Add an EARLY_AP_STARTUP option to start APs earlier during boot.

Currently, Application Processors (non-boot CPUs) are started by
MD code at SI_SUB_CPU, but they are kept waiting in a "pen" until
SI_SUB_SMP at which point they are released to run kernel threads.
SI_SUB_SMP is one of the last SYSINIT levels, so APs don't enter
the scheduler and start running threads until fairly late in the
boot.

This change moves SI_SUB_SMP up to just before software interrupt
threads are created allowing the APs to start executing kernel
threads much sooner (before any devices are probed).  This allows
several initialization routines that need to perform initialization
on all CPUs to now perform that initialization in one step rather
than having to defer the AP initialization to a second SYSINIT run
at SI_SUB_SMP.  It also permits all CPUs to be available for
handling interrupts before any devices are probed.

This last feature fixes a problem on with interrupt vector exhaustion.
Specifically, in the old model all device interrupts were routed
onto the boot CPU during boot.  Later after the APs were released at
SI_SUB_SMP, interrupts were redistributed across all CPUs.

However, several drivers for multiqueue hardware allocate N interrupts
per CPU in the system.  In a system with many CPUs, just a few drivers
doing this could exhaust the available pool of interrupt vectors on
the boot CPU as each driver was allocating N * mp_ncpu vectors on the
boot CPU.  Now, drivers will allocate interrupts on their desired CPUs
during boot meaning that only N interrupts are allocated from the boot
CPU instead of N * mp_ncpu.

Some other bits of code can also be simplified as smp_started is
now true much earlier and will now always be true for these bits of
code.  This removes the need to treat the single-CPU boot environment
as a special case.

As a transition aid, the new behavior is available under a new kernel
option (EARLY_AP_STARTUP).  This will allow the option to be turned off
if need be during initial testing.  I plan to enable this on x86 by
default in a followup commit in the next few days and to have all
platforms moved over before 11.0.  Once the transition is complete,
the option will be removed along with the !EARLY_AP_STARTUP code.

These changes have only been tested on x86.  Other platform maintainers
are encouraged to port their architectures over as well.  The main
things to check for are any uses of smp_started in MD code that can be
simplified and SI_SUB_SMP SYSINITs in MD code that can be removed in
the EARLY_AP_STARTUP case (e.g. the interrupt shuffling).

PR:		kern/199321
Reviewed by:	markj, gnn, kib
Sponsored by:	Netflix

22 files changed, 256 insertions, 28 deletions

diff --git a/sys/cddl/dev/dtrace/amd64/dtrace_subr.c b/sys/cddl/dev/dtrace/amd64/dtrace_subr.c
index 1d81d42..df39d2e 100644
--- a/sys/cddl/dev/dtrace/amd64/dtrace_subr.c
+++ b/sys/cddl/dev/dtrace/amd64/dtrace_subr.c
@@ -246,6 +246,26 @@ static uint64_t	nsec_scale;
 /* See below for the explanation of this macro. */
 #define SCALE_SHIFT	28
 
+static void
+dtrace_gethrtime_init_cpu(void *arg)
+{
+	uintptr_t cpu = (uintptr_t) arg;
+
+	if (cpu == curcpu)
+		tgt_cpu_tsc = rdtsc();
+	else
+		hst_cpu_tsc = rdtsc();
+}
+
+#ifdef EARLY_AP_STARTUP
+static void
+dtrace_gethrtime_init(void *arg)
+{
+	struct pcpu *pc;
+	uint64_t tsc_f;
+	cpuset_t map;
+	int i;
+#else
 /*
  * Get the frequency and scale factor as early as possible so that they can be
  * used for boot-time tracing.
@@ -254,6 +274,7 @@ static void
 dtrace_gethrtime_init_early(void *arg)
 {
 	uint64_t tsc_f;
+#endif
 
 	/*
 	 * Get TSC frequency known at this moment.
@@ -282,27 +303,18 @@ dtrace_gethrtime_init_early(void *arg)
 	 *   (terahertz) values;
 	 */
 	nsec_scale = ((uint64_t)NANOSEC << SCALE_SHIFT) / tsc_f;
+#ifndef EARLY_AP_STARTUP
 }
 SYSINIT(dtrace_gethrtime_init_early, SI_SUB_CPU, SI_ORDER_ANY,
     dtrace_gethrtime_init_early, NULL);
 
 static void
-dtrace_gethrtime_init_cpu(void *arg)
-{
-	uintptr_t cpu = (uintptr_t) arg;
-
-	if (cpu == curcpu)
-		tgt_cpu_tsc = rdtsc();
-	else
-		hst_cpu_tsc = rdtsc();
-}
-
-static void
 dtrace_gethrtime_init(void *arg)
 {
 	struct pcpu *pc;
 	cpuset_t map;
 	int i;
+#endif
 
