xref: /titanic_41/usr/src/uts/i86pc/os/trap.c (revision 5aefb6555731130ca4fd295960123d71f2d21fe8)
1 /*
2  * CDDL HEADER START
3  *
4  * The contents of this file are subject to the terms of the
5  * Common Development and Distribution License, Version 1.0 only
6  * (the "License").  You may not use this file except in compliance
7  * with the License.
8  *
9  * You can obtain a copy of the license at usr/src/OPENSOLARIS.LICENSE
10  * or http://www.opensolaris.org/os/licensing.
11  * See the License for the specific language governing permissions
12  * and limitations under the License.
13  *
14  * When distributing Covered Code, include this CDDL HEADER in each
15  * file and include the License file at usr/src/OPENSOLARIS.LICENSE.
16  * If applicable, add the following below this CDDL HEADER, with the
17  * fields enclosed by brackets "[]" replaced with your own identifying
18  * information: Portions Copyright [yyyy] [name of copyright owner]
19  *
20  * CDDL HEADER END
21  */
22 
23 /*
24  * Copyright 2006 Sun Microsystems, Inc.  All rights reserved.
25  * Use is subject to license terms.
26  */
27 
28 /*	Copyright (c) 1990, 1991 UNIX System Laboratories, Inc. */
29 /*	Copyright (c) 1984, 1986, 1987, 1988, 1989, 1990 AT&T   */
30 /*		All Rights Reserved   				*/
31 /*								*/
32 /*	Copyright (c) 1987, 1988 Microsoft Corporation  	*/
33 /*		All Rights Reserved   				*/
34 /*								*/
35 
36 #pragma ident	"%Z%%M%	%I%	%E% SMI"
37 
38 #include <sys/types.h>
39 #include <sys/sysmacros.h>
40 #include <sys/param.h>
41 #include <sys/signal.h>
42 #include <sys/systm.h>
43 #include <sys/user.h>
44 #include <sys/proc.h>
45 #include <sys/disp.h>
46 #include <sys/class.h>
47 #include <sys/core.h>
48 #include <sys/syscall.h>
49 #include <sys/cpuvar.h>
50 #include <sys/vm.h>
51 #include <sys/sysinfo.h>
52 #include <sys/fault.h>
53 #include <sys/stack.h>
54 #include <sys/mmu.h>
55 #include <sys/psw.h>
56 #include <sys/regset.h>
57 #include <sys/fp.h>
58 #include <sys/trap.h>
59 #include <sys/kmem.h>
60 #include <sys/vtrace.h>
61 #include <sys/cmn_err.h>
62 #include <sys/prsystm.h>
63 #include <sys/mutex_impl.h>
64 #include <sys/machsystm.h>
65 #include <sys/archsystm.h>
66 #include <sys/sdt.h>
67 #include <sys/avintr.h>
68 #include <sys/kobj.h>
69 
70 #include <vm/hat.h>
71 
72 #include <vm/seg_kmem.h>
73 #include <vm/as.h>
74 #include <vm/seg.h>
75 #include <vm/hat_pte.h>
76 
77 #include <sys/procfs.h>
78 
79 #include <sys/reboot.h>
80 #include <sys/debug.h>
81 #include <sys/debugreg.h>
82 #include <sys/modctl.h>
83 #include <sys/aio_impl.h>
84 #include <sys/tnf.h>
85 #include <sys/tnf_probe.h>
86 #include <sys/cred.h>
87 #include <sys/mman.h>
88 #include <sys/x86_archext.h>
89 #include <sys/copyops.h>
90 #include <c2/audit.h>
91 #include <sys/ftrace.h>
92 #include <sys/panic.h>
93 #include <sys/traptrace.h>
94 #include <sys/ontrap.h>
95 #include <sys/cpc_impl.h>
96 
97 #define	USER	0x10000		/* user-mode flag added to trap type */
98 
99 static const char *trap_type_mnemonic[] = {
100 	"de",	"db",	"2",	"bp",
101 	"of",	"br",	"ud",	"nm",
102 	"df",	"9",	"ts",	"np",
103 	"ss",	"gp",	"pf",	"15",
104 	"mf",	"ac",	"mc",	"xf"
105 };
106 
107 static const char *trap_type[] = {
108 	"Divide error",				/* trap id 0 	*/
109 	"Debug",				/* trap id 1	*/
110 	"NMI interrupt",			/* trap id 2	*/
111 	"Breakpoint",				/* trap id 3 	*/
112 	"Overflow",				/* trap id 4 	*/
113 	"BOUND range exceeded",			/* trap id 5 	*/
114 	"Invalid opcode",			/* trap id 6 	*/
115 	"Device not available",			/* trap id 7 	*/
116 	"Double fault",				/* trap id 8 	*/
117 	"Coprocessor segment overrun",		/* trap id 9 	*/
118 	"Invalid TSS",				/* trap id 10 	*/
119 	"Segment not present",			/* trap id 11 	*/
120 	"Stack segment fault",			/* trap id 12 	*/
121 	"General protection",			/* trap id 13 	*/
122 	"Page fault",				/* trap id 14 	*/
123 	"Reserved",				/* trap id 15 	*/
124 	"x87 floating point error",		/* trap id 16 	*/
125 	"Alignment check",			/* trap id 17 	*/
126 	"Machine check",			/* trap id 18	*/
127 	"SIMD floating point exception",	/* trap id 19	*/
128 };
129 
130 #define	TRAP_TYPES	(sizeof (trap_type) / sizeof (trap_type[0]))
131 
132 int tudebug = 0;
133 int tudebugbpt = 0;
134 int tudebugfpe = 0;
135 int tudebugsse = 0;
136 
137 #if defined(TRAPDEBUG) || defined(lint)
138 int tdebug = 0;
139 int lodebug = 0;
140 int faultdebug = 0;
141 #else
142 #define	tdebug	0
143 #define	lodebug	0
144 #define	faultdebug	0
145 #endif /* defined(TRAPDEBUG) || defined(lint) */
146 
147 #if defined(TRAPTRACE)
148 static void dump_ttrace(void);
149 #endif	/* TRAPTRACE */
150 static void dumpregs(struct regs *);
151 static void showregs(uint_t, struct regs *, caddr_t);
152 static void dump_tss(void);
153 static int kern_gpfault(struct regs *);
154 
155 struct trap_info {
156 	struct regs *trap_regs;
157 	uint_t trap_type;
158 	caddr_t trap_addr;
159 };
160 
161 /*ARGSUSED*/
162 static int
163 die(uint_t type, struct regs *rp, caddr_t addr, processorid_t cpuid)
164 {
165 	struct trap_info ti;
166 	const char *trap_name, *trap_mnemonic;
167 
168 	if (type < TRAP_TYPES) {
169 		trap_name = trap_type[type];
170 		trap_mnemonic = trap_type_mnemonic[type];
171 	} else {
172 		trap_name = "trap";
173 		trap_mnemonic = "-";
174 	}
175 
176 #ifdef TRAPTRACE
177 	TRAPTRACE_FREEZE;
178 #endif
179 
180 	ti.trap_regs = rp;
181 	ti.trap_type = type & ~USER;
182 	ti.trap_addr = addr;
183 
184 	curthread->t_panic_trap = &ti;
185 
186 	if (type == T_PGFLT && addr < (caddr_t)KERNELBASE) {
187 		panic("BAD TRAP: type=%x (#%s %s) rp=%p addr=%p "
188 		    "occurred in module \"%s\" due to %s",
189 		    type, trap_mnemonic, trap_name, (void *)rp, (void *)addr,
190 		    mod_containing_pc((caddr_t)rp->r_pc),
191 		    addr < (caddr_t)PAGESIZE ?
192 		    "a NULL pointer dereference" :
193 		    "an illegal access to a user address");
194 	} else
195 		panic("BAD TRAP: type=%x (#%s %s) rp=%p addr=%p",
196 		    type, trap_mnemonic, trap_name, (void *)rp, (void *)addr);
197 	return (0);
198 }
199 
200 /*
201  * Rewrite the instruction at pc to be an int $T_SYSCALLINT instruction.
202  *
203  * int <vector> is two bytes: 0xCD <vector>
204  */
205 
206 #define	SLOW_SCALL_SIZE	2
207 
208 static int
209 rewrite_syscall(caddr_t pc)
210 {
211 	uchar_t instr[SLOW_SCALL_SIZE] = { 0xCD, T_SYSCALLINT };
212 
213 	if (uwrite(curthread->t_procp, instr, SLOW_SCALL_SIZE,
214 	    (uintptr_t)pc) != 0)
215 		return (1);
216 
217 	return (0);
218 }
219 
220 /*
221  * Test to see if the instruction at pc is sysenter or syscall. The second
222  * argument should be the x86 feature flag corresponding to the expected
223  * instruction.
224  *
225  * sysenter is two bytes: 0x0F 0x34
226  * syscall is two bytes:  0x0F 0x05
227  */
228 
229 #define	FAST_SCALL_SIZE	2
230 
231 static int
232 instr_is_fast_syscall(caddr_t pc, int which)
233 {
234 	uchar_t instr[FAST_SCALL_SIZE];
235 
236 	ASSERT(which == X86_SEP || which == X86_ASYSC);
237 
238 	if (copyin_nowatch(pc, (caddr_t)instr, FAST_SCALL_SIZE) != 0 ||
239 	    instr[0] != 0x0F)
240 		return (0);
241 
242 	if ((which == X86_SEP && instr[1] == 0x34) ||
243 	    (which == X86_ASYSC && instr[1] == 0x05))
244 		return (1);
245 
246 	return (0);
247 }
248 
249 /*
250  * Test to see if the instruction at pc is a system call instruction.
251  *
252  * The bytes of an lcall instruction used for the syscall trap.
253  * static uchar_t lcall[7] = { 0x9a, 0, 0, 0, 0, 0x7, 0 };
254  * static uchar_t lcallalt[7] = { 0x9a, 0, 0, 0, 0, 0x27, 0 };
255  */
256 
257 #define	LCALLSIZE	7
258 
259 static int
260 instr_is_syscall(caddr_t pc)
261 {
262 	uchar_t instr[LCALLSIZE];
263 
264 	if (copyin_nowatch(pc, (caddr_t)instr, LCALLSIZE) == 0 &&
265 	    instr[0] == 0x9a &&
266 	    instr[1] == 0 &&
267 	    instr[2] == 0 &&
268 	    instr[3] == 0 &&
269 	    instr[4] == 0 &&
270 	    (instr[5] == 0x7 || instr[5] == 0x27) &&
271 	    instr[6] == 0)
272 		return (1);
273 
274 	return (0);
275 }
276 
277 #ifdef __amd64
278 
279 /*
280  * In the first revisions of AMD64 CPUs produced by AMD, the LAHF and
281  * SAHF instructions were not implemented in 64bit mode. Later revisions
282  * did implement these instructions. An extension to the cpuid instruction
283  * was added to check for the capability of executing these instructions
284  * in 64bit mode.
