]> git.saurik.com Git - apple/xnu.git/blobdiff - osfmk/i386/mp.c
xnu-1228.5.20.tar.gz
[apple/xnu.git] / osfmk / i386 / mp.c
index a243e7619aad368f2da2b456182fdb5426a417b0..fb7afc4d05b9387fa8a0f798ec7744c4cf37d397 100644 (file)
@@ -1,5 +1,5 @@
 /*
 /*
- * Copyright (c) 2000-2005 Apple Computer, Inc. All rights reserved.
+ * Copyright (c) 2000-2007 Apple Inc. All rights reserved.
  *
  * @APPLE_OSREFERENCE_LICENSE_HEADER_START@
  * 
  *
  * @APPLE_OSREFERENCE_LICENSE_HEADER_START@
  * 
@@ -63,6 +63,7 @@
 #include <i386/machine_cpu.h>
 #include <i386/misc_protos.h>
 #include <i386/mtrr.h>
 #include <i386/machine_cpu.h>
 #include <i386/misc_protos.h>
 #include <i386/mtrr.h>
+#include <i386/vmx/vmx_cpu.h>
 #include <i386/postcode.h>
 #include <i386/perfmon.h>
 #include <i386/cpu_threads.h>
 #include <i386/postcode.h>
 #include <i386/perfmon.h>
 #include <i386/cpu_threads.h>
@@ -71,6 +72,7 @@
 #include <i386/machine_routines.h>
 #include <i386/pmCPU.h>
 #include <i386/hpet.h>
 #include <i386/machine_routines.h>
 #include <i386/pmCPU.h>
 #include <i386/hpet.h>
+#include <i386/machine_check.h>
 
 #include <chud/chud_xnu.h>
 #include <chud/chud_xnu_private.h>
 
 #include <chud/chud_xnu.h>
 #include <chud/chud_xnu_private.h>
@@ -118,27 +120,41 @@ volatile boolean_t        mp_kdb_trap = FALSE;
 volatile long  mp_kdb_ncpus = 0;
 #endif
 
 volatile long  mp_kdb_ncpus = 0;
 #endif
 
-static void    mp_kdp_wait(void);
+static void    mp_kdp_wait(boolean_t flush);
 static void    mp_rendezvous_action(void);
 static void    mp_rendezvous_action(void);
+static void    mp_broadcast_action(void);
 
 
-static int     NMIInterruptHandler(void *regs);
+static int             NMIInterruptHandler(x86_saved_state_t *regs);
 static boolean_t       cpu_signal_pending(int cpu, mp_event_t event);
 static boolean_t       cpu_signal_pending(int cpu, mp_event_t event);
-static void    cpu_NMI_interrupt(int cpu);
+static void            cpu_NMI_interrupt(int cpu);
 
 boolean_t      smp_initialized = FALSE;
 
 boolean_t      smp_initialized = FALSE;
+boolean_t      force_immediate_debugger_NMI = FALSE;
 
 decl_simple_lock_data(,mp_kdp_lock);
 
 decl_mutex_data(static, mp_cpu_boot_lock);
 
 /* Variables needed for MP rendezvous. */
 
 decl_simple_lock_data(,mp_kdp_lock);
 
 decl_mutex_data(static, mp_cpu_boot_lock);
 
 /* 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 int             mp_rv_ncpus;
 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 int             mp_rv_ncpus;
-static volatile long   mp_rv_waiters[2];
-decl_simple_lock_data(,mp_rv_lock);
+                       /* 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)));
+
+/* Variables needed for MP broadcast. */
+static void        (*mp_bc_action_func)(void *arg);
+static void        *mp_bc_func_arg;
+static int     mp_bc_ncpus;
+static volatile long   mp_bc_count;
+decl_mutex_data(static, mp_bc_lock);
+
+static void    mp_cpus_call_action(void); 
 
 int            lapic_to_cpu[MAX_CPUS];
 int            cpu_to_lapic[MAX_CPUS];
 
 int            lapic_to_cpu[MAX_CPUS];
 int            cpu_to_lapic[MAX_CPUS];
@@ -238,6 +254,7 @@ smp_init(void)
        simple_lock_init(&mp_kdp_lock, 0);
        simple_lock_init(&mp_rv_lock, 0);
        mutex_init(&mp_cpu_boot_lock, 0);
        simple_lock_init(&mp_kdp_lock, 0);
        simple_lock_init(&mp_rv_lock, 0);
        mutex_init(&mp_cpu_boot_lock, 0);
+       mutex_init(&mp_bc_lock, 0);
        console_init();
 