 	/* The current CPU is the reference one. */
 	sched_pin();
@@ -323,8 +335,13 @@ dtrace_gethrtime_init(void *arg)
 	}
 	sched_unpin();
 }
+#ifdef EARLY_AP_STARTUP
+SYSINIT(dtrace_gethrtime_init, SI_SUB_DTRACE, SI_ORDER_ANY,
+    dtrace_gethrtime_init, NULL);
+#else
 SYSINIT(dtrace_gethrtime_init, SI_SUB_SMP, SI_ORDER_ANY, dtrace_gethrtime_init,
     NULL);
+#endif
 
 /*
  * DTrace needs a high resolution time function which can
diff --git a/sys/cddl/dev/dtrace/dtrace_load.c b/sys/cddl/dev/dtrace/dtrace_load.c
index c49de55..7f7dd79 100644
--- a/sys/cddl/dev/dtrace/dtrace_load.c
+++ b/sys/cddl/dev/dtrace/dtrace_load.c
@@ -22,6 +22,7 @@
  *
  */
 
+#ifndef EARLY_AP_STARTUP
 static void
 dtrace_ap_start(void *dummy)
 {
@@ -41,11 +42,15 @@ dtrace_ap_start(void *dummy)
 }
 
 SYSINIT(dtrace_ap_start, SI_SUB_SMP, SI_ORDER_ANY, dtrace_ap_start, NULL);
+#endif
 
 static void
 dtrace_load(void *dummy)
 {
 	dtrace_provider_id_t id;
+#ifdef EARLY_AP_STARTUP
+	int i;
+#endif
 
 	/* Hook into the trap handler. */
 	dtrace_trap_func = dtrace_trap;
@@ -142,8 +147,14 @@ dtrace_load(void *dummy)
 
 	mutex_enter(&cpu_lock);
 
+#ifdef EARLY_AP_STARTUP
+	CPU_FOREACH(i) {
+		(void) dtrace_cpu_setup(CPU_CONFIG, i);
+	}
+#else
 	/* Setup the boot CPU */
 	(void) dtrace_cpu_setup(CPU_CONFIG, 0);
+#endif
 
 	mutex_exit(&cpu_lock);
 
diff --git a/sys/cddl/dev/dtrace/i386/dtrace_subr.c b/sys/cddl/dev/dtrace/i386/dtrace_subr.c
index a2263ec..b70eab9 100644
--- a/sys/cddl/dev/dtrace/i386/dtrace_subr.c
+++ b/sys/cddl/dev/dtrace/i386/dtrace_subr.c
@@ -248,6 +248,26 @@ static uint64_t	nsec_scale;
 /* See below for the explanation of this macro. */
 #define SCALE_SHIFT	28
 
+static void
+dtrace_gethrtime_init_cpu(void *arg)
+{
+	uintptr_t cpu = (uintptr_t) arg;
+
+	if (cpu == curcpu)
+		tgt_cpu_tsc = rdtsc();
+	else
+		hst_cpu_tsc = rdtsc();
+}
+
+#ifdef EARLY_AP_STARTUP
+static void
+dtrace_gethrtime_init(void *arg)
+{
+	struct pcpu *pc;
+	uint64_t tsc_f;
+	cpuset_t map;
+	int i;
+#else
 /*
  * Get the frequency and scale factor as early as possible so that they can be
  * used for boot-time tracing.
@@ -256,6 +276,7 @@ static void
 dtrace_gethrtime_init_early(void *arg)
 {
 	uint64_t tsc_f;
+#endif
 
