xref: /illumos-gate/usr/src/uts/i86pc/os/trap.c (revision 4e5b757f)
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 2007 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/psw.h>
54 #include <sys/regset.h>
55 #include <sys/fp.h>
56 #include <sys/trap.h>
57 #include <sys/kmem.h>
58 #include <sys/vtrace.h>
59 #include <sys/cmn_err.h>
60 #include <sys/prsystm.h>
61 #include <sys/mutex_impl.h>
62 #include <sys/machsystm.h>
63 #include <sys/archsystm.h>
64 #include <sys/sdt.h>
65 #include <sys/avintr.h>
66 #include <sys/kobj.h>
67 
68 #include <vm/hat.h>
69 
70 #include <vm/seg_kmem.h>
71 #include <vm/as.h>
72 #include <vm/seg.h>
73 #include <vm/hat_pte.h>
74 #include <vm/hat_i86.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 #include <sys/bootconf.h>
96 #include <sys/bootinfo.h>
97 #include <sys/promif.h>
98 #include <sys/mach_mmu.h>
99 
100 #define	USER	0x10000		/* user-mode flag added to trap type */
101 
102 static const char *trap_type_mnemonic[] = {
103 	"de",	"db",	"2",	"bp",
104 	"of",	"br",	"ud",	"nm",
105 	"df",	"9",	"ts",	"np",
106 	"ss",	"gp",	"pf",	"15",
107 	"mf",	"ac",	"mc",	"xf"
108 };
109 
110 static const char *trap_type[] = {
111 	"Divide error",				/* trap id 0 	*/
112 	"Debug",				/* trap id 1	*/
113 	"NMI interrupt",			/* trap id 2	*/
114 	"Breakpoint",				/* trap id 3 	*/
115 	"Overflow",				/* trap id 4 	*/
116 	"BOUND range exceeded",			/* trap id 5 	*/
117 	"Invalid opcode",			/* trap id 6 	*/
118 	"Device not available",			/* trap id 7 	*/
119 	"Double fault",				/* trap id 8 	*/
120 	"Coprocessor segment overrun",		/* trap id 9 	*/
121 	"Invalid TSS",				/* trap id 10 	*/
122 	"Segment not present",			/* trap id 11 	*/
123 	"Stack segment fault",			/* trap id 12 	*/
124 	"General protection",			/* trap id 13 	*/
125 	"Page fault",				/* trap id 14 	*/
126 	"Reserved",				/* trap id 15 	*/
127 	"x87 floating point error",		/* trap id 16 	*/
128 	"Alignment check",			/* trap id 17 	*/
129 	"Machine check",			/* trap id 18	*/
130 	"SIMD floating point exception",	/* trap id 19	*/
131 };
132 
133 #define	TRAP_TYPES	(sizeof (trap_type) / sizeof (trap_type[0]))
134 
135 #define	SLOW_SCALL_SIZE	2
136 #define	FAST_SCALL_SIZE	2
137 
138 int tudebug = 0;
139 int tudebugbpt = 0;
140 int tudebugfpe = 0;
141 int tudebugsse = 0;
142 
143 #if defined(TRAPDEBUG) || defined(lint)
144 int tdebug = 0;
145 int lodebug = 0;
146 int faultdebug = 0;
147 #else
148 #define	tdebug	0
149 #define	lodebug	0
150 #define	faultdebug	0
151 #endif /* defined(TRAPDEBUG) || defined(lint) */
152 
153 #if defined(TRAPTRACE)
154 static void dump_ttrace(void);
155 #endif	/* TRAPTRACE */
156 static void dumpregs(struct regs *);
157 static void showregs(uint_t, struct regs *, caddr_t);
158 static void dump_tss(void);
159 static int kern_gpfault(struct regs *);
160 
161 struct trap_info {
162 	struct regs *trap_regs;
163 	uint_t trap_type;
164 	caddr_t trap_addr;
165 };
166 
167 /*ARGSUSED*/
168 static int
169 die(uint_t type, struct regs *rp, caddr_t addr, processorid_t cpuid)
170 {
171 	struct trap_info ti;
172 	const char *trap_name, *trap_mnemonic;
173 
174 	if (type < TRAP_TYPES) {
175 		trap_name = trap_type[type];
176 		trap_mnemonic = trap_type_mnemonic[type];
177 	} else {
178 		trap_name = "trap";
179 		trap_mnemonic = "-";
180 	}
181 
182 #ifdef TRAPTRACE
183 	TRAPTRACE_FREEZE;
184 #endif
185 
186 	ti.trap_regs = rp;
187 	ti.trap_type = type & ~USER;
188 	ti.trap_addr = addr;
189 
190 	curthread->t_panic_trap = &ti;
191 
192 	if (type == T_PGFLT && addr < (caddr_t)KERNELBASE) {
193 		panic("BAD TRAP: type=%x (#%s %s) rp=%p addr=%p "
194 		    "occurred in module \"%s\" due to %s",
195 		    type, trap_mnemonic, trap_name, (void *)rp, (void *)addr,
196 		    mod_containing_pc((caddr_t)rp->r_pc),
197 		    addr < (caddr_t)PAGESIZE ?
198 		    "a NULL pointer dereference" :
199 		    "an illegal access to a user address");
200 	} else
201 		panic("BAD TRAP: type=%x (#%s %s) rp=%p addr=%p",
202 		    type, trap_mnemonic, trap_name, (void *)rp, (void *)addr);
203 	return (0);
204 }
205 
206 /*
207  * Rewrite the instruction at pc to be an int $T_SYSCALLINT instruction.
208  *
209  * int <vector> is two bytes: 0xCD <vector>
210  */
211 
212 static int
213 rewrite_syscall(caddr_t pc)
214 {
215 	uchar_t instr[SLOW_SCALL_SIZE] = { 0xCD, T_SYSCALLINT };
216 
217 	if (uwrite(curthread->t_procp, instr, SLOW_SCALL_SIZE,
218 	    (uintptr_t)pc) != 0)
219 		return (1);
220 
221 	return (0);
222 }
223 
224 /*
225  * Test to see if the instruction at pc is sysenter or syscall. The second
226  * argument should be the x86 feature flag corresponding to the expected
227  * instruction.
228  *
229  * sysenter is two bytes: 0x0F 0x34
230  * syscall is two bytes:  0x0F 0x05
231  * int $T_SYSCALLINT is two bytes: 0xCD 0x91
232  */
233 
234 static int
235 instr_is_other_syscall(caddr_t pc, int which)
236 {
237 	uchar_t instr[FAST_SCALL_SIZE];
238 
239 	ASSERT(which == X86_SEP || which == X86_ASYSC || which == 0xCD);
240 
241 	if (copyin_nowatch(pc, (caddr_t)instr, FAST_SCALL_SIZE) != 0)
242 		return (0);
243 
244 	switch (which) {
245 	case X86_SEP:
246 		if (instr[0] == 0x0F && instr[1] == 0x34)
247 			return (1);
248 		break;
249 	case X86_ASYSC:
250 		if (instr[0] == 0x0F && instr[1] == 0x05)
251 			return (1);
252 		break;
253 	case 0xCD:
254 		if (instr[0] == 0xCD && instr[1] == T_SYSCALLINT)
255 			return (1);
256 		break;
257 	}
258 
259 	return (0);
260 }
261 
262 static const char *
263 syscall_insn_string(int syscall_insn)
264 {
265 	switch (syscall_insn) {
266 	case X86_SEP:
267 		return ("sysenter");
268 	case X86_ASYSC:
269 		return ("syscall");
270 	case 0xCD:
271 		return ("int");
272 	default:
273 		return ("Unknown");
274 	}
275 }
276 
277 static int
278 ldt_rewrite_syscall(struct regs *rp, proc_t *p, int syscall_insn)
279 {
280 	caddr_t	linearpc;
281 	int return_code = 0;
282 
283 	mutex_enter(&p->p_ldtlock);	/* Must be held across linear_pc() */
284 
285 	if (linear_pc(rp, p, &linearpc) == 0) {
286 
287 		/*
288 		 * If another thread beat us here, it already changed
289 		 * this site to the slower (int) syscall instruction.
290 		 */
291 		if (instr_is_other_syscall(linearpc, 0xCD)) {
292 			return_code = 1;
293 		} else if (instr_is_other_syscall(linearpc, syscall_insn)) {
294 
295 			if (rewrite_syscall(linearpc) == 0) {
296 				return_code = 1;
297 			}
298 #ifdef DEBUG
299 			else
300 				cmn_err(CE_WARN, "failed to rewrite %s "
301 				    "instruction in process %d",
302 				    syscall_insn_string(syscall_insn),
303 				    p->p_pid);
304 #endif /* DEBUG */
305 		}
306 	}
307 
308 	mutex_exit(&p->p_ldtlock);	/* Must be held across linear_pc() */
309 
310 	return (return_code);
311 }
312 
313 /*
314  * Test to see if the instruction at pc is a system call instruction.
315  *
316  * The bytes of an lcall instruction used for the syscall trap.
317  * static uchar_t lcall[7] = { 0x9a, 0, 0, 0, 0, 0x7, 0 };
318  * static uchar_t lcallalt[7] = { 0x9a, 0, 0, 0, 0, 0x27, 0 };
319  */
320 
321 #define	LCALLSIZE	7
322 
323 static int
324 instr_is_lcall_syscall(caddr_t pc)
325 {
326 	uchar_t instr[LCALLSIZE];
327 
328 	if (copyin_nowatch(pc, (caddr_t)instr, LCALLSIZE) == 0 &&
329 	    instr[0] == 0x9a &&
330 	    instr[1] == 0 &&
331 	    instr[2] == 0 &&
332 	    instr[3] == 0 &&
333 	    instr[4] == 0 &&
334 	    (instr[5] == 0x7 || instr[5] == 0x27) &&
335 	    instr[6] == 0)
336 		return (1);
337 
338 	return (0);
339 }
340 
341 #ifdef __amd64
342 
343 /*
344  * In the first revisions of amd64 CPUs produced by AMD, the LAHF and
345  * SAHF instructions were not implemented in 64-bit mode. Later revisions
346  * did implement these instructions. An extension to the cpuid instruction
347  * was added to check for the capability of executing these instructions
348  * in 64-bit mode.
349  *
350  * Intel originally did not implement these instructions in EM64T either,
351  * but added them in later revisions.
352  *
353  * So, there are different chip revisions by both vendors out there that
354  * may or may not implement these instructions. The easy solution is to
355  * just always emulate these instructions on demand.
356  *
357  * SAHF == store %ah in the lower 8 bits of %rflags (opcode 0x9e)
358  * LAHF == load the lower 8 bits of %rflags into %ah (opcode 0x9f)
359  */
360 
361 #define	LSAHFSIZE 1
362 
363 static int
364 instr_is_lsahf(caddr_t pc, uchar_t *instr)
365 {
366 	if (copyin_nowatch(pc, (caddr_t)instr, LSAHFSIZE) == 0 &&
367 	    (*instr == 0x9e || *instr == 0x9f))
368 		return (1);
369 	return (0);
370 }
371 
372 /*
373  * Emulate the LAHF and SAHF instructions. The reference manuals define
374  * these instructions to always load/store bit 1 as a 1, and bits 3 and 5
375  * as a 0. The other, defined, bits are copied (the PS_ICC bits and PS_P).