        /* Local APIC? */
        console_init();
 
        /* Local APIC? */
@@ -257,7 +274,8 @@ smp_init(void)
 
        /* Establish a map to the local apic */
        lapic_start = vm_map_min(kernel_map);
 
        /* Establish a map to the local apic */
        lapic_start = vm_map_min(kernel_map);
-       result = vm_map_find_space(kernel_map, &lapic_start,
+       result = vm_map_find_space(kernel_map,
+                                  (vm_map_address_t *) &lapic_start,
                                   round_page(LAPIC_SIZE), 0,
                                   VM_MAKE_TAG(VM_MEMORY_IOKIT), &entry);
        if (result != KERN_SUCCESS) {
                                   round_page(LAPIC_SIZE), 0,
                                   VM_MAKE_TAG(VM_MEMORY_IOKIT), &entry);
        if (result != KERN_SUCCESS) {
@@ -613,7 +631,7 @@ lapic_interrupt(int interrupt, x86_saved_state_t *state)
        int     retval = 0;
 
        /* Did we just field an interruption for the HPET comparator? */
        int     retval = 0;
 
        /* Did we just field an interruption for the HPET comparator? */
-       if(current_cpu_datap()->cpu_pmHpetVec == ((uint32_t)interrupt - 0x40)) {
+       if(x86_core()->HpetVec == ((uint32_t)interrupt - 0x40)) {
                /* Yes, go handle it... */
                retval = HPETInterrupt();
                /* Was it really handled? */
                /* Yes, go handle it... */
                retval = HPETInterrupt();
                /* Was it really handled? */
@@ -780,8 +798,7 @@ intel_startCPU(
                cpu_shutdown();
                return KERN_SUCCESS;
        } else {
                cpu_shutdown();
                return KERN_SUCCESS;
        } else {
-               kprintf("Started cpu %d (lapic id %p)\n", slot_num, lapic);
-               printf("Started CPU %02d\n", slot_num);
+               kprintf("Started cpu %d (lapic id %08x)\n", slot_num, lapic);
                return KERN_SUCCESS;
        }
 }
                return KERN_SUCCESS;
        }
 }
@@ -867,7 +884,7 @@ cpu_signal_handler(x86_saved_state_t *regs)
  * XXX 64-bit state?
  */
                        sync_iss_to_iks(saved_state32(regs));
  * XXX 64-bit state?
  */
                        sync_iss_to_iks(saved_state32(regs));
-                       mp_kdp_wait();
+                       mp_kdp_wait(TRUE);
                } else
 #endif /* MACH_KDP */
                if (i_bit(MP_TLB_FLUSH, my_word)) {
                } else
 #endif /* MACH_KDP */
                if (i_bit(MP_TLB_FLUSH, my_word)) {
@@ -890,10 +907,18 @@ cpu_signal_handler(x86_saved_state_t *regs)
                        DBGLOG(cpu_handle,my_cpu,MP_RENDEZVOUS);
                        i_bit_clear(MP_RENDEZVOUS, my_word);
                        mp_rendezvous_action();
                        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_CHUD, my_word)) {
                        DBGLOG(cpu_handle,my_cpu,MP_CHUD);
                        i_bit_clear(MP_CHUD, my_word);
                        chudxnu_cpu_signal_handler();
+               } else if (i_bit(MP_CALL, my_word)) {
+                       DBGLOG(cpu_handle,my_cpu,MP_CALL);
+                       i_bit_clear(MP_CALL, my_word);
+                       mp_cpus_call_action();
                }
        } while (*my_word);
 
                }
        } while (*my_word);
 
@@ -901,66 +926,78 @@ cpu_signal_handler(x86_saved_state_t *regs)
 
 }
 
 
 }
 
-
-/* We want this to show up in backtraces, so mark it noinline
+/* We want this to show up in backtraces, hence marked noinline.
  */
 static int __attribute__((noinline))
  */
 static int __attribute__((noinline))
-NMIInterruptHandler(void *regs)
+NMIInterruptHandler(x86_saved_state_t *regs)
 {
        boolean_t state = ml_set_interrupts_enabled(FALSE);
        sync_iss_to_iks_unconditionally(regs);
 {
        boolean_t state = ml_set_interrupts_enabled(FALSE);
        sync_iss_to_iks_unconditionally(regs);
-       mp_kdp_wait();
+       mp_kdp_wait(FALSE);
        (void) ml_set_interrupts_enabled(state);
        return 1;
 }
 
