]> git.saurik.com Git - apple/xnu.git/blobdiff - osfmk/i386/mp.c
xnu-6153.11.26.tar.gz
[apple/xnu.git] / osfmk / i386 / mp.c
index e11c8f6c5a6039d1eed6dbce1c794f1f1d7fe170..b6654cc393b0f2ebf1ebb7196cc710041069e804 100644 (file)
@@ -1,8 +1,8 @@
 /*
- * Copyright (c) 2000-2009 Apple Inc. All rights reserved.
+ * Copyright (c) 2000-2018 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
  * 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,
  * 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_kdb.h>
 #include <mach_kdp.h>
+#include <kdp/kdp_internal.h>
 #include <mach_ldebug.h>
-#include <gprof.h>
 
 #include <mach/mach_types.h>
 #include <mach/kern_return.h>
 #include <kern/cpu_number.h>
 #include <kern/cpu_data.h>
 #include <kern/assert.h>
+#include <kern/lock_group.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/mp.h>
 #include <i386/mp_events.h>
 #include <i386/lapic.h>
-#include <i386/ipl.h>
 #include <i386/cpuid.h>
 #include <i386/fpu.h>
 #include <i386/machine_cpu.h>
-#include <i386/mtrr.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>
-#if MACH_KDB
-#include <machine/db_machdep.h>
-#include <ddb/db_aout.h>
-#include <ddb/db_access.h>
-#include <ddb/db_sym.h>
-#include <ddb/db_variables.h>
-#include <ddb/db_command.h>
-#include <ddb/db_output.h>
-#include <ddb/db_expr.h>
-#endif
 
-#if    MP_DEBUG
-#define PAUSE          delay(1000000)
-#define DBG(x...)      kprintf(x)
+#include <console/serial_protos.h>
+
+#if MONOTONIC
+#include <kern/monotonic.h>
+#endif /* MONOTONIC */
+
+#if     MP_DEBUG
+#define PAUSE           delay(1000000)
+#define DBG(x...)       kprintf(x)
 #else
 #define DBG(x...)
 #define PAUSE
-#endif /* MP_DEBUG */
+#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)
 
-void           slave_boot_init(void);
+#define ABS(v)          (((v) > 0)?(v):-(v))
 
-#if MACH_KDB
-static void    mp_kdb_wait(void);
-volatile boolean_t     mp_kdb_trap = FALSE;
-volatile long  mp_kdb_ncpus = 0;
-#endif
+void            slave_boot_init(void);
+void            i386_cpu_IPI(int cpu);
 
-static void    mp_kdp_wait(boolean_t flush, boolean_t isNMI);
-static void    mp_rendezvous_action(void);
-static void    mp_broadcast_action(void);
+#if MACH_KDP
+static void     mp_kdp_wait(boolean_t flush, boolean_t isNMI);
+#endif /* MACH_KDP */
 
-static boolean_t       cpu_signal_pending(int cpu, mp_event_t event);
-static int             cpu_signal_handler(x86_saved_state_t *regs);
-static int             NMIInterruptHandler(x86_saved_state_t *regs);
+#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;
+#if DEBUG || DEVELOPMENT
+boolean_t               mp_interrupt_watchdog_enabled = TRUE;
+uint32_t                mp_interrupt_watchdog_events = 0;
+#endif
 
-boolean_t              smp_initialized = FALSE;
-volatile boolean_t     force_immediate_debugger_NMI = FALSE;
-volatile boolean_t     pmap_tlb_flush_timeout = FALSE;
-decl_simple_lock_data(,mp_kdp_lock);
+decl_simple_lock_data(, debugger_callback_lock);
+struct debugger_callback *debugger_callback = NULL;
 
 decl_lck_mtx_data(static, mp_cpu_boot_lock);
-lck_mtx_ext_t  mp_cpu_boot_lock_ext;
+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;
+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;
@@ -153,83 +161,121 @@ static struct _kdp_xcpu_call_func {
 /* 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 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;
-
-static void    mp_cpus_call_action(void); 
-static void    mp_call_PM(void);
-
-char           mp_slave_stack[PAGE_SIZE] __attribute__((aligned(PAGE_SIZE))); // Temp stack for slave init
-
-
-#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;
-
-extern void    slave_pstart(void);
+lck_mtx_ext_t   mp_bc_lock_ext;
+static  volatile int    debugger_cpu = -1;
+volatile long    NMIPI_acks = 0;
+volatile long    NMI_count = 0;
+static NMI_reason_t     NMI_panic_reason = NONE;
+static int              vector_timed_out;
+
+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);
+
+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);
+void NMIPI_enable(boolean_t);
+
+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);
+       simple_lock_init(&debugger_callback_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();
 
-       /* Local APIC? */
-       if (!lapic_probe())
+       if (!i386_smp_init(LAPIC_NMI_INTERRUPT, NMIInterruptHandler,
+           LAPIC_VECTOR(INTERPROCESSOR), cpu_signal_handler)) {
                return;
-
-       lapic_init();
-       lapic_configure();
-       lapic_set_intr_func(LAPIC_NMI_INTERRUPT,  NMIInterruptHandler);
-       lapic_set_intr_func(LAPIC_VECTOR(INTERPROCESSOR), cpu_signal_handler);
+       }
 
        cpu_thread_init();
 
-       GPROF_INIT();
        DBGLOG_CPU_INIT(master_cpu);
 
-       install_real_mode_bootstrap(slave_pstart);
+       mp_cpus_call_init();
+       mp_cpus_call_cpu_init(master_cpu);
+
+#if DEBUG || DEVELOPMENT
+       if (PE_parse_boot_argn("interrupt_watchdog",
+           &mp_interrupt_watchdog_enabled,
+           sizeof(mp_interrupt_watchdog_enabled))) {
+               kprintf("Interrupt watchdog %sabled\n",
+                   mp_interrupt_watchdog_enabled ? "en" : "dis");
+       }
+#endif
 
+       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;
+       while (iters-- > 0) {
+               if (cpu_datap(slot_num)->cpu_running) {
+                       break;
+               }
                delay(usecdelay);
        }
 }
@@ -240,101 +286,162 @@ mp_wait_for_cpu_up(int slot_num, unsigned int iters, unsigned int usecdelay)
 kern_return_t
 intel_startCPU_fast(int slot_num)
 {
-       kern_return_t   rc;
+       kern_return_t   rc;
 
        /*
         * Try to perform a fast restart
         */
        rc = pmCPUExitHalt(slot_num);
-       if (rc != KERN_SUCCESS)
+       if (rc != KERN_SUCCESS) {
                /*
                 * The CPU was not eligible for a fast restart.
                 */
-               return(rc);
+               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);
+       if (cpu_datap(slot_num)->cpu_running) {
+               return KERN_SUCCESS;
+       } else {
+               return KERN_FAILURE;
+       }
 }
 
-typedef struct {
-       int     target_cpu;
-       int     target_lapic;
-       int     starter_cpu;
-} processor_start_info_t;
+static void
+started_cpu(void)
+{
+       /* Here on the started cpu with cpu_running set TRUE */
 
-static processor_start_info_t start_info;
+       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;
+       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)
+       if (cpu_number() != psip->starter_cpu) {
                return;
+       }
 