376  *
377  * Note that %ah is bits 8-15 of %rax.
378  */
379 static void
380 emulate_lsahf(struct regs *rp, uchar_t instr)
381 {
382 	if (instr == 0x9e) {
383 		/* sahf. Copy bits from %ah to flags. */
384 		rp->r_ps = (rp->r_ps & ~0xff) |
385 		    ((rp->r_rax >> 8) & PSL_LSAHFMASK) | PS_MB1;
386 	} else {
387 		/* lahf. Copy bits from flags to %ah. */
388 		rp->r_rax = (rp->r_rax & ~0xff00) |
389 		    (((rp->r_ps & PSL_LSAHFMASK) | PS_MB1) << 8);
390 	}
391 	rp->r_pc += LSAHFSIZE;
392 }
393 #endif /* __amd64 */
394 
395 #ifdef OPTERON_ERRATUM_91
396 
397 /*
398  * Test to see if the instruction at pc is a prefetch instruction.
399  *
400  * The first byte of prefetch instructions is always 0x0F.
401  * The second byte is 0x18 for regular prefetch or 0x0D for AMD 3dnow prefetch.
402  * The third byte is between 0 and 3 inclusive.
403  */
404 
405 #define	PREFETCHSIZE 3
406 
407 static int
408 cmp_to_prefetch(uchar_t *p)
409 {
410 	if (*p == 0x0F && (*(p+1) == 0x18 || *(p+1) == 0x0D) && *(p+2) <= 3)
411 		return (1);
412 	return (0);
413 }
414 
415 static int
416 instr_is_prefetch(caddr_t pc)
417 {
418 	uchar_t instr[PREFETCHSIZE];
419 	int	error;
420 
421 	error = copyin_nowatch(pc, (caddr_t)instr, PREFETCHSIZE);
422 
423 	if (error == 0 && cmp_to_prefetch(instr))
424 		return (1);
425 	return (0);
426 }
427 
428 #endif /* OPTERON_ERRATUM_91 */
429 
430 /*
431  * Called from the trap handler when a processor trap occurs.
432  *
433  * Note: All user-level traps that might call stop() must exit
434  * trap() by 'goto out' or by falling through.
435  * Note Also: trap() is usually called with interrupts enabled, (PS_IE == 1)
436  * however, there are paths that arrive here with PS_IE == 0 so special care
437  * must be taken in those cases.
438  */
439 void
440 trap(struct regs *rp, caddr_t addr, processorid_t cpuid)
441 {
442 	kthread_t *ct = curthread;
443 	enum seg_rw rw;
444 	unsigned type;
445 	proc_t *p = ttoproc(ct);
446 	klwp_t *lwp = ttolwp(ct);
447 	uintptr_t lofault;
448 	faultcode_t pagefault(), res, errcode;
449 	enum fault_type fault_type;
450 	k_siginfo_t siginfo;
451 	uint_t fault = 0;
452 	int mstate;
453 	int sicode = 0;
454 	int watchcode;
455 	int watchpage;
456 	caddr_t vaddr;
457 	int singlestep_twiddle;
458 	size_t sz;
459 	int ta;
460 #ifdef __amd64
461 	uchar_t instr;
462 #endif
463 
464 	ASSERT_STACK_ALIGNED();
465 
466 	type = rp->r_trapno;
467 	CPU_STATS_ADDQ(CPU, sys, trap, 1);
468 	ASSERT(ct->t_schedflag & TS_DONT_SWAP);
469 
470 	if (type == T_PGFLT) {
471 
472 		errcode = rp->r_err;
473 		if (errcode & PF_ERR_WRITE)
474 			rw = S_WRITE;
475 		else if ((caddr_t)rp->r_pc == addr ||
476 		    (mmu.pt_nx != 0 && (errcode & PF_ERR_EXEC)))
477 			rw = S_EXEC;
478 		else
479 			rw = S_READ;
480 
481 #if defined(__i386)
482 		/*
483 		 * Pentium Pro work-around
484 		 */
485 		if ((errcode & PF_ERR_PROT) && pentiumpro_bug4046376) {
486 			uint_t	attr;
487 			uint_t	priv_violation;
488 			uint_t	access_violation;
489 
490 			if (hat_getattr(addr < (caddr_t)kernelbase ?
491 			    curproc->p_as->a_hat : kas.a_hat, addr, &attr)
492 			    == -1) {
493 				errcode &= ~PF_ERR_PROT;
494 			} else {
495 				priv_violation = (errcode & PF_ERR_USER) &&
496 					!(attr & PROT_USER);
497 				access_violation = (errcode & PF_ERR_WRITE) &&
498 					!(attr & PROT_WRITE);
499 				if (!priv_violation && !access_violation)
500 					goto cleanup;
501 			}
502 		}
503 #endif /* __i386 */
504 
505 	} else if (type == T_SGLSTP && lwp != NULL)
506 		lwp->lwp_pcb.pcb_drstat = (uintptr_t)addr;
507 
508 	if (tdebug)
509 		showregs(type, rp, addr);
510 
511 	if (USERMODE(rp->r_cs)) {
512 		/*
513 		 * Set up the current cred to use during this trap. u_cred
514 		 * no longer exists.  t_cred is used instead.
515 		 * The current process credential applies to the thread for
516 		 * the entire trap.  If trapping from the kernel, this
517 		 * should already be set up.
518 		 */
519 		if (ct->t_cred != p->p_cred) {
520 			cred_t *oldcred = ct->t_cred;
521 			/*
522 			 * DTrace accesses t_cred in probe context.  t_cred
523 			 * must always be either NULL, or point to a valid,
524 			 * allocated cred structure.
525 			 */
526 			ct->t_cred = crgetcred();
527 			crfree(oldcred);
528 		}
529 		ASSERT(lwp != NULL);
530 		type |= USER;
531 		ASSERT(lwptoregs(lwp) == rp);
532 		lwp->lwp_state = LWP_SYS;
533 
534 		switch (type) {
535 		case T_PGFLT + USER:
536 			if ((caddr_t)rp->r_pc == addr)
537 				mstate = LMS_TFAULT;
538 			else
539 				mstate = LMS_DFAULT;
540 			break;
541 		default:
542 			mstate = LMS_TRAP;
543 			break;
544 		}
545 		/* Kernel probe */
546 		TNF_PROBE_1(thread_state, "thread", /* CSTYLED */,
547 		    tnf_microstate, state, mstate);
548 		mstate = new_mstate(ct, mstate);
549 
550 		bzero(&siginfo, sizeof (siginfo));
551 	}
552 
553 	switch (type) {
554 	case T_PGFLT + USER:
555 	case T_SGLSTP:
556 	case T_SGLSTP + USER:
557 	case T_BPTFLT + USER:
558 		break;
559 
560 	default:
561 		FTRACE_2("trap(): type=0x%lx, regs=0x%lx",
562 		    (ulong_t)type, (ulong_t)rp);
563 		break;
564 	}
565 
566 	switch (type) {
567 	default:
568 		if (type & USER) {
569 			if (tudebug)
570 				showregs(type, rp, (caddr_t)0);
571 			printf("trap: Unknown trap type %d in user mode\n",
572 			    type & ~USER);
573 			siginfo.si_signo = SIGILL;
574 			siginfo.si_code  = ILL_ILLTRP;
575 			siginfo.si_addr  = (caddr_t)rp->r_pc;
576 			siginfo.si_trapno = type & ~USER;
577 			fault = FLTILL;
578 			break;
579 		} else {
580 			(void) die(type, rp, addr, cpuid);
581 			/*NOTREACHED*/
582 		}
583 
584 	case T_PGFLT:		/* system page fault */
585 		/*
586 		 * If we're under on_trap() protection (see <sys/ontrap.h>),
587 		 * set ot_trap and longjmp back to the on_trap() call site.
588 		 */
589 		if ((ct->t_ontrap != NULL) &&
590 		    (ct->t_ontrap->ot_prot & OT_DATA_ACCESS)) {
591 			ct->t_ontrap->ot_trap |= OT_DATA_ACCESS;
592 			longjmp(&curthread->t_ontrap->ot_jmpbuf);
593 		}
594 
595 		/*
596 		 * See if we can handle as pagefault. Save lofault
597 		 * across this. Here we assume that an address
598 		 * less than KERNELBASE is a user fault.
599 		 * We can do this as copy.s routines verify that the
600 		 * starting address is less than KERNELBASE before
601 		 * starting and because we know that we always have
602 		 * KERNELBASE mapped as invalid to serve as a "barrier".
603 		 */
604 		lofault = ct->t_lofault;
605 		ct->t_lofault = 0;
606 
607 		mstate = new_mstate(ct, LMS_KFAULT);
608 
609 		if (addr < (caddr_t)kernelbase) {
610 			res = pagefault(addr,
611 			    (errcode & PF_ERR_PROT)? F_PROT: F_INVAL, rw, 0);
612 			if (res == FC_NOMAP &&
613 			    addr < p->p_usrstack &&
614 			    grow(addr))
615 				res = 0;
616 		} else {
617 			res = pagefault(addr,
618 			    (errcode & PF_ERR_PROT)? F_PROT: F_INVAL, rw, 1);
619 		}
620 		(void) new_mstate(ct, mstate);
621 
622 		/*
623 		 * Restore lofault. If we resolved the fault, exit.
624 		 * If we didn't and lofault wasn't set, die.
625 		 */
626 		ct->t_lofault = lofault;
627 		if (res == 0)
628 			goto cleanup;
629 
630 #if defined(OPTERON_ERRATUM_93) && defined(_LP64)
631 		if (lofault == 0 && opteron_erratum_93) {
632 			/*
633 			 * Workaround for Opteron Erratum 93. On return from
634 			 * a System Managment Interrupt at a HLT instruction
635 			 * the %rip might be truncated to a 32 bit value.
636 			 * BIOS is supposed to fix this, but some don't.
637 			 * If this occurs we simply restore the high order bits.
638 			 * The HLT instruction is 1 byte of 0xf4.
639 			 */
640 			uintptr_t	rip = rp->r_pc;
641 
642 			if ((rip & 0xfffffffful) == rip) {
643 				rip |= 0xfffffffful << 32;
644 				if (hat_getpfnum(kas.a_hat, (caddr_t)rip) !=
645 				    PFN_INVALID &&
646 				    (*(uchar_t *)rip == 0xf4 ||
647 				    *(uchar_t *)(rip - 1) == 0xf4)) {
648 					rp->r_pc = rip;
649 					goto cleanup;
650 				}
651 			}
652 		}
653 #endif /* OPTERON_ERRATUM_93 && _LP64 */
654 
655 #ifdef OPTERON_ERRATUM_91
656 		if (lofault == 0 && opteron_erratum_91) {
657 			/*
658 			 * Workaround for Opteron Erratum 91. Prefetches may
659 			 * generate a page fault (they're not supposed to do
660 			 * that!). If this occurs we simply return back to the
661 			 * instruction.
