xref: /titanic_41/usr/src/uts/i86pc/os/trap.c (revision 74e20cfe817b82802b16fac8690dadcda76f54f5)
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 (the "License").
6  * You may not use this file except in compliance with the License.
7  *
8  * You can obtain a copy of the license at usr/src/OPENSOLARIS.LICENSE
9  * or http://www.opensolaris.org/os/licensing.
10  * See the License for the specific language governing permissions
11  * and limitations under the License.
12  *
13  * When distributing Covered Code, include this CDDL HEADER in each
14  * file and include the License file at usr/src/OPENSOLARIS.LICENSE.
15  * If applicable, add the following below this CDDL HEADER, with the
16  * fields enclosed by brackets "[]" replaced with your own identifying
17  * information: Portions Copyright [yyyy] [name of copyright owner]
18  *
19  * CDDL HEADER END
20  */
21 
22 /*
23  * Copyright 2006 Sun Microsystems, Inc.  All rights reserved.
24  * Use is subject to license terms.
25  */
26 
27 /*	Copyright (c) 1990, 1991 UNIX System Laboratories, Inc. */
28 /*	Copyright (c) 1984, 1986, 1987, 1988, 1989, 1990 AT&T   */
29 /*		All Rights Reserved   				*/
30 /*								*/
31 /*	Copyright (c) 1987, 1988 Microsoft Corporation  	*/
32 /*		All Rights Reserved   				*/
33 /*								*/
34 
35 #pragma ident	"%Z%%M%	%I%	%E% SMI"
36 
37 #include <sys/types.h>
38 #include <sys/sysmacros.h>
39 #include <sys/param.h>
40 #include <sys/signal.h>
41 #include <sys/systm.h>
42 #include <sys/user.h>
43 #include <sys/proc.h>
44 #include <sys/disp.h>
45 #include <sys/class.h>
46 #include <sys/core.h>
47 #include <sys/syscall.h>
48 #include <sys/cpuvar.h>
49 #include <sys/vm.h>
50 #include <sys/sysinfo.h>
51 #include <sys/fault.h>
52 #include <sys/stack.h>
53 #include <sys/mmu.h>
54 #include <sys/psw.h>
55 #include <sys/regset.h>
56 #include <sys/fp.h>
57 #include <sys/trap.h>
58 #include <sys/kmem.h>
59 #include <sys/vtrace.h>
60 #include <sys/cmn_err.h>
61 #include <sys/prsystm.h>
62 #include <sys/mutex_impl.h>
63 #include <sys/machsystm.h>
64 #include <sys/archsystm.h>
65 #include <sys/sdt.h>
66 #include <sys/avintr.h>
67 #include <sys/kobj.h>
68 
69 #include <vm/hat.h>
70 
71 #include <vm/seg_kmem.h>
72 #include <vm/as.h>
73 #include <vm/seg.h>
74 #include <vm/hat_pte.h>
75 
76 #include <sys/procfs.h>
77 
78 #include <sys/reboot.h>
79 #include <sys/debug.h>
80 #include <sys/debugreg.h>
81 #include <sys/modctl.h>
82 #include <sys/aio_impl.h>
83 #include <sys/tnf.h>
84 #include <sys/tnf_probe.h>
85 #include <sys/cred.h>
86 #include <sys/mman.h>
87 #include <sys/x86_archext.h>
88 #include <sys/copyops.h>
89 #include <c2/audit.h>
90 #include <sys/ftrace.h>
91 #include <sys/panic.h>
92 #include <sys/traptrace.h>
93 #include <sys/ontrap.h>
94 #include <sys/cpc_impl.h>
95 
96 #define	USER	0x10000		/* user-mode flag added to trap type */
97 
98 static const char *trap_type_mnemonic[] = {
99 	"de",	"db",	"2",	"bp",
100 	"of",	"br",	"ud",	"nm",
101 	"df",	"9",	"ts",	"np",
102 	"ss",	"gp",	"pf",	"15",
103 	"mf",	"ac",	"mc",	"xf"
104 };
105 
106 static const char *trap_type[] = {
107 	"Divide error",				/* trap id 0 	*/
108 	"Debug",				/* trap id 1	*/
109 	"NMI interrupt",			/* trap id 2	*/
110 	"Breakpoint",				/* trap id 3 	*/
111 	"Overflow",				/* trap id 4 	*/
112 	"BOUND range exceeded",			/* trap id 5 	*/
113 	"Invalid opcode",			/* trap id 6 	*/
114 	"Device not available",			/* trap id 7 	*/
115 	"Double fault",				/* trap id 8 	*/
116 	"Coprocessor segment overrun",		/* trap id 9 	*/
117 	"Invalid TSS",				/* trap id 10 	*/
118 	"Segment not present",			/* trap id 11 	*/
119 	"Stack segment fault",			/* trap id 12 	*/
120 	"General protection",			/* trap id 13 	*/
121 	"Page fault",				/* trap id 14 	*/
122 	"Reserved",				/* trap id 15 	*/
123 	"x87 floating point error",		/* trap id 16 	*/
124 	"Alignment check",			/* trap id 17 	*/
125 	"Machine check",			/* trap id 18	*/
126 	"SIMD floating point exception",	/* trap id 19	*/
127 };
128 
129 #define	TRAP_TYPES	(sizeof (trap_type) / sizeof (trap_type[0]))
130 
131 int tudebug = 0;
132 int tudebugbpt = 0;
133 int tudebugfpe = 0;
134 int tudebugsse = 0;
135 
136 #if defined(TRAPDEBUG) || defined(lint)
137 int tdebug = 0;
138 int lodebug = 0;
139 int faultdebug = 0;
140 #else
141 #define	tdebug	0
142 #define	lodebug	0
143 #define	faultdebug	0
144 #endif /* defined(TRAPDEBUG) || defined(lint) */
145 
146 #if defined(TRAPTRACE)
147 static void dump_ttrace(void);
148 #endif	/* TRAPTRACE */
149 static void dumpregs(struct regs *);
150 static void showregs(uint_t, struct regs *, caddr_t);
151 static void dump_tss(void);
152 static int kern_gpfault(struct regs *);
153 
154 struct trap_info {
155 	struct regs *trap_regs;
156 	uint_t trap_type;
157 	caddr_t trap_addr;
158 };
159 
160 /*ARGSUSED*/
161 static int
162 die(uint_t type, struct regs *rp, caddr_t addr, processorid_t cpuid)
163 {
164 	struct trap_info ti;
165 	const char *trap_name, *trap_mnemonic;
166 
167 	if (type < TRAP_TYPES) {
168 		trap_name = trap_type[type];
169 		trap_mnemonic = trap_type_mnemonic[type];
170 	} else {
171 		trap_name = "trap";
172 		trap_mnemonic = "-";
173 	}
174 
175 #ifdef TRAPTRACE
176 	TRAPTRACE_FREEZE;
177 #endif
178 
179 	ti.trap_regs = rp;
180 	ti.trap_type = type & ~USER;
181 	ti.trap_addr = addr;
182 
183 	curthread->t_panic_trap = &ti;
184 
185 	if (type == T_PGFLT && addr < (caddr_t)KERNELBASE) {
186 		panic("BAD TRAP: type=%x (#%s %s) rp=%p addr=%p "
187 		    "occurred in module \"%s\" due to %s",
188 		    type, trap_mnemonic, trap_name, (void *)rp, (void *)addr,
189 		    mod_containing_pc((caddr_t)rp->r_pc),
190 		    addr < (caddr_t)PAGESIZE ?
191 		    "a NULL pointer dereference" :
192 		    "an illegal access to a user address");
193 	} else
194 		panic("BAD TRAP: type=%x (#%s %s) rp=%p addr=%p",
195 		    type, trap_mnemonic, trap_name, (void *)rp, (void *)addr);
196 	return (0);
197 }
198 
199 /*
200  * Rewrite the instruction at pc to be an int $T_SYSCALLINT instruction.
201  *
202  * int <vector> is two bytes: 0xCD <vector>
203  */
204 
205 #define	SLOW_SCALL_SIZE	2
206 
207 static int
208 rewrite_syscall(caddr_t pc)
209 {
210 	uchar_t instr[SLOW_SCALL_SIZE] = { 0xCD, T_SYSCALLINT };
211 
212 	if (uwrite(curthread->t_procp, instr, SLOW_SCALL_SIZE,
213 	    (uintptr_t)pc) != 0)
214 		return (1);
215 
216 	return (0);
217 }
218 
219 /*
220  * Test to see if the instruction at pc is sysenter or syscall. The second
221  * argument should be the x86 feature flag corresponding to the expected
222  * instruction.
223  *
224  * sysenter is two bytes: 0x0F 0x34
225  * syscall is two bytes:  0x0F 0x05
226  */
227 
228 #define	FAST_SCALL_SIZE	2
229 
230 static int
231 instr_is_fast_syscall(caddr_t pc, int which)
232 {
233 	uchar_t instr[FAST_SCALL_SIZE];
234 
235 	ASSERT(which == X86_SEP || which == X86_ASYSC);
236 
237 	if (copyin_nowatch(pc, (caddr_t)instr, FAST_SCALL_SIZE) != 0 ||
238 	    instr[0] != 0x0F)
239 		return (0);
240 
241 	if ((which == X86_SEP && instr[1] == 0x34) ||
242 	    (which == X86_ASYSC && instr[1] == 0x05))
243 		return (1);
244 
245 	return (0);
246 }
247 
248 /*
249  * Test to see if the instruction at pc is a system call instruction.
250  *
251  * The bytes of an lcall instruction used for the syscall trap.
252  * static uchar_t lcall[7] = { 0x9a, 0, 0, 0, 0, 0x7, 0 };
253  * static uchar_t lcallalt[7] = { 0x9a, 0, 0, 0, 0, 0x27, 0 };
254  */
255 
256 #define	LCALLSIZE	7
257 
258 static int
259 instr_is_syscall(caddr_t pc)
260 {
261 	uchar_t instr[LCALLSIZE];
262 
263 	if (copyin_nowatch(pc, (caddr_t)instr, LCALLSIZE) == 0 &&
264 	    instr[0] == 0x9a &&
265 	    instr[1] == 0 &&
266 	    instr[2] == 0 &&
267 	    instr[3] == 0 &&
268 	    instr[4] == 0 &&
269 	    (instr[5] == 0x7 || instr[5] == 0x27) &&
270 	    instr[6] == 0)
271 		return (1);
272 
273 	return (0);
274 }
275 
276 #ifdef __amd64
277 
278 /*
279  * In the first revisions of AMD64 CPUs produced by AMD, the LAHF and
280  * SAHF instructions were not implemented in 64bit mode. Later revisions
281  * did implement these instructions. An extension to the cpuid instruction
282  * was added to check for the capability of executing these instructions
283  * in 64bit mode.
