xref: /illumos-gate/usr/src/uts/i86pc/os/trap.c (revision fd6d41c5025e9fb45a115fc82d86e9983d1e9fd6)
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 (c) 1992, 2010, Oracle and/or its affiliates. All rights reserved.
24  */
25 
26 /*	Copyright (c) 1990, 1991 UNIX System Laboratories, Inc. */
27 /*	Copyright (c) 1984, 1986, 1987, 1988, 1989, 1990 AT&T   */
28 /*		All Rights Reserved   				*/
29 /*								*/
30 /*	Copyright (c) 1987, 1988 Microsoft Corporation  	*/
31 /*		All Rights Reserved   				*/
32 /*								*/
33 
34 /*
35  * Copyright 2018 Joyent, Inc.
36  */
37 
38 #include <sys/types.h>
39 #include <sys/sysmacros.h>
40 #include <sys/param.h>
41 #include <sys/signal.h>
42 #include <sys/systm.h>
43 #include <sys/user.h>
44 #include <sys/proc.h>
45 #include <sys/disp.h>
46 #include <sys/class.h>
47 #include <sys/core.h>
48 #include <sys/syscall.h>
49 #include <sys/cpuvar.h>
50 #include <sys/vm.h>
51 #include <sys/sysinfo.h>
52 #include <sys/fault.h>
53 #include <sys/stack.h>
54 #include <sys/psw.h>
55 #include <sys/regset.h>
56 #include <sys/fp.h>
57 #include <sys/trap.h>
58 #include <sys/kmem.h>
59 #include <sys/vtrace.h>
60 #include <sys/cmn_err.h>
61 #include <sys/prsystm.h>
62 #include <sys/mutex_impl.h>
63 #include <sys/machsystm.h>
64 #include <sys/archsystm.h>
65 #include <sys/sdt.h>
66 #include <sys/avintr.h>
67 #include <sys/kobj.h>
68 
69 #include <vm/hat.h>
70 
71 #include <vm/seg_kmem.h>
72 #include <vm/as.h>
73 #include <vm/seg.h>
74 #include <vm/hat_pte.h>
75 #include <vm/hat_i86.h>
76 
77 #include <sys/procfs.h>
78 
79 #include <sys/reboot.h>
80 #include <sys/debug.h>
81 #include <sys/debugreg.h>
82 #include <sys/modctl.h>
83 #include <sys/aio_impl.h>
84 #include <sys/tnf.h>
85 #include <sys/tnf_probe.h>
86 #include <sys/cred.h>
87 #include <sys/mman.h>
88 #include <sys/x86_archext.h>
89 #include <sys/copyops.h>
90 #include <c2/audit.h>
91 #include <sys/ftrace.h>
92 #include <sys/panic.h>
93 #include <sys/traptrace.h>
94 #include <sys/ontrap.h>
95 #include <sys/cpc_impl.h>
96 #include <sys/bootconf.h>
97 #include <sys/bootinfo.h>
98 #include <sys/promif.h>
99 #include <sys/mach_mmu.h>
100 #if defined(__xpv)
101 #include <sys/hypervisor.h>
102 #endif
103 #include <sys/contract/process_impl.h>
104 
105 #define	USER	0x10000		/* user-mode flag added to trap type */
106 
107 static const char *trap_type_mnemonic[] = {
108 	"de",	"db",	"2",	"bp",
109 	"of",	"br",	"ud",	"nm",
110 	"df",	"9",	"ts",	"np",
111 	"ss",	"gp",	"pf",	"15",
112 	"mf",	"ac",	"mc",	"xf"
113 };
114 
115 static const char *trap_type[] = {
116 	"Divide error",				/* trap id 0 	*/
117 	"Debug",				/* trap id 1	*/
118 	"NMI interrupt",			/* trap id 2	*/
119 	"Breakpoint",				/* trap id 3 	*/
120 	"Overflow",				/* trap id 4 	*/
121 	"BOUND range exceeded",			/* trap id 5 	*/
122 	"Invalid opcode",			/* trap id 6 	*/
123 	"Device not available",			/* trap id 7 	*/
124 	"Double fault",				/* trap id 8 	*/
125 	"Coprocessor segment overrun",		/* trap id 9 	*/
126 	"Invalid TSS",				/* trap id 10 	*/
127 	"Segment not present",			/* trap id 11 	*/
128 	"Stack segment fault",			/* trap id 12 	*/
129 	"General protection",			/* trap id 13 	*/
130 	"Page fault",				/* trap id 14 	*/
131 	"Reserved",				/* trap id 15 	*/
132 	"x87 floating point error",		/* trap id 16 	*/
133 	"Alignment check",			/* trap id 17 	*/
134 	"Machine check",			/* trap id 18	*/
135 	"SIMD floating point exception",	/* trap id 19	*/
136 };
137 
138 #define	TRAP_TYPES	(sizeof (trap_type) / sizeof (trap_type[0]))
139 
140 #define	SLOW_SCALL_SIZE	2
141 #define	FAST_SCALL_SIZE	2
142 
143 int tudebug = 0;
144 int tudebugbpt = 0;
145 int tudebugfpe = 0;
146 int tudebugsse = 0;
147 
148 #if defined(TRAPDEBUG) || defined(lint)
149 int tdebug = 0;
150 int lodebug = 0;
151 int faultdebug = 0;
152 #else
153 #define	tdebug	0
154 #define	lodebug	0
155 #define	faultdebug	0
156 #endif /* defined(TRAPDEBUG) || defined(lint) */
157 
158 #if defined(TRAPTRACE)
159 /*
160  * trap trace record for cpu0 is allocated here.