662 			 */
663 			caddr_t		pc = (caddr_t)rp->r_pc;
664 
665 			/*
666 			 * If the faulting PC is not mapped, this is a
667 			 * legitimate kernel page fault that must result in a
668 			 * panic. If the faulting PC is mapped, it could contain
669 			 * a prefetch instruction. Check for that here.
670 			 */
671 			if (hat_getpfnum(kas.a_hat, pc) != PFN_INVALID) {
672 				if (cmp_to_prefetch((uchar_t *)pc)) {
673 #ifdef DEBUG
674 					cmn_err(CE_WARN, "Opteron erratum 91 "
675 					    "occurred: kernel prefetch"
676 					    " at %p generated a page fault!",
677 					    (void *)rp->r_pc);
678 #endif /* DEBUG */
679 					goto cleanup;
680 				}
681 			}
682 			(void) die(type, rp, addr, cpuid);
683 		}
684 #endif /* OPTERON_ERRATUM_91 */
685 
686 		if (lofault == 0)
687 			(void) die(type, rp, addr, cpuid);
688 
689 		/*
690 		 * Cannot resolve fault.  Return to lofault.
691 		 */
692 		if (lodebug) {
693 			showregs(type, rp, addr);
694 			traceregs(rp);
695 		}
696 		if (FC_CODE(res) == FC_OBJERR)
697 			res = FC_ERRNO(res);
698 		else
699 			res = EFAULT;
700 		rp->r_r0 = res;
701 		rp->r_pc = ct->t_lofault;
702 		goto cleanup;
703 
704 	case T_PGFLT + USER:	/* user page fault */
705 		if (faultdebug) {
706 			char *fault_str;
707 
708 			switch (rw) {
709 			case S_READ:
710 				fault_str = "read";
711 				break;
712 			case S_WRITE:
713 				fault_str = "write";
714 				break;
715 			case S_EXEC:
716 				fault_str = "exec";
717 				break;
718 			default:
719 				fault_str = "";
720 				break;
721 			}
722 			printf("user %s fault:  addr=0x%lx errcode=0x%x\n",
723 			    fault_str, (uintptr_t)addr, errcode);
724 		}
725 
726 #if defined(OPTERON_ERRATUM_100) && defined(_LP64)
727 		/*
728 		 * Workaround for AMD erratum 100
729 		 *
730 		 * A 32-bit process may receive a page fault on a non
731 		 * 32-bit address by mistake. The range of the faulting
732 		 * address will be
733 		 *
734 		 *	0xffffffff80000000 .. 0xffffffffffffffff or
735 		 *	0x0000000100000000 .. 0x000000017fffffff
736 		 *
737 		 * The fault is always due to an instruction fetch, however
738 		 * the value of r_pc should be correct (in 32 bit range),
739 		 * so we ignore the page fault on the bogus address.
740 		 */
741 		if (p->p_model == DATAMODEL_ILP32 &&
742 		    (0xffffffff80000000 <= (uintptr_t)addr ||
743 		    (0x100000000 <= (uintptr_t)addr &&
744 		    (uintptr_t)addr <= 0x17fffffff))) {
745 			if (!opteron_erratum_100)
746 				panic("unexpected erratum #100");
747 			if (rp->r_pc <= 0xffffffff)
748 				goto out;
749 		}
750 #endif /* OPTERON_ERRATUM_100 && _LP64 */
751 
752 		ASSERT(!(curthread->t_flag & T_WATCHPT));
753 		watchpage = (pr_watch_active(p) && pr_is_watchpage(addr, rw));
754 #ifdef __i386
755 		/*
756 		 * In 32-bit mode, the lcall (system call) instruction fetches
757 		 * one word from the stack, at the stack pointer, because of the
758 		 * way the call gate is constructed.  This is a bogus
759 		 * read and should not be counted as a read watchpoint.
760 		 * We work around the problem here by testing to see if
761 		 * this situation applies and, if so, simply jumping to
762 		 * the code in locore.s that fields the system call trap.
763 		 * The registers on the stack are already set up properly
764 		 * due to the match between the call gate sequence and the
765 		 * trap gate sequence.  We just have to adjust the pc.
766 		 */
767 		if (watchpage && addr == (caddr_t)rp->r_sp &&
768 		    rw == S_READ && instr_is_lcall_syscall((caddr_t)rp->r_pc)) {
769 			extern void watch_syscall(void);
770 
771 			rp->r_pc += LCALLSIZE;
772 			watch_syscall();	/* never returns */
773 			/* NOTREACHED */
774 		}
775 #endif /* __i386 */
776 		vaddr = addr;
777 		if (!watchpage || (sz = instr_size(rp, &vaddr, rw)) <= 0)
778 			fault_type = (errcode & PF_ERR_PROT)? F_PROT: F_INVAL;
779 		else if ((watchcode = pr_is_watchpoint(&vaddr, &ta,
780 		    sz, NULL, rw)) != 0) {
781 			if (ta) {
782 				do_watch_step(vaddr, sz, rw,
783 					watchcode, rp->r_pc);
784 				fault_type = F_INVAL;
785 			} else {
786 				bzero(&siginfo, sizeof (siginfo));
787 				siginfo.si_signo = SIGTRAP;
788 				siginfo.si_code = watchcode;
789 				siginfo.si_addr = vaddr;
790 				siginfo.si_trapafter = 0;
791 				siginfo.si_pc = (caddr_t)rp->r_pc;
792 				fault = FLTWATCH;
793 				break;
794 			}
795 		} else {
796 			/* XXX pr_watch_emul() never succeeds (for now) */
797 			if (rw != S_EXEC && pr_watch_emul(rp, vaddr, rw))
798 				goto out;
799 			do_watch_step(vaddr, sz, rw, 0, 0);
800 			fault_type = F_INVAL;
801 		}
802 
803 		res = pagefault(addr, fault_type, rw, 0);
804 
805 		/*
806 		 * If pagefault() succeeded, ok.
807 		 * Otherwise attempt to grow the stack.
808 		 */
809 		if (res == 0 ||
810 		    (res == FC_NOMAP &&
811 		    addr < p->p_usrstack &&
812 		    grow(addr))) {
813 			lwp->lwp_lastfault = FLTPAGE;
814 			lwp->lwp_lastfaddr = addr;
815 			if (prismember(&p->p_fltmask, FLTPAGE)) {
816 				bzero(&siginfo, sizeof (siginfo));
817 				siginfo.si_addr = addr;
818 				(void) stop_on_fault(FLTPAGE, &siginfo);
819 			}
820 			goto out;
821 		} else if (res == FC_PROT && addr < p->p_usrstack &&
822 		    (mmu.pt_nx != 0 && (errcode & PF_ERR_EXEC))) {
823 			report_stack_exec(p, addr);
824 		}
825 
826 #ifdef OPTERON_ERRATUM_91
827 		/*
828 		 * Workaround for Opteron Erratum 91. Prefetches may generate a
829 		 * page fault (they're not supposed to do that!). If this
830 		 * occurs we simply return back to the instruction.
831 		 *
832 		 * We rely on copyin to properly fault in the page with r_pc.
833 		 */
834 		if (opteron_erratum_91 &&
835 		    addr != (caddr_t)rp->r_pc &&
836 		    instr_is_prefetch((caddr_t)rp->r_pc)) {
837 #ifdef DEBUG
838 			cmn_err(CE_WARN, "Opteron erratum 91 occurred: "
839 			    "prefetch at %p in pid %d generated a trap!",
840 			    (void *)rp->r_pc, p->p_pid);
841 #endif /* DEBUG */
842 			goto out;
843 		}
844 #endif /* OPTERON_ERRATUM_91 */
845 
846 		if (tudebug)
847 			showregs(type, rp, addr);
848 		/*
849 		 * In the case where both pagefault and grow fail,
850 		 * set the code to the value provided by pagefault.
851 		 * We map all errors returned from pagefault() to SIGSEGV.
852 		 */
853 		bzero(&siginfo, sizeof (siginfo));
854 		siginfo.si_addr = addr;
855 		switch (FC_CODE(res)) {
856 		case FC_HWERR:
857 		case FC_NOSUPPORT:
858 			siginfo.si_signo = SIGBUS;
859 			siginfo.si_code = BUS_ADRERR;
860 			fault = FLTACCESS;
861 			break;
862 		case FC_ALIGN:
863 			siginfo.si_signo = SIGBUS;
864 			siginfo.si_code = BUS_ADRALN;
865 			fault = FLTACCESS;
866 			break;
867 		case FC_OBJERR:
868 			if ((siginfo.si_errno = FC_ERRNO(res)) != EINTR) {
869 				siginfo.si_signo = SIGBUS;
870 				siginfo.si_code = BUS_OBJERR;
871 				fault = FLTACCESS;
872 			}
873 			break;
874 		default:	/* FC_NOMAP or FC_PROT */
875 			siginfo.si_signo = SIGSEGV;
876 			siginfo.si_code =
877 			    (res == FC_NOMAP)? SEGV_MAPERR : SEGV_ACCERR;
878 			fault = FLTBOUNDS;
879 			break;
880 		}
881 		break;
882 
883 	case T_ILLINST + USER:	/* invalid opcode fault */
884 		/*
885 		 * If the syscall instruction is disabled due to LDT usage, a
886 		 * user program that attempts to execute it will trigger a #ud
887 		 * trap. Check for that case here. If this occurs on a CPU which
888 		 * doesn't even support syscall, the result of all of this will
889 		 * be to emulate that particular instruction.
890 		 */
891 		if (p->p_ldt != NULL &&
892 		    ldt_rewrite_syscall(rp, p, X86_ASYSC))
893 			goto out;
894 
895 #ifdef __amd64
896 		/*
897 		 * Emulate the LAHF and SAHF instructions if needed.
898 		 * See the instr_is_lsahf function for details.