284  *
285  * Intel originally did not implement these instructions in EM64T either,
286  * but added them in later revisions.
287  *
288  * So, there are different chip revisions by both vendors out there that
289  * may or may not implement these instructions. The easy solution is to
290  * just always emulate these instructions on demand.
291  *
292  * SAHF == store %ah in the lower 8 bits of %rflags (opcode 0x9e)
293  * LAHF == load the lower 8 bits of %rflags into %ah (opcode 0x9f)
294  */
295 
296 #define	LSAHFSIZE 1
297 
298 static int
299 instr_is_lsahf(caddr_t pc, uchar_t *instr)
300 {
301 	if (copyin_nowatch(pc, (caddr_t)instr, LSAHFSIZE) == 0 &&
302 	    (*instr == 0x9e || *instr == 0x9f))
303 		return (1);
304 	return (0);
305 }
306 
307 /*
308  * Emulate the LAHF and SAHF instructions. The reference manuals define
309  * these instructions to always load/store bit 1 as a 1, and bits 3 and 5
310  * as a 0. The other, defined, bits are copied (the PS_ICC bits and PS_P).
311  *
312  * Note that %ah is bits 8-15 of %rax.
313  */
314 static void
315 emulate_lsahf(struct regs *rp, uchar_t instr)
316 {
317 	if (instr == 0x9e) {
318 		/* sahf. Copy bits from %ah to flags. */
319 		rp->r_ps = (rp->r_ps & ~0xff) |
320 		    ((rp->r_rax >> 8) & PSL_LSAHFMASK) | PS_MB1;
321 	} else {
322 		/* lahf. Copy bits from flags to %ah. */
323 		rp->r_rax = (rp->r_rax & ~0xff00) |
324 		    (((rp->r_ps & PSL_LSAHFMASK) | PS_MB1) << 8);
325 	}
326 	rp->r_pc += LSAHFSIZE;
327 }
328 #endif /* __amd64 */
329 
330 #ifdef OPTERON_ERRATUM_91
331 
332 /*
333  * Test to see if the instruction at pc is a prefetch instruction.
334  *
335  * The first byte of prefetch instructions is always 0x0F.
336  * The second byte is 0x18 for regular prefetch or 0x0D for AMD 3dnow prefetch.
337  * The third byte is between 0 and 3 inclusive.
338  */
339 
340 #define	PREFETCHSIZE 3
341 
342 static int
343 cmp_to_prefetch(uchar_t *p)
344 {
345 	if (*p == 0x0F && (*(p+1) == 0x18 || *(p+1) == 0x0D) && *(p+2) <= 3)
346 		return (1);
347 	return (0);
348 }
349 
350 static int
351 instr_is_prefetch(caddr_t pc)
352 {
353 	uchar_t instr[PREFETCHSIZE];
354 	int	error;
355 
356 	error = copyin_nowatch(pc, (caddr_t)instr, PREFETCHSIZE);
357 
358 	if (error == 0 && cmp_to_prefetch(instr))
359 		return (1);
360 	return (0);
361 }
362 
363 #endif /* OPTERON_ERRATUM_91 */
364 
365 /*
366  * Called from the trap handler when a processor trap occurs.
367  *
368  * Note: All user-level traps that might call stop() must exit
369  * trap() by 'goto out' or by falling through.
370  */
371 void
372 trap(struct regs *rp, caddr_t addr, processorid_t cpuid)
373 {
374 	kthread_t *cur_thread = curthread;
375 	enum seg_rw rw;
376 	unsigned type;
377 	proc_t *p = ttoproc(cur_thread);
378 	klwp_t *lwp = ttolwp(cur_thread);
379 	uintptr_t lofault;
380 	faultcode_t pagefault(), res, errcode;
381 	enum fault_type fault_type;
382 	k_siginfo_t siginfo;
383 	uint_t fault = 0;
384 	int mstate;
385 	int sicode = 0;
386 	int watchcode;
387 	int watchpage;
388 	caddr_t vaddr;
389 	int singlestep_twiddle;
390 	size_t sz;
391 	int ta;
392 #ifdef __amd64
393 	uchar_t instr;
394 #endif
395 
396 	ASSERT_STACK_ALIGNED();
397 
398 	type = rp->r_trapno;
399 	CPU_STATS_ADDQ(CPU, sys, trap, 1);
400 
401 	ASSERT(cur_thread->t_schedflag & TS_DONT_SWAP);
402 
403 	if (type == T_PGFLT) {
404 
405 		errcode = rp->r_err;
406 		if (errcode & PF_ERR_WRITE)
407 			rw = S_WRITE;
408 		else if ((caddr_t)rp->r_pc == addr ||
409 		    (mmu.pt_nx != 0 && (errcode & PF_ERR_EXEC)))
410 			rw = S_EXEC;
411 		else
412 			rw = S_READ;
413 
414 #if defined(__i386)
415 		/*
416 		 * Pentium Pro work-around
417 		 */
418 		if ((errcode & PF_ERR_PROT) && pentiumpro_bug4046376) {
419 			uint_t	attr;
420 			uint_t	priv_violation;
421 			uint_t	access_violation;
422 
423 			if (hat_getattr(addr < (caddr_t)kernelbase ?
424 			    curproc->p_as->a_hat : kas.a_hat, addr, &attr)
425 			    == -1) {
426 				errcode &= ~PF_ERR_PROT;
427 			} else {
428 				priv_violation = (errcode & PF_ERR_USER) &&
429 					!(attr & PROT_USER);
430 				access_violation = (errcode & PF_ERR_WRITE) &&
431 					!(attr & PROT_WRITE);
432 				if (!priv_violation && !access_violation)
433 					goto cleanup;
434 			}
435 		}
436 #endif /* __i386 */
437 
438 	}
439 
440 	if (tdebug)
441 		showregs(type, rp, addr);
442 
443 	if (USERMODE(rp->r_cs)) {
444 		/*
445 		 * Set up the current cred to use during this trap. u_cred
446 		 * no longer exists.  t_cred is used instead.
447 		 * The current process credential applies to the thread for
448 		 * the entire trap.  If trapping from the kernel, this
449 		 * should already be set up.
450 		 */
451 		if (cur_thread->t_cred != p->p_cred) {
452 			cred_t *oldcred = cur_thread->t_cred;
453 			/*
454 			 * DTrace accesses t_cred in probe context.  t_cred
455 			 * must always be either NULL, or point to a valid,
456 			 * allocated cred structure.
457 			 */
458 			cur_thread->t_cred = crgetcred();
459 			crfree(oldcred);
460 		}
461 		ASSERT(lwp != NULL);
462 		type |= USER;
463 		ASSERT(lwptoregs(lwp) == rp);
464 		lwp->lwp_state = LWP_SYS;
465 
466 		switch (type) {
467 		case T_PGFLT + USER:
468 			if ((caddr_t)rp->r_pc == addr)
469 				mstate = LMS_TFAULT;
470 			else
471 				mstate = LMS_DFAULT;
472 			break;
473 		default:
474 			mstate = LMS_TRAP;
475 			break;
476 		}
477 		/* Kernel probe */
478 		TNF_PROBE_1(thread_state, "thread", /* CSTYLED */,
479 		    tnf_microstate, state, mstate);
480 		mstate = new_mstate(cur_thread, mstate);
481 
482 		bzero(&siginfo, sizeof (siginfo));
483 	}
484 
485 	switch (type) {
486 	case T_PGFLT + USER:
487 	case T_SGLSTP:
488 	case T_SGLSTP + USER:
489 	case T_BPTFLT + USER:
490 		break;
491 
492 	default:
493 		FTRACE_2("trap(): type=0x%lx, regs=0x%lx",
494 		    (ulong_t)type, (ulong_t)rp);
495 		break;
496 	}
497 
498 	switch (type) {
499 	default:
500 		if (type & USER) {
501 			if (tudebug)
502 				showregs(type, rp, (caddr_t)0);
503 			printf("trap: Unknown trap type %d in user mode\n",
504 			    type & ~USER);
505 			siginfo.si_signo = SIGILL;
506 			siginfo.si_code  = ILL_ILLTRP;
507 			siginfo.si_addr  = (caddr_t)rp->r_pc;
508 			siginfo.si_trapno = type & ~USER;
509 			fault = FLTILL;
510 			break;
511 		} else {
512 			(void) die(type, rp, addr, cpuid);
513 			/*NOTREACHED*/
514 		}
515 
516 	case T_PGFLT:		/* system page fault */
517 		/*
518 		 * If we're under on_trap() protection (see <sys/ontrap.h>),
519 		 * set ot_trap and longjmp back to the on_trap() call site.
520 		 */
521 		if ((cur_thread->t_ontrap != NULL) &&
522 		    (cur_thread->t_ontrap->ot_prot & OT_DATA_ACCESS)) {
523 			curthread->t_ontrap->ot_trap |= OT_DATA_ACCESS;
524 			longjmp(&curthread->t_ontrap->ot_jmpbuf);
525 		}
526 
527 		/*
528 		 * See if we can handle as pagefault. Save lofault
529 		 * across this. Here we assume that an address
530 		 * less than KERNELBASE is a user fault.
531 		 * We can do this as copy.s routines verify that the
532 		 * starting address is less than KERNELBASE before
533 		 * starting and because we know that we always have
534 		 * KERNELBASE mapped as invalid to serve as a "barrier".
535 		 */
536 		lofault = cur_thread->t_lofault;
537 		cur_thread->t_lofault = 0;
538 
539 		mstate = new_mstate(cur_thread, LMS_KFAULT);
540 
541 		if (addr < (caddr_t)kernelbase) {
542 			res = pagefault(addr,
543 			    (errcode & PF_ERR_PROT)? F_PROT: F_INVAL, rw, 0);
544 			if (res == FC_NOMAP &&
545 			    addr < p->p_usrstack &&
546 			    grow(addr))
547 				res = 0;
548 		} else {
549 			res = pagefault(addr,
550 			    (errcode & PF_ERR_PROT)? F_PROT: F_INVAL, rw, 1);
551 		}
552 		(void) new_mstate(cur_thread, mstate);
553 
554 		/*
555 		 * Restore lofault. If we resolved the fault, exit.
556 		 * If we didn't and lofault wasn't set, die.
557 		 */
558 		cur_thread->t_lofault = lofault;
559 		if (res == 0)
560 			goto cleanup;
561 
562 #if defined(OPTERON_ERRATUM_93) && defined(_LP64)
563 		if (lofault == 0 && opteron_erratum_93) {
564 			/*
565 			 * Workaround for Opteron Erratum 93. On return from
566 			 * a System Managment Interrupt at a HLT instruction
567 			 * the %rip might be truncated to a 32 bit value.
568 			 * BIOS is supposed to fix this, but some don't.