 	/*
 	 * Get TSC frequency known at this moment.
@@ -284,27 +305,18 @@ dtrace_gethrtime_init_early(void *arg)
 	 *   (terahertz) values;
 	 */
 	nsec_scale = ((uint64_t)NANOSEC << SCALE_SHIFT) / tsc_f;
+#ifndef EARLY_AP_STARTUP
 }
 SYSINIT(dtrace_gethrtime_init_early, SI_SUB_CPU, SI_ORDER_ANY,
     dtrace_gethrtime_init_early, NULL);
 
 static void
-dtrace_gethrtime_init_cpu(void *arg)
-{
-	uintptr_t cpu = (uintptr_t) arg;
-
-	if (cpu == curcpu)
-		tgt_cpu_tsc = rdtsc();
-	else
-		hst_cpu_tsc = rdtsc();
-}
-
-static void
 dtrace_gethrtime_init(void *arg)
 {
 	cpuset_t map;
 	struct pcpu *pc;
 	int i;
+#endif
 
 	/* The current CPU is the reference one. */
 	sched_pin();
@@ -325,8 +337,13 @@ dtrace_gethrtime_init(void *arg)
 	}
 	sched_unpin();
 }
+#ifdef EARLY_AP_STARTUP
+SYSINIT(dtrace_gethrtime_init, SI_SUB_DTRACE, SI_ORDER_ANY,
+    dtrace_gethrtime_init, NULL);
+#else
 SYSINIT(dtrace_gethrtime_init, SI_SUB_SMP, SI_ORDER_ANY, dtrace_gethrtime_init,
     NULL);
+#endif
 
 /*
  * DTrace needs a high resolution time function which can
diff --git a/sys/cddl/dev/dtrace/powerpc/dtrace_subr.c b/sys/cddl/dev/dtrace/powerpc/dtrace_subr.c
index 3546147..310a553 100644
--- a/sys/cddl/dev/dtrace/powerpc/dtrace_subr.c
+++ b/sys/cddl/dev/dtrace/powerpc/dtrace_subr.c
@@ -218,8 +218,13 @@ dtrace_gethrtime_init(void *arg)
 	}
 	sched_unpin();
 }
-
-SYSINIT(dtrace_gethrtime_init, SI_SUB_SMP, SI_ORDER_ANY, dtrace_gethrtime_init, NULL);
+#ifdef EARLY_AP_STARTUP
+SYSINIT(dtrace_gethrtime_init, SI_SUB_DTRACE, SI_ORDER_ANY,
+    dtrace_gethrtime_init, NULL);
+#else
+SYSINIT(dtrace_gethrtime_init, SI_SUB_SMP, SI_ORDER_ANY, dtrace_gethrtime_init,
+    NULL);
+#endif
 
 /*
  * DTrace needs a high resolution time function which can
diff --git a/sys/conf/NOTES b/sys/conf/NOTES
index 51e81f1..45732da 100644
--- a/sys/conf/NOTES
+++ b/sys/conf/NOTES
@@ -223,6 +223,12 @@ options 	SCHED_STATS
 # Mandatory:
 options 	SMP			# Symmetric MultiProcessor Kernel
 
+# EARLY_AP_STARTUP releases the Application Processors earlier in the
+# kernel startup process (before devices are probed) rather than at the
+# end.  This is a temporary option for use during the transition from
+# late to early AP startup.
+options		EARLY_AP_STARTUP
+
 # MAXCPU defines the maximum number of CPUs that can boot in the system.
 # A default value should be already present, for every architecture.
 options 	MAXCPU=32
diff --git a/sys/conf/options b/sys/conf/options
index 6dd51d7..f06a092 100644
--- a/sys/conf/options
+++ b/sys/conf/options
@@ -620,6 +620,7 @@ DEBUG_MEMGUARD		opt_vm.h
 DEBUG_REDZONE		opt_vm.h
 
 # Standard SMP options
+EARLY_AP_STARTUP	opt_global.h
 SMP			opt_global.h
 
 # Size of the kernel message buffer
diff --git a/sys/dev/acpica/acpi.c b/sys/dev/acpica/acpi.c
index b29b3c4..cbba9bf 100644
--- a/sys/dev/acpica/acpi.c
+++ b/sys/dev/acpica/acpi.c
@@ -2856,11 +2856,18 @@ acpi_EnterSleepState(struct acpi_softc *sc, int state)
     stop_all_proc();
     EVENTHANDLER_INVOKE(power_suspend);
 
+#ifdef EARLY_AP_STARTUP
+    MPASS(mp_ncpus == 1 || smp_started);
+    thread_lock(curthread);
+    sched_bind(curthread, 0);
+    thread_unlock(curthread);
+#else
     if (smp_started) {
 	thread_lock(curthread);
 	sched_bind(curthread, 0);
 	thread_unlock(curthread);
     }
+#endif
 
