu, flags);
+	CTR6(KTR_CALLOUT, "%sscheduled %p func %p arg %p in %d.%08x",
+	    cancelled ? "re" : "", c, c->c_func, c->c_arg, (int)(to_sbt >> 32),
+	    (u_int)(to_sbt & 0xffffffff));
+	CC_UNLOCK(cc);
+
+	return (cancelled);
 }
 
 /*
@@ -1176,105 +1090,204 @@ callout_schedule(struct callout *c, int to_ticks)
 }
 
 int
-callout_stop(struct callout *c)
+_callout_stop_safe(struct callout *c, int safe)
 {
-	/* get callback stopped, if any */
-	return (callout_restart_async(c, NULL, NULL, NULL));
-}
-
-static void
-callout_drain_function(void *arg)
-{
-	wakeup(arg);
-}
-
-int
-callout_drain_async(struct callout *c, callout_func_t *fn, void *arg)
-{
-	/* get callback stopped, if any */
-	return (callout_restart_async(c, NULL, fn, arg) & 2);
-}
-
-int
-callout_drain(struct callout *c)
-{
-	int cancelled;
-
-	WITNESS_WARN(WARN_GIANTOK | WARN_SLEEPOK, NULL,
-	    "Draining callout");
-
-	callout_lock_client(c->c_flags, c->c_lock);
-
-	/* at this point the "c->c_cpu" field is not changing */
+	struct callout_cpu *cc, *old_cc;
+	struct lock_class *class;
+	int direct, sq_locked, use_lock;
 
-	cancelled = callout_drain_async(c, &callout_drain_function, c);
-
-	if (cancelled != 0) {
-		struct callout_cpu *cc;
-		int direct;
+	/*
+	 * Some old subsystems don't hold Giant while running a callout_stop(),
+	 * so just discard this check for the moment.
+	 */
+	if (!safe && c->c_lock != NULL) {
+		if (c->c_lock == &Giant.lock_object)
+			use_lock = mtx_owned(&Giant);
+		else {
+			use_lock = 1;
+			class = LOCK_CLASS(c->c_lock);
+			class->lc_assert(c->c_lock, LA_XLOCKED);
+		}
+	} else
+		use_lock = 0;
+	direct = (c->c_flags & CALLOUT_DIRECT) != 0;
+	sq_locked = 0;
+	old_cc = NULL;
+again:
+	cc = callout_lock(c);
 
-		CTR3(KTR_CALLOUT, "need to drain %p func %p arg %p",
-		    c, c->c_func, c->c_arg);
+	/*
+	 * If the callout was migrating while the callout cpu lock was
+	 * dropped,  just drop the sleepqueue lock and check the states
+	 * again.
+	 */
+	if (sq_locked != 0 && cc != old_cc) {
+#ifdef SMP
+		CC_UNLOCK(cc);
+		sleepq_release(&old_cc->cc_exec_entity[direct].cc_waiting);
+		sq_locked = 0;
+		old_cc = NULL;
+		goto again;
+#else
+		panic("migration should not happen");
+#endif
+	}
 
-		cc = callout_lock(c);
-		direct = ((c->c_flags & CALLOUT_DIRECT) != 0);
+	/*
+	 * If the callout isn't pending, it's not on the queue, so
+	 * don't attempt to remove it from the queue.  We can try to
+	 * stop it by other means however.
+	 */
+	if (!(c->c_flags & CALLOUT_PENDING)) {
+		c->c_flags &= ~CALLOUT_ACTIVE;
 
 		/*
-		 * We've gotten our callout CPU lock, it is safe to
-		 * drop the initial lock:
+		 * If it wasn't on the queue and it isn't the current
+		 * callout, then we can't stop it, so just bail.
 		 */
-		callout_unlock_client(c->c_flags, c->c_lock);
-
-		/* Wait for drain to complete */
+		if (cc->cc_exec_entity[direct].cc_curr != c) {
+			CTR3(KTR_CALLOUT, "failed to stop %p func %p arg %p",
+			    c, c->c_func, c->c_arg);
+			CC_UNLOCK(cc);
+			if (sq_locked)
+				sleepq_release(
+				    &cc->cc_exec_entity[direct].cc_waiting);
+			return (0);
+		}
 
-		while (cc->cc_exec_entity[direct].cc_curr == c)
-			msleep_spin(c, (struct mtx *)&cc->cc_lock, "codrain", 0);
+		if (safe) {
+			/*
+			 * The current callout is running (or just
+			 * about to run) and blocking is allowed, so
+			 * just wait for the current invocation to
+			 * finish.
+			 */
+			while (cc->cc_exec_entity[direct].cc_curr == c) {
+				/*
+				 * Use direct calls to sleepqueue interface
+				 * instead of cv/msleep in order to avoid
+				 * a LOR between cc_lock and sleepqueue
+				 * chain spinlocks.  This piece of code
+				 * emulates a msleep_spin() call actually.
+				 *
+				 * If we already have the sleepqueue chain
+				 * locked, then we can safely block.  If we
+				 * don't already have it locked, however,
+				 * we have to drop the cc_lock to lock
+				 * it.  This opens several races, so we
+				 * restart at the beginning once we have
+				 * both locks.  If nothing has changed, then
+				 * we will end up back here with sq_locked
+				 * set.
+				 */
+				if (!sq_locked) {
+					CC_UNLOCK(cc);
+					sleepq_lock(
+					&cc->cc_exec_entity[direct].cc_waiting);
+					sq_locked = 1;
+					old_cc = cc;
+					goto again;
+				}
 
