xnu-3248.60.10.tar.gz

[apple/xnu.git] / osfmk / i386 / mp.c

/*
 * Copyright (c) 2000-2012 Apple Inc. All rights reserved.
 *
 * @APPLE_OSREFERENCE_LICENSE_HEADER_START@
 * 
 * This file contains Original Code and/or Modifications of Original Code
 * as defined in and that are subject to the Apple Public Source License
 * Version 2.0 (the 'License'). You may not use this file except in
 * compliance with the License. The rights granted to you under the License
 * may not be used to create, or enable the creation or redistribution of,
 * unlawful or unlicensed copies of an Apple operating system, or to
 * circumvent, violate, or enable the circumvention or violation of, any
 * terms of an Apple operating system software license agreement.
 * 
 * Please obtain a copy of the License at
 * http://www.opensource.apple.com/apsl/ and read it before using this file.
 * 
 * The Original Code and all software distributed under the License are
 * distributed on an 'AS IS' basis, WITHOUT WARRANTY OF ANY KIND, EITHER
 * EXPRESS OR IMPLIED, AND APPLE HEREBY DISCLAIMS ALL SUCH WARRANTIES,
 * INCLUDING WITHOUT LIMITATION, ANY WARRANTIES OF MERCHANTABILITY,
 * FITNESS FOR A PARTICULAR PURPOSE, QUIET ENJOYMENT OR NON-INFRINGEMENT.
 * Please see the License for the specific language governing rights and
 * limitations under the License.
 * 
 * @APPLE_OSREFERENCE_LICENSE_HEADER_END@
 */
/*
 * @OSF_COPYRIGHT@
 */

#include <mach_rt.h>
#include <mach_kdp.h>
#include <mach_ldebug.h>
#include <gprof.h>

#include <mach/mach_types.h>
#include <mach/kern_return.h>

#include <kern/kern_types.h>
#include <kern/startup.h>
#include <kern/timer_queue.h>
#include <kern/processor.h>
#include <kern/cpu_number.h>
#include <kern/cpu_data.h>
#include <kern/assert.h>
#include <kern/machine.h>
#include <kern/pms.h>
#include <kern/misc_protos.h>
#include <kern/timer_call.h>
#include <kern/kalloc.h>
#include <kern/queue.h>
#include <prng/random.h>

#include <vm/vm_map.h>
#include <vm/vm_kern.h>

#include <profiling/profile-mk.h>

#include <i386/bit_routines.h>
#include <i386/proc_reg.h>
#include <i386/cpu_threads.h>
#include <i386/mp_desc.h>
#include <i386/misc_protos.h>
#include <i386/trap.h>
#include <i386/postcode.h>
#include <i386/machine_routines.h>
#include <i386/mp.h>
#include <i386/mp_events.h>
#include <i386/lapic.h>
#include <i386/cpuid.h>
#include <i386/fpu.h>
#include <i386/machine_cpu.h>
#include <i386/pmCPU.h>
#if CONFIG_MCA
#include <i386/machine_check.h>
#endif
#include <i386/acpi.h>

#include <chud/chud_xnu.h>
#include <chud/chud_xnu_private.h>

#include <sys/kdebug.h>

#include <console/serial_protos.h>

#if     MP_DEBUG
#define PAUSE           delay(1000000)
#define DBG(x...)       kprintf(x)
#else
#define DBG(x...)
#define PAUSE
#endif  /* MP_DEBUG */

/* Debugging/test trace events: */
#define TRACE_MP_TLB_FLUSH              MACHDBG_CODE(DBG_MACH_MP, 0)
#define TRACE_MP_CPUS_CALL              MACHDBG_CODE(DBG_MACH_MP, 1)
#define TRACE_MP_CPUS_CALL_LOCAL        MACHDBG_CODE(DBG_MACH_MP, 2)
#define TRACE_MP_CPUS_CALL_ACTION       MACHDBG_CODE(DBG_MACH_MP, 3)
#define TRACE_MP_CPUS_CALL_NOBUF        MACHDBG_CODE(DBG_MACH_MP, 4)
#define TRACE_MP_CPU_FAST_START         MACHDBG_CODE(DBG_MACH_MP, 5)
#define TRACE_MP_CPU_START              MACHDBG_CODE(DBG_MACH_MP, 6)
#define TRACE_MP_CPU_DEACTIVATE         MACHDBG_CODE(DBG_MACH_MP, 7)

#define ABS(v)          (((v) > 0)?(v):-(v))

void            slave_boot_init(void);
void            i386_cpu_IPI(int cpu);

#if MACH_KDP
static void     mp_kdp_wait(boolean_t flush, boolean_t isNMI);
#endif /* MACH_KDP */
static void     mp_rendezvous_action(void);
static void     mp_broadcast_action(void);

#if MACH_KDP
static boolean_t        cpu_signal_pending(int cpu, mp_event_t event);
#endif /* MACH_KDP */
static int              NMIInterruptHandler(x86_saved_state_t *regs);

boolean_t               smp_initialized = FALSE;
uint32_t                TSC_sync_margin = 0xFFF;
volatile boolean_t      force_immediate_debugger_NMI = FALSE;
volatile boolean_t      pmap_tlb_flush_timeout = FALSE;
decl_simple_lock_data(,mp_kdp_lock);

decl_lck_mtx_data(static, mp_cpu_boot_lock);
lck_mtx_ext_t   mp_cpu_boot_lock_ext;

/* Variables needed for MP rendezvous. */
decl_simple_lock_data(,mp_rv_lock);
static void     (*mp_rv_setup_func)(void *arg);
static void     (*mp_rv_action_func)(void *arg);
static void     (*mp_rv_teardown_func)(void *arg);
static void     *mp_rv_func_arg;
static volatile int     mp_rv_ncpus;
                        /* Cache-aligned barriers: */
static volatile long    mp_rv_entry    __attribute__((aligned(64)));
static volatile long    mp_rv_exit     __attribute__((aligned(64)));
static volatile long    mp_rv_complete __attribute__((aligned(64)));

volatile        uint64_t        debugger_entry_time;
volatile        uint64_t        debugger_exit_time;
#if MACH_KDP
#include <kdp/kdp.h>
extern int kdp_snapshot;
static struct _kdp_xcpu_call_func {
        kdp_x86_xcpu_func_t func;
        void     *arg0, *arg1;
        volatile long     ret;
        volatile uint16_t cpu;
} kdp_xcpu_call_func = {
        .cpu  = KDP_XCPU_NONE
};

#endif

/* Variables needed for MP broadcast. */
static void        (*mp_bc_action_func)(void *arg);
static void        *mp_bc_func_arg;
static int      mp_bc_ncpus;
static volatile long   mp_bc_count;
decl_lck_mtx_data(static, mp_bc_lock);
lck_mtx_ext_t   mp_bc_lock_ext;
static  volatile int    debugger_cpu = -1;
volatile long    NMIPI_acks = 0;
volatile long    NMI_count = 0;

extern void     NMI_cpus(void);

static void     mp_cpus_call_init(void); 
static void     mp_cpus_call_action(void); 
static void     mp_call_PM(void);

static boolean_t        mp_cpus_call_wait_timeout = FALSE;

char            mp_slave_stack[PAGE_SIZE] __attribute__((aligned(PAGE_SIZE))); // Temp stack for slave init

/* PAL-related routines */
boolean_t i386_smp_init(int nmi_vector, i386_intr_func_t nmi_handler, 
                int ipi_vector, i386_intr_func_t ipi_handler);
void i386_start_cpu(int lapic_id, int cpu_num);
void i386_send_NMI(int cpu);

#if GPROF
/*
 * Initialize dummy structs for profiling. These aren't used but
 * allows hertz_tick() to be built with GPROF defined.
 */
struct profile_vars _profile_vars;
struct profile_vars *_profile_vars_cpus[MAX_CPUS] = { &_profile_vars };
#define GPROF_INIT()                                                    \
{                                                                       \
        int     i;                                                      \
                                                                        \
        /* Hack to initialize pointers to unused profiling structs */   \
        for (i = 1; i < MAX_CPUS; i++)                          \
                _profile_vars_cpus[i] = &_profile_vars;                 \
}
#else
#define GPROF_INIT()
#endif /* GPROF */

static lck_grp_t        smp_lck_grp;
static lck_grp_attr_t   smp_lck_grp_attr;