569 			 * If this occurs we simply restore the high order bits.
570 			 * The HLT instruction is 1 byte of 0xf4.
571 			 */
572 			uintptr_t	rip = rp->r_pc;
573 
574 			if ((rip & 0xfffffffful) == rip) {
575 				rip |= 0xfffffffful << 32;
576 				if (hat_getpfnum(kas.a_hat, (caddr_t)rip) !=
577 				    PFN_INVALID &&
578 				    (*(uchar_t *)rip == 0xf4 ||
579 				    *(uchar_t *)(rip - 1) == 0xf4)) {
580 					rp->r_pc = rip;
581 					goto cleanup;
582 				}
583 			}
584 		}
585 #endif /* OPTERON_ERRATUM_93 && _LP64 */
586 
587 #ifdef OPTERON_ERRATUM_91
588 		if (lofault == 0 && opteron_erratum_91) {
589 			/*
590 			 * Workaround for Opteron Erratum 91. Prefetches may
591 			 * generate a page fault (they're not supposed to do
592 			 * that!). If this occurs we simply return back to the
593 			 * instruction.
594 			 */
595 			caddr_t		pc = (caddr_t)rp->r_pc;
596 
597 			/*
598 			 * If the faulting PC is not mapped, this is a
599 			 * legitimate kernel page fault that must result in a
600 			 * panic. If the faulting PC is mapped, it could contain
601 			 * a prefetch instruction. Check for that here.
602 			 */
603 			if (hat_getpfnum(kas.a_hat, pc) != PFN_INVALID) {
604 				if (cmp_to_prefetch((uchar_t *)pc)) {
605 #ifdef DEBUG
606 					cmn_err(CE_WARN, "Opteron erratum 91 "
607 					    "occurred: kernel prefetch"
608 					    " at %p generated a page fault!",
609 					    (void *)rp->r_pc);
610 #endif /* DEBUG */
611 					goto cleanup;
612 				}
613 			}
614 			(void) die(type, rp, addr, cpuid);
615 		}
616 #endif /* OPTERON_ERRATUM_91 */
617 
618 		if (lofault == 0)
619 			(void) die(type, rp, addr, cpuid);
620 
621 		/*
622 		 * Cannot resolve fault.  Return to lofault.
623 		 */
624 		if (lodebug) {
625 			showregs(type, rp, addr);
626 			traceregs(rp);
627 		}
628 		if (FC_CODE(res) == FC_OBJERR)
629 			res = FC_ERRNO(res);
630 		else
631 			res = EFAULT;
632 		rp->r_r0 = res;
633 		rp->r_pc = cur_thread->t_lofault;
634 		goto cleanup;
635 
636 	case T_PGFLT + USER:	/* user page fault */
637 		if (faultdebug) {
638 			char *fault_str;
639 
640 			switch (rw) {
641 			case S_READ:
642 				fault_str = "read";
643 				break;
644 			case S_WRITE:
645 				fault_str = "write";
646 				break;
647 			case S_EXEC:
648 				fault_str = "exec";
649 				break;
650 			default:
651 				fault_str = "";
652 				break;
653 			}
654 			printf("user %s fault:  addr=0x%lx errcode=0x%x\n",
655 			    fault_str, (uintptr_t)addr, errcode);
656 		}
657 
658 #if defined(OPTERON_ERRATUM_100) && defined(_LP64)
659 		/*
660 		 * Workaround for AMD erratum 100
661 		 *
662 		 * A 32-bit process may receive a page fault on a non
663 		 * 32-bit address by mistake. The range of the faulting
664 		 * address will be
665 		 *
666 		 *	0xffffffff80000000 .. 0xffffffffffffffff or
667 		 *	0x0000000100000000 .. 0x000000017fffffff
668 		 *
669 		 * The fault is always due to an instruction fetch, however
670 		 * the value of r_pc should be correct (in 32 bit range),
671 		 * so we ignore the page fault on the bogus address.
672 		 */
673 		if (p->p_model == DATAMODEL_ILP32 &&
674 		    (0xffffffff80000000 <= (uintptr_t)addr ||
675 		    (0x100000000 <= (uintptr_t)addr &&
676 		    (uintptr_t)addr <= 0x17fffffff))) {
677 			if (!opteron_erratum_100)
678 				panic("unexpected erratum #100");
679 			if (rp->r_pc <= 0xffffffff)
680 				goto out;
681 		}
682 #endif /* OPTERON_ERRATUM_100 && _LP64 */
683 
684 		ASSERT(!(curthread->t_flag & T_WATCHPT));
685 		watchpage = (pr_watch_active(p) && pr_is_watchpage(addr, rw));
686 #ifdef __i386
687 		/*
688 		 * In 32-bit mode, the lcall (system call) instruction fetches
689 		 * one word from the stack, at the stack pointer, because of the
690 		 * way the call gate is constructed.  This is a bogus
691 		 * read and should not be counted as a read watchpoint.
692 		 * We work around the problem here by testing to see if
693 		 * this situation applies and, if so, simply jumping to
694 		 * the code in locore.s that fields the system call trap.
695 		 * The registers on the stack are already set up properly
696 		 * due to the match between the call gate sequence and the
697 		 * trap gate sequence.  We just have to adjust the pc.
698 		 */
699 		if (watchpage && addr == (caddr_t)rp->r_sp &&
700 		    rw == S_READ && instr_is_syscall((caddr_t)rp->r_pc)) {
701 			extern void watch_syscall(void);
702 
703 			rp->r_pc += LCALLSIZE;
704 			watch_syscall();	/* never returns */
705 			/* NOTREACHED */
706 		}
707 #endif /* __i386 */
708 		vaddr = addr;
709 		if (!watchpage || (sz = instr_size(rp, &vaddr, rw)) <= 0)
710 			fault_type = (errcode & PF_ERR_PROT)? F_PROT: F_INVAL;
711 		else if ((watchcode = pr_is_watchpoint(&vaddr, &ta,
712 		    sz, NULL, rw)) != 0) {
713 			if (ta) {
714 				do_watch_step(vaddr, sz, rw,
715 					watchcode, rp->r_pc);
716 				fault_type = F_INVAL;
717 			} else {
718 				bzero(&siginfo, sizeof (siginfo));
719 				siginfo.si_signo = SIGTRAP;
720 				siginfo.si_code = watchcode;
721 				siginfo.si_addr = vaddr;
722 				siginfo.si_trapafter = 0;
723 				siginfo.si_pc = (caddr_t)rp->r_pc;
724 				fault = FLTWATCH;
725 				break;
726 			}
727 		} else {
728 			/* XXX pr_watch_emul() never succeeds (for now) */
729 			if (rw != S_EXEC && pr_watch_emul(rp, vaddr, rw))
730 				goto out;
731 			do_watch_step(vaddr, sz, rw, 0, 0);
732 			fault_type = F_INVAL;
733 		}
734 
735 		res = pagefault(addr, fault_type, rw, 0);
736 
737 		/*
738 		 * If pagefault() succeeded, ok.
739 		 * Otherwise attempt to grow the stack.
740 		 */
741 		if (res == 0 ||
742 		    (res == FC_NOMAP &&
743 		    addr < p->p_usrstack &&
744 		    grow(addr))) {
745 			lwp->lwp_lastfault = FLTPAGE;
746 			lwp->lwp_lastfaddr = addr;
747 			if (prismember(&p->p_fltmask, FLTPAGE)) {
748 				bzero(&siginfo, sizeof (siginfo));
749 				siginfo.si_addr = addr;
750 				(void) stop_on_fault(FLTPAGE, &siginfo);
751 			}
752 			goto out;
753 		} else if (res == FC_PROT && addr < p->p_usrstack &&
754 		    (mmu.pt_nx != 0 && (errcode & PF_ERR_EXEC))) {
755 			report_stack_exec(p, addr);
756 		}
757 
758 #ifdef OPTERON_ERRATUM_91
759 		/*
760 		 * Workaround for Opteron Erratum 91. Prefetches may generate a
761 		 * page fault (they're not supposed to do that!). If this
762 		 * occurs we simply return back to the instruction.
763 		 *
764 		 * We rely on copyin to properly fault in the page with r_pc.
765 		 */
766 		if (opteron_erratum_91 &&
767 		    addr != (caddr_t)rp->r_pc &&
768 		    instr_is_prefetch((caddr_t)rp->r_pc)) {
769 #ifdef DEBUG
770 			cmn_err(CE_WARN, "Opteron erratum 91 occurred: "
771 			    "prefetch at %p in pid %d generated a trap!",
772 			    (void *)rp->r_pc, p->p_pid);
773 #endif /* DEBUG */
774 			goto out;
775 		}
776 #endif /* OPTERON_ERRATUM_91 */
777 
778 		if (tudebug)
779 			showregs(type, rp, addr);
780 		/*
781 		 * In the case where both pagefault and grow fail,
782 		 * set the code to the value provided by pagefault.
783 		 * We map all errors returned from pagefault() to SIGSEGV.
784 		 */
785 		bzero(&siginfo, sizeof (siginfo));
786 		siginfo.si_addr = addr;
787 		switch (FC_CODE(res)) {
788 		case FC_HWERR:
789 		case FC_NOSUPPORT:
790 			siginfo.si_signo = SIGBUS;
791 			siginfo.si_code = BUS_ADRERR;
792 			fault = FLTACCESS;
793 			break;
794 		case FC_ALIGN:
795 			siginfo.si_signo = SIGBUS;
796 			siginfo.si_code = BUS_ADRALN;
797 			fault = FLTACCESS;
798 			break;
799 		case FC_OBJERR:
800 			if ((siginfo.si_errno = FC_ERRNO(res)) != EINTR) {
801 				siginfo.si_signo = SIGBUS;
802 				siginfo.si_code = BUS_OBJERR;
803 				fault = FLTACCESS;
804 			}
805 			break;
806 		default:	/* FC_NOMAP or FC_PROT */
807 			siginfo.si_signo = SIGSEGV;
808 			siginfo.si_code =
809 			    (res == FC_NOMAP)? SEGV_MAPERR : SEGV_ACCERR;
810 			fault = FLTBOUNDS;
811 			break;
812 		}
813 		break;
814 
815 	case T_ILLINST + USER:	/* invalid opcode fault */
816 		/*
817 		 * If the syscall instruction is disabled due to LDT usage, a
818 		 * user program that attempts to execute it will trigger a #ud
819 		 * trap. Check for that case here. If this occurs on a CPU which
820 		 * doesn't even support syscall, the result of all of this will
821 		 * be to emulate that particular instruction.