899 		 */
900 		if (p->p_model == DATAMODEL_LP64 &&
901 		    instr_is_lsahf((caddr_t)rp->r_pc, &instr)) {
902 			emulate_lsahf(rp, instr);
903 			goto out;
904 		}
905 #endif
906 
907 		/*FALLTHROUGH*/
908 
909 		if (tudebug)
910 			showregs(type, rp, (caddr_t)0);
911 		siginfo.si_signo = SIGILL;
912 		siginfo.si_code  = ILL_ILLOPC;
913 		siginfo.si_addr  = (caddr_t)rp->r_pc;
914 		fault = FLTILL;
915 		break;
916 
917 	case T_ZERODIV + USER:		/* integer divide by zero */
918 		if (tudebug && tudebugfpe)
919 			showregs(type, rp, (caddr_t)0);
920 		siginfo.si_signo = SIGFPE;
921 		siginfo.si_code  = FPE_INTDIV;
922 		siginfo.si_addr  = (caddr_t)rp->r_pc;
923 		fault = FLTIZDIV;
924 		break;
925 
926 	case T_OVFLW + USER:	/* integer overflow */
927 		if (tudebug && tudebugfpe)
928 			showregs(type, rp, (caddr_t)0);
929 		siginfo.si_signo = SIGFPE;
930 		siginfo.si_code  = FPE_INTOVF;
931 		siginfo.si_addr  = (caddr_t)rp->r_pc;
932 		fault = FLTIOVF;
933 		break;
934 
935 	case T_NOEXTFLT + USER:	/* math coprocessor not available */
936 		if (tudebug && tudebugfpe)
937 			showregs(type, rp, addr);
938 		if (fpnoextflt(rp)) {
939 			siginfo.si_signo = SIGFPE;
940 			siginfo.si_code  = ILL_ILLOPC;
941 			siginfo.si_addr  = (caddr_t)rp->r_pc;
942 			fault = FLTFPE;
943 		}
944 		break;
945 
946 	case T_EXTOVRFLT:	/* extension overrun fault */
947 		/* check if we took a kernel trap on behalf of user */
948 		{
949 			extern  void ndptrap_frstor(void);
950 			if (rp->r_pc != (uintptr_t)ndptrap_frstor)
951 				(void) die(type, rp, addr, cpuid);
952 			type |= USER;
953 		}
954 		/*FALLTHROUGH*/
955 	case T_EXTOVRFLT + USER:	/* extension overrun fault */
956 		if (tudebug && tudebugfpe)
957 			showregs(type, rp, addr);
958 		if (fpextovrflt(rp)) {
959 			siginfo.si_signo = SIGSEGV;
960 			siginfo.si_code  = SEGV_MAPERR;
961 			siginfo.si_addr  = (caddr_t)rp->r_pc;
962 			fault = FLTBOUNDS;
963 		}
964 		break;
965 
966 	case T_EXTERRFLT:	/* x87 floating point exception pending */
967 		/* check if we took a kernel trap on behalf of user */
968 		{
969 			extern  void ndptrap_frstor(void);
970 			if (rp->r_pc != (uintptr_t)ndptrap_frstor)
971 				(void) die(type, rp, addr, cpuid);
972 			type |= USER;
973 		}
974 		/*FALLTHROUGH*/
975 
976 	case T_EXTERRFLT + USER: /* x87 floating point exception pending */
977 		if (tudebug && tudebugfpe)
978 			showregs(type, rp, addr);
979 		if (sicode = fpexterrflt(rp)) {
980 			siginfo.si_signo = SIGFPE;
981 			siginfo.si_code  = sicode;
982 			siginfo.si_addr  = (caddr_t)rp->r_pc;
983 			fault = FLTFPE;
984 		}
985 		break;
986 
987 	case T_SIMDFPE + USER:		/* SSE and SSE2 exceptions */
988 		if (tudebug && tudebugsse)
989 			showregs(type, rp, addr);
990 		if ((x86_feature & (X86_SSE|X86_SSE2)) == 0) {
991 			/*
992 			 * There are rumours that some user instructions
993 			 * on older CPUs can cause this trap to occur; in
994 			 * which case send a SIGILL instead of a SIGFPE.
995 			 */
996 			siginfo.si_signo = SIGILL;
997 			siginfo.si_code  = ILL_ILLTRP;
998 			siginfo.si_addr  = (caddr_t)rp->r_pc;
999 			siginfo.si_trapno = type & ~USER;
1000 			fault = FLTILL;
1001 		} else if ((sicode = fpsimderrflt(rp)) != 0) {
1002 			siginfo.si_signo = SIGFPE;
1003 			siginfo.si_code = sicode;
1004 			siginfo.si_addr = (caddr_t)rp->r_pc;
1005 			fault = FLTFPE;
1006 		}
1007 		break;
1008 
1009 	case T_BPTFLT:	/* breakpoint trap */
1010 		/*
1011 		 * Kernel breakpoint traps should only happen when kmdb is
1012 		 * active, and even then, it'll have interposed on the IDT, so
1013 		 * control won't get here.  If it does, we've hit a breakpoint
1014 		 * without the debugger, which is very strange, and very
1015 		 * fatal.
1016 		 */
1017 		if (tudebug && tudebugbpt)
1018 			showregs(type, rp, (caddr_t)0);
1019 
1020 		(void) die(type, rp, addr, cpuid);
1021 		break;
1022 
1023 	case T_SGLSTP: /* single step/hw breakpoint exception */
1024 
1025 		/* Now evaluate how we got here */
1026 		if (lwp != NULL && (lwp->lwp_pcb.pcb_drstat & DR_SINGLESTEP)) {
1027 			/*
1028 			 * i386 single-steps even through lcalls which
1029 			 * change the privilege level. So we take a trap at
1030 			 * the first instruction in privileged mode.
1031 			 *
1032 			 * Set a flag to indicate that upon completion of
1033 			 * the system call, deal with the single-step trap.
1034 			 *
1035 			 * The same thing happens for sysenter, too.
1036 			 */
1037 			singlestep_twiddle = 0;
1038 			if (rp->r_pc == (uintptr_t)sys_sysenter ||
1039 			    rp->r_pc == (uintptr_t)brand_sys_sysenter) {
1040 				singlestep_twiddle = 1;
1041 #if defined(__amd64)
1042 				/*
1043 				 * Since we are already on the kernel's
1044 				 * %gs, on 64-bit systems the sysenter case
1045 				 * needs to adjust the pc to avoid
1046 				 * executing the swapgs instruction at the
1047 				 * top of the handler.
1048 				 */
1049 				if (rp->r_pc == (uintptr_t)sys_sysenter)
1050 					rp->r_pc = (uintptr_t)
1051 					    _sys_sysenter_post_swapgs;
1052 				else
1053 					rp->r_pc = (uintptr_t)
1054 					    _brand_sys_sysenter_post_swapgs;
1055 #endif
1056 			}
1057 #if defined(__i386)
1058 			else if (rp->r_pc == (uintptr_t)sys_call ||
1059 			    rp->r_pc == (uintptr_t)brand_sys_call) {
1060 				singlestep_twiddle = 1;
1061 			}
1062 #endif
1063 			else {
1064 				/* not on sysenter/syscall; uregs available */
1065 				if (tudebug && tudebugbpt)
1066 					showregs(type, rp, (caddr_t)0);
1067 			}
1068 			if (singlestep_twiddle) {
1069 				rp->r_ps &= ~PS_T; /* turn off trace */
1070 				lwp->lwp_pcb.pcb_flags |= DEBUG_PENDING;
1071 				ct->t_post_sys = 1;
1072 				aston(curthread);
1073 				goto cleanup;
1074 			}
1075 		}
1076 		/* XXX - needs review on debugger interface? */
1077 		if (boothowto & RB_DEBUG)
1078 			debug_enter((char *)NULL);
1079 		else
1080 			(void) die(type, rp, addr, cpuid);
1081 		break;
1082 
1083 	case T_NMIFLT:	/* NMI interrupt */
1084 		printf("Unexpected NMI in system mode\n");
1085 		goto cleanup;
1086 
1087 	case T_NMIFLT + USER:	/* NMI interrupt */
1088 		printf("Unexpected NMI in user mode\n");
1089 		break;
1090 
1091 	case T_GPFLT:	/* general protection violation */
1092 #if defined(__amd64)
1093 		/*
1094 		 * On amd64, we can get a #gp from referencing addresses
1095 		 * in the virtual address hole e.g. from a copyin
1096 		 * or in update_sregs while updating user semgent registers.
1097 		 */
1098 
1099 		/*
1100 		 * If we're under on_trap() protection (see <sys/ontrap.h>),
1101 		 * set ot_trap and longjmp back to the on_trap() call site.
1102 		 */
1103 		if (ct->t_ontrap != NULL) {
1104 			if (ct->t_ontrap->ot_prot & OT_DATA_ACCESS)
1105 				ct->t_ontrap->ot_trap |= OT_DATA_ACCESS;
1106 
1107 			if (ct->t_ontrap->ot_prot & OT_SEGMENT_ACCESS)
1108 				ct->t_ontrap->ot_trap |= OT_SEGMENT_ACCESS;
1109 			longjmp(&curthread->t_ontrap->ot_jmpbuf);
1110 		}
1111 
1112 		/*
1113 		 * If we're under lofault protection (copyin etc.),
1114 		 * longjmp back to lofault with an EFAULT.
1115 		 */
1116 		if (ct->t_lofault) {
1117 			/*
1118 			 * Fault is not resolvable, so just return to lofault
1119 			 */
1120 			if (lodebug) {
1121 				showregs(type, rp, addr);
1122 				traceregs(rp);
1123 			}
1124 			rp->r_r0 = EFAULT;
1125 			rp->r_pc = ct->t_lofault;
1126 			goto cleanup;
1127 		}
1128 		/*FALLTHROUGH*/
1129 #endif
1130 	case T_SEGFLT:	/* segment not present fault */
1131 #if defined(__amd64)
1132 		/*
1133 		 * One example of this is #NP in update_sregs while
1134 		 * attempting to update a user segment register
1135 		 * that points to a descriptor that is marked not
1136 		 * present.
1137 		 */
1138 		if (ct->t_ontrap != NULL &&
1139 		    ct->t_ontrap->ot_prot & OT_SEGMENT_ACCESS) {
1140 			ct->t_ontrap->ot_trap |= OT_SEGMENT_ACCESS;
1141 			longjmp(&curthread->t_ontrap->ot_jmpbuf);
1142 		}
1143 #endif	/* __amd64 */
1144 		/*FALLTHROUGH*/
1145 	case T_STKFLT:	/* stack fault */
1146 	case T_TSSFLT:	/* invalid TSS fault */
1147 		if (tudebug)
1148 			showregs(type, rp, (caddr_t)0);
1149 		if (kern_gpfault(rp))
1150 			(void) die(type, rp, addr, cpuid);
1151 		goto cleanup;
1152 		/*FALLTHROUGH*/
1153 
1154 	/*
1155 	 * ONLY 32-bit PROCESSES can USE a PRIVATE LDT! 64-bit apps
1156 	 * should have no need for them, so we put a stop to it here.
1157 	 *
1158 	 * So: not-present fault is ONLY valid for 32-bit processes with
1159 	 * a private LDT trying to do a system call. Emulate it.
1160 	 *
1161 	 * #gp fault is ONLY valid for 32-bit processes also, which DO NOT
1162 	 * have a private LDT, and are trying to do a system call. Emulate it.