285  *
286  * Intel originally did not implement these instructions in EM64T either,
287  * but added them in later revisions.
288  *
289  * So, there are different chip revisions by both vendors out there that
290  * may or may not implement these instructions. The easy solution is to
291  * just always emulate these instructions on demand.
292  *
293  * SAHF == store %ah in the lower 8 bits of %rflags (opcode 0x9e)
294  * LAHF == load the lower 8 bits of %rflags into %ah (opcode 0x9f)
295  */
296 
297 #define	LSAHFSIZE 1
298 
299 static int
300 instr_is_lsahf(caddr_t pc, uchar_t *instr)
301 {
302 	if (copyin_nowatch(pc, (caddr_t)instr, LSAHFSIZE) == 0 &&
303 	    (*instr == 0x9e || *instr == 0x9f))
304 		return (1);
305 	return (0);
306 }
307 
308 /*
309  * Emulate the LAHF and SAHF instructions. The reference manuals define
310  * these instructions to always load/store bit 1 as a 1, and bits 3 and 5
311  * as a 0. The other, defined, bits are copied (the PS_ICC bits and PS_P).
312  *
313  * Note that %ah is bits 8-15 of %rax.
314  */
315 static void
316 emulate_lsahf(struct regs *rp, uchar_t instr)
317 {
318 	if (instr == 0x9e) {
319 		/* sahf. Copy bits from %ah to flags. */
320 		rp->r_ps = (rp->r_ps & ~0xff) |
321 		    ((rp->r_rax >> 8) & PSL_LSAHFMASK) | PS_MB1;
322 	} else {
323 		/* lahf. Copy bits from flags to %ah. */
324 		rp->r_rax = (rp->r_rax & ~0xff00) |
325 		    (((rp->r_ps & PSL_LSAHFMASK) | PS_MB1) << 8);
326 	}
327 	rp->r_pc += LSAHFSIZE;
328 }
329 #endif /* __amd64 */
330 
331 #ifdef OPTERON_ERRATUM_91
332 
333 /*
334  * Test to see if the instruction at pc is a prefetch instruction.
335  *
336  * The first byte of prefetch instructions is always 0x0F.
337  * The second byte is 0x18 for regular prefetch or 0x0D for AMD 3dnow prefetch.
338  * The third byte is between 0 and 3 inclusive.
339  */
340 
341 #define	PREFETCHSIZE 3
342 
343 static int
344 cmp_to_prefetch(uchar_t *p)
345 {
346 	if (*p == 0x0F && (*(p+1) == 0x18 || *(p+1) == 0x0D) && *(p+2) <= 3)
347 		return (1);
348 	return (0);
349 }
350 
351 static int
352 instr_is_prefetch(caddr_t pc)
353 {
354 	uchar_t instr[PREFETCHSIZE];
355 	int	error;
356 
357 	error = copyin_nowatch(pc, (caddr_t)instr, PREFETCHSIZE);
358 
359 	if (error == 0 && cmp_to_prefetch(instr))
360 		return (1);
361 	return (0);
362 }
363 
364 #endif /* OPTERON_ERRATUM_91 */
365 
366 /*
367  * Called from the trap handler when a processor trap occurs.
368  *
369  * Note: All user-level traps that might call stop() must exit
370  * trap() by 'goto out' or by falling through.
371  */
372 void
373 trap(struct regs *rp, caddr_t addr, processorid_t cpuid)
374 {
375 	kthread_t *cur_thread = curthread;
376 	enum seg_rw rw;
377 	unsigned type;
378 	extern int stop_on_fault(uint_t, k_siginfo_t *);
379 	proc_t *p = ttoproc(cur_thread);
380 	klwp_t *lwp = ttolwp(cur_thread);
381 	uintptr_t lofault;
382 	faultcode_t pagefault(), res, errcode;
383 	enum fault_type fault_type;
384 	k_siginfo_t siginfo;
385 	uint_t fault = 0;
386 	int mstate;
387 	int sicode = 0;
388 	int watchcode;
389 	int watchpage;
390 	caddr_t vaddr;
391 	size_t sz;
392 	int ta;
393 #ifdef __amd64
394 	uchar_t instr;
395 #endif
396 
397 	ASSERT_STACK_ALIGNED();
398 
399 	type = rp->r_trapno;
400 	CPU_STATS_ADDQ(CPU, sys, trap, 1);
401 
402 	ASSERT(cur_thread->t_schedflag & TS_DONT_SWAP);
403 
404 	if (type == T_PGFLT) {
405 
406 		errcode = rp->r_err;
407 		if (errcode & PF_ERR_WRITE)
408 			rw = S_WRITE;
409 		else if ((caddr_t)rp->r_pc == addr ||
410 		    (mmu.pt_nx != 0 && (errcode & PF_ERR_EXEC)))
411 			rw = S_EXEC;
412 		else
413 			rw = S_READ;
414 
415 #if defined(__i386)
416 		/*
417 		 * Pentium Pro work-around
418 		 */
419 		if ((errcode & PF_ERR_PROT) && pentiumpro_bug4046376) {
420 			uint_t	attr;
421 			uint_t	priv_violation;
422 			uint_t	access_violation;
423 
424 			if (hat_getattr(addr < (caddr_t)kernelbase ?
425 			    curproc->p_as->a_hat : kas.a_hat, addr, &attr)
426 			    == -1) {
427 				errcode &= ~PF_ERR_PROT;
428 			} else {
429 				priv_violation = (errcode & PF_ERR_USER) &&
430 					!(attr & PROT_USER);
431 				access_violation = (errcode & PF_ERR_WRITE) &&
432 					!(attr & PROT_WRITE);
433 				if (!priv_violation && !access_violation)
434 					goto cleanup;
435 			}
436 		}
437 #endif /* __i386 */
438 
439 	}
440 
441 	if (tdebug)
442 		showregs(type, rp, addr);
443 
444 	if (USERMODE(rp->r_cs)) {
445 		/*
446 		 * Set up the current cred to use during this trap. u_cred
447 		 * no longer exists.  t_cred is used instead.
448 		 * The current process credential applies to the thread for
449 		 * the entire trap.  If trapping from the kernel, this
450 		 * should already be set up.
451 		 */
452 		if (cur_thread->t_cred != p->p_cred) {
453 			cred_t *oldcred = cur_thread->t_cred;
454 			/*
455 			 * DTrace accesses t_cred in probe context.  t_cred
456 			 * must always be either NULL, or point to a valid,
457 			 * allocated cred structure.
458 			 */
459 			cur_thread->t_cred = crgetcred();
460 			crfree(oldcred);
461 		}
462 		ASSERT(lwp != NULL);
463 		type |= USER;
464 		ASSERT(lwptoregs(lwp) == rp);
465 		lwp->lwp_state = LWP_SYS;
466 
467 		switch (type) {
468 		case T_PGFLT + USER:
469 			if ((caddr_t)rp->r_pc == addr)
470 				mstate = LMS_TFAULT;
471 			else
472 				mstate = LMS_DFAULT;
473 			break;
474 		default:
475 			mstate = LMS_TRAP;
476 			break;
477 		}
478 		/* Kernel probe */
479 		TNF_PROBE_1(thread_state, "thread", /* CSTYLED */,
480 		    tnf_microstate, state, mstate);
481 		mstate = new_mstate(cur_thread, mstate);
482 
483 		bzero(&siginfo, sizeof (siginfo));
484 	}
485 
486 	switch (type) {
487 	case T_PGFLT + USER:
488 	case T_SGLSTP:
489 	case T_SGLSTP + USER:
490 	case T_BPTFLT + USER:
491 		break;
492 
493 	default:
494 		FTRACE_2("trap(): type=0x%lx, regs=0x%lx",
495 		    (ulong_t)type, (ulong_t)rp);
496 		break;
497 	}
498 
499 	switch (type) {
500 	default:
501 		if (type & USER) {
502 			if (tudebug)
503 				showregs(type, rp, (caddr_t)0);
504 			printf("trap: Unknown trap type %d in user mode\n",
505 			    type & ~USER);
506 			siginfo.si_signo = SIGILL;
507 			siginfo.si_code  = ILL_ILLTRP;
508 			siginfo.si_addr  = (caddr_t)rp->r_pc;
509 			siginfo.si_trapno = type & ~USER;
510 			fault = FLTILL;
511 			break;
512 		} else {
513 			(void) die(type, rp, addr, cpuid);
514 			/*NOTREACHED*/
515 		}
516 
517 	case T_PGFLT:		/* system page fault */
518 		/*
519 		 * If we're under on_trap() protection (see <sys/ontrap.h>),
520 		 * set ot_trap and longjmp back to the on_trap() call site.
521 		 */
522 		if ((cur_thread->t_ontrap != NULL) &&
523 		    (cur_thread->t_ontrap->ot_prot & OT_DATA_ACCESS)) {
524 			curthread->t_ontrap->ot_trap |= OT_DATA_ACCESS;
525 			longjmp(&curthread->t_ontrap->ot_jmpbuf);
526 		}
527 
528 		/*
529 		 * See if we can handle as pagefault. Save lofault
530 		 * across this. Here we assume that an address
531 		 * less than KERNELBASE is a user fault.
532 		 * We can do this as copy.s routines verify that the
533 		 * starting address is less than KERNELBASE before
534 		 * starting and because we know that we always have
535 		 * KERNELBASE mapped as invalid to serve as a "barrier".
536 		 */
537 		lofault = cur_thread->t_lofault;
538 		cur_thread->t_lofault = 0;
539 
540 		mstate = new_mstate(cur_thread, LMS_KFAULT);
541 
542 		if (addr < (caddr_t)kernelbase) {
543 			res = pagefault(addr,
544 			    (errcode & PF_ERR_PROT)? F_PROT: F_INVAL, rw, 0);
545 			if (res == FC_NOMAP &&
546 			    addr < p->p_usrstack &&
547 			    grow(addr))
548 				res = 0;
549 		} else {
550 			res = pagefault(addr,
551 			    (errcode & PF_ERR_PROT)? F_PROT: F_INVAL, rw, 1);
552 		}
553 		(void) new_mstate(cur_thread, mstate);
554 
555 		/*
556 		 * Restore lofault. If we resolved the fault, exit.
557 		 * If we didn't and lofault wasn't set, die.