822 		 */
823 		if (p->p_ldt != NULL &&
824 		    instr_is_fast_syscall((caddr_t)rp->r_pc, X86_ASYSC)) {
825 			if (rewrite_syscall((caddr_t)rp->r_pc) == 0)
826 				goto out;
827 #ifdef DEBUG
828 			else
829 				cmn_err(CE_WARN, "failed to rewrite syscall "
830 				    "instruction in process %d",
831 				    curthread->t_procp->p_pid);
832 #endif /* DEBUG */
833 		}
834 
835 #ifdef __amd64
836 		/*
837 		 * Emulate the LAHF and SAHF instructions if needed.
838 		 * See the instr_is_lsahf function for details.
839 		 */
840 		if (p->p_model == DATAMODEL_LP64 &&
841 		    instr_is_lsahf((caddr_t)rp->r_pc, &instr)) {
842 			emulate_lsahf(rp, instr);
843 			goto out;
844 		}
845 #endif
846 
847 		/*FALLTHROUGH*/
848 
849 		if (tudebug)
850 			showregs(type, rp, (caddr_t)0);
851 		siginfo.si_signo = SIGILL;
852 		siginfo.si_code  = ILL_ILLOPC;
853 		siginfo.si_addr  = (caddr_t)rp->r_pc;
854 		fault = FLTILL;
855 		break;
856 
857 	case T_ZERODIV + USER:		/* integer divide by zero */
858 		if (tudebug && tudebugfpe)
859 			showregs(type, rp, (caddr_t)0);
860 		siginfo.si_signo = SIGFPE;
861 		siginfo.si_code  = FPE_INTDIV;
862 		siginfo.si_addr  = (caddr_t)rp->r_pc;
863 		fault = FLTIZDIV;
864 		break;
865 
866 	case T_OVFLW + USER:	/* integer overflow */
867 		if (tudebug && tudebugfpe)
868 			showregs(type, rp, (caddr_t)0);
869 		siginfo.si_signo = SIGFPE;
870 		siginfo.si_code  = FPE_INTOVF;
871 		siginfo.si_addr  = (caddr_t)rp->r_pc;
872 		fault = FLTIOVF;
873 		break;
874 
875 	case T_NOEXTFLT + USER:	/* math coprocessor not available */
876 		if (tudebug && tudebugfpe)
877 			showregs(type, rp, addr);
878 		if (fpnoextflt(rp)) {
879 			siginfo.si_signo = SIGFPE;
880 			siginfo.si_code  = ILL_ILLOPC;
881 			siginfo.si_addr  = (caddr_t)rp->r_pc;
882 			fault = FLTFPE;
883 		}
884 		break;
885 
886 	case T_EXTOVRFLT:	/* extension overrun fault */
887 		/* check if we took a kernel trap on behalf of user */
888 		{
889 			extern  void ndptrap_frstor(void);
890 			if (rp->r_pc != (uintptr_t)ndptrap_frstor)
891 				(void) die(type, rp, addr, cpuid);
892 			type |= USER;
893 		}
894 		/*FALLTHROUGH*/
895 	case T_EXTOVRFLT + USER:	/* extension overrun fault */
896 		if (tudebug && tudebugfpe)
897 			showregs(type, rp, addr);
898 		if (fpextovrflt(rp)) {
899 			siginfo.si_signo = SIGSEGV;
900 			siginfo.si_code  = SEGV_MAPERR;
901 			siginfo.si_addr  = (caddr_t)rp->r_pc;
902 			fault = FLTBOUNDS;
903 		}
904 		break;
905 
906 	case T_EXTERRFLT:	/* x87 floating point exception pending */
907 		/* check if we took a kernel trap on behalf of user */
908 		{
909 			extern  void ndptrap_frstor(void);
910 			if (rp->r_pc != (uintptr_t)ndptrap_frstor)
911 				(void) die(type, rp, addr, cpuid);
912 			type |= USER;
913 		}
914 		/*FALLTHROUGH*/
915 
916 	case T_EXTERRFLT + USER: /* x87 floating point exception pending */
917 		if (tudebug && tudebugfpe)
918 			showregs(type, rp, addr);
919 		if (sicode = fpexterrflt(rp)) {
920 			siginfo.si_signo = SIGFPE;
921 			siginfo.si_code  = sicode;
922 			siginfo.si_addr  = (caddr_t)rp->r_pc;
923 			fault = FLTFPE;
924 		}
925 		break;
926 
927 	case T_SIMDFPE + USER:		/* SSE and SSE2 exceptions */
928 		if (tudebug && tudebugsse)
929 			showregs(type, rp, addr);
930 		if ((x86_feature & (X86_SSE|X86_SSE2)) == 0) {
931 			/*
932 			 * There are rumours that some user instructions
933 			 * on older CPUs can cause this trap to occur; in
934 			 * which case send a SIGILL instead of a SIGFPE.
935 			 */
936 			siginfo.si_signo = SIGILL;
937 			siginfo.si_code  = ILL_ILLTRP;
938 			siginfo.si_addr  = (caddr_t)rp->r_pc;
939 			siginfo.si_trapno = type & ~USER;
940 			fault = FLTILL;
941 		} else if ((sicode = fpsimderrflt(rp)) != 0) {
942 			siginfo.si_signo = SIGFPE;
943 			siginfo.si_code = sicode;
944 			siginfo.si_addr = (caddr_t)rp->r_pc;
945 			fault = FLTFPE;
946 		}
947 		break;
948 
949 	case T_BPTFLT:	/* breakpoint trap */
950 		/*
951 		 * Kernel breakpoint traps should only happen when kmdb is
952 		 * active, and even then, it'll have interposed on the IDT, so
953 		 * control won't get here.  If it does, we've hit a breakpoint
954 		 * without the debugger, which is very strange, and very
955 		 * fatal.
956 		 */
957 		if (tudebug && tudebugbpt)
958 			showregs(type, rp, (caddr_t)0);
959 
960 		(void) die(type, rp, addr, cpuid);
961 		break;
962 
963 	case T_SGLSTP: /* single step/hw breakpoint exception */
964 
965 		/* Now evaluate how we got here */
966 		if (lwp != NULL && (lwp->lwp_pcb.pcb_drstat & DR_SINGLESTEP)) {
967 			/*
968 			 * i386 single-steps even through lcalls which
969 			 * change the privilege level. So we take a trap at
970 			 * the first instruction in privileged mode.
971 			 *
972 			 * Set a flag to indicate that upon completion of
973 			 * the system call, deal with the single-step trap.
974 			 *
975 			 * The same thing happens for sysenter, too.
976 			 */
977 			singlestep_twiddle = 0;
978 			if (rp->r_pc == (uintptr_t)sys_sysenter ||
979 			    rp->r_pc == (uintptr_t)brand_sys_sysenter) {
980 				singlestep_twiddle = 1;
981 #if defined(__amd64)
982 				/*
983 				 * Since we are already on the kernel's
984 				 * %gs, on 64-bit systems the sysenter case
985 				 * needs to adjust the pc to avoid
986 				 * executing the swapgs instruction at the
987 				 * top of the handler.
988 				 */
989 				if (rp->r_pc == (uintptr_t)sys_sysenter)
990 					rp->r_pc = (uintptr_t)
991 					    _sys_sysenter_post_swapgs;
992 				else
993 					rp->r_pc = (uintptr_t)
994 					    _brand_sys_sysenter_post_swapgs;
995 #endif
996 			}
997 #if defined(__i386)
998 			else if (rp->r_pc == (uintptr_t)sys_call ||
999 			    rp->r_pc == (uintptr_t)brand_sys_call) {
1000 				singlestep_twiddle = 1;
1001 			}
1002 #endif
1003 			else {
1004 				/* not on sysenter/syscall; uregs available */
1005 				if (tudebug && tudebugbpt)
1006 					showregs(type, rp, (caddr_t)0);
1007 			}
1008 			if (singlestep_twiddle) {
1009 				rp->r_ps &= ~PS_T; /* turn off trace */
1010 				lwp->lwp_pcb.pcb_flags |= DEBUG_PENDING;
1011 				cur_thread->t_post_sys = 1;
1012 				aston(curthread);
1013 				goto cleanup;
1014 			}
1015 		}
1016 		/* XXX - needs review on debugger interface? */
1017 		if (boothowto & RB_DEBUG)
1018 			debug_enter((char *)NULL);
1019 		else
1020 			(void) die(type, rp, addr, cpuid);
1021 		break;
1022 
1023 	case T_NMIFLT:	/* NMI interrupt */
1024 		printf("Unexpected NMI in system mode\n");
1025 		goto cleanup;
1026 
1027 	case T_NMIFLT + USER:	/* NMI interrupt */
1028 		printf("Unexpected NMI in user mode\n");
1029 		break;
1030 
1031 	case T_GPFLT:	/* general protection violation */
1032 #if defined(__amd64)
1033 		/*
1034 		 * On amd64, we can get a #gp from referencing addresses
1035 		 * in the virtual address hole e.g. from a copyin.
1036 		 */
1037 
1038 		/*
1039 		 * If we're under on_trap() protection (see <sys/ontrap.h>),
1040 		 * set ot_trap and longjmp back to the on_trap() call site.
1041 		 */
1042 		if ((cur_thread->t_ontrap != NULL) &&
1043 		    (cur_thread->t_ontrap->ot_prot & OT_DATA_ACCESS)) {
1044 			curthread->t_ontrap->ot_trap |= OT_DATA_ACCESS;
1045 			longjmp(&curthread->t_ontrap->ot_jmpbuf);
1046 		}
1047 
1048 		/*
1049 		 * If we're under lofault protection (copyin etc.),
1050 		 * longjmp back to lofault with an EFAULT.
1051 		 */
1052 		if (cur_thread->t_lofault) {
1053 			/*
1054 			 * Fault is not resolvable, so just return to lofault
1055 			 */
1056 			if (lodebug) {
1057 				showregs(type, rp, addr);
1058 				traceregs(rp);
1059 			}
1060 			rp->r_r0 = EFAULT;
1061 			rp->r_pc = cur_thread->t_lofault;
1062 			goto cleanup;
1063 		}
1064 		/*FALLTHROUGH*/
1065 #endif
1066 	case T_STKFLT:	/* stack fault */
1067 	case T_TSSFLT:	/* invalid TSS fault */
1068 	case T_SEGFLT:	/* segment not present fault */
1069 		if (tudebug)
1070 			showregs(type, rp, (caddr_t)0);
1071 		if (kern_gpfault(rp))
1072 			(void) die(type, rp, addr, cpuid);
1073 		goto cleanup;
1074 		/*FALLTHROUGH*/
1075 
1076 /*
1077  * ONLY 32-bit PROCESSES can USE a PRIVATE LDT! 64-bit apps should have
1078  * no legacy need for them, so we put a stop to it here.
1079  *
1080  * So: not-present fault is ONLY valid for 32-bit processes with a private LDT
1081  * trying to do a system call. Emulate it.
1082  *
1083  * #gp fault is ONLY valid for 32-bit processes also, which DO NOT have private
1084  * LDT, and are trying to do a system call. Emulate it.
