xref: /titanic_41/usr/src/uts/i86pc/os/trap.c (revision d24234c24aeaca4ca56ee3ac2794507968f274c4)
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 2009 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 #include <sys/types.h>
36 #include <sys/sysmacros.h>
37 #include <sys/param.h>
38 #include <sys/signal.h>
39 #include <sys/systm.h>
40 #include <sys/user.h>
41 #include <sys/proc.h>
42 #include <sys/disp.h>
43 #include <sys/class.h>
44 #include <sys/core.h>
45 #include <sys/syscall.h>
46 #include <sys/cpuvar.h>
47 #include <sys/vm.h>
48 #include <sys/sysinfo.h>
49 #include <sys/fault.h>
50 #include <sys/stack.h>
51 #include <sys/psw.h>
52 #include <sys/regset.h>
53 #include <sys/fp.h>
54 #include <sys/trap.h>
55 #include <sys/kmem.h>
56 #include <sys/vtrace.h>
57 #include <sys/cmn_err.h>
58 #include <sys/prsystm.h>
59 #include <sys/mutex_impl.h>
60 #include <sys/machsystm.h>
61 #include <sys/archsystm.h>
62 #include <sys/sdt.h>
63 #include <sys/avintr.h>
64 #include <sys/kobj.h>
65 
66 #include <vm/hat.h>
67 
68 #include <vm/seg_kmem.h>
69 #include <vm/as.h>
70 #include <vm/seg.h>
71 #include <vm/hat_pte.h>
72 #include <vm/hat_i86.h>
73 
74 #include <sys/procfs.h>
75 
76 #include <sys/reboot.h>
77 #include <sys/debug.h>
78 #include <sys/debugreg.h>
79 #include <sys/modctl.h>
80 #include <sys/aio_impl.h>
81 #include <sys/tnf.h>
82 #include <sys/tnf_probe.h>
83 #include <sys/cred.h>
84 #include <sys/mman.h>
85 #include <sys/x86_archext.h>
86 #include <sys/copyops.h>
87 #include <c2/audit.h>
88 #include <sys/ftrace.h>
89 #include <sys/panic.h>
90 #include <sys/traptrace.h>
91 #include <sys/ontrap.h>
92 #include <sys/cpc_impl.h>
93 #include <sys/bootconf.h>
94 #include <sys/bootinfo.h>
95 #include <sys/promif.h>
96 #include <sys/mach_mmu.h>
97 #if defined(__xpv)
98 #include <sys/hypervisor.h>
99 #endif
100 #include <sys/contract/process_impl.h>
101 
102 #define	USER	0x10000		/* user-mode flag added to trap type */
103 
104 static const char *trap_type_mnemonic[] = {
105 	"de",	"db",	"2",	"bp",
106 	"of",	"br",	"ud",	"nm",
107 	"df",	"9",	"ts",	"np",
108 	"ss",	"gp",	"pf",	"15",
109 	"mf",	"ac",	"mc",	"xf"
110 };
111 
112 static const char *trap_type[] = {
113 	"Divide error",				/* trap id 0 	*/
114 	"Debug",				/* trap id 1	*/
115 	"NMI interrupt",			/* trap id 2	*/
116 	"Breakpoint",				/* trap id 3 	*/
117 	"Overflow",				/* trap id 4 	*/
118 	"BOUND range exceeded",			/* trap id 5 	*/
119 	"Invalid opcode",			/* trap id 6 	*/
120 	"Device not available",			/* trap id 7 	*/
121 	"Double fault",				/* trap id 8 	*/
122 	"Coprocessor segment overrun",		/* trap id 9 	*/
123 	"Invalid TSS",				/* trap id 10 	*/
124 	"Segment not present",			/* trap id 11 	*/
125 	"Stack segment fault",			/* trap id 12 	*/
126 	"General protection",			/* trap id 13 	*/
127 	"Page fault",				/* trap id 14 	*/
128 	"Reserved",				/* trap id 15 	*/
129 	"x87 floating point error",		/* trap id 16 	*/
130 	"Alignment check",			/* trap id 17 	*/
131 	"Machine check",			/* trap id 18	*/
132 	"SIMD floating point exception",	/* trap id 19	*/
133 };
134 
135 #define	TRAP_TYPES	(sizeof (trap_type) / sizeof (trap_type[0]))
136 
137 #define	SLOW_SCALL_SIZE	2
138 #define	FAST_SCALL_SIZE	2
139 
140 int tudebug = 0;
141 int tudebugbpt = 0;
142 int tudebugfpe = 0;
143 int tudebugsse = 0;
144 
145 #if defined(TRAPDEBUG) || defined(lint)
146 int tdebug = 0;
147 int lodebug = 0;
148 int faultdebug = 0;
149 #else
150 #define	tdebug	0
151 #define	lodebug	0
152 #define	faultdebug	0
153 #endif /* defined(TRAPDEBUG) || defined(lint) */
154 
155 #if defined(TRAPTRACE)
156 /*
157  * trap trace record for cpu0 is allocated here.
158  * trap trace records for non-boot cpus are allocated in mp_startup_init().
159  */
160 static trap_trace_rec_t trap_tr0[TRAPTR_NENT];
161 trap_trace_ctl_t trap_trace_ctl[NCPU] = {
162 	{
163 	    (uintptr_t)trap_tr0,			/* next record */
164 	    (uintptr_t)trap_tr0,			/* first record */
165 	    (uintptr_t)(trap_tr0 + TRAPTR_NENT),	/* limit */
166 	    (uintptr_t)0				/* current */
167 	},
168 };
169 
170 /*
171  * default trap buffer size
172  */
173 size_t trap_trace_bufsize = TRAPTR_NENT * sizeof (trap_trace_rec_t);
174 int trap_trace_freeze = 0;
175 int trap_trace_off = 0;
176 
177 /*
178  * A dummy TRAPTRACE entry to use after death.
179  */
180 trap_trace_rec_t trap_trace_postmort;
181 
182 static void dump_ttrace(void);
183 #endif	/* TRAPTRACE */
184 static void dumpregs(struct regs *);
185 static void showregs(uint_t, struct regs *, caddr_t);
186 static int kern_gpfault(struct regs *);
187 
188 /*ARGSUSED*/
189 static int
190 die(uint_t type, struct regs *rp, caddr_t addr, processorid_t cpuid)
191 {
192 	struct panic_trap_info ti;
193 	const char *trap_name, *trap_mnemonic;
194 
195 	if (type < TRAP_TYPES) {
196 		trap_name = trap_type[type];
197 		trap_mnemonic = trap_type_mnemonic[type];
198 	} else {
199 		trap_name = "trap";
200 		trap_mnemonic = "-";
201 	}
202 
203 #ifdef TRAPTRACE
204 	TRAPTRACE_FREEZE;
205 #endif
206 
207 	ti.trap_regs = rp;
208 	ti.trap_type = type & ~USER;
209 	ti.trap_addr = addr;
210 
211 	curthread->t_panic_trap = &ti;
212 
213 	if (type == T_PGFLT && addr < (caddr_t)KERNELBASE) {
214 		panic("BAD TRAP: type=%x (#%s %s) rp=%p addr=%p "
215 		    "occurred in module \"%s\" due to %s",
216 		    type, trap_mnemonic, trap_name, (void *)rp, (void *)addr,
217 		    mod_containing_pc((caddr_t)rp->r_pc),
218 		    addr < (caddr_t)PAGESIZE ?
219 		    "a NULL pointer dereference" :
220 		    "an illegal access to a user address");
221 	} else
222 		panic("BAD TRAP: type=%x (#%s %s) rp=%p addr=%p",
223 		    type, trap_mnemonic, trap_name, (void *)rp, (void *)addr);
224 	return (0);
225 }
226 
227 /*
228  * Rewrite the instruction at pc to be an int $T_SYSCALLINT instruction.
229  *
230  * int <vector> is two bytes: 0xCD <vector>
231  */
232 
233 static int
234 rewrite_syscall(caddr_t pc)
235 {
236 	uchar_t instr[SLOW_SCALL_SIZE] = { 0xCD, T_SYSCALLINT };
237 
238 	if (uwrite(curthread->t_procp, instr, SLOW_SCALL_SIZE,
239 	    (uintptr_t)pc) != 0)
240 		return (1);
241 
242 	return (0);
243 }
244 
245 /*
246  * Test to see if the instruction at pc is sysenter or syscall. The second
247  * argument should be the x86 feature flag corresponding to the expected
248  * instruction.
249  *
250  * sysenter is two bytes: 0x0F 0x34
251  * syscall is two bytes:  0x0F 0x05
252  * int $T_SYSCALLINT is two bytes: 0xCD 0x91
253  */
254 
255 static int
256 instr_is_other_syscall(caddr_t pc, int which)
257 {
258 	uchar_t instr[FAST_SCALL_SIZE];
259 
260 	ASSERT(which == X86_SEP || which == X86_ASYSC || which == 0xCD);
261 
262 	if (copyin_nowatch(pc, (caddr_t)instr, FAST_SCALL_SIZE) != 0)
263 		return (0);
264 
265 	switch (which) {
266 	case X86_SEP:
267 		if (instr[0] == 0x0F && instr[1] == 0x34)
268 			return (1);
269 		break;
270 	case X86_ASYSC:
271 		if (instr[0] == 0x0F && instr[1] == 0x05)
272 			return (1);
273 		break;
274 	case 0xCD:
275 		if (instr[0] == 0xCD && instr[1] == T_SYSCALLINT)
276 			return (1);
277 		break;
278 	}
279 
280 	return (0);
281 }
282 
283 static const char *
284 syscall_insn_string(int syscall_insn)
285 {
286 	switch (syscall_insn) {
287 	case X86_SEP:
288 		return ("sysenter");
289 	case X86_ASYSC:
290 		return ("syscall");
291 	case 0xCD:
292 		return ("int");
293 	default:
294 		return ("Unknown");
295 	}
296 }
297 
298 static int
299 ldt_rewrite_syscall(struct regs *rp, proc_t *p, int syscall_insn)
300 {
301 	caddr_t	linearpc;
302 	int return_code = 0;
303 
304 	mutex_enter(&p->p_ldtlock);	/* Must be held across linear_pc() */
305 
306 	if (linear_pc(rp, p, &linearpc) == 0) {
307 
308 		/*
309 		 * If another thread beat us here, it already changed
310 		 * this site to the slower (int) syscall instruction.
311 		 */
312 		if (instr_is_other_syscall(linearpc, 0xCD)) {
313 			return_code = 1;
314 		} else if (instr_is_other_syscall(linearpc, syscall_insn)) {
315 
316 			if (rewrite_syscall(linearpc) == 0) {
317 				return_code = 1;
318 			}
319 #ifdef DEBUG
320 			else
321 				cmn_err(CE_WARN, "failed to rewrite %s "
322 				    "instruction in process %d",
323 				    syscall_insn_string(syscall_insn),
324 				    p->p_pid);
325 #endif /* DEBUG */
326 		}
327 	}
328 
329 	mutex_exit(&p->p_ldtlock);	/* Must be held across linear_pc() */
330 
331 	return (return_code);
332 }
333 
334 /*
335  * Test to see if the instruction at pc is a system call instruction.
336  *
337  * The bytes of an lcall instruction used for the syscall trap.