#define NUM_CPU_WARM_CALLS      20
struct timer_call       cpu_warm_call_arr[NUM_CPU_WARM_CALLS];
queue_head_t            cpu_warm_call_list;
decl_simple_lock_data(static, cpu_warm_lock);

typedef struct cpu_warm_data {
        timer_call_t    cwd_call;
        uint64_t        cwd_deadline;
        int             cwd_result;
} *cpu_warm_data_t;

static void             cpu_prewarm_init(void);
static void             cpu_warm_timer_call_func(call_entry_param_t p0, call_entry_param_t p1);
static void             _cpu_warm_setup(void *arg);
static timer_call_t     grab_warm_timer_call(void);
static void             free_warm_timer_call(timer_call_t call);

void
smp_init(void)
{
        simple_lock_init(&mp_kdp_lock, 0);
        simple_lock_init(&mp_rv_lock, 0);
        lck_grp_attr_setdefault(&smp_lck_grp_attr);
        lck_grp_init(&smp_lck_grp, "i386_smp", &smp_lck_grp_attr);
        lck_mtx_init_ext(&mp_cpu_boot_lock, &mp_cpu_boot_lock_ext, &smp_lck_grp, LCK_ATTR_NULL);
        lck_mtx_init_ext(&mp_bc_lock, &mp_bc_lock_ext, &smp_lck_grp, LCK_ATTR_NULL);
        console_init();

        if(!i386_smp_init(LAPIC_NMI_INTERRUPT, NMIInterruptHandler, 
                                LAPIC_VECTOR(INTERPROCESSOR), cpu_signal_handler))
                return;

        cpu_thread_init();

        GPROF_INIT();
        DBGLOG_CPU_INIT(master_cpu);

        mp_cpus_call_init();
        mp_cpus_call_cpu_init(master_cpu);

        if (PE_parse_boot_argn("TSC_sync_margin",
                                        &TSC_sync_margin, sizeof(TSC_sync_margin))) {
                kprintf("TSC sync Margin 0x%x\n", TSC_sync_margin);
        } else if (cpuid_vmm_present()) {
                kprintf("TSC sync margin disabled\n");
                TSC_sync_margin = 0;
        }
        smp_initialized = TRUE;

        cpu_prewarm_init();

        return;
}

typedef struct {
        int                     target_cpu;
        int                     target_lapic;
        int                     starter_cpu;
} processor_start_info_t;
static processor_start_info_t   start_info        __attribute__((aligned(64)));

/* 
 * Cache-alignment is to avoid cross-cpu false-sharing interference.
 */
static volatile long            tsc_entry_barrier __attribute__((aligned(64)));
static volatile long            tsc_exit_barrier  __attribute__((aligned(64)));
static volatile uint64_t        tsc_target        __attribute__((aligned(64)));

/*
 * Poll a CPU to see when it has marked itself as running.
 */
static void
mp_wait_for_cpu_up(int slot_num, unsigned int iters, unsigned int usecdelay)
{
        while (iters-- > 0) {
                if (cpu_datap(slot_num)->cpu_running)
                        break;
                delay(usecdelay);
        }
}

/*
 * Quickly bring a CPU back online which has been halted.
 */
kern_return_t
intel_startCPU_fast(int slot_num)
{
        kern_return_t   rc;

        /*
         * Try to perform a fast restart
         */
        rc = pmCPUExitHalt(slot_num);
        if (rc != KERN_SUCCESS)
                /*
                 * The CPU was not eligible for a fast restart.
                 */
                return(rc);

        KERNEL_DEBUG_CONSTANT(
                TRACE_MP_CPU_FAST_START | DBG_FUNC_START,
                slot_num, 0, 0, 0, 0);

        /*
         * Wait until the CPU is back online.
         */
        mp_disable_preemption();
    
        /*
         * We use short pauses (1us) for low latency.  30,000 iterations is
         * longer than a full restart would require so it should be more
         * than long enough.
         */

        mp_wait_for_cpu_up(slot_num, 30000, 1);
        mp_enable_preemption();

        KERNEL_DEBUG_CONSTANT(
                TRACE_MP_CPU_FAST_START | DBG_FUNC_END,
                slot_num, cpu_datap(slot_num)->cpu_running, 0, 0, 0);

        /*
         * Check to make sure that the CPU is really running.  If not,
         * go through the slow path.
         */
        if (cpu_datap(slot_num)->cpu_running)
                return(KERN_SUCCESS);
        else
                return(KERN_FAILURE);
}

static void
started_cpu(void)
{
        /* Here on the started cpu with cpu_running set TRUE */

        if (TSC_sync_margin &&
            start_info.target_cpu == cpu_number()) {
                /*
                 * I've just started-up, synchronize again with the starter cpu
                 * and then snap my TSC.
                 */
                tsc_target   = 0;
                atomic_decl(&tsc_entry_barrier, 1);
                while (tsc_entry_barrier != 0)
                        ;       /* spin for starter and target at barrier */
                tsc_target = rdtsc64();
                atomic_decl(&tsc_exit_barrier, 1);
        }
}

static void
start_cpu(void *arg)
{
        int                     i = 1000;
        processor_start_info_t  *psip = (processor_start_info_t *) arg;

        /* Ignore this if the current processor is not the starter */
        if (cpu_number() != psip->starter_cpu)
                return;

        DBG("start_cpu(%p) about to start cpu %d, lapic %d\n",
                arg, psip->target_cpu, psip->target_lapic);

        KERNEL_DEBUG_CONSTANT(
                TRACE_MP_CPU_START | DBG_FUNC_START,
                psip->target_cpu,
                psip->target_lapic, 0, 0, 0);

        i386_start_cpu(psip->target_lapic, psip->target_cpu);

#ifdef  POSTCODE_DELAY
        /* Wait much longer if postcodes are displayed for a delay period. */
        i *= 10000;
#endif
        DBG("start_cpu(%p) about to wait for cpu %d\n",
                arg, psip->target_cpu);

        mp_wait_for_cpu_up(psip->target_cpu, i*100, 100);

        KERNEL_DEBUG_CONSTANT(
                TRACE_MP_CPU_START | DBG_FUNC_END,
                psip->target_cpu,
                cpu_datap(psip->target_cpu)->cpu_running, 0, 0, 0);

        if (TSC_sync_margin &&
            cpu_datap(psip->target_cpu)->cpu_running) {
                /*
                 * Compare the TSC from the started processor with ours.
                 * Report and log/panic if it diverges by more than
                 * TSC_sync_margin (TSC_SYNC_MARGIN) ticks. This margin
                 * can be overriden by boot-arg (with 0 meaning no checking).
                 */
                uint64_t        tsc_starter;
                int64_t         tsc_delta;
                atomic_decl(&tsc_entry_barrier, 1);
                while (tsc_entry_barrier != 0)
                        ;       /* spin for both processors at barrier */
                tsc_starter = rdtsc64();
                atomic_decl(&tsc_exit_barrier, 1);
                while (tsc_exit_barrier != 0)
                        ;       /* spin for target to store its TSC */
                tsc_delta = tsc_target - tsc_starter;
                kprintf("TSC sync for cpu %d: 0x%016llx delta 0x%llx (%lld)\n",
                        psip->target_cpu, tsc_target, tsc_delta, tsc_delta);
                if (ABS(tsc_delta) > (int64_t) TSC_sync_margin) { 
#if DEBUG
                        panic(
#else
                        printf(
#endif
                                "Unsynchronized  TSC for cpu %d: "
                                        "0x%016llx, delta 0x%llx\n",
                                psip->target_cpu, tsc_target, tsc_delta);
                }
        }
}

kern_return_t
intel_startCPU(
        int     slot_num)
{
        int             lapic = cpu_to_lapic[slot_num];
        boolean_t       istate;

        assert(lapic != -1);

        DBGLOG_CPU_INIT(slot_num);

        DBG("intel_startCPU(%d) lapic_id=%d\n", slot_num, lapic);
        DBG("IdlePTD(%p): 0x%x\n", &IdlePTD, (int) (uintptr_t)IdlePTD);

        /*
         * Initialize (or re-initialize) the descriptor tables for this cpu.
         * Propagate processor mode to slave.
         */
        cpu_desc_init64(cpu_datap(slot_num));

        /* Serialize use of the slave boot stack, etc. */
        lck_mtx_lock(&mp_cpu_boot_lock);

        istate = ml_set_interrupts_enabled(FALSE);
        if (slot_num == get_cpu_number()) {
                ml_set_interrupts_enabled(istate);
                lck_mtx_unlock(&mp_cpu_boot_lock);
                return KERN_SUCCESS;
        }

        start_info.starter_cpu  = cpu_number();
        start_info.target_cpu   = slot_num;
        start_info.target_lapic = lapic;
        tsc_entry_barrier = 2;
        tsc_exit_barrier = 2;