161  * trap trace records for non-boot cpus are allocated in mp_startup_init().
162  */
163 static trap_trace_rec_t trap_tr0[TRAPTR_NENT];
164 trap_trace_ctl_t trap_trace_ctl[NCPU] = {
165 	{
166 	    (uintptr_t)trap_tr0,			/* next record */
167 	    (uintptr_t)trap_tr0,			/* first record */
168 	    (uintptr_t)(trap_tr0 + TRAPTR_NENT),	/* limit */
169 	    (uintptr_t)0				/* current */
170 	},
171 };
172 
173 /*
174  * default trap buffer size
175  */
176 size_t trap_trace_bufsize = TRAPTR_NENT * sizeof (trap_trace_rec_t);
177 int trap_trace_freeze = 0;
178 int trap_trace_off = 0;
179 
180 /*
181  * A dummy TRAPTRACE entry to use after death.
182  */
183 trap_trace_rec_t trap_trace_postmort;
184 
185 static void dump_ttrace(void);
186 #endif	/* TRAPTRACE */
187 static void dumpregs(struct regs *);
188 static void showregs(uint_t, struct regs *, caddr_t);
189 static int kern_gpfault(struct regs *);
190 
191 /*ARGSUSED*/
192 static int
193 die(uint_t type, struct regs *rp, caddr_t addr, processorid_t cpuid)
194 {
195 	struct panic_trap_info ti;
196 	const char *trap_name, *trap_mnemonic;
197 
198 	if (type < TRAP_TYPES) {
199 		trap_name = trap_type[type];
200 		trap_mnemonic = trap_type_mnemonic[type];
201 	} else {
202 		trap_name = "trap";
203 		trap_mnemonic = "-";
204 	}
205 
206 #ifdef TRAPTRACE
207 	TRAPTRACE_FREEZE;
208 #endif
209 
210 	ti.trap_regs = rp;
211 	ti.trap_type = type & ~USER;
212 	ti.trap_addr = addr;
213 
214 	curthread->t_panic_trap = &ti;
215 
216 	if (type == T_PGFLT && addr < (caddr_t)kernelbase) {
217 		panic("BAD TRAP: type=%x (#%s %s) rp=%p addr=%p "
218 		    "occurred in module \"%s\" due to %s",
219 		    type, trap_mnemonic, trap_name, (void *)rp, (void *)addr,
220 		    mod_containing_pc((caddr_t)rp->r_pc),
221 		    addr < (caddr_t)PAGESIZE ?
222 		    "a NULL pointer dereference" :
223 		    "an illegal access to a user address");
224 	} else
225 		panic("BAD TRAP: type=%x (#%s %s) rp=%p addr=%p",
226 		    type, trap_mnemonic, trap_name, (void *)rp, (void *)addr);
227 	return (0);
228 }
229 
230 /*
231  * Rewrite the instruction at pc to be an int $T_SYSCALLINT instruction.
232  *
233  * int <vector> is two bytes: 0xCD <vector>
234  */
235 
236 static int
237 rewrite_syscall(caddr_t pc)
238 {
239 	uchar_t instr[SLOW_SCALL_SIZE] = { 0xCD, T_SYSCALLINT };
240 
241 	if (uwrite(curthread->t_procp, instr, SLOW_SCALL_SIZE,
242 	    (uintptr_t)pc) != 0)
243 		return (1);
244 
245 	return (0);
246 }
247 
248 /*
249  * Test to see if the instruction at pc is sysenter or syscall. The second
250  * argument should be the x86 feature flag corresponding to the expected
251  * instruction.
252  *
253  * sysenter is two bytes: 0x0F 0x34
254  * syscall is two bytes:  0x0F 0x05
255  * int $T_SYSCALLINT is two bytes: 0xCD 0x91
256  */
257 
258 static int
259 instr_is_other_syscall(caddr_t pc, int which)
260 {
261 	uchar_t instr[FAST_SCALL_SIZE];
262 
263 	ASSERT(which == X86FSET_SEP || which == X86FSET_ASYSC || which == 0xCD);
264 
265 	if (copyin_nowatch(pc, (caddr_t)instr, FAST_SCALL_SIZE) != 0)
266 		return (0);
267 
268 	switch (which) {
269 	case X86FSET_SEP:
270 		if (instr[0] == 0x0F && instr[1] == 0x34)
271 			return (1);
272 		break;
273 	case X86FSET_ASYSC:
274 		if (instr[0] == 0x0F && instr[1] == 0x05)
275 			return (1);
276 		break;
277 	case 0xCD:
278 		if (instr[0] == 0xCD && instr[1] == T_SYSCALLINT)
279 			return (1);
280 		break;
281 	}
282 
283 	return (0);
284 }
285 
286 static const char *
287 syscall_insn_string(int syscall_insn)
288 {
289 	switch (syscall_insn) {
290 	case X86FSET_SEP:
291 		return ("sysenter");
292 	case X86FSET_ASYSC:
293 		return ("syscall");
294 	case 0xCD:
295 		return ("int");
296 	default:
297 		return ("Unknown");
298 	}
299 }
300 
301 static int
302 ldt_rewrite_syscall(struct regs *rp, proc_t *p, int syscall_insn)
303 {
304 	caddr_t	linearpc;
305 	int return_code = 0;
306 
307 	mutex_enter(&p->p_ldtlock);	/* Must be held across linear_pc() */
308 
309 	if (linear_pc(rp, p, &linearpc) == 0) {
310 
311 		/*
312 		 * If another thread beat us here, it already changed
313 		 * this site to the slower (int) syscall instruction.