-       LAPIC_WRITE(ICRD, psip->target_lapic << LAPIC_ICRD_DEST_SHIFT);
-       LAPIC_WRITE(ICR, LAPIC_ICR_DM_INIT);
-       delay(100);
+       DBG("start_cpu(%p) about to start cpu %d, lapic %d\n",
+           arg, psip->target_cpu, psip->target_lapic);
 
-       LAPIC_WRITE(ICRD, psip->target_lapic << LAPIC_ICRD_DEST_SHIFT);
-       LAPIC_WRITE(ICR, LAPIC_ICR_DM_STARTUP|(REAL_MODE_BOOTSTRAP_OFFSET>>12));
+       KERNEL_DEBUG_CONSTANT(
+               TRACE_MP_CPU_START | DBG_FUNC_START,
+               psip->target_cpu,
+               psip->target_lapic, 0, 0, 0);
 
-#ifdef POSTCODE_DELAY
+       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
-       mp_wait_for_cpu_up(psip->target_cpu, i*100, 100);
-}
+       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);
 
-extern char    prot_mode_gdt[];
-extern char    slave_boot_base[];
-extern char real_mode_bootstrap_base[];
-extern char real_mode_bootstrap_end[];
-extern char    slave_boot_end[];
+       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     slot_num)
 {
-       int             lapic = cpu_to_lapic[slot_num];
-       boolean_t       istate;
+       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) IdlePTD);
+       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.
         */
-       if (cpu_mode_is64bit())
-               cpu_desc_init64(cpu_datap(slot_num));
-       else
-               cpu_desc_init(cpu_datap(slot_num));
+       cpu_desc_init(cpu_datap(slot_num));
 
        /* Serialize use of the slave boot stack, etc. */
        lck_mtx_lock(&mp_cpu_boot_lock);
@@ -349,6 +456,8 @@ intel_startCPU(
        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
@@ -357,6 +466,8 @@ intel_startCPU(
         */
        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);
 
@@ -372,36 +483,41 @@ intel_startCPU(
        }
 }
 
-#if    MP_DEBUG
-cpu_signal_event_log_t *cpu_signal[MAX_CPUS];
-cpu_signal_event_log_t *cpu_handle[MAX_CPUS];
+#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 */
+#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)
 {
-       int             my_cpu;
-       volatile int    *my_word;
-#if    MACH_KDB && MACH_ASSERT
-       int             i=100;
-#endif /* MACH_KDB && MACH_ASSERT */
+#if     !MACH_KDP
+#pragma unused (regs)
+#endif /* !MACH_KDP */
+       int             my_cpu;
+       volatile int    *my_word;
 
-       mp_disable_preemption();
+       SCHED_STATS_IPI(current_processor());
 
        my_cpu = cpu_number();
-       my_word = &current_cpu_datap()->cpu_signals;
+       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_KDB && MACH_ASSERT
-               if (i-- <= 0)
-                   Debugger("cpu_signal_handler: signals did not clear");
-#endif /* MACH_KDB && MACH_ASSERT */
-#if    MACH_KDP
+#if     MACH_KDP
                if (i_bit(MP_KDP, my_word)) {
-                       DBGLOG(cpu_handle,my_cpu,MP_KDP);
+                       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
@@ -409,127 +525,127 @@ cpu_signal_handler(x86_saved_state_t *regs)
  * 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 */
+#endif  /* MACH_KDP */
                if (i_bit(MP_TLB_FLUSH, my_word)) {
-                       DBGLOG(cpu_handle,my_cpu,MP_TLB_FLUSH);
+                       DBGLOG(cpu_handle, my_cpu, MP_TLB_FLUSH);
                        i_bit_clear(MP_TLB_FLUSH, my_word);
                        pmap_update_interrupt();
-               } 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));
-#if    MACH_KDB
-               } else if (i_bit(MP_KDB, my_word)) {
-
-                       i_bit_clear(MP_KDB, my_word);
-                       current_cpu_datap()->cpu_kdb_is_slave++;
-                       mp_kdb_wait();
-                       current_cpu_datap()->cpu_kdb_is_slave--;
-#endif /* MACH_KDB */
-               } 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);
+                       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);
+                       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);
 
-       mp_enable_preemption();
-
        return 0;
 }
 
-static int
+extern void kprintf_break_lock(void);
+int
 NMIInterruptHandler(x86_saved_state_t *regs)
 {
-       void    *stackptr;
-       
+       void            *stackptr;
+       char            pstr[256];
+       uint64_t        now = mach_absolute_time();
+
+       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);
-#if defined (__i386__)
-       __asm__ volatile("movl %%ebp, %0" : "=m" (stackptr));
-#elif defined (__x86_64__)
-       __asm__ volatile("movq %%rbp, %0" : "=m" (stackptr));
-#endif
+       __asm__ volatile ("movq %%rbp, %0" : "=m" (stackptr));
 
-       if (cpu_number() == debugger_cpu)
-                       goto NMExit;
+       if (cpu_number() == debugger_cpu) {
+               goto NMExit;
+       }
 
-       if (pmap_tlb_flush_timeout == TRUE && current_cpu_datap()->cpu_tlb_invalid) {
-               char pstr[128];
-               snprintf(&pstr[0], sizeof(pstr), "Panic(CPU %d): Unresponsive processor\n", cpu_number());
-               panic_i386_backtrace(stackptr, 16, &pstr[0], TRUE, regs);
+       if (NMI_panic_reason == SPINLOCK_TIMEOUT) {
+               snprintf(&pstr[0], sizeof(pstr),
+                   "Panic(CPU %d, time %llu): NMIPI for spinlock acquisition timeout, spinlock: %p, spinlock owner: %p, current_thread: %p, spinlock_owner_cpu: 0x%x\n",
+                   cpu_number(), now, 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 (NMI_panic_reason == TLB_FLUSH_TIMEOUT) {
+               snprintf(&pstr[0], sizeof(pstr),
+                   "Panic(CPU %d, time %llu): NMIPI for unresponsive processor: TLB flush timeout, TLB state:0x%x\n",
+                   cpu_number(), now, current_cpu_datap()->cpu_tlb_invalid);
+               panic_i386_backtrace(stackptr, 48, &pstr[0], TRUE, regs);
+       } else if (NMI_panic_reason == CROSSCALL_TIMEOUT) {
+               snprintf(&pstr[0], sizeof(pstr),
+                   "Panic(CPU %d, time %llu): NMIPI for unresponsive processor: cross-call timeout\n",
+                   cpu_number(), now);
+               panic_i386_backtrace(stackptr, 64, &pstr[0], TRUE, regs);
+       } else if (NMI_panic_reason == INTERRUPT_WATCHDOG) {
+               snprintf(&pstr[0], sizeof(pstr),
+                   "Panic(CPU %d, time %llu): NMIPI for unresponsive processor: interrupt watchdog for vector 0x%x\n",
+                   cpu_number(), now, vector_timed_out);
+               panic_i386_backtrace(stackptr, 64, &pstr[0], TRUE, regs);
        }
 