558 		 */
559 		cur_thread->t_lofault = lofault;
560 		if (res == 0)
561 			goto cleanup;
562 
563 #if defined(OPTERON_ERRATUM_93) && defined(_LP64)
564 		if (lofault == 0 && opteron_erratum_93) {
565 			/*
566 			 * Workaround for Opteron Erratum 93. On return from
567 			 * a System Managment Interrupt at a HLT instruction
568 			 * the %rip might be truncated to a 32 bit value.
569 			 * BIOS is supposed to fix this, but some don't.
570 			 * If this occurs we simply restore the high order bits.
571 			 * The HLT instruction is 1 byte of 0xf4.
572 			 */
573 			uintptr_t	rip = rp->r_pc;
574 
575 			if ((rip & 0xfffffffful) == rip) {
576 				rip |= 0xfffffffful << 32;
577 				if (hat_getpfnum(kas.a_hat, (caddr_t)rip) !=
578 				    PFN_INVALID &&
579 				    (*(uchar_t *)rip == 0xf4 ||
580 				    *(uchar_t *)(rip - 1) == 0xf4)) {
581 					rp->r_pc = rip;
582 					goto cleanup;
583 				}
584 			}
585 		}
586 #endif /* OPTERON_ERRATUM_93 && _LP64 */
587 
588 #ifdef OPTERON_ERRATUM_91
589 		if (lofault == 0 && opteron_erratum_91) {
590 			/*
591 			 * Workaround for Opteron Erratum 91. Prefetches may
592 			 * generate a page fault (they're not supposed to do
593 			 * that!). If this occurs we simply return back to the
594 			 * instruction.
595 			 */
596 			caddr_t		pc = (caddr_t)rp->r_pc;
597 
598 			/*
599 			 * If the faulting PC is not mapped, this is a
600 			 * legitimate kernel page fault that must result in a
601 			 * panic. If the faulting PC is mapped, it could contain
602 			 * a prefetch instruction. Check for that here.
603 			 */
604 			if (hat_getpfnum(kas.a_hat, pc) != PFN_INVALID) {
605 				if (cmp_to_prefetch((uchar_t *)pc)) {
606 #ifdef DEBUG
607 					cmn_err(CE_WARN, "Opteron erratum 91 "
608 					    "occurred: kernel prefetch"
609 					    " at %p generated a page fault!",
610 					    (void *)rp->r_pc);
611 #endif /* DEBUG */
612 					goto cleanup;
613 				}
614 			}
615 			(void) die(type, rp, addr, cpuid);
616 		}
617 #endif /* OPTERON_ERRATUM_91 */
618 
619 		if (lofault == 0)
620 			(void) die(type, rp, addr, cpuid);
621 
622 		/*
623 		 * Cannot resolve fault.  Return to lofault.
624 		 */
625 		if (lodebug) {
626 			showregs(type, rp, addr);
627 			traceregs(rp);
628 		}
629 		if (FC_CODE(res) == FC_OBJERR)
630 			res = FC_ERRNO(res);
631 		else
632 			res = EFAULT;
633 		rp->r_r0 = res;
634 		rp->r_pc = cur_thread->t_lofault;
635 		goto cleanup;
636 
637 	case T_PGFLT + USER:	/* user page fault */
638 		if (faultdebug) {
639 			char *fault_str;
640 
641 			switch (rw) {
642 			case S_READ:
643 				fault_str = "read";
644 				break;
645 			case S_WRITE:
646 				fault_str = "write";
647 				break;
648 			case S_EXEC:
649 				fault_str = "exec";
650 				break;
651 			default:
652 				fault_str = "";
653 				break;
654 			}
655 			printf("user %s fault:  addr=0x%lx errcode=0x%x\n",
656 			    fault_str, (uintptr_t)addr, errcode);
657 		}
658 
659 #if defined(OPTERON_ERRATUM_100) && defined(_LP64)
660 		/*
661 		 * Workaround for AMD erratum 100
662 		 *
663 		 * A 32-bit process may receive a page fault on a non
664 		 * 32-bit address by mistake. The range of the faulting
665 		 * address will be
666 		 *
667 		 *	0xffffffff80000000 .. 0xffffffffffffffff or
668 		 *	0x0000000100000000 .. 0x000000017fffffff
669 		 *
670 		 * The fault is always due to an instruction fetch, however
671 		 * the value of r_pc should be correct (in 32 bit range),
672 		 * so we ignore the page fault on the bogus address.
673 		 */
674 		if (p->p_model == DATAMODEL_ILP32 &&
675 		    (0xffffffff80000000 <= (uintptr_t)addr ||
676 		    (0x100000000 <= (uintptr_t)addr &&
677 		    (uintptr_t)addr <= 0x17fffffff))) {
678 			if (!opteron_erratum_100)
679 				panic("unexpected erratum #100");
680 			if (rp->r_pc <= 0xffffffff)
681 				goto out;
682 		}
683 #endif /* OPTERON_ERRATUM_100 && _LP64 */
684 
685 		ASSERT(!(curthread->t_flag & T_WATCHPT));
686 		watchpage = (pr_watch_active(p) && pr_is_watchpage(addr, rw));
687 #ifdef __i386
688 		/*
689 		 * In 32-bit mode, the lcall (system call) instruction fetches
690 		 * one word from the stack, at the stack pointer, because of the
691 		 * way the call gate is constructed.  This is a bogus
692 		 * read and should not be counted as a read watchpoint.
693 		 * We work around the problem here by testing to see if
694 		 * this situation applies and, if so, simply jumping to
695 		 * the code in locore.s that fields the system call trap.
696 		 * The registers on the stack are already set up properly
697 		 * due to the match between the call gate sequence and the
698 		 * trap gate sequence.  We just have to adjust the pc.
699 		 */
700 		if (watchpage && addr == (caddr_t)rp->r_sp &&
701 		    rw == S_READ && instr_is_syscall((caddr_t)rp->r_pc)) {
702 			extern void watch_syscall(void);
703 
704 			rp->r_pc += LCALLSIZE;
705 			watch_syscall();	/* never returns */
706 			/* NOTREACHED */
707 		}
708 #endif /* __i386 */
709 		vaddr = addr;
710 		if (!watchpage || (sz = instr_size(rp, &vaddr, rw)) <= 0)
711 			fault_type = (errcode & PF_ERR_PROT)? F_PROT: F_INVAL;
712 		else if ((watchcode = pr_is_watchpoint(&vaddr, &ta,
713 		    sz, NULL, rw)) != 0) {
714 			if (ta) {
715 				do_watch_step(vaddr, sz, rw,
716 					watchcode, rp->r_pc);
717 				fault_type = F_INVAL;
718 			} else {
719 				bzero(&siginfo, sizeof (siginfo));
720 				siginfo.si_signo = SIGTRAP;
721 				siginfo.si_code = watchcode;
722 				siginfo.si_addr = vaddr;
723 				siginfo.si_trapafter = 0;
724 				siginfo.si_pc = (caddr_t)rp->r_pc;
725 				fault = FLTWATCH;
726 				break;
727 			}
728 		} else {
729 			/* XXX pr_watch_emul() never succeeds (for now) */
730 			if (rw != S_EXEC && pr_watch_emul(rp, vaddr, rw))
731 				goto out;
732 			do_watch_step(vaddr, sz, rw, 0, 0);
733 			fault_type = F_INVAL;
734 		}
735 
736 		res = pagefault(addr, fault_type, rw, 0);
737 
738 		/*
739 		 * If pagefault() succeeded, ok.
740 		 * Otherwise attempt to grow the stack.
741 		 */
742 		if (res == 0 ||
743 		    (res == FC_NOMAP &&
744 		    addr < p->p_usrstack &&
745 		    grow(addr))) {
746 			lwp->lwp_lastfault = FLTPAGE;
747 			lwp->lwp_lastfaddr = addr;
748 			if (prismember(&p->p_fltmask, FLTPAGE)) {
749 				bzero(&siginfo, sizeof (siginfo));
750 				siginfo.si_addr = addr;
751 				(void) stop_on_fault(FLTPAGE, &siginfo);
752 			}
753 			goto out;
754 		} else if (res == FC_PROT && addr < p->p_usrstack &&
755 		    (mmu.pt_nx != 0 && (errcode & PF_ERR_EXEC))) {
756 			report_stack_exec(p, addr);
757 		}
758 
759 #ifdef OPTERON_ERRATUM_91
760 		/*
761 		 * Workaround for Opteron Erratum 91. Prefetches may generate a
762 		 * page fault (they're not supposed to do that!). If this
763 		 * occurs we simply return back to the instruction.
764 		 *
765 		 * We rely on copyin to properly fault in the page with r_pc.
766 		 */
767 		if (opteron_erratum_91 &&
768 		    addr != (caddr_t)rp->r_pc &&
769 		    instr_is_prefetch((caddr_t)rp->r_pc)) {
770 #ifdef DEBUG
771 			cmn_err(CE_WARN, "Opteron erratum 91 occurred: "
772 			    "prefetch at %p in pid %d generated a trap!",
773 			    (void *)rp->r_pc, p->p_pid);
774 #endif /* DEBUG */
775 			goto out;
776 		}
777 #endif /* OPTERON_ERRATUM_91 */
778 
779 		if (tudebug)
780 			showregs(type, rp, addr);
781 		/*
782 		 * In the case where both pagefault and grow fail,
783 		 * set the code to the value provided by pagefault.
784 		 * We map all errors returned from pagefault() to SIGSEGV.
785 		 */
786 		bzero(&siginfo, sizeof (siginfo));
787 		siginfo.si_addr = addr;
788 		switch (FC_CODE(res)) {
789 		case FC_HWERR:
790 		case FC_NOSUPPORT:
791 			siginfo.si_signo = SIGBUS;
792 			siginfo.si_code = BUS_ADRERR;
793 			fault = FLTACCESS;
794 			break;
795 		case FC_ALIGN:
796 			siginfo.si_signo = SIGBUS;
797 			siginfo.si_code = BUS_ADRALN;
798 			fault = FLTACCESS;
799 			break;
800 		case FC_OBJERR:
801 			if ((siginfo.si_errno = FC_ERRNO(res)) != EINTR) {
802 				siginfo.si_signo = SIGBUS;
803 				siginfo.si_code = BUS_OBJERR;
804 				fault = FLTACCESS;
805 			}
806 			break;
807 		default:	/* FC_NOMAP or FC_PROT */
808 			siginfo.si_signo = SIGSEGV;
809 			siginfo.si_code =
810 			    (res == FC_NOMAP)? SEGV_MAPERR : SEGV_ACCERR;
811 			fault = FLTBOUNDS;
812 			break;
813 		}
814 		break;
815 
816 	case T_ILLINST + USER:	/* invalid opcode fault */
817 		/*
818 		 * If the syscall instruction is disabled due to LDT usage, a
819 		 * user program that attempts to execute it will trigger a #ud
820 		 * trap. Check for that case here. If this occurs on a CPU which
821 		 * doesn't even support syscall, the result of all of this will
822 		 * be to emulate that particular instruction.