        /*
         * Perform the processor startup sequence with all running
         * processors rendezvous'ed. This is required during periods when
         * the cache-disable bit is set for MTRR/PAT initialization.
         */
        mp_rendezvous_no_intrs(start_cpu, (void *) &start_info);

        start_info.target_cpu = 0;

        ml_set_interrupts_enabled(istate);
        lck_mtx_unlock(&mp_cpu_boot_lock);

        if (!cpu_datap(slot_num)->cpu_running) {
                kprintf("Failed to start CPU %02d\n", slot_num);
                printf("Failed to start CPU %02d, rebooting...\n", slot_num);
                delay(1000000);
                halt_cpu();
                return KERN_SUCCESS;
        } else {
                kprintf("Started cpu %d (lapic id %08x)\n", slot_num, lapic);
                return KERN_SUCCESS;
        }
}

#if     MP_DEBUG
cpu_signal_event_log_t  *cpu_signal[MAX_CPUS];
cpu_signal_event_log_t  *cpu_handle[MAX_CPUS];

MP_EVENT_NAME_DECL();

#endif  /* MP_DEBUG */

/*
 * Note: called with NULL state when polling for TLB flush and cross-calls.
 */
int
cpu_signal_handler(x86_saved_state_t *regs)
{
#if     !MACH_KDP
#pragma unused (regs)
#endif /* !MACH_KDP */
        int             my_cpu;
        volatile int    *my_word;

        SCHED_STATS_IPI(current_processor());

        my_cpu = cpu_number();
        my_word = &cpu_data_ptr[my_cpu]->cpu_signals;
        /* Store the initial set of signals for diagnostics. New
         * signals could arrive while these are being processed
         * so it's no more than a hint.
         */

        cpu_data_ptr[my_cpu]->cpu_prior_signals = *my_word;

        do {
#if     MACH_KDP
                if (i_bit(MP_KDP, my_word)) {
                        DBGLOG(cpu_handle,my_cpu,MP_KDP);
                        i_bit_clear(MP_KDP, my_word);
/* Ensure that the i386_kernel_state at the base of the
 * current thread's stack (if any) is synchronized with the
 * context at the moment of the interrupt, to facilitate
 * access through the debugger.
 */
                        sync_iss_to_iks(regs);
                        if (pmsafe_debug && !kdp_snapshot)
                                pmSafeMode(&current_cpu_datap()->lcpu, PM_SAFE_FL_SAFE);
                        mp_kdp_wait(TRUE, FALSE);
                        if (pmsafe_debug && !kdp_snapshot)
                                pmSafeMode(&current_cpu_datap()->lcpu, PM_SAFE_FL_NORMAL);
                } else
#endif  /* MACH_KDP */
                if (i_bit(MP_TLB_FLUSH, my_word)) {
                        DBGLOG(cpu_handle,my_cpu,MP_TLB_FLUSH);
                        i_bit_clear(MP_TLB_FLUSH, my_word);
                        pmap_update_interrupt();
                } else if (i_bit(MP_RENDEZVOUS, my_word)) {
                        DBGLOG(cpu_handle,my_cpu,MP_RENDEZVOUS);
                        i_bit_clear(MP_RENDEZVOUS, my_word);
                        mp_rendezvous_action();
                } else if (i_bit(MP_BROADCAST, my_word)) {
                        DBGLOG(cpu_handle,my_cpu,MP_BROADCAST);
                        i_bit_clear(MP_BROADCAST, my_word);
                        mp_broadcast_action();
                } else if (i_bit(MP_CHUD, my_word)) {
                        DBGLOG(cpu_handle,my_cpu,MP_CHUD);
                        i_bit_clear(MP_CHUD, my_word);
                        chudxnu_cpu_signal_handler();
                } else if (i_bit(MP_CALL, my_word)) {
                        DBGLOG(cpu_handle,my_cpu,MP_CALL);
                        i_bit_clear(MP_CALL, my_word);
                        mp_cpus_call_action();
                } else if (i_bit(MP_CALL_PM, my_word)) {
                        DBGLOG(cpu_handle,my_cpu,MP_CALL_PM);
                        i_bit_clear(MP_CALL_PM, my_word);
                        mp_call_PM();
                }
                if (regs == NULL) {
                        /* Called to poll only for cross-calls and TLB flush */
                        break;
                } else if (i_bit(MP_AST, my_word)) {
                        DBGLOG(cpu_handle,my_cpu,MP_AST);
                        i_bit_clear(MP_AST, my_word);
                        ast_check(cpu_to_processor(my_cpu));
                }
        } while (*my_word);

        return 0;
}

extern void kprintf_break_lock(void);
static int
NMIInterruptHandler(x86_saved_state_t *regs)
{
        void            *stackptr;

        if (panic_active() && !panicDebugging) {
                if (pmsafe_debug)
                        pmSafeMode(&current_cpu_datap()->lcpu, PM_SAFE_FL_SAFE);
                for(;;)
                        cpu_pause();
        }

        atomic_incl(&NMIPI_acks, 1);
        atomic_incl(&NMI_count, 1);
        sync_iss_to_iks_unconditionally(regs);
        __asm__ volatile("movq %%rbp, %0" : "=m" (stackptr));

        if (cpu_number() == debugger_cpu)
                goto NMExit;

        if (spinlock_timed_out) {
                char pstr[192];
                snprintf(&pstr[0], sizeof(pstr), "Panic(CPU %d): NMIPI for spinlock acquisition timeout, spinlock: %p, spinlock owner: %p, current_thread: %p, spinlock_owner_cpu: 0x%x\n", cpu_number(), spinlock_timed_out, (void *) spinlock_timed_out->interlock.lock_data, current_thread(), spinlock_owner_cpu);
                panic_i386_backtrace(stackptr, 64, &pstr[0], TRUE, regs);
        } else if (mp_cpus_call_wait_timeout) {
                char pstr[192];
                snprintf(&pstr[0], sizeof(pstr), "Panic(CPU %d): Unresponsive processor, this CPU timed-out during cross-call\n", cpu_number());
                panic_i386_backtrace(stackptr, 64, &pstr[0], TRUE, regs);
        } else if (pmap_tlb_flush_timeout == TRUE) {
                char pstr[128];
                snprintf(&pstr[0], sizeof(pstr), "Panic(CPU %d): Unresponsive processor (this CPU did not acknowledge interrupts) TLB state:0x%x\n", cpu_number(), current_cpu_datap()->cpu_tlb_invalid);
                panic_i386_backtrace(stackptr, 48, &pstr[0], TRUE, regs);
        } 

#if MACH_KDP
        if (pmsafe_debug && !kdp_snapshot)
                pmSafeMode(&current_cpu_datap()->lcpu, PM_SAFE_FL_SAFE);
        current_cpu_datap()->cpu_NMI_acknowledged = TRUE;
        i_bit_clear(MP_KDP, &current_cpu_datap()->cpu_signals);
        if (pmap_tlb_flush_timeout ||
            spinlock_timed_out ||
            mp_cpus_call_wait_timeout ||
            panic_active()) {
                mp_kdp_wait(FALSE, TRUE);
        } else if (virtualized && (debug_boot_arg & DB_NMI)) {
                /*
                 * Under a VMM with the debug boot-arg set, drop into kdp.
                 * Since an NMI is involved, there's a risk of contending with
                 * a panic. And side-effects of NMIs may result in entry into, 
                 * and continuing from, the debugger being unreliable.
                 */
                kprintf_break_lock();
                kprintf("Debugger entry requested by NMI\n");
                kdp_i386_trap(T_DEBUG, saved_state64(regs), 0, 0);
                printf("Debugger entry requested by NMI\n");
        } else {
                mp_kdp_wait(FALSE, FALSE);
        }
        if (pmsafe_debug && !kdp_snapshot)
                pmSafeMode(&current_cpu_datap()->lcpu, PM_SAFE_FL_NORMAL);
#endif
NMExit: 
        return 1;
}


/*
 * cpu_interrupt is really just to be used by the scheduler to
 * get a CPU's attention it may not always issue an IPI.  If an
 * IPI is always needed then use i386_cpu_IPI.
 */
void
cpu_interrupt(int cpu)
{
        boolean_t did_IPI = FALSE;

        if (smp_initialized
            && pmCPUExitIdle(cpu_datap(cpu))) {
                i386_cpu_IPI(cpu);
                did_IPI = TRUE;
        }