 #ifdef MP_DEBUG
 extern int     max_lock_loops;
        (void) ml_set_interrupts_enabled(state);
        return 1;
 }
 
 #ifdef MP_DEBUG
 extern int     max_lock_loops;
+int            trappedalready = 0;     /* (BRINGUP */
 #endif /* MP_DEBUG */
 
 #endif /* MP_DEBUG */
 
-int trappedalready = 0;        /* (BRINGUP */
-
-void
-cpu_interrupt(int cpu)
+static void
+i386_cpu_IPI(int cpu)
 {
        boolean_t       state;
        
 {
        boolean_t       state;
        
+#ifdef MP_DEBUG
        if(cpu_datap(cpu)->cpu_signals & 6) {   /* (BRINGUP) */
        if(cpu_datap(cpu)->cpu_signals & 6) {   /* (BRINGUP) */
-               kprintf("cpu_interrupt: sending enter debugger signal (%08X) to cpu %d\n", cpu_datap(cpu)->cpu_signals, cpu);
+               kprintf("i386_cpu_IPI: sending enter debugger signal (%08X) to cpu %d\n", cpu_datap(cpu)->cpu_signals, cpu);
        }
        }
-
-       if (smp_initialized) {
+#endif /* MP_DEBUG */
 
 #if MACH_KDB
 
 #if MACH_KDB
-//             if(!trappedalready && (cpu_datap(cpu)->cpu_signals & 6)) {      /* (BRINGUP) */
-//                     if(kdb_cpu != cpu_number()) {
-//                             trappedalready = 1;
-//                             panic("cpu_interrupt: 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);
-//                     }
-//             }
+#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);
+               }
+       }
+#endif /* MP_DEBUG */
 #endif
 
 #endif
 
-               /* Wait for previous interrupt to be delivered... */
+       /* Wait for previous interrupt to be delivered... */
 #ifdef MP_DEBUG
 #ifdef MP_DEBUG
-               int     pending_busy_count = 0;
-               while (LAPIC_REG(ICR) & LAPIC_ICR_DS_PENDING) {
-                       if (++pending_busy_count > max_lock_loops)
-                               panic("cpus_interrupt() deadlock\n");
+       int     pending_busy_count = 0;
+       while (LAPIC_REG(ICR) & LAPIC_ICR_DS_PENDING) {
+               if (++pending_busy_count > max_lock_loops)
+                       panic("i386_cpu_IPI() deadlock\n");
 #else
 #else
-               while (LAPIC_REG(ICR) & LAPIC_ICR_DS_PENDING) {
+       while (LAPIC_REG(ICR) & LAPIC_ICR_DS_PENDING) {
 #endif /* MP_DEBUG */
 #endif /* MP_DEBUG */
-                       cpu_pause();
-               }
-
-               state = ml_set_interrupts_enabled(FALSE);
-               LAPIC_REG(ICRD) =
-                       cpu_to_lapic[cpu] << LAPIC_ICRD_DEST_SHIFT;
-               LAPIC_REG(ICR)  =
-                       LAPIC_VECTOR(INTERPROCESSOR) | LAPIC_ICR_DM_FIXED;
-               (void) ml_set_interrupts_enabled(state);
+               cpu_pause();
        }
 
        }
 
+       state = ml_set_interrupts_enabled(FALSE);
+       LAPIC_REG(ICRD) =
+               cpu_to_lapic[cpu] << LAPIC_ICRD_DEST_SHIFT;
+       LAPIC_REG(ICR)  =
+               LAPIC_VECTOR(INTERPROCESSOR) | LAPIC_ICR_DM_FIXED;
+       (void) ml_set_interrupts_enabled(state);
+}
+
+/*
+ * cpu_interrupt is really just to be used by the scheduler to
+ * get a CPU's attention it may not always issue an IPI.  If an
+ * IPI is always needed then use i386_cpu_IPI.
+ */
+void
+cpu_interrupt(int cpu)
+{
+       if (smp_initialized
+           && pmCPUExitIdle(cpu_datap(cpu))) {
+               i386_cpu_IPI(cpu);
+       }
 }
 