338  * static uchar_t lcall[7] = { 0x9a, 0, 0, 0, 0, 0x7, 0 };
339  * static uchar_t lcallalt[7] = { 0x9a, 0, 0, 0, 0, 0x27, 0 };
340  */
341 
342 #define	LCALLSIZE	7
343 
344 static int
345 instr_is_lcall_syscall(caddr_t pc)
346 {
347 	uchar_t instr[LCALLSIZE];
348 
349 	if (copyin_nowatch(pc, (caddr_t)instr, LCALLSIZE) == 0 &&
350 	    instr[0] == 0x9a &&
351 	    instr[1] == 0 &&
352 	    instr[2] == 0 &&
353 	    instr[3] == 0 &&
354 	    instr[4] == 0 &&
355 	    (instr[5] == 0x7 || instr[5] == 0x27) &&
356 	    instr[6] == 0)
357 		return (1);
358 
359 	return (0);
360 }
361 
362 #ifdef __amd64
363 
364 /*
365  * In the first revisions of amd64 CPUs produced by AMD, the LAHF and
366  * SAHF instructions were not implemented in 64-bit mode. Later revisions
367  * did implement these instructions. An extension to the cpuid instruction
368  * was added to check for the capability of executing these instructions
369  * in 64-bit mode.
370  *
371  * Intel originally did not implement these instructions in EM64T either,
372  * but added them in later revisions.
373  *
374  * So, there are different chip revisions by both vendors out there that
375  * may or may not implement these instructions. The easy solution is to
376  * just always emulate these instructions on demand.
377  *
378  * SAHF == store %ah in the lower 8 bits of %rflags (opcode 0x9e)
379  * LAHF == load the lower 8 bits of %rflags into %ah (opcode 0x9f)
380  */
381 
382 #define	LSAHFSIZE 1
383 
384 static int
385 instr_is_lsahf(caddr_t pc, uchar_t *instr)
386 {
387 	if (copyin_nowatch(pc, (caddr_t)instr, LSAHFSIZE) == 0 &&
388 	    (*instr == 0x9e || *instr == 0x9f))
389 		return (1);
390 	return (0);
391 }
392 
393 /*
394  * Emulate the LAHF and SAHF instructions. The reference manuals define
395  * these instructions to always load/store bit 1 as a 1, and bits 3 and 5
396  * as a 0. The other, defined, bits are copied (the PS_ICC bits and PS_P).
397  *
398  * Note that %ah is bits 8-15 of %rax.
399  */
400 static void
401 emulate_lsahf(struct regs *rp, uchar_t instr)
402 {
403 	if (instr == 0x9e) {
404 		/* sahf. Copy bits from %ah to flags. */
405 		rp->r_ps = (rp->r_ps & ~0xff) |
406 		    ((rp->r_rax >> 8) & PSL_LSAHFMASK) | PS_MB1;
407 	} else {
408 		/* lahf. Copy bits from flags to %ah. */
409 		rp->r_rax = (rp->r_rax & ~0xff00) |
410 		    (((rp->r_ps & PSL_LSAHFMASK) | PS_MB1) << 8);
411 	}
412 	rp->r_pc += LSAHFSIZE;
413 }
414 #endif /* __amd64 */
415 
416 #ifdef OPTERON_ERRATUM_91
417 
418 /*
419  * Test to see if the instruction at pc is a prefetch instruction.
420  *
421  * The first byte of prefetch instructions is always 0x0F.
422  * The second byte is 0x18 for regular prefetch or 0x0D for AMD 3dnow prefetch.
423  * The third byte (ModRM) contains the register field bits (bits 3-5).
424  * These bits must be between 0 and 3 inclusive for regular prefetch and
425  * 0 and 1 inclusive for AMD 3dnow prefetch.
426  *
427  * In 64-bit mode, there may be a one-byte REX prefex (0x40-0x4F).
428  */
429 
430 static int
431 cmp_to_prefetch(uchar_t *p)
432 {
433 #ifdef _LP64
434 	if ((p[0] & 0xF0) == 0x40)	/* 64-bit REX prefix */
435 		p++;
436 #endif
437 	return ((p[0] == 0x0F && p[1] == 0x18 && ((p[2] >> 3) & 7) <= 3) ||
438 	    (p[0] == 0x0F && p[1] == 0x0D && ((p[2] >> 3) & 7) <= 1));
439 }
440 
441 static int
442 instr_is_prefetch(caddr_t pc)
443 {
444 	uchar_t instr[4];	/* optional REX prefix plus 3-byte opcode */
445 
446 	return (copyin_nowatch(pc, instr, sizeof (instr)) == 0 &&
447 	    cmp_to_prefetch(instr));
448 }
449 
450 #endif /* OPTERON_ERRATUM_91 */
451 
452 /*
453  * Called from the trap handler when a processor trap occurs.
454  *
455  * Note: All user-level traps that might call stop() must exit
456  * trap() by 'goto out' or by falling through.
457  * Note Also: trap() is usually called with interrupts enabled, (PS_IE == 1)
458  * however, there are paths that arrive here with PS_IE == 0 so special care
459  * must be taken in those cases.
460  */
461 void
462 trap(struct regs *rp, caddr_t addr, processorid_t cpuid)
463 {
464 	kthread_t *ct = curthread;
465 	enum seg_rw rw;
466 	unsigned type;
467 	proc_t *p = ttoproc(ct);
468 	klwp_t *lwp = ttolwp(ct);
469 	uintptr_t lofault;
470 	faultcode_t pagefault(), res, errcode;
471 	enum fault_type fault_type;
472 	k_siginfo_t siginfo;
473 	uint_t fault = 0;
474 	int mstate;
475 	int sicode = 0;
476 	int watchcode;
477 	int watchpage;
478 	caddr_t vaddr;
479 	int singlestep_twiddle;
480 	size_t sz;
481 	int ta;
482 #ifdef __amd64
483 	uchar_t instr;
484 #endif
485 
486 	ASSERT_STACK_ALIGNED();
487 
488 	type = rp->r_trapno;
489 	CPU_STATS_ADDQ(CPU, sys, trap, 1);
490 	ASSERT(ct->t_schedflag & TS_DONT_SWAP);
491 
492 	if (type == T_PGFLT) {
493 
494 		errcode = rp->r_err;
495 		if (errcode & PF_ERR_WRITE)
496 			rw = S_WRITE;
497 		else if ((caddr_t)rp->r_pc == addr ||
498 		    (mmu.pt_nx != 0 && (errcode & PF_ERR_EXEC)))
499 			rw = S_EXEC;
500 		else
501 			rw = S_READ;
502 
503 #if defined(__i386)
504 		/*
505 		 * Pentium Pro work-around
506 		 */
507 		if ((errcode & PF_ERR_PROT) && pentiumpro_bug4046376) {
508 			uint_t	attr;
509 			uint_t	priv_violation;
510 			uint_t	access_violation;
511 
512 			if (hat_getattr(addr < (caddr_t)kernelbase ?
513 			    curproc->p_as->a_hat : kas.a_hat, addr, &attr)
514 			    == -1) {
515 				errcode &= ~PF_ERR_PROT;
516 			} else {
517 				priv_violation = (errcode & PF_ERR_USER) &&
518 				    !(attr & PROT_USER);
519 				access_violation = (errcode & PF_ERR_WRITE) &&
520 				    !(attr & PROT_WRITE);
521 				if (!priv_violation && !access_violation)
522 					goto cleanup;
523 			}
524 		}
525 #endif /* __i386 */
526 
527 	} else if (type == T_SGLSTP && lwp != NULL)
528 		lwp->lwp_pcb.pcb_drstat = (uintptr_t)addr;
529 
530 	if (tdebug)
531 		showregs(type, rp, addr);
532 
533 	if (USERMODE(rp->r_cs)) {
534 		/*
535 		 * Set up the current cred to use during this trap. u_cred
536 		 * no longer exists.  t_cred is used instead.
537 		 * The current process credential applies to the thread for
538 		 * the entire trap.  If trapping from the kernel, this
539 		 * should already be set up.
540 		 */
541 		if (ct->t_cred != p->p_cred) {
542 			cred_t *oldcred = ct->t_cred;
543 			/*
544 			 * DTrace accesses t_cred in probe context.  t_cred
545 			 * must always be either NULL, or point to a valid,
546 			 * allocated cred structure.
547 			 */
548 			ct->t_cred = crgetcred();
549 			crfree(oldcred);
550 		}
551 		ASSERT(lwp != NULL);
552 		type |= USER;
553 		ASSERT(lwptoregs(lwp) == rp);
554 		lwp->lwp_state = LWP_SYS;
555 
556 		switch (type) {
557 		case T_PGFLT + USER:
558 			if ((caddr_t)rp->r_pc == addr)
559 				mstate = LMS_TFAULT;
560 			else
561 				mstate = LMS_DFAULT;
562 			break;
563 		default:
564 			mstate = LMS_TRAP;
565 			break;
566 		}
567 		/* Kernel probe */
568 		TNF_PROBE_1(thread_state, "thread", /* CSTYLED */,
569 		    tnf_microstate, state, mstate);
570 		mstate = new_mstate(ct, mstate);
571 
572 		bzero(&siginfo, sizeof (siginfo));
573 	}
574 
575 	switch (type) {
576 	case T_PGFLT + USER:
577 	case T_SGLSTP:
578 	case T_SGLSTP + USER:
579 	case T_BPTFLT + USER:
580 		break;
581 
582 	default:
583 		FTRACE_2("trap(): type=0x%lx, regs=0x%lx",
584 		    (ulong_t)type, (ulong_t)rp);
585 		break;
586 	}
587 
588 	switch (type) {
589 	case T_SIMDFPE:
590 		/* Make sure we enable interrupts before die()ing */
591 		sti();	/* The SIMD exception comes in via cmninttrap */
592 		/*FALLTHROUGH*/
593 	default:
594 		if (type & USER) {
595 			if (tudebug)
596 				showregs(type, rp, (caddr_t)0);
597 			printf("trap: Unknown trap type %d in user mode\n",
598 			    type & ~USER);
599 			siginfo.si_signo = SIGILL;
600 			siginfo.si_code  = ILL_ILLTRP;
601 			siginfo.si_addr  = (caddr_t)rp->r_pc;
602 			siginfo.si_trapno = type & ~USER;
603 			fault = FLTILL;
604 			break;
605 		} else {
606 			(void) die(type, rp, addr, cpuid);
607 			/*NOTREACHED*/
608 		}
609 
610 	case T_PGFLT:		/* system page fault */
611 		/*
612 		 * If we're under on_trap() protection (see <sys/ontrap.h>),
613 		 * set ot_trap and bounce back to the on_trap() call site
614 		 * via the installed trampoline.
615 		 */
616 		if ((ct->t_ontrap != NULL) &&
617 		    (ct->t_ontrap->ot_prot & OT_DATA_ACCESS)) {
618 			ct->t_ontrap->ot_trap |= OT_DATA_ACCESS;
619 			rp->r_pc = ct->t_ontrap->ot_trampoline;
620 			goto cleanup;
621 		}
622 
623 		/*
624 		 * See if we can handle as pagefault. Save lofault
625 		 * across this. Here we assume that an address
626 		 * less than KERNELBASE is a user fault.
627 		 * We can do this as copy.s routines verify that the
628 		 * starting address is less than KERNELBASE before
629 		 * starting and because we know that we always have
630 		 * KERNELBASE mapped as invalid to serve as a "barrier".
