xnu-4570.1.46.tar.gz

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

diff --git a/osfmk/i386/trap.c b/osfmk/i386/trap.c
index a2a805a033dad5eb33d79d621aadfa4403ddf150..ad07c88874d3da1fca1c985719f2fd2dd1e34f7a 100644 (file)
--- a/osfmk/i386/trap.c
+++ b/osfmk/i386/trap.c
@@ -1,5 +1,5 @@
 /*
 /*

- * Copyright (c) 2000-2010 Apple Inc. All rights reserved.
+ * Copyright (c) 2000-2016 Apple Inc. All rights reserved.

  *
  * @APPLE_OSREFERENCE_LICENSE_HEADER_START@
  * 
  *
  * @APPLE_OSREFERENCE_LICENSE_HEADER_START@
  * 


@@ -89,8 +89,11 @@
 #include <kern/spl.h>
 #include <kern/misc_protos.h>
 #include <kern/debug.h>
 #include <kern/spl.h>
 #include <kern/misc_protos.h>
 #include <kern/debug.h>

-
+#if CONFIG_TELEMETRY
+#include <kern/telemetry.h>
+#endif

 #include <sys/kdebug.h>
 #include <sys/kdebug.h>

+#include <prng/random.h>

 
 #include <string.h>
 
 
 #include <string.h>
 


@@ -103,7 +106,7 @@
 #include <mach/i386/syscall_sw.h>
 
 #include <libkern/OSDebug.h>
 #include <mach/i386/syscall_sw.h>
 
 #include <libkern/OSDebug.h>

-
+#include <i386/cpu_threads.h>

 #include <machine/pal_routines.h>
 
 extern void throttle_lowpri_io(int);
 #include <machine/pal_routines.h>
 
 extern void throttle_lowpri_io(int);


@@ -113,14 +116,8 @@ extern void kprint_state(x86_saved_state64_t *saved_state);
  * Forward declarations
  */
 static void user_page_fault_continue(kern_return_t kret);
  * Forward declarations
  */
 static void user_page_fault_continue(kern_return_t kret);

-#ifdef __i386__
-static void panic_trap(x86_saved_state32_t *saved_state);
-static void set_recovery_ip(x86_saved_state32_t *saved_state, vm_offset_t ip);
-extern void panic_64(x86_saved_state_t *, int, const char *, boolean_t);
-#else
-static void panic_trap(x86_saved_state64_t *saved_state);
+static void panic_trap(x86_saved_state64_t *saved_state, uint32_t pl, kern_return_t fault_result);

 static void set_recovery_ip(x86_saved_state64_t *saved_state, vm_offset_t ip);
 static void set_recovery_ip(x86_saved_state64_t *saved_state, vm_offset_t ip);

-#endif

 
 volatile perfCallback perfTrapHook = NULL; /* Pointer to CHUD trap hook routine */
 
 
 volatile perfCallback perfTrapHook = NULL; /* Pointer to CHUD trap hook routine */
 


@@ -132,7 +129,9 @@ extern boolean_t dtrace_tally_fault(user_addr_t);
 #endif
 
 extern boolean_t pmap_smep_enabled;
 #endif
 
 extern boolean_t pmap_smep_enabled;

+extern boolean_t pmap_smap_enabled;

 
 

+__attribute__((noreturn))

 void
 thread_syscall_return(
         kern_return_t ret)
 void
 thread_syscall_return(
         kern_return_t ret)


@@ -193,7 +192,14 @@ thread_syscall_return(
                                ret);
 #endif
        }
                                ret);
 #endif
        }

-       throttle_lowpri_io(TRUE);
+
+#if DEBUG || DEVELOPMENT
+       kern_allocation_name_t
+       prior __assert_only = thread_get_kernel_state(thr_act)->allocation_name;
+       assertf(prior == NULL, "thread_set_allocation_name(\"%s\") not cleared", kern_allocation_get_name(prior));
+#endif /* DEBUG || DEVELOPMENT */
+
+       throttle_lowpri_io(1);

 
        thread_exception_return();
         /*NOTREACHED*/
 
        thread_exception_return();
         /*NOTREACHED*/


@@ -221,10 +227,6 @@ user_page_fault_continue(
                vaddr = uregs->cr2;
        }
 
                vaddr = uregs->cr2;
        }
 

-       if (__probable((kr == KERN_SUCCESS) || (kr == KERN_ABORTED))) {
-               thread_exception_return();
-               /*NOTREACHED*/
-       }

 
        /* PAL debug hook */
        pal_dbg_page_fault( thread, vaddr, kr );
 
        /* PAL debug hook */
        pal_dbg_page_fault( thread, vaddr, kr );


@@ -264,9 +266,6 @@ kprint_state(x86_saved_state64_t    *saved_state)
        kprintf("      r10    0x%llx\n", saved_state->r10);
        kprintf("      r8     0x%llx\n", saved_state->r8);
        kprintf("      r9     0x%llx\n", saved_state->r9);     
        kprintf("      r10    0x%llx\n", saved_state->r10);
        kprintf("      r8     0x%llx\n", saved_state->r8);
        kprintf("      r9     0x%llx\n", saved_state->r9);     

-       kprintf("      v_arg6 0x%llx\n", saved_state->v_arg6);
-       kprintf("      v_arg7 0x%llx\n", saved_state->v_arg7);
-       kprintf("      v_arg8 0x%llx\n", saved_state->v_arg8);

 
        kprintf("      cr2    0x%llx\n", saved_state->cr2);
        kprintf("real  cr2    0x%lx\n", get_cr2());
 
        kprintf("      cr2    0x%llx\n", saved_state->cr2);
        kprintf("real  cr2    0x%lx\n", get_cr2());