     /*
      * Be sure to hold Giant across DEVICE_SUSPEND/RESUME since non-MPSAFE
@@ -2991,11 +2998,17 @@ backout:
 
     mtx_unlock(&Giant);
 
+#ifdef EARLY_AP_STARTUP
+    thread_lock(curthread);
+    sched_unbind(curthread);
+    thread_unlock(curthread);
+#else
     if (smp_started) {
 	thread_lock(curthread);
 	sched_unbind(curthread);
 	thread_unlock(curthread);
     }
+#endif
 
     resume_all_proc();
 
diff --git a/sys/dev/acpica/acpi_cpu.c b/sys/dev/acpica/acpi_cpu.c
index 218da4e..00c951f 100644
--- a/sys/dev/acpica/acpi_cpu.c
+++ b/sys/dev/acpica/acpi_cpu.c
@@ -439,8 +439,12 @@ acpi_cpu_postattach(void *unused __unused)
     free(devices, M_TEMP);
 
     if (attached) {
+#ifdef EARLY_AP_STARTUP
+	acpi_cpu_startup(NULL);
+#else
 	/* Queue post cpu-probing task handler */
 	AcpiOsExecute(OSL_NOTIFY_HANDLER, acpi_cpu_startup, NULL);
+#endif
     }
 }
 
diff --git a/sys/dev/hwpmc/hwpmc_mod.c b/sys/dev/hwpmc/hwpmc_mod.c
index c0caab6..48bb23e 100644
--- a/sys/dev/hwpmc/hwpmc_mod.c
+++ b/sys/dev/hwpmc/hwpmc_mod.c
@@ -334,7 +334,11 @@ static moduledata_t pmc_mod = {
 	&pmc_syscall_mod
 };
 
+#ifdef EARLY_AP_STARTUP
+DECLARE_MODULE(pmc, pmc_mod, SI_SUB_SYSCALLS, SI_ORDER_ANY);
+#else
 DECLARE_MODULE(pmc, pmc_mod, SI_SUB_SMP, SI_ORDER_ANY);
+#endif
 MODULE_VERSION(pmc, PMC_VERSION);
 
 #ifdef	HWPMC_DEBUG
diff --git a/sys/dev/hyperv/vmbus/hv_vmbus_drv_freebsd.c b/sys/dev/hyperv/vmbus/hv_vmbus_drv_freebsd.c
index bcdf0cb..d0bdfc4 100644
--- a/sys/dev/hyperv/vmbus/hv_vmbus_drv_freebsd.c
+++ b/sys/dev/hyperv/vmbus/hv_vmbus_drv_freebsd.c
@@ -519,6 +519,7 @@ vmbus_attach(device_t dev)
 		device_printf(dev, "VMBUS: attach dev: %p\n", dev);
 	vmbus_devp = dev;
 
+#ifndef EARLY_AP_STARTUP
 	/* 
 	 * If the system has already booted and thread
 	 * scheduling is possible indicated by the global
@@ -526,6 +527,7 @@ vmbus_attach(device_t dev)
 	 * initialization directly.
 	 */
 	if (!cold)
+#endif
 		vmbus_bus_init();
 
 	bus_generic_probe(dev);
@@ -538,6 +540,7 @@ vmbus_init(void)
 	if (vm_guest != VM_GUEST_HV)
 		return;
 
+#ifndef EARLY_AP_STARTUP
 	/* 
 	 * If the system has already booted and thread
 	 * scheduling is possible, as indicated by the
@@ -545,6 +548,7 @@ vmbus_init(void)
 	 * initialization directly.
 	 */
 	if (!cold) 
+#endif
 		vmbus_bus_init();
 }
 
@@ -611,6 +615,9 @@ vmbus_modevent(module_t mod, int what, void *arg)
 	switch (what) {
 
 	case MOD_LOAD:
+#ifdef EARLY_AP_STARTUP
+		vmbus_init();
+#endif
 		vmbus_mod_load();
 		break;
 	case MOD_UNLOAD:
@@ -649,6 +656,7 @@ DRIVER_MODULE(vmbus, acpi, vmbus_driver, vmbus_devclass, vmbus_modevent, 0);
 MODULE_DEPEND(vmbus, acpi, 1, 1, 1);
 MODULE_VERSION(vmbus, 1);
 