 #if MACH_KDP
-       mp_kdp_wait(FALSE, pmap_tlb_flush_timeout);
-#endif
-NMExit:        
-       return 1;
-}
-
-#ifdef MP_DEBUG
-int    max_lock_loops = 100000000;
-int            trappedalready = 0;     /* (BRINGUP) */
-#endif /* MP_DEBUG */
-
-static void
-i386_cpu_IPI(int cpu)
-{
-       boolean_t       state;
-       
-#ifdef MP_DEBUG
-       if(cpu_datap(cpu)->cpu_signals & 6) {   /* (BRINGUP) */
-               kprintf("i386_cpu_IPI: sending enter debugger signal (%08X) to cpu %d\n", cpu_datap(cpu)->cpu_signals, cpu);
-       }
-#endif /* MP_DEBUG */
-
-#if MACH_KDB
-#ifdef MP_DEBUG
-       if(!trappedalready && (cpu_datap(cpu)->cpu_signals & 6)) {      /* (BRINGUP) */
-               if(kdb_cpu != cpu_number()) {
-                       trappedalready = 1;
-                       panic("i386_cpu_IPI: sending enter debugger signal (%08X) to cpu %d and I do not own debugger, owner = %08X\n", 
-                               cpu_datap(cpu)->cpu_signals, cpu, kdb_cpu);
+       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 (panic_active() || NMI_panic_reason != NONE) {
+               mp_kdp_wait(FALSE, TRUE);
+       } else if (!mp_kdp_trap &&
+           !mp_kdp_is_NMI &&
+           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.
+                */
+               if (__sync_bool_compare_and_swap(&mp_kdp_is_NMI, FALSE, TRUE)) {
+                       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");
+                       mp_kdp_is_NMI = FALSE;
+               } else {
+                       mp_kdp_wait(FALSE, FALSE);
                }
+       } else {
+               mp_kdp_wait(FALSE, FALSE);
        }
-#endif /* MP_DEBUG */
-#endif
-
-       /* Wait for previous interrupt to be delivered... */
-#ifdef MP_DEBUG
-       int     pending_busy_count = 0;
-       while (LAPIC_READ(ICR) & LAPIC_ICR_DS_PENDING) {
-               if (++pending_busy_count > max_lock_loops)
-                       panic("i386_cpu_IPI() deadlock\n");
-#else
-       while (LAPIC_READ(ICR) & LAPIC_ICR_DS_PENDING) {
-#endif /* MP_DEBUG */
-               cpu_pause();
+       if (pmsafe_debug && !kdp_snapshot) {
+               pmSafeMode(&current_cpu_datap()->lcpu, PM_SAFE_FL_NORMAL);
        }
-
-       state = ml_set_interrupts_enabled(FALSE);
-       LAPIC_WRITE(ICRD, cpu_to_lapic[cpu] << LAPIC_ICRD_DEST_SHIFT);
-       LAPIC_WRITE(ICR, LAPIC_VECTOR(INTERPROCESSOR) | LAPIC_ICR_DM_FIXED);
-       (void) ml_set_interrupts_enabled(state);
+#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
@@ -538,10 +654,15 @@ i386_cpu_IPI(int cpu)
 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);
 }
 
 /*
@@ -550,29 +671,53 @@ cpu_interrupt(int cpu)
 void
 cpu_NMI_interrupt(int cpu)
 {
-       boolean_t       state;
-
        if (smp_initialized) {
-               state = ml_set_interrupts_enabled(FALSE);
-/* Program the interrupt command register */
-               LAPIC_WRITE(ICRD, cpu_to_lapic[cpu] << LAPIC_ICRD_DEST_SHIFT);
-/* The vector is ignored in this case--the target CPU will enter on the
- * NMI vector.
- */
-               LAPIC_WRITE(ICR, LAPIC_VECTOR(INTERPROCESSOR)|LAPIC_ICR_DM_NMI);
-               (void) ml_set_interrupts_enabled(state);
+               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);
+       NMIPI_enable(TRUE);
+       for (cpu = 0; cpu < real_ncpus; cpu++) {
+               if (!cpu_is_running(cpu)) {
+                       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;
        }
+       NMIPI_enable(FALSE);
+
+       ml_set_interrupts_enabled(intrs_enabled);
 }
 
-static void    (* volatile mp_PM_func)(void) = NULL;
+static void(*volatile mp_PM_func)(void) = NULL;
 
 static void
 mp_call_PM(void)
 {
        assert(!ml_get_interrupts_enabled());
 
-       if (mp_PM_func != NULL)
+       if (mp_PM_func != NULL) {
                mp_PM_func();
+       }
 }
 
 void
@@ -581,10 +726,11 @@ cpu_PM_interrupt(int cpu)
        assert(!ml_get_interrupts_enabled());
 
        if (mp_PM_func != NULL) {
-               if (cpu == cpu_number())
+               if (cpu == cpu_number()) {
                        mp_PM_func();
-               else
+               } else {
                        i386_signal_cpu(cpu, MP_CALL_PM, ASYNC);
+               }
        }
 }
 
@@ -597,77 +743,98 @@ PM_interrupt_register(void (*fn)(void))
 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;
+       volatile int    *signals = &cpu_datap(cpu)->cpu_signals;
+       uint64_t        tsc_timeout;
 
-       
-       if (!cpu_datap(cpu)->cpu_running)
+
+       if (!cpu_datap(cpu)->cpu_running) {
                return;
+       }
 
-       if (event == MP_TLB_FLUSH)
-               KERNEL_DEBUG(0xef800020 | DBG_FUNC_START, cpu, 0, 0, 0, 0);
+       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 = rdtsc64() + (1000*1000*1000);
+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);
+                           cpu, event);
                        goto again;
                }
        }
-       if (event == MP_TLB_FLUSH)
-               KERNEL_DEBUG(0xef800020 | DBG_FUNC_END, cpu, 0, 0, 0, 0);
+       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.
+ * Helper function called when busy-waiting: panic if too long
+ * a TSC-based time has elapsed since the start of the spin.
  */
-void
-i386_signal_cpus(mp_event_t event, mp_sync_t mode)
+static boolean_t
+mp_spin_timeout(uint64_t tsc_start)
 {
-       unsigned int    cpu;
-       unsigned int    my_cpu = cpu_number();
-
-       assert(hw_lock_held((hw_lock_t)&x86_topo_lock));
+       uint64_t        tsc_timeout;
 
-       for (cpu = 0; cpu < real_ncpus; cpu++) {
-               if (cpu == my_cpu || !cpu_datap(cpu)->cpu_running)
-                       continue;
-               i386_signal_cpu(cpu, event, mode);
+       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 ? LockTimeOutTSC << 2
+               : LockTimeOutTSC << 4;
+       return rdtsc64() > tsc_start + tsc_timeout;
 }
 
 /*
- * Return the number of running cpus.
- * Called with the topology locked.
+ * Helper function to take a spinlock while ensuring that incoming IPIs
+ * are still serviced if interrupts are masked while we spin.
+ * Returns current interrupt state.
  */
-int
-i386_active_cpus(void)
+boolean_t
+mp_safe_spin_lock(usimple_lock_t lock)
 {
-       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++;
+       if (ml_get_interrupts_enabled()) {
+               simple_lock(lock, LCK_GRP_NULL);
+               return TRUE;
+       } else {
+               uint64_t tsc_spin_start = rdtsc64();
+               while (!simple_lock_try(lock, LCK_GRP_NULL)) {
+                       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);
+                               NMIPI_panic(cpu_to_cpumask(lock_cpu), SPINLOCK_TIMEOUT);
+                               panic("mp_safe_spin_lock() timed out, lock: %p, owner thread: 0x%lx, current_thread: %p, owner on CPU 0x%x, time: %llu",
+                                   lock, lowner, current_thread(), lock_cpu, mach_absolute_time());
+                       }
+               }
+               return FALSE;
        }
-       return(ncpus);
 }
 