314 		 */
315 		if (instr_is_other_syscall(linearpc, 0xCD)) {
316 			return_code = 1;
317 		} else if (instr_is_other_syscall(linearpc, syscall_insn)) {
318 
319 			if (rewrite_syscall(linearpc) == 0) {
320 				return_code = 1;
321 			}
322 #ifdef DEBUG
323 			else
324 				cmn_err(CE_WARN, "failed to rewrite %s "
325 				    "instruction in process %d",
326 				    syscall_insn_string(syscall_insn),
327 				    p->p_pid);
328 #endif /* DEBUG */
329 		}
330 	}
331 
332 	mutex_exit(&p->p_ldtlock);	/* Must be held across linear_pc() */
333 
334 	return (return_code);
335 }
336 
337 /*
338  * Test to see if the instruction at pc is a system call instruction.
339  *
340  * The bytes of an lcall instruction used for the syscall trap.
341  * static uchar_t lcall[7] = { 0x9a, 0, 0, 0, 0, 0x7, 0 };
342  * static uchar_t lcallalt[7] = { 0x9a, 0, 0, 0, 0, 0x27, 0 };
343  */
344 
345 #define	LCALLSIZE	7
346 
347 static int
348 instr_is_lcall_syscall(caddr_t pc)
349 {
350 	uchar_t instr[LCALLSIZE];
351 
352 	if (copyin_nowatch(pc, (caddr_t)instr, LCALLSIZE) == 0 &&
353 	    instr[0] == 0x9a &&
354 	    instr[1] == 0 &&
355 	    instr[2] == 0 &&
356 	    instr[3] == 0 &&
357 	    instr[4] == 0 &&
358 	    (instr[5] == 0x7 || instr[5] == 0x27) &&
359 	    instr[6] == 0)
360 		return (1);
361 
362 	return (0);
363 }
364 
365 #ifdef __amd64
366 
367 /*
368  * In the first revisions of amd64 CPUs produced by AMD, the LAHF and
369  * SAHF instructions were not implemented in 64-bit mode. Later revisions
370  * did implement these instructions. An extension to the cpuid instruction
371  * was added to check for the capability of executing these instructions
372  * in 64-bit mode.
373  *
374  * Intel originally did not implement these instructions in EM64T either,
375  * but added them in later revisions.
376  *
377  * So, there are different chip revisions by both vendors out there that
378  * may or may not implement these instructions. The easy solution is to
379  * just always emulate these instructions on demand.
380  *
381  * SAHF == store %ah in the lower 8 bits of %rflags (opcode 0x9e)
382  * LAHF == load the lower 8 bits of %rflags into %ah (opcode 0x9f)
383  */
384 
385 #define	LSAHFSIZE 1
386 
387 static int
388 instr_is_lsahf(caddr_t pc, uchar_t *instr)
389 {
390 	if (copyin_nowatch(pc, (caddr_t)instr, LSAHFSIZE) == 0 &&
391 	    (*instr == 0x9e || *instr == 0x9f))
392 		return (1);
393 	return (0);
394 }
395 
396 /*
397  * Emulate the LAHF and SAHF instructions. The reference manuals define
398  * these instructions to always load/store bit 1 as a 1, and bits 3 and 5
399  * as a 0. The other, defined, bits are copied (the PS_ICC bits and PS_P).
400  *
401  * Note that %ah is bits 8-15 of %rax.
402  */
403 static void
404 emulate_lsahf(struct regs *rp, uchar_t instr)
405 {
406 	if (instr == 0x9e) {
407 		/* sahf. Copy bits from %ah to flags. */
408 		rp->r_ps = (rp->r_ps & ~0xff) |
409 		    ((rp->r_rax >> 8) & PSL_LSAHFMASK) | PS_MB1;
410 	} else {
411 		/* lahf. Copy bits from flags to %ah. */
412 		rp->r_rax = (rp->r_rax & ~0xff00) |
413 		    (((rp->r_ps & PSL_LSAHFMASK) | PS_MB1) << 8);
414 	}
415 	rp->r_pc += LSAHFSIZE;
416 }
417 #endif /* __amd64 */
418 
419 #ifdef OPTERON_ERRATUM_91
420 
421 /*
422  * Test to see if the instruction at pc is a prefetch instruction.
423  *
424  * The first byte of prefetch instructions is always 0x0F.
425  * The second byte is 0x18 for regular prefetch or 0x0D for AMD 3dnow prefetch.
426  * The third byte (ModRM) contains the register field bits (bits 3-5).
427  * These bits must be between 0 and 3 inclusive for regular prefetch and
428  * 0 and 1 inclusive for AMD 3dnow prefetch.
429  *
430  * In 64-bit mode, there may be a one-byte REX prefex (0x40-0x4F).