        KERNEL_DEBUG_CONSTANT(MACHDBG_CODE(DBG_MACH_SCHED, MACH_REMOTE_AST), cpu, did_IPI, 0, 0, 0);
}

/*
 * Send a true NMI via the local APIC to the specified CPU.
 */
void
cpu_NMI_interrupt(int cpu)
{
        if (smp_initialized) {
                i386_send_NMI(cpu);
        }
}

void
NMI_cpus(void)
{
        unsigned int    cpu;
        boolean_t       intrs_enabled;
        uint64_t        tsc_timeout;

        intrs_enabled = ml_set_interrupts_enabled(FALSE);

        for (cpu = 0; cpu < real_ncpus; cpu++) {
                if (!cpu_datap(cpu)->cpu_running)
                        continue;
                cpu_datap(cpu)->cpu_NMI_acknowledged = FALSE;
                cpu_NMI_interrupt(cpu);
                tsc_timeout = !machine_timeout_suspended() ?
                                rdtsc64() + (1000 * 1000 * 1000 * 10ULL) :
                                ~0ULL;
                while (!cpu_datap(cpu)->cpu_NMI_acknowledged) {
                        handle_pending_TLB_flushes();
                        cpu_pause();
                        if (rdtsc64() > tsc_timeout)
                                panic("NMI_cpus() timeout cpu %d", cpu);
                }
                cpu_datap(cpu)->cpu_NMI_acknowledged = FALSE;
        }

        ml_set_interrupts_enabled(intrs_enabled);
}

static void     (* volatile mp_PM_func)(void) = NULL;

static void
mp_call_PM(void)
{
        assert(!ml_get_interrupts_enabled());

        if (mp_PM_func != NULL)
                mp_PM_func();
}

void
cpu_PM_interrupt(int cpu)
{
        assert(!ml_get_interrupts_enabled());

        if (mp_PM_func != NULL) {
                if (cpu == cpu_number())
                        mp_PM_func();
                else
                        i386_signal_cpu(cpu, MP_CALL_PM, ASYNC);
        }
}

void
PM_interrupt_register(void (*fn)(void))
{
        mp_PM_func = fn;
}

void
i386_signal_cpu(int cpu, mp_event_t event, mp_sync_t mode)
{
        volatile int    *signals = &cpu_datap(cpu)->cpu_signals;
        uint64_t        tsc_timeout;

        
        if (!cpu_datap(cpu)->cpu_running)
                return;

        if (event == MP_TLB_FLUSH)
                KERNEL_DEBUG(TRACE_MP_TLB_FLUSH | DBG_FUNC_START, cpu, 0, 0, 0, 0);

        DBGLOG(cpu_signal, cpu, event);
        
        i_bit_set(event, signals);
        i386_cpu_IPI(cpu);
        if (mode == SYNC) {
           again:
                tsc_timeout = !machine_timeout_suspended() ?
                                        rdtsc64() + (1000*1000*1000) :
                                        ~0ULL;
                while (i_bit(event, signals) && rdtsc64() < tsc_timeout) {
                        cpu_pause();
                }
                if (i_bit(event, signals)) {
                        DBG("i386_signal_cpu(%d, 0x%x, SYNC) timed out\n",
                                cpu, event);
                        goto again;
                }
        }
        if (event == MP_TLB_FLUSH)
                KERNEL_DEBUG(TRACE_MP_TLB_FLUSH | DBG_FUNC_END, cpu, 0, 0, 0, 0);
}

/*
 * Send event to all running cpus.
 * Called with the topology locked.
 */
void
i386_signal_cpus(mp_event_t event, mp_sync_t mode)
{
        unsigned int    cpu;
        unsigned int    my_cpu = cpu_number();

        assert(hw_lock_held((hw_lock_t)&x86_topo_lock));

        for (cpu = 0; cpu < real_ncpus; cpu++) {
                if (cpu == my_cpu || !cpu_datap(cpu)->cpu_running)
                        continue;
                i386_signal_cpu(cpu, event, mode);
        }
}

/*
 * Return the number of running cpus.
 * Called with the topology locked.
 */
int
i386_active_cpus(void)
{
        unsigned int    cpu;
        unsigned int    ncpus = 0;

        assert(hw_lock_held((hw_lock_t)&x86_topo_lock));

        for (cpu = 0; cpu < real_ncpus; cpu++) {
                if (cpu_datap(cpu)->cpu_running)
                        ncpus++;
        }
        return(ncpus);
}

/*
 * Helper function called when busy-waiting: panic if too long
 * a TSC-based time has elapsed since the start of the spin.
 */
static boolean_t
mp_spin_timeout(uint64_t tsc_start)
{
        uint64_t        tsc_timeout;

        cpu_pause();
        if (machine_timeout_suspended())
                return FALSE;

        /*
         * The timeout is 4 * the spinlock timeout period
         * unless we have serial console printing (kprintf) enabled
         * in which case we allow an even greater margin.
         */
        tsc_timeout = disable_serial_output ? (uint64_t) LockTimeOutTSC << 2
                                            : (uint64_t) LockTimeOutTSC << 4;
        return  (rdtsc64() > tsc_start + tsc_timeout);
}

/*
 * Helper function to take a spinlock while ensuring that incoming IPIs
 * are still serviced if interrupts are masked while we spin.
 */
static boolean_t
mp_safe_spin_lock(usimple_lock_t lock)
{
        if (ml_get_interrupts_enabled()) {
                simple_lock(lock);
                return TRUE;
        } else {
                uint64_t tsc_spin_start = rdtsc64();
                while (!simple_lock_try(lock)) {
                        cpu_signal_handler(NULL);
                        if (mp_spin_timeout(tsc_spin_start)) {
                                uint32_t lock_cpu;
                                uintptr_t lowner = (uintptr_t)
                                                   lock->interlock.lock_data;
                                spinlock_timed_out = lock;
                                lock_cpu = spinlock_timeout_NMI(lowner);
                                panic("mp_safe_spin_lock() timed out,"
                                      " lock: %p, owner thread: 0x%lx,"
                                      " current_thread: %p, owner on CPU 0x%x",
                                      lock, lowner,
                                      current_thread(), lock_cpu);
                        }
                }
                return FALSE;
        } 
}

/*
 * All-CPU rendezvous:
 *      - CPUs are signalled,
 *      - all execute the setup function (if specified),
 *      - rendezvous (i.e. all cpus reach a barrier),
 *      - all execute the action function (if specified),
 *      - rendezvous again,
 *      - execute the teardown function (if specified), and then
 *      - resume.
 *
 * Note that the supplied external functions _must_ be reentrant and aware
 * that they are running in parallel and in an unknown lock context.
 */

static void
mp_rendezvous_action(void)
{
        boolean_t       intrs_enabled;
        uint64_t        tsc_spin_start;

        /* setup function */
        if (mp_rv_setup_func != NULL)
                mp_rv_setup_func(mp_rv_func_arg);

        intrs_enabled = ml_get_interrupts_enabled();

        /* spin on entry rendezvous */
        atomic_incl(&mp_rv_entry, 1);
        tsc_spin_start = rdtsc64();

        while (mp_rv_entry < mp_rv_ncpus) {
                /* poll for pesky tlb flushes if interrupts disabled */
                if (!intrs_enabled)
                        handle_pending_TLB_flushes();
                if (mp_spin_timeout(tsc_spin_start)) {
                        panic("mp_rv_action() entry: %ld of %d responses, start: 0x%llx, cur: 0x%llx", mp_rv_entry, mp_rv_ncpus, tsc_spin_start, rdtsc64());
                }
        }

        /* action function */
        if (mp_rv_action_func != NULL)
                mp_rv_action_func(mp_rv_func_arg);

        /* spin on exit rendezvous */
        atomic_incl(&mp_rv_exit, 1);
        tsc_spin_start = rdtsc64();
        while (mp_rv_exit < mp_rv_ncpus) {
                if (!intrs_enabled)
                        handle_pending_TLB_flushes();
                if (mp_spin_timeout(tsc_spin_start))
                        panic("mp_rv_action() exit: %ld of %d responses, start: 0x%llx, cur: 0x%llx", mp_rv_exit, mp_rv_ncpus, tsc_spin_start, rdtsc64());
        }

        /* teardown function */
        if (mp_rv_teardown_func != NULL)
                mp_rv_teardown_func(mp_rv_func_arg);

        /* Bump completion count */
        atomic_incl(&mp_rv_complete, 1);
}

void
mp_rendezvous(void (*setup_func)(void *), 
              void (*action_func)(void *),
              void (*teardown_func)(void *),
              void *arg)
{
        uint64_t        tsc_spin_start;

        if (!smp_initialized) {
                if (setup_func != NULL)
                        setup_func(arg);
                if (action_func != NULL)
                        action_func(arg);
                if (teardown_func != NULL)
                        teardown_func(arg);
                return;
        }
                