 /*
  * All-CPU rendezvous:
- *     - CPUs are signalled,
+ *      - 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),
@@ -680,62 +847,90 @@ i386_active_cpus(void)
  */
 
 static void
-mp_rendezvous_action(void)
+mp_rendezvous_action(__unused void *null)
 {
-       boolean_t intrs_enabled;
+       boolean_t       intrs_enabled;
+       uint64_t        tsc_spin_start;
+
+       /*
+        * Note that mp_rv_lock was acquired by the thread that initiated the
+        * rendezvous and must have been acquired before we enter
+        * mp_rendezvous_action().
+        */
+       current_cpu_datap()->cpu_rendezvous_in_progress = TRUE;
 
        /* setup function */
-       if (mp_rv_setup_func != NULL)
+       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)
+               if (!intrs_enabled) {
                        handle_pending_TLB_flushes();
-               cpu_pause();
+               }
+               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)
+       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)
+               if (!intrs_enabled) {
                        handle_pending_TLB_flushes();
-               cpu_pause();
+               }
+               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)
+       if (mp_rv_teardown_func != NULL) {
                mp_rv_teardown_func(mp_rv_func_arg);
+       }
+
+       current_cpu_datap()->cpu_rendezvous_in_progress = FALSE;
 
        /* 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)
+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)
+               if (setup_func != NULL) {
                        setup_func(arg);
-               if (action_func != NULL)
+               }
+               if (action_func != NULL) {
                        action_func(arg);
-               if (teardown_func != NULL)
+               }
+               if (teardown_func != NULL) {
                        teardown_func(arg);
+               }
                return;
        }
-               
+
        /* obtain rendezvous lock */
-       simple_lock(&mp_rv_lock);
+       mp_rendezvous_lock();
 
        /* set static function pointers */
        mp_rv_setup_func = setup_func;
@@ -751,23 +946,23 @@ mp_rendezvous(void (*setup_func)(void *),
         * signal other processors, which will call mp_rendezvous_action()
         * with interrupts disabled
         */
-       simple_lock(&x86_topo_lock);
-       mp_rv_ncpus = i386_active_cpus();
-       i386_signal_cpus(MP_RENDEZVOUS, ASYNC);
-       simple_unlock(&x86_topo_lock);
+       mp_rv_ncpus = mp_cpus_call(CPUMASK_OTHERS, NOSYNC, &mp_rendezvous_action, NULL) + 1;
 
        /* call executor function on this cpu */
-       mp_rendezvous_action();
+       mp_rendezvous_action(NULL);
 
        /*
         * 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) {
-               cpu_pause();
+               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;
@@ -775,6 +970,18 @@ mp_rendezvous(void (*setup_func)(void *),
        mp_rv_func_arg = NULL;
 
        /* release lock */
+       mp_rendezvous_unlock();
+}
+
+void
+mp_rendezvous_lock(void)
+{
+       (void) mp_safe_spin_lock(&mp_rv_lock);
+}
+
+void
+mp_rendezvous_unlock(void)
+{
        simple_unlock(&mp_rv_lock);
 }
 
@@ -808,208 +1015,510 @@ teardown_restore_intrs(__unused void * param_not_used)
  */
 void
 mp_rendezvous_no_intrs(
-             void (*action_func)(void *),
-             void *arg)
+       void (*action_func)(void *),
+       void *arg)
 {
        mp_rendezvous(setup_disable_intrs,
-                     action_func,
-                     teardown_restore_intrs,
-                     arg);     
+           action_func,
+           teardown_restore_intrs,
+           arg);
 }
 
-void
-handle_pending_TLB_flushes(void)
-{
-       volatile int    *my_word = &current_cpu_datap()->cpu_signals;
 
-       if (i_bit(MP_TLB_FLUSH, my_word)) {
-               DBGLOG(cpu_handle, cpu_number(), MP_TLB_FLUSH);
-               i_bit_clear(MP_TLB_FLUSH, my_word);
-               pmap_update_interrupt();
-       }
-}
+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;
 
-/*
- * This is called from cpu_signal_handler() to process an MP_CALL signal.
- */
-static void
-mp_cpus_call_action(void)
+
+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)
 {
-       if (mp_rv_action_func != NULL)
-               mp_rv_action_func(mp_rv_func_arg);
-       atomic_incl(&mp_rv_complete, 1);
+       boolean_t       intrs_enabled;
+
+       intrs_enabled = ml_set_interrupts_enabled(FALSE);
+       simple_lock(&cqp->lock, LCK_GRP_NULL);
+
+       return intrs_enabled;
 }
 
 /*
- * mp_cpus_call() runs a given function on cpus specified in a given cpu mask.
- * If the mode is SYNC, the function is called serially on the target cpus
- * in logical cpu order. If the mode is ASYNC, the function is called in
- * parallel over the specified cpus.
- * The action function may be NULL.
- * The cpu mask may include the local cpu. Offline cpus are ignored.
- * Return does not occur until the function has completed on all cpus.
- * The return value is the number of cpus on which the function was called.
+ * Deliver an NMIPI to a set of processors to cause them to panic .
  */
-cpu_t
-mp_cpus_call(
-       cpumask_t       cpus,
-       mp_sync_t       mode,
-        void           (*action_func)(void *),
-        void           *arg)
+void
+NMIPI_panic(cpumask_t cpu_mask, NMI_reason_t why)
 {
-       cpu_t           cpu;
-       boolean_t       intrs_enabled = ml_get_interrupts_enabled();
-       boolean_t       call_self = FALSE;
-
-       if (!smp_initialized) {
-               if ((cpus & CPUMASK_SELF) == 0)
-                       return 0;
-               if (action_func != NULL) {
-                       (void) ml_set_interrupts_enabled(FALSE);
-                       action_func(arg);
-                       ml_set_interrupts_enabled(intrs_enabled);
-               }
-               return 1;
-       }
-               
-       /* obtain rendezvous lock */
-       simple_lock(&mp_rv_lock);
+       unsigned int cpu;
+       cpumask_t cpu_bit;
+       uint64_t deadline;
 
-       /* Use the rendezvous data structures for this call */
-       mp_rv_action_func = action_func;
-       mp_rv_func_arg = arg;
-       mp_rv_ncpus = 0;
-       mp_rv_complete = 0;
+       NMIPI_enable(TRUE);
+       NMI_panic_reason = why;
 