1085  */
1086 	case T_SEGFLT + USER:	/* segment not present fault */
1087 	case T_GPFLT + USER:	/* general protection violation */
1088 #ifdef _SYSCALL32_IMPL
1089 		if (p->p_model != DATAMODEL_NATIVE) {
1090 #endif /* _SYSCALL32_IMPL */
1091 		if (instr_is_syscall((caddr_t)rp->r_pc)) {
1092 			if (type == T_SEGFLT + USER)
1093 				ASSERT(p->p_ldt != NULL);
1094 
1095 			if ((p->p_ldt == NULL && type == T_GPFLT + USER) ||
1096 			    type == T_SEGFLT + USER) {
1097 
1098 			/*
1099 			 * The user attempted a system call via the obsolete
1100 			 * call gate mechanism. Because the process doesn't have
1101 			 * an LDT (i.e. the ldtr contains 0), a #gp results.
1102 			 * Emulate the syscall here, just as we do above for a
1103 			 * #np trap.
1104 			 */
1105 
1106 			/*
1107 			 * Since this is a not-present trap, rp->r_pc points to
1108 			 * the trapping lcall instruction. We need to bump it
1109 			 * to the next insn so the app can continue on.
1110 			 */
1111 			rp->r_pc += LCALLSIZE;
1112 			lwp->lwp_regs = rp;
1113 
1114 			/*
1115 			 * Normally the microstate of the LWP is forced back to
1116 			 * LMS_USER by the syscall handlers. Emulate that
1117 			 * behavior here.
1118 			 */
1119 			mstate = LMS_USER;
1120 
1121 			dosyscall();
1122 			goto out;
1123 			}
1124 		}
1125 #ifdef _SYSCALL32_IMPL
1126 		}
1127 #endif /* _SYSCALL32_IMPL */
1128 		/*
1129 		 * If the current process is using a private LDT and the
1130 		 * trapping instruction is sysenter, the sysenter instruction
1131 		 * has been disabled on the CPU because it destroys segment
1132 		 * registers. If this is the case, rewrite the instruction to
1133 		 * be a safe system call and retry it. If this occurs on a CPU
1134 		 * which doesn't even support sysenter, the result of all of
1135 		 * this will be to emulate that particular instruction.
1136 		 */
1137 		if (p->p_ldt != NULL &&
1138 		    instr_is_fast_syscall((caddr_t)rp->r_pc, X86_SEP)) {
1139 			if (rewrite_syscall((caddr_t)rp->r_pc) == 0)
1140 				goto out;
1141 #ifdef DEBUG
1142 			else
1143 				cmn_err(CE_WARN, "failed to rewrite sysenter "
1144 				    "instruction in process %d",
1145 				    curthread->t_procp->p_pid);
1146 #endif /* DEBUG */
1147 		}
1148 		/*FALLTHROUGH*/
1149 
1150 	case T_BOUNDFLT + USER:	/* bound fault */
1151 	case T_STKFLT + USER:	/* stack fault */
1152 	case T_TSSFLT + USER:	/* invalid TSS fault */
1153 		if (tudebug)
1154 			showregs(type, rp, (caddr_t)0);
1155 		siginfo.si_signo = SIGSEGV;
1156 		siginfo.si_code  = SEGV_MAPERR;
1157 		siginfo.si_addr  = (caddr_t)rp->r_pc;
1158 		fault = FLTBOUNDS;
1159 		break;
1160 
1161 	case T_ALIGNMENT + USER:	/* user alignment error (486) */
1162 		if (tudebug)
1163 			showregs(type, rp, (caddr_t)0);
1164 		bzero(&siginfo, sizeof (siginfo));
1165 		siginfo.si_signo = SIGBUS;
1166 		siginfo.si_code = BUS_ADRALN;
1167 		siginfo.si_addr = (caddr_t)rp->r_pc;
1168 		fault = FLTACCESS;
1169 		break;
1170 
1171 	case T_SGLSTP + USER: /* single step/hw breakpoint exception */
1172 		if (tudebug && tudebugbpt)
1173 			showregs(type, rp, (caddr_t)0);
1174 
1175 		/* Was it single-stepping? */
1176 		if (lwp->lwp_pcb.pcb_drstat & DR_SINGLESTEP) {
1177 			pcb_t *pcb = &lwp->lwp_pcb;
1178 
1179 			rp->r_ps &= ~PS_T;
1180 			/*
1181 			 * If both NORMAL_STEP and WATCH_STEP are in effect,
1182 			 * give precedence to WATCH_STEP.  If neither is set,
1183 			 * user must have set the PS_T bit in %efl; treat this
1184 			 * as NORMAL_STEP.
1185 			 */
1186 			if ((fault = undo_watch_step(&siginfo)) == 0 &&
1187 			    ((pcb->pcb_flags & NORMAL_STEP) ||
1188 			    !(pcb->pcb_flags & WATCH_STEP))) {
1189 				siginfo.si_signo = SIGTRAP;
1190 				siginfo.si_code = TRAP_TRACE;
1191 				siginfo.si_addr = (caddr_t)rp->r_pc;
1192 				fault = FLTTRACE;
1193 			}
1194 			pcb->pcb_flags &= ~(NORMAL_STEP|WATCH_STEP);
1195 		} else {
1196 			cmn_err(CE_WARN,
1197 			    "Unexpected INT 1 in user mode, dr6=%lx",
1198 			    lwp->lwp_pcb.pcb_drstat);
1199 		}
1200 		break;
1201 
1202 	case T_BPTFLT + USER:	/* breakpoint trap */
1203 		if (tudebug && tudebugbpt)
1204 			showregs(type, rp, (caddr_t)0);
1205 		/*
1206 		 * int 3 (the breakpoint instruction) leaves the pc referring
1207 		 * to the address one byte after the breakpointed address.
1208 		 * If the P_PR_BPTADJ flag has been set via /proc, We adjust
1209 		 * it back so it refers to the breakpointed address.
1210 		 */
1211 		if (p->p_proc_flag & P_PR_BPTADJ)
1212 			rp->r_pc--;
1213 		siginfo.si_signo = SIGTRAP;
1214 		siginfo.si_code  = TRAP_BRKPT;
1215 		siginfo.si_addr  = (caddr_t)rp->r_pc;
1216 		fault = FLTBPT;
1217 		break;
1218 
1219 	case T_AST:
1220 		/*
1221 		 * This occurs only after the cs register has been made to
1222 		 * look like a kernel selector, either through debugging or
1223 		 * possibly by functions like setcontext().  The thread is
1224 		 * about to cause a general protection fault at common_iret()
1225 		 * in locore.  We let that happen immediately instead of
1226 		 * doing the T_AST processing.
1227 		 */
1228 		goto cleanup;
1229 
1230 	case T_AST + USER:		/* profiling or resched pseudo trap */
1231 		if (lwp->lwp_pcb.pcb_flags & CPC_OVERFLOW) {
1232 			lwp->lwp_pcb.pcb_flags &= ~CPC_OVERFLOW;
1233 			if (kcpc_overflow_ast()) {
1234 				/*
1235 				 * Signal performance counter overflow
1236 				 */
1237 				if (tudebug)
1238 					showregs(type, rp, (caddr_t)0);
1239 				bzero(&siginfo, sizeof (siginfo));
1240 				siginfo.si_signo = SIGEMT;
1241 				siginfo.si_code = EMT_CPCOVF;
1242 				siginfo.si_addr = (caddr_t)rp->r_pc;
1243 				fault = FLTCPCOVF;
1244 			}
1245 		}
1246 		break;
1247 	}
1248 
1249 	/*
1250 	 * We can't get here from a system trap
1251 	 */
1252 	ASSERT(type & USER);
1253 
1254 	if (fault) {
1255 		/*
1256 		 * Remember the fault and fault adddress
1257 		 * for real-time (SIGPROF) profiling.
1258 		 */
1259 		lwp->lwp_lastfault = fault;
1260 		lwp->lwp_lastfaddr = siginfo.si_addr;
1261 
1262 		DTRACE_PROC2(fault, int, fault, ksiginfo_t *, &siginfo);
1263 
1264 		/*
1265 		 * If a debugger has declared this fault to be an
1266 		 * event of interest, stop the lwp.  Otherwise just
1267 		 * deliver the associated signal.
1268 		 */
1269 		if (siginfo.si_signo != SIGKILL &&
1270 		    prismember(&p->p_fltmask, fault) &&
1271 		    stop_on_fault(fault, &siginfo) == 0)
1272 			siginfo.si_signo = 0;
1273 	}
1274 
1275 	if (siginfo.si_signo)
1276 		trapsig(&siginfo, (fault == FLTCPCOVF)? 0 : 1);
1277 
1278 	if (lwp->lwp_oweupc)
1279 		profil_tick(rp->r_pc);
1280 
1281 	if (cur_thread->t_astflag | cur_thread->t_sig_check) {
1282 		/*
1283 		 * Turn off the AST flag before checking all the conditions that
1284 		 * may have caused an AST.  This flag is on whenever a signal or
1285 		 * unusual condition should be handled after the next trap or
1286 		 * syscall.
1287 		 */
1288 		astoff(cur_thread);
1289 		/*
1290 		 * If a single-step trap occurred on a syscall (see above)
1291 		 * recognize it now.  Do this before checking for signals
1292 		 * because deferred_singlestep_trap() may generate a SIGTRAP to
1293 		 * the LWP or may otherwise mark the LWP to call issig(FORREAL).
1294 		 */
1295 		if (lwp->lwp_pcb.pcb_flags & DEBUG_PENDING)
1296 			deferred_singlestep_trap((caddr_t)rp->r_pc);
1297 
1298 		cur_thread->t_sig_check = 0;
1299 
1300 		mutex_enter(&p->p_lock);
1301 		if (curthread->t_proc_flag & TP_CHANGEBIND) {
1302 			timer_lwpbind();
1303 			curthread->t_proc_flag &= ~TP_CHANGEBIND;
1304 		}
1305 		mutex_exit(&p->p_lock);
1306 
1307 		/*
1308 		 * for kaio requests that are on the per-process poll queue,
1309 		 * aiop->aio_pollq, they're AIO_POLL bit is set, the kernel
1310 		 * should copyout their result_t to user memory. by copying
1311 		 * out the result_t, the user can poll on memory waiting
1312 		 * for the kaio request to complete.
1313 		 */
1314 		if (p->p_aio)
1315 			aio_cleanup(0);
1316 		/*
1317 		 * If this LWP was asked to hold, call holdlwp(), which will
1318 		 * stop.  holdlwps() sets this up and calls pokelwps() which
1319 		 * sets the AST flag.