        /* obtain rendezvous lock */
        (void) mp_safe_spin_lock(&mp_rv_lock);

        /* set static function pointers */
        mp_rv_setup_func = setup_func;
        mp_rv_action_func = action_func;
        mp_rv_teardown_func = teardown_func;
        mp_rv_func_arg = arg;

        mp_rv_entry    = 0;
        mp_rv_exit     = 0;
        mp_rv_complete = 0;

        /*
         * signal other processors, which will call mp_rendezvous_action()
         * with interrupts disabled
         */
        (void) mp_safe_spin_lock(&x86_topo_lock);
        mp_rv_ncpus = i386_active_cpus();
        i386_signal_cpus(MP_RENDEZVOUS, ASYNC);
        simple_unlock(&x86_topo_lock);

        /* call executor function on this cpu */
        mp_rendezvous_action();

        /*
         * Spin for everyone to complete.
         * This is necessary to ensure that all processors have proceeded
         * from the exit barrier before we release the rendezvous structure.
         */
        tsc_spin_start = rdtsc64();
        while (mp_rv_complete < mp_rv_ncpus) {
                if (mp_spin_timeout(tsc_spin_start))
                        panic("mp_rendezvous() timeout: %ld of %d responses, start: 0x%llx, cur: 0x%llx", mp_rv_complete, mp_rv_ncpus, tsc_spin_start, rdtsc64());
        }
        
        /* Tidy up */
        mp_rv_setup_func = NULL;
        mp_rv_action_func = NULL;
        mp_rv_teardown_func = NULL;
        mp_rv_func_arg = NULL;

        /* release lock */
        simple_unlock(&mp_rv_lock);
}

void
mp_rendezvous_break_lock(void)
{
        simple_lock_init(&mp_rv_lock, 0);
}

static void
setup_disable_intrs(__unused void * param_not_used)
{
        /* disable interrupts before the first barrier */
        boolean_t intr = ml_set_interrupts_enabled(FALSE);

        current_cpu_datap()->cpu_iflag = intr;
        DBG("CPU%d: %s\n", get_cpu_number(), __FUNCTION__);
}

static void
teardown_restore_intrs(__unused void * param_not_used)
{
        /* restore interrupt flag following MTRR changes */
        ml_set_interrupts_enabled(current_cpu_datap()->cpu_iflag);
        DBG("CPU%d: %s\n", get_cpu_number(), __FUNCTION__);
}

/*
 * A wrapper to mp_rendezvous() to call action_func() with interrupts disabled.
 * This is exported for use by kexts.
 */
void
mp_rendezvous_no_intrs(
              void (*action_func)(void *),
              void *arg)
{
        mp_rendezvous(setup_disable_intrs,
                      action_func,
                      teardown_restore_intrs,
                      arg);     
}


typedef struct {
        queue_chain_t   link;                   /* queue linkage */
        void            (*func)(void *,void *); /* routine to call */
        void            *arg0;                  /* routine's 1st arg */
        void            *arg1;                  /* routine's 2nd arg */
        cpumask_t       *maskp;                 /* completion response mask */
} mp_call_t;


typedef struct {
        queue_head_t            queue;
        decl_simple_lock_data(, lock);
} mp_call_queue_t;
#define MP_CPUS_CALL_BUFS_PER_CPU       MAX_CPUS
static mp_call_queue_t  mp_cpus_call_freelist;
static mp_call_queue_t  mp_cpus_call_head[MAX_CPUS];

static inline boolean_t
mp_call_head_lock(mp_call_queue_t *cqp)
{
        boolean_t       intrs_enabled;

        intrs_enabled = ml_set_interrupts_enabled(FALSE);
        simple_lock(&cqp->lock);

        return intrs_enabled;
}

void
mp_cpus_NMIPI(cpumask_t cpu_mask) {
        unsigned int cpu, cpu_bit;
        uint64_t deadline;

        for (cpu = 0, cpu_bit = 1; cpu < real_ncpus; cpu++, cpu_bit <<= 1) {
                if (cpu_mask & cpu_bit)
                        cpu_NMI_interrupt(cpu);
        }
        deadline = mach_absolute_time() + (LockTimeOut);
        while (mach_absolute_time() < deadline)
                cpu_pause();
}

#if MACH_ASSERT
static inline boolean_t
mp_call_head_is_locked(mp_call_queue_t *cqp)
{
        return !ml_get_interrupts_enabled() &&
                hw_lock_held((hw_lock_t)&cqp->lock);
}
#endif

static inline void
mp_call_head_unlock(mp_call_queue_t *cqp, boolean_t intrs_enabled)
{
        simple_unlock(&cqp->lock);
        ml_set_interrupts_enabled(intrs_enabled);
}

static inline mp_call_t *
mp_call_alloc(void)
{
        mp_call_t       *callp = NULL;
        boolean_t       intrs_enabled;
        mp_call_queue_t *cqp = &mp_cpus_call_freelist;

        intrs_enabled = mp_call_head_lock(cqp);
        if (!queue_empty(&cqp->queue))
                queue_remove_first(&cqp->queue, callp, typeof(callp), link);
        mp_call_head_unlock(cqp, intrs_enabled);

        return callp;
}

static inline void
mp_call_free(mp_call_t *callp)
{
        boolean_t       intrs_enabled;
        mp_call_queue_t *cqp = &mp_cpus_call_freelist;

        intrs_enabled = mp_call_head_lock(cqp);
        queue_enter_first(&cqp->queue, callp, typeof(callp), link);
        mp_call_head_unlock(cqp, intrs_enabled);
}

static inline mp_call_t *
mp_call_dequeue_locked(mp_call_queue_t *cqp)
{
        mp_call_t       *callp = NULL;

        assert(mp_call_head_is_locked(cqp));
        if (!queue_empty(&cqp->queue))
                queue_remove_first(&cqp->queue, callp, typeof(callp), link);
        return callp;
}

static inline void
mp_call_enqueue_locked(
        mp_call_queue_t *cqp,
        mp_call_t       *callp)
{
        queue_enter(&cqp->queue, callp, typeof(callp), link);
}

/* Called on the boot processor to initialize global structures */
static void
mp_cpus_call_init(void)
{
        mp_call_queue_t *cqp = &mp_cpus_call_freelist;

        DBG("mp_cpus_call_init()\n");
        simple_lock_init(&cqp->lock, 0);
        queue_init(&cqp->queue);
}

/*
 * Called at processor registration to add call buffers to the free list
 * and to initialize the per-cpu call queue.
 */
void
mp_cpus_call_cpu_init(int cpu)
{
        int             i;
        mp_call_queue_t *cqp = &mp_cpus_call_head[cpu];
        mp_call_t       *callp;

        simple_lock_init(&cqp->lock, 0);
        queue_init(&cqp->queue);
        for (i = 0; i < MP_CPUS_CALL_BUFS_PER_CPU; i++) {
                callp = (mp_call_t *) kalloc(sizeof(mp_call_t));
                mp_call_free(callp);
        }

        DBG("mp_cpus_call_init(%d) done\n", cpu);
}

/*
 * This is called from cpu_signal_handler() to process an MP_CALL signal.
 * And also from i386_deactivate_cpu() when a cpu is being taken offline.
 */
static void
mp_cpus_call_action(void)
{
        mp_call_queue_t *cqp;
        boolean_t       intrs_enabled;
        mp_call_t       *callp;
        mp_call_t       call;

        assert(!ml_get_interrupts_enabled());
        cqp = &mp_cpus_call_head[cpu_number()];
        intrs_enabled = mp_call_head_lock(cqp);
        while ((callp = mp_call_dequeue_locked(cqp)) != NULL) {
                /* Copy call request to the stack to free buffer */
                call = *callp;
                mp_call_free(callp);
                if (call.func != NULL) {
                        mp_call_head_unlock(cqp, intrs_enabled);
                        KERNEL_DEBUG_CONSTANT(
                                TRACE_MP_CPUS_CALL_ACTION,
                                VM_KERNEL_UNSLIDE(call.func), VM_KERNEL_UNSLIDE_OR_PERM(call.arg0),
                                VM_KERNEL_UNSLIDE_OR_PERM(call.arg1), VM_KERNEL_ADDRPERM(call.maskp), 0);
                        call.func(call.arg0, call.arg1);
                        (void) mp_call_head_lock(cqp);
                }
                if (call.maskp != NULL)
                        i_bit_set(cpu_number(), call.maskp);
        }
        mp_call_head_unlock(cqp, intrs_enabled);
}