823 		 */
824 		if (p->p_ldt != NULL &&
825 		    instr_is_fast_syscall((caddr_t)rp->r_pc, X86_ASYSC)) {
826 			if (rewrite_syscall((caddr_t)rp->r_pc) == 0)
827 				goto out;
828 #ifdef DEBUG
829 			else
830 				cmn_err(CE_WARN, "failed to rewrite syscall "
831 				    "instruction in process %d",
832 				    curthread->t_procp->p_pid);
833 #endif /* DEBUG */
834 		}
835 
836 #ifdef __amd64
837 		/*
838 		 * Emulate the LAHF and SAHF instructions if needed.
839 		 * See the instr_is_lsahf function for details.
840 		 */
841 		if (p->p_model == DATAMODEL_LP64 &&
842 		    instr_is_lsahf((caddr_t)rp->r_pc, &instr)) {
843 			emulate_lsahf(rp, instr);
844 			goto out;
845 		}
846 #endif
847 
848 		/*FALLTHROUGH*/
849 
850 		if (tudebug)
851 			showregs(type, rp, (caddr_t)0);
852 		siginfo.si_signo = SIGILL;
853 		siginfo.si_code  = ILL_ILLOPC;
854 		siginfo.si_addr  = (caddr_t)rp->r_pc;
855 		fault = FLTILL;
856 		break;
857 
858 	case T_ZERODIV + USER:		/* integer divide by zero */
859 		if (tudebug && tudebugfpe)
860 			showregs(type, rp, (caddr_t)0);
861 		siginfo.si_signo = SIGFPE;
862 		siginfo.si_code  = FPE_INTDIV;
863 		siginfo.si_addr  = (caddr_t)rp->r_pc;
864 		fault = FLTIZDIV;
865 		break;
866 
867 	case T_OVFLW + USER:	/* integer overflow */
868 		if (tudebug && tudebugfpe)
869 			showregs(type, rp, (caddr_t)0);
870 		siginfo.si_signo = SIGFPE;
871 		siginfo.si_code  = FPE_INTOVF;
872 		siginfo.si_addr  = (caddr_t)rp->r_pc;
873 		fault = FLTIOVF;
874 		break;
875 
876 	case T_NOEXTFLT + USER:	/* math coprocessor not available */
877 		if (tudebug && tudebugfpe)
878 			showregs(type, rp, addr);
879 		if (fpnoextflt(rp)) {
880 			siginfo.si_signo = SIGFPE;
881 			siginfo.si_code  = ILL_ILLOPC;
882 			siginfo.si_addr  = (caddr_t)rp->r_pc;
883 			fault = FLTFPE;
884 		}
885 		break;
886 
887 	case T_EXTOVRFLT:	/* extension overrun fault */
888 		/* check if we took a kernel trap on behalf of user */
889 		{
890 			extern  void ndptrap_frstor(void);
891 			if (rp->r_pc != (uintptr_t)ndptrap_frstor)
892 				(void) die(type, rp, addr, cpuid);
893 			type |= USER;
894 		}
895 		/*FALLTHROUGH*/
896 	case T_EXTOVRFLT + USER:	/* extension overrun fault */
897 		if (tudebug && tudebugfpe)
898 			showregs(type, rp, addr);
899 		if (fpextovrflt(rp)) {
900 			siginfo.si_signo = SIGSEGV;
901 			siginfo.si_code  = SEGV_MAPERR;
902 			siginfo.si_addr  = (caddr_t)rp->r_pc;
903 			fault = FLTBOUNDS;
904 		}
905 		break;
906 
907 	case T_EXTERRFLT:	/* x87 floating point exception pending */
908 		/* check if we took a kernel trap on behalf of user */
909 		{
910 			extern  void ndptrap_frstor(void);
911 			if (rp->r_pc != (uintptr_t)ndptrap_frstor)
912 				(void) die(type, rp, addr, cpuid);
913 			type |= USER;
914 		}
915 		/*FALLTHROUGH*/
916 
917 	case T_EXTERRFLT + USER: /* x87 floating point exception pending */
918 		if (tudebug && tudebugfpe)
919 			showregs(type, rp, addr);
920 		if (sicode = fpexterrflt(rp)) {
921 			siginfo.si_signo = SIGFPE;
922 			siginfo.si_code  = sicode;
923 			siginfo.si_addr  = (caddr_t)rp->r_pc;
924 			fault = FLTFPE;
925 		}
926 		break;
927 
928 	case T_SIMDFPE + USER:		/* SSE and SSE2 exceptions */
929 		if (tudebug && tudebugsse)
930 			showregs(type, rp, addr);
931 		if ((x86_feature & (X86_SSE|X86_SSE2)) == 0) {
932 			/*
933 			 * There are rumours that some user instructions
934 			 * on older CPUs can cause this trap to occur; in
935 			 * which case send a SIGILL instead of a SIGFPE.
936 			 */
937 			siginfo.si_signo = SIGILL;
938 			siginfo.si_code  = ILL_ILLTRP;
939 			siginfo.si_addr  = (caddr_t)rp->r_pc;
940 			siginfo.si_trapno = type & ~USER;
941 			fault = FLTILL;
942 		} else if ((sicode = fpsimderrflt(rp)) != 0) {
943 			siginfo.si_signo = SIGFPE;
944 			siginfo.si_code = sicode;
945 			siginfo.si_addr = (caddr_t)rp->r_pc;
946 			fault = FLTFPE;
947 		}
948 		break;
949 
950 	case T_BPTFLT:	/* breakpoint trap */
951 		/*
952 		 * Kernel breakpoint traps should only happen when kmdb is
953 		 * active, and even then, it'll have interposed on the IDT, so
954 		 * control won't get here.  If it does, we've hit a breakpoint
955 		 * without the debugger, which is very strange, and very
956 		 * fatal.
957 		 */
958 		if (tudebug && tudebugbpt)
959 			showregs(type, rp, (caddr_t)0);
960 
961 		(void) die(type, rp, addr, cpuid);
962 		break;
963 
964 	case T_SGLSTP: /* single step/hw breakpoint exception */
965 		if (tudebug && tudebugbpt)
966 			showregs(type, rp, (caddr_t)0);
967 
968 		/* Now evaluate how we got here */
969 		if (lwp != NULL && (lwp->lwp_pcb.pcb_drstat & DR_SINGLESTEP)) {
970 			/*
971 			 * i386 single-steps even through lcalls which
972 			 * change the privilege level. So we take a trap at
973 			 * the first instruction in privileged mode.
974 			 *
975 			 * Set a flag to indicate that upon completion of
976 			 * the system call, deal with the single-step trap.
977 			 *
978 			 * The same thing happens for sysenter, too.
979 			 */
980 #if defined(__amd64)
981 			if (rp->r_pc == (uintptr_t)sys_sysenter) {
982 				/*
983 				 * Adjust the pc so that we don't execute the
984 				 * swapgs instruction at the head of the
985 				 * handler and completely confuse things.
986 				 */
987 				rp->r_pc = (uintptr_t)
988 				    _sys_sysenter_post_swapgs;
989 #elif defined(__i386)
990 			if (rp->r_pc == (uintptr_t)sys_call ||
991 			    rp->r_pc == (uintptr_t)sys_sysenter) {
992 #endif
993 				rp->r_ps &= ~PS_T; /* turn off trace */
994 				lwp->lwp_pcb.pcb_flags |= DEBUG_PENDING;
995 				cur_thread->t_post_sys = 1;
996 				goto cleanup;
997 			}
998 		}
999 		/* XXX - needs review on debugger interface? */
1000 		if (boothowto & RB_DEBUG)
1001 			debug_enter((char *)NULL);
1002 		else
1003 			(void) die(type, rp, addr, cpuid);
1004 		break;
1005 
1006 	case T_NMIFLT:	/* NMI interrupt */
1007 		printf("Unexpected NMI in system mode\n");
1008 		goto cleanup;
1009 
1010 	case T_NMIFLT + USER:	/* NMI interrupt */
1011 		printf("Unexpected NMI in user mode\n");
1012 		break;
1013 
1014 	case T_GPFLT:	/* general protection violation */
1015 #if defined(__amd64)
1016 		/*
1017 		 * On amd64, we can get a #gp from referencing addresses
1018 		 * in the virtual address hole e.g. from a copyin.
1019 		 */
1020 
1021 		/*
1022 		 * If we're under on_trap() protection (see <sys/ontrap.h>),
1023 		 * set ot_trap and longjmp back to the on_trap() call site.
1024 		 */
1025 		if ((cur_thread->t_ontrap != NULL) &&
1026 		    (cur_thread->t_ontrap->ot_prot & OT_DATA_ACCESS)) {
1027 			curthread->t_ontrap->ot_trap |= OT_DATA_ACCESS;
1028 			longjmp(&curthread->t_ontrap->ot_jmpbuf);
1029 		}
1030 
1031 		/*
1032 		 * If we're under lofault protection (copyin etc.),
1033 		 * longjmp back to lofault with an EFAULT.
1034 		 */
1035 		if (cur_thread->t_lofault) {
1036 			/*
1037 			 * Fault is not resolvable, so just return to lofault
1038 			 */
1039 			if (lodebug) {
1040 				showregs(type, rp, addr);
1041 				traceregs(rp);
1042 			}
1043 			rp->r_r0 = EFAULT;
1044 			rp->r_pc = cur_thread->t_lofault;
1045 			goto cleanup;
1046 		}
1047 		/*FALLTHROUGH*/
1048 #endif
1049 	case T_STKFLT:	/* stack fault */
1050 	case T_TSSFLT:	/* invalid TSS fault */
1051 	case T_SEGFLT:	/* segment not present fault */
1052 		if (tudebug)
1053 			showregs(type, rp, (caddr_t)0);
1054 		if (kern_gpfault(rp))
1055 			(void) die(type, rp, addr, cpuid);
1056 		goto cleanup;
1057 		/*FALLTHROUGH*/
1058 
1059 /*
1060  * ONLY 32-bit PROCESSES can USE a PRIVATE LDT! 64-bit apps should have
1061  * no legacy need for them, so we put a stop to it here.
1062  *
1063  * So: not-present fault is ONLY valid for 32-bit processes with a private LDT
1064  * trying to do a system call. Emulate it.
1065  *
1066  * #gp fault is ONLY valid for 32-bit processes also, which DO NOT have private
1067  * LDT, and are trying to do a system call. Emulate it.