 /*
 }
 
 /*
@@ -973,16 +1010,16 @@ cpu_NMI_interrupt(int cpu)
 
        if (smp_initialized) {
                state = ml_set_interrupts_enabled(FALSE);
 
        if (smp_initialized) {
                state = ml_set_interrupts_enabled(FALSE);
+/* Program the interrupt command register */
                LAPIC_REG(ICRD) =
                        cpu_to_lapic[cpu] << LAPIC_ICRD_DEST_SHIFT;
                LAPIC_REG(ICRD) =
                        cpu_to_lapic[cpu] << LAPIC_ICRD_DEST_SHIFT;
-/* The vector is ignored in this case, the other CPU will come in on the
+/* The vector is ignored in this case--the target CPU will enter on the
  * NMI vector.
  */
                LAPIC_REG(ICR)  =
                        LAPIC_VECTOR(INTERPROCESSOR) | LAPIC_ICR_DM_NMI;
                (void) ml_set_interrupts_enabled(state);
        }
  * NMI vector.
  */
                LAPIC_REG(ICR)  =
                        LAPIC_VECTOR(INTERPROCESSOR) | LAPIC_ICR_DM_NMI;
                (void) ml_set_interrupts_enabled(state);
        }
-
 }
 
 void
 }
 
 void
@@ -1001,7 +1038,7 @@ i386_signal_cpu(int cpu, mp_event_t event, mp_sync_t mode)
        DBGLOG(cpu_signal, cpu, event);
        
        i_bit_set(event, signals);
        DBGLOG(cpu_signal, cpu, event);
        
        i_bit_set(event, signals);
-       cpu_interrupt(cpu);
+       i386_cpu_IPI(cpu);
        if (mode == SYNC) {
           again:
                tsc_timeout = rdtsc64() + (1000*1000*1000);
        if (mode == SYNC) {
           again:
                tsc_timeout = rdtsc64() + (1000*1000*1000);
@@ -1018,12 +1055,18 @@ i386_signal_cpu(int cpu, mp_event_t event, mp_sync_t mode)
                KERNEL_DEBUG(0xef800020 | DBG_FUNC_END, cpu, 0, 0, 0, 0);
 }
 
                KERNEL_DEBUG(0xef800020 | DBG_FUNC_END, cpu, 0, 0, 0, 0);
 }
 
+/*
+ * Send event to all running cpus.
+ * Called with the topology locked.
+ */
 void
 i386_signal_cpus(mp_event_t event, mp_sync_t mode)
 {
        unsigned int    cpu;
        unsigned int    my_cpu = cpu_number();
 
 void
 i386_signal_cpus(mp_event_t event, mp_sync_t mode)
 {
        unsigned int    cpu;
        unsigned int    my_cpu = cpu_number();
 
+       assert(hw_lock_held(&x86_topo_lock));
+
        for (cpu = 0; cpu < real_ncpus; cpu++) {
                if (cpu == my_cpu || !cpu_datap(cpu)->cpu_running)
                        continue;
        for (cpu = 0; cpu < real_ncpus; cpu++) {
                if (cpu == my_cpu || !cpu_datap(cpu)->cpu_running)
                        continue;
@@ -1031,12 +1074,18 @@ i386_signal_cpus(mp_event_t event, mp_sync_t mode)
        }
 }
 
        }
 }
 
+/*
+ * Return the number of running cpus.
+ * Called with the topology locked.
+ */
 int
 i386_active_cpus(void)
 {
        unsigned int    cpu;
        unsigned int    ncpus = 0;
 
 int
 i386_active_cpus(void)
 {
        unsigned int    cpu;
        unsigned int    ncpus = 0;
 
+       assert(hw_lock_held(&x86_topo_lock));
+
        for (cpu = 0; cpu < real_ncpus; cpu++) {
                if (cpu_datap(cpu)->cpu_running)
                        ncpus++;
        for (cpu = 0; cpu < real_ncpus; cpu++) {
                if (cpu_datap(cpu)->cpu_running)
                        ncpus++;
@@ -1061,29 +1110,39 @@ i386_active_cpus(void)
 static void
 mp_rendezvous_action(void)
 {
 static void
 mp_rendezvous_action(void)
 {
+       boolean_t intrs_enabled;
 
        /* setup function */
        if (mp_rv_setup_func != NULL)
                mp_rv_setup_func(mp_rv_func_arg);
 