@@ -339,6 +338,9 @@ void interrupt_populate_latency_stats(char *buf, unsigned bufsize) {
                snprintf(buf, bufsize, "0x%x 0x%x 0x%llx", tcpu, cpu_data_ptr[tcpu]->cpu_max_observed_int_latency_vector, cpu_data_ptr[tcpu]->cpu_max_observed_int_latency);
 }
 
                snprintf(buf, bufsize, "0x%x 0x%x 0x%llx", tcpu, cpu_data_ptr[tcpu]->cpu_max_observed_int_latency_vector, cpu_data_ptr[tcpu]->cpu_max_observed_int_latency);
 }
 

+uint32_t interrupt_timer_coalescing_enabled = 1;
+uint64_t interrupt_coalesced_timers;
+

 /*
  * Handle interrupts:
  *  - local APIC interrupts (IPIs, timers, etc) are handled by the kernel,
 /*
  * Handle interrupts:
  *  - local APIC interrupts (IPIs, timers, etc) are handled by the kernel,


@@ -353,36 +355,25 @@ interrupt(x86_saved_state_t *state)
        boolean_t       user_mode = FALSE;
        int             ipl;
        int             cnum = cpu_number();
        boolean_t       user_mode = FALSE;
        int             ipl;
        int             cnum = cpu_number();

-       int             itype = 0;
-       
-       if (is_saved_state64(state) == TRUE) {
-               x86_saved_state64_t     *state64;
+       cpu_data_t      *cdp = cpu_data_ptr[cnum];
+       int             itype = DBG_INTR_TYPE_UNKNOWN;

 
 

-               state64 = saved_state64(state);
-               rip = state64->isf.rip;
-               rsp = state64->isf.rsp;
-               interrupt_num = state64->isf.trapno;
-#ifdef __x86_64__
-               if(state64->isf.cs & 0x03)
-#endif
-                       user_mode = TRUE;
-       } else {
-               x86_saved_state32_t     *state32;
-
-               state32 = saved_state32(state);
-               if (state32->cs & 0x03)
-                       user_mode = TRUE;
-               rip = state32->eip;
-               rsp = state32->uesp;
-               interrupt_num = state32->trapno;
-       }
+        x86_saved_state64_t    *state64 = saved_state64(state);
+       rip = state64->isf.rip;
+       rsp = state64->isf.rsp;
+       interrupt_num = state64->isf.trapno;
+       if(state64->isf.cs & 0x03)
+               user_mode = TRUE;
+
+       if (cpu_data_ptr[cnum]->lcpu.package->num_idle == topoParms.nLThreadsPerPackage)
+               cpu_data_ptr[cnum]->cpu_hwIntpexits[interrupt_num]++;

 
        if (interrupt_num == (LAPIC_DEFAULT_INTERRUPT_BASE + LAPIC_INTERPROCESSOR_INTERRUPT))
 
        if (interrupt_num == (LAPIC_DEFAULT_INTERRUPT_BASE + LAPIC_INTERPROCESSOR_INTERRUPT))

-               itype = 1;
+               itype = DBG_INTR_TYPE_IPI;

        else if (interrupt_num == (LAPIC_DEFAULT_INTERRUPT_BASE + LAPIC_TIMER_INTERRUPT))
        else if (interrupt_num == (LAPIC_DEFAULT_INTERRUPT_BASE + LAPIC_TIMER_INTERRUPT))

-               itype = 2;
+               itype = DBG_INTR_TYPE_TIMER;

        else
        else

-               itype = 3;
+               itype = DBG_INTR_TYPE_OTHER;

 
        KERNEL_DEBUG_CONSTANT_IST(KDEBUG_TRACE, 
                MACHDBG_CODE(DBG_MACH_EXCP_INTR, 0) | DBG_FUNC_START,
 
        KERNEL_DEBUG_CONSTANT_IST(KDEBUG_TRACE, 
                MACHDBG_CODE(DBG_MACH_EXCP_INTR, 0) | DBG_FUNC_START,


@@ -392,35 +383,61 @@ interrupt(x86_saved_state_t *state)
 
        SCHED_STATS_INTERRUPT(current_processor());
 
 
        SCHED_STATS_INTERRUPT(current_processor());
 

+#if CONFIG_TELEMETRY
+       if (telemetry_needs_record) {
+               telemetry_mark_curthread(user_mode);
+       }
+#endif
+

        ipl = get_preemption_level();
        
        /*
         * Handle local APIC interrupts
         * else call platform expert for devices.
         */
        ipl = get_preemption_level();
        
        /*
         * Handle local APIC interrupts
         * else call platform expert for devices.
         */

-       if (!lapic_interrupt(interrupt_num, state))
+       if (!lapic_interrupt(interrupt_num, state)) {

                PE_incoming_interrupt(interrupt_num);
                PE_incoming_interrupt(interrupt_num);

+       }

 
        if (__improbable(get_preemption_level() != ipl)) {
                panic("Preemption level altered by interrupt vector 0x%x: initial 0x%x, final: 0x%x\n", interrupt_num, ipl, get_preemption_level());
        }
 
 
 
        if (__improbable(get_preemption_level() != ipl)) {
                panic("Preemption level altered by interrupt vector 0x%x: initial 0x%x, final: 0x%x\n", interrupt_num, ipl, get_preemption_level());
        }
 
 

-       KERNEL_DEBUG_CONSTANT_IST(KDEBUG_TRACE, 
-               MACHDBG_CODE(DBG_MACH_EXCP_INTR, 0) | DBG_FUNC_END,
-               interrupt_num, 0, 0, 0, 0);
-
-       if (cpu_data_ptr[cnum]->cpu_nested_istack) {
-               cpu_data_ptr[cnum]->cpu_nested_istack_events++;
+       if (__improbable(cdp->cpu_nested_istack)) {
+               cdp->cpu_nested_istack_events++;