631 		 */
632 		lofault = ct->t_lofault;
633 		ct->t_lofault = 0;
634 
635 		mstate = new_mstate(ct, LMS_KFAULT);
636 
637 		if (addr < (caddr_t)kernelbase) {
638 			res = pagefault(addr,
639 			    (errcode & PF_ERR_PROT)? F_PROT: F_INVAL, rw, 0);
640 			if (res == FC_NOMAP &&
641 			    addr < p->p_usrstack &&
642 			    grow(addr))
643 				res = 0;
644 		} else {
645 			res = pagefault(addr,
646 			    (errcode & PF_ERR_PROT)? F_PROT: F_INVAL, rw, 1);
647 		}
648 		(void) new_mstate(ct, mstate);
649 
650 		/*
651 		 * Restore lofault. If we resolved the fault, exit.
652 		 * If we didn't and lofault wasn't set, die.
653 		 */
654 		ct->t_lofault = lofault;
655 		if (res == 0)
656 			goto cleanup;
657 
658 #if defined(OPTERON_ERRATUM_93) && defined(_LP64)
659 		if (lofault == 0 && opteron_erratum_93) {
660 			/*
661 			 * Workaround for Opteron Erratum 93. On return from
662 			 * a System Managment Interrupt at a HLT instruction
663 			 * the %rip might be truncated to a 32 bit value.
664 			 * BIOS is supposed to fix this, but some don't.
665 			 * If this occurs we simply restore the high order bits.
666 			 * The HLT instruction is 1 byte of 0xf4.
667 			 */
668 			uintptr_t	rip = rp->r_pc;
669 
670 			if ((rip & 0xfffffffful) == rip) {
671 				rip |= 0xfffffffful << 32;
672 				if (hat_getpfnum(kas.a_hat, (caddr_t)rip) !=
673 				    PFN_INVALID &&
674 				    (*(uchar_t *)rip == 0xf4 ||
675 				    *(uchar_t *)(rip - 1) == 0xf4)) {
676 					rp->r_pc = rip;
677 					goto cleanup;
678 				}
679 			}
680 		}
681 #endif /* OPTERON_ERRATUM_93 && _LP64 */
682 
683 #ifdef OPTERON_ERRATUM_91
684 		if (lofault == 0 && opteron_erratum_91) {
685 			/*
686 			 * Workaround for Opteron Erratum 91. Prefetches may
687 			 * generate a page fault (they're not supposed to do
688 			 * that!). If this occurs we simply return back to the
689 			 * instruction.
690 			 */
691 			caddr_t		pc = (caddr_t)rp->r_pc;
692 
693 			/*
694 			 * If the faulting PC is not mapped, this is a
695 			 * legitimate kernel page fault that must result in a
696 			 * panic. If the faulting PC is mapped, it could contain
697 			 * a prefetch instruction. Check for that here.
698 			 */
699 			if (hat_getpfnum(kas.a_hat, pc) != PFN_INVALID) {
700 				if (cmp_to_prefetch((uchar_t *)pc)) {
701 #ifdef DEBUG
702 					cmn_err(CE_WARN, "Opteron erratum 91 "
703 					    "occurred: kernel prefetch"
704 					    " at %p generated a page fault!",
705 					    (void *)rp->r_pc);
706 #endif /* DEBUG */
707 					goto cleanup;
708 				}
709 			}
710 			(void) die(type, rp, addr, cpuid);
711 		}
712 #endif /* OPTERON_ERRATUM_91 */
713 
714 		if (lofault == 0)
715 			(void) die(type, rp, addr, cpuid);
716 
717 		/*
718 		 * Cannot resolve fault.  Return to lofault.
719 		 */
720 		if (lodebug) {
721 			showregs(type, rp, addr);
722 			traceregs(rp);
723 		}
724 		if (FC_CODE(res) == FC_OBJERR)
725 			res = FC_ERRNO(res);
726 		else
727 			res = EFAULT;
728 		rp->r_r0 = res;
729 		rp->r_pc = ct->t_lofault;
730 		goto cleanup;
731 
732 	case T_PGFLT + USER:	/* user page fault */
733 		if (faultdebug) {
734 			char *fault_str;
735 
736 			switch (rw) {
737 			case S_READ:
738 				fault_str = "read";
739 				break;
740 			case S_WRITE:
741 				fault_str = "write";
742 				break;
743 			case S_EXEC:
744 				fault_str = "exec";
745 				break;
746 			default:
747 				fault_str = "";
748 				break;
749 			}
750 			printf("user %s fault:  addr=0x%lx errcode=0x%x\n",
751 			    fault_str, (uintptr_t)addr, errcode);
752 		}
753 
754 #if defined(OPTERON_ERRATUM_100) && defined(_LP64)
755 		/*
756 		 * Workaround for AMD erratum 100
757 		 *
758 		 * A 32-bit process may receive a page fault on a non
759 		 * 32-bit address by mistake. The range of the faulting
760 		 * address will be
761 		 *
762 		 *	0xffffffff80000000 .. 0xffffffffffffffff or
763 		 *	0x0000000100000000 .. 0x000000017fffffff
764 		 *
765 		 * The fault is always due to an instruction fetch, however
766 		 * the value of r_pc should be correct (in 32 bit range),
767 		 * so we ignore the page fault on the bogus address.
768 		 */
769 		if (p->p_model == DATAMODEL_ILP32 &&
770 		    (0xffffffff80000000 <= (uintptr_t)addr ||
771 		    (0x100000000 <= (uintptr_t)addr &&
772 		    (uintptr_t)addr <= 0x17fffffff))) {
773 			if (!opteron_erratum_100)
774 				panic("unexpected erratum #100");
775 			if (rp->r_pc <= 0xffffffff)
776 				goto out;
777 		}
778 #endif /* OPTERON_ERRATUM_100 && _LP64 */
779 
780 		ASSERT(!(curthread->t_flag & T_WATCHPT));
781 		watchpage = (pr_watch_active(p) && pr_is_watchpage(addr, rw));
782 #ifdef __i386
783 		/*
784 		 * In 32-bit mode, the lcall (system call) instruction fetches
785 		 * one word from the stack, at the stack pointer, because of the
786 		 * way the call gate is constructed.  This is a bogus
787 		 * read and should not be counted as a read watchpoint.
788 		 * We work around the problem here by testing to see if
789 		 * this situation applies and, if so, simply jumping to
790 		 * the code in locore.s that fields the system call trap.
791 		 * The registers on the stack are already set up properly
792 		 * due to the match between the call gate sequence and the
793 		 * trap gate sequence.  We just have to adjust the pc.
794 		 */
795 		if (watchpage && addr == (caddr_t)rp->r_sp &&
796 		    rw == S_READ && instr_is_lcall_syscall((caddr_t)rp->r_pc)) {
797 			extern void watch_syscall(void);
798 
799 			rp->r_pc += LCALLSIZE;
800 			watch_syscall();	/* never returns */
801 			/* NOTREACHED */
802 		}
803 #endif /* __i386 */
804 		vaddr = addr;
805 		if (!watchpage || (sz = instr_size(rp, &vaddr, rw)) <= 0)
806 			fault_type = (errcode & PF_ERR_PROT)? F_PROT: F_INVAL;
807 		else if ((watchcode = pr_is_watchpoint(&vaddr, &ta,
808 		    sz, NULL, rw)) != 0) {
809 			if (ta) {
810 				do_watch_step(vaddr, sz, rw,
811 				    watchcode, rp->r_pc);
812 				fault_type = F_INVAL;
813 			} else {
814 				bzero(&siginfo, sizeof (siginfo));
815 				siginfo.si_signo = SIGTRAP;
816 				siginfo.si_code = watchcode;
817 				siginfo.si_addr = vaddr;
818 				siginfo.si_trapafter = 0;
819 				siginfo.si_pc = (caddr_t)rp->r_pc;
820 				fault = FLTWATCH;
821 				break;
822 			}
823 		} else {
824 			/* XXX pr_watch_emul() never succeeds (for now) */
825 			if (rw != S_EXEC && pr_watch_emul(rp, vaddr, rw))
826 				goto out;
827 			do_watch_step(vaddr, sz, rw, 0, 0);
828 			fault_type = F_INVAL;
829 		}
830 
831 		res = pagefault(addr, fault_type, rw, 0);
832 
833 		/*
834 		 * If pagefault() succeeded, ok.
835 		 * Otherwise attempt to grow the stack.
836 		 */
837 		if (res == 0 ||
838 		    (res == FC_NOMAP &&
839 		    addr < p->p_usrstack &&
840 		    grow(addr))) {
841 			lwp->lwp_lastfault = FLTPAGE;
842 			lwp->lwp_lastfaddr = addr;
843 			if (prismember(&p->p_fltmask, FLTPAGE)) {
844 				bzero(&siginfo, sizeof (siginfo));
845 				siginfo.si_addr = addr;
846 				(void) stop_on_fault(FLTPAGE, &siginfo);
847 			}
848 			goto out;
849 		} else if (res == FC_PROT && addr < p->p_usrstack &&
850 		    (mmu.pt_nx != 0 && (errcode & PF_ERR_EXEC))) {
851 			report_stack_exec(p, addr);
852 		}
853 
854 #ifdef OPTERON_ERRATUM_91
855 		/*
856 		 * Workaround for Opteron Erratum 91. Prefetches may generate a
857 		 * page fault (they're not supposed to do that!). If this
858 		 * occurs we simply return back to the instruction.
859 		 *
860 		 * We rely on copyin to properly fault in the page with r_pc.
861 		 */
862 		if (opteron_erratum_91 &&
863 		    addr != (caddr_t)rp->r_pc &&
864 		    instr_is_prefetch((caddr_t)rp->r_pc)) {
865 #ifdef DEBUG
866 			cmn_err(CE_WARN, "Opteron erratum 91 occurred: "
867 			    "prefetch at %p in pid %d generated a trap!",
868 			    (void *)rp->r_pc, p->p_pid);
869 #endif /* DEBUG */
870 			goto out;
871 		}
872 #endif /* OPTERON_ERRATUM_91 */
873 
874 		if (tudebug)
875 			showregs(type, rp, addr);
876 		/*
877 		 * In the case where both pagefault and grow fail,
878 		 * set the code to the value provided by pagefault.
879 		 * We map all errors returned from pagefault() to SIGSEGV.
880 		 */
881 		bzero(&siginfo, sizeof (siginfo));
882 		siginfo.si_addr = addr;
883 		switch (FC_CODE(res)) {
884 		case FC_HWERR:
885 		case FC_NOSUPPORT:
886 			siginfo.si_signo = SIGBUS;
887 			siginfo.si_code = BUS_ADRERR;
888 			fault = FLTACCESS;
889 			break;
890 		case FC_ALIGN:
891 			siginfo.si_signo = SIGBUS;
892 			siginfo.si_code = BUS_ADRALN;
893 			fault = FLTACCESS;
894 			break;
895 		case FC_OBJERR:
896 			if ((siginfo.si_errno = FC_ERRNO(res)) != EINTR) {
897 				siginfo.si_signo = SIGBUS;
898 				siginfo.si_code = BUS_OBJERR;
899 				fault = FLTACCESS;
900 			}
901 			break;
902 		default:	/* FC_NOMAP or FC_PROT */
903 			siginfo.si_signo = SIGSEGV;
904 			siginfo.si_code =
905 			    (res == FC_NOMAP)? SEGV_MAPERR : SEGV_ACCERR;
906 			fault = FLTBOUNDS;
907 			break;
908 		}
909 		break;
910 
911 	case T_ILLINST + USER:	/* invalid opcode fault */
912 		/*
913 		 * If the syscall instruction is disabled due to LDT usage, a
914 		 * user program that attempts to execute it will trigger a #ud
915 		 * trap. Check for that case here. If this occurs on a CPU which
916 		 * doesn't even support syscall, the result of all of this will
917 		 * be to emulate that particular instruction.