        /* setup function */
        if (mp_rv_setup_func != NULL)
                mp_rv_setup_func(mp_rv_func_arg);
+
+       intrs_enabled = ml_get_interrupts_enabled();
+
        /* spin on entry rendezvous */
        /* spin on entry rendezvous */
-       atomic_incl(&mp_rv_waiters[0], 1);
-       while (mp_rv_waiters[0] < mp_rv_ncpus) {
-               boolean_t intr = ml_set_interrupts_enabled(FALSE);
-               /* poll for pesky tlb flushes */
-               handle_pending_TLB_flushes();
-               ml_set_interrupts_enabled(intr);
+       atomic_incl(&mp_rv_entry, 1);
+       while (mp_rv_entry < mp_rv_ncpus) {
+               /* poll for pesky tlb flushes if interrupts disabled */
+               if (!intrs_enabled)
+                       handle_pending_TLB_flushes();
                cpu_pause();
        }
        /* action function */
        if (mp_rv_action_func != NULL)
                mp_rv_action_func(mp_rv_func_arg);
        /* spin on exit rendezvous */
                cpu_pause();
        }
        /* action function */
        if (mp_rv_action_func != NULL)
                mp_rv_action_func(mp_rv_func_arg);
        /* spin on exit rendezvous */
-       atomic_incl(&mp_rv_waiters[1], 1);
-       while (mp_rv_waiters[1] < mp_rv_ncpus)
+       atomic_incl(&mp_rv_exit, 1);
+       while (mp_rv_exit < mp_rv_ncpus) {
+               if (!intrs_enabled)
+                       handle_pending_TLB_flushes();
                cpu_pause();
                cpu_pause();
+       }
+
        /* teardown function */
        if (mp_rv_teardown_func != NULL)
                mp_rv_teardown_func(mp_rv_func_arg);
        /* teardown function */
        if (mp_rv_teardown_func != NULL)
                mp_rv_teardown_func(mp_rv_func_arg);
+
+       /* Bump completion count */
+       atomic_incl(&mp_rv_complete, 1);
 }
 
 void
 }
 
 void
@@ -1112,19 +1171,37 @@ mp_rendezvous(void (*setup_func)(void *),
        mp_rv_teardown_func = teardown_func;
        mp_rv_func_arg = arg;
 
        mp_rv_teardown_func = teardown_func;
        mp_rv_func_arg = arg;
 
-       mp_rv_waiters[0] = 0;           /* entry rendezvous count */
-       mp_rv_waiters[1] = 0;           /* exit  rendezvous count */
-       mp_rv_ncpus = i386_active_cpus();
+       mp_rv_entry    = 0;
+       mp_rv_exit     = 0;
+       mp_rv_complete = 0;
 
        /*
         * signal other processors, which will call mp_rendezvous_action()
         * with interrupts disabled
         */
 
        /*
         * 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);
        i386_signal_cpus(MP_RENDEZVOUS, ASYNC);
+       simple_unlock(&x86_topo_lock);
 
        /* call executor function on this cpu */
        mp_rendezvous_action();
 
 
        /* call executor function on this cpu */
        mp_rendezvous_action();
 
+       /*
+        * Spin for everyone to complete.
+        * This is necessary to ensure that all processors have proceeded
+        * from the exit barrier before we release the rendezvous structure.
+        */
+       while (mp_rv_complete < mp_rv_ncpus) {
+               cpu_pause();
+       }
+       
+       /* Tidy up */
+       mp_rv_setup_func = NULL;
+       mp_rv_action_func = NULL;
+       mp_rv_teardown_func = NULL;
+       mp_rv_func_arg = NULL;
+
        /* release lock */
        simple_unlock(&mp_rv_lock);
 }
        /* release lock */
        simple_unlock(&mp_rv_lock);
 }
@@ -1173,17 +1250,236 @@ handle_pending_TLB_flushes(void)
 {
        volatile int    *my_word = &current_cpu_datap()->cpu_signals;
 
 {
        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);
+       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();
        }
 }
 
                i_bit_clear(MP_TLB_FLUSH, my_word);
                pmap_update_interrupt();
        }
 }
 