1320 		 *
1321 		 * Also check TP_EXITLWP, since this is used by fresh new LWPs
1322 		 * through lwp_rtt().  That flag is set if the lwp_create(2)
1323 		 * syscall failed after creating the LWP.
1324 		 */
1325 		if (ISHOLD(p))
1326 			holdlwp();
1327 
1328 		/*
1329 		 * All code that sets signals and makes ISSIG evaluate true must
1330 		 * set t_astflag afterwards.
1331 		 */
1332 		if (ISSIG_PENDING(cur_thread, lwp, p)) {
1333 			if (issig(FORREAL))
1334 				psig();
1335 			cur_thread->t_sig_check = 1;
1336 		}
1337 
1338 		if (cur_thread->t_rprof != NULL) {
1339 			realsigprof(0, 0);
1340 			cur_thread->t_sig_check = 1;
1341 		}
1342 
1343 		/*
1344 		 * /proc can't enable/disable the trace bit itself
1345 		 * because that could race with the call gate used by
1346 		 * system calls via "lcall". If that happened, an
1347 		 * invalid EFLAGS would result. prstep()/prnostep()
1348 		 * therefore schedule an AST for the purpose.
1349 		 */
1350 		if (lwp->lwp_pcb.pcb_flags & REQUEST_STEP) {
1351 			lwp->lwp_pcb.pcb_flags &= ~REQUEST_STEP;
1352 			rp->r_ps |= PS_T;
1353 		}
1354 		if (lwp->lwp_pcb.pcb_flags & REQUEST_NOSTEP) {
1355 			lwp->lwp_pcb.pcb_flags &= ~REQUEST_NOSTEP;
1356 			rp->r_ps &= ~PS_T;
1357 		}
1358 	}
1359 
1360 out:	/* We can't get here from a system trap */
1361 	ASSERT(type & USER);
1362 
1363 	if (ISHOLD(p))
1364 		holdlwp();
1365 
1366 	/*
1367 	 * Set state to LWP_USER here so preempt won't give us a kernel
1368 	 * priority if it occurs after this point.  Call CL_TRAPRET() to
1369 	 * restore the user-level priority.
1370 	 *
1371 	 * It is important that no locks (other than spinlocks) be entered
1372 	 * after this point before returning to user mode (unless lwp_state
1373 	 * is set back to LWP_SYS).
1374 	 */
1375 	lwp->lwp_state = LWP_USER;
1376 
1377 	if (cur_thread->t_trapret) {
1378 		cur_thread->t_trapret = 0;
1379 		thread_lock(cur_thread);
1380 		CL_TRAPRET(cur_thread);
1381 		thread_unlock(cur_thread);
1382 	}
1383 	if (CPU->cpu_runrun)
1384 		preempt();
1385 	(void) new_mstate(cur_thread, mstate);
1386 
1387 	/* Kernel probe */
1388 	TNF_PROBE_1(thread_state, "thread", /* CSTYLED */,
1389 	    tnf_microstate, state, LMS_USER);
1390 
1391 	return;
1392 
1393 cleanup:	/* system traps end up here */
1394 	ASSERT(!(type & USER));
1395 }
1396 
1397 /*
1398  * Patch non-zero to disable preemption of threads in the kernel.
1399  */
1400 int IGNORE_KERNEL_PREEMPTION = 0;	/* XXX - delete this someday */
1401 
1402 struct kpreempt_cnts {		/* kernel preemption statistics */
1403 	int	kpc_idle;	/* executing idle thread */
1404 	int	kpc_intr;	/* executing interrupt thread */
1405 	int	kpc_clock;	/* executing clock thread */
1406 	int	kpc_blocked;	/* thread has blocked preemption (t_preempt) */
1407 	int	kpc_notonproc;	/* thread is surrendering processor */
1408 	int	kpc_inswtch;	/* thread has ratified scheduling decision */
1409 	int	kpc_prilevel;	/* processor interrupt level is too high */
1410 	int	kpc_apreempt;	/* asynchronous preemption */
1411 	int	kpc_spreempt;	/* synchronous preemption */
1412 } kpreempt_cnts;
1413 
1414 /*
1415  * kernel preemption: forced rescheduling, preempt the running kernel thread.
1416  *	the argument is old PIL for an interrupt,
1417  *	or the distingished value KPREEMPT_SYNC.
1418  */
1419 void
1420 kpreempt(int asyncspl)
1421 {
1422 	kthread_t *cur_thread = curthread;
1423 
1424 	if (IGNORE_KERNEL_PREEMPTION) {
1425 		aston(CPU->cpu_dispthread);
1426 		return;
1427 	}
1428 
1429 	/*
1430 	 * Check that conditions are right for kernel preemption
1431 	 */
1432 	do {
1433 		if (cur_thread->t_preempt) {
1434 			/*
1435 			 * either a privileged thread (idle, panic, interrupt)
1436 			 *	or will check when t_preempt is lowered
1437 			 */
1438 			if (cur_thread->t_pri < 0)
1439 				kpreempt_cnts.kpc_idle++;
1440 			else if (cur_thread->t_flag & T_INTR_THREAD) {
1441 				kpreempt_cnts.kpc_intr++;
1442 				if (cur_thread->t_pil == CLOCK_LEVEL)
1443 					kpreempt_cnts.kpc_clock++;
1444 			} else
1445 				kpreempt_cnts.kpc_blocked++;
1446 			aston(CPU->cpu_dispthread);
1447 			return;
1448 		}
1449 		if (cur_thread->t_state != TS_ONPROC ||
1450 		    cur_thread->t_disp_queue != CPU->cpu_disp) {
1451 			/* this thread will be calling swtch() shortly */
1452 			kpreempt_cnts.kpc_notonproc++;
1453 			if (CPU->cpu_thread != CPU->cpu_dispthread) {
1454 				/* already in swtch(), force another */
1455 				kpreempt_cnts.kpc_inswtch++;
1456 				siron();
1457 			}
1458 			return;
1459 		}
1460 		if (getpil() >= DISP_LEVEL) {
1461 			/*
1462 			 * We can't preempt this thread if it is at
1463 			 * a PIL >= DISP_LEVEL since it may be holding
1464 			 * a spin lock (like sched_lock).
1465 			 */
1466 			siron();	/* check back later */
1467 			kpreempt_cnts.kpc_prilevel++;
1468 			return;
1469 		}
1470 
1471 		if (asyncspl != KPREEMPT_SYNC)
1472 			kpreempt_cnts.kpc_apreempt++;
1473 		else
1474 			kpreempt_cnts.kpc_spreempt++;
1475 
1476 		cur_thread->t_preempt++;
1477 		preempt();
1478 		cur_thread->t_preempt--;
1479 	} while (CPU->cpu_kprunrun);
1480 }
1481 
1482 /*
1483  * Print out debugging info.
1484  */
1485 static void
1486 showregs(uint_t type, struct regs *rp, caddr_t addr)
1487 {
1488 	int s;
1489 
1490 	s = spl7();
1491 	type &= ~USER;
1492 	if (u.u_comm[0])
1493 		printf("%s: ", u.u_comm);
1494 	if (type < TRAP_TYPES)
1495 		printf("#%s %s\n", trap_type_mnemonic[type], trap_type[type]);
1496 	else
1497 		switch (type) {
1498 		case T_SYSCALL:
1499 			printf("Syscall Trap:\n");
1500 			break;
1501 		case T_AST:
1502 			printf("AST\n");
1503 			break;
1504 		default:
1505 			printf("Bad Trap = %d\n", type);
1506 			break;
1507 		}
1508 	if (type == T_PGFLT) {
1509 		printf("Bad %s fault at addr=0x%lx\n",
1510 		    USERMODE(rp->r_cs) ? "user": "kernel", (uintptr_t)addr);
1511 	} else if (addr) {
1512 		printf("addr=0x%lx\n", (uintptr_t)addr);
1513 	}
1514 
1515 	printf("pid=%d, pc=0x%lx, sp=0x%lx, eflags=0x%lx\n",
1516 	    (ttoproc(curthread) && ttoproc(curthread)->p_pidp) ?
1517 	    ttoproc(curthread)->p_pid : 0, rp->r_pc, rp->r_sp, rp->r_ps);
1518 
1519 #if defined(__lint)
1520 	/*
1521 	 * this clause can be deleted when lint bug 4870403 is fixed
1522 	 * (lint thinks that bit 32 is illegal in a %b format string)
1523 	 */
1524 	printf("cr0: %x cr4: %b\n",
1525 	    (uint_t)getcr0(), (uint_t)getcr4(), FMT_CR4);
1526 #else
1527 	printf("cr0: %b cr4: %b\n",
1528 	    (uint_t)getcr0(), FMT_CR0, (uint_t)getcr4(), FMT_CR4);
1529 #endif
1530 
1531 #if defined(__amd64)
1532 	printf("cr2: %lx cr3: %lx cr8: %lx\n", getcr2(), getcr3(), getcr8());
1533 #elif defined(__i386)
1534 	printf("cr2: %lx cr3: %lx\n", getcr2(), getcr3());
1535 #endif
1536 
1537 	dumpregs(rp);
1538 	splx(s);
1539 }
1540 
1541 static void
1542 dumpregs(struct regs *rp)
1543 {
1544 #if defined(__amd64)
1545 	const char fmt[] = "\t%3s: %16lx %3s: %16lx %3s: %16lx\n";
1546 
1547 	printf(fmt, "rdi", rp->r_rdi, "rsi", rp->r_rsi, "rdx", rp->r_rdx);
1548 	printf(fmt, "rcx", rp->r_rcx, " r8", rp->r_r8, " r9", rp->r_r9);
1549 	printf(fmt, "rax", rp->r_rax, "rbx", rp->r_rbx, "rbp", rp->r_rbp);
1550 	printf(fmt, "r10", rp->r_r10, "r11", rp->r_r11, "r12", rp->r_r12);
1551 	printf(fmt, "r13", rp->r_r13, "r14", rp->r_r14, "r15", rp->r_r15);
1552 
1553 	printf(fmt, "fsb", rp->r_fsbase, "gsb", rp->r_gsbase, " ds", rp->r_ds);
1554 	printf(fmt, " es", rp->r_es, " fs", rp->r_fs, " gs", rp->r_gs);
1555 
1556 	printf(fmt, "trp", rp->r_trapno, "err", rp->r_err, "rip", rp->r_rip);
1557 	printf(fmt, " cs", rp->r_cs, "rfl", rp->r_rfl, "rsp", rp->r_rsp);
1558 
1559 	printf("\t%3s: %16lx\n", " ss", rp->r_ss);
1560 
1561 #elif defined(__i386)
1562 	const char fmt[] = "\t%3s: %8lx %3s: %8lx %3s: %8lx %3s: %8lx\n";
1563 
1564 	printf(fmt, " gs", rp->r_gs, " fs", rp->r_fs,
1565 	    " es", rp->r_es, " ds", rp->r_ds);
1566 	printf(fmt, "edi", rp->r_edi, "esi", rp->r_esi,
1567 	    "ebp", rp->r_ebp, "esp", rp->r_esp);
1568 	printf(fmt, "ebx", rp->r_ebx, "edx", rp->r_edx,
1569 	    "ecx", rp->r_ecx, "eax", rp->r_eax);
1570 	printf(fmt, "trp", rp->r_trapno, "err", rp->r_err,
1571 	    "eip", rp->r_eip, " cs", rp->r_cs);
1572 	printf("\t%3s: %8lx %3s: %8lx %3s: %8lx\n",
1573 	    "efl", rp->r_efl, "usp", rp->r_uesp, " ss", rp->r_ss);
1574 
1575 #endif	/* __i386 */
1576 }
1577 
1578 /*
1579  * Handle #gp faults in kernel mode.