918 		 */
919 		if (p->p_ldt != NULL &&
920 		    ldt_rewrite_syscall(rp, p, X86_ASYSC))
921 			goto out;
922 
923 #ifdef __amd64
924 		/*
925 		 * Emulate the LAHF and SAHF instructions if needed.
926 		 * See the instr_is_lsahf function for details.
927 		 */
928 		if (p->p_model == DATAMODEL_LP64 &&
929 		    instr_is_lsahf((caddr_t)rp->r_pc, &instr)) {
930 			emulate_lsahf(rp, instr);
931 			goto out;
932 		}
933 #endif
934 
935 		/*FALLTHROUGH*/
936 
937 		if (tudebug)
938 			showregs(type, rp, (caddr_t)0);
939 		siginfo.si_signo = SIGILL;
940 		siginfo.si_code  = ILL_ILLOPC;
941 		siginfo.si_addr  = (caddr_t)rp->r_pc;
942 		fault = FLTILL;
943 		break;
944 
945 	case T_ZERODIV + USER:		/* integer divide by zero */
946 		if (tudebug && tudebugfpe)
947 			showregs(type, rp, (caddr_t)0);
948 		siginfo.si_signo = SIGFPE;
949 		siginfo.si_code  = FPE_INTDIV;
950 		siginfo.si_addr  = (caddr_t)rp->r_pc;
951 		fault = FLTIZDIV;
952 		break;
953 
954 	case T_OVFLW + USER:	/* integer overflow */
955 		if (tudebug && tudebugfpe)
956 			showregs(type, rp, (caddr_t)0);
957 		siginfo.si_signo = SIGFPE;
958 		siginfo.si_code  = FPE_INTOVF;
959 		siginfo.si_addr  = (caddr_t)rp->r_pc;
960 		fault = FLTIOVF;
961 		break;
962 
963 	case T_NOEXTFLT + USER:	/* math coprocessor not available */
964 		if (tudebug && tudebugfpe)
965 			showregs(type, rp, addr);
966 		if (fpnoextflt(rp)) {
967 			siginfo.si_signo = SIGILL;
968 			siginfo.si_code  = ILL_ILLOPC;
969 			siginfo.si_addr  = (caddr_t)rp->r_pc;
970 			fault = FLTILL;
971 		}
972 		break;
973 
974 	case T_EXTOVRFLT:	/* extension overrun fault */
975 		/* check if we took a kernel trap on behalf of user */
976 		{
977 			extern  void ndptrap_frstor(void);
978 			if (rp->r_pc != (uintptr_t)ndptrap_frstor) {
979 				sti(); /* T_EXTOVRFLT comes in via cmninttrap */
980 				(void) die(type, rp, addr, cpuid);
981 			}
982 			type |= USER;
983 		}
984 		/*FALLTHROUGH*/
985 	case T_EXTOVRFLT + USER:	/* extension overrun fault */
986 		if (tudebug && tudebugfpe)
987 			showregs(type, rp, addr);
988 		if (fpextovrflt(rp)) {
989 			siginfo.si_signo = SIGSEGV;
990 			siginfo.si_code  = SEGV_MAPERR;
991 			siginfo.si_addr  = (caddr_t)rp->r_pc;
992 			fault = FLTBOUNDS;
993 		}
994 		break;
995 
996 	case T_EXTERRFLT:	/* x87 floating point exception pending */
997 		/* check if we took a kernel trap on behalf of user */
998 		{
999 			extern  void ndptrap_frstor(void);
1000 			if (rp->r_pc != (uintptr_t)ndptrap_frstor) {
1001 				sti(); /* T_EXTERRFLT comes in via cmninttrap */
1002 				(void) die(type, rp, addr, cpuid);
1003 			}
1004 			type |= USER;
1005 		}
1006 		/*FALLTHROUGH*/
1007 
1008 	case T_EXTERRFLT + USER: /* x87 floating point exception pending */
1009 		if (tudebug && tudebugfpe)
1010 			showregs(type, rp, addr);
1011 		if (sicode = fpexterrflt(rp)) {
1012 			siginfo.si_signo = SIGFPE;
1013 			siginfo.si_code  = sicode;
1014 			siginfo.si_addr  = (caddr_t)rp->r_pc;
1015 			fault = FLTFPE;
1016 		}
1017 		break;
1018 
1019 	case T_SIMDFPE + USER:		/* SSE and SSE2 exceptions */
1020 		if (tudebug && tudebugsse)
1021 			showregs(type, rp, addr);
1022 		if ((x86_feature & (X86_SSE|X86_SSE2)) == 0) {
1023 			/*
1024 			 * There are rumours that some user instructions
1025 			 * on older CPUs can cause this trap to occur; in
1026 			 * which case send a SIGILL instead of a SIGFPE.
1027 			 */
1028 			siginfo.si_signo = SIGILL;
1029 			siginfo.si_code  = ILL_ILLTRP;
1030 			siginfo.si_addr  = (caddr_t)rp->r_pc;
1031 			siginfo.si_trapno = type & ~USER;
1032 			fault = FLTILL;
1033 		} else if ((sicode = fpsimderrflt(rp)) != 0) {
1034 			siginfo.si_signo = SIGFPE;
1035 			siginfo.si_code = sicode;
1036 			siginfo.si_addr = (caddr_t)rp->r_pc;
1037 			fault = FLTFPE;
1038 		}
1039 
1040 		sti();	/* The SIMD exception comes in via cmninttrap */
1041 		break;
1042 
1043 	case T_BPTFLT:	/* breakpoint trap */
1044 		/*
1045 		 * Kernel breakpoint traps should only happen when kmdb is
1046 		 * active, and even then, it'll have interposed on the IDT, so
1047 		 * control won't get here.  If it does, we've hit a breakpoint
1048 		 * without the debugger, which is very strange, and very
1049 		 * fatal.
1050 		 */
1051 		if (tudebug && tudebugbpt)
1052 			showregs(type, rp, (caddr_t)0);
1053 
1054 		(void) die(type, rp, addr, cpuid);
1055 		break;
1056 
1057 	case T_SGLSTP: /* single step/hw breakpoint exception */
1058 
1059 		/* Now evaluate how we got here */
1060 		if (lwp != NULL && (lwp->lwp_pcb.pcb_drstat & DR_SINGLESTEP)) {
1061 			/*
1062 			 * i386 single-steps even through lcalls which
1063 			 * change the privilege level. So we take a trap at
1064 			 * the first instruction in privileged mode.
1065 			 *
1066 			 * Set a flag to indicate that upon completion of
1067 			 * the system call, deal with the single-step trap.
1068 			 *
1069 			 * The same thing happens for sysenter, too.
1070 			 */
1071 			singlestep_twiddle = 0;
1072 			if (rp->r_pc == (uintptr_t)sys_sysenter ||
1073 			    rp->r_pc == (uintptr_t)brand_sys_sysenter) {
1074 				singlestep_twiddle = 1;
1075 #if defined(__amd64)
1076 				/*
1077 				 * Since we are already on the kernel's
1078 				 * %gs, on 64-bit systems the sysenter case
1079 				 * needs to adjust the pc to avoid
1080 				 * executing the swapgs instruction at the
1081 				 * top of the handler.
1082 				 */
1083 				if (rp->r_pc == (uintptr_t)sys_sysenter)
1084 					rp->r_pc = (uintptr_t)
1085 					    _sys_sysenter_post_swapgs;
1086 				else
1087 					rp->r_pc = (uintptr_t)
1088 					    _brand_sys_sysenter_post_swapgs;
1089 #endif
1090 			}
1091 #if defined(__i386)
1092 			else if (rp->r_pc == (uintptr_t)sys_call ||
1093 			    rp->r_pc == (uintptr_t)brand_sys_call) {
1094 				singlestep_twiddle = 1;
1095 			}
1096 #endif
1097 			else {
1098 				/* not on sysenter/syscall; uregs available */
1099 				if (tudebug && tudebugbpt)
1100 					showregs(type, rp, (caddr_t)0);
1101 			}
1102 			if (singlestep_twiddle) {
1103 				rp->r_ps &= ~PS_T; /* turn off trace */
1104 				lwp->lwp_pcb.pcb_flags |= DEBUG_PENDING;
1105 				ct->t_post_sys = 1;
1106 				aston(curthread);
1107 				goto cleanup;
1108 			}
1109 		}
1110 		/* XXX - needs review on debugger interface? */
1111 		if (boothowto & RB_DEBUG)
1112 			debug_enter((char *)NULL);
1113 		else
1114 			(void) die(type, rp, addr, cpuid);
1115 		break;
1116 
1117 	case T_NMIFLT:	/* NMI interrupt */
1118 		printf("Unexpected NMI in system mode\n");
1119 		goto cleanup;
1120 
1121 	case T_NMIFLT + USER:	/* NMI interrupt */
1122 		printf("Unexpected NMI in user mode\n");
1123 		break;
1124 
1125 	case T_GPFLT:	/* general protection violation */
1126 		/*
1127 		 * Any #GP that occurs during an on_trap .. no_trap bracket
1128 		 * with OT_DATA_ACCESS or OT_SEGMENT_ACCESS protection,
1129 		 * or in a on_fault .. no_fault bracket, is forgiven
1130 		 * and we trampoline.  This protection is given regardless
1131 		 * of whether we are 32/64 bit etc - if a distinction is
1132 		 * required then define new on_trap protection types.
1133 		 *
1134 		 * On amd64, we can get a #gp from referencing addresses
1135 		 * in the virtual address hole e.g. from a copyin or in
1136 		 * update_sregs while updating user segment registers.
1137 		 *
1138 		 * On the 32-bit hypervisor we could also generate one in
1139 		 * mfn_to_pfn by reaching around or into where the hypervisor
1140 		 * lives which is protected by segmentation.
1141 		 */
1142 
1143 		/*
1144 		 * If we're under on_trap() protection (see <sys/ontrap.h>),
1145 		 * set ot_trap and trampoline back to the on_trap() call site
1146 		 * for OT_DATA_ACCESS or OT_SEGMENT_ACCESS.
1147 		 */
1148 		if (ct->t_ontrap != NULL) {
1149 			int ttype =  ct->t_ontrap->ot_prot &
1150 			    (OT_DATA_ACCESS | OT_SEGMENT_ACCESS);
1151 
1152 			if (ttype != 0) {
1153 				ct->t_ontrap->ot_trap |= ttype;
1154 				if (tudebug)
1155 					showregs(type, rp, (caddr_t)0);
1156 				rp->r_pc = ct->t_ontrap->ot_trampoline;
1157 				goto cleanup;
1158 			}
1159 		}
1160 
1161 		/*
1162 		 * If we're under lofault protection (copyin etc.),
1163 		 * longjmp back to lofault with an EFAULT.
1164 		 */
1165 		if (ct->t_lofault) {
1166 			/*
1167 			 * Fault is not resolvable, so just return to lofault
1168 			 */
1169 			if (lodebug) {
1170 				showregs(type, rp, addr);
1171 				traceregs(rp);
1172 			}
1173 			rp->r_r0 = EFAULT;
1174 			rp->r_pc = ct->t_lofault;
1175 			goto cleanup;
1176 		}
1177 
1178 		/*
1179 		 * We fall through to the next case, which repeats
1180 		 * the OT_SEGMENT_ACCESS check which we've already
1181 		 * done, so we'll always fall through to the
1182 		 * T_STKFLT case.