1068  */
1069 	case T_SEGFLT + USER:	/* segment not present fault */
1070 	case T_GPFLT + USER:	/* general protection violation */
1071 #ifdef _SYSCALL32_IMPL
1072 		if (p->p_model != DATAMODEL_NATIVE) {
1073 #endif /* _SYSCALL32_IMPL */
1074 		if (instr_is_syscall((caddr_t)rp->r_pc)) {
1075 			if (type == T_SEGFLT + USER)
1076 				ASSERT(p->p_ldt != NULL);
1077 
1078 			if ((p->p_ldt == NULL && type == T_GPFLT + USER) ||
1079 			    type == T_SEGFLT + USER) {
1080 
1081 			/*
1082 			 * The user attempted a system call via the obsolete
1083 			 * call gate mechanism. Because the process doesn't have
1084 			 * an LDT (i.e. the ldtr contains 0), a #gp results.
1085 			 * Emulate the syscall here, just as we do above for a
1086 			 * #np trap.
1087 			 */
1088 
1089 			/*
1090 			 * Since this is a not-present trap, rp->r_pc points to
1091 			 * the trapping lcall instruction. We need to bump it
1092 			 * to the next insn so the app can continue on.
1093 			 */
1094 			rp->r_pc += LCALLSIZE;
1095 			lwp->lwp_regs = rp;
1096 
1097 			/*
1098 			 * Normally the microstate of the LWP is forced back to
1099 			 * LMS_USER by the syscall handlers. Emulate that
1100 			 * behavior here.
1101 			 */
1102 			mstate = LMS_USER;
1103 
1104 			dosyscall();
1105 			goto out;
1106 			}
1107 		}
1108 #ifdef _SYSCALL32_IMPL
1109 		}
1110 #endif /* _SYSCALL32_IMPL */
1111 		/*
1112 		 * If the current process is using a private LDT and the
1113 		 * trapping instruction is sysenter, the sysenter instruction
1114 		 * has been disabled on the CPU because it destroys segment
1115 		 * registers. If this is the case, rewrite the instruction to
1116 		 * be a safe system call and retry it. If this occurs on a CPU
1117 		 * which doesn't even support sysenter, the result of all of
1118 		 * this will be to emulate that particular instruction.
1119 		 */
1120 		if (p->p_ldt != NULL &&
1121 		    instr_is_fast_syscall((caddr_t)rp->r_pc, X86_SEP)) {
1122 			if (rewrite_syscall((caddr_t)rp->r_pc) == 0)
1123 				goto out;
1124 #ifdef DEBUG
1125 			else
1126 				cmn_err(CE_WARN, "failed to rewrite sysenter "
1127 				    "instruction in process %d",
1128 				    curthread->t_procp->p_pid);
1129 #endif /* DEBUG */
1130 		}
1131 		/*FALLTHROUGH*/
1132 
1133 	case T_BOUNDFLT + USER:	/* bound fault */
1134 	case T_STKFLT + USER:	/* stack fault */
1135 	case T_TSSFLT + USER:	/* invalid TSS fault */
1136 		if (tudebug)
1137 			showregs(type, rp, (caddr_t)0);
1138 		siginfo.si_signo = SIGSEGV;
1139 		siginfo.si_code  = SEGV_MAPERR;
1140 		siginfo.si_addr  = (caddr_t)rp->r_pc;
1141 		fault = FLTBOUNDS;
1142 		break;
1143 
1144 	case T_ALIGNMENT + USER:	/* user alignment error (486) */
1145 		if (tudebug)
1146 			showregs(type, rp, (caddr_t)0);
1147 		bzero(&siginfo, sizeof (siginfo));
1148 		siginfo.si_signo = SIGBUS;
1149 		siginfo.si_code = BUS_ADRALN;
1150 		siginfo.si_addr = (caddr_t)rp->r_pc;
1151 		fault = FLTACCESS;
1152 		break;
1153 
1154 	case T_SGLSTP + USER: /* single step/hw breakpoint exception */
1155 		if (tudebug && tudebugbpt)
1156 			showregs(type, rp, (caddr_t)0);
1157 
1158 		/* Was it single-stepping? */
1159 		if (lwp->lwp_pcb.pcb_drstat & DR_SINGLESTEP) {
1160 			pcb_t *pcb = &lwp->lwp_pcb;
1161 
1162 			rp->r_ps &= ~PS_T;
1163 			/*
1164 			 * If both NORMAL_STEP and WATCH_STEP are in effect,
1165 			 * give precedence to NORMAL_STEP.  If neither is set,
1166 			 * user must have set the PS_T bit in %efl; treat this
1167 			 * as NORMAL_STEP.
1168 			 */
1169 			if ((pcb->pcb_flags & NORMAL_STEP) ||
1170 			    !(pcb->pcb_flags & WATCH_STEP)) {
1171 				siginfo.si_signo = SIGTRAP;
1172 				siginfo.si_code = TRAP_TRACE;
1173 				siginfo.si_addr = (caddr_t)rp->r_pc;
1174 				fault = FLTTRACE;
1175 				if (pcb->pcb_flags & WATCH_STEP)
1176 					(void) undo_watch_step(NULL);
1177 			} else {
1178 				fault = undo_watch_step(&siginfo);
1179 			}
1180 			pcb->pcb_flags &= ~(NORMAL_STEP|WATCH_STEP);
1181 		} else {
1182 			cmn_err(CE_WARN,
1183 			    "Unexpected INT 1 in user mode, dr6=%lx",
1184 			    lwp->lwp_pcb.pcb_drstat);
1185 		}
1186 		break;
1187 
1188 	case T_BPTFLT + USER:	/* breakpoint trap */
1189 		if (tudebug && tudebugbpt)
1190 			showregs(type, rp, (caddr_t)0);
1191 		/*
1192 		 * int 3 (the breakpoint instruction) leaves the pc referring
1193 		 * to the address one byte after the breakpointed address.
1194 		 * If the P_PR_BPTADJ flag has been set via /proc, We adjust
1195 		 * it back so it refers to the breakpointed address.
1196 		 */
1197 		if (p->p_proc_flag & P_PR_BPTADJ)
1198 			rp->r_pc--;
1199 		siginfo.si_signo = SIGTRAP;
1200 		siginfo.si_code  = TRAP_BRKPT;
1201 		siginfo.si_addr  = (caddr_t)rp->r_pc;
1202 		fault = FLTBPT;
1203 		break;
1204 
1205 	case T_AST:
1206 		/*
1207 		 * This occurs only after the cs register has been made to
1208 		 * look like a kernel selector, either through debugging or
1209 		 * possibly by functions like setcontext().  The thread is
1210 		 * about to cause a general protection fault at common_iret()
1211 		 * in locore.  We let that happen immediately instead of
1212 		 * doing the T_AST processing.
1213 		 */
1214 		goto cleanup;
1215 
1216 	case T_AST + USER:		/* profiling or resched pseudo trap */
1217 		if (lwp->lwp_pcb.pcb_flags & CPC_OVERFLOW) {
1218 			lwp->lwp_pcb.pcb_flags &= ~CPC_OVERFLOW;
1219 			if (kcpc_overflow_ast()) {
1220 				/*
1221 				 * Signal performance counter overflow
1222 				 */
1223 				if (tudebug)
1224 					showregs(type, rp, (caddr_t)0);
1225 				bzero(&siginfo, sizeof (siginfo));
1226 				siginfo.si_signo = SIGEMT;
1227 				siginfo.si_code = EMT_CPCOVF;
1228 				siginfo.si_addr = (caddr_t)rp->r_pc;
1229 				fault = FLTCPCOVF;
1230 			}
1231 		}
1232 		break;
1233 	}
1234 
1235 	/*
1236 	 * We can't get here from a system trap
1237 	 */
1238 	ASSERT(type & USER);
1239 
1240 	if (fault) {
1241 		/*
1242 		 * Remember the fault and fault adddress
1243 		 * for real-time (SIGPROF) profiling.
1244 		 */
1245 		lwp->lwp_lastfault = fault;
1246 		lwp->lwp_lastfaddr = siginfo.si_addr;
1247 
1248 		DTRACE_PROC2(fault, int, fault, ksiginfo_t *, &siginfo);
1249 
1250 		/*
1251 		 * If a debugger has declared this fault to be an
1252 		 * event of interest, stop the lwp.  Otherwise just
1253 		 * deliver the associated signal.
1254 		 */
1255 		if (siginfo.si_signo != SIGKILL &&
1256 		    prismember(&p->p_fltmask, fault) &&
1257 		    stop_on_fault(fault, &siginfo) == 0)
1258 			siginfo.si_signo = 0;
1259 	}
1260 
1261 	if (siginfo.si_signo)
1262 		trapsig(&siginfo, (fault == FLTCPCOVF)? 0 : 1);
1263 
1264 	if (lwp->lwp_oweupc)
1265 		profil_tick(rp->r_pc);
1266 
1267 	if (cur_thread->t_astflag | cur_thread->t_sig_check) {
1268 		/*
1269 		 * Turn off the AST flag before checking all the conditions that
1270 		 * may have caused an AST.  This flag is on whenever a signal or
1271 		 * unusual condition should be handled after the next trap or
1272 		 * syscall.
1273 		 */
1274 		astoff(cur_thread);
1275 		cur_thread->t_sig_check = 0;
1276 
1277 		mutex_enter(&p->p_lock);
1278 		if (curthread->t_proc_flag & TP_CHANGEBIND) {
1279 			timer_lwpbind();
1280 			curthread->t_proc_flag &= ~TP_CHANGEBIND;
1281 		}
1282 		mutex_exit(&p->p_lock);
1283 
1284 		/*
1285 		 * for kaio requests that are on the per-process poll queue,
1286 		 * aiop->aio_pollq, they're AIO_POLL bit is set, the kernel
1287 		 * should copyout their result_t to user memory. by copying
1288 		 * out the result_t, the user can poll on memory waiting
1289 		 * for the kaio request to complete.
1290 		 */
1291 		if (p->p_aio)
1292 			aio_cleanup(0);
1293 		/*
1294 		 * If this LWP was asked to hold, call holdlwp(), which will
1295 		 * stop.  holdlwps() sets this up and calls pokelwps() which
1296 		 * sets the AST flag.
1297 		 *
1298 		 * Also check TP_EXITLWP, since this is used by fresh new LWPs
1299 		 * through lwp_rtt().  That flag is set if the lwp_create(2)
1300 		 * syscall failed after creating the LWP.
1301 		 */
1302 		if (ISHOLD(p))
1303 			holdlwp();
1304 
1305 		/*
1306 		 * All code that sets signals and makes ISSIG evaluate true must
1307 		 * set t_astflag afterwards.