431  */
432 
433 static int
434 cmp_to_prefetch(uchar_t *p)
435 {
436 #ifdef _LP64
437 	if ((p[0] & 0xF0) == 0x40)	/* 64-bit REX prefix */
438 		p++;
439 #endif
440 	return ((p[0] == 0x0F && p[1] == 0x18 && ((p[2] >> 3) & 7) <= 3) ||
441 	    (p[0] == 0x0F && p[1] == 0x0D && ((p[2] >> 3) & 7) <= 1));
442 }
443 
444 static int
445 instr_is_prefetch(caddr_t pc)
446 {
447 	uchar_t instr[4];	/* optional REX prefix plus 3-byte opcode */
448 
449 	return (copyin_nowatch(pc, instr, sizeof (instr)) == 0 &&
450 	    cmp_to_prefetch(instr));
451 }
452 
453 #endif /* OPTERON_ERRATUM_91 */
454 
455 /*
456  * Called from the trap handler when a processor trap occurs.
457  *
458  * Note: All user-level traps that might call stop() must exit
459  * trap() by 'goto out' or by falling through.
460  * Note Also: trap() is usually called with interrupts enabled, (PS_IE == 1)
461  * however, there are paths that arrive here with PS_IE == 0 so special care
462  * must be taken in those cases.
463  */
464 void
465 trap(struct regs *rp, caddr_t addr, processorid_t cpuid)
466 {
467 	kthread_t *ct = curthread;
468 	enum seg_rw rw;
469 	unsigned type;
470 	proc_t *p = ttoproc(ct);
471 	klwp_t *lwp = ttolwp(ct);
472 	uintptr_t lofault;
473 	label_t *onfault;
474 	faultcode_t pagefault(), res, errcode;
475 	enum fault_type fault_type;
476 	k_siginfo_t siginfo;
477 	uint_t fault = 0;
478 	int mstate;
479 	int sicode = 0;
480 	int watchcode;
481 	int watchpage;
482 	caddr_t vaddr;
483 	size_t sz;
484 	int ta;
485 #ifdef __amd64
486 	uchar_t instr;
487 #endif
488 
489 	ASSERT_STACK_ALIGNED();
490 
491 	type = rp->r_trapno;
492 	CPU_STATS_ADDQ(CPU, sys, trap, 1);
493 	ASSERT(ct->t_schedflag & TS_DONT_SWAP);
494 
495 	if (type == T_PGFLT) {
496 
497 		errcode = rp->r_err;
498 		if (errcode & PF_ERR_WRITE)
499 			rw = S_WRITE;
500 		else if ((caddr_t)rp->r_pc == addr ||
501 		    (mmu.pt_nx != 0 && (errcode & PF_ERR_EXEC)))
502 			rw = S_EXEC;
503 		else
504 			rw = S_READ;
505 
506 #if defined(__i386)
507 		/*
508 		 * Pentium Pro work-around
509 		 */
510 		if ((errcode & PF_ERR_PROT) && pentiumpro_bug4046376) {
511 			uint_t	attr;
512 			uint_t	priv_violation;
513 			uint_t	access_violation;
514 
515 			if (hat_getattr(addr < (caddr_t)kernelbase ?
516 			    curproc->p_as->a_hat : kas.a_hat, addr, &attr)
517 			    == -1) {
518 				errcode &= ~PF_ERR_PROT;
519 			} else {
520 				priv_violation = (errcode & PF_ERR_USER) &&
521 				    !(attr & PROT_USER);
522 				access_violation = (errcode & PF_ERR_WRITE) &&
523 				    !(attr & PROT_WRITE);
524 				if (!priv_violation && !access_violation)
525 					goto cleanup;
526 			}
527 		}
528 #endif /* __i386 */
529 
530 	} else if (type == T_SGLSTP && lwp != NULL)
531 		lwp->lwp_pcb.pcb_drstat = (uintptr_t)addr;
532 
533 	if (tdebug)
534 		showregs(type, rp, addr);
535 
536 	if (USERMODE(rp->r_cs)) {
537 		/*
538 		 * Set up the current cred to use during this trap. u_cred
539 		 * no longer exists.  t_cred is used instead.
540 		 * The current process credential applies to the thread for
541 		 * the entire trap.  If trapping from the kernel, this
542 		 * should already be set up.
543 		 */
544 		if (ct->t_cred != p->p_cred) {
545 			cred_t *oldcred = ct->t_cred;
546 			/*
547 			 * DTrace accesses t_cred in probe context.  t_cred
548 			 * must always be either NULL, or point to a valid,
549 			 * allocated cred structure.