+#ifndef EARLY_AP_STARTUP
 /* We want to be started after SMP is initialized */
 SYSINIT(vmb_init, SI_SUB_SMP + 1, SI_ORDER_FIRST, vmbus_init, NULL);
-
+#endif
diff --git a/sys/dev/xen/control/control.c b/sys/dev/xen/control/control.c
index a62b2ca..ae13c6c 100644
--- a/sys/dev/xen/control/control.c
+++ b/sys/dev/xen/control/control.c
@@ -202,11 +202,18 @@ xctrl_suspend()
 	stop_all_proc();
 	EVENTHANDLER_INVOKE(power_suspend);
 
+#ifdef EARLY_AP_STARTUP
+	MPASS(mp_ncpus == 1 || smp_started);
+	thread_lock(curthread);
+	sched_bind(curthread, 0);
+	thread_unlock(curthread);
+#else
 	if (smp_started) {
 		thread_lock(curthread);
 		sched_bind(curthread, 0);
 		thread_unlock(curthread);
 	}
+#endif
 	KASSERT((PCPU_GET(cpuid) == 0), ("Not running on CPU#0"));
 
 	/*
@@ -227,6 +234,17 @@ xctrl_suspend()
 	}
 
 #ifdef SMP
+#ifdef EARLY_AP_STARTUP
+	/*
+	 * Suspend other CPUs. This prevents IPIs while we
+	 * are resuming, and will allow us to reset per-cpu
+	 * vcpu_info on resume.
+	 */
+	cpu_suspend_map = all_cpus;
+	CPU_CLR(PCPU_GET(cpuid), &cpu_suspend_map);
+	if (!CPU_EMPTY(&cpu_suspend_map))
+		suspend_cpus(cpu_suspend_map);
+#else
 	CPU_ZERO(&cpu_suspend_map);	/* silence gcc */
 	if (smp_started) {
 		/*
@@ -240,6 +258,7 @@ xctrl_suspend()
 			suspend_cpus(cpu_suspend_map);
 	}
 #endif
+#endif
 
 	/*
 	 * Prevent any races with evtchn_interrupt() handler.
@@ -285,11 +304,17 @@ xctrl_suspend()
 	timecounter->tc_get_timecount(timecounter);
 	inittodr(time_second);
 
+#ifdef EARLY_AP_STARTUP
+	thread_lock(curthread);
+	sched_unbind(curthread);
+	thread_unlock(curthread);
+#else
 	if (smp_started) {
 		thread_lock(curthread);
 		sched_unbind(curthread);
 		thread_unlock(curthread);
 	}
+#endif
 
 	resume_all_proc();
 
diff --git a/sys/geom/eli/g_eli.c b/sys/geom/eli/g_eli.c
index 912a5c5..403d0b6 100644
--- a/sys/geom/eli/g_eli.c
+++ b/sys/geom/eli/g_eli.c
@@ -479,7 +479,9 @@ g_eli_worker(void *arg)
 
 	wr = arg;
 	sc = wr->w_softc;
-#ifdef SMP
+#ifdef EARLY_AP_STARTUP
+	MPASS(!sc->sc_cpubind || smp_started);
+#elif defined(SMP)
 	/* Before sched_bind() to a CPU, wait for all CPUs to go on-line. */
 	if (sc->sc_cpubind) {
 		while (!smp_started)
diff --git a/sys/kern/kern_clock.c b/sys/kern/kern_clock.c
index 1f4b129..17b85bc 100644
--- a/sys/kern/kern_clock.c
+++ b/sys/kern/kern_clock.c
@@ -391,6 +391,10 @@ static void
 initclocks(dummy)
 	void *dummy;
 {
+#ifdef EARLY_AP_STARTUP
+	struct proc *p;
+	struct thread *td;
+#endif
 	register int i;
 