1580  *
1581  * One legitimate way this can happen is if we attempt to update segment
1582  * registers to naughty values on the way out of the kernel.
1583  *
1584  * This can happen in a couple of ways: someone - either accidentally or
1585  * on purpose - creates (setcontext(2), lwp_create(2)) or modifies
1586  * (signal(2)) a ucontext that contains silly segment register values.
1587  * Or someone - either accidentally or on purpose - modifies the prgregset_t
1588  * of a subject process via /proc to contain silly segment register values.
1589  *
1590  * (The unfortunate part is that we can end up discovering the bad segment
1591  * register value in the middle of an 'iret' after we've popped most of the
1592  * stack.  So it becomes quite difficult to associate an accurate ucontext
1593  * with the lwp, because the act of taking the #gp trap overwrites most of
1594  * what we were going to send the lwp.)
1595  *
1596  * OTOH if it turns out that's -not- the problem, and we're -not- an lwp
1597  * trying to return to user mode and we get a #gp fault, then we need
1598  * to die() -- which will happen if we return non-zero from this routine.
1599  */
1600 static int
1601 kern_gpfault(struct regs *rp)
1602 {
1603 	kthread_t *t = curthread;
1604 	proc_t *p = ttoproc(t);
1605 	klwp_t *lwp = ttolwp(t);
1606 	struct regs tmpregs, *trp = NULL;
1607 	caddr_t pc = (caddr_t)rp->r_pc;
1608 	int v;
1609 
1610 	extern void _sys_rtt(), sr_sup();
1611 
1612 #if defined(__amd64)
1613 	extern void _update_sregs(), _update_sregs_done();
1614 	static const uint8_t iretq_insn[2] = { 0x48, 0xcf };
1615 
1616 #elif defined(__i386)
1617 	static const uint8_t iret_insn[1] = { 0xcf };
1618 
1619 	/*
1620 	 * Note carefully the appallingly awful dependency between
1621 	 * the instruction sequence used in __SEGREGS_POP and these
1622 	 * instructions encoded here.
1623 	 *
1624 	 * XX64	Add some commentary to locore.s/privregs.h to document this.
1625 	 */
1626 	static const uint8_t movw_0_esp_gs[4] = { 0x8e, 0x6c, 0x24, 0x0 };
1627 	static const uint8_t movw_4_esp_fs[4] = { 0x8e, 0x64, 0x24, 0x4 };
1628 	static const uint8_t movw_8_esp_es[4] = { 0x8e, 0x44, 0x24, 0x8 };
1629 	static const uint8_t movw_c_esp_ds[4] = { 0x8e, 0x5c, 0x24, 0xc };
1630 #endif
1631 	/*
1632 	 * if we're not an lwp, or the pc range is outside _sys_rtt, then
1633 	 * we should immediately be die()ing horribly
1634 	 */
1635 	if (lwp == NULL ||
1636 	    (uintptr_t)pc < (uintptr_t)_sys_rtt ||
1637 	    (uintptr_t)pc > (uintptr_t)sr_sup)
1638 		return (1);
1639 
1640 	/*
1641 	 * So at least we're in the right part of the kernel.
1642 	 *
1643 	 * Disassemble the instruction at the faulting pc.
1644 	 * Once we know what it is, we carefully reconstruct the stack
1645 	 * based on the order in which the stack is deconstructed in
1646 	 * _sys_rtt. Ew.
1647 	 */
1648 
1649 #if defined(__amd64)
1650 
1651 	if (bcmp(pc, iretq_insn, sizeof (iretq_insn)) == 0) {
1652 		/*
1653 		 * We took the #gp while trying to perform the iretq.
1654 		 * This means that either %cs or %ss are bad.
1655 		 * All we know for sure is that most of the general
1656 		 * registers have been restored, including the
1657 		 * segment registers, and all we have left on the
1658 		 * topmost part of the lwp's stack are the
1659 		 * registers that the iretq was unable to consume.
1660 		 *
1661 		 * All the rest of the state was crushed by the #gp
1662 		 * which pushed -its- registers atop our old save area
1663 		 * (because we had to decrement the stack pointer, sigh) so
1664 		 * all that we can try and do is to reconstruct the
1665 		 * crushed frame from the #gp trap frame itself.
1666 		 */
1667 		trp = &tmpregs;
1668 		trp->r_ss = lwptoregs(lwp)->r_ss;
1669 		trp->r_sp = lwptoregs(lwp)->r_sp;
1670 		trp->r_ps = lwptoregs(lwp)->r_ps;
1671 		trp->r_cs = lwptoregs(lwp)->r_cs;
1672 		trp->r_pc = lwptoregs(lwp)->r_pc;
1673 		bcopy(rp, trp, offsetof(struct regs, r_pc));
1674 
1675 		/*
1676 		 * Validate simple math
1677 		 */
1678 		ASSERT(trp->r_pc == lwptoregs(lwp)->r_pc);
1679 		ASSERT(trp->r_err == rp->r_err);
1680 
1681 	} else if ((lwp->lwp_pcb.pcb_flags & RUPDATE_PENDING) != 0 &&
1682 	    pc >= (caddr_t)_update_sregs &&
1683 	    pc < (caddr_t)_update_sregs_done) {
1684 		/*
1685 		 * This is the common case -- we're trying to load
1686 		 * a bad segment register value in the only section
1687 		 * of kernel code that ever loads segment registers.
1688 		 *
1689 		 * We don't need to do anything at this point because
1690 		 * the pcb contains all the pending segment register
1691 		 * state, and the regs are still intact because we
1692 		 * didn't adjust the stack pointer yet.  Given the fidelity
1693 		 * of all this, we could conceivably send a signal
1694 		 * to the lwp, rather than core-ing.
1695 		 */
1696 		trp = lwptoregs(lwp);
1697 		ASSERT((caddr_t)trp == (caddr_t)rp->r_sp);
1698 	}
1699 
1700 #elif defined(__i386)
1701 
1702 	if (bcmp(pc, iret_insn, sizeof (iret_insn)) == 0) {
1703 		/*
1704 		 * We took the #gp while trying to perform the iret.
1705 		 * This means that either %cs or %ss are bad.
1706 		 * All we know for sure is that most of the general
1707 		 * registers have been restored, including the
1708 		 * segment registers, and all we have left on the
1709 		 * topmost part of the lwp's stack are the registers that
1710 		 * the iret was unable to consume.
1711 		 *
1712 		 * All the rest of the state was crushed by the #gp
1713 		 * which pushed -its- registers atop our old save area
1714 		 * (because we had to decrement the stack pointer, sigh) so
1715 		 * all that we can try and do is to reconstruct the
1716 		 * crushed frame from the #gp trap frame itself.
1717 		 */
1718 		trp = &tmpregs;
1719 		trp->r_ss = lwptoregs(lwp)->r_ss;
1720 		trp->r_sp = lwptoregs(lwp)->r_sp;
1721 		trp->r_ps = lwptoregs(lwp)->r_ps;
1722 		trp->r_cs = lwptoregs(lwp)->r_cs;
1723 		trp->r_pc = lwptoregs(lwp)->r_pc;
1724 		bcopy(rp, trp, offsetof(struct regs, r_pc));
1725 
1726 		ASSERT(trp->r_pc == lwptoregs(lwp)->r_pc);
1727 		ASSERT(trp->r_err == rp->r_err);
1728 
1729 	} else {
1730 		/*
1731 		 * Segment registers are reloaded in _sys_rtt
1732 		 * via the following sequence:
1733 		 *
1734 		 *	movw	0(%esp), %gs
1735 		 *	movw	4(%esp), %fs
1736 		 *	movw	8(%esp), %es
1737 		 *	movw	12(%esp), %ds
1738 		 *	addl	$16, %esp
1739 		 *
1740 		 * Thus if any of them fault, we know the user
1741 		 * registers are left unharmed on the stack.
1742 		 */
1743 		if (bcmp(pc, movw_0_esp_gs, sizeof (movw_0_esp_gs)) == 0 ||
1744 		    bcmp(pc, movw_4_esp_fs, sizeof (movw_4_esp_fs)) == 0 ||
1745 		    bcmp(pc, movw_8_esp_es, sizeof (movw_8_esp_es)) == 0 ||
1746 		    bcmp(pc, movw_c_esp_ds, sizeof (movw_c_esp_ds)) == 0)
1747 			trp = lwptoregs(lwp);
1748 	}
1749 #endif	/* __amd64 */
1750 
1751 	if (trp == NULL)
1752 		return (1);
1753 
1754 	/*
1755 	 * If we get to here, we're reasonably confident that we've
1756 	 * correctly decoded what happened on the way out of the kernel.
1757 	 * Rewrite the lwp's registers so that we can create a core dump
1758 	 * the (at least vaguely) represents the mcontext we were
1759 	 * being asked to restore when things went so terribly wrong.
1760 	 */
1761 
1762 	/*
1763 	 * Make sure that we have a meaningful %trapno and %err.
1764 	 */
1765 	trp->r_trapno = rp->r_trapno;
1766 	trp->r_err = rp->r_err;
1767 
1768 	if ((caddr_t)trp != (caddr_t)lwptoregs(lwp))
1769 		bcopy(trp, lwptoregs(lwp), sizeof (*trp));
1770 
1771 	mutex_enter(&p->p_lock);
1772 	lwp->lwp_cursig = SIGSEGV;
1773 	mutex_exit(&p->p_lock);
1774 
1775 	/*
1776 	 * Terminate all LWPs but don't discard them.  If another lwp beat us to
1777 	 * the punch by calling exit(), evaporate now.