550 			 */
551 			ct->t_cred = crgetcred();
552 			crfree(oldcred);
553 		}
554 		ASSERT(lwp != NULL);
555 		type |= USER;
556 		ASSERT(lwptoregs(lwp) == rp);
557 		lwp->lwp_state = LWP_SYS;
558 
559 		switch (type) {
560 		case T_PGFLT + USER:
561 			if ((caddr_t)rp->r_pc == addr)
562 				mstate = LMS_TFAULT;
563 			else
564 				mstate = LMS_DFAULT;
565 			break;
566 		default:
567 			mstate = LMS_TRAP;
568 			break;
569 		}
570 		/* Kernel probe */
571 		TNF_PROBE_1(thread_state, "thread", /* CSTYLED */,
572 		    tnf_microstate, state, mstate);
573 		mstate = new_mstate(ct, mstate);
574 
575 		bzero(&siginfo, sizeof (siginfo));
576 	}
577 
578 	switch (type) {
579 	case T_PGFLT + USER:
580 	case T_SGLSTP:
581 	case T_SGLSTP + USER:
582 	case T_BPTFLT + USER:
583 		break;
584 
585 	default:
586 		FTRACE_2("trap(): type=0x%lx, regs=0x%lx",
587 		    (ulong_t)type, (ulong_t)rp);
588 		break;
589 	}
590 
591 	switch (type) {
592 	case T_SIMDFPE:
593 		/* Make sure we enable interrupts before die()ing */
594 		sti();	/* The SIMD exception comes in via cmninttrap */
595 		/*FALLTHROUGH*/
596 	default:
597 		if (type & USER) {
598 			if (tudebug)
599 				showregs(type, rp, (caddr_t)0);
600 			printf("trap: Unknown trap type %d in user mode\n",
601 			    type & ~USER);
602 			siginfo.si_signo = SIGILL;
603 			siginfo.si_code  = ILL_ILLTRP;
604 			siginfo.si_addr  = (caddr_t)rp->r_pc;
605 			siginfo.si_trapno = type & ~USER;
606 			fault = FLTILL;
607 		} else {
608 			(void) die(type, rp, addr, cpuid);
609 			/*NOTREACHED*/
610 		}
611 		break;
612 
613 	case T_PGFLT:		/* system page fault */
614 		/*
615 		 * If we're under on_trap() protection (see <sys/ontrap.h>),
616 		 * set ot_trap and bounce back to the on_trap() call site
617 		 * via the installed trampoline.
618 		 */
619 		if ((ct->t_ontrap != NULL) &&
620 		    (ct->t_ontrap->ot_prot & OT_DATA_ACCESS)) {
621 			ct->t_ontrap->ot_trap |= OT_DATA_ACCESS;
622 			rp->r_pc = ct->t_ontrap->ot_trampoline;
623 			goto cleanup;
624 		}
625 
626 		/*
627 		 * If we have an Instruction fault in kernel mode, then that
628 		 * means we've tried to execute a user page (SMEP) or both of
629 		 * PAE and NXE are enabled. In either case, given that it's a
630 		 * kernel fault, we should panic immediately and not try to make
631 		 * any more forward progress. This indicates a bug in the
632 		 * kernel, which if execution continued, could be exploited to
633 		 * wreak havoc on the system.
634 		 */
635 		if (errcode & PF_ERR_EXEC) {
636 			(void) die(type, rp, addr, cpuid);
637 		}
638 
639 		/*
640 		 * We need to check if SMAP is in play. If SMAP is in play, then
641 		 * any access to a user page will show up as a protection
642 		 * violation. To see if SMAP is enabled we first check if it's a
643 		 * user address and whether we have the feature flag set. If we
644 		 * do and the interrupted registers do not allow for user
645 		 * accesses (PS_ACHK is not enabled), then we need to die
646 		 * immediately.
647 		 */
648 		if (addr < (caddr_t)kernelbase &&
649 		    is_x86_feature(x86_featureset, X86FSET_SMAP) == B_TRUE &&
650 		    (rp->r_ps & PS_ACHK) == 0) {
651 			(void) die(type, rp, addr, cpuid);
652 		}
653 
654 		/*
655 		 * See if we can handle as pagefault. Save lofault and onfault
656 		 * across this. Here we assume that an address less than
657 		 * KERNELBASE is a user fault.  We can do this as copy.s
658 		 * routines verify that the starting address is less than
659 		 * KERNELBASE before starting and because we know that we
660 		 * always have KERNELBASE mapped as invalid to serve as a
661 		 * "barrier".
662 		 */
663 		lofault = ct->t_lofault;
664 		onfault = ct->t_onfault;
665 		ct->t_lofault = 0;
666 
667 		mstate = new_mstate(ct, LMS_KFAULT);
668 
669 		if (addr < (caddr_t)kernelbase) {
670 			res = pagefault(addr,
671 			    (errcode & PF_ERR_PROT)? F_PROT: F_INVAL, rw, 0);
672 			if (res == FC_NOMAP &&
673 			    addr < p->p_usrstack &&
674 			    grow(addr))
675 				res = 0;
676 		} else {
677 			res = pagefault(addr,
678 			    (errcode & PF_ERR_PROT)? F_PROT: F_INVAL, rw, 1);
679 		}
680 		(void) new_mstate(ct, mstate);
681 
682 		/*
683 		 * Restore lofault and onfault. If we resolved the fault, exit.
684 		 * If we didn't and lofault wasn't set, die.
685 		 */
686 		ct->t_lofault = lofault;
687 		ct->t_onfault = onfault;
688 		if (res == 0)
689 			goto cleanup;
690 
691 #if defined(OPTERON_ERRATUM_93) && defined(_LP64)
692 		if (lofault == 0 && opteron_erratum_93) {
693 			/*
694 			 * Workaround for Opteron Erratum 93. On return from
695 			 * a System Managment Interrupt at a HLT instruction
696 			 * the %rip might be truncated to a 32 bit value.
697 			 * BIOS is supposed to fix this, but some don't.
698 			 * If this occurs we simply restore the high order bits.
699 			 * The HLT instruction is 1 byte of 0xf4.