1183 		 */
1184 		/*FALLTHROUGH*/
1185 	case T_SEGFLT:	/* segment not present fault */
1186 		/*
1187 		 * One example of this is #NP in update_sregs while
1188 		 * attempting to update a user segment register
1189 		 * that points to a descriptor that is marked not
1190 		 * present.
1191 		 */
1192 		if (ct->t_ontrap != NULL &&
1193 		    ct->t_ontrap->ot_prot & OT_SEGMENT_ACCESS) {
1194 			ct->t_ontrap->ot_trap |= OT_SEGMENT_ACCESS;
1195 			if (tudebug)
1196 				showregs(type, rp, (caddr_t)0);
1197 			rp->r_pc = ct->t_ontrap->ot_trampoline;
1198 			goto cleanup;
1199 		}
1200 		/*FALLTHROUGH*/
1201 	case T_STKFLT:	/* stack fault */
1202 	case T_TSSFLT:	/* invalid TSS fault */
1203 		if (tudebug)
1204 			showregs(type, rp, (caddr_t)0);
1205 		if (kern_gpfault(rp))
1206 			(void) die(type, rp, addr, cpuid);
1207 		goto cleanup;
1208 
1209 	/*
1210 	 * ONLY 32-bit PROCESSES can USE a PRIVATE LDT! 64-bit apps
1211 	 * should have no need for them, so we put a stop to it here.
1212 	 *
1213 	 * So: not-present fault is ONLY valid for 32-bit processes with
1214 	 * a private LDT trying to do a system call. Emulate it.
1215 	 *
1216 	 * #gp fault is ONLY valid for 32-bit processes also, which DO NOT
1217 	 * have a private LDT, and are trying to do a system call. Emulate it.
1218 	 */
1219 
1220 	case T_SEGFLT + USER:	/* segment not present fault */
1221 	case T_GPFLT + USER:	/* general protection violation */
1222 #ifdef _SYSCALL32_IMPL
1223 		if (p->p_model != DATAMODEL_NATIVE) {
1224 #endif /* _SYSCALL32_IMPL */
1225 		if (instr_is_lcall_syscall((caddr_t)rp->r_pc)) {
1226 			if (type == T_SEGFLT + USER)
1227 				ASSERT(p->p_ldt != NULL);
1228 
1229 			if ((p->p_ldt == NULL && type == T_GPFLT + USER) ||
1230 			    type == T_SEGFLT + USER) {
1231 
1232 			/*
1233 			 * The user attempted a system call via the obsolete
1234 			 * call gate mechanism. Because the process doesn't have
1235 			 * an LDT (i.e. the ldtr contains 0), a #gp results.
1236 			 * Emulate the syscall here, just as we do above for a
1237 			 * #np trap.
1238 			 */
1239 
1240 			/*
1241 			 * Since this is a not-present trap, rp->r_pc points to
1242 			 * the trapping lcall instruction. We need to bump it
1243 			 * to the next insn so the app can continue on.
1244 			 */
1245 			rp->r_pc += LCALLSIZE;
1246 			lwp->lwp_regs = rp;
1247 
1248 			/*
1249 			 * Normally the microstate of the LWP is forced back to
1250 			 * LMS_USER by the syscall handlers. Emulate that
1251 			 * behavior here.
1252 			 */
1253 			mstate = LMS_USER;
1254 
1255 			dosyscall();
1256 			goto out;
1257 			}
1258 		}
1259 #ifdef _SYSCALL32_IMPL
1260 		}
1261 #endif /* _SYSCALL32_IMPL */
1262 		/*
1263 		 * If the current process is using a private LDT and the
1264 		 * trapping instruction is sysenter, the sysenter instruction
1265 		 * has been disabled on the CPU because it destroys segment
1266 		 * registers. If this is the case, rewrite the instruction to
1267 		 * be a safe system call and retry it. If this occurs on a CPU
1268 		 * which doesn't even support sysenter, the result of all of
1269 		 * this will be to emulate that particular instruction.
1270 		 */
1271 		if (p->p_ldt != NULL &&
1272 		    ldt_rewrite_syscall(rp, p, X86_SEP))
1273 			goto out;
1274 
1275 		/*FALLTHROUGH*/
1276 
1277 	case T_BOUNDFLT + USER:	/* bound fault */
1278 	case T_STKFLT + USER:	/* stack fault */
1279 	case T_TSSFLT + USER:	/* invalid TSS fault */
1280 		if (tudebug)
1281 			showregs(type, rp, (caddr_t)0);
1282 		siginfo.si_signo = SIGSEGV;
1283 		siginfo.si_code  = SEGV_MAPERR;
1284 		siginfo.si_addr  = (caddr_t)rp->r_pc;
1285 		fault = FLTBOUNDS;
1286 		break;
1287 
1288 	case T_ALIGNMENT + USER:	/* user alignment error (486) */
1289 		if (tudebug)
1290 			showregs(type, rp, (caddr_t)0);
1291 		bzero(&siginfo, sizeof (siginfo));
1292 		siginfo.si_signo = SIGBUS;
1293 		siginfo.si_code = BUS_ADRALN;
1294 		siginfo.si_addr = (caddr_t)rp->r_pc;
1295 		fault = FLTACCESS;
1296 		break;
1297 
1298 	case T_SGLSTP + USER: /* single step/hw breakpoint exception */
1299 		if (tudebug && tudebugbpt)
1300 			showregs(type, rp, (caddr_t)0);
1301 
1302 		/* Was it single-stepping? */
1303 		if (lwp->lwp_pcb.pcb_drstat & DR_SINGLESTEP) {
1304 			pcb_t *pcb = &lwp->lwp_pcb;
1305 
1306 			rp->r_ps &= ~PS_T;
1307 			/*
1308 			 * If both NORMAL_STEP and WATCH_STEP are in effect,
1309 			 * give precedence to WATCH_STEP.  If neither is set,
1310 			 * user must have set the PS_T bit in %efl; treat this
1311 			 * as NORMAL_STEP.
1312 			 */
1313 			if ((fault = undo_watch_step(&siginfo)) == 0 &&
1314 			    ((pcb->pcb_flags & NORMAL_STEP) ||
1315 			    !(pcb->pcb_flags & WATCH_STEP))) {
1316 				siginfo.si_signo = SIGTRAP;
1317 				siginfo.si_code = TRAP_TRACE;
1318 				siginfo.si_addr = (caddr_t)rp->r_pc;
1319 				fault = FLTTRACE;
1320 			}
1321 			pcb->pcb_flags &= ~(NORMAL_STEP|WATCH_STEP);
1322 		} else {
1323 			cmn_err(CE_WARN,
1324 			    "Unexpected INT 1 in user mode, dr6=%lx",
1325 			    lwp->lwp_pcb.pcb_drstat);
1326 		}
1327 		break;
1328 
1329 	case T_BPTFLT + USER:	/* breakpoint trap */
1330 		if (tudebug && tudebugbpt)
1331 			showregs(type, rp, (caddr_t)0);
1332 		/*
1333 		 * int 3 (the breakpoint instruction) leaves the pc referring
1334 		 * to the address one byte after the breakpointed address.
1335 		 * If the P_PR_BPTADJ flag has been set via /proc, We adjust
1336 		 * it back so it refers to the breakpointed address.
1337 		 */
1338 		if (p->p_proc_flag & P_PR_BPTADJ)
1339 			rp->r_pc--;
1340 		siginfo.si_signo = SIGTRAP;
1341 		siginfo.si_code  = TRAP_BRKPT;
1342 		siginfo.si_addr  = (caddr_t)rp->r_pc;
1343 		fault = FLTBPT;
1344 		break;
1345 
1346 	case T_AST:
1347 		/*
1348 		 * This occurs only after the cs register has been made to
1349 		 * look like a kernel selector, either through debugging or
1350 		 * possibly by functions like setcontext().  The thread is
1351 		 * about to cause a general protection fault at common_iret()
1352 		 * in locore.  We let that happen immediately instead of
1353 		 * doing the T_AST processing.
1354 		 */
1355 		goto cleanup;
1356 
1357 	case T_AST + USER:	/* profiling, resched, h/w error pseudo trap */
1358 		if (lwp->lwp_pcb.pcb_flags & ASYNC_HWERR) {
1359 			proc_t *p = ttoproc(curthread);
1360 			extern void print_msg_hwerr(ctid_t ct_id, proc_t *p);
1361 
1362 			lwp->lwp_pcb.pcb_flags &= ~ASYNC_HWERR;
1363 			print_msg_hwerr(p->p_ct_process->conp_contract.ct_id,
1364 			    p);
1365 			contract_process_hwerr(p->p_ct_process, p);
1366 			siginfo.si_signo = SIGKILL;
1367 			siginfo.si_code = SI_NOINFO;
1368 		} else if (lwp->lwp_pcb.pcb_flags & CPC_OVERFLOW) {
1369 			lwp->lwp_pcb.pcb_flags &= ~CPC_OVERFLOW;
1370 			if (kcpc_overflow_ast()) {
1371 				/*
1372 				 * Signal performance counter overflow
1373 				 */
1374 				if (tudebug)
1375 					showregs(type, rp, (caddr_t)0);
1376 				bzero(&siginfo, sizeof (siginfo));
1377 				siginfo.si_signo = SIGEMT;
1378 				siginfo.si_code = EMT_CPCOVF;
1379 				siginfo.si_addr = (caddr_t)rp->r_pc;
1380 				fault = FLTCPCOVF;
1381 			}
1382 		}
1383 
1384 		break;
1385 	}
1386 
1387 	/*
1388 	 * We can't get here from a system trap
1389 	 */
1390 	ASSERT(type & USER);
1391 
1392 	if (fault) {
1393 		/* We took a fault so abort single step. */
1394 		lwp->lwp_pcb.pcb_flags &= ~(NORMAL_STEP|WATCH_STEP);
1395 		/*
1396 		 * Remember the fault and fault adddress
1397 		 * for real-time (SIGPROF) profiling.
1398 		 */
1399 		lwp->lwp_lastfault = fault;
1400 		lwp->lwp_lastfaddr = siginfo.si_addr;
1401 
1402 		DTRACE_PROC2(fault, int, fault, ksiginfo_t *, &siginfo);
1403 
1404 		/*
1405 		 * If a debugger has declared this fault to be an
1406 		 * event of interest, stop the lwp.  Otherwise just
1407 		 * deliver the associated signal.
1408 		 */
1409 		if (siginfo.si_signo != SIGKILL &&
1410 		    prismember(&p->p_fltmask, fault) &&
1411 		    stop_on_fault(fault, &siginfo) == 0)
1412 			siginfo.si_signo = 0;
1413 	}
1414 
1415 	if (siginfo.si_signo)
1416 		trapsig(&siginfo, (fault != FLTFPE && fault != FLTCPCOVF));
1417 
1418 	if (lwp->lwp_oweupc)
1419 		profil_tick(rp->r_pc);
1420 
1421 	if (ct->t_astflag | ct->t_sig_check) {
1422 		/*
1423 		 * Turn off the AST flag before checking all the conditions that
1424 		 * may have caused an AST.  This flag is on whenever a signal or
1425 		 * unusual condition should be handled after the next trap or
1426 		 * syscall.
1427 		 */
1428 		astoff(ct);
1429 		/*
1430 		 * If a single-step trap occurred on a syscall (see above)
1431 		 * recognize it now.  Do this before checking for signals
1432 		 * because deferred_singlestep_trap() may generate a SIGTRAP to
1433 		 * the LWP or may otherwise mark the LWP to call issig(FORREAL).