        }
        else  {
        }
        else  {

-               uint64_t int_latency = mach_absolute_time() - cpu_data_ptr[cnum]->cpu_int_event_time;
-               if (ilat_assert && (int_latency > interrupt_latency_cap) && !machine_timeout_suspended()) {
-                       panic("Interrupt vector 0x%x exceeded interrupt latency threshold, 0x%llx absolute time delta, prior signals: 0x%x, current signals: 0x%x", interrupt_num, int_latency, cpu_data_ptr[cnum]->cpu_prior_signals, cpu_data_ptr[cnum]->cpu_signals);
+               uint64_t ctime = mach_absolute_time();
+               uint64_t int_latency = ctime - cdp->cpu_int_event_time;
+               uint64_t esdeadline, ehdeadline;
+               /* Attempt to process deferred timers in the context of
+                * this interrupt, unless interrupt time has already exceeded
+                * TCOAL_ILAT_THRESHOLD.
+                */
+#define TCOAL_ILAT_THRESHOLD (30000ULL)
+
+               if ((int_latency < TCOAL_ILAT_THRESHOLD) &&
+                   interrupt_timer_coalescing_enabled) {
+                       esdeadline = cdp->rtclock_timer.queue.earliest_soft_deadline;
+                       ehdeadline = cdp->rtclock_timer.deadline;
+                       if ((ctime >= esdeadline) && (ctime < ehdeadline)) {
+                               interrupt_coalesced_timers++;
+                               TCOAL_DEBUG(0x88880000 | DBG_FUNC_START, ctime, esdeadline, ehdeadline, interrupt_coalesced_timers, 0);
+                               rtclock_intr(state);
+                               TCOAL_DEBUG(0x88880000 | DBG_FUNC_END, ctime, esdeadline, interrupt_coalesced_timers, 0, 0);
+                       } else {
+                               TCOAL_DEBUG(0x77770000, ctime, cdp->rtclock_timer.queue.earliest_soft_deadline, cdp->rtclock_timer.deadline, interrupt_coalesced_timers, 0);
+                       }

                }
                }

-               if (int_latency > cpu_data_ptr[cnum]->cpu_max_observed_int_latency) {
-                       cpu_data_ptr[cnum]->cpu_max_observed_int_latency = int_latency;
-                       cpu_data_ptr[cnum]->cpu_max_observed_int_latency_vector = interrupt_num;
+
+               if (__improbable(ilat_assert && (int_latency > interrupt_latency_cap) && !machine_timeout_suspended())) {
+                       panic("Interrupt vector 0x%x exceeded interrupt latency threshold, 0x%llx absolute time delta, prior signals: 0x%x, current signals: 0x%x", interrupt_num, int_latency, cdp->cpu_prior_signals, cdp->cpu_signals);
+               }
+
+               if (__improbable(int_latency > cdp->cpu_max_observed_int_latency)) {
+                       cdp->cpu_max_observed_int_latency = int_latency;
+                       cdp->cpu_max_observed_int_latency_vector = interrupt_num;

                }
        }
 
                }
        }
 


@@ -428,17 +445,26 @@ interrupt(x86_saved_state_t *state)
         * Having serviced the interrupt first, look at the interrupted stack depth.
         */
        if (!user_mode) {
         * Having serviced the interrupt first, look at the interrupted stack depth.
         */
        if (!user_mode) {

-               uint64_t depth = cpu_data_ptr[cnum]->cpu_kernel_stack
-                                + sizeof(struct x86_kernel_state)
+               uint64_t depth = cdp->cpu_kernel_stack
+                                + sizeof(struct thread_kernel_state)

                                 + sizeof(struct i386_exception_link *)
                                 - rsp;
                                 + sizeof(struct i386_exception_link *)
                                 - rsp;

-               if (depth > kernel_stack_depth_max) {
+               if (__improbable(depth > kernel_stack_depth_max)) {

                        kernel_stack_depth_max = (vm_offset_t)depth;
                        KERNEL_DEBUG_CONSTANT(
                                MACHDBG_CODE(DBG_MACH_SCHED, MACH_STACK_DEPTH),
                                (long) depth, (long) VM_KERNEL_UNSLIDE(rip), 0, 0, 0);
                }
        }
                        kernel_stack_depth_max = (vm_offset_t)depth;
                        KERNEL_DEBUG_CONSTANT(
                                MACHDBG_CODE(DBG_MACH_SCHED, MACH_STACK_DEPTH),
                                (long) depth, (long) VM_KERNEL_UNSLIDE(rip), 0, 0, 0);
                }
        }

+       
+       if (cnum == master_cpu)
+               ml_entropy_collect();
+
+       KERNEL_DEBUG_CONSTANT_IST(KDEBUG_TRACE, 
+               MACHDBG_CODE(DBG_MACH_EXCP_INTR, 0) | DBG_FUNC_END,
+               interrupt_num, 0, 0, 0, 0);
+
+       assert(ml_get_interrupts_enabled() == FALSE);

 }
 
 static inline void
 }
 
 static inline void


@@ -464,16 +490,13 @@ kernel_trap(
        x86_saved_state_t       *state,
        uintptr_t *lo_spp)
 {
        x86_saved_state_t       *state,
        uintptr_t *lo_spp)
 {

-#ifdef __i386__
-       x86_saved_state32_t     *saved_state;
-#else

        x86_saved_state64_t     *saved_state;
        x86_saved_state64_t     *saved_state;

-#endif

        int                     code;
        user_addr_t             vaddr;
        int                     type;
        vm_map_t                map = 0;        /* protected by T_PAGE_FAULT */
        kern_return_t           result = KERN_FAILURE;
        int                     code;
        user_addr_t             vaddr;
        int                     type;
        vm_map_t                map = 0;        /* protected by T_PAGE_FAULT */
        kern_return_t           result = KERN_FAILURE;

+       kern_return_t           fault_result = KERN_SUCCESS;

        thread_t                thread;
        ast_t                   *myast;
        boolean_t               intr;
        thread_t                thread;
        ast_t                   *myast;
        boolean_t               intr;


@@ -483,26 +506,11 @@ kernel_trap(
 #if NCOPY_WINDOWS > 0
        int                     fault_in_copy_window = -1;
 #endif
 #if NCOPY_WINDOWS > 0
        int                     fault_in_copy_window = -1;
 #endif