 	/*
@@ -415,6 +419,35 @@ initclocks(dummy)
 	 * sign problems sooner.
 	 */
 	ticks = INT_MAX - (hz * 10 * 60);
+
+#ifdef EARLY_AP_STARTUP
+	/*
+	 * Fixup the tick counts in any blocked or sleeping threads to
+	 * account for the jump above.
+	 */
+	sx_slock(&allproc_lock);
+	FOREACH_PROC_IN_SYSTEM(p) {
+		PROC_LOCK(p);
+		if (p->p_state == PRS_NEW) {
+			PROC_UNLOCK(p);
+			continue;
+		}
+		FOREACH_THREAD_IN_PROC(p, td) {
+			thread_lock(td);
+			if (TD_ON_LOCK(td)) {
+				MPASS(td->td_blktick == 0);
+				td->td_blktick = ticks;
+			}
+			if (TD_ON_SLEEPQ(td)) {
+				MPASS(td->td_slptick == 0);
+				td->td_slptick = ticks;
+			}
+			thread_unlock(td);
+		}
+		PROC_UNLOCK(p);
+	}
+	sx_sunlock(&allproc_lock);
+#endif
 }
 
 /*
diff --git a/sys/kern/kern_clocksource.c b/sys/kern/kern_clocksource.c
index cc7847c..7f7769d 100644
--- a/sys/kern/kern_clocksource.c
+++ b/sys/kern/kern_clocksource.c
@@ -322,9 +322,16 @@ timercb(struct eventtimer *et, void *arg)
 	    curcpu, (int)(now >> 32), (u_int)(now & 0xffffffff));
 
 #ifdef SMP
+#ifdef EARLY_AP_STARTUP
+	MPASS(mp_ncpus == 1 || smp_started);
+#endif
 	/* Prepare broadcasting to other CPUs for non-per-CPU timers. */
 	bcast = 0;
+#ifdef EARLY_AP_STARTUP
+	if ((et->et_flags & ET_FLAGS_PERCPU) == 0) {
+#else
 	if ((et->et_flags & ET_FLAGS_PERCPU) == 0 && smp_started) {
+#endif
 		CPU_FOREACH(cpu) {
 			state = DPCPU_ID_PTR(cpu, timerstate);
 			ET_HW_LOCK(state);
@@ -485,12 +492,17 @@ configtimer(int start)
 			nexttick = next;
 		else
 			nexttick = -1;
+#ifdef EARLY_AP_STARTUP
+		MPASS(mp_ncpus == 1 || smp_started);
+#endif
 		CPU_FOREACH(cpu) {
 			state = DPCPU_ID_PTR(cpu, timerstate);
 			state->now = now;
+#ifndef EARLY_AP_STARTUP
 			if (!smp_started && cpu != CPU_FIRST())
 				state->nextevent = SBT_MAX;
 			else
+#endif
 				state->nextevent = next;
 			if (periodic)
 				state->nexttick = next;
@@ -513,8 +525,13 @@ configtimer(int start)
 	}
 	ET_HW_UNLOCK(DPCPU_PTR(timerstate));
 #ifdef SMP
+#ifdef EARLY_AP_STARTUP
+	/* If timer is global we are done. */
+	if ((timer->et_flags & ET_FLAGS_PERCPU) == 0) {
+#else
 	/* If timer is global or there is no other CPUs yet - we are done. */
 	if ((timer->et_flags & ET_FLAGS_PERCPU) == 0 || !smp_started) {
+#endif
 		critical_exit();
 		return;
 	}
diff --git a/sys/kern/kern_cpu.c b/sys/kern/kern_cpu.c
index 293b940..4e66b13 100644
--- a/sys/kern/kern_cpu.c
+++ b/sys/kern/kern_cpu.c
@@ -259,6 +259,9 @@ cf_set_method(device_t dev, const struct cf_level *level, int priority)
 	CF_MTX_LOCK(&sc->lock);
 
 #ifdef SMP
+#ifdef EARLY_AP_STARTUP
+	MPASS(mp_ncpus == 1 || smp_started);
+#else
 	/*
 	 * If still booting and secondary CPUs not started yet, don't allow
 	 * changing the frequency until they're online.  This is because we
@@ -271,6 +274,7 @@ cf_set_method(device_t dev, const struct cf_level *level, int priority)
 		error = ENXIO;
 		goto out;
 	}
+#endif
 #endif /* SMP */
 
 	/*
diff --git a/sys/net/netisr.c b/sys/net/netisr.c
index 14de112..492a851 100644
--- a/sys/net/netisr.c
+++ b/sys/net/netisr.c
@@ -1119,6 +1119,10 @@ netisr_start_swi(u_int cpuid, struct pcpu *pc)
 static void
 netisr_init(void *arg)
 {
+#ifdef EARLY_AP_STARTUP
+	struct pcpu *pc;
+#endif
+
 	KASSERT(curcpu == 0, ("%s: not on CPU 0", __func__));
 