+				/*
+				 * Migration could be cancelled here, but
+				 * as long as it is still not sure when it
+				 * will be packed up, just let softclock()
+				 * take care of it.
+				 */
+				cc->cc_exec_entity[direct].cc_waiting = true;
+				DROP_GIANT();
+				CC_UNLOCK(cc);
+				sleepq_add(
+				    &cc->cc_exec_entity[direct].cc_waiting,
+				    &cc->cc_lock.lock_object, "codrain",
+				    SLEEPQ_SLEEP, 0);
+				sleepq_wait(
+				    &cc->cc_exec_entity[direct].cc_waiting,
+					     0);
+				sq_locked = 0;
+				old_cc = NULL;
+
+				/* Reacquire locks previously released. */
+				PICKUP_GIANT();
+				CC_LOCK(cc);
+			}
+		} else if (use_lock &&
+			    !cc->cc_exec_entity[direct].cc_cancel) {
+			/*
+			 * The current callout is waiting for its
+			 * lock which we hold.  Cancel the callout
+			 * and return.  After our caller drops the
+			 * lock, the callout will be skipped in
+			 * softclock().
+			 */
+			cc->cc_exec_entity[direct].cc_cancel = true;
+			CTR3(KTR_CALLOUT, "cancelled %p func %p arg %p",
+			    c, c->c_func, c->c_arg);
+			KASSERT(!cc_cce_migrating(cc, direct),
+			    ("callout wrongly scheduled for migration"));
+			CC_UNLOCK(cc);
+			KASSERT(!sq_locked, ("sleepqueue chain locked"));
+			return (1);
+		} else if ((c->c_flags & CALLOUT_DFRMIGRATION) != 0) {
+			c->c_flags &= ~CALLOUT_DFRMIGRATION;
+			CTR3(KTR_CALLOUT, "postponing stop %p func %p arg %p",
+			    c, c->c_func, c->c_arg);
+			CC_UNLOCK(cc);
+			return (1);
+		}
+		CTR3(KTR_CALLOUT, "failed to stop %p func %p arg %p",
+		    c, c->c_func, c->c_arg);
 		CC_UNLOCK(cc);
-	} else {
-		callout_unlock_client(c->c_flags, c->c_lock);
+		KASSERT(!sq_locked, ("sleepqueue chain still locked"));
+		return (0);
 	}
+	if (sq_locked)
+		sleepq_release(&cc->cc_exec_entity[direct].cc_waiting);
+
+	c->c_flags &= ~(CALLOUT_ACTIVE | CALLOUT_PENDING);
 
 	CTR3(KTR_CALLOUT, "cancelled %p func %p arg %p",
 	    c, c->c_func, c->c_arg);
+	if ((c->c_flags & CALLOUT_PROCESSED) == 0) {
+		if (cc->cc_exec_next_dir == c)
+			cc->cc_exec_next_dir = LIST_NEXT(c, c_links.le);
+		LIST_REMOVE(c, c_links.le);
+	} else
+		TAILQ_REMOVE(&cc->cc_expireq, c, c_links.tqe);
+	callout_cc_del(c, cc);
 
-	return (cancelled & 1);
+	CC_UNLOCK(cc);
+	return (1);
 }
 
 void
 callout_init(struct callout *c, int mpsafe)
 {
+	bzero(c, sizeof *c);
 	if (mpsafe) {
-		_callout_init_lock(c, NULL, CALLOUT_RETURNUNLOCKED);
+		c->c_lock = NULL;
+		c->c_flags = CALLOUT_RETURNUNLOCKED;
 	} else {
-		_callout_init_lock(c, &Giant.lock_object, 0);
+		c->c_lock = &Giant.lock_object;
+		c->c_flags = 0;
 	}
+	c->c_cpu = timeout_cpu;
 }
 
 void
 _callout_init_lock(struct callout *c, struct lock_object *lock, int flags)
 {
 	bzero(c, sizeof *c);
-	KASSERT((flags & ~CALLOUT_RETURNUNLOCKED) == 0,
-	    ("callout_init_lock: bad flags 0x%08x", flags));
-	flags &= CALLOUT_RETURNUNLOCKED;
-	if (lock != NULL) {
-		struct lock_class *class = LOCK_CLASS(lock);
-		if (class == &lock_class_mtx_sleep)
-			flags |= CALLOUT_SET_LC(CALLOUT_LC_MUTEX);
-		else if (class == &lock_class_mtx_spin)
-			flags |= CALLOUT_SET_LC(CALLOUT_LC_SPIN);
-		else if (class == &lock_class_rm)
-			flags |= CALLOUT_SET_LC(CALLOUT_LC_RM);
-		else if (class == &lock_class_rw)
-			flags |= CALLOUT_SET_LC(CALLOUT_LC_RW);
-		else
-			panic("callout_init_lock: Unsupported lock class '%s'\n", class->lc_name);
-	} else {
-		flags |= CALLOUT_SET_LC(CALLOUT_LC_UNUSED_0);
-	}
 	c->c_lock = lock;
-	c->c_flags = flags;
+	KASSERT((flags & ~(CALLOUT_RETURNUNLOCKED | CALLOUT_SHAREDLOCK)) == 0,
+	    ("callout_init_lock: bad flags %d", flags));
+	KASSERT(lock != NULL || (flags & CALLOUT_RETURNUNLOCKED) == 0,
+	    ("callout_init_lock: CALLOUT_RETURNUNLOCKED with no lock"));
+	KASSERT(lock == NULL || !(LOCK_CLASS(lock)->lc_flags &
+	    (LC_SPINLOCK | LC_SLEEPABLE)), ("%s: invalid lock class",
+	    __func__));
+	c->c_flags = flags & (CALLOUT_RETURNUNLOCKED | CALLOUT_SHAREDLOCK);
 	c->c_cpu = timeout_cpu;
 }