+/*
+ * This is called from cpu_signal_handler() to process an MP_CALL signal.
+ */
+static void
+mp_cpus_call_action(void)
+{
+       if (mp_rv_action_func != NULL)
+               mp_rv_action_func(mp_rv_func_arg);
+       atomic_incl(&mp_rv_complete, 1);
+}
+
+/*
+ * 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.
+ */
+cpu_t
+mp_cpus_call(
+       cpumask_t       cpus,
+       mp_sync_t       mode,
+        void           (*action_func)(void *),
+        void           *arg)
+{
+       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);
+
+       /* 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;
+
+       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)
+                       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);
+                       }
+               }
+       }
+       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);
+               }
+               while (mp_rv_complete < mp_rv_ncpus) {
+                       if (!intrs_enabled)
+                               handle_pending_TLB_flushes();
+                       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)
+{
+   /* 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((volatile long *)&mp_bc_count, 1))
+       thread_wakeup(((event_t)(unsigned int *) &mp_bc_count));
+}
+
+/*
+ * mp_broadcast() runs a given function on all active cpus.
+ * The caller blocks until the functions has run on all cpus.
+ * The caller will also block if there is another pending braodcast.
+ */
+void
+mp_broadcast(
+         void (*action_func)(void *),
+         void *arg)
+{
+   if (!smp_initialized) {
+       if (action_func != NULL)
+                  action_func(arg);
+       return;
+   }
+       
+   /* obtain broadcast lock */
+   mutex_lock(&mp_bc_lock);
+
+   /* set static function pointers */
+   mp_bc_action_func = action_func;
+   mp_bc_func_arg = arg;
+
+   assert_wait(&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 */
+   mutex_unlock(&mp_bc_lock);
+}
+
+void
+i386_activate_cpu(void)
+{
+       cpu_data_t      *cdp = current_cpu_datap();
+
+       assert(!ml_get_interrupts_enabled());
+
+       if (!smp_initialized) {
+               cdp->cpu_running = TRUE;
+               return;
+       }
+
+       simple_lock(&x86_topo_lock);
+       cdp->cpu_running = TRUE;
+       simple_unlock(&x86_topo_lock);
+}
+
+void
+i386_deactivate_cpu(void)
+{
+       cpu_data_t      *cdp = current_cpu_datap();
+
+       assert(!ml_get_interrupts_enabled());
+
+       simple_lock(&x86_topo_lock);
+       cdp->cpu_running = FALSE;
+       simple_unlock(&x86_topo_lock);
+
+       /*
+        * In case a rendezvous/braodcast/call was initiated to this cpu
+        * before we cleared cpu_running, we must perform any actions due.
+        */
+       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 */
+}
+
+int    pmsafe_debug    = 1;
 
 #if    MACH_KDP
 volatile boolean_t     mp_kdp_trap = FALSE;
 
 #if    MACH_KDP
 volatile boolean_t     mp_kdp_trap = FALSE;
-volatile long          mp_kdp_ncpus;
+volatile unsigned long         mp_kdp_ncpus;
 boolean_t              mp_kdp_state;
 
 
 boolean_t              mp_kdp_state;
 
 
@@ -1205,10 +1501,13 @@ mp_kdp_enter(void)
        mp_kdp_state = ml_set_interrupts_enabled(FALSE);
        simple_lock(&mp_kdp_lock);
 
        mp_kdp_state = ml_set_interrupts_enabled(FALSE);
        simple_lock(&mp_kdp_lock);
 
+       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");
        while (mp_kdp_trap) {
                simple_unlock(&mp_kdp_lock);
                DBG("mp_kdp_enter() race lost\n");
-               mp_kdp_wait();
+               mp_kdp_wait(TRUE);
                simple_lock(&mp_kdp_lock);
        }
        mp_kdp_ncpus = 1;       /* self */
                simple_lock(&mp_kdp_lock);
        }
        mp_kdp_ncpus = 1;       /* self */
@@ -1219,43 +1518,57 @@ mp_kdp_enter(void)
         * Deliver a nudge to other cpus, counting how many
         */
        DBG("mp_kdp_enter() signaling other processors\n");
         * Deliver a nudge to other cpus, counting how many
         */
        DBG("mp_kdp_enter() signaling other processors\n");
-       for (ncpus = 1, cpu = 0; cpu < real_ncpus; cpu++) {
-               if (cpu == my_cpu || !cpu_datap(cpu)->cpu_running)
-                       continue;
-               ncpus++;
-               i386_signal_cpu(cpu, MP_KDP, ASYNC);
-       }
-       /*
-        * Wait other processors to synchronize
-        */
-       DBG("mp_kdp_enter() waiting for (%d) processors to suspend\n", ncpus);
+       if (force_immediate_debugger_NMI == FALSE) {
+               for (ncpus = 1, cpu = 0; cpu < real_ncpus; cpu++) {
+                       if (cpu == my_cpu || !cpu_datap(cpu)->cpu_running)
+                               continue;
+                       ncpus++;
+                       i386_signal_cpu(cpu, MP_KDP, ASYNC);
+               }
+               /*
+                * Wait other processors to synchronize
+                */
+               DBG("mp_kdp_enter() waiting for (%d) processors to suspend\n", ncpus);
 