700 			 */
701 			uintptr_t	rip = rp->r_pc;
702 
703 			if ((rip & 0xfffffffful) == rip) {
704 				rip |= 0xfffffffful << 32;
705 				if (hat_getpfnum(kas.a_hat, (caddr_t)rip) !=
706 				    PFN_INVALID &&
707 				    (*(uchar_t *)rip == 0xf4 ||
708 				    *(uchar_t *)(rip - 1) == 0xf4)) {
709 					rp->r_pc = rip;
710 					goto cleanup;
711 				}
712 			}
713 		}
714 #endif /* OPTERON_ERRATUM_93 && _LP64 */
715 
716 #ifdef OPTERON_ERRATUM_91
717 		if (lofault == 0 && opteron_erratum_91) {
718 			/*
719 			 * Workaround for Opteron Erratum 91. Prefetches may
720 			 * generate a page fault (they're not supposed to do
721 			 * that!). If this occurs we simply return back to the
722 			 * instruction.
723 			 */
724 			caddr_t		pc = (caddr_t)rp->r_pc;
725 
726 			/*
727 			 * If the faulting PC is not mapped, this is a
728 			 * legitimate kernel page fault that must result in a
729 			 * panic. If the faulting PC is mapped, it could contain
730 			 * a prefetch instruction. Check for that here.
731 			 */
732 			if (hat_getpfnum(kas.a_hat, pc) != PFN_INVALID) {
733 				if (cmp_to_prefetch((uchar_t *)pc)) {
734 #ifdef DEBUG
735 					cmn_err(CE_WARN, "Opteron erratum 91 "
736 					    "occurred: kernel prefetch"
737 					    " at %p generated a page fault!",
738 					    (void *)rp->r_pc);
739 #endif /* DEBUG */
740 					goto cleanup;
741 				}
742 			}
743 			(void) die(type, rp, addr, cpuid);
744 		}
745 #endif /* OPTERON_ERRATUM_91 */
746 
747 		if (lofault == 0)
748 			(void) die(type, rp, addr, cpuid);
749 
750 		/*
751 		 * Cannot resolve fault.  Return to lofault.
752 		 */
753 		if (lodebug) {
754 			showregs(type, rp, addr);
755 			traceregs(rp);
756 		}
757 		if (FC_CODE(res) == FC_OBJERR)
758 			res = FC_ERRNO(res);
759 		else
760 			res = EFAULT;
761 		rp->r_r0 = res;
762 		rp->r_pc = ct->t_lofault;
763 		goto cleanup;
764 
765 	case T_PGFLT + USER:	/* user page fault */
766 		if (faultdebug) {
767 			char *fault_str;
768 
769 			switch (rw) {
770 			case S_READ:
771 				fault_str = "read";
772 				break;
773 			case S_WRITE:
774 				fault_str = "write";
775 				break;
776 			case S_EXEC:
777 				fault_str = "exec";
778 				break;
779 			default:
780 				fault_str = "";
781 				break;
782 			}
783 			printf("user %s fault:  addr=0x%lx errcode=0x%x\n",
784 			    fault_str, (uintptr_t)addr, errcode);
785 		}
786 
787 #if defined(OPTERON_ERRATUM_100) && defined(_LP64)
788 		/*
789 		 * Workaround for AMD erratum 100
790 		 *
791 		 * A 32-bit process may receive a page fault on a non
792 		 * 32-bit address by mistake. The range of the faulting
793 		 * address will be
794 		 *
795 		 *	0xffffffff80000000 .. 0xffffffffffffffff or
796 		 *	0x0000000100000000 .. 0x000000017fffffff
797 		 *
798 		 * The fault is always due to an instruction fetch, however
799 		 * the value of r_pc should be correct (in 32 bit range),
800 		 * so we ignore the page fault on the bogus address.
801 		 */
802 		if (p->p_model == DATAMODEL_ILP32 &&
803 		    (0xffffffff80000000 <= (uintptr_t)addr ||
804 		    (0x100000000 <= (uintptr_t)addr &&
805 		    (uintptr_t)addr <= 0x17fffffff))) {
806 			if (!opteron_erratum_100)
807 				panic("unexpected erratum #100");
808 			if (rp->r_pc <= 0xffffffff)
809 				goto out;
810 		}
811 #endif /* OPTERON_ERRATUM_100 && _LP64 */
812 
813 		ASSERT(!(curthread->t_flag & T_WATCHPT));
814 		watchpage = (pr_watch_active(p) && pr_is_watchpage(addr, rw));
815 #ifdef __i386
816 		/*
817 		 * In 32-bit mode, the lcall (system call) instruction fetches
818 		 * one word from the stack, at the stack pointer, because of the
819 		 * way the call gate is constructed.  This is a bogus
820 		 * read and should not be counted as a read watchpoint.
821 		 * We work around the problem here by testing to see if
822 		 * this situation applies and, if so, simply jumping to
823 		 * the code in locore.s that fields the system call trap.
824 		 * The registers on the stack are already set up properly
825 		 * due to the match between the call gate sequence and the
826 		 * trap gate sequence.  We just have to adjust the pc.