-       int                     is_user = 0;
-       
-       thread = current_thread();
-
-#ifdef __i386__
-       if (__improbable(is_saved_state64(state))) {
-               panic_64(state, 0, "Kernel trap with 64-bit state", FALSE);
-       }
-
-       saved_state = saved_state32(state);
+       int                     is_user;
+       int                     trap_pl = get_preemption_level();

 
 

-       /* Record cpu where state was captured (trampolines don't set this) */
-       saved_state->cpu = cpu_number();
+       thread = current_thread();

 
 

-       vaddr = (user_addr_t)saved_state->cr2;
-       type  = saved_state->trapno;
-       code  = saved_state->err & 0xffff;
-       intr  = (saved_state->efl & EFL_IF) != 0;       /* state of ints at trap */
-       kern_ip = (vm_offset_t)saved_state->eip;
-#else

        if (__improbable(is_saved_state32(state)))
                panic("kernel_trap(%p) with 32-bit state", state);
        saved_state = saved_state64(state);
        if (__improbable(is_saved_state32(state)))
                panic("kernel_trap(%p) with 32-bit state", state);
        saved_state = saved_state64(state);


@@ -515,10 +523,11 @@ kernel_trap(
        code  = (int)(saved_state->isf.err & 0xffff);
        intr  = (saved_state->isf.rflags & EFL_IF) != 0;        /* state of ints at trap */
        kern_ip = (vm_offset_t)saved_state->isf.rip;
        code  = (int)(saved_state->isf.err & 0xffff);
        intr  = (saved_state->isf.rflags & EFL_IF) != 0;        /* state of ints at trap */
        kern_ip = (vm_offset_t)saved_state->isf.rip;

-#endif

 
        myast = ast_pending();
 
 
        myast = ast_pending();
 

+       is_user = (vaddr < VM_MAX_USER_PAGE_ADDRESS);
+

        perfASTCallback astfn = perfASTHook;
        if (__improbable(astfn != NULL)) {
                if (*myast & AST_CHUD_ALL)
        perfASTCallback astfn = perfASTHook;
        if (__improbable(astfn != NULL)) {
                if (*myast & AST_CHUD_ALL)


@@ -526,20 +535,11 @@ kernel_trap(
        } else
                *myast &= ~AST_CHUD_ALL;
 
        } else
                *myast &= ~AST_CHUD_ALL;
 

-       /*
-        * Is there a hook?
-        */
-       perfCallback fn = perfTrapHook;
-       if (__improbable(fn != NULL)) {
-               if (fn(type, NULL, 0, 0) == KERN_SUCCESS) {
-                       /*
-                        * If it succeeds, we are done...
-                        */
-                       return;
-               }
-       }

 
 #if CONFIG_DTRACE
 
 #if CONFIG_DTRACE

+       /*
+        * Is there a DTrace hook?
+        */     

        if (__improbable(tempDTraceTrapHook != NULL)) {
                if (tempDTraceTrapHook(type, state, lo_spp, 0) == KERN_SUCCESS) {
                        /*
        if (__improbable(tempDTraceTrapHook != NULL)) {
                if (tempDTraceTrapHook(type, state, lo_spp, 0) == KERN_SUCCESS) {
                        /*


@@ -556,14 +556,21 @@ kernel_trap(
         * as soon we possibly can to hold latency down
         */
        if (__improbable(T_PREEMPT == type)) {
         * as soon we possibly can to hold latency down
         */
        if (__improbable(T_PREEMPT == type)) {

-               ast_taken(AST_PREEMPTION, FALSE);
+               ast_taken_kernel();

 
                KERNEL_DEBUG_CONSTANT_IST(KDEBUG_TRACE, 
                        (MACHDBG_CODE(DBG_MACH_EXCP_KTRAP_x86, type)) | DBG_FUNC_NONE,
                        0, 0, 0, VM_KERNEL_UNSLIDE(kern_ip), 0);
                return;
        }
 
                KERNEL_DEBUG_CONSTANT_IST(KDEBUG_TRACE, 
                        (MACHDBG_CODE(DBG_MACH_EXCP_KTRAP_x86, type)) | DBG_FUNC_NONE,
                        0, 0, 0, VM_KERNEL_UNSLIDE(kern_ip), 0);
                return;
        }

-       
+
+       user_addr_t     kd_vaddr = is_user ? vaddr : VM_KERNEL_UNSLIDE(vaddr);
+       KERNEL_DEBUG_CONSTANT_IST(KDEBUG_TRACE,
+               (MACHDBG_CODE(DBG_MACH_EXCP_KTRAP_x86, type)) | DBG_FUNC_NONE,
+               (unsigned)(kd_vaddr >> 32), (unsigned)kd_vaddr, is_user,
+               VM_KERNEL_UNSLIDE(kern_ip), 0);
+
+

        if (T_PAGE_FAULT == type) {
                /*
                 * assume we're faulting in the kernel map
        if (T_PAGE_FAULT == type) {
                /*
                 * assume we're faulting in the kernel map


@@ -598,13 +605,11 @@ kernel_trap(
                                        map = thread->map;
                                        fault_in_copy_window = window_index;
                                }
                                        map = thread->map;
                                        fault_in_copy_window = window_index;
                                }

-                               is_user = -1;

                        }
 #else
                        if (__probable(vaddr < VM_MAX_USER_PAGE_ADDRESS)) {
                                /* fault occurred in userspace */
                                map = thread->map;
                        }
 #else
                        if (__probable(vaddr < VM_MAX_USER_PAGE_ADDRESS)) {
                                /* fault occurred in userspace */
                                map = thread->map;

-                               is_user = -1;

 
                                /* Intercept a potential Supervisor Mode Execute
                                 * Protection fault. These criteria identify
 
                                /* Intercept a potential Supervisor Mode Execute
                                 * Protection fault. These criteria identify


@@ -613,7 +618,19 @@ kernel_trap(
                                 * (The VM could just redrive a SMEP fault, hence
                                 * the intercept).
                                 */
                                 * (The VM could just redrive a SMEP fault, hence
                                 * the intercept).
                                 */