1434 		 */
1435 		if (lwp->lwp_pcb.pcb_flags & DEBUG_PENDING)
1436 			deferred_singlestep_trap((caddr_t)rp->r_pc);
1437 
1438 		ct->t_sig_check = 0;
1439 
1440 		mutex_enter(&p->p_lock);
1441 		if (curthread->t_proc_flag & TP_CHANGEBIND) {
1442 			timer_lwpbind();
1443 			curthread->t_proc_flag &= ~TP_CHANGEBIND;
1444 		}
1445 		mutex_exit(&p->p_lock);
1446 
1447 		/*
1448 		 * for kaio requests that are on the per-process poll queue,
1449 		 * aiop->aio_pollq, they're AIO_POLL bit is set, the kernel
1450 		 * should copyout their result_t to user memory. by copying
1451 		 * out the result_t, the user can poll on memory waiting
1452 		 * for the kaio request to complete.
1453 		 */
1454 		if (p->p_aio)
1455 			aio_cleanup(0);
1456 		/*
1457 		 * If this LWP was asked to hold, call holdlwp(), which will
1458 		 * stop.  holdlwps() sets this up and calls pokelwps() which
1459 		 * sets the AST flag.
1460 		 *
1461 		 * Also check TP_EXITLWP, since this is used by fresh new LWPs
1462 		 * through lwp_rtt().  That flag is set if the lwp_create(2)
1463 		 * syscall failed after creating the LWP.
1464 		 */
1465 		if (ISHOLD(p))
1466 			holdlwp();
1467 
1468 		/*
1469 		 * All code that sets signals and makes ISSIG evaluate true must
1470 		 * set t_astflag afterwards.
1471 		 */
1472 		if (ISSIG_PENDING(ct, lwp, p)) {
1473 			if (issig(FORREAL))
1474 				psig();
1475 			ct->t_sig_check = 1;
1476 		}
1477 
1478 		if (ct->t_rprof != NULL) {
1479 			realsigprof(0, 0, 0);
1480 			ct->t_sig_check = 1;
1481 		}
1482 
1483 		/*
1484 		 * /proc can't enable/disable the trace bit itself
1485 		 * because that could race with the call gate used by
1486 		 * system calls via "lcall". If that happened, an
1487 		 * invalid EFLAGS would result. prstep()/prnostep()
1488 		 * therefore schedule an AST for the purpose.
1489 		 */
1490 		if (lwp->lwp_pcb.pcb_flags & REQUEST_STEP) {
1491 			lwp->lwp_pcb.pcb_flags &= ~REQUEST_STEP;
1492 			rp->r_ps |= PS_T;
1493 		}
1494 		if (lwp->lwp_pcb.pcb_flags & REQUEST_NOSTEP) {
1495 			lwp->lwp_pcb.pcb_flags &= ~REQUEST_NOSTEP;
1496 			rp->r_ps &= ~PS_T;
1497 		}
1498 	}
1499 
1500 out:	/* We can't get here from a system trap */
1501 	ASSERT(type & USER);
1502 
1503 	if (ISHOLD(p))
1504 		holdlwp();
1505 
1506 	/*
1507 	 * Set state to LWP_USER here so preempt won't give us a kernel
1508 	 * priority if it occurs after this point.  Call CL_TRAPRET() to
1509 	 * restore the user-level priority.
1510 	 *
1511 	 * It is important that no locks (other than spinlocks) be entered
1512 	 * after this point before returning to user mode (unless lwp_state
1513 	 * is set back to LWP_SYS).
1514 	 */
1515 	lwp->lwp_state = LWP_USER;
1516 
1517 	if (ct->t_trapret) {
1518 		ct->t_trapret = 0;
1519 		thread_lock(ct);
1520 		CL_TRAPRET(ct);
1521 		thread_unlock(ct);
1522 	}
1523 	if (CPU->cpu_runrun || curthread->t_schedflag & TS_ANYWAITQ)
1524 		preempt();
1525 	prunstop();
1526 	(void) new_mstate(ct, mstate);
1527 
1528 	/* Kernel probe */
1529 	TNF_PROBE_1(thread_state, "thread", /* CSTYLED */,
1530 	    tnf_microstate, state, LMS_USER);
1531 
1532 	return;
1533 
1534 cleanup:	/* system traps end up here */
1535 	ASSERT(!(type & USER));
1536 }
1537 
1538 /*
1539  * Patch non-zero to disable preemption of threads in the kernel.
1540  */
1541 int IGNORE_KERNEL_PREEMPTION = 0;	/* XXX - delete this someday */
1542 
1543 struct kpreempt_cnts {		/* kernel preemption statistics */
1544 	int	kpc_idle;	/* executing idle thread */
1545 	int	kpc_intr;	/* executing interrupt thread */
1546 	int	kpc_clock;	/* executing clock thread */
1547 	int	kpc_blocked;	/* thread has blocked preemption (t_preempt) */
1548 	int	kpc_notonproc;	/* thread is surrendering processor */
1549 	int	kpc_inswtch;	/* thread has ratified scheduling decision */
1550 	int	kpc_prilevel;	/* processor interrupt level is too high */
1551 	int	kpc_apreempt;	/* asynchronous preemption */
1552 	int	kpc_spreempt;	/* synchronous preemption */
1553 } kpreempt_cnts;
1554 
1555 /*
1556  * kernel preemption: forced rescheduling, preempt the running kernel thread.
1557  *	the argument is old PIL for an interrupt,
1558  *	or the distingished value KPREEMPT_SYNC.
1559  */
1560 void
1561 kpreempt(int asyncspl)
1562 {
1563 	kthread_t *ct = curthread;
1564 
1565 	if (IGNORE_KERNEL_PREEMPTION) {
1566 		aston(CPU->cpu_dispthread);
1567 		return;
1568 	}
1569 
1570 	/*
1571 	 * Check that conditions are right for kernel preemption
1572 	 */
1573 	do {
1574 		if (ct->t_preempt) {
1575 			/*
1576 			 * either a privileged thread (idle, panic, interrupt)
1577 			 * or will check when t_preempt is lowered
1578 			 * We need to specifically handle the case where
1579 			 * the thread is in the middle of swtch (resume has
1580 			 * been called) and has its t_preempt set
1581 			 * [idle thread and a thread which is in kpreempt
1582 			 * already] and then a high priority thread is
1583 			 * available in the local dispatch queue.
1584 			 * In this case the resumed thread needs to take a
1585 			 * trap so that it can call kpreempt. We achieve
1586 			 * this by using siron().
1587 			 * How do we detect this condition:
1588 			 * idle thread is running and is in the midst of
1589 			 * resume: curthread->t_pri == -1 && CPU->dispthread
1590 			 * != CPU->thread
1591 			 * Need to ensure that this happens only at high pil
1592 			 * resume is called at high pil
1593 			 * Only in resume_from_idle is the pil changed.
1594 			 */
1595 			if (ct->t_pri < 0) {
1596 				kpreempt_cnts.kpc_idle++;
1597 				if (CPU->cpu_dispthread != CPU->cpu_thread)
1598 					siron();
1599 			} else if (ct->t_flag & T_INTR_THREAD) {
1600 				kpreempt_cnts.kpc_intr++;
1601 				if (ct->t_pil == CLOCK_LEVEL)
1602 					kpreempt_cnts.kpc_clock++;
1603 			} else {
1604 				kpreempt_cnts.kpc_blocked++;
1605 				if (CPU->cpu_dispthread != CPU->cpu_thread)
1606 					siron();
1607 			}
1608 			aston(CPU->cpu_dispthread);
1609 			return;
1610 		}
1611 		if (ct->t_state != TS_ONPROC ||
1612 		    ct->t_disp_queue != CPU->cpu_disp) {
1613 			/* this thread will be calling swtch() shortly */
1614 			kpreempt_cnts.kpc_notonproc++;
1615 			if (CPU->cpu_thread != CPU->cpu_dispthread) {
1616 				/* already in swtch(), force another */
1617 				kpreempt_cnts.kpc_inswtch++;
1618 				siron();
1619 			}
1620 			return;
1621 		}
1622 		if (getpil() >= DISP_LEVEL) {
1623 			/*
1624 			 * We can't preempt this thread if it is at
1625 			 * a PIL >= DISP_LEVEL since it may be holding
1626 			 * a spin lock (like sched_lock).
1627 			 */
1628 			siron();	/* check back later */
1629 			kpreempt_cnts.kpc_prilevel++;
1630 			return;
1631 		}
1632 		if (!interrupts_enabled()) {
1633 			/*
1634 			 * Can't preempt while running with ints disabled
1635 			 */
1636 			kpreempt_cnts.kpc_prilevel++;
1637 			return;
1638 		}
1639 		if (asyncspl != KPREEMPT_SYNC)
1640 			kpreempt_cnts.kpc_apreempt++;
1641 		else
1642 			kpreempt_cnts.kpc_spreempt++;
1643 
1644 		ct->t_preempt++;
1645 		preempt();
1646 		ct->t_preempt--;
1647 	} while (CPU->cpu_kprunrun);
1648 }
1649 
1650 /*
1651  * Print out debugging info.
1652  */
1653 static void
1654 showregs(uint_t type, struct regs *rp, caddr_t addr)
1655 {
1656 	int s;
1657 
1658 	s = spl7();
1659 	type &= ~USER;
1660 	if (PTOU(curproc)->u_comm[0])
1661 		printf("%s: ", PTOU(curproc)->u_comm);
1662 	if (type < TRAP_TYPES)
1663 		printf("#%s %s\n", trap_type_mnemonic[type], trap_type[type]);
1664 	else
1665 		switch (type) {
1666 		case T_SYSCALL:
1667 			printf("Syscall Trap:\n");
1668 			break;
1669 		case T_AST:
1670 			printf("AST\n");
1671 			break;
1672 		default:
1673 			printf("Bad Trap = %d\n", type);
1674 			break;
1675 		}
1676 	if (type == T_PGFLT) {
1677 		printf("Bad %s fault at addr=0x%lx\n",
1678 		    USERMODE(rp->r_cs) ? "user": "kernel", (uintptr_t)addr);
1679 	} else if (addr) {
1680 		printf("addr=0x%lx\n", (uintptr_t)addr);
1681 	}
1682 
1683 	printf("pid=%d, pc=0x%lx, sp=0x%lx, eflags=0x%lx\n",
1684 	    (ttoproc(curthread) && ttoproc(curthread)->p_pidp) ?