827 		 */
828 		if (watchpage && addr == (caddr_t)rp->r_sp &&
829 		    rw == S_READ && instr_is_lcall_syscall((caddr_t)rp->r_pc)) {
830 			extern void watch_syscall(void);
831 
832 			rp->r_pc += LCALLSIZE;
833 			watch_syscall();	/* never returns */
834 			/* NOTREACHED */
835 		}
836 #endif /* __i386 */
837 		vaddr = addr;
838 		if (!watchpage || (sz = instr_size(rp, &vaddr, rw)) <= 0)
839 			fault_type = (errcode & PF_ERR_PROT)? F_PROT: F_INVAL;
840 		else if ((watchcode = pr_is_watchpoint(&vaddr, &ta,
841 		    sz, NULL, rw)) != 0) {
842 			if (ta) {
843 				do_watch_step(vaddr, sz, rw,
844 				    watchcode, rp->r_pc);
845 				fault_type = F_INVAL;
846 			} else {
847 				bzero(&siginfo, sizeof (siginfo));
848 				siginfo.si_signo = SIGTRAP;
849 				siginfo.si_code = watchcode;
850 				siginfo.si_addr = vaddr;
851 				siginfo.si_trapafter = 0;
852 				siginfo.si_pc = (caddr_t)rp->r_pc;
853 				fault = FLTWATCH;
854 				break;
855 			}
856 		} else {
857 			/* XXX pr_watch_emul() never succeeds (for now) */
858 			if (rw != S_EXEC && pr_watch_emul(rp, vaddr, rw))
859 				goto out;
860 			do_watch_step(vaddr, sz, rw, 0, 0);
861 			fault_type = F_INVAL;
862 		}
863 
864 		res = pagefault(addr, fault_type, rw, 0);
865 
866 		/*
867 		 * If pagefault() succeeded, ok.
868 		 * Otherwise attempt to grow the stack.
869 		 */
870 		if (res == 0 ||
871 		    (res == FC_NOMAP &&
872 		    addr < p->p_usrstack &&
873 		    grow(addr))) {
874 			lwp->lwp_lastfault = FLTPAGE;
875 			lwp->lwp_lastfaddr = addr;
876 			if (prismember(&p->p_fltmask, FLTPAGE)) {
877 				bzero(&siginfo, sizeof (siginfo));
878 				siginfo.si_addr = addr;
879 				(void) stop_on_fault(FLTPAGE, &siginfo);
880 			}
881 			goto out;
882 		} else if (res == FC_PROT && addr < p->p_usrstack &&
883 		    (mmu.pt_nx != 0 && (errcode & PF_ERR_EXEC))) {
884 			report_stack_exec(p, addr);
885 		}
886 
887 #ifdef OPTERON_ERRATUM_91
888 		/*
889 		 * Workaround for Opteron Erratum 91. Prefetches may generate a
890 		 * page fault (they're not supposed to do that!). If this
891 		 * occurs we simply return back to the instruction.
892 		 *
893 		 * We rely on copyin to properly fault in the page with r_pc.
894 		 */
895 		if (opteron_erratum_91 &&
896 		    addr != (caddr_t)rp->r_pc &&
897 		    instr_is_prefetch((caddr_t)rp->r_pc)) {
898 #ifdef DEBUG
899 			cmn_err(CE_WARN, "Opteron erratum 91 occurred: "
900 			    "prefetch at %p in pid %d generated a trap!",
901 			    (void *)rp->r_pc, p->p_pid);
902 #endif /* DEBUG */
903 			goto out;
904 		}
905 #endif /* OPTERON_ERRATUM_91 */
906 
907 		if (tudebug)
908 			showregs(type, rp, addr);
909 		/*
910 		 * In the case where both pagefault and grow fail,
911 		 * set the code to the value provided by pagefault.
912 		 * We map all errors returned from pagefault() to SIGSEGV.
913 		 */
914 		bzero(&siginfo, sizeof (siginfo));
915 		siginfo.si_addr = addr;
916 		switch (FC_CODE(res)) {
917 		case FC_HWERR:
918 		case FC_NOSUPPORT:
919 			siginfo.si_signo = SIGBUS;
920 			siginfo.si_code = BUS_ADRERR;
921 			fault = FLTACCESS;
922 			break;
923 		case FC_ALIGN:
924 			siginfo.si_signo = SIGBUS;
925 			siginfo.si_code = BUS_ADRALN;
926 			fault = FLTACCESS;
927 			break;
928 		case FC_OBJERR:
929 			if ((siginfo.si_errno = FC_ERRNO(res)) != EINTR) {
930 				siginfo.si_signo = SIGBUS;
931 				siginfo.si_code = BUS_OBJERR;
932 				fault = FLTACCESS;
933 			}
934 			break;
935 		default:	/* FC_NOMAP or FC_PROT */
936 			siginfo.si_signo = SIGSEGV;
937 			siginfo.si_code =
938 			    (res == FC_NOMAP)? SEGV_MAPERR : SEGV_ACCERR;
939 			fault = FLTBOUNDS;
940 			break;
941 		}
942 		break;
943 
944 	case T_ILLINST + USER:	/* invalid opcode fault */
945 		/*
946 		 * If the syscall instruction is disabled due to LDT usage, a
947 		 * user program that attempts to execute it will trigger a #ud
948 		 * trap. Check for that case here. If this occurs on a CPU which
949 		 * doesn't even support syscall, the result of all of this will
950 		 * be to emulate that particular instruction.
951 		 */
952 		if (p->p_ldt != NULL &&
953 		    ldt_rewrite_syscall(rp, p, X86FSET_ASYSC))
954 			goto out;
955 
956 #ifdef __amd64
957 		/*
958 		 * Emulate the LAHF and SAHF instructions if needed.
959 		 * See the instr_is_lsahf function for details.