1163 	 */
1164 
1165 	case T_SEGFLT + USER:	/* segment not present fault */
1166 	case T_GPFLT + USER:	/* general protection violation */
1167 #ifdef _SYSCALL32_IMPL
1168 		if (p->p_model != DATAMODEL_NATIVE) {
1169 #endif /* _SYSCALL32_IMPL */
1170 		if (instr_is_lcall_syscall((caddr_t)rp->r_pc)) {
1171 			if (type == T_SEGFLT + USER)
1172 				ASSERT(p->p_ldt != NULL);
1173 
1174 			if ((p->p_ldt == NULL && type == T_GPFLT + USER) ||
1175 			    type == T_SEGFLT + USER) {
1176 
1177 			/*
1178 			 * The user attempted a system call via the obsolete
1179 			 * call gate mechanism. Because the process doesn't have
1180 			 * an LDT (i.e. the ldtr contains 0), a #gp results.
1181 			 * Emulate the syscall here, just as we do above for a
1182 			 * #np trap.
1183 			 */
1184 
1185 			/*
1186 			 * Since this is a not-present trap, rp->r_pc points to
1187 			 * the trapping lcall instruction. We need to bump it
1188 			 * to the next insn so the app can continue on.
1189 			 */
1190 			rp->r_pc += LCALLSIZE;
1191 			lwp->lwp_regs = rp;
1192 
1193 			/*
1194 			 * Normally the microstate of the LWP is forced back to
1195 			 * LMS_USER by the syscall handlers. Emulate that
1196 			 * behavior here.
1197 			 */
1198 			mstate = LMS_USER;
1199 
1200 			dosyscall();
1201 			goto out;
1202 			}
1203 		}
1204 #ifdef _SYSCALL32_IMPL
1205 		}
1206 #endif /* _SYSCALL32_IMPL */
1207 		/*
1208 		 * If the current process is using a private LDT and the
1209 		 * trapping instruction is sysenter, the sysenter instruction
1210 		 * has been disabled on the CPU because it destroys segment
1211 		 * registers. If this is the case, rewrite the instruction to
1212 		 * be a safe system call and retry it. If this occurs on a CPU
1213 		 * which doesn't even support sysenter, the result of all of
1214 		 * this will be to emulate that particular instruction.
1215 		 */
1216 		if (p->p_ldt != NULL &&
1217 		    ldt_rewrite_syscall(rp, p, X86_SEP))
1218 			goto out;
1219 
1220 		/*FALLTHROUGH*/
1221 
1222 	case T_BOUNDFLT + USER:	/* bound fault */
1223 	case T_STKFLT + USER:	/* stack fault */
1224 	case T_TSSFLT + USER:	/* invalid TSS fault */
1225 		if (tudebug)
1226 			showregs(type, rp, (caddr_t)0);
1227 		siginfo.si_signo = SIGSEGV;
1228 		siginfo.si_code  = SEGV_MAPERR;
1229 		siginfo.si_addr  = (caddr_t)rp->r_pc;
1230 		fault = FLTBOUNDS;
1231 		break;
1232 
1233 	case T_ALIGNMENT + USER:	/* user alignment error (486) */
1234 		if (tudebug)
1235 			showregs(type, rp, (caddr_t)0);
1236 		bzero(&siginfo, sizeof (siginfo));
1237 		siginfo.si_signo = SIGBUS;
1238 		siginfo.si_code = BUS_ADRALN;
1239 		siginfo.si_addr = (caddr_t)rp->r_pc;
1240 		fault = FLTACCESS;
1241 		break;
1242 
1243 	case T_SGLSTP + USER: /* single step/hw breakpoint exception */
1244 		if (tudebug && tudebugbpt)
1245 			showregs(type, rp, (caddr_t)0);
1246 
1247 		/* Was it single-stepping? */
1248 		if (lwp->lwp_pcb.pcb_drstat & DR_SINGLESTEP) {
1249 			pcb_t *pcb = &lwp->lwp_pcb;
1250 
1251 			rp->r_ps &= ~PS_T;
1252 			/*
1253 			 * If both NORMAL_STEP and WATCH_STEP are in effect,
1254 			 * give precedence to WATCH_STEP.  If neither is set,
1255 			 * user must have set the PS_T bit in %efl; treat this
1256 			 * as NORMAL_STEP.
1257 			 */
1258 			if ((fault = undo_watch_step(&siginfo)) == 0 &&
1259 			    ((pcb->pcb_flags & NORMAL_STEP) ||
1260 			    !(pcb->pcb_flags & WATCH_STEP))) {
1261 				siginfo.si_signo = SIGTRAP;
1262 				siginfo.si_code = TRAP_TRACE;
1263 				siginfo.si_addr = (caddr_t)rp->r_pc;
1264 				fault = FLTTRACE;
1265 			}
1266 			pcb->pcb_flags &= ~(NORMAL_STEP|WATCH_STEP);
1267 		} else {
1268 			cmn_err(CE_WARN,
1269 			    "Unexpected INT 1 in user mode, dr6=%lx",
1270 			    lwp->lwp_pcb.pcb_drstat);
1271 		}
1272 		break;
1273 
1274 	case T_BPTFLT + USER:	/* breakpoint trap */
1275 		if (tudebug && tudebugbpt)
1276 			showregs(type, rp, (caddr_t)0);
1277 		/*
1278 		 * int 3 (the breakpoint instruction) leaves the pc referring
1279 		 * to the address one byte after the breakpointed address.
1280 		 * If the P_PR_BPTADJ flag has been set via /proc, We adjust
1281 		 * it back so it refers to the breakpointed address.
1282 		 */
1283 		if (p->p_proc_flag & P_PR_BPTADJ)
1284 			rp->r_pc--;
1285 		siginfo.si_signo = SIGTRAP;
1286 		siginfo.si_code  = TRAP_BRKPT;
1287 		siginfo.si_addr  = (caddr_t)rp->r_pc;
1288 		fault = FLTBPT;
1289 		break;
1290 
1291 	case T_AST:
1292 		/*
1293 		 * This occurs only after the cs register has been made to
1294 		 * look like a kernel selector, either through debugging or
1295 		 * possibly by functions like setcontext().  The thread is
1296 		 * about to cause a general protection fault at common_iret()
1297 		 * in locore.  We let that happen immediately instead of
1298 		 * doing the T_AST processing.
1299 		 */
1300 		goto cleanup;
1301 
1302 	case T_AST + USER:		/* profiling or resched pseudo trap */
1303 		if (lwp->lwp_pcb.pcb_flags & CPC_OVERFLOW) {
1304 			lwp->lwp_pcb.pcb_flags &= ~CPC_OVERFLOW;
1305 			if (kcpc_overflow_ast()) {
1306 				/*
1307 				 * Signal performance counter overflow
1308 				 */
1309 				if (tudebug)
1310 					showregs(type, rp, (caddr_t)0);
1311 				bzero(&siginfo, sizeof (siginfo));
1312 				siginfo.si_signo = SIGEMT;
1313 				siginfo.si_code = EMT_CPCOVF;
1314 				siginfo.si_addr = (caddr_t)rp->r_pc;
1315 				fault = FLTCPCOVF;
1316 			}
1317 		}
1318 		break;
1319 	}
1320 
1321 	/*
1322 	 * We can't get here from a system trap
1323 	 */
1324 	ASSERT(type & USER);
1325 
1326 	if (fault) {
1327 		/* We took a fault so abort single step. */
1328 		lwp->lwp_pcb.pcb_flags &= ~(NORMAL_STEP|WATCH_STEP);
1329 		/*
1330 		 * Remember the fault and fault adddress
1331 		 * for real-time (SIGPROF) profiling.
1332 		 */
1333 		lwp->lwp_lastfault = fault;
1334 		lwp->lwp_lastfaddr = siginfo.si_addr;
1335 
1336 		DTRACE_PROC2(fault, int, fault, ksiginfo_t *, &siginfo);
1337 
1338 		/*
1339 		 * If a debugger has declared this fault to be an
1340 		 * event of interest, stop the lwp.  Otherwise just
1341 		 * deliver the associated signal.
1342 		 */
1343 		if (siginfo.si_signo != SIGKILL &&
1344 		    prismember(&p->p_fltmask, fault) &&
1345 		    stop_on_fault(fault, &siginfo) == 0)
1346 			siginfo.si_signo = 0;
1347 	}
1348 
1349 	if (siginfo.si_signo)
1350 		trapsig(&siginfo, (fault == FLTCPCOVF)? 0 : 1);
1351 
1352 	if (lwp->lwp_oweupc)
1353 		profil_tick(rp->r_pc);
1354 
1355 	if (ct->t_astflag | ct->t_sig_check) {
1356 		/*
1357 		 * Turn off the AST flag before checking all the conditions that
1358 		 * may have caused an AST.  This flag is on whenever a signal or
1359 		 * unusual condition should be handled after the next trap or
1360 		 * syscall.
1361 		 */
1362 		astoff(ct);
1363 		/*
1364 		 * If a single-step trap occurred on a syscall (see above)
1365 		 * recognize it now.  Do this before checking for signals
1366 		 * because deferred_singlestep_trap() may generate a SIGTRAP to
1367 		 * the LWP or may otherwise mark the LWP to call issig(FORREAL).
1368 		 */
1369 		if (lwp->lwp_pcb.pcb_flags & DEBUG_PENDING)
1370 			deferred_singlestep_trap((caddr_t)rp->r_pc);
1371 
1372 		ct->t_sig_check = 0;
1373 
1374 		mutex_enter(&p->p_lock);
1375 		if (curthread->t_proc_flag & TP_CHANGEBIND) {
1376 			timer_lwpbind();
1377 			curthread->t_proc_flag &= ~TP_CHANGEBIND;
1378 		}
1379 		mutex_exit(&p->p_lock);
1380 
1381 		/*
1382 		 * for kaio requests that are on the per-process poll queue,
1383 		 * aiop->aio_pollq, they're AIO_POLL bit is set, the kernel
1384 		 * should copyout their result_t to user memory. by copying
1385 		 * out the result_t, the user can poll on memory waiting
1386 		 * for the kaio request to complete.
1387 		 */
1388 		if (p->p_aio)
1389 			aio_cleanup(0);
1390 		/*
1391 		 * If this LWP was asked to hold, call holdlwp(), which will
1392 		 * stop.  holdlwps() sets this up and calls pokelwps() which
1393 		 * sets the AST flag.
1394 		 *
1395 		 * Also check TP_EXITLWP, since this is used by fresh new LWPs
1396 		 * through lwp_rtt().  That flag is set if the lwp_create(2)
1397 		 * syscall failed after creating the LWP.
1398 		 */
1399 		if (ISHOLD(p))
1400 			holdlwp();
1401 
1402 		/*
1403 		 * All code that sets signals and makes ISSIG evaluate true must
1404 		 * set t_astflag afterwards.