1685 	    ttoproc(curthread)->p_pid : 0, rp->r_pc, rp->r_sp, rp->r_ps);
1686 
1687 #if defined(__lint)
1688 	/*
1689 	 * this clause can be deleted when lint bug 4870403 is fixed
1690 	 * (lint thinks that bit 32 is illegal in a %b format string)
1691 	 */
1692 	printf("cr0: %x cr4: %b\n",
1693 	    (uint_t)getcr0(), (uint_t)getcr4(), FMT_CR4);
1694 #else
1695 	printf("cr0: %b cr4: %b\n",
1696 	    (uint_t)getcr0(), FMT_CR0, (uint_t)getcr4(), FMT_CR4);
1697 #endif	/* __lint */
1698 
1699 	printf("cr2: %lx", getcr2());
1700 #if !defined(__xpv)
1701 	printf("cr3: %lx", getcr3());
1702 #if defined(__amd64)
1703 	printf("cr8: %lx\n", getcr8());
1704 #endif
1705 #endif
1706 	printf("\n");
1707 
1708 	dumpregs(rp);
1709 	splx(s);
1710 }
1711 
1712 static void
1713 dumpregs(struct regs *rp)
1714 {
1715 #if defined(__amd64)
1716 	const char fmt[] = "\t%3s: %16lx %3s: %16lx %3s: %16lx\n";
1717 
1718 	printf(fmt, "rdi", rp->r_rdi, "rsi", rp->r_rsi, "rdx", rp->r_rdx);
1719 	printf(fmt, "rcx", rp->r_rcx, " r8", rp->r_r8, " r9", rp->r_r9);
1720 	printf(fmt, "rax", rp->r_rax, "rbx", rp->r_rbx, "rbp", rp->r_rbp);
1721 	printf(fmt, "r10", rp->r_r10, "r11", rp->r_r11, "r12", rp->r_r12);
1722 	printf(fmt, "r13", rp->r_r13, "r14", rp->r_r14, "r15", rp->r_r15);
1723 
1724 	printf(fmt, "fsb", rdmsr(MSR_AMD_FSBASE), "gsb", rdmsr(MSR_AMD_GSBASE),
1725 	    " ds", rp->r_ds);
1726 	printf(fmt, " es", rp->r_es, " fs", rp->r_fs, " gs", rp->r_gs);
1727 
1728 	printf(fmt, "trp", rp->r_trapno, "err", rp->r_err, "rip", rp->r_rip);
1729 	printf(fmt, " cs", rp->r_cs, "rfl", rp->r_rfl, "rsp", rp->r_rsp);
1730 
1731 	printf("\t%3s: %16lx\n", " ss", rp->r_ss);
1732 
1733 #elif defined(__i386)
1734 	const char fmt[] = "\t%3s: %8lx %3s: %8lx %3s: %8lx %3s: %8lx\n";
1735 
1736 	printf(fmt, " gs", rp->r_gs, " fs", rp->r_fs,
1737 	    " es", rp->r_es, " ds", rp->r_ds);
1738 	printf(fmt, "edi", rp->r_edi, "esi", rp->r_esi,
1739 	    "ebp", rp->r_ebp, "esp", rp->r_esp);
1740 	printf(fmt, "ebx", rp->r_ebx, "edx", rp->r_edx,
1741 	    "ecx", rp->r_ecx, "eax", rp->r_eax);
1742 	printf(fmt, "trp", rp->r_trapno, "err", rp->r_err,
1743 	    "eip", rp->r_eip, " cs", rp->r_cs);
1744 	printf("\t%3s: %8lx %3s: %8lx %3s: %8lx\n",
1745 	    "efl", rp->r_efl, "usp", rp->r_uesp, " ss", rp->r_ss);
1746 
1747 #endif	/* __i386 */
1748 }
1749 
1750 /*
1751  * Test to see if the instruction is iret on i386 or iretq on amd64.
1752  *
1753  * On the hypervisor we can only test for nopop_sys_rtt_syscall. If true
1754  * then we are in the context of hypervisor's failsafe handler because it
1755  * tried to iret and failed due to a bad selector. See xen_failsafe_callback.
1756  */
1757 static int
1758 instr_is_iret(caddr_t pc)
1759 {
1760 
1761 #if defined(__xpv)
1762 	extern void nopop_sys_rtt_syscall(void);
1763 	return ((pc == (caddr_t)nopop_sys_rtt_syscall) ? 1 : 0);
1764 
1765 #else
1766 
1767 #if defined(__amd64)
1768 	static const uint8_t iret_insn[2] = { 0x48, 0xcf };	/* iretq */
1769 
1770 #elif defined(__i386)
1771 	static const uint8_t iret_insn[1] = { 0xcf };		/* iret */
1772 #endif	/* __i386 */
1773 	return (bcmp(pc, iret_insn, sizeof (iret_insn)) == 0);
1774 
1775 #endif	/* __xpv */
1776 }
1777 
1778 #if defined(__i386)
1779 
1780 /*
1781  * Test to see if the instruction is part of __SEGREGS_POP
1782  *
1783  * Note carefully the appallingly awful dependency between
1784  * the instruction sequence used in __SEGREGS_POP and these
1785  * instructions encoded here.
1786  */
1787 static int
1788 instr_is_segregs_pop(caddr_t pc)
1789 {
1790 	static const uint8_t movw_0_esp_gs[4] = { 0x8e, 0x6c, 0x24, 0x0 };
1791 	static const uint8_t movw_4_esp_fs[4] = { 0x8e, 0x64, 0x24, 0x4 };
1792 	static const uint8_t movw_8_esp_es[4] = { 0x8e, 0x44, 0x24, 0x8 };
1793 	static const uint8_t movw_c_esp_ds[4] = { 0x8e, 0x5c, 0x24, 0xc };
1794 
1795 	if (bcmp(pc, movw_0_esp_gs, sizeof (movw_0_esp_gs)) == 0 ||
1796 	    bcmp(pc, movw_4_esp_fs, sizeof (movw_4_esp_fs)) == 0 ||
1797 	    bcmp(pc, movw_8_esp_es, sizeof (movw_8_esp_es)) == 0 ||
1798 	    bcmp(pc, movw_c_esp_ds, sizeof (movw_c_esp_ds)) == 0)
1799 		return (1);
1800 
1801 	return (0);
1802 }
1803 
1804 #endif	/* __i386 */
1805 
1806 /*
1807  * Test to see if the instruction is part of _sys_rtt.
1808  *
1809  * Again on the hypervisor if we try to IRET to user land with a bad code
1810  * or stack selector we will get vectored through xen_failsafe_callback.
1811  * In which case we assume we got here via _sys_rtt since we only allow
1812  * IRET to user land to take place in _sys_rtt.
1813  */
1814 static int
1815 instr_is_sys_rtt(caddr_t pc)
1816 {
1817 	extern void _sys_rtt(), _sys_rtt_end();
1818 
1819 	if ((uintptr_t)pc < (uintptr_t)_sys_rtt ||
1820 	    (uintptr_t)pc > (uintptr_t)_sys_rtt_end)
1821 		return (0);
1822 
1823 	return (1);
1824 }
1825 
1826 /*
1827  * Handle #gp faults in kernel mode.
1828  *
1829  * One legitimate way this can happen is if we attempt to update segment
1830  * registers to naughty values on the way out of the kernel.
1831  *
1832  * This can happen in a couple of ways: someone - either accidentally or
1833  * on purpose - creates (setcontext(2), lwp_create(2)) or modifies
1834  * (signal(2)) a ucontext that contains silly segment register values.
1835  * Or someone - either accidentally or on purpose - modifies the prgregset_t
1836  * of a subject process via /proc to contain silly segment register values.
1837  *
1838  * (The unfortunate part is that we can end up discovering the bad segment
1839  * register value in the middle of an 'iret' after we've popped most of the
1840  * stack.  So it becomes quite difficult to associate an accurate ucontext
1841  * with the lwp, because the act of taking the #gp trap overwrites most of
1842  * what we were going to send the lwp.)
1843  *
1844  * OTOH if it turns out that's -not- the problem, and we're -not- an lwp
1845  * trying to return to user mode and we get a #gp fault, then we need
1846  * to die() -- which will happen if we return non-zero from this routine.
1847  */
1848 static int
1849 kern_gpfault(struct regs *rp)
1850 {
1851 	kthread_t *t = curthread;
1852 	proc_t *p = ttoproc(t);
1853 	klwp_t *lwp = ttolwp(t);
1854 	struct regs tmpregs, *trp = NULL;
1855 	caddr_t pc = (caddr_t)rp->r_pc;
1856 	int v;
1857 
1858 	/*
1859 	 * if we're not an lwp, or in the case of running native the
1860 	 * pc range is outside _sys_rtt, then we should immediately
1861 	 * be die()ing horribly.
1862 	 */
1863 	if (lwp == NULL || !instr_is_sys_rtt(pc))
1864 		return (1);
1865 
1866 	/*
1867 	 * So at least we're in the right part of the kernel.
1868 	 *
1869 	 * Disassemble the instruction at the faulting pc.
1870 	 * Once we know what it is, we carefully reconstruct the stack
1871 	 * based on the order in which the stack is deconstructed in
1872 	 * _sys_rtt. Ew.
1873 	 */
1874 	if (instr_is_iret(pc)) {
1875 		/*
1876 		 * We took the #gp while trying to perform the IRET.
1877 		 * This means that either %cs or %ss are bad.
1878 		 * All we know for sure is that most of the general
1879 		 * registers have been restored, including the
1880 		 * segment registers, and all we have left on the
1881 		 * topmost part of the lwp's stack are the
1882 		 * registers that the iretq was unable to consume.
1883 		 *
1884 		 * All the rest of the state was crushed by the #gp
1885 		 * which pushed -its- registers atop our old save area
1886 		 * (because we had to decrement the stack pointer, sigh) so
1887 		 * all that we can try and do is to reconstruct the
1888 		 * crushed frame from the #gp trap frame itself.
1889 		 */
1890 		trp = &tmpregs;
1891 		trp->r_ss = lwptoregs(lwp)->r_ss;
1892 		trp->r_sp = lwptoregs(lwp)->r_sp;
1893 		trp->r_ps = lwptoregs(lwp)->r_ps;
1894 		trp->r_cs = lwptoregs(lwp)->r_cs;
1895 		trp->r_pc = lwptoregs(lwp)->r_pc;
1896 		bcopy(rp, trp, offsetof(struct regs, r_pc));
1897 
1898 		/*
1899 		 * Validate simple math
1900 		 */
1901 		ASSERT(trp->r_pc == lwptoregs(lwp)->r_pc);
1902 		ASSERT(trp->r_err == rp->r_err);
1903 
1904 
1905 
1906 	}
1907 
1908 #if defined(__amd64)
1909 	if (trp == NULL && lwp->lwp_pcb.pcb_rupdate != 0) {
1910 
1911 		/*
1912 		 * This is the common case -- we're trying to load
1913 		 * a bad segment register value in the only section
1914 		 * of kernel code that ever loads segment registers.
1915 		 *
1916 		 * We don't need to do anything at this point because
1917 		 * the pcb contains all the pending segment register
1918 		 * state, and the regs are still intact because we
1919 		 * didn't adjust the stack pointer yet.  Given the fidelity
1920 		 * of all this, we could conceivably send a signal
1921 		 * to the lwp, rather than core-ing.
1922 		 */
1923 		trp = lwptoregs(lwp);
1924 		ASSERT((caddr_t)trp == (caddr_t)rp->r_sp);
1925 	}
1926 
1927 #elif defined(__i386)
1928 
1929 	if (trp == NULL && instr_is_segregs_pop(pc))
1930 		trp = lwptoregs(lwp);
1931 
1932 #endif	/* __i386 */
1933 
1934 	if (trp == NULL)
1935 		return (1);
1936 
1937 	/*
1938 	 * If we get to here, we're reasonably confident that we've
1939 	 * correctly decoded what happened on the way out of the kernel.
1940 	 * Rewrite the lwp's registers so that we can create a core dump
1941 	 * the (at least vaguely) represents the mcontext we were
1942 	 * being asked to restore when things went so terribly wrong.
1943 	 */
1944 
1945 	/*
1946 	 * Make sure that we have a meaningful %trapno and %err.
1947 	 */
1948 	trp->r_trapno = rp->r_trapno;
1949 	trp->r_err = rp->r_err;
1950 
1951 	if ((caddr_t)trp != (caddr_t)lwptoregs(lwp))
1952 		bcopy(trp, lwptoregs(lwp), sizeof (*trp));
1953 
1954 
1955 	mutex_enter(&p->p_lock);
1956 	lwp->lwp_cursig = SIGSEGV;
1957 	mutex_exit(&p->p_lock);
1958 
1959 	/*
1960 	 * Terminate all LWPs but don't discard them.  If another lwp beat
1961 	 * us to the punch by calling exit(), evaporate now.