-       simple_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)
+       for (cpu = 0, cpu_bit = 1; cpu < real_ncpus; cpu++, cpu_bit <<= 1) {
+               if ((cpu_mask & cpu_bit) == 0) {
                        continue;
-               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) {
-                               (void) ml_set_interrupts_enabled(FALSE);
-                               action_func(arg);
-                               ml_set_interrupts_enabled(intrs_enabled);
-                       }
-               } else {
-                       /*
-                        * Bump count of other cpus called and signal this cpu.
-                        * Note: we signal asynchronously regardless of mode
-                        * because we wait on mp_rv_complete either here
-                        * (if mode == SYNC) or later (if mode == ASYNC).
-                        * While spinning, poll for TLB flushes if interrupts
-                        * are disabled.
-                        */
-                       mp_rv_ncpus++;
-                       i386_signal_cpu(cpu, MP_CALL, ASYNC);
-                       if (mode == SYNC) {
-                               simple_unlock(&x86_topo_lock);
-                               while (mp_rv_complete < mp_rv_ncpus) {
-                                       if (!intrs_enabled)
-                                               handle_pending_TLB_flushes();
-                                       cpu_pause();
-                               }
-                               simple_lock(&x86_topo_lock);
-                       }
                }
+               cpu_datap(cpu)->cpu_NMI_acknowledged = FALSE;
+               cpu_NMI_interrupt(cpu);
        }
-       simple_unlock(&x86_topo_lock);
 
-       /*
-        * If calls are being made asynchronously,
-        * make the local call now if needed, and then
-        * wait for all other cpus to finish their calls.
-        */
-       if (mode == ASYNC) {
-               if (call_self && action_func != NULL) {
-                       (void) ml_set_interrupts_enabled(FALSE);
-                       action_func(arg);
-                       ml_set_interrupts_enabled(intrs_enabled);
+       /* Wait (only so long) for NMi'ed cpus to respond */
+       deadline = mach_absolute_time() + LockTimeOut;
+       for (cpu = 0, cpu_bit = 1; cpu < real_ncpus; cpu++, cpu_bit <<= 1) {
+               if ((cpu_mask & cpu_bit) == 0) {
+                       continue;
                }
-               while (mp_rv_complete < mp_rv_ncpus) {
-                       if (!intrs_enabled)
-                               handle_pending_TLB_flushes();
+               while (!cpu_datap(cpu)->cpu_NMI_acknowledged &&
+                   mach_absolute_time() < deadline) {
                        cpu_pause();
                }
        }
-       
-       /* Determine the number of cpus called */
-       cpu = mp_rv_ncpus + (call_self ? 1 : 0);
-
-       simple_unlock(&mp_rv_lock);
-
-       return cpu;
 }
 
-static void
-mp_broadcast_action(void)
+#if MACH_ASSERT
+static inline boolean_t
+mp_call_head_is_locked(mp_call_queue_t *cqp)
 {
-   /* 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));
+       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);
+}
+
+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;
+
+       assert(ml_get_interrupts_enabled() == 0 || get_preemption_level() != 0);
+       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;
+
+                       cpus_unresponsive = cpus_called & ~(*cpus_responded);
+                       NMIPI_panic(cpus_unresponsive, CROSSCALL_TIMEOUT);
+                       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)
+{
+       cpu_t           cpu = 0;
+       boolean_t       intrs_enabled = FALSE;
+       boolean_t       call_self = FALSE;
+       cpumask_t       cpus_called = 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_is_running(cpu)) {
+                       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.
+                        */
+                       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 (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 start: 0x%llx, cur: 0x%llx",
+                                           tsc_spin_start, rdtsc64());
+                               }
+                               goto queue_call;
+                       }
+                       callp->maskp = (mode == NOSYNC) ? NULL : &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);
+               }
+       }
+
+       /* 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);
+       }
+
+       /* Safe to allow pre-emption now */
+       mp_enable_preemption();
+
+out:
+       if (call_self) {
+               cpus_called |= cpu_to_cpumask(cpu);
+               cpus_call_count++;
+       }
+
+       if (cpus_calledp) {
+               *cpus_calledp = cpus_called;
+       }
+
+       KERNEL_DEBUG_CONSTANT(
+               TRACE_MP_CPUS_CALL | DBG_FUNC_END,
+               cpus_call_count, cpus_called, 0, 0, 0);
+
+       return (cpu_t) cpus_call_count;
+}
+
+
+static void
+mp_broadcast_action(__unused void *null)
+{
+       /* 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.
+ * The caller will also block if there is another pending broadcast.
  */
 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 (*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()
+        */
+       mp_bc_count = real_ncpus;                       /* assume max possible active */
+       mp_bc_ncpus = mp_cpus_call(CPUMASK_ALL, NOSYNC, *mp_broadcast_action, NULL);
+       atomic_decl(&mp_bc_count, real_ncpus - mp_bc_ncpus); /* subtract inactive */
+
+       /* block for other 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_is_running(cpu)) {
+                       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();
+       cpu_data_t      *cdp = current_cpu_datap();
 
        assert(!ml_get_interrupts_enabled());
 
@@ -1018,56 +1527,79 @@ i386_activate_cpu(void)
                return;
        }
 
-       simple_lock(&x86_topo_lock);
+       mp_safe_spin_lock(&x86_topo_lock);
        cdp->cpu_running = TRUE;
+       started_cpu();
+       pmap_tlbi_range(0, ~0ULL, true, 0);
        simple_unlock(&x86_topo_lock);
 }
 
-extern void etimer_timer_expire(void   *arg);
-
 void
 i386_deactivate_cpu(void)
 {
-       cpu_data_t      *cdp = current_cpu_datap();
+       cpu_data_t      *cdp = current_cpu_datap();
 
        assert(!ml_get_interrupts_enabled());
 
-       simple_lock(&x86_topo_lock);
+       KERNEL_DEBUG_CONSTANT(
+               TRACE_MP_CPU_DEACTIVATE | DBG_FUNC_START,
+               0, 0, 0, 0, 0);
+
+       mp_safe_spin_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);
-       cdp->rtclock_timer.deadline = EndOfAllTime;
-       mp_cpus_call(cpu_to_cpumask(master_cpu), ASYNC, etimer_timer_expire, NULL);
+       mp_cpus_call(cpu_to_cpumask(master_cpu), ASYNC, timer_queue_expire_local, NULL);
+
+#if MONOTONIC
+       mt_cpu_down(cdp);
+#endif /* MONOTONIC */
 
        /*
-        * In case a rendezvous/braodcast/call was initiated to this cpu
-        * before we cleared cpu_running, we must perform any actions due.
+        * 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.
         */
-       if (i_bit(MP_RENDEZVOUS, &cdp->cpu_signals))
-               mp_rendezvous_action();
-       if (i_bit(MP_BROADCAST, &cdp->cpu_signals))
-               mp_broadcast_action();
-       if (i_bit(MP_CALL, &cdp->cpu_signals))
-               mp_cpus_call_action();
-       cdp->cpu_signals = 0;                   /* all clear */
+       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;
+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;
+#if     MACH_KDP
+volatile boolean_t      mp_kdp_trap = FALSE;
+volatile boolean_t      mp_kdp_is_NMI = FALSE;
+volatile unsigned long  mp_kdp_ncpus;
+boolean_t               mp_kdp_state;
 
 
 void
-mp_kdp_enter(void)
+mp_kdp_enter(boolean_t proceed_on_failure)
 {
-       unsigned int    cpu;
-       unsigned int    ncpus;
-       unsigned int    my_cpu;
-       uint64_t        tsc_timeout;
+       unsigned int    cpu;
+       unsigned int    ncpus = 0;
+       unsigned int    my_cpu;
+       uint64_t        tsc_timeout;
 
        DBG("mp_kdp_enter()\n");
 