-                               if (__improbable((code == (T_PF_PROT | T_PF_EXECUTE)) && (pmap_smep_enabled) && (saved_state->isf.rip == vaddr))) {
+                               if (__improbable((code == (T_PF_PROT | T_PF_EXECUTE)) &&
+                                       (pmap_smep_enabled) && (saved_state->isf.rip == vaddr))) {
+                                       goto debugger_entry;
+                               }
+
+                               /*
+                                * Additionally check for SMAP faults...
+                                * which are characterized by page-present and
+                                * the AC bit unset (i.e. not from copyin/out path).
+                                */
+                               if (__improbable(code & T_PF_PROT &&
+                                                pmap_smap_enabled &&
+                                                (saved_state->isf.rflags & EFL_AC) == 0)) {

                                        goto debugger_entry;
                                }
 
                                        goto debugger_entry;
                                }
 


@@ -629,16 +646,14 @@ kernel_trap(
                                        set_cr3_raw(map->pmap->pm_cr3);
                                        return;
                                }
                                        set_cr3_raw(map->pmap->pm_cr3);
                                        return;
                                }

+                               if (__improbable(vaddr < PAGE_SIZE) &&
+                                   ((thread->machine.specFlags & CopyIOActive) == 0)) {
+                                       goto debugger_entry;
+                               }

                        }
 #endif
                }
        }
                        }
 #endif
                }
        }

-       user_addr_t     kd_vaddr = is_user ? vaddr : VM_KERNEL_UNSLIDE(vaddr);  
-       KERNEL_DEBUG_CONSTANT_IST(KDEBUG_TRACE, 
-               (MACHDBG_CODE(DBG_MACH_EXCP_KTRAP_x86, type)) | DBG_FUNC_NONE,
-               (unsigned)(kd_vaddr >> 32), (unsigned)kd_vaddr, is_user,
-               VM_KERNEL_UNSLIDE(kern_ip), 0);
-

 
        (void) ml_set_interrupts_enabled(intr);
 
 
        (void) ml_set_interrupts_enabled(intr);
 


@@ -660,11 +675,7 @@ kernel_trap(
                fpSSEexterrflt();
                return;
            case T_DEBUG:
                fpSSEexterrflt();
                return;
            case T_DEBUG:

-#ifdef __i386__
-                   if ((saved_state->efl & EFL_TF) == 0 && NO_WATCHPOINTS)
-#else

                    if ((saved_state->isf.rflags & EFL_TF) == 0 && NO_WATCHPOINTS)
                    if ((saved_state->isf.rflags & EFL_TF) == 0 && NO_WATCHPOINTS)

-#endif

                    {
                            /* We've somehow encountered a debug
                             * register match that does not belong
                    {
                            /* We've somehow encountered a debug
                             * register match that does not belong


@@ -698,15 +709,13 @@ kernel_trap(
 
                if (code & T_PF_WRITE)
                        prot |= VM_PROT_WRITE;
 
                if (code & T_PF_WRITE)
                        prot |= VM_PROT_WRITE;

-#if     PAE

                if (code & T_PF_EXECUTE)
                        prot |= VM_PROT_EXECUTE;
                if (code & T_PF_EXECUTE)
                        prot |= VM_PROT_EXECUTE;

-#endif

 
 

-               result = vm_fault(map,
-                                 vm_map_trunc_page(vaddr),
+               fault_result = result = vm_fault(map,
+                                 vaddr,

                                  prot,
                                  prot,

-                                 FALSE, 
+                                 FALSE, VM_KERN_MEMORY_NONE,

                                  THREAD_UNINT, NULL, 0);
 
                if (result == KERN_SUCCESS) {
                                  THREAD_UNINT, NULL, 0);
 
                if (result == KERN_SUCCESS) {


@@ -770,84 +779,31 @@ debugger_entry:
                 */
                sync_iss_to_iks(state);
 #if  MACH_KDP
                 */
                sync_iss_to_iks(state);
 #if  MACH_KDP

-               if (current_debugger != KDB_CUR_DB) {
-                       if (kdp_i386_trap(type, saved_state, result, (vm_offset_t)vaddr))
-                               return;
-               }
+               if (kdp_i386_trap(type, saved_state, result, (vm_offset_t)vaddr))
+                       return;

 #endif
        }
        pal_cli();
 #endif
        }
        pal_cli();

-       panic_trap(saved_state);
+       panic_trap(saved_state, trap_pl, fault_result);

        /*
         * NO RETURN
         */
 }
 
 
        /*
         * NO RETURN
         */
 }
 
 

-#ifdef __i386__
-static void
-set_recovery_ip(x86_saved_state32_t  *saved_state, vm_offset_t ip)
-{
-        saved_state->eip = ip;
-}
-#else

 static void
 set_recovery_ip(x86_saved_state64_t  *saved_state, vm_offset_t ip)
 {
         saved_state->isf.rip = ip;
 }
 static void
 set_recovery_ip(x86_saved_state64_t  *saved_state, vm_offset_t ip)
 {
         saved_state->isf.rip = ip;
 }

-#endif
-

 
 