1405 		 */
1406 		if (ISSIG_PENDING(ct, lwp, p)) {
1407 			if (issig(FORREAL))
1408 				psig();
1409 			ct->t_sig_check = 1;
1410 		}
1411 
1412 		if (ct->t_rprof != NULL) {
1413 			realsigprof(0, 0);
1414 			ct->t_sig_check = 1;
1415 		}
1416 
1417 		/*
1418 		 * /proc can't enable/disable the trace bit itself
1419 		 * because that could race with the call gate used by
1420 		 * system calls via "lcall". If that happened, an
1421 		 * invalid EFLAGS would result. prstep()/prnostep()
1422 		 * therefore schedule an AST for the purpose.
1423 		 */
1424 		if (lwp->lwp_pcb.pcb_flags & REQUEST_STEP) {
1425 			lwp->lwp_pcb.pcb_flags &= ~REQUEST_STEP;
1426 			rp->r_ps |= PS_T;
1427 		}
1428 		if (lwp->lwp_pcb.pcb_flags & REQUEST_NOSTEP) {
1429 			lwp->lwp_pcb.pcb_flags &= ~REQUEST_NOSTEP;
1430 			rp->r_ps &= ~PS_T;
1431 		}
1432 	}
1433 
1434 out:	/* We can't get here from a system trap */
1435 	ASSERT(type & USER);
1436 
1437 	if (ISHOLD(p))
1438 		holdlwp();
1439 
1440 	/*
1441 	 * Set state to LWP_USER here so preempt won't give us a kernel
1442 	 * priority if it occurs after this point.  Call CL_TRAPRET() to
1443 	 * restore the user-level priority.
1444 	 *
1445 	 * It is important that no locks (other than spinlocks) be entered
1446 	 * after this point before returning to user mode (unless lwp_state
1447 	 * is set back to LWP_SYS).
1448 	 */
1449 	lwp->lwp_state = LWP_USER;
1450 
1451 	if (ct->t_trapret) {
1452 		ct->t_trapret = 0;
1453 		thread_lock(ct);
1454 		CL_TRAPRET(ct);
1455 		thread_unlock(ct);
1456 	}
1457 	if (CPU->cpu_runrun || curthread->t_schedflag & TS_ANYWAITQ)
1458 		preempt();
1459 	(void) new_mstate(ct, mstate);
1460 
1461 	/* Kernel probe */
1462 	TNF_PROBE_1(thread_state, "thread", /* CSTYLED */,
1463 	    tnf_microstate, state, LMS_USER);
1464 
1465 	return;
1466 
1467 cleanup:	/* system traps end up here */
1468 	ASSERT(!(type & USER));
1469 }
1470 
1471 /*
1472  * Patch non-zero to disable preemption of threads in the kernel.
1473  */
1474 int IGNORE_KERNEL_PREEMPTION = 0;	/* XXX - delete this someday */
1475 
1476 struct kpreempt_cnts {		/* kernel preemption statistics */
1477 	int	kpc_idle;	/* executing idle thread */
1478 	int	kpc_intr;	/* executing interrupt thread */
1479 	int	kpc_clock;	/* executing clock thread */
1480 	int	kpc_blocked;	/* thread has blocked preemption (t_preempt) */
1481 	int	kpc_notonproc;	/* thread is surrendering processor */
1482 	int	kpc_inswtch;	/* thread has ratified scheduling decision */
1483 	int	kpc_prilevel;	/* processor interrupt level is too high */
1484 	int	kpc_apreempt;	/* asynchronous preemption */
1485 	int	kpc_spreempt;	/* synchronous preemption */
1486 } kpreempt_cnts;
1487 
1488 /*
1489  * kernel preemption: forced rescheduling, preempt the running kernel thread.
1490  *	the argument is old PIL for an interrupt,
1491  *	or the distingished value KPREEMPT_SYNC.
1492  */
1493 void
1494 kpreempt(int asyncspl)
1495 {
1496 	kthread_t *ct = curthread;
1497 
1498 	if (IGNORE_KERNEL_PREEMPTION) {
1499 		aston(CPU->cpu_dispthread);
1500 		return;
1501 	}
1502 
1503 	/*
1504 	 * Check that conditions are right for kernel preemption
1505 	 */
1506 	do {
1507 		if (ct->t_preempt) {
1508 			/*
1509 			 * either a privileged thread (idle, panic, interrupt)
1510 			 *	or will check when t_preempt is lowered
1511 			 */
1512 			if (ct->t_pri < 0)
1513 				kpreempt_cnts.kpc_idle++;
1514 			else if (ct->t_flag & T_INTR_THREAD) {
1515 				kpreempt_cnts.kpc_intr++;
1516 				if (ct->t_pil == CLOCK_LEVEL)
1517 					kpreempt_cnts.kpc_clock++;
1518 			} else
1519 				kpreempt_cnts.kpc_blocked++;
1520 			aston(CPU->cpu_dispthread);
1521 			return;
1522 		}
1523 		if (ct->t_state != TS_ONPROC ||
1524 		    ct->t_disp_queue != CPU->cpu_disp) {
1525 			/* this thread will be calling swtch() shortly */
1526 			kpreempt_cnts.kpc_notonproc++;
1527 			if (CPU->cpu_thread != CPU->cpu_dispthread) {
1528 				/* already in swtch(), force another */
1529 				kpreempt_cnts.kpc_inswtch++;
1530 				siron();
1531 			}
1532 			return;
1533 		}
1534 		if (getpil() >= DISP_LEVEL) {
1535 			/*
1536 			 * We can't preempt this thread if it is at
1537 			 * a PIL >= DISP_LEVEL since it may be holding
1538 			 * a spin lock (like sched_lock).
1539 			 */
1540 			siron();	/* check back later */
1541 			kpreempt_cnts.kpc_prilevel++;
1542 			return;
1543 		}
1544 		if (!interrupts_enabled()) {
1545 			/*
1546 			 * Can't preempt while running with ints disabled
1547 			 */
1548 			kpreempt_cnts.kpc_prilevel++;
1549 			return;
1550 		}
1551 		if (asyncspl != KPREEMPT_SYNC)
1552 			kpreempt_cnts.kpc_apreempt++;
1553 		else
1554 			kpreempt_cnts.kpc_spreempt++;
1555 
1556 		ct->t_preempt++;
1557 		preempt();
1558 		ct->t_preempt--;
1559 	} while (CPU->cpu_kprunrun);
1560 }
1561 
1562 /*
1563  * Print out debugging info.
1564  */
1565 static void
1566 showregs(uint_t type, struct regs *rp, caddr_t addr)
1567 {
1568 	int s;
1569 
1570 	s = spl7();
1571 	type &= ~USER;
1572 	if (PTOU(curproc)->u_comm[0])
1573 		printf("%s: ", PTOU(curproc)->u_comm);
1574 	if (type < TRAP_TYPES)
1575 		printf("#%s %s\n", trap_type_mnemonic[type], trap_type[type]);
1576 	else
1577 		switch (type) {
1578 		case T_SYSCALL:
1579 			printf("Syscall Trap:\n");
1580 			break;
1581 		case T_AST:
1582 			printf("AST\n");
1583 			break;
1584 		default:
1585 			printf("Bad Trap = %d\n", type);
1586 			break;
1587 		}
1588 	if (type == T_PGFLT) {
1589 		printf("Bad %s fault at addr=0x%lx\n",
1590 		    USERMODE(rp->r_cs) ? "user": "kernel", (uintptr_t)addr);
1591 	} else if (addr) {
1592 		printf("addr=0x%lx\n", (uintptr_t)addr);
1593 	}
1594 
1595 	printf("pid=%d, pc=0x%lx, sp=0x%lx, eflags=0x%lx\n",
1596 	    (ttoproc(curthread) && ttoproc(curthread)->p_pidp) ?
1597 	    ttoproc(curthread)->p_pid : 0, rp->r_pc, rp->r_sp, rp->r_ps);
1598 
1599 #if defined(__lint)
1600 	/*
1601 	 * this clause can be deleted when lint bug 4870403 is fixed
1602 	 * (lint thinks that bit 32 is illegal in a %b format string)
1603 	 */
1604 	printf("cr0: %x cr4: %b\n",
1605 	    (uint_t)getcr0(), (uint_t)getcr4(), FMT_CR4);
1606 #else
1607 	printf("cr0: %b cr4: %b\n",
1608 	    (uint_t)getcr0(), FMT_CR0, (uint_t)getcr4(), FMT_CR4);
1609 #endif	/* __lint */
1610 
1611 #if defined(__amd64)
1612 	printf("cr2: %lx cr3: %lx cr8: %lx\n", getcr2(), getcr3(), getcr8());
1613 #elif defined(__i386)
1614 	printf("cr2: %lx cr3: %lx\n", getcr2(), getcr3());
1615 #endif
1616 
1617 	dumpregs(rp);
1618 	splx(s);
1619 }
1620 
1621 static void
1622 dumpregs(struct regs *rp)
1623 {
1624 #if defined(__amd64)
1625 	const char fmt[] = "\t%3s: %16lx %3s: %16lx %3s: %16lx\n";
1626 
1627 	printf(fmt, "rdi", rp->r_rdi, "rsi", rp->r_rsi, "rdx", rp->r_rdx);
1628 	printf(fmt, "rcx", rp->r_rcx, " r8", rp->r_r8, " r9", rp->r_r9);
1629 	printf(fmt, "rax", rp->r_rax, "rbx", rp->r_rbx, "rbp", rp->r_rbp);
1630 	printf(fmt, "r10", rp->r_r10, "r11", rp->r_r11, "r12", rp->r_r12);
1631 	printf(fmt, "r13", rp->r_r13, "r14", rp->r_r14, "r15", rp->r_r15);
1632 
1633 	printf(fmt, "fsb", rdmsr(MSR_AMD_FSBASE), "gsb", rdmsr(MSR_AMD_GSBASE),
1634 	    " ds", rp->r_ds);
1635 	printf(fmt, " es", rp->r_es, " fs", rp->r_fs, " gs", rp->r_gs);
1636 
1637 	printf(fmt, "trp", rp->r_trapno, "err", rp->r_err, "rip", rp->r_rip);
1638 	printf(fmt, " cs", rp->r_cs, "rfl", rp->r_rfl, "rsp", rp->r_rsp);
1639 
1640 	printf("\t%3s: %16lx\n", " ss", rp->r_ss);
1641 
1642 #elif defined(__i386)
1643 	const char fmt[] = "\t%3s: %8lx %3s: %8lx %3s: %8lx %3s: %8lx\n";
1644 
1645 	printf(fmt, " gs", rp->r_gs, " fs", rp->r_fs,
1646 	    " es", rp->r_es, " ds", rp->r_ds);
1647 	printf(fmt, "edi", rp->r_edi, "esi", rp->r_esi,
1648 	    "ebp", rp->r_ebp, "esp", rp->r_esp);
1649 	printf(fmt, "ebx", rp->r_ebx, "edx", rp->r_edx,
1650 	    "ecx", rp->r_ecx, "eax", rp->r_eax);
1651 	printf(fmt, "trp", rp->r_trapno, "err", rp->r_err,
1652 	    "eip", rp->r_eip, " cs", rp->r_cs);
1653 	printf("\t%3s: %8lx %3s: %8lx %3s: %8lx\n",
1654 	    "efl", rp->r_efl, "usp", rp->r_uesp, " ss", rp->r_ss);
1655 
1656 #endif	/* __i386 */
1657 }
1658 
1659 /*
1660  * Handle #gp faults in kernel mode.