@@ -1077,33 +1609,63 @@ mp_kdp_enter(void)
         * stopping others.
         */
        mp_kdp_state = ml_set_interrupts_enabled(FALSE);
-       simple_lock(&mp_kdp_lock);
-       debugger_entry_time = mach_absolute_time();
-       if (pmsafe_debug)
-           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");
+       my_cpu = cpu_number();
+
+       if (my_cpu == (unsigned) debugger_cpu) {
+               kprintf("\n\nRECURSIVE DEBUGGER ENTRY DETECTED\n\n");
+               kdp_reset();
+               return;
+       }
+
+       uint64_t start_time = cpu_datap(my_cpu)->debugger_entry_time = mach_absolute_time();
+       int locked = 0;
+       while (!locked || mp_kdp_trap) {
+               if (locked) {
+                       simple_unlock(&x86_topo_lock);
+               }
+               if (proceed_on_failure) {
+                       if (mach_absolute_time() - start_time > 500000000ll) {
+                               paniclog_append_noflush("mp_kdp_enter() can't get x86_topo_lock! Debugging anyway! #YOLO\n");
+                               break;
+                       }
+                       locked = simple_lock_try(&x86_topo_lock, LCK_GRP_NULL);
+                       if (!locked) {
+                               cpu_pause();
+                       }
+               } else {
+                       mp_safe_spin_lock(&x86_topo_lock);
+                       locked = TRUE;
+               }
+
+               if (locked && mp_kdp_trap) {
+                       simple_unlock(&x86_topo_lock);
+                       DBG("mp_kdp_enter() race lost\n");
 #if MACH_KDP
-               mp_kdp_wait(TRUE, FALSE);
+                       mp_kdp_wait(TRUE, FALSE);
 #endif
-               simple_lock(&mp_kdp_lock);
+                       locked = FALSE;
+               }
        }
-       my_cpu = cpu_number();
+
+       if (pmsafe_debug && !kdp_snapshot) {
+               pmSafeMode(&current_cpu_datap()->lcpu, PM_SAFE_FL_SAFE);
+       }
+
        debugger_cpu = my_cpu;
-       mp_kdp_ncpus = 1;       /* self */
+       ncpus = 1;
+       atomic_incl((volatile long *)&mp_kdp_ncpus, 1);
        mp_kdp_trap = TRUE;
-       simple_unlock(&mp_kdp_lock);
+       debugger_entry_time = cpu_datap(my_cpu)->debugger_entry_time;
 
        /*
         * Deliver a nudge to other cpus, counting how many
         */
        DBG("mp_kdp_enter() signaling other processors\n");
        if (force_immediate_debugger_NMI == FALSE) {
-               for (ncpus = 1, cpu = 0; cpu < real_ncpus; cpu++) {
-                       if (cpu == my_cpu || !cpu_datap(cpu)->cpu_running)
+               for (cpu = 0; cpu < real_ncpus; cpu++) {
+                       if (cpu == my_cpu || !cpu_is_running(cpu)) {
                                continue;
+                       }
                        ncpus++;
                        i386_signal_cpu(cpu, MP_KDP, ASYNC);
                }
@@ -1118,7 +1680,7 @@ mp_kdp_enter(void)
                 * "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);
+               tsc_timeout = rdtsc64() + (LockTimeOutTSC);
 
                while (mp_kdp_ncpus != ncpus && rdtsc64() < tsc_timeout) {
                        /*
@@ -1131,69 +1693,134 @@ mp_kdp_enter(void)
                        cpu_pause();
                }
                /* If we've timed out, and some processor(s) are still unresponsive,
-                * interrupt them with an NMI via the local APIC.
+                * interrupt them with an NMI via the local APIC, iff a panic is
+                * in progress.
                 */
+               if (panic_active()) {
+                       NMIPI_enable(TRUE);
+               }
                if (mp_kdp_ncpus != ncpus) {
+                       unsigned int wait_cycles = 0;
+                       if (proceed_on_failure) {
+                               paniclog_append_noflush("mp_kdp_enter() timed-out on cpu %d, NMI-ing\n", my_cpu);
+                       } else {
+                               DBG("mp_kdp_enter() timed-out on cpu %d, NMI-ing\n", my_cpu);
+                       }
                        for (cpu = 0; cpu < real_ncpus; cpu++) {
-                               if (cpu == my_cpu || !cpu_datap(cpu)->cpu_running)
+                               if (cpu == my_cpu || !cpu_is_running(cpu)) {
                                        continue;
-                               if (cpu_signal_pending(cpu, MP_KDP))
+                               }
+                               if (cpu_signal_pending(cpu, MP_KDP)) {
+                                       cpu_datap(cpu)->cpu_NMI_acknowledged = FALSE;
                                        cpu_NMI_interrupt(cpu);
+                               }
+                       }
+                       /* Wait again for the same timeout */
+                       tsc_timeout = rdtsc64() + (LockTimeOutTSC);
+                       while (mp_kdp_ncpus != ncpus && rdtsc64() < tsc_timeout) {
+                               handle_pending_TLB_flushes();
+                               cpu_pause();
+                               ++wait_cycles;
+                       }
+                       if (mp_kdp_ncpus != ncpus) {
+                               paniclog_append_noflush("mp_kdp_enter() NMI pending on cpus:");
+                               for (cpu = 0; cpu < real_ncpus; cpu++) {
+                                       if (cpu_is_running(cpu) && !cpu_datap(cpu)->cpu_NMI_acknowledged) {
+                                               paniclog_append_noflush(" %d", cpu);
+                                       }
+                               }
+                               paniclog_append_noflush("\n");
+                               if (proceed_on_failure) {
+                                       paniclog_append_noflush("mp_kdp_enter() timed-out during %s wait after NMI;"
+                                           "expected %u acks but received %lu after %u loops in %llu ticks\n",
+                                           (locked ? "locked" : "unlocked"), ncpus, mp_kdp_ncpus, wait_cycles, LockTimeOutTSC);
+                               } else {
+                                       panic("mp_kdp_enter() timed-out during %s wait after NMI;"
+                                           "expected %u acks but received %lu after %u loops in %llu ticks",
+                                           (locked ? "locked" : "unlocked"), ncpus, mp_kdp_ncpus, wait_cycles, LockTimeOutTSC);
+                               }
                        }
                }
-       }
-       else
+       } else {
                for (cpu = 0; cpu < real_ncpus; cpu++) {
-                       if (cpu == my_cpu || !cpu_datap(cpu)->cpu_running)
+                       if (cpu == my_cpu || !cpu_is_running(cpu)) {
                                continue;
+                       }
                        cpu_NMI_interrupt(cpu);
                }
+       }
+
+       if (locked) {
+               simple_unlock(&x86_topo_lock);
+       }
+
+       DBG("mp_kdp_enter() %d processors done %s\n",
+           (int)mp_kdp_ncpus, (mp_kdp_ncpus == ncpus) ? "OK" : "timed out");
 
-       DBG("mp_kdp_enter() %u processors done %s\n",
-           mp_kdp_ncpus, (mp_kdp_ncpus == ncpus) ? "OK" : "timed out");
-       
        postcode(MP_KDP_ENTER);
 }
 