 	NETISR_LOCK_INIT();
@@ -1149,10 +1153,20 @@ netisr_init(void *arg)
 		netisr_bindthreads = 0;
 	}
 #endif
+
+#ifdef EARLY_AP_STARTUP
+	STAILQ_FOREACH(pc, &cpuhead, pc_allcpu) {
+		if (nws_count >= netisr_maxthreads)
+			break;
+		netisr_start_swi(pc->pc_cpuid, pc);
+	}
+#else
 	netisr_start_swi(curcpu, pcpu_find(curcpu));
+#endif
 }
 SYSINIT(netisr_init, SI_SUB_SOFTINTR, SI_ORDER_FIRST, netisr_init, NULL);
 
+#ifndef EARLY_AP_STARTUP
 /*
  * Start worker threads for additional CPUs.  No attempt to gracefully handle
  * work reassignment, we don't yet support dynamic reconfiguration.
@@ -1172,6 +1186,7 @@ netisr_start(void *arg)
 	}
 }
 SYSINIT(netisr_start, SI_SUB_SMP, SI_ORDER_MIDDLE, netisr_start, NULL);
+#endif
 
 /*
  * Sysctl monitoring for netisr: query a list of registered protocols.
diff --git a/sys/sys/kernel.h b/sys/sys/kernel.h
index 02628ba..c6dd848 100644
--- a/sys/sys/kernel.h
+++ b/sys/sys/kernel.h
@@ -118,7 +118,10 @@ enum sysinit_sub_id {
 	SI_SUB_SCHED_IDLE	= 0x2600000,	/* required idle procs */
 	SI_SUB_MBUF		= 0x2700000,	/* mbuf subsystem */
 	SI_SUB_INTR		= 0x2800000,	/* interrupt threads */
-	SI_SUB_SOFTINTR		= 0x2800001,	/* start soft interrupt thread */
+#ifdef EARLY_AP_STARTUP
+	SI_SUB_SMP		= 0x2900000,	/* start the APs*/
+#endif
+	SI_SUB_SOFTINTR		= 0x2A00000,	/* start soft interrupt thread */
 	SI_SUB_DEVFS		= 0x2F00000,	/* devfs ready for devices */
 	SI_SUB_INIT_IF		= 0x3000000,	/* prep for net interfaces */
 	SI_SUB_NETGRAPH		= 0x3010000,	/* Let Netgraph initialize */
@@ -154,7 +157,9 @@ enum sysinit_sub_id {
 	SI_SUB_KTHREAD_BUF	= 0xea00000,	/* buffer daemon*/
 	SI_SUB_KTHREAD_UPDATE	= 0xec00000,	/* update daemon*/
 	SI_SUB_KTHREAD_IDLE	= 0xee00000,	/* idle procs*/
+#ifndef EARLY_AP_STARTUP
 	SI_SUB_SMP		= 0xf000000,	/* start the APs*/
+#endif	
 	SI_SUB_RACCTD		= 0xf100000,	/* start racctd*/
 	SI_SUB_LAST		= 0xfffffff	/* final initialization */
 };
diff --git a/sys/x86/isa/clock.c b/sys/x86/isa/clock.c
index cced5e34..fa4ca5b 100644
--- a/sys/x86/isa/clock.c
+++ b/sys/x86/isa/clock.c
@@ -475,8 +475,27 @@ startrtclock()
 void
 cpu_initclocks(void)
 {
+#ifdef EARLY_AP_STARTUP
+	struct thread *td;
+	int i;
 
+	td = curthread;
 	cpu_initclocks_bsp();
+	CPU_FOREACH(i) {
+		if (i == 0)
+			continue;
+		thread_lock(td);
+		sched_bind(td, i);
+		thread_unlock(td);
+		cpu_initclocks_ap();
+	}
+	thread_lock(td);
+	if (sched_is_bound(td))
+		sched_unbind(td);
+	thread_unlock(td);
+#else
+	cpu_initclocks_bsp();
+#endif
 }
 
 static int
diff --git a/sys/x86/x86/intr_machdep.c b/sys/x86/x86/intr_machdep.c
index 36a3441..1bc3038 100644
--- a/sys/x86/x86/intr_machdep.c
+++ b/sys/x86/x86/intr_machdep.c
@@ -77,7 +77,7 @@ static struct mtx intr_table_lock;
 static struct mtx intrcnt_lock;
 static TAILQ_HEAD(pics_head, pic) pics;
 