1778 	 */
1779 	proc_is_exiting(p);
1780 	if (exitlwps(1) != 0) {
1781 		mutex_enter(&p->p_lock);
1782 		lwp_exit();
1783 	}
1784 
1785 #ifdef C2_AUDIT
1786 	if (audit_active)		/* audit core dump */
1787 		audit_core_start(SIGSEGV);
1788 #endif
1789 	v = core(SIGSEGV, B_FALSE);
1790 #ifdef C2_AUDIT
1791 	if (audit_active)		/* audit core dump */
1792 		audit_core_finish(v ? CLD_KILLED : CLD_DUMPED);
1793 #endif
1794 	exit(v ? CLD_KILLED : CLD_DUMPED, SIGSEGV);
1795 	return (0);
1796 }
1797 
1798 /*
1799  * dump_tss() - Display the TSS structure
1800  */
1801 
1802 #if defined(__amd64)
1803 
1804 static void
1805 dump_tss(void)
1806 {
1807 	const char tss_fmt[] = "tss.%s:\t0x%p\n";  /* Format string */
1808 	struct tss *tss = CPU->cpu_tss;
1809 
1810 	printf(tss_fmt, "tss_rsp0", (void *)tss->tss_rsp0);
1811 	printf(tss_fmt, "tss_rsp1", (void *)tss->tss_rsp1);
1812 	printf(tss_fmt, "tss_rsp2", (void *)tss->tss_rsp2);
1813 
1814 	printf(tss_fmt, "tss_ist1", (void *)tss->tss_ist1);
1815 	printf(tss_fmt, "tss_ist2", (void *)tss->tss_ist2);
1816 	printf(tss_fmt, "tss_ist3", (void *)tss->tss_ist3);
1817 	printf(tss_fmt, "tss_ist4", (void *)tss->tss_ist4);
1818 	printf(tss_fmt, "tss_ist5", (void *)tss->tss_ist5);
1819 	printf(tss_fmt, "tss_ist6", (void *)tss->tss_ist6);
1820 	printf(tss_fmt, "tss_ist7", (void *)tss->tss_ist7);
1821 }
1822 
1823 #elif defined(__i386)
1824 
1825 static void
1826 dump_tss(void)
1827 {
1828 	const char tss_fmt[] = "tss.%s:\t0x%p\n";  /* Format string */
1829 	struct tss *tss = CPU->cpu_tss;
1830 
1831 	printf(tss_fmt, "tss_link", (void *)(uintptr_t)tss->tss_link);
1832 	printf(tss_fmt, "tss_esp0", (void *)(uintptr_t)tss->tss_esp0);
1833 	printf(tss_fmt, "tss_ss0", (void *)(uintptr_t)tss->tss_ss0);
1834 	printf(tss_fmt, "tss_esp1", (void *)(uintptr_t)tss->tss_esp1);
1835 	printf(tss_fmt, "tss_ss1", (void *)(uintptr_t)tss->tss_ss1);
1836 	printf(tss_fmt, "tss_esp2", (void *)(uintptr_t)tss->tss_esp2);
1837 	printf(tss_fmt, "tss_ss2", (void *)(uintptr_t)tss->tss_ss2);
1838 	printf(tss_fmt, "tss_cr3", (void *)(uintptr_t)tss->tss_cr3);
1839 	printf(tss_fmt, "tss_eip", (void *)(uintptr_t)tss->tss_eip);
1840 	printf(tss_fmt, "tss_eflags", (void *)(uintptr_t)tss->tss_eflags);
1841 	printf(tss_fmt, "tss_eax", (void *)(uintptr_t)tss->tss_eax);
1842 	printf(tss_fmt, "tss_ebx", (void *)(uintptr_t)tss->tss_ebx);
1843 	printf(tss_fmt, "tss_ecx", (void *)(uintptr_t)tss->tss_ecx);
1844 	printf(tss_fmt, "tss_edx", (void *)(uintptr_t)tss->tss_edx);
1845 	printf(tss_fmt, "tss_esp", (void *)(uintptr_t)tss->tss_esp);
1846 }
1847 
1848 #endif	/* __amd64 */
1849 
1850 #if defined(TRAPTRACE)
1851 
1852 int ttrace_nrec = 0;		/* number of records to dump out */
1853 int ttrace_dump_nregs = 5;	/* dump out this many records with regs too */
1854 
1855 /*
1856  * Dump out the last ttrace_nrec traptrace records on each CPU
1857  */
1858 static void
1859 dump_ttrace(void)
1860 {
1861 	trap_trace_ctl_t *ttc;
1862 	trap_trace_rec_t *rec;
1863 	uintptr_t current;
1864 	int i, j, k;
1865 	int n = NCPU;
1866 #if defined(__amd64)
1867 	const char banner[] =
1868 		"\ncpu          address    timestamp "
1869 		"type  vc  handler   pc\n";
1870 	const char fmt1[] = "%3d %016lx %12llx ";
1871 #elif defined(__i386)
1872 	const char banner[] =
1873 		"\ncpu address     timestamp type  vc  handler   pc\n";
1874 	const char fmt1[] = "%3d %08lx %12llx ";
1875 #endif
1876 	const char fmt2[] = "%4s %3x ";
1877 	const char fmt3[] = "%8s ";
1878 
1879 	if (ttrace_nrec == 0)
1880 		return;
1881 
1882 	printf(banner);
1883 
1884 	for (i = 0; i < n; i++) {
1885 		ttc = &trap_trace_ctl[i];
1886 		if (ttc->ttc_first == NULL)
1887 			continue;
1888 
1889 		current = ttc->ttc_next - sizeof (trap_trace_rec_t);
1890 		for (j = 0; j < ttrace_nrec; j++) {
1891 			struct sysent	*sys;
1892 			struct autovec	*vec;
1893 			extern struct av_head autovect[];
1894 			int type;
1895 			ulong_t	off;
1896 			char *sym, *stype;
1897 
1898 			if (current < ttc->ttc_first)
1899 				current =
1900 				    ttc->ttc_limit - sizeof (trap_trace_rec_t);
1901 
1902 			if (current == NULL)
1903 				continue;
1904 
1905 			rec = (trap_trace_rec_t *)current;
1906 
1907 			if (rec->ttr_stamp == 0)
1908 				break;
1909 
1910 			printf(fmt1, i, (uintptr_t)rec, rec->ttr_stamp);
1911 
1912 			switch (rec->ttr_marker) {
1913 			case TT_SYSCALL:
1914 			case TT_SYSENTER:
1915 			case TT_SYSC:
1916 			case TT_SYSC64:
1917 #if defined(__amd64)
1918 				sys = &sysent32[rec->ttr_sysnum];
1919 				switch (rec->ttr_marker) {
1920 				case TT_SYSC64:
1921 					sys = &sysent[rec->ttr_sysnum];
1922 					/*FALLTHROUGH*/
1923 #elif defined(__i386)
1924 				sys = &sysent[rec->ttr_sysnum];
1925 				switch (rec->ttr_marker) {
1926 				case TT_SYSC64:
1927 #endif
1928 				case TT_SYSC:
1929 					stype = "sysc";	/* syscall */
1930 					break;
1931 				case TT_SYSCALL:
1932 					stype = "lcal";	/* lcall */
1933 					break;
1934 				case TT_SYSENTER:
1935 					stype = "syse";	/* sysenter */
1936 					break;
1937 				default:
1938 					break;
1939 				}
1940 				printf(fmt2, "sysc", rec->ttr_sysnum);
1941 				if (sys != NULL) {
1942 					sym = kobj_getsymname(
1943 					    (uintptr_t)sys->sy_callc,
1944 					    &off);
1945 					if (sym != NULL)
1946 						printf("%s ", sym);
1947 					else
1948 						printf("%p ", sys->sy_callc);
1949 				} else {
1950 					printf("unknown ");
1951 				}
1952 				break;
1953 
1954 			case TT_INTERRUPT:
1955 				printf(fmt2, "intr", rec->ttr_vector);
1956 				vec = (&autovect[rec->ttr_vector])->avh_link;
1957 				if (vec != NULL) {
1958 					sym = kobj_getsymname(
1959 					    (uintptr_t)vec->av_vector, &off);
1960 					if (sym != NULL)
1961 						printf("%s ", sym);
1962 					else
1963 						printf("%p ", vec->av_vector);
1964 				} else {
1965 					printf("unknown ");
1966 				}
1967 				break;
1968 
1969 			case TT_TRAP:
1970 				type = rec->ttr_regs.r_trapno;
1971 				printf(fmt2, "trap", type);
1972 				printf("#%s ", type < TRAP_TYPES ?
1973 				    trap_type_mnemonic[type] : "trap");
1974 				break;
1975 
1976 			default:
1977 				break;
1978 			}
1979 
1980 			sym = kobj_getsymname(rec->ttr_regs.r_pc, &off);
1981 			if (sym != NULL)
1982 				printf("%s+%lx\n", sym, off);
1983 			else
1984 				printf("%lx\n", rec->ttr_regs.r_pc);
1985 
1986 			if (ttrace_dump_nregs-- > 0) {
1987 				int s;
1988 
1989 				if (rec->ttr_marker == TT_INTERRUPT)
1990 					printf(
1991 					    "\t\tipl %x spl %x pri %x\n",
1992 					    rec->ttr_ipl,
1993 					    rec->ttr_spl,
1994 					    rec->ttr_pri);
1995 
1996 				dumpregs(&rec->ttr_regs);
1997 
1998 				printf("\t%3s: %p\n\n", " ct",
1999 				    (void *)rec->ttr_curthread);
2000 
2001 				/*
2002 				 * print out the pc stack that we recorded
2003 				 * at trap time (if any)
2004 				 */
2005 				for (s = 0; s < rec->ttr_sdepth; s++) {
2006 					uintptr_t fullpc;
2007 
2008 					if (s >= TTR_STACK_DEPTH) {
2009 						printf("ttr_sdepth corrupt\n");
2010 						break;
2011 					}
2012 
2013 					fullpc = (uintptr_t)rec->ttr_stack[s];
2014 
2015 					sym = kobj_getsymname(fullpc, &off);
2016 					if (sym != NULL)
2017 						printf("-> %s+0x%lx()\n",
2018 						    sym, off);
2019 					else
2020 						printf("-> 0x%lx()\n", fullpc);
2021 				}
2022 				printf("\n");
2023 			}
2024 			current -= sizeof (trap_trace_rec_t);
2025 		}
2026 	}
2027 }
2028 
2029 #endif	/* TRAPTRACE */
2030 
2031 void
2032 panic_showtrap(struct trap_info *tip)
2033 {
2034 	showregs(tip->trap_type, tip->trap_regs, tip->trap_addr);
2035 
2036 #if defined(TRAPTRACE)
2037 	dump_ttrace();
2038 #endif	/* TRAPTRACE */
2039 
2040 	if (tip->trap_type == T_DBLFLT)
2041 		dump_tss();
2042 }
2043 
2044 void
2045 panic_savetrap(panic_data_t *pdp, struct trap_info *tip)
2046 {
2047 	panic_saveregs(pdp, tip->trap_regs);
2048 }
2049