1661  *
1662  * One legitimate way this can happen is if we attempt to update segment
1663  * registers to naughty values on the way out of the kernel.
1664  *
1665  * This can happen in a couple of ways: someone - either accidentally or
1666  * on purpose - creates (setcontext(2), lwp_create(2)) or modifies
1667  * (signal(2)) a ucontext that contains silly segment register values.
1668  * Or someone - either accidentally or on purpose - modifies the prgregset_t
1669  * of a subject process via /proc to contain silly segment register values.
1670  *
1671  * (The unfortunate part is that we can end up discovering the bad segment
1672  * register value in the middle of an 'iret' after we've popped most of the
1673  * stack.  So it becomes quite difficult to associate an accurate ucontext
1674  * with the lwp, because the act of taking the #gp trap overwrites most of
1675  * what we were going to send the lwp.)
1676  *
1677  * OTOH if it turns out that's -not- the problem, and we're -not- an lwp
1678  * trying to return to user mode and we get a #gp fault, then we need
1679  * to die() -- which will happen if we return non-zero from this routine.
1680  */
1681 static int
1682 kern_gpfault(struct regs *rp)
1683 {
1684 	kthread_t *t = curthread;
1685 	proc_t *p = ttoproc(t);
1686 	klwp_t *lwp = ttolwp(t);
1687 	struct regs tmpregs, *trp = NULL;
1688 	caddr_t pc = (caddr_t)rp->r_pc;
1689 	int v;
1690 
1691 	extern void _sys_rtt(), sr_sup();
1692 
1693 #if defined(__amd64)
1694 	static const uint8_t iretq_insn[2] = { 0x48, 0xcf };
1695 
1696 #elif defined(__i386)
1697 	static const uint8_t iret_insn[1] = { 0xcf };
1698 
1699 	/*
1700 	 * Note carefully the appallingly awful dependency between
1701 	 * the instruction sequence used in __SEGREGS_POP and these
1702 	 * instructions encoded here.
1703 	 *
1704 	 * XX64	Add some commentary to locore.s/privregs.h to document this.
1705 	 */
1706 	static const uint8_t movw_0_esp_gs[4] = { 0x8e, 0x6c, 0x24, 0x0 };
1707 	static const uint8_t movw_4_esp_fs[4] = { 0x8e, 0x64, 0x24, 0x4 };
1708 	static const uint8_t movw_8_esp_es[4] = { 0x8e, 0x44, 0x24, 0x8 };
1709 	static const uint8_t movw_c_esp_ds[4] = { 0x8e, 0x5c, 0x24, 0xc };
1710 #endif
1711 	/*
1712 	 * if we're not an lwp, or the pc range is outside _sys_rtt, then
1713 	 * we should immediately be die()ing horribly
1714 	 */
1715 	if (lwp == NULL ||
1716 	    (uintptr_t)pc < (uintptr_t)_sys_rtt ||
1717 	    (uintptr_t)pc > (uintptr_t)sr_sup)
1718 		return (1);
1719 
1720 	/*
1721 	 * So at least we're in the right part of the kernel.
1722 	 *
1723 	 * Disassemble the instruction at the faulting pc.
1724 	 * Once we know what it is, we carefully reconstruct the stack
1725 	 * based on the order in which the stack is deconstructed in
1726 	 * _sys_rtt. Ew.
1727 	 */
1728 
1729 #if defined(__amd64)
1730 
1731 	if (bcmp(pc, iretq_insn, sizeof (iretq_insn)) == 0) {
1732 		/*
1733 		 * We took the #gp while trying to perform the iretq.
1734 		 * This means that either %cs or %ss are bad.
1735 		 * All we know for sure is that most of the general
1736 		 * registers have been restored, including the
1737 		 * segment registers, and all we have left on the
1738 		 * topmost part of the lwp's stack are the
1739 		 * registers that the iretq was unable to consume.
1740 		 *
1741 		 * All the rest of the state was crushed by the #gp
1742 		 * which pushed -its- registers atop our old save area
1743 		 * (because we had to decrement the stack pointer, sigh) so
1744 		 * all that we can try and do is to reconstruct the
1745 		 * crushed frame from the #gp trap frame itself.
1746 		 */
1747 		trp = &tmpregs;
1748 		trp->r_ss = lwptoregs(lwp)->r_ss;
1749 		trp->r_sp = lwptoregs(lwp)->r_sp;
1750 		trp->r_ps = lwptoregs(lwp)->r_ps;
1751 		trp->r_cs = lwptoregs(lwp)->r_cs;
1752 		trp->r_pc = lwptoregs(lwp)->r_pc;
1753 		bcopy(rp, trp, offsetof(struct regs, r_pc));
1754 
1755 		/*
1756 		 * Validate simple math
1757 		 */
1758 		ASSERT(trp->r_pc == lwptoregs(lwp)->r_pc);
1759 		ASSERT(trp->r_err == rp->r_err);
1760 
1761 	}
1762 
1763 #elif defined(__i386)
1764 
1765 	if (bcmp(pc, iret_insn, sizeof (iret_insn)) == 0) {
1766 		/*
1767 		 * We took the #gp while trying to perform the iret.
1768 		 * This means that either %cs or %ss are bad.
1769 		 * All we know for sure is that most of the general
1770 		 * registers have been restored, including the
1771 		 * segment registers, and all we have left on the
1772 		 * topmost part of the lwp's stack are the registers that
1773 		 * the iret was unable to consume.
1774 		 *
1775 		 * All the rest of the state was crushed by the #gp
1776 		 * which pushed -its- registers atop our old save area
1777 		 * (because we had to decrement the stack pointer, sigh) so
1778 		 * all that we can try and do is to reconstruct the
1779 		 * crushed frame from the #gp trap frame itself.
1780 		 */
1781 		trp = &tmpregs;
1782 		trp->r_ss = lwptoregs(lwp)->r_ss;
1783 		trp->r_sp = lwptoregs(lwp)->r_sp;
1784 		trp->r_ps = lwptoregs(lwp)->r_ps;
1785 		trp->r_cs = lwptoregs(lwp)->r_cs;
1786 		trp->r_pc = lwptoregs(lwp)->r_pc;
1787 		bcopy(rp, trp, offsetof(struct regs, r_pc));
1788 
1789 		ASSERT(trp->r_pc == lwptoregs(lwp)->r_pc);
1790 		ASSERT(trp->r_err == rp->r_err);
1791 
1792 	} else {
1793 		/*
1794 		 * Segment registers are reloaded in _sys_rtt
1795 		 * via the following sequence:
1796 		 *
1797 		 *	movw	0(%esp), %gs
1798 		 *	movw	4(%esp), %fs
1799 		 *	movw	8(%esp), %es
1800 		 *	movw	12(%esp), %ds
1801 		 *	addl	$16, %esp
1802 		 *
1803 		 * Thus if any of them fault, we know the user
1804 		 * registers are left unharmed on the stack.
1805 		 */
1806 		if (bcmp(pc, movw_0_esp_gs, sizeof (movw_0_esp_gs)) == 0 ||
1807 		    bcmp(pc, movw_4_esp_fs, sizeof (movw_4_esp_fs)) == 0 ||
1808 		    bcmp(pc, movw_8_esp_es, sizeof (movw_8_esp_es)) == 0 ||
1809 		    bcmp(pc, movw_c_esp_ds, sizeof (movw_c_esp_ds)) == 0)
1810 			trp = lwptoregs(lwp);
1811 	}
1812 #endif	/* __amd64 */
1813 
1814 	if (trp == NULL)
1815 		return (1);
1816 
1817 	/*
1818 	 * If we get to here, we're reasonably confident that we've
1819 	 * correctly decoded what happened on the way out of the kernel.
1820 	 * Rewrite the lwp's registers so that we can create a core dump
1821 	 * the (at least vaguely) represents the mcontext we were
1822 	 * being asked to restore when things went so terribly wrong.
1823 	 */
1824 
1825 	/*
1826 	 * Make sure that we have a meaningful %trapno and %err.
1827 	 */
1828 	trp->r_trapno = rp->r_trapno;
1829 	trp->r_err = rp->r_err;
1830 
1831 	if ((caddr_t)trp != (caddr_t)lwptoregs(lwp))
1832 		bcopy(trp, lwptoregs(lwp), sizeof (*trp));
1833 
1834 	mutex_enter(&p->p_lock);
1835 	lwp->lwp_cursig = SIGSEGV;
1836 	mutex_exit(&p->p_lock);
1837 
1838 	/*
1839 	 * Terminate all LWPs but don't discard them.  If another lwp beat us to
1840 	 * the punch by calling exit(), evaporate now.