960 		 */
961 		if (p->p_model == DATAMODEL_LP64 &&
962 		    instr_is_lsahf((caddr_t)rp->r_pc, &instr)) {
963 			emulate_lsahf(rp, instr);
964 			goto out;
965 		}
966 #endif
967 
968 		/*FALLTHROUGH*/
969 
970 		if (tudebug)
971 			showregs(type, rp, (caddr_t)0);
972 		siginfo.si_signo = SIGILL;
973 		siginfo.si_code  = ILL_ILLOPC;
974 		siginfo.si_addr  = (caddr_t)rp->r_pc;
975 		fault = FLTILL;
976 		break;
977 
978 	case T_ZERODIV + USER:		/* integer divide by zero */
979 		if (tudebug && tudebugfpe)
980 			showregs(type, rp, (caddr_t)0);
981 		siginfo.si_signo = SIGFPE;
982 		siginfo.si_code  = FPE_INTDIV;
983 		siginfo.si_addr  = (caddr_t)rp->r_pc;
984 		fault = FLTIZDIV;
985 		break;
986 
987 	case T_OVFLW + USER:	/* integer overflow */
988 		if (tudebug && tudebugfpe)
989 			showregs(type, rp, (caddr_t)0);
990 		siginfo.si_signo = SIGFPE;
991 		siginfo.si_code  = FPE_INTOVF;
992 		siginfo.si_addr  = (caddr_t)rp->r_pc;
993 		fault = FLTIOVF;
994 		break;
995 
996 	/*
997 	 * When using an eager FPU on x86, the #NM trap is no longer meaningful.
998 	 * Userland should not be able to trigger it. Anything that does
999 	 * represents a fatal error in the kernel and likely in the register
1000 	 * state of the system. User FPU state should always be valid.
1001 	 */
1002 	case T_NOEXTFLT + USER:	/* math coprocessor not available */
1003 	case T_NOEXTFLT:
1004 		(void) die(type, rp, addr, cpuid);
1005 		break;
1006 
1007 	/*
1008 	 * Kernel threads leveraging floating point need to mask the exceptions
1009 	 * or ensure that they cannot happen. There is no recovery from this.
1010 	 */
1011 	case T_EXTERRFLT:	/* x87 floating point exception pending */
1012 		sti(); /* T_EXTERRFLT comes in via cmninttrap */
1013 		(void) die(type, rp, addr, cpuid);
1014 		break;
1015 
1016 	case T_EXTERRFLT + USER: /* x87 floating point exception pending */
1017 		if (tudebug && tudebugfpe)
1018 			showregs(type, rp, addr);
1019 		if (sicode = fpexterrflt(rp)) {
1020 			siginfo.si_signo = SIGFPE;
1021 			siginfo.si_code  = sicode;
1022 			siginfo.si_addr  = (caddr_t)rp->r_pc;
1023 			fault = FLTFPE;
1024 		}
1025 		break;
1026 
1027 	case T_SIMDFPE + USER:		/* SSE and SSE2 exceptions */
1028 		if (tudebug && tudebugsse)
1029 			showregs(type, rp, addr);
1030 		if (!is_x86_feature(x86_featureset, X86FSET_SSE) &&
1031 		    !is_x86_feature(x86_featureset, X86FSET_SSE2)) {
1032 			/*
1033 			 * There are rumours that some user instructions
1034 			 * on older CPUs can cause this trap to occur; in
1035 			 * which case send a SIGILL instead of a SIGFPE.
1036 			 */
1037 			siginfo.si_signo = SIGILL;
1038 			siginfo.si_code  = ILL_ILLTRP;
1039 			siginfo.si_addr  = (caddr_t)rp->r_pc;
1040 			siginfo.si_trapno = type & ~USER;
1041 			fault = FLTILL;
1042 		} else if ((sicode = fpsimderrflt(rp)) != 0) {
1043 			siginfo.si_signo = SIGFPE;
1044 			siginfo.si_code = sicode;
1045 			siginfo.si_addr = (caddr_t)rp->r_pc;
1046 			fault = FLTFPE;
1047 		}
1048 
1049 		sti();	/* The SIMD exception comes in via cmninttrap */
1050 		break;
1051 
1052 	case T_BPTFLT:	/* breakpoint trap */
1053 		/*
1054 		 * Kernel breakpoint traps should only happen when kmdb is
1055 		 * active, and even then, it'll have interposed on the IDT, so
1056 		 * control won't get here.  If it does, we've hit a breakpoint
1057 		 * without the debugger, which is very strange, and very
1058 		 * fatal.
1059 		 */
1060 		if (tudebug && tudebugbpt)
1061 			showregs(type, rp, (caddr_t)0);
1062 
1063 		(void) die(type, rp, addr, cpuid);
1064 		break;
1065 
1066 	case T_SGLSTP: /* single step/hw breakpoint exception */
1067 
1068 #if !defined(__xpv)
1069 		/*
1070 		 * We'd never normally get here, as kmdb handles its own single
1071 		 * step traps.  There is one nasty exception though, as
1072 		 * described in more detail in sys_sysenter().  Note that
1073 		 * checking for all four locations covers both the KPTI and the
1074 		 * non-KPTI cases correctly: the former will never be found at
1075 		 * (brand_)sys_sysenter, and vice versa.
1076 		 */
1077 		if (lwp != NULL && (lwp->lwp_pcb.pcb_drstat & DR_SINGLESTEP)) {
1078 			if (rp->r_pc == (greg_t)brand_sys_sysenter ||
1079 			    rp->r_pc == (greg_t)sys_sysenter ||
1080 			    rp->