1308 		 */
1309 		if (ISSIG_PENDING(cur_thread, lwp, p)) {
1310 			if (issig(FORREAL))
1311 				psig();
1312 			cur_thread->t_sig_check = 1;
1313 		}
1314 
1315 		if (cur_thread->t_rprof != NULL) {
1316 			realsigprof(0, 0);
1317 			cur_thread->t_sig_check = 1;
1318 		}
1319 
1320 		/*
1321 		 * /proc can't enable/disable the trace bit itself
1322 		 * because that could race with the call gate used by
1323 		 * system calls via "lcall". If that happened, an
1324 		 * invalid EFLAGS would result. prstep()/prnostep()
1325 		 * therefore schedule an AST for the purpose.
1326 		 */
1327 		if (lwp->lwp_pcb.pcb_flags & REQUEST_STEP) {
1328 			lwp->lwp_pcb.pcb_flags &= ~REQUEST_STEP;
1329 			rp->r_ps |= PS_T;
1330 		}
1331 		if (lwp->lwp_pcb.pcb_flags & REQUEST_NOSTEP) {
1332 			lwp->lwp_pcb.pcb_flags &= ~REQUEST_NOSTEP;
1333 			rp->r_ps &= ~PS_T;
1334 		}
1335 	}
1336 
1337 out:	/* We can't get here from a system trap */
1338 	ASSERT(type & USER);
1339 
1340 	if (ISHOLD(p))
1341 		holdlwp();
1342 
1343 	/*
1344 	 * Set state to LWP_USER here so preempt won't give us a kernel
1345 	 * priority if it occurs after this point.  Call CL_TRAPRET() to
1346 	 * restore the user-level priority.
1347 	 *
1348 	 * It is important that no locks (other than spinlocks) be entered
1349 	 * after this point before returning to user mode (unless lwp_state
1350 	 * is set back to LWP_SYS).
1351 	 */
1352 	lwp->lwp_state = LWP_USER;
1353 
1354 	if (cur_thread->t_trapret) {
1355 		cur_thread->t_trapret = 0;
1356 		thread_lock(cur_thread);
1357 		CL_TRAPRET(cur_thread);
1358 		thread_unlock(cur_thread);
1359 	}
1360 	if (CPU->cpu_runrun)
1361 		preempt();
1362 	(void) new_mstate(cur_thread, mstate);
1363 
1364 	/* Kernel probe */
1365 	TNF_PROBE_1(thread_state, "thread", /* CSTYLED */,
1366 	    tnf_microstate, state, LMS_USER);
1367 
1368 	return;
1369 
1370 cleanup:	/* system traps end up here */
1371 	ASSERT(!(type & USER));
1372 }
1373 
1374 /*
1375  * Patch non-zero to disable preemption of threads in the kernel.
1376  */
1377 int IGNORE_KERNEL_PREEMPTION = 0;	/* XXX - delete this someday */
1378 
1379 struct kpreempt_cnts {		/* kernel preemption statistics */
1380 	int	kpc_idle;	/* executing idle thread */
1381 	int	kpc_intr;	/* executing interrupt thread */
1382 	int	kpc_clock;	/* executing clock thread */
1383 	int	kpc_blocked;	/* thread has blocked preemption (t_preempt) */
1384 	int	kpc_notonproc;	/* thread is surrendering processor */
1385 	int	kpc_inswtch;	/* thread has ratified scheduling decision */
1386 	int	kpc_prilevel;	/* processor interrupt level is too high */
1387 	int	kpc_apreempt;	/* asynchronous preemption */
1388 	int	kpc_spreempt;	/* synchronous preemption */
1389 } kpreempt_cnts;
1390 
1391 /*
1392  * kernel preemption: forced rescheduling, preempt the running kernel thread.
1393  *	the argument is old PIL for an interrupt,
1394  *	or the distingished value KPREEMPT_SYNC.
1395  */
1396 void
1397 kpreempt(int asyncspl)
1398 {
1399 	kthread_t *cur_thread = curthread;
1400 
1401 	if (IGNORE_KERNEL_PREEMPTION) {
1402 		aston(CPU->cpu_dispthread);
1403 		return;
1404 	}
1405 
1406 	/*
1407 	 * Check that conditions are right for kernel preemption
1408 	 */
1409 	do {
1410 		if (cur_thread->t_preempt) {
1411 			/*
1412 			 * either a privileged thread (idle, panic, interrupt)
1413 			 *	or will check when t_preempt is lowered
1414 			 */
1415 			if (cur_thread->t_pri < 0)
1416 				kpreempt_cnts.kpc_idle++;
1417 			else if (cur_thread->t_flag & T_INTR_THREAD) {
1418 				kpreempt_cnts.kpc_intr++;
1419 				if (cur_thread->t_pil == CLOCK_LEVEL)
1420 					kpreempt_cnts.kpc_clock++;
1421 			} else
1422 				kpreempt_cnts.kpc_blocked++;
1423 			aston(CPU->cpu_dispthread);
1424 			return;
1425 		}
1426 		if (cur_thread->t_state != TS_ONPROC ||
1427 		    cur_thread->t_disp_queue != CPU->cpu_disp) {
1428 			/* this thread will be calling swtch() shortly */
1429 			kpreempt_cnts.kpc_notonproc++;
1430 			if (CPU->cpu_thread != CPU->cpu_dispthread) {
1431 				/* already in swtch(), force another */
1432 				kpreempt_cnts.kpc_inswtch++;
1433 				siron();
1434 			}
1435 			return;
1436 		}
1437 		if (getpil() >= DISP_LEVEL) {
1438 			/*
1439 			 * We can't preempt this thread if it is at
1440 			 * a PIL >= DISP_LEVEL since it may be holding
1441 			 * a spin lock (like sched_lock).
1442 			 */
1443 			siron();	/* check back later */
1444 			kpreempt_cnts.kpc_prilevel++;
1445 			return;
1446 		}
1447 
1448 		if (asyncspl != KPREEMPT_SYNC)
1449 			kpreempt_cnts.kpc_apreempt++;
1450 		else
1451 			kpreempt_cnts.kpc_spreempt++;
1452 
1453 		cur_thread->t_preempt++;
1454 		preempt();
1455 		cur_thread->t_preempt--;
1456 	} while (CPU->cpu_kprunrun);
1457 }
1458 
1459 /*
1460  * Print out debugging info.
1461  */
1462 static void
1463 showregs(uint_t type, struct regs *rp, caddr_t addr)
1464 {
1465 	int s;
1466 
1467 	s = spl7();
1468 	type &= ~USER;
1469 	if (u.u_comm[0])
1470 		printf("%s: ", u.u_comm);
1471 	if (type < TRAP_TYPES)
1472 		printf("#%s %s\n", trap_type_mnemonic[type], trap_type[type]);
1473 	else
1474 		switch (type) {
1475 		case T_SYSCALL:
1476 			printf("Syscall Trap:\n");
1477 			break;
1478 		case T_AST:
1479 			printf("AST\n");
1480 			break;
1481 		default:
1482 			printf("Bad Trap = %d\n", type);
1483 			break;
1484 		}
1485 	if (type == T_PGFLT) {
1486 		printf("Bad %s fault at addr=0x%lx\n",
1487 		    USERMODE(rp->r_cs) ? "user": "kernel", (uintptr_t)addr);
1488 	} else if (addr) {
1489 		printf("addr=0x%lx\n", (uintptr_t)addr);
1490 	}
1491 
1492 	printf("pid=%d, pc=0x%lx, sp=0x%lx, eflags=0x%lx\n",
1493 	    (ttoproc(curthread) && ttoproc(curthread)->p_pidp) ?
1494 	    ttoproc(curthread)->p_pid : 0, rp->r_pc, rp->r_sp, rp->r_ps);
1495 
1496 #if defined(__lint)
1497 	/*
1498 	 * this clause can be deleted when lint bug 4870403 is fixed
1499 	 * (lint thinks that bit 32 is illegal in a %b format string)
1500 	 */
1501 	printf("cr0: %x cr4: %b\n",
1502 	    (uint_t)getcr0(), (uint_t)getcr4(), FMT_CR4);
1503 #else
1504 	printf("cr0: %b cr4: %b\n",
1505 	    (uint_t)getcr0(), FMT_CR0, (uint_t)getcr4(), FMT_CR4);
1506 #endif
1507 
1508 #if defined(__amd64)
1509 	printf("cr2: %lx cr3: %lx cr8: %lx\n", getcr2(), getcr3(), getcr8());
1510 #elif defined(__i386)
1511 	printf("cr2: %lx cr3: %lx\n", getcr2(), getcr3());
1512 #endif
1513 
1514 	dumpregs(rp);
1515 	splx(s);
1516 }
1517 
1518 static void
1519 dumpregs(struct regs *rp)
1520 {
1521 #if defined(__amd64)
1522 	const char fmt[] = "\t%3s: %16lx %3s: %16lx %3s: %16lx\n";
1523 
1524 	printf(fmt, "rdi", rp->r_rdi, "rsi", rp->r_rsi, "rdx", rp->r_rdx);
1525 	printf(fmt, "rcx", rp->r_rcx, " r8", rp->r_r8, " r9", rp->r_r9);
1526 	printf(fmt, "rax", rp->r_rax, "rbx", rp->r_rbx, "rbp", rp->r_rbp);
1527 	printf(fmt, "r10", rp->r_r10, "r11", rp->r_r11, "r12", rp->r_r12);
1528 	printf(fmt, "r13", rp->r_r13, "r14", rp->r_r14, "r15", rp->r_r15);
1529 
1530 	printf(fmt, "fsb", rp->r_fsbase, "gsb", rp->r_gsbase, " ds", rp->r_ds);
1531 	printf(fmt, " es", rp->r_es, " fs", rp->r_fs, " gs", rp->r_gs);
1532 
1533 	printf(fmt, "trp", rp->r_trapno, "err", rp->r_err, "rip", rp->r_rip);
1534 	printf(fmt, " cs", rp->r_cs, "rfl", rp->r_rfl, "rsp", rp->r_rsp);
1535 
1536 	printf("\t%3s: %16lx\n", " ss", rp->r_ss);
1537 
1538 #elif defined(__i386)
1539 	const char fmt[] = "\t%3s: %8lx %3s: %8lx %3s: %8lx %3s: %8lx\n";
1540 
1541 	printf(fmt, " gs", rp->r_gs, " fs", rp->r_fs,
1542 	    " es", rp->r_es, " ds", rp->r_ds);
1543 	printf(fmt, "edi", rp->r_edi, "esi", rp->r_esi,
1544 	    "ebp", rp->r_ebp, "esp", rp->r_esp);
1545 	printf(fmt, "ebx", rp->r_ebx, "edx", rp->r_edx,
1546 	    "ecx", rp->r_ecx, "eax", rp->r_eax);
1547 	printf(fmt, "trp", rp->r_trapno, "err", rp->r_err,
1548 	    "eip", rp->r_eip, " cs", rp->r_cs);
1549 	printf("\t%3s: %8lx %3s: %8lx %3s: %8lx\n",
1550 	    "efl", rp->r_efl, "usp", rp->r_uesp, " ss", rp->r_ss);
1551 
1552 #endif	/* __i386 */
1553 }
1554 
1555 /*
1556  * Handle #gp faults in kernel mode.