-#ifdef SMP
+#if defined(SMP) && !defined(EARLY_AP_STARTUP)
 static int assign_cpu;
 #endif
 
@@ -320,11 +320,16 @@ intr_assign_cpu(void *arg, int cpu)
 	struct intsrc *isrc;
 	int error;
 
+#ifdef EARLY_AP_STARTUP
+	MPASS(mp_ncpus == 1 || smp_started);
+	if (cpu != NOCPU) {
+#else
 	/*
 	 * Don't do anything during early boot.  We will pick up the
 	 * assignment once the APs are started.
 	 */
 	if (assign_cpu && cpu != NOCPU) {
+#endif
 		isrc = arg;
 		mtx_lock(&intr_table_lock);
 		error = isrc->is_pic->pic_assign_cpu(isrc, cpu_apic_ids[cpu]);
@@ -502,9 +507,13 @@ intr_next_cpu(void)
 {
 	u_int apic_id;
 
+#ifdef EARLY_AP_STARTUP
+	MPASS(mp_ncpus == 1 || smp_started);
+#else
 	/* Leave all interrupts on the BSP during boot. */
 	if (!assign_cpu)
 		return (PCPU_GET(apic_id));
+#endif
 
 	mtx_lock_spin(&icu_lock);
 	apic_id = cpu_apic_ids[current_cpu];
@@ -546,6 +555,7 @@ intr_add_cpu(u_int cpu)
 	CPU_SET(cpu, &intr_cpus);
 }
 
+#ifndef EARLY_AP_STARTUP
 /*
  * Distribute all the interrupt sources among the available CPUs once the
  * AP's have been launched.
@@ -586,6 +596,7 @@ intr_shuffle_irqs(void *arg __unused)
 }
 SYSINIT(intr_shuffle_irqs, SI_SUB_SMP, SI_ORDER_SECOND, intr_shuffle_irqs,
     NULL);
+#endif
 #else
 /*
  * Always route interrupts to the current processor in the UP case.
diff --git a/sys/x86/x86/local_apic.c b/sys/x86/x86/local_apic.c
index 1a02626..0a77679 100644
--- a/sys/x86/x86/local_apic.c
+++ b/sys/x86/x86/local_apic.c
@@ -749,6 +749,10 @@ native_lapic_enable_pmc(void)
 
 	lvts[APIC_LVT_PMC].lvt_masked = 0;
 
+#ifdef EARLY_AP_STARTUP
+	MPASS(mp_ncpus == 1 || smp_started);
+	smp_rendezvous(NULL, lapic_update_pmc, NULL, NULL);
+#else
 #ifdef SMP
 	/*
 	 * If hwpmc was loaded at boot time then the APs may not be
@@ -760,6 +764,7 @@ native_lapic_enable_pmc(void)
 	else
 #endif
 		lapic_update_pmc(NULL);
+#endif
 	return (1);
 #else
 	return (0);
diff --git a/sys/x86/x86/mca.c b/sys/x86/x86/mca.c
index b3d6066..d005180 100644
--- a/sys/x86/x86/mca.c
+++ b/sys/x86/x86/mca.c
@@ -726,7 +726,11 @@ mca_startup(void *dummy)
 
 	callout_reset(&mca_timer, mca_ticks * hz, mca_periodic_scan, NULL);
 }
+#ifdef EARLY_AP_STARTUP
+SYSINIT(mca_startup, SI_SUB_KICK_SCHEDULER, SI_ORDER_ANY, mca_startup, NULL);
+#else
 SYSINIT(mca_startup, SI_SUB_SMP, SI_ORDER_ANY, mca_startup, NULL);
+#endif
 
 #ifdef DEV_APIC
 static void
diff --git a/sys/x86/x86/mp_x86.c b/sys/x86/x86/mp_x86.c
index 7ad9b96..d2eb2e9 100644
--- a/sys/x86/x86/mp_x86.c
+++ b/sys/x86/x86/mp_x86.c
@@ -933,8 +933,10 @@ init_secondary_tail(void)
 	while (atomic_load_acq_int(&smp_started) == 0)
 		ia32_pause();
 
+#ifndef EARLY_AP_STARTUP
 	/* Start per-CPU event timers. */
 	cpu_initclocks_ap();
+#endif
 
 	sched_throw(NULL);