/*
 * mp_cpus_call() runs a given function on cpus specified in a given cpu mask.
 * Possible modes are:
 *  SYNC:   function is called serially on target cpus in logical cpu order
 *          waiting for each call to be acknowledged before proceeding
 *  ASYNC:  function call is queued to the specified cpus
 *          waiting for all calls to complete in parallel before returning
 *  NOSYNC: function calls are queued
 *          but we return before confirmation of calls completing. 
 * The action function may be NULL.
 * The cpu mask may include the local cpu. Offline cpus are ignored.
 * The return value is the number of cpus on which the call was made or queued.
 */
cpu_t
mp_cpus_call(
        cpumask_t       cpus,
        mp_sync_t       mode,
        void            (*action_func)(void *),
        void            *arg)
{
        return mp_cpus_call1(
                        cpus,
                        mode,
                        (void (*)(void *,void *))action_func,
                        arg,
                        NULL,
                        NULL,
                        NULL);
}

static void
mp_cpus_call_wait(boolean_t     intrs_enabled,
                  cpumask_t     cpus_called,
                  cpumask_t     *cpus_responded)
{
        mp_call_queue_t         *cqp;
        uint64_t                tsc_spin_start;

        cqp = &mp_cpus_call_head[cpu_number()];

        tsc_spin_start = rdtsc64();
        while (*cpus_responded != cpus_called) {
                if (!intrs_enabled) {
                        /* Sniffing w/o locking */
                        if (!queue_empty(&cqp->queue))
                                mp_cpus_call_action();
                        cpu_signal_handler(NULL);
                }
                if (mp_spin_timeout(tsc_spin_start)) {
                        cpumask_t       cpus_unresponsive;

                        mp_cpus_call_wait_timeout = TRUE;
                        cpus_unresponsive = cpus_called & ~(*cpus_responded);
                        mp_cpus_NMIPI(cpus_unresponsive);
                        panic("mp_cpus_call_wait() timeout, cpus: 0x%llx",
                                cpus_unresponsive);
                }
        }
}

cpu_t
mp_cpus_call1(
        cpumask_t       cpus,
        mp_sync_t       mode,
        void            (*action_func)(void *, void *),
        void            *arg0,
        void            *arg1,
        cpumask_t       *cpus_calledp,
        cpumask_t       *cpus_notcalledp)
{
        cpu_t           cpu;
        boolean_t       intrs_enabled = FALSE;
        boolean_t       call_self = FALSE;
        cpumask_t       cpus_called = 0;
        cpumask_t       cpus_notcalled = 0;
        cpumask_t       cpus_responded = 0;
        long            cpus_call_count = 0;
        uint64_t        tsc_spin_start;
        boolean_t       topo_lock;

        KERNEL_DEBUG_CONSTANT(
                TRACE_MP_CPUS_CALL | DBG_FUNC_START,
                cpus, mode, VM_KERNEL_UNSLIDE(action_func), VM_KERNEL_UNSLIDE_OR_PERM(arg0), VM_KERNEL_UNSLIDE_OR_PERM(arg1));

        if (!smp_initialized) {
                if ((cpus & CPUMASK_SELF) == 0)
                        goto out;
                if (action_func != NULL) {
                        intrs_enabled = ml_set_interrupts_enabled(FALSE);
                        action_func(arg0, arg1);
                        ml_set_interrupts_enabled(intrs_enabled);
                }
                call_self = TRUE;
                goto out;
        }

        /*
         * Queue the call for each non-local requested cpu.
         * This is performed under the topo lock to prevent changes to
         * cpus online state and to prevent concurrent rendezvouses --
         * although an exception is made if we're calling only the master
         * processor since that always remains active. Note: this exception
         * is expected for longterm timer nosync cross-calls to the master cpu.
         */ 
        mp_disable_preemption();
        intrs_enabled = ml_get_interrupts_enabled();
        topo_lock = (cpus != cpu_to_cpumask(master_cpu));
        if (topo_lock) {
                ml_set_interrupts_enabled(FALSE);
                (void) mp_safe_spin_lock(&x86_topo_lock);
        }
        for (cpu = 0; cpu < (cpu_t) real_ncpus; cpu++) {
                if (((cpu_to_cpumask(cpu) & cpus) == 0) ||
                    !cpu_datap(cpu)->cpu_running)
                        continue;
                tsc_spin_start = rdtsc64();
                if (cpu == (cpu_t) cpu_number()) {
                        /*
                         * We don't IPI ourself and if calling asynchronously,
                         * we defer our call until we have signalled all others.
                         */
                        call_self = TRUE;
                        if (mode == SYNC && action_func != NULL) {
                                KERNEL_DEBUG_CONSTANT(
                                        TRACE_MP_CPUS_CALL_LOCAL,
                                        VM_KERNEL_UNSLIDE(action_func),
                                        VM_KERNEL_UNSLIDE_OR_PERM(arg0), VM_KERNEL_UNSLIDE_OR_PERM(arg1), 0, 0);
                                action_func(arg0, arg1);
                        }
                } else {
                        /*
                         * Here to queue a call to cpu and IPI.
                         * Spinning for request buffer unless NOSYNC.
                         */
                        mp_call_t       *callp = NULL;
                        mp_call_queue_t *cqp = &mp_cpus_call_head[cpu];
                        boolean_t       intrs_inner;

                queue_call:
                        if (callp == NULL)
                                callp = mp_call_alloc();
                        intrs_inner = mp_call_head_lock(cqp);
                        if (mode == NOSYNC) {
                                if (callp == NULL) {
                                        cpus_notcalled |= cpu_to_cpumask(cpu);
                                        mp_call_head_unlock(cqp, intrs_inner);
                                        KERNEL_DEBUG_CONSTANT(
                                                TRACE_MP_CPUS_CALL_NOBUF,
                                                cpu, 0, 0, 0, 0);
                                        continue;
                                }
                                callp->maskp = NULL;
                        } else {
                                if (callp == NULL) {
                                        mp_call_head_unlock(cqp, intrs_inner);
                                        KERNEL_DEBUG_CONSTANT(
                                                TRACE_MP_CPUS_CALL_NOBUF,
                                                cpu, 0, 0, 0, 0);
                                        if (!intrs_inner) {
                                                /* Sniffing w/o locking */
                                                if (!queue_empty(&cqp->queue))
                                                        mp_cpus_call_action();
                                                handle_pending_TLB_flushes();
                                        }
                                        if (mp_spin_timeout(tsc_spin_start))
                                                panic("mp_cpus_call1() timeout");
                                        goto queue_call;
                                }
                                callp->maskp = &cpus_responded;
                        }
                        callp->func = action_func;
                        callp->arg0 = arg0;
                        callp->arg1 = arg1;
                        mp_call_enqueue_locked(cqp, callp);
                        cpus_call_count++;
                        cpus_called |= cpu_to_cpumask(cpu);
                        i386_signal_cpu(cpu, MP_CALL, ASYNC);
                        mp_call_head_unlock(cqp, intrs_inner);
                        if (mode == SYNC) {
                                mp_cpus_call_wait(intrs_inner, cpus_called, &cpus_responded);
                        }
                }
        }
        if (topo_lock) {
                simple_unlock(&x86_topo_lock);
                ml_set_interrupts_enabled(intrs_enabled);
        }

        /* Call locally if mode not SYNC */
        if (mode != SYNC && call_self ) {
                KERNEL_DEBUG_CONSTANT(
                        TRACE_MP_CPUS_CALL_LOCAL,
                        VM_KERNEL_UNSLIDE(action_func), VM_KERNEL_UNSLIDE_OR_PERM(arg0), VM_KERNEL_UNSLIDE_OR_PERM(arg1), 0, 0);
                if (action_func != NULL) {
                        ml_set_interrupts_enabled(FALSE);
                        action_func(arg0, arg1);
                        ml_set_interrupts_enabled(intrs_enabled);
                }
        }

        /* Safe to allow pre-emption now */
        mp_enable_preemption();

        /* For ASYNC, now wait for all signaled cpus to complete their calls */
        if (mode == ASYNC)
                mp_cpus_call_wait(intrs_enabled, cpus_called, &cpus_responded);

out:
        if (call_self){
                cpus_called |= cpu_to_cpumask(cpu);
                cpus_call_count++;
        }

        if (cpus_calledp)
                *cpus_calledp = cpus_called;
        if (cpus_notcalledp)
                *cpus_notcalledp = cpus_notcalled;