1557  *
1558  * One legitimate way this can happen is if we attempt to update segment
1559  * registers to naughty values on the way out of the kernel.
1560  *
1561  * This can happen in a couple of ways: someone - either accidentally or
1562  * on purpose - creates (setcontext(2), lwp_create(2)) or modifies
1563  * (signal(2)) a ucontext that contains silly segment register values.
1564  * Or someone - either accidentally or on purpose - modifies the prgregset_t
1565  * of a subject process via /proc to contain silly segment register values.
1566  *
1567  * (The unfortunate part is that we can end up discovering the bad segment
1568  * register value in the middle of an 'iret' after we've popped most of the
1569  * stack.  So it becomes quite difficult to associate an accurate ucontext
1570  * with the lwp, because the act of taking the #gp trap overwrites most of
1571  * what we were going to send the lwp.)
1572  *
1573  * OTOH if it turns out that's -not- the problem, and we're -not- an lwp
1574  * trying to return to user mode and we get a #gp fault, then we need
1575  * to die() -- which will happen if we return non-zero from this routine.
1576  */
1577 static int
1578 kern_gpfault(struct regs *rp)
1579 {
1580 	kthread_t *t = curthread;
1581 	proc_t *p = ttoproc(t);
1582 	klwp_t *lwp = ttolwp(t);
1583 	struct regs tmpregs, *trp = NULL;
1584 	caddr_t pc = (caddr_t)rp->r_pc;
1585 	int v;
1586 
1587 	extern void _sys_rtt(), sr_sup();
1588 
1589 #if defined(__amd64)
1590 	extern void _update_sregs(), _update_sregs_done();
1591 	static const uint8_t iretq_insn[2] = { 0x48, 0xcf };
1592 
1593 #elif defined(__i386)
1594 	static const uint8_t iret_insn[1] = { 0xcf };
1595 
1596 	/*
1597 	 * Note carefully the appallingly awful dependency between
1598 	 * the instruction sequence used in __SEGREGS_POP and these
1599 	 * instructions encoded here.
1600 	 *
1601 	 * XX64	Add some commentary to locore.s/privregs.h to document this.
1602 	 */
1603 	static const uint8_t movw_0_esp_gs[4] = { 0x8e, 0x6c, 0x24, 0x0 };
1604 	static const uint8_t movw_4_esp_fs[4] = { 0x8e, 0x64, 0x24, 0x4 };
1605 	static const uint8_t movw_8_esp_es[4] = { 0x8e, 0x44, 0x24, 0x8 };
1606 	static const uint8_t movw_c_esp_ds[4] = { 0x8e, 0x5c, 0x24, 0xc };
1607 #endif
1608 	/*
1609 	 * if we're not an lwp, or the pc range is outside _sys_rtt, then
1610 	 * we should immediately be die()ing horribly
1611 	 */
1612 	if (lwp == NULL ||
1613 	    (uintptr_t)pc < (uintptr_t)_sys_rtt ||
1614 	    (uintptr_t)pc > (uintptr_t)sr_sup)
1615 		return (1);
1616 
1617 	/*
1618 	 * So at least we're in the right part of the kernel.
1619 	 *
1620 	 * Disassemble the instruction at the faulting pc.
1621 	 * Once we know what it is, we carefully reconstruct the stack
1622 	 * based on the order in which the stack is deconstructed in
1623 	 * _sys_rtt. Ew.
1624 	 */
1625 
1626 #if defined(__amd64)
1627 
1628 	if (bcmp(pc, iretq_insn, sizeof (iretq_insn)) == 0) {
1629 		/*
1630 		 * We took the #gp while trying to perform the iretq.
1631 		 * This means that either %cs or %ss are bad.
1632 		 * All we know for sure is that most of the general
1633 		 * registers have been restored, including the
1634 		 * segment registers, and all we have left on the
1635 		 * topmost part of the lwp's stack are the
1636 		 * registers that the iretq was unable to consume.
1637 		 *
1638 		 * All the rest of the state was crushed by the #gp
1639 		 * which pushed -its- registers atop our old save area
1640 		 * (because we had to decrement the stack pointer, sigh) so
1641 		 * all that we can try and do is to reconstruct the
1642 		 * crushed frame from the #gp trap frame itself.
1643 		 */
1644 		trp = &tmpregs;
1645 		trp->r_ss = lwptoregs(lwp)->r_ss;
1646 		trp->r_sp = lwptoregs(lwp)->r_sp;
1647 		trp->r_ps = lwptoregs(lwp)->r_ps;
1648 		trp->r_cs = lwptoregs(lwp)->r_cs;
1649 		trp->r_pc = lwptoregs(lwp)->r_pc;
1650 		bcopy(rp, trp, offsetof(struct regs, r_pc));
1651 
1652 		/*
1653 		 * Validate simple math
1654 		 */
1655 		ASSERT(trp->r_pc == lwptoregs(lwp)->r_pc);
1656 		ASSERT(trp->r_err == rp->r_err);
1657 
1658 	} else if ((lwp->lwp_pcb.pcb_flags & RUPDATE_PENDING) != 0 &&
1659 	    pc >= (caddr_t)_update_sregs &&
1660 	    pc < (caddr_t)_update_sregs_done) {
1661 		/*
1662 		 * This is the common case -- we're trying to load
1663 		 * a bad segment register value in the only section
1664 		 * of kernel code that ever loads segment registers.
1665 		 *
1666 		 * We don't need to do anything at this point because
1667 		 * the pcb contains all the pending segment register
1668 		 * state, and the regs are still intact because we
1669 		 * didn't adjust the stack pointer yet.  Given the fidelity
1670 		 * of all this, we could conceivably send a signal
1671 		 * to the lwp, rather than core-ing.
1672 		 */
1673 		trp = lwptoregs(lwp);
1674 		ASSERT((caddr_t)trp == (caddr_t)rp->r_sp);
1675 	}
1676 
1677 #elif defined(__i386)
1678 
1679 	if (bcmp(pc, iret_insn, sizeof (iret_insn)) == 0) {
1680 		/*
1681 		 * We took the #gp while trying to perform the iret.
1682 		 * This means that either %cs or %ss are bad.
1683 		 * All we know for sure is that most of the general
1684 		 * registers have been restored, including the
1685 		 * segment registers, and all we have left on the
1686 		 * topmost part of the lwp's stack are the registers that
1687 		 * the iret was unable to consume.
1688 		 *
1689 		 * All the rest of the state was crushed by the #gp
1690 		 * which pushed -its- registers atop our old save area
1691 		 * (because we had to decrement the stack pointer, sigh) so
1692 		 * all that we can try and do is to reconstruct the
1693 		 * crushed frame from the #gp trap frame itself.
1694 		 */
1695 		trp = &tmpregs;
1696 		trp->r_ss = lwptoregs(lwp)->r_ss;
1697 		trp->r_sp = lwptoregs(lwp)->r_sp;
1698 		trp->r_ps = lwptoregs(lwp)->r_ps;
1699 		trp->r_cs = lwptoregs(lwp)->r_cs;
1700 		trp->r_pc = lwptoregs(lwp)->r_pc;
1701 		bcopy(rp, trp, offsetof(struct regs, r_pc));
1702 
1703 		ASSERT(trp->r_pc == lwptoregs(lwp)->r_pc);
1704 		ASSERT(trp->r_err == rp->r_err);
1705 
1706 	} else {
1707 		/*
1708 		 * Segment registers are reloaded in _sys_rtt
1709 		 * via the following sequence:
1710 		 *
1711 		 *	movw	0(%esp), %gs
1712 		 *	movw	4(%esp), %fs
1713 		 *	movw	8(%esp), %es
1714 		 *	movw	12(%esp), %ds
1715 		 *	addl	$16, %esp
1716 		 *
1717 		 * Thus if any of them fault, we know the user
1718 		 * registers are left unharmed on the stack.
1719 		 */
1720 		if (bcmp(pc, movw_0_esp_gs, sizeof (movw_0_esp_gs)) == 0 ||
1721 		    bcmp(pc, movw_4_esp_fs, sizeof (movw_4_esp_fs)) == 0 ||
1722 		    bcmp(pc, movw_8_esp_es, sizeof (movw_8_esp_es)) == 0 ||
1723 		    bcmp(pc, movw_c_esp_ds, sizeof (movw_c_esp_ds)) == 0)
1724 			trp = lwptoregs(lwp);
1725 	}
1726 #endif	/* __amd64 */
1727 
1728 	if (trp == NULL)
1729 		return (1);
1730 
1731 	/*
1732 	 * If we get to here, we're reasonably confident that we've
1733 	 * correctly decoded what happened on the way out of the kernel.
1734 	 * Rewrite the lwp's registers so that we can create a core dump
1735 	 * the (at least vaguely) represents the mcontext we were
1736 	 * being asked to restore when things went so terribly wrong.
1737 	 */
1738 
1739 	/*
1740 	 * Make sure that we have a meaningful %trapno and %err.
1741 	 */
1742 	trp->r_trapno = rp->r_trapno;
1743 	trp->r_err = rp->r_err;
1744 
1745 	if ((caddr_t)trp != (caddr_t)lwptoregs(lwp))
1746 		bcopy(trp, lwptoregs(lwp), sizeof (*trp));
1747 
1748 	mutex_enter(&p->p_lock);
1749 	lwp->lwp_cursig = SIGSEGV;
1750 	mutex_exit(&p->p_lock);
1751 
1752 	/*
1753 	 * Terminate all LWPs but don't discard them.  If another lwp beat us to
1754 	 * the punch by calling exit(), evaporate now.