 
-       tsc_timeout = rdtsc64() + (ncpus * 100 * 1000 * 1000);
+               /*
+                * This timeout is rather arbitrary; we don't want to NMI
+                * processors that are executing at potentially
+                * "unsafe-to-interrupt" points such as the trampolines,
+                * but neither do we want to lose state by waiting too long.
+                */
+               tsc_timeout = rdtsc64() + (ncpus * 1000 * 1000);
 
 
-       while (mp_kdp_ncpus != ncpus && rdtsc64() < tsc_timeout) {
-               /*
-                * A TLB shootdown request may be pending... this would
-                * result in the requesting processor waiting in
-                * PMAP_UPDATE_TLBS() until this processor deals with it.
-                * Process it, so it can now enter mp_kdp_wait()
+               while (mp_kdp_ncpus != ncpus && rdtsc64() < tsc_timeout) {
+                       /*
+                        * A TLB shootdown request may be pending--this would
+                        * result in the requesting processor waiting in
+                        * PMAP_UPDATE_TLBS() until this processor deals with it.
+                        * Process it, so it can now enter mp_kdp_wait()
+                        */
+                       handle_pending_TLB_flushes();
+                       cpu_pause();
+               }
+               /* If we've timed out, and some processor(s) are still unresponsive,
+                * interrupt them with an NMI via the local APIC.
                 */
                 */
-               handle_pending_TLB_flushes();
-               cpu_pause();
+               if (mp_kdp_ncpus != ncpus) {
+                       for (cpu = 0; cpu < real_ncpus; cpu++) {
+                               if (cpu == my_cpu || !cpu_datap(cpu)->cpu_running)
+                                       continue;
+                               if (cpu_signal_pending(cpu, MP_KDP))
+                                       cpu_NMI_interrupt(cpu);
+                       }
+               }
        }
        }
-/* If we've timed out, and some processor(s) are still unresponsive,
- * interrupt them with an NMI via the local APIC.
- */
-       if (mp_kdp_ncpus != ncpus) {
+       else
                for (cpu = 0; cpu < real_ncpus; cpu++) {
                        if (cpu == my_cpu || !cpu_datap(cpu)->cpu_running)
                                continue;
                for (cpu = 0; cpu < real_ncpus; cpu++) {
                        if (cpu == my_cpu || !cpu_datap(cpu)->cpu_running)
                                continue;
-                       if (cpu_signal_pending(cpu, MP_KDP))
-                           cpu_NMI_interrupt(cpu);
+                       cpu_NMI_interrupt(cpu);
                }
                }
-       }
 
 
-       DBG("mp_kdp_enter() %d processors done %s\n",
-               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);
 }
        
        postcode(MP_KDP_ENTER);
 }
@@ -1270,26 +1583,38 @@ cpu_signal_pending(int cpu, mp_event_t event)
                retval = TRUE;
        return retval;
 }
                retval = TRUE;
        return retval;
 }
 
 static void
 
 static void
-mp_kdp_wait(void)
+mp_kdp_wait(boolean_t flush)
 {
        DBG("mp_kdp_wait()\n");
 {
        DBG("mp_kdp_wait()\n");
-
+       /* If an I/O port has been specified as a debugging aid, issue a read */
        panic_io_port_read();
 
        panic_io_port_read();
 
-       atomic_incl(&mp_kdp_ncpus, 1);
+       /* If we've trapped due to a machine-check, save MCA registers */
+       mca_check_save();
+
+       if (pmsafe_debug)
+           pmSafeMode(&current_cpu_datap()->lcpu, PM_SAFE_FL_SAFE);
+
+       atomic_incl((volatile long *)&mp_kdp_ncpus, 1);
        while (mp_kdp_trap) {
                /*
        while (mp_kdp_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.
+                * 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()
                 */
                 * Process it, so it can now enter mp_kdp_wait()
                 */
-               handle_pending_TLB_flushes();
-
+               if (flush)
+                       handle_pending_TLB_flushes();
                cpu_pause();
        }
                cpu_pause();
        }
-       atomic_decl(&mp_kdp_ncpus, 1);
+
+       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");
 }
 