        KERNEL_DEBUG_CONSTANT(
                TRACE_MP_CPUS_CALL | DBG_FUNC_END,
                cpus_call_count, cpus_called, cpus_notcalled, 0, 0);

        return (cpu_t) cpus_call_count;
}


static void
mp_broadcast_action(void)
{
   /* call action function */
   if (mp_bc_action_func != NULL)
       mp_bc_action_func(mp_bc_func_arg);

   /* if we're the last one through, wake up the instigator */
   if (atomic_decl_and_test(&mp_bc_count, 1))
       thread_wakeup(((event_t)(uintptr_t) &mp_bc_count));
}

/*
 * mp_broadcast() runs a given function on all active cpus.
 * The caller blocks until the functions has run on all cpus.
 * The caller will also block if there is another pending braodcast.
 */
void
mp_broadcast(
         void (*action_func)(void *),
         void *arg)
{
   if (!smp_initialized) {
       if (action_func != NULL)
                   action_func(arg);
       return;
   }
       
   /* obtain broadcast lock */
   lck_mtx_lock(&mp_bc_lock);

   /* set static function pointers */
   mp_bc_action_func = action_func;
   mp_bc_func_arg = arg;

   assert_wait((event_t)(uintptr_t)&mp_bc_count, THREAD_UNINT);

   /*
    * signal other processors, which will call mp_broadcast_action()
    */
   simple_lock(&x86_topo_lock);
   mp_bc_ncpus = i386_active_cpus();   /* total including this cpu */
   mp_bc_count = mp_bc_ncpus;
   i386_signal_cpus(MP_BROADCAST, ASYNC);

   /* call executor function on this cpu */
   mp_broadcast_action();
   simple_unlock(&x86_topo_lock);

   /* block for all cpus to have run action_func */
   if (mp_bc_ncpus > 1)
       thread_block(THREAD_CONTINUE_NULL);
   else
       clear_wait(current_thread(), THREAD_AWAKENED);
       
   /* release lock */
   lck_mtx_unlock(&mp_bc_lock);
}

void
mp_cpus_kick(cpumask_t cpus)
{
        cpu_t           cpu;
        boolean_t       intrs_enabled = FALSE;

        intrs_enabled = ml_set_interrupts_enabled(FALSE);
        mp_safe_spin_lock(&x86_topo_lock);

        for (cpu = 0; cpu < (cpu_t) real_ncpus; cpu++) {
                if ((cpu == (cpu_t) cpu_number())
                        || ((cpu_to_cpumask(cpu) & cpus) == 0)
                        || (!cpu_datap(cpu)->cpu_running))
                {
                                continue;
                }

                lapic_send_ipi(cpu, LAPIC_VECTOR(KICK));
        }

        simple_unlock(&x86_topo_lock);
        ml_set_interrupts_enabled(intrs_enabled);
}

void
i386_activate_cpu(void)
{
        cpu_data_t      *cdp = current_cpu_datap();

        assert(!ml_get_interrupts_enabled());

        if (!smp_initialized) {
                cdp->cpu_running = TRUE;
                return;
        }

        simple_lock(&x86_topo_lock);
        cdp->cpu_running = TRUE;
        started_cpu();
        simple_unlock(&x86_topo_lock);
        flush_tlb_raw();
}

void
i386_deactivate_cpu(void)
{
        cpu_data_t      *cdp = current_cpu_datap();

        assert(!ml_get_interrupts_enabled());
 
        KERNEL_DEBUG_CONSTANT(
                TRACE_MP_CPU_DEACTIVATE | DBG_FUNC_START,
                0, 0, 0, 0, 0);

        simple_lock(&x86_topo_lock);
        cdp->cpu_running = FALSE;
        simple_unlock(&x86_topo_lock);

        /*
         * Move all of this cpu's timers to the master/boot cpu,
         * and poke it in case there's a sooner deadline for it to schedule.
         */
        timer_queue_shutdown(&cdp->rtclock_timer.queue);
        mp_cpus_call(cpu_to_cpumask(master_cpu), ASYNC, timer_queue_expire_local, NULL);

        /*
         * Open an interrupt window
         * and ensure any pending IPI or timer is serviced
         */
        mp_disable_preemption();
        ml_set_interrupts_enabled(TRUE);

        while (cdp->cpu_signals && x86_lcpu()->rtcDeadline != EndOfAllTime)
                cpu_pause();
        /*
         * Ensure there's no remaining timer deadline set
         * - AICPM may have left one active.
         */
        setPop(0);

        ml_set_interrupts_enabled(FALSE);
        mp_enable_preemption();

        KERNEL_DEBUG_CONSTANT(
                TRACE_MP_CPU_DEACTIVATE | DBG_FUNC_END,
                0, 0, 0, 0, 0);
}

int     pmsafe_debug    = 1;

#if     MACH_KDP
volatile boolean_t      mp_kdp_trap = FALSE;
volatile unsigned long  mp_kdp_ncpus;
boolean_t               mp_kdp_state;


void
mp_kdp_enter(void)
{
        unsigned int    cpu;
        unsigned int    ncpus = 0;
        unsigned int    my_cpu;
        uint64_t        tsc_timeout;

        DBG("mp_kdp_enter()\n");

#if DEBUG
        if (!smp_initialized)
                simple_lock_init(&mp_kdp_lock, 0);
#endif

        /*
         * Here to enter the debugger.
         * In case of races, only one cpu is allowed to enter kdp after
         * stopping others.
         */
        mp_kdp_state = ml_set_interrupts_enabled(FALSE);
        my_cpu = cpu_number();

        if (my_cpu == (unsigned) debugger_cpu) {
                kprintf("\n\nRECURSIVE DEBUGGER ENTRY DETECTED\n\n");
                kdp_reset();
                return;
        }

        cpu_datap(my_cpu)->debugger_entry_time = mach_absolute_time();
        simple_lock(&mp_kdp_lock);

        if (pmsafe_debug && !kdp_snapshot)
            pmSafeMode(&current_cpu_datap()->lcpu, PM_SAFE_FL_SAFE);

        while (mp_kdp_trap) {
                simple_unlock(&mp_kdp_lock);
                DBG("mp_kdp_enter() race lost\n");
#if MACH_KDP
                mp_kdp_wait(TRUE, FALSE);
#endif
                simple_lock(&mp_kdp_lock);
        }
        debugger_cpu = my_cpu;
        ncpus = 1;
        mp_kdp_ncpus = 1;       /* self */
        mp_kdp_trap = TRUE;
        debugger_entry_time = cpu_datap(my_cpu)->debugger_entry_time;
        simple_unlock(&mp_kdp_lock);

        /*
         * Deliver a nudge to other cpus, counting how many
         */
        DBG("mp_kdp_enter() signaling other processors\n");
        if (force_immediate_debugger_NMI == FALSE) {
                for (cpu = 0; cpu < real_ncpus; cpu++) {
                        if (cpu == my_cpu || !cpu_datap(cpu)->cpu_running)
                                continue;
                        ncpus++;
                        i386_signal_cpu(cpu, MP_KDP, ASYNC);
                }
                /*
                 * Wait other processors to synchronize
                 */
                DBG("mp_kdp_enter() waiting for (%d) processors to suspend\n", ncpus);

                /*
                 * This timeout is rather arbitrary; we don't want to NMI
                 * processors that are executing at potentially
                 * "unsafe-to-interrupt" points such as the trampolines,
                 * but neither do we want to lose state by waiting too long.
                 */
                tsc_timeout = rdtsc64() + (ncpus * 1000 * 1000 * 10ULL);

                if (virtualized)
                        tsc_timeout = ~0ULL;

                while (mp_kdp_ncpus != ncpus && rdtsc64() < tsc_timeout) {
                        /*
                         * A TLB shootdown request may be pending--this would
                         * result in the requesting processor waiting in
                         * PMAP_UPDATE_TLBS() until this processor deals with it.
                         * Process it, so it can now enter mp_kdp_wait()
                         */
                        handle_pending_TLB_flushes();
                        cpu_pause();
                }
                /* If we've timed out, and some processor(s) are still unresponsive,
                 * interrupt them with an NMI via the local APIC.
                 */
                if (mp_kdp_ncpus != ncpus) {
                        for (cpu = 0; cpu < real_ncpus; cpu++) {
                                if (cpu == my_cpu || !cpu_datap(cpu)->cpu_running)
                                        continue;
                                if (cpu_signal_pending(cpu, MP_KDP))
                                        cpu_NMI_interrupt(cpu);
                        }
                }
        }
        else
                for (cpu = 0; cpu < real_ncpus; cpu++) {
                        if (cpu == my_cpu || !cpu_datap(cpu)->cpu_running)
                                continue;
                        cpu_NMI_interrupt(cpu);
                }