+boolean_t
+mp_kdp_all_cpus_halted()
+{
+       unsigned int ncpus = 0, cpu = 0, my_cpu = 0;
+
+       my_cpu = cpu_number();
+       ncpus = 1; /* current CPU */
+       for (cpu = 0; cpu < real_ncpus; cpu++) {
+               if (cpu == my_cpu || !cpu_is_running(cpu)) {
+                       continue;
+               }
+               ncpus++;
+       }
+
+       return mp_kdp_ncpus == ncpus;
+}
+
 static boolean_t
 cpu_signal_pending(int cpu, mp_event_t event)
 {
-       volatile int    *signals = &cpu_datap(cpu)->cpu_signals;
+       volatile int    *signals = &cpu_datap(cpu)->cpu_signals;
        boolean_t retval = FALSE;
 
-       if (i_bit(event, signals))
+       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)
+long
+kdp_x86_xcpu_invoke(const uint16_t lcpu, kdp_x86_xcpu_func_t func,
+    void *arg0, void *arg1)
 {
-       if (lcpu > (real_ncpus - 1))
+       if (lcpu > (real_ncpus - 1)) {
                return -1;
+       }
 
-        if (func == NULL)
+       if (func == NULL) {
                return -1;
+       }
 
        kdp_xcpu_call_func.func = func;
-        kdp_xcpu_call_func.ret  = -1;
+       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)
+       while (kdp_xcpu_call_func.cpu != KDP_XCPU_NONE) {
                cpu_pause();
-        return kdp_xcpu_call_func.ret;
+       }
+       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.ret =
                    kdp_xcpu_call_func.func(kdp_xcpu_call_func.arg0,
-                                           kdp_xcpu_call_func.arg1,
-                                           cpu_number());
+                   kdp_xcpu_call_func.arg1,
+                   cpu_number());
                kdp_xcpu_call_func.cpu = KDP_XCPU_NONE;
        }
 }
@@ -1202,35 +1829,29 @@ 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
 
-       if (pmsafe_debug)
-           pmSafeMode(&current_cpu_datap()->lcpu, PM_SAFE_FL_SAFE);
-
        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)
+               if (flush) {
                        handle_pending_TLB_flushes();
+               }
 
                kdp_x86_xcpu_poll();
                cpu_pause();
        }
 
-       if (pmsafe_debug)
-           pmSafeMode(&current_cpu_datap()->lcpu, PM_SAFE_FL_NORMAL);
-
        atomic_decl((volatile long *)&mp_kdp_ncpus, 1);
        DBG("mp_kdp_wait() done\n");
 }
@@ -1245,221 +1866,302 @@ mp_kdp_exit(void)
        debugger_exit_time = mach_absolute_time();
 
        mp_kdp_trap = FALSE;
-       __asm__ volatile("mfence");
+       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();
+               handle_pending_TLB_flushes();
 
                cpu_pause();
        }
 
-       if (pmsafe_debug)
-           pmSafeMode(&current_cpu_datap()->lcpu, PM_SAFE_FL_NORMAL);
+       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);
+       postcode(MP_KDP_EXIT);
 }
-#endif /* MACH_KDP */
+
+#endif  /* MACH_KDP */
 
 boolean_t
-mp_recent_debugger_activity() {
-       return (((mach_absolute_time() - debugger_entry_time) < LastDebuggerEntryAllowance) ||
-           ((mach_absolute_time() - debugger_exit_time) < LastDebuggerEntryAllowance));
+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)
+       __unused processor_t    processor)
 {
 }
 
 void
 cause_ast_check(
-       processor_t     processor)
+       processor_t     processor)
 {
-       int     cpu = processor->cpu_id;
+       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);
        }
 }
 
-#if MACH_KDB
-/*
- * invoke kdb on slave processors 
- */
-
 void
-remote_kdb(void)
-{
-       unsigned int    my_cpu = cpu_number();
-       unsigned int    cpu;
-       int kdb_ncpus;
-       uint64_t tsc_timeout = 0;
-       
-       mp_kdb_trap = TRUE;
-       mp_kdb_ncpus = 1;
-       for (kdb_ncpus = 1, cpu = 0; cpu < real_ncpus; cpu++) {
-               if (cpu == my_cpu || !cpu_datap(cpu)->cpu_running)
-                       continue;
-               kdb_ncpus++;
-               i386_signal_cpu(cpu, MP_KDB, ASYNC);
-       }
-       DBG("remote_kdb() waiting for (%d) processors to suspend\n",kdb_ncpus);
-
-       tsc_timeout = rdtsc64() + (kdb_ncpus * 100 * 1000 * 1000);
+slave_machine_init(void *param)
+{
+       /*
+        * Here in process context, but with interrupts disabled.
+        */
+       DBG("slave_machine_init() CPU%d\n", get_cpu_number());
 
-       while (mp_kdb_ncpus != kdb_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()
+       if (param == FULL_SLAVE_INIT) {
+               /*
+                * Cold start
                 */
-               handle_pending_TLB_flushes();
-
-               cpu_pause();
+               clock_init();
        }
-       DBG("mp_kdp_enter() %d processors done %s\n",
-               mp_kdb_ncpus, (mp_kdb_ncpus == kdb_ncpus) ? "OK" : "timed out");
+       cpu_machine_init();     /* Interrupts enabled hereafter */
+}
+
+#undef cpu_number
+int
+cpu_number(void)
+{
+       return get_cpu_number();
 }
 
 static void
-mp_kdb_wait(void)
+cpu_prewarm_init()
 {
-       DBG("mp_kdb_wait()\n");
+       int i;
 
-       /* If an I/O port has been specified as a debugging aid, issue a read */
-       panic_io_port_read();
+       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]);
+       }
+}
 
-       atomic_incl(&mp_kdb_ncpus, 1);
-       while (mp_kdb_trap) {
-               /*
-                * 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();
+static timer_call_t
+grab_warm_timer_call()
+{
+       spl_t x;
+       timer_call_t call = NULL;
 
-               cpu_pause();
+       x = splsched();
+       simple_lock(&cpu_warm_lock, LCK_GRP_NULL);
+       if (!queue_empty(&cpu_warm_call_list)) {
+               call = (timer_call_t) dequeue_head(&cpu_warm_call_list);
        }
-       atomic_decl((volatile long *)&mp_kdb_ncpus, 1);
-       DBG("mp_kdb_wait() done\n");
+       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, LCK_GRP_NULL);
+       enqueue_head(&cpu_warm_call_list, (queue_entry_t)call);
+       simple_unlock(&cpu_warm_lock);
+       splx(x);
 }
 
 /*
- * Clear kdb interrupt
+ * Runs in timer call context (interrupts disabled).
  */
-
-void
-clear_kdb_intr(void)
+static void
+cpu_warm_timer_call_func(
+       call_entry_param_t p0,
+       __unused call_entry_param_t p1)
 {
-       mp_disable_preemption();
-       i_bit_clear(MP_KDB, &current_cpu_datap()->cpu_signals);
-       mp_enable_preemption();
+       free_warm_timer_call((timer_call_t)p0);
+       return;
 }
 
-void
-mp_kdb_exit(void)
+/*
+ * Runs with interrupts disabled on the CPU we wish to warm (i.e. CPU 0).
+ */
+static void
+_cpu_warm_setup(
+       void *arg)
 {
-       DBG("mp_kdb_exit()\n");
-       atomic_decl((volatile long *)&mp_kdb_ncpus, 1);
-       mp_kdb_trap = FALSE;
-       __asm__ volatile("mfence");
+       cpu_warm_data_t cwdp = (cpu_warm_data_t)arg;
 