-#ifdef __i386__

 static void
 static void

-panic_trap(x86_saved_state32_t *regs)
-{
-       const char *trapname = "Unknown";
-       pal_cr_t        cr0, cr2, cr3, cr4;
-
-       pal_get_control_registers( &cr0, &cr2, &cr3, &cr4 );
-
-       /*
-        * Issue an I/O port read if one has been requested - this is an
-        * event logic analyzers can use as a trigger point.
-        */
-       panic_io_port_read();
-
-       kprintf("panic trap number 0x%x, eip 0x%x\n", regs->trapno, regs->eip);
-       kprintf("cr0 0x%08x cr2 0x%08x cr3 0x%08x cr4 0x%08x\n",
-               cr0, cr2, cr3, cr4);
-
-       if (regs->trapno < TRAP_TYPES)
-               trapname = trap_type[regs->trapno];
-#undef panic
-       panic("Kernel trap at 0x%08x, type %d=%s, registers:\n"
-             "CR0: 0x%08x, CR2: 0x%08x, CR3: 0x%08x, CR4: 0x%08x\n"
-             "EAX: 0x%08x, EBX: 0x%08x, ECX: 0x%08x, EDX: 0x%08x\n"
-             "CR2: 0x%08x, EBP: 0x%08x, ESI: 0x%08x, EDI: 0x%08x\n"
-             "EFL: 0x%08x, EIP: 0x%08x, CS:  0x%08x, DS:  0x%08x\n"
-             "Error code: 0x%08x%s\n",
-             regs->eip, regs->trapno, trapname, cr0, cr2, cr3, cr4,
-             regs->eax,regs->ebx,regs->ecx,regs->edx,
-             regs->cr2,regs->ebp,regs->esi,regs->edi,
-             regs->efl,regs->eip,regs->cs & 0xFFFF, regs->ds & 0xFFFF, regs->err,
-             virtualized ? " VMM" : "");
-       /*
-        * This next statement is not executed,
-        * but it's needed to stop the compiler using tail call optimization
-        * for the panic call - which confuses the subsequent backtrace.
-        */
-       cr0 = 0;
-}
-#else
-
-
-static void
-panic_trap(x86_saved_state64_t *regs)
+panic_trap(x86_saved_state64_t *regs, uint32_t pl, kern_return_t fault_result)

 {
        const char      *trapname = "Unknown";
        pal_cr_t        cr0, cr2, cr3, cr4;
        boolean_t       potential_smep_fault = FALSE, potential_kernel_NX_fault = FALSE;
 {
        const char      *trapname = "Unknown";
        pal_cr_t        cr0, cr2, cr3, cr4;
        boolean_t       potential_smep_fault = FALSE, potential_kernel_NX_fault = FALSE;

+       boolean_t       potential_smap_fault = FALSE;

 
        pal_get_control_registers( &cr0, &cr2, &cr3, &cr4 );
        assert(ml_get_interrupts_enabled() == FALSE);
 
        pal_get_control_registers( &cr0, &cr2, &cr3, &cr4 );
        assert(ml_get_interrupts_enabled() == FALSE);


@@ -872,6 +828,12 @@ panic_trap(x86_saved_state64_t *regs)
                } else if (regs->isf.rip >= VM_MIN_KERNEL_AND_KEXT_ADDRESS) {
                        potential_kernel_NX_fault = TRUE;
                }
                } else if (regs->isf.rip >= VM_MIN_KERNEL_AND_KEXT_ADDRESS) {
                        potential_kernel_NX_fault = TRUE;
                }

+       } else if (pmap_smap_enabled &&
+                  regs->isf.trapno == T_PAGE_FAULT &&
+                  regs->isf.err & T_PF_PROT &&
+                  regs->cr2 < VM_MAX_USER_PAGE_ADDRESS &&
+                  regs->isf.rip >= VM_MIN_KERNEL_AND_KEXT_ADDRESS) {
+               potential_smap_fault = TRUE;

        }
 
 #undef panic
        }
 
 #undef panic


@@ -882,7 +844,7 @@ panic_trap(x86_saved_state64_t *regs)
              "R8:  0x%016llx, R9:  0x%016llx, R10: 0x%016llx, R11: 0x%016llx\n"
              "R12: 0x%016llx, R13: 0x%016llx, R14: 0x%016llx, R15: 0x%016llx\n"
              "RFL: 0x%016llx, RIP: 0x%016llx, CS:  0x%016llx, SS:  0x%016llx\n"
              "R8:  0x%016llx, R9:  0x%016llx, R10: 0x%016llx, R11: 0x%016llx\n"
              "R12: 0x%016llx, R13: 0x%016llx, R14: 0x%016llx, R15: 0x%016llx\n"
              "RFL: 0x%016llx, RIP: 0x%016llx, CS:  0x%016llx, SS:  0x%016llx\n"

-             "Fault CR2: 0x%016llx, Error code: 0x%016llx, Fault CPU: 0x%x%s%s%s\n",
+             "Fault CR2: 0x%016llx, Error code: 0x%016llx, Fault CPU: 0x%x%s%s%s%s, PL: %d, VF: %d\n",

              regs->isf.rip, regs->isf.trapno, trapname,
              cr0, cr2, cr3, cr4,
              regs->rax, regs->rbx, regs->rcx, regs->rdx,
              regs->isf.rip, regs->isf.trapno, trapname,
              cr0, cr2, cr3, cr4,
              regs->rax, regs->rbx, regs->rcx, regs->rdx,


@@ -893,7 +855,10 @@ panic_trap(x86_saved_state64_t *regs)
              regs->isf.ss & 0xFFFF,regs->cr2, regs->isf.err, regs->isf.cpu,
              virtualized ? " VMM" : "",
              potential_kernel_NX_fault ? " Kernel NX fault" : "",
              regs->isf.ss & 0xFFFF,regs->cr2, regs->isf.err, regs->isf.cpu,
              virtualized ? " VMM" : "",
              potential_kernel_NX_fault ? " Kernel NX fault" : "",

-             potential_smep_fault ? " SMEP/User NX fault" : "");
+             potential_smep_fault ? " SMEP/User NX fault" : "",
+             potential_smap_fault ? " SMAP fault" : "",
+             pl,
+             fault_result);

        /*
         * This next statement is not executed,
         * but it's needed to stop the compiler using tail call optimization
        /*
         * This next statement is not executed,
         * but it's needed to stop the compiler using tail call optimization


@@ -901,7 +866,6 @@ panic_trap(x86_saved_state64_t *regs)
         */
        cr0 = 0;
 }
         */
        cr0 = 0;
 }

-#endif

 
 #if CONFIG_DTRACE
 extern kern_return_t dtrace_user_probe(x86_saved_state_t *);
 
 #if CONFIG_DTRACE
 extern kern_return_t dtrace_user_probe(x86_saved_state_t *);