1841 	 */
1842 	proc_is_exiting(p);
1843 	if (exitlwps(1) != 0) {
1844 		mutex_enter(&p->p_lock);
1845 		lwp_exit();
1846 	}
1847 
1848 #ifdef C2_AUDIT
1849 	if (audit_active)		/* audit core dump */
1850 		audit_core_start(SIGSEGV);
1851 #endif
1852 	v = core(SIGSEGV, B_FALSE);
1853 #ifdef C2_AUDIT
1854 	if (audit_active)		/* audit core dump */
1855 		audit_core_finish(v ? CLD_KILLED : CLD_DUMPED);
1856 #endif
1857 	exit(v ? CLD_KILLED : CLD_DUMPED, SIGSEGV);
1858 	return (0);
1859 }
1860 
1861 /*
1862  * dump_tss() - Display the TSS structure
1863  */
1864 
1865 #if defined(__amd64)
1866 
1867 static void
1868 dump_tss(void)
1869 {
1870 	const char tss_fmt[] = "tss.%s:\t0x%p\n";  /* Format string */
1871 	struct tss *tss = CPU->cpu_tss;
1872 
1873 	printf(tss_fmt, "tss_rsp0", (void *)tss->tss_rsp0);
1874 	printf(tss_fmt, "tss_rsp1", (void *)tss->tss_rsp1);
1875 	printf(tss_fmt, "tss_rsp2", (void *)tss->tss_rsp2);
1876 
1877 	printf(tss_fmt, "tss_ist1", (void *)tss->tss_ist1);
1878 	printf(tss_fmt, "tss_ist2", (void *)tss->tss_ist2);
1879 	printf(tss_fmt, "tss_ist3", (void *)tss->tss_ist3);
1880 	printf(tss_fmt, "tss_ist4", (void *)tss->tss_ist4);
1881 	printf(tss_fmt, "tss_ist5", (void *)tss->tss_ist5);
1882 	printf(tss_fmt, "tss_ist6", (void *)tss->tss_ist6);
1883 	printf(tss_fmt, "tss_ist7", (void *)tss->tss_ist7);
1884 }
1885 
1886 #elif defined(__i386)
1887 
1888 static void
1889 dump_tss(void)
1890 {
1891 	const char tss_fmt[] = "tss.%s:\t0x%p\n";  /* Format string */
1892 	struct tss *tss = CPU->cpu_tss;
1893 
1894 	printf(tss_fmt, "tss_link", (void *)(uintptr_t)tss->tss_link);
1895 	printf(tss_fmt, "tss_esp0", (void *)(uintptr_t)tss->tss_esp0);
1896 	printf(tss_fmt, "tss_ss0", (void *)(uintptr_t)tss->tss_ss0);
1897 	printf(tss_fmt, "tss_esp1", (void *)(uintptr_t)tss->tss_esp1);
1898 	printf(tss_fmt, "tss_ss1", (void *)(uintptr_t)tss->tss_ss1);
1899 	printf(tss_fmt, "tss_esp2", (void *)(uintptr_t)tss->tss_esp2);
1900 	printf(tss_fmt, "tss_ss2", (void *)(uintptr_t)tss->tss_ss2);
1901 	printf(tss_fmt, "tss_cr3", (void *)(uintptr_t)tss->tss_cr3);
1902 	printf(tss_fmt, "tss_eip", (void *)(uintptr_t)tss->tss_eip);
1903 	printf(tss_fmt, "tss_eflags", (void *)(uintptr_t)tss->tss_eflags);
1904 	printf(tss_fmt, "tss_eax", (void *)(uintptr_t)tss->tss_eax);
1905 	printf(tss_fmt, "tss_ebx", (void *)(uintptr_t)tss->tss_ebx);
1906 	printf(tss_fmt, "tss_ecx", (void *)(uintptr_t)tss->tss_ecx);
1907 	printf(tss_fmt, "tss_edx", (void *)(uintptr_t)tss->tss_edx);
1908 	printf(tss_fmt, "tss_esp", (void *)(uintptr_t)tss->tss_esp);
1909 }
1910 
1911 #endif	/* __amd64 */
1912 
1913 #if defined(TRAPTRACE)
1914 
1915 int ttrace_nrec = 0;		/* number of records to dump out */
1916 int ttrace_dump_nregs = 5;	/* dump out this many records with regs too */
1917 
1918 /*
1919  * Dump out the last ttrace_nrec traptrace records on each CPU
1920  */
1921 static void
1922 dump_ttrace(void)
1923 {
1924 	trap_trace_ctl_t *ttc;
1925 	trap_trace_rec_t *rec;
1926 	uintptr_t current;
1927 	int i, j, k;
1928 	int n = NCPU;
1929 #if defined(__amd64)
1930 	const char banner[] =
1931 		"\ncpu          address    timestamp "
1932 		"type  vc  handler   pc\n";
1933 	const char fmt1[] = "%3d %016lx %12llx ";
1934 #elif defined(__i386)
1935 	const char banner[] =
1936 		"\ncpu  address     timestamp type  vc  handler   pc\n";
1937 	const char fmt1[] = "%3d %08lx %12llx ";
1938 #endif
1939 	const char fmt2[] = "%4s %3x ";
1940 	const char fmt3[] = "%8s ";
1941 
1942 	if (ttrace_nrec == 0)
1943 		return;
1944 
1945 	printf(banner);
1946 
1947 	for (i = 0; i < n; i++) {
1948 		ttc = &trap_trace_ctl[i];
1949 		if (ttc->ttc_first == NULL)
1950 			continue;
1951 
1952 		current = ttc->ttc_next - sizeof (trap_trace_rec_t);
1953 		for (j = 0; j < ttrace_nrec; j++) {
1954 			struct sysent	*sys;
1955 			struct autovec	*vec;
1956 			extern struct av_head autovect[];
1957 			int type;
1958 			ulong_t	off;
1959 			char *sym, *stype;
1960 
1961 			if (current < ttc->ttc_first)
1962 				current =
1963 				    ttc->ttc_limit - sizeof (trap_trace_rec_t);
1964 
1965 			if (current == NULL)
1966 				continue;
1967 
1968 			rec = (trap_trace_rec_t *)current;
1969 
1970 			if (rec->ttr_stamp == 0)
1971 				break;
1972 
1973 			printf(fmt1, i, (uintptr_t)rec, rec->ttr_stamp);
1974 
1975 			switch (rec->ttr_marker) {
1976 			case TT_SYSCALL:
1977 			case TT_SYSENTER:
1978 			case TT_SYSC:
1979 			case TT_SYSC64:
1980 #if defined(__amd64)
1981 				sys = &sysent32[rec->ttr_sysnum];
1982 				switch (rec->ttr_marker) {
1983 				case TT_SYSC64:
1984 					sys = &sysent[rec->ttr_sysnum];
1985 					/*FALLTHROUGH*/
1986 #elif defined(__i386)
1987 				sys = &sysent[rec->ttr_sysnum];
1988 				switch (rec->ttr_marker) {
1989 				case TT_SYSC64:
1990 #endif
1991 				case TT_SYSC:
1992 					stype = "sysc";	/* syscall */
1993 					break;
1994 				case TT_SYSCALL:
1995 					stype = "lcal";	/* lcall */
1996 					break;
1997 				case TT_SYSENTER:
1998 					stype = "syse";	/* sysenter */
1999 					break;
2000 				default:
2001 					break;
2002 				}
2003 				printf(fmt2, "sysc", rec->ttr_sysnum);
2004 				if (sys != NULL) {
2005 					sym = kobj_getsymname(
2006 					    (uintptr_t)sys->sy_callc,
2007 					    &off);
2008 					if (sym != NULL)
2009 						printf(fmt3, sym);
2010 					else
2011 						printf("%p ", sys->sy_callc);
2012 				} else {
2013 					printf(fmt3, "unknown");
2014 				}
2015 				break;
2016 
2017 			case TT_INTERRUPT:
2018 				printf(fmt2, "intr", rec->ttr_vector);
2019 				vec = (&autovect[rec->ttr_vector])->avh_link;
2020 				if (vec != NULL) {
2021 					sym = kobj_getsymname(
2022 					    (uintptr_t)vec->av_vector, &off);
2023 					if (sym != NULL)
2024 						printf(fmt3, sym);
2025 					else
2026 						printf("%p ", vec->av_vector);
2027 				} else {
2028 					printf(fmt3, "unknown ");
2029 				}
2030 				break;
2031 
2032 			case TT_TRAP:
2033 			case TT_EVENT:
2034 				type = rec->ttr_regs.r_trapno;
2035 				printf(fmt2, "trap", type);
2036 				if (type < TRAP_TYPES)
2037 					printf("     #%s ",
2038 					    trap_type_mnemonic[type]);
2039 				else
2040 					switch (type) {
2041 					case T_AST:
2042 						printf(fmt3, "ast");
2043 						break;
2044 					default:
2045 						printf(fmt3, "");
2046 						break;
2047 					}
2048 				break;
2049 
2050 			case TT_XCALL:
2051 				printf(fmt2, "xcal",
2052 				    rec->ttr_info.xc_entry.xce_marker);
2053 				printf(fmt3, "");
2054 				break;
2055 
2056 			default:
2057 				break;
2058 			}
2059 
2060 			sym = kobj_getsymname(rec->ttr_regs.r_pc, &off);
2061 			if (sym != NULL)
2062 				printf("%s+%lx\n", sym, off);
2063 			else
2064 				printf("%lx\n", rec->ttr_regs.r_pc);
2065 
2066 			if (ttrace_dump_nregs-- > 0) {
2067 				int s;
2068 
2069 				if (rec->ttr_marker == TT_INTERRUPT)
2070 					printf(
2071 					    "\t\tipl %x spl %x pri %x\n",
2072 					    rec->ttr_ipl,
2073 					    rec->ttr_spl,
2074 					    rec->ttr_pri);
2075 
2076 				dumpregs(&rec->ttr_regs);
2077 
2078 				printf("\t%3s: %p\n\n", " ct",
2079 				    (void *)rec->ttr_curthread);
2080 
2081 				/*
2082 				 * print out the pc stack that we recorded
2083 				 * at trap time (if any)
2084 				 */
2085 				for (s = 0; s < rec->ttr_sdepth; s++) {
2086 					uintptr_t fullpc;
2087 
2088 					if (s >= TTR_STACK_DEPTH) {
2089 						printf("ttr_sdepth corrupt\n");
2090 						break;
2091 					}
2092 
2093 					fullpc = (uintptr_t)rec->ttr_stack[s];
2094 
2095 					sym = kobj_getsymname(fullpc, &off);
2096 					if (sym != NULL)
2097 						printf("-> %s+0x%lx()\n",
2098 						    sym, off);
2099 					else
2100 						printf("-> 0x%lx()\n", fullpc);
2101 				}
2102 				printf("\n");
2103 			}
2104 			current -= sizeof (trap_trace_rec_t);
2105 		}
2106 	}
2107 }
2108 
2109 /*
2110  * Help with constructing traptrace records in C
2111  */
2112 trap_trace_rec_t *
2113 trap_trace_get_traceptr(uint8_t marker, ulong_t pc, ulong_t sp)
2114 {
2115 	trap_trace_rec_t *ttr;
2116 
2117 	if (trap_trace_freeze)
2118 		ttr = &trap_trace_postmort;
2119 	else {
2120 		trap_trace_ctl_t *ttc = &trap_trace_ctl[CPU->cpu_id];
2121 
2122 		ttr = (void *)ttc->ttc_next;
2123 
2124 		if (ttc->ttc_next >= ttc->ttc_limit)
2125 			ttc->ttc_next = ttc->ttc_first;
2126 		else
2127 			ttc->ttc_next += sizeof (trap_trace_rec_t);
2128 	}
2129 
2130 	ttr->ttr_regs.r_sp = sp;
2131 	ttr->ttr_regs.r_pc = pc;
2132 	ttr->ttr_cr2 = getcr2();
2133 	ttr->ttr_curthread = (uintptr_t)curthread;
2134 	ttr->ttr_stamp = tsc_read();
2135 	ttr->ttr_marker = marker;
2136 	return (ttr);
2137 }
2138 
2139 #endif	/* TRAPTRACE */
2140 
2141 void
2142 panic_showtrap(struct trap_info *tip)
2143 {
2144 	showregs(tip->trap_type, tip->trap_regs, tip->trap_addr);
2145 
2146 #if defined(TRAPTRACE)
2147 	dump_ttrace();
2148 #endif	/* TRAPTRACE */
2149 
2150 	if (tip->trap_type == T_DBLFLT)
2151 		dump_tss();
2152 }
2153 
2154 void
2155 panic_savetrap(panic_data_t *pdp, struct trap_info *tip)
2156 {
2157 	panic_saveregs(pdp, tip->trap_regs);
2158 }
2159