-       while (mp_kdb_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();
-       }
+       timer_call_enter(cwdp->cwd_call, cwdp->cwd_deadline, TIMER_CALL_SYS_CRITICAL | TIMER_CALL_LOCAL);
+       cwdp->cwd_result = 0;
 
-       DBG("mp_kdb_exit() done\n");
+       return;
 }
 
-#endif /* MACH_KDB */
-
-void
-slave_machine_init(void *param)
+/*
+ * 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__);
+       }
+
        /*
-        * Here in process context, but with interrupts disabled.
+        * If the platform doesn't need our help, say that we succeeded.
         */
-       DBG("slave_machine_init() CPU%d\n", get_cpu_number());
+       if (!ml_get_interrupt_prewake_applicable()) {
+               return KERN_SUCCESS;
+       }
 
-       if (param == FULL_SLAVE_INIT) {
-               /*
-                * Cold start
-                */
-               clock_init();
+       /*
+        * Grab a timer call to use.
+        */
+       call = grab_warm_timer_call();
+       if (call == NULL) {
+               return KERN_RESOURCE_SHORTAGE;
+       }
 
-               cpu_machine_init();     /* Interrupts enabled hereafter */
+       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;
        }
 }
 
-#undef cpu_number
-int cpu_number(void)
+#if DEBUG || DEVELOPMENT
+void
+kernel_spin(uint64_t spin_ns)
 {
-       return get_cpu_number();
-}
+       boolean_t       istate;
+       uint64_t        spin_abs;
+       uint64_t        deadline;
+       cpu_data_t      *cdp;
 
-#if    MACH_KDB
-#include <ddb/db_output.h>
+       kprintf("kernel_spin(%llu) spinning uninterruptibly\n", spin_ns);
+       istate = ml_set_interrupts_enabled(FALSE);
+       cdp = current_cpu_datap();
+       nanoseconds_to_absolutetime(spin_ns, &spin_abs);
 
-#define TRAP_DEBUG 0 /* Must match interrupt.s and spl.s */
+       /* Fake interrupt handler entry for testing mp_interrupt_watchdog() */
+       cdp->cpu_int_event_time = mach_absolute_time();
+       cdp->cpu_int_state = (void *) USER_STATE(current_thread());
 
+       deadline = mach_absolute_time() + spin_ns;
+       while (mach_absolute_time() < deadline) {
+               cpu_pause();
+       }
 
-#if    TRAP_DEBUG
-#define MTRAPS 100
-struct mp_trap_hist_struct {
-       unsigned char type;
-       unsigned char data[5];
-} trap_hist[MTRAPS], *cur_trap_hist = trap_hist,
-    *max_trap_hist = &trap_hist[MTRAPS];
+       cdp->cpu_int_event_time = 0;
+       cdp->cpu_int_state = NULL;
 
-void db_trap_hist(void);
+       ml_set_interrupts_enabled(istate);
+       kprintf("kernel_spin() continuing\n");
+}
 
 /*
- * SPL:
- *     1: new spl
- *     2: old spl
- *     3: new tpr
- *     4: old tpr
- * INT:
- *     1: int vec
- *     2: old spl
- *     3: new spl
- *     4: post eoi tpr
- *     5: exit tpr
+ * Called from the scheduler's maintenance thread,
+ * scan running processors for long-running ISRs and:
+ *  - panic if longer than LockTimeOut, or
+ *  - log if more than a quantum.
  */
-
 void
-db_trap_hist(void)
-{
-       int i,j;
-       for(i=0;i<MTRAPS;i++)
-           if (trap_hist[i].type == 1 || trap_hist[i].type == 2) {
-                   db_printf("%s%s",
-                             (&trap_hist[i]>=cur_trap_hist)?"*":" ",
-                             (trap_hist[i].type == 1)?"SPL":"INT");
-                   for(j=0;j<5;j++)
-                       db_printf(" %02x", trap_hist[i].data[j]);
-                   db_printf("\n");
-           }
-               
-}
-#endif /* TRAP_DEBUG */
-#endif /* MACH_KDB */
+mp_interrupt_watchdog(void)
+{
+       cpu_t                   cpu;
+       boolean_t               intrs_enabled = FALSE;
+       uint16_t                cpu_int_num;
+       uint64_t                cpu_int_event_time;
+       uint64_t                cpu_rip;
+       uint64_t                cpu_int_duration;
+       uint64_t                now;
+       x86_saved_state_t       *cpu_int_state;
+
+       if (__improbable(!mp_interrupt_watchdog_enabled)) {
+               return;
+       }
+
+       intrs_enabled = ml_set_interrupts_enabled(FALSE);
+       now = mach_absolute_time();
+       /*
+        * While timeouts are not suspended,
+        * check all other processors for long outstanding interrupt handling.
+        */
+       for (cpu = 0;
+           cpu < (cpu_t) real_ncpus && !machine_timeout_suspended();
+           cpu++) {
+               if ((cpu == (cpu_t) cpu_number()) ||
+                   (!cpu_is_running(cpu))) {
+                       continue;
+               }
+               cpu_int_event_time = cpu_datap(cpu)->cpu_int_event_time;
+               if (cpu_int_event_time == 0) {
+                       continue;
+               }
+               if (__improbable(now < cpu_int_event_time)) {
+                       continue;       /* skip due to inter-processor skew */
+               }
+               cpu_int_state = cpu_datap(cpu)->cpu_int_state;
+               if (__improbable(cpu_int_state == NULL)) {
+                       /* The interrupt may have been dismissed */
+                       continue;
+               }
 
+               /* Here with a cpu handling an interrupt */
+
+               cpu_int_duration = now - cpu_int_event_time;
+               if (__improbable(cpu_int_duration > LockTimeOut)) {
+                       cpu_int_num = saved_state64(cpu_int_state)->isf.trapno;
+                       cpu_rip = saved_state64(cpu_int_state)->isf.rip;
+                       vector_timed_out = cpu_int_num;
+                       NMIPI_panic(cpu_to_cpumask(cpu), INTERRUPT_WATCHDOG);
+                       panic("Interrupt watchdog, "
+                           "cpu: %d interrupt: 0x%x time: %llu..%llu state: %p RIP: 0x%llx",
+                           cpu, cpu_int_num, cpu_int_event_time, now, cpu_int_state, cpu_rip);
+                       /* NOT REACHED */
+               } else if (__improbable(cpu_int_duration > (uint64_t) std_quantum)) {
+                       mp_interrupt_watchdog_events++;
+                       cpu_int_num = saved_state64(cpu_int_state)->isf.trapno;
+                       cpu_rip = saved_state64(cpu_int_state)->isf.rip;
+                       ml_set_interrupts_enabled(intrs_enabled);
+                       printf("Interrupt watchdog, "
+                           "cpu: %d interrupt: 0x%x time: %llu..%llu RIP: 0x%llx\n",
+                           cpu, cpu_int_num, cpu_int_event_time, now, cpu_rip);
+                       return;
+               }
+       }
+
+       ml_set_interrupts_enabled(intrs_enabled);
+}
+#endif