@@ -996,11 +960,14 @@ user_trap(
                        return; /* If it succeeds, we are done... */
        }
 
                        return; /* If it succeeds, we are done... */
        }
 

+#if CONFIG_DTRACE

        /*
         * DTrace does not consume all user traps, only INT_3's for now.
         * Avoid needlessly calling tempDTraceTrapHook here, and let the
         * INT_3 case handle them.
         */
        /*
         * DTrace does not consume all user traps, only INT_3's for now.
         * Avoid needlessly calling tempDTraceTrapHook here, and let the
         * INT_3 case handle them.
         */

+#endif
+       

        DEBUG_KPRINT_SYSCALL_MASK(1,
                "user_trap: type=0x%x(%s) err=0x%x cr2=%p rip=%p\n",
                type, trap_type[type], err, (void *)(long) vaddr, (void *)(long) rip);
        DEBUG_KPRINT_SYSCALL_MASK(1,
                "user_trap: type=0x%x(%s) err=0x%x cr2=%p rip=%p\n",
                type, trap_type[type], err, (void *)(long) vaddr, (void *)(long) rip);


@@ -1059,6 +1026,9 @@ user_trap(
                break;
 
            case T_INVALID_OPCODE:
                break;
 
            case T_INVALID_OPCODE:

+#if !defined(RC_HIDE_XNU_J137)
+               fpUDflt(rip);   /* May return from exception directly */
+#endif

                exc = EXC_BAD_INSTRUCTION;
                code = EXC_I386_INVOP;
                break;
                exc = EXC_BAD_INSTRUCTION;
                code = EXC_I386_INVOP;
                break;


@@ -1110,21 +1080,25 @@ user_trap(
                break;
 
            case T_PAGE_FAULT:
                break;
 
            case T_PAGE_FAULT:

-               prot = VM_PROT_READ;
+           {
+                   prot = VM_PROT_READ;

 
                if (err & T_PF_WRITE)
                        prot |= VM_PROT_WRITE;
 
                if (err & T_PF_WRITE)
                        prot |= VM_PROT_WRITE;

-#if     PAE

                if (__improbable(err & T_PF_EXECUTE))
                        prot |= VM_PROT_EXECUTE;
                if (__improbable(err & T_PF_EXECUTE))
                        prot |= VM_PROT_EXECUTE;

-#endif
-               kret = vm_fault(thread->map, vm_map_trunc_page(vaddr),
-                                prot, FALSE,
-                                THREAD_ABORTSAFE, NULL, 0);
+               kret = vm_fault(thread->map,
+                               vaddr,
+                               prot, FALSE, VM_KERN_MEMORY_NONE,
+                               THREAD_ABORTSAFE, NULL, 0);
+
+               if (__probable((kret == KERN_SUCCESS) || (kret == KERN_ABORTED))) {
+                       thread_exception_return();
+                       /*NOTREACHED*/
+               }

 
                user_page_fault_continue(kret);
 
                user_page_fault_continue(kret);

-       
-               /* NOTREACHED */
+           }   /* NOTREACHED */

                break;
 
            case T_SSE_FLOAT_ERROR:
                break;
 
            case T_SSE_FLOAT_ERROR:


@@ -1160,29 +1134,6 @@ user_trap(
        /* NOTREACHED */
 }
 
        /* NOTREACHED */
 }
 

-
-/*
- * Handle AST traps for i386.
- */
-
-extern void     log_thread_action (thread_t, char *);
-
-void
-i386_astintr(int preemption)
-{
-       ast_t           mask = AST_ALL;
-       spl_t           s;
-
-       if (preemption)
-               mask = AST_PREEMPTION;
-
-       s = splsched();
-
-       ast_taken(mask, s);
-
-       splx(s);
-}
-

 /*
  * Handle exceptions for i386.
  *
 /*
  * Handle exceptions for i386.
  *


@@ -1210,46 +1161,33 @@ i386_exception(
 }
 
 
 }
 
 

-/* Synchronize a thread's i386_kernel_state (if any) with the given
- * i386_saved_state_t obtained from the trap/IPI handler; called in
+/* Synchronize a thread's x86_kernel_state (if any) with the given
+ * x86_saved_state_t obtained from the trap/IPI handler; called in

  * kernel_trap() prior to entering the debugger, and when receiving
  * kernel_trap() prior to entering the debugger, and when receiving

- * an "MP_KDP" IPI.
+ * an "MP_KDP" IPI. Called with null saved_state if an incoming IPI
+ * was detected from the kernel while spinning with interrupts masked.

  */
   
 void
 sync_iss_to_iks(x86_saved_state_t *saved_state)
 {
  */
   
 void
 sync_iss_to_iks(x86_saved_state_t *saved_state)
 {

-       struct x86_kernel_state *iks;
+       struct x86_kernel_state *iks = NULL;

        vm_offset_t kstack;
        boolean_t record_active_regs = FALSE;
 
        /* The PAL may have a special way to sync registers */
        vm_offset_t kstack;
        boolean_t record_active_regs = FALSE;
 
        /* The PAL may have a special way to sync registers */

-       if( saved_state->flavor == THREAD_STATE_NONE )
+       if (saved_state && saved_state->flavor == THREAD_STATE_NONE)

                pal_get_kern_regs( saved_state );
 
                pal_get_kern_regs( saved_state );
 

-       if ((kstack = current_thread()->kernel_stack) != 0) {
-#ifdef __i386__
-               x86_saved_state32_t     *regs = saved_state32(saved_state);
-#else
+       if (current_thread() != NULL && 
+           (kstack = current_thread()->kernel_stack) != 0) {

                x86_saved_state64_t     *regs = saved_state64(saved_state);
                x86_saved_state64_t     *regs = saved_state64(saved_state);

-#endif

 
                iks = STACK_IKS(kstack);
 
                /* Did we take the trap/interrupt in kernel mode? */
 
                iks = STACK_IKS(kstack);
 