1962 	 */
1963 	proc_is_exiting(p);
1964 	if (exitlwps(1) != 0) {
1965 		mutex_enter(&p->p_lock);
1966 		lwp_exit();
1967 	}
1968 
1969 	if (audit_active)		/* audit core dump */
1970 		audit_core_start(SIGSEGV);
1971 	v = core(SIGSEGV, B_FALSE);
1972 	if (audit_active)		/* audit core dump */
1973 		audit_core_finish(v ? CLD_KILLED : CLD_DUMPED);
1974 	exit(v ? CLD_KILLED : CLD_DUMPED, SIGSEGV);
1975 	return (0);
1976 }
1977 
1978 /*
1979  * dump_tss() - Display the TSS structure
1980  */
1981 
1982 #if !defined(__xpv)
1983 #if defined(__amd64)
1984 
1985 static void
1986 dump_tss(void)
1987 {
1988 	const char tss_fmt[] = "tss.%s:\t0x%p\n";  /* Format string */
1989 	struct tss *tss = CPU->cpu_tss;
1990 
1991 	printf(tss_fmt, "tss_rsp0", (void *)tss->tss_rsp0);
1992 	printf(tss_fmt, "tss_rsp1", (void *)tss->tss_rsp1);
1993 	printf(tss_fmt, "tss_rsp2", (void *)tss->tss_rsp2);
1994 
1995 	printf(tss_fmt, "tss_ist1", (void *)tss->tss_ist1);
1996 	printf(tss_fmt, "tss_ist2", (void *)tss->tss_ist2);
1997 	printf(tss_fmt, "tss_ist3", (void *)tss->tss_ist3);
1998 	printf(tss_fmt, "tss_ist4", (void *)tss->tss_ist4);
1999 	printf(tss_fmt, "tss_ist5", (void *)tss->tss_ist5);
2000 	printf(tss_fmt, "tss_ist6", (void *)tss->tss_ist6);
2001 	printf(tss_fmt, "tss_ist7", (void *)tss->tss_ist7);
2002 }
2003 
2004 #elif defined(__i386)
2005 
2006 static void
2007 dump_tss(void)
2008 {
2009 	const char tss_fmt[] = "tss.%s:\t0x%p\n";  /* Format string */
2010 	struct tss *tss = CPU->cpu_tss;
2011 
2012 	printf(tss_fmt, "tss_link", (void *)(uintptr_t)tss->tss_link);
2013 	printf(tss_fmt, "tss_esp0", (void *)(uintptr_t)tss->tss_esp0);
2014 	printf(tss_fmt, "tss_ss0", (void *)(uintptr_t)tss->tss_ss0);
2015 	printf(tss_fmt, "tss_esp1", (void *)(uintptr_t)tss->tss_esp1);
2016 	printf(tss_fmt, "tss_ss1", (void *)(uintptr_t)tss->tss_ss1);
2017 	printf(tss_fmt, "tss_esp2", (void *)(uintptr_t)tss->tss_esp2);
2018 	printf(tss_fmt, "tss_ss2", (void *)(uintptr_t)tss->tss_ss2);
2019 	printf(tss_fmt, "tss_cr3", (void *)(uintptr_t)tss->tss_cr3);
2020 	printf(tss_fmt, "tss_eip", (void *)(uintptr_t)tss->tss_eip);
2021 	printf(tss_fmt, "tss_eflags", (void *)(uintptr_t)tss->tss_eflags);
2022 	printf(tss_fmt, "tss_eax", (void *)(uintptr_t)tss->tss_eax);
2023 	printf(tss_fmt, "tss_ebx", (void *)(uintptr_t)tss->tss_ebx);
2024 	printf(tss_fmt, "tss_ecx", (void *)(uintptr_t)tss->tss_ecx);
2025 	printf(tss_fmt, "tss_edx", (void *)(uintptr_t)tss->tss_edx);
2026 	printf(tss_fmt, "tss_esp", (void *)(uintptr_t)tss->tss_esp);
2027 }
2028 
2029 #endif	/* __amd64 */
2030 #endif	/* !__xpv */
2031 
2032 #if defined(TRAPTRACE)
2033 
2034 int ttrace_nrec = 10;		/* number of records to dump out */
2035 int ttrace_dump_nregs = 0;	/* dump out this many records with regs too */
2036 
2037 /*
2038  * Dump out the last ttrace_nrec traptrace records on each CPU
2039  */
2040 static void
2041 dump_ttrace(void)
2042 {
2043 	trap_trace_ctl_t *ttc;
2044 	trap_trace_rec_t *rec;
2045 	uintptr_t current;
2046 	int i, j, k;
2047 	int n = NCPU;
2048 #if defined(__amd64)
2049 	const char banner[] =
2050 	    "\ncpu          address    timestamp "
2051 	    "type  vc  handler   pc\n";
2052 	const char fmt1[] = "%3d %016lx %12llx ";
2053 #elif defined(__i386)
2054 	const char banner[] =
2055 	    "\ncpu  address     timestamp type  vc  handler   pc\n";
2056 	const char fmt1[] = "%3d %08lx %12llx ";
2057 #endif
2058 	const char fmt2[] = "%4s %3x ";
2059 	const char fmt3[] = "%8s ";
2060 
2061 	if (ttrace_nrec == 0)
2062 		return;
2063 
2064 	printf(banner);
2065 
2066 	for (i = 0; i < n; i++) {
2067 		ttc = &trap_trace_ctl[i];
2068 		if (ttc->ttc_first == NULL)
2069 			continue;
2070 
2071 		current = ttc->ttc_next - sizeof (trap_trace_rec_t);
2072 		for (j = 0; j < ttrace_nrec; j++) {
2073 			struct sysent	*sys;
2074 			struct autovec	*vec;
2075 			extern struct av_head autovect[];
2076 			int type;
2077 			ulong_t	off;
2078 			char *sym, *stype;
2079 
2080 			if (current < ttc->ttc_first)
2081 				current =
2082 				    ttc->ttc_limit - sizeof (trap_trace_rec_t);
2083 
2084 			if (current == NULL)
2085 				continue;
2086 
2087 			rec = (trap_trace_rec_t *)current;
2088 
2089 			if (rec->ttr_stamp == 0)
2090 				break;
2091 
2092 			printf(fmt1, i, (uintptr_t)rec, rec->ttr_stamp);
2093 
2094 			switch (rec->ttr_marker) {
2095 			case TT_SYSCALL:
2096 			case TT_SYSENTER:
2097 			case TT_SYSC:
2098 			case TT_SYSC64:
2099 #if defined(__amd64)
2100 				sys = &sysent32[rec->ttr_sysnum];
2101 				switch (rec->ttr_marker) {
2102 				case TT_SYSC64:
2103 					sys = &sysent[rec->ttr_sysnum];
2104 					/*FALLTHROUGH*/
2105 #elif defined(__i386)
2106 				sys = &sysent[rec->ttr_sysnum];
2107 				switch (rec->ttr_marker) {
2108 				case TT_SYSC64:
2109 #endif
2110 				case TT_SYSC:
2111 					stype = "sysc";	/* syscall */
2112 					break;
2113 				case TT_SYSCALL:
2114 					stype = "lcal";	/* lcall */
2115 					break;
2116 				case TT_SYSENTER:
2117 					stype = "syse";	/* sysenter */
2118 					break;
2119 				default:
2120 					break;
2121 				}
2122 				printf(fmt2, "sysc", rec->ttr_sysnum);
2123 				if (sys != NULL) {
2124 					sym = kobj_getsymname(
2125 					    (uintptr_t)sys->sy_callc,
2126 					    &off);
2127 					if (sym != NULL)
2128 						printf(fmt3, sym);
2129 					else
2130 						printf("%p ", sys->sy_callc);
2131 				} else {
2132 					printf(fmt3, "unknown");
2133 				}
2134 				break;
2135 
2136 			case TT_INTERRUPT:
2137 				printf(fmt2, "intr", rec->ttr_vector);
2138 				vec = (&autovect[rec->ttr_vector])->avh_link;
2139 				if (vec != NULL) {
2140 					sym = kobj_getsymname(
2141 					    (uintptr_t)vec->av_vector, &off);
2142 					if (sym != NULL)
2143 						printf(fmt3, sym);
2144 					else
2145 						printf("%p ", vec->av_vector);
2146 				} else {
2147 					printf(fmt3, "unknown ");
2148 				}
2149 				break;
2150 
2151 			case TT_TRAP:
2152 			case TT_EVENT:
2153 				type = rec->ttr_regs.r_trapno;
2154 				printf(fmt2, "trap", type);
2155 				if (type < TRAP_TYPES)
2156 					printf("     #%s ",
2157 					    trap_type_mnemonic[type]);
2158 				else
2159 					switch (type) {
2160 					case T_AST:
2161 						printf(fmt3, "ast");
2162 						break;
2163 					default:
2164 						printf(fmt3, "");
2165 						break;
2166 					}
2167 				break;
2168 
2169 			default:
2170 				break;
2171 			}
2172 
2173 			sym = kobj_getsymname(rec->ttr_regs.r_pc, &off);
2174 			if (sym != NULL)
2175 				printf("%s+%lx\n", sym, off);
2176 			else
2177 				printf("%lx\n", rec->ttr_regs.r_pc);
2178 
2179 			if (ttrace_dump_nregs-- > 0) {
2180 				int s;
2181 
2182 				if (rec->ttr_marker == TT_INTERRUPT)
2183 					printf(
2184 					    "\t\tipl %x spl %x pri %x\n",
2185 					    rec->ttr_ipl,
2186 					    rec->ttr_spl,
2187 					    rec->ttr_pri);
2188 
2189 				dumpregs(&rec->ttr_regs);
2190 
2191 				printf("\t%3s: %p\n\n", " ct",
2192 				    (void *)rec->ttr_curthread);
2193 
2194 				/*
2195 				 * print out the pc stack that we recorded
2196 				 * at trap time (if any)
2197 				 */
2198 				for (s = 0; s < rec->ttr_sdepth; s++) {
2199 					uintptr_t fullpc;
2200 
2201 					if (s >= TTR_STACK_DEPTH) {
2202 						printf("ttr_sdepth corrupt\n");
2203 						break;
2204 					}
2205 
2206 					fullpc = (uintptr_t)rec->ttr_stack[s];
2207 
2208 					sym = kobj_getsymname(fullpc, &off);
2209 					if (sym != NULL)
2210 						printf("-> %s+0x%lx()\n",
2211 						    sym, off);
2212 					else
2213 						printf("-> 0x%lx()\n", fullpc);
2214 				}
2215 				printf("\n");
2216 			}
2217 			current -= sizeof (trap_trace_rec_t);
2218 		}
2219 	}
2220 }
2221 
2222 #endif	/* TRAPTRACE */
2223 
2224 void
2225 panic_showtrap(struct panic_trap_info *tip)
2226 {
2227 	showregs(tip->trap_type, tip->trap_regs, tip->trap_addr);
2228 
2229 #if defined(TRAPTRACE)
2230 	dump_ttrace();
2231 #endif
2232 
2233 #if !defined(__xpv)
2234 	if (tip->trap_type == T_DBLFLT)
2235 		dump_tss();
2236 #endif
2237 }
2238 
2239 void
2240 panic_savetrap(panic_data_t *pdp, struct panic_trap_info *tip)
2241 {
2242 	panic_saveregs(pdp, tip->trap_regs);
2243 }
2244