1755 	 */
1756 	proc_is_exiting(p);
1757 	if (exitlwps(1) != 0) {
1758 		mutex_enter(&p->p_lock);
1759 		lwp_exit();
1760 	}
1761 
1762 #ifdef C2_AUDIT
1763 	if (audit_active)		/* audit core dump */
1764 		audit_core_start(SIGSEGV);
1765 #endif
1766 	v = core(SIGSEGV, B_FALSE);
1767 #ifdef C2_AUDIT
1768 	if (audit_active)		/* audit core dump */
1769 		audit_core_finish(v ? CLD_KILLED : CLD_DUMPED);
1770 #endif
1771 	exit(v ? CLD_KILLED : CLD_DUMPED, SIGSEGV);
1772 	return (0);
1773 }
1774 
1775 /*
1776  * dump_tss() - Display the TSS structure
1777  */
1778 
1779 #if defined(__amd64)
1780 
1781 static void
1782 dump_tss(void)
1783 {
1784 	const char tss_fmt[] = "tss.%s:\t0x%p\n";  /* Format string */
1785 	struct tss *tss = CPU->cpu_tss;
1786 
1787 	printf(tss_fmt, "tss_rsp0", (void *)tss->tss_rsp0);
1788 	printf(tss_fmt, "tss_rsp1", (void *)tss->tss_rsp1);
1789 	printf(tss_fmt, "tss_rsp2", (void *)tss->tss_rsp2);
1790 
1791 	printf(tss_fmt, "tss_ist1", (void *)tss->tss_ist1);
1792 	printf(tss_fmt, "tss_ist2", (void *)tss->tss_ist2);
1793 	printf(tss_fmt, "tss_ist3", (void *)tss->tss_ist3);
1794 	printf(tss_fmt, "tss_ist4", (void *)tss->tss_ist4);
1795 	printf(tss_fmt, "tss_ist5", (void *)tss->tss_ist5);
1796 	printf(tss_fmt, "tss_ist6", (void *)tss->tss_ist6);
1797 	printf(tss_fmt, "tss_ist7", (void *)tss->tss_ist7);
1798 }
1799 
1800 #elif defined(__i386)
1801 
1802 static void
1803 dump_tss(void)
1804 {
1805 	const char tss_fmt[] = "tss.%s:\t0x%p\n";  /* Format string */
1806 	struct tss *tss = CPU->cpu_tss;
1807 
1808 	printf(tss_fmt, "tss_link", (void *)(uintptr_t)tss->tss_link);
1809 	printf(tss_fmt, "tss_esp0", (void *)(uintptr_t)tss->tss_esp0);
1810 	printf(tss_fmt, "tss_ss0", (void *)(uintptr_t)tss->tss_ss0);
1811 	printf(tss_fmt, "tss_esp1", (void *)(uintptr_t)tss->tss_esp1);
1812 	printf(tss_fmt, "tss_ss1", (void *)(uintptr_t)tss->tss_ss1);
1813 	printf(tss_fmt, "tss_esp2", (void *)(uintptr_t)tss->tss_esp2);
1814 	printf(tss_fmt, "tss_ss2", (void *)(uintptr_t)tss->tss_ss2);
1815 	printf(tss_fmt, "tss_cr3", (void *)(uintptr_t)tss->tss_cr3);
1816 	printf(tss_fmt, "tss_eip", (void *)(uintptr_t)tss->tss_eip);
1817 	printf(tss_fmt, "tss_eflags", (void *)(uintptr_t)tss->tss_eflags);
1818 	printf(tss_fmt, "tss_eax", (void *)(uintptr_t)tss->tss_eax);
1819 	printf(tss_fmt, "tss_ebx", (void *)(uintptr_t)tss->tss_ebx);
1820 	printf(tss_fmt, "tss_ecx", (void *)(uintptr_t)tss->tss_ecx);
1821 	printf(tss_fmt, "tss_edx", (void *)(uintptr_t)tss->tss_edx);
1822 	printf(tss_fmt, "tss_esp", (void *)(uintptr_t)tss->tss_esp);
1823 }
1824 
1825 #endif	/* __amd64 */
1826 
1827 #if defined(TRAPTRACE)
1828 
1829 int ttrace_nrec = 0;		/* number of records to dump out */
1830 int ttrace_dump_nregs = 5;	/* dump out this many records with regs too */
1831 
1832 /*
1833  * Dump out the last ttrace_nrec traptrace records on each CPU
1834  */
1835 static void
1836 dump_ttrace(void)
1837 {
1838 	trap_trace_ctl_t *ttc;
1839 	trap_trace_rec_t *rec;
1840 	uintptr_t current;
1841 	int i, j, k;
1842 	int n = NCPU;
1843 #if defined(__amd64)
1844 	const char banner[] =
1845 		"\ncpu          address    timestamp "
1846 		"type  vc  handler   pc\n";
1847 	const char fmt1[] = "%3d %016lx %12llx ";
1848 #elif defined(__i386)
1849 	const char banner[] =
1850 		"\ncpu address     timestamp type  vc  handler   pc\n";
1851 	const char fmt1[] = "%3d %08lx %12llx ";
1852 #endif
1853 	const char fmt2[] = "%4s %3x ";
1854 	const char fmt3[] = "%8s ";
1855 
1856 	if (ttrace_nrec == 0)
1857 		return;
1858 
1859 	printf(banner);
1860 
1861 	for (i = 0; i < n; i++) {
1862 		ttc = &trap_trace_ctl[i];
1863 		if (ttc->ttc_first == NULL)
1864 			continue;
1865 
1866 		current = ttc->ttc_next - sizeof (trap_trace_rec_t);
1867 		for (j = 0; j < ttrace_nrec; j++) {
1868 			struct sysent	*sys;
1869 			struct autovec	*vec;
1870 			extern struct av_head autovect[];
1871 			int type;
1872 			ulong_t	off;
1873 			char *sym, *stype;
1874 
1875 			if (current < ttc->ttc_first)
1876 				current =
1877 				    ttc->ttc_limit - sizeof (trap_trace_rec_t);
1878 
1879 			if (current == NULL)
1880 				continue;
1881 
1882 			rec = (trap_trace_rec_t *)current;
1883 
1884 			if (rec->ttr_stamp == 0)
1885 				break;
1886 
1887 			printf(fmt1, i, (uintptr_t)rec, rec->ttr_stamp);
1888 
1889 			switch (rec->ttr_marker) {
1890 			case TT_SYSCALL:
1891 			case TT_SYSENTER:
1892 			case TT_SYSC:
1893 			case TT_SYSC64:
1894 #if defined(__amd64)
1895 				sys = &sysent32[rec->ttr_sysnum];
1896 				switch (rec->ttr_marker) {
1897 				case TT_SYSC64:
1898 					sys = &sysent[rec->ttr_sysnum];
1899 					/*FALLTHROUGH*/
1900 #elif defined(__i386)
1901 				sys = &sysent[rec->ttr_sysnum];
1902 				switch (rec->ttr_marker) {
1903 				case TT_SYSC64:
1904 #endif
1905 				case TT_SYSC:
1906 					stype = "sysc";	/* syscall */
1907 					break;
1908 				case TT_SYSCALL:
1909 					stype = "lcal";	/* lcall */
1910 					break;
1911 				case TT_SYSENTER:
1912 					stype = "syse";	/* sysenter */
1913 					break;
1914 				default:
1915 					break;
1916 				}
1917 				printf(fmt2, "sysc", rec->ttr_sysnum);
1918 				if (sys != NULL) {
1919 					sym = kobj_getsymname(
1920 					    (uintptr_t)sys->sy_callc,
1921 					    &off);
1922 					if (sym != NULL)
1923 						printf("%s ", sym);
1924 					else
1925 						printf("%p ", sys->sy_callc);
1926 				} else {
1927 					printf("unknown ");
1928 				}
1929 				break;
1930 
1931 			case TT_INTERRUPT:
1932 				printf(fmt2, "intr", rec->ttr_vector);
1933 				vec = (&autovect[rec->ttr_vector])->avh_link;
1934 				if (vec != NULL) {
1935 					sym = kobj_getsymname(
1936 					    (uintptr_t)vec->av_vector, &off);
1937 					if (sym != NULL)
1938 						printf("%s ", sym);
1939 					else
1940 						printf("%p ", vec->av_vector);
1941 				} else {
1942 					printf("unknown ");
1943 				}
1944 				break;
1945 
1946 			case TT_TRAP:
1947 				type = rec->ttr_regs.r_trapno;
1948 				printf(fmt2, "trap", type);
1949 				printf("#%s ", type < TRAP_TYPES ?
1950 				    trap_type_mnemonic[type] : "trap");
1951 				break;
1952 
1953 			default:
1954 				break;
1955 			}
1956 
1957 			sym = kobj_getsymname(rec->ttr_regs.r_pc, &off);
1958 			if (sym != NULL)
1959 				printf("%s+%lx\n", sym, off);
1960 			else
1961 				printf("%lx\n", rec->ttr_regs.r_pc);
1962 
1963 			if (ttrace_dump_nregs-- > 0) {
1964 				int s;
1965 
1966 				if (rec->ttr_marker == TT_INTERRUPT)
1967 					printf(
1968 					    "\t\tipl %x spl %x pri %x\n",
1969 					    rec->ttr_ipl,
1970 					    rec->ttr_spl,
1971 					    rec->ttr_pri);
1972 
1973 				dumpregs(&rec->ttr_regs);
1974 
1975 				printf("\t%3s: %p\n\n", " ct",
1976 				    (void *)rec->ttr_curthread);
1977 
1978 				/*
1979 				 * print out the pc stack that we recorded
1980 				 * at trap time (if any)
1981 				 */
1982 				for (s = 0; s < rec->ttr_sdepth; s++) {
1983 					uintptr_t fullpc;
1984 
1985 					if (s >= TTR_STACK_DEPTH) {
1986 						printf("ttr_sdepth corrupt\n");
1987 						break;
1988 					}
1989 
1990 					fullpc = (uintptr_t)rec->ttr_stack[s];
1991 
1992 					sym = kobj_getsymname(fullpc, &off);
1993 					if (sym != NULL)
1994 						printf("-> %s+0x%lx()\n",
1995 						    sym, off);
1996 					else
1997 						printf("-> 0x%lx()\n", fullpc);
1998 				}
1999 				printf("\n");
2000 			}
2001 			current -= sizeof (trap_trace_rec_t);
2002 		}
2003 	}
2004 }
2005 
2006 #endif	/* TRAPTRACE */
2007 
2008 void
2009 panic_showtrap(struct trap_info *tip)
2010 {
2011 	showregs(tip->trap_type, tip->trap_regs, tip->trap_addr);
2012 
2013 #if defined(TRAPTRACE)
2014 	dump_ttrace();
2015 #endif	/* TRAPTRACE */
2016 
2017 	if (tip->trap_type == T_DBLFLT)
2018 		dump_tss();
2019 }
2020 
2021 void
2022 panic_savetrap(panic_data_t *pdp, struct trap_info *tip)
2023 {
2024 	panic_saveregs(pdp, tip->trap_regs);
2025 }
2026