                /* Did we take the trap/interrupt in kernel mode? */

-#ifdef __i386__
-               if (regs == USER_REGS32(current_thread()))
-                       record_active_regs = TRUE;
-               else {
-                       iks->k_ebx = regs->ebx;
-                       iks->k_esp = (int)regs;
-                       iks->k_ebp = regs->ebp;
-                       iks->k_edi = regs->edi;
-                       iks->k_esi = regs->esi;
-                       iks->k_eip = regs->eip;
-               }
-#else
-               if (regs == USER_REGS64(current_thread()))
+               if (saved_state == NULL || /* NULL => polling in kernel */
+                   regs == USER_REGS64(current_thread()))

                        record_active_regs = TRUE;
                else {
                        iks->k_rbx = regs->rbx;
                        record_active_regs = TRUE;
                else {
                        iks->k_rbx = regs->rbx;


@@ -1261,20 +1199,9 @@ sync_iss_to_iks(x86_saved_state_t *saved_state)
                        iks->k_r15 = regs->r15;
                        iks->k_rip = regs->isf.rip;
                }
                        iks->k_r15 = regs->r15;
                        iks->k_rip = regs->isf.rip;
                }

-#endif

        }
 
        if (record_active_regs == TRUE) {
        }
 
        if (record_active_regs == TRUE) {

-#ifdef __i386__
-               /* Show the trap handler path */
-               __asm__ volatile("movl %%ebx, %0" : "=m" (iks->k_ebx));
-               __asm__ volatile("movl %%esp, %0" : "=m" (iks->k_esp));
-               __asm__ volatile("movl %%ebp, %0" : "=m" (iks->k_ebp));
-               __asm__ volatile("movl %%edi, %0" : "=m" (iks->k_edi));
-               __asm__ volatile("movl %%esi, %0" : "=m" (iks->k_esi));
-               /* "Current" instruction pointer */
-               __asm__ volatile("movl $1f, %0\n1:" : "=m" (iks->k_eip));
-#else

                /* Show the trap handler path */
                __asm__ volatile("movq %%rbx, %0" : "=m" (iks->k_rbx));
                __asm__ volatile("movq %%rsp, %0" : "=m" (iks->k_rsp));
                /* Show the trap handler path */
                __asm__ volatile("movq %%rbx, %0" : "=m" (iks->k_rbx));
                __asm__ volatile("movq %%rsp, %0" : "=m" (iks->k_rsp));


@@ -1288,7 +1215,6 @@ sync_iss_to_iks(x86_saved_state_t *saved_state)
                                 : "=m" (iks->k_rip)
                                 :
                                 : "rax");
                                 : "=m" (iks->k_rip)
                                 :
                                 : "rax");

-#endif

        }
 }
 
        }
 }
 


@@ -1305,16 +1231,6 @@ sync_iss_to_iks_unconditionally(__unused x86_saved_state_t *saved_state) {
 
        if ((kstack = current_thread()->kernel_stack) != 0) {
                iks = STACK_IKS(kstack);
 
        if ((kstack = current_thread()->kernel_stack) != 0) {
                iks = STACK_IKS(kstack);

-#ifdef __i386__
-               /* Display the trap handler path */
-               __asm__ volatile("movl %%ebx, %0" : "=m" (iks->k_ebx));
-               __asm__ volatile("movl %%esp, %0" : "=m" (iks->k_esp));
-               __asm__ volatile("movl %%ebp, %0" : "=m" (iks->k_ebp));
-               __asm__ volatile("movl %%edi, %0" : "=m" (iks->k_edi));
-               __asm__ volatile("movl %%esi, %0" : "=m" (iks->k_esi));
-               /* "Current" instruction pointer */
-               __asm__ volatile("movl $1f, %0\n1:" : "=m" (iks->k_eip));
-#else

                /* Display the trap handler path */
                __asm__ volatile("movq %%rbx, %0" : "=m" (iks->k_rbx));
                __asm__ volatile("movq %%rsp, %0" : "=m" (iks->k_rsp));
                /* Display the trap handler path */
                __asm__ volatile("movq %%rbx, %0" : "=m" (iks->k_rbx));
                __asm__ volatile("movq %%rsp, %0" : "=m" (iks->k_rsp));


@@ -1325,6 +1241,29 @@ sync_iss_to_iks_unconditionally(__unused x86_saved_state_t *saved_state) {
                __asm__ volatile("movq %%r15, %0" : "=m" (iks->k_r15));
                /* "Current" instruction pointer */
                __asm__ volatile("leaq 1f(%%rip), %%rax; mov %%rax, %0\n1:" : "=m" (iks->k_rip)::"rax");
                __asm__ volatile("movq %%r15, %0" : "=m" (iks->k_r15));
                /* "Current" instruction pointer */
                __asm__ volatile("leaq 1f(%%rip), %%rax; mov %%rax, %0\n1:" : "=m" (iks->k_rip)::"rax");

-#endif

        }
 }
        }
 }

+
+#if DEBUG
+extern void    thread_exception_return_internal(void) __dead2;
+
+void thread_exception_return(void) {
+       thread_t thread = current_thread();
+       ml_set_interrupts_enabled(FALSE);
+       if (thread_is_64bit(thread) != task_has_64BitAddr(thread->task)) {
+               panic("Task/thread bitness mismatch %p %p, task: %d, thread: %d", thread, thread->task, thread_is_64bit(thread),  task_has_64BitAddr(thread->task));
+       }
+
+       if (thread_is_64bit(thread)) {
+               if ((gdt_desc_p(USER64_CS)->access & ACC_PL_U) == 0) {
+                       panic("64-GDT mismatch %p, descriptor: %p", thread, gdt_desc_p(USER64_CS));
+               }
+       } else {
+                       if ((gdt_desc_p(USER_CS)->access & ACC_PL_U) == 0) {
+                               panic("32-GDT mismatch %p, descriptor: %p", thread, gdt_desc_p(USER_CS));
+
+               }
+       }
+       thread_exception_return_internal();
+}
+#endif