        DBG("mp_kdp_enter() %d processors done %s\n",
            (int)mp_kdp_ncpus, (mp_kdp_ncpus == ncpus) ? "OK" : "timed out");
        
        postcode(MP_KDP_ENTER);
}

static boolean_t
cpu_signal_pending(int cpu, mp_event_t event)
{
        volatile int    *signals = &cpu_datap(cpu)->cpu_signals;
        boolean_t retval = FALSE;

        if (i_bit(event, signals))
                retval = TRUE;
        return retval;
}

long kdp_x86_xcpu_invoke(const uint16_t lcpu, kdp_x86_xcpu_func_t func,
                         void *arg0, void *arg1)
{
        if (lcpu > (real_ncpus - 1))
                return -1;

        if (func == NULL)
                return -1;

        kdp_xcpu_call_func.func = func;
        kdp_xcpu_call_func.ret  = -1;
        kdp_xcpu_call_func.arg0 = arg0;
        kdp_xcpu_call_func.arg1 = arg1;
        kdp_xcpu_call_func.cpu  = lcpu;
        DBG("Invoking function %p on CPU %d\n", func, (int32_t)lcpu);
        while (kdp_xcpu_call_func.cpu != KDP_XCPU_NONE)
                cpu_pause();
        return kdp_xcpu_call_func.ret;
}

static void
kdp_x86_xcpu_poll(void)
{
        if ((uint16_t)cpu_number() == kdp_xcpu_call_func.cpu) {
            kdp_xcpu_call_func.ret = 
                    kdp_xcpu_call_func.func(kdp_xcpu_call_func.arg0,
                                            kdp_xcpu_call_func.arg1,
                                            cpu_number());
                kdp_xcpu_call_func.cpu = KDP_XCPU_NONE;
        }
}

static void
mp_kdp_wait(boolean_t flush, boolean_t isNMI)
{
        DBG("mp_kdp_wait()\n");
        /* If an I/O port has been specified as a debugging aid, issue a read */
        panic_io_port_read();
        current_cpu_datap()->debugger_ipi_time = mach_absolute_time();
#if CONFIG_MCA
        /* If we've trapped due to a machine-check, save MCA registers */
        mca_check_save();
#endif

        atomic_incl((volatile long *)&mp_kdp_ncpus, 1);
        while (mp_kdp_trap || (isNMI == TRUE)) {
                /*
                 * A TLB shootdown request may be pending--this would result
                 * in the requesting processor waiting in PMAP_UPDATE_TLBS()
                 * until this processor handles it.
                 * Process it, so it can now enter mp_kdp_wait()
                 */
                if (flush)
                        handle_pending_TLB_flushes();

                kdp_x86_xcpu_poll();
                cpu_pause();
        }

        atomic_decl((volatile long *)&mp_kdp_ncpus, 1);
        DBG("mp_kdp_wait() done\n");
}

void
mp_kdp_exit(void)
{
        DBG("mp_kdp_exit()\n");
        debugger_cpu = -1;
        atomic_decl((volatile long *)&mp_kdp_ncpus, 1);

        debugger_exit_time = mach_absolute_time();

        mp_kdp_trap = FALSE;
        mfence();

        /* Wait other processors to stop spinning. XXX needs timeout */
        DBG("mp_kdp_exit() waiting for processors to resume\n");
        while (mp_kdp_ncpus > 0) {
                /*
                 * a TLB shootdown request may be pending... this would result in the requesting
                 * processor waiting in PMAP_UPDATE_TLBS() until this processor deals with it.
                 * Process it, so it can now enter mp_kdp_wait()
                 */
                handle_pending_TLB_flushes();

                cpu_pause();
        }

        if (pmsafe_debug && !kdp_snapshot)
            pmSafeMode(&current_cpu_datap()->lcpu, PM_SAFE_FL_NORMAL);

        debugger_exit_time = mach_absolute_time();

        DBG("mp_kdp_exit() done\n");
        (void) ml_set_interrupts_enabled(mp_kdp_state);
        postcode(0);
}
#endif  /* MACH_KDP */

boolean_t
mp_recent_debugger_activity(void) {
        uint64_t abstime = mach_absolute_time();
        return (((abstime - debugger_entry_time) < LastDebuggerEntryAllowance) ||
            ((abstime - debugger_exit_time) < LastDebuggerEntryAllowance));
}

/*ARGSUSED*/
void
init_ast_check(
        __unused processor_t    processor)
{
}

void
cause_ast_check(
        processor_t     processor)
{
        int     cpu = processor->cpu_id;

        if (cpu != cpu_number()) {
                i386_signal_cpu(cpu, MP_AST, ASYNC);
                KERNEL_DEBUG_CONSTANT(MACHDBG_CODE(DBG_MACH_SCHED, MACH_REMOTE_AST), cpu, 1, 0, 0, 0);
        }
}

void
slave_machine_init(void *param)
{
        /*
         * Here in process context, but with interrupts disabled.
         */
        DBG("slave_machine_init() CPU%d\n", get_cpu_number());

        if (param == FULL_SLAVE_INIT) {
                /*
                 * Cold start
                 */
                clock_init();
        }
        cpu_machine_init();     /* Interrupts enabled hereafter */
}

#undef cpu_number
int cpu_number(void)
{
        return get_cpu_number();
}

static void
cpu_prewarm_init()
{
        int i;

        simple_lock_init(&cpu_warm_lock, 0);
        queue_init(&cpu_warm_call_list);
        for (i = 0; i < NUM_CPU_WARM_CALLS; i++) {
                enqueue_head(&cpu_warm_call_list, (queue_entry_t)&cpu_warm_call_arr[i]);
        }
}

static timer_call_t
grab_warm_timer_call()
{
        spl_t x;
        timer_call_t call = NULL;

        x = splsched();
        simple_lock(&cpu_warm_lock);
        if (!queue_empty(&cpu_warm_call_list)) {
                call = (timer_call_t) dequeue_head(&cpu_warm_call_list);
        }
        simple_unlock(&cpu_warm_lock);
        splx(x);

        return call;
}

static void
free_warm_timer_call(timer_call_t call)
{
        spl_t x;

        x = splsched();
        simple_lock(&cpu_warm_lock);
        enqueue_head(&cpu_warm_call_list, (queue_entry_t)call);
        simple_unlock(&cpu_warm_lock);
        splx(x);
}

/*
 * Runs in timer call context (interrupts disabled).
 */
static void
cpu_warm_timer_call_func(
                call_entry_param_t p0,
                __unused call_entry_param_t p1)
{
        free_warm_timer_call((timer_call_t)p0);
        return;
}

/*
 * Runs with interrupts disabled on the CPU we wish to warm (i.e. CPU 0).
 */
static void
_cpu_warm_setup(
                void *arg)
{
        cpu_warm_data_t cwdp = (cpu_warm_data_t)arg;

        timer_call_enter(cwdp->cwd_call, cwdp->cwd_deadline, TIMER_CALL_SYS_CRITICAL | TIMER_CALL_LOCAL);
        cwdp->cwd_result = 0;

        return;
}

/*
 * Not safe to call with interrupts disabled.
 */
kern_return_t
ml_interrupt_prewarm(
        uint64_t        deadline)
{
        struct cpu_warm_data cwd;
        timer_call_t call;
        cpu_t ct;

        if (ml_get_interrupts_enabled() == FALSE) {
                panic("%s: Interrupts disabled?\n", __FUNCTION__);
        }

        /* 
         * If the platform doesn't need our help, say that we succeeded. 
         */
        if (!ml_get_interrupt_prewake_applicable()) {
                return KERN_SUCCESS;
        }

        /*
         * Grab a timer call to use.
         */
        call = grab_warm_timer_call();
        if (call == NULL) {
                return KERN_RESOURCE_SHORTAGE;
        }

        timer_call_setup(call, cpu_warm_timer_call_func, call);
        cwd.cwd_call = call;
        cwd.cwd_deadline = deadline;
        cwd.cwd_result = 0;

        /*
         * For now, non-local interrupts happen on the master processor.
         */
        ct = mp_cpus_call(cpu_to_cpumask(master_cpu), SYNC, _cpu_warm_setup, &cwd);
        if (ct == 0) {
                free_warm_timer_call(call);
                return KERN_FAILURE;
        } else {
                return cwd.cwd_result;
        }
}