        DBG("mp_kdp_wait() done\n");
 }
 
@@ -1297,7 +1622,7 @@ void
 mp_kdp_exit(void)
 {
        DBG("mp_kdp_exit()\n");
 mp_kdp_exit(void)
 {
        DBG("mp_kdp_exit()\n");
-       atomic_decl(&mp_kdp_ncpus, 1);
+       atomic_decl((volatile long *)&mp_kdp_ncpus, 1);
        mp_kdp_trap = FALSE;
        __asm__ volatile("mfence");
 
        mp_kdp_trap = FALSE;
        __asm__ volatile("mfence");
 
@@ -1313,6 +1638,10 @@ mp_kdp_exit(void)
 
                cpu_pause();
        }
 
                cpu_pause();
        }
+
+       if (pmsafe_debug)
+           pmSafeMode(&current_cpu_datap()->lcpu, PM_SAFE_FL_NORMAL);
+
        DBG("mp_kdp_exit() done\n");
        (void) ml_set_interrupts_enabled(mp_kdp_state);
        postcode(0);
        DBG("mp_kdp_exit() done\n");
        (void) ml_set_interrupts_enabled(mp_kdp_state);
        postcode(0);
@@ -1381,6 +1710,7 @@ mp_kdb_wait(void)
 {
        DBG("mp_kdb_wait()\n");
 
 {
        DBG("mp_kdb_wait()\n");
 
+       /* If an I/O port has been specified as a debugging aid, issue a read */
        panic_io_port_read();
 
        atomic_incl(&mp_kdb_ncpus, 1);
        panic_io_port_read();
 
        atomic_incl(&mp_kdb_ncpus, 1);
@@ -1394,7 +1724,7 @@ mp_kdb_wait(void)
 
                cpu_pause();
        }
 
                cpu_pause();
        }
-       atomic_decl(&mp_kdb_ncpus, 1);
+       atomic_decl((volatile long *)&mp_kdb_ncpus, 1);
        DBG("mp_kdb_wait() done\n");
 }
 
        DBG("mp_kdb_wait() done\n");
 }
 
@@ -1414,7 +1744,7 @@ void
 mp_kdb_exit(void)
 {
        DBG("mp_kdb_exit()\n");
 mp_kdb_exit(void)
 {
        DBG("mp_kdb_exit()\n");
-       atomic_decl(&mp_kdb_ncpus, 1);
+       atomic_decl((volatile long *)&mp_kdb_ncpus, 1);
        mp_kdb_trap = FALSE;
        __asm__ volatile("mfence");
 
        mp_kdb_trap = FALSE;
        __asm__ volatile("mfence");
 
@@ -1428,6 +1758,7 @@ mp_kdb_exit(void)
 
                cpu_pause();
        }
 
                cpu_pause();
        }
+
        DBG("mp_kdb_exit() done\n");
 }
 
        DBG("mp_kdb_exit() done\n");
 }
 
@@ -1451,16 +1782,12 @@ i386_init_slave(void)
                get_cpu_number(), get_cpu_phys_number());
 
        assert(!ml_get_interrupts_enabled());
                get_cpu_number(), get_cpu_phys_number());
 
        assert(!ml_get_interrupts_enabled());
-       if (cpu_mode_is64bit()) {
-               cpu_IA32e_enable(current_cpu_datap());
-               cpu_desc_load64(current_cpu_datap());
-               fast_syscall_init64();
-       } else {
-               fast_syscall_init();
-       }
 
 
-       lapic_init();
+       cpu_mode_init(current_cpu_datap());
+
+       mca_cpu_init();
 
 
+       lapic_init();
        LAPIC_DUMP();
        LAPIC_CPU_MAP_DUMP();
 
        LAPIC_DUMP();
        LAPIC_CPU_MAP_DUMP();
 
@@ -1468,9 +1795,12 @@ i386_init_slave(void)
 
        mtrr_update_cpu();
 
 
        mtrr_update_cpu();
 
+       /* resume VT operation */
+       vmx_resume();
+
        pat_init();
 
        pat_init();
 
-       cpu_thread_init();
+       cpu_thread_init();      /* not strictly necessary */
 
        cpu_init();     /* Sets cpu_running which starter cpu waits for */ 
 
 
        cpu_init();     /* Sets cpu_running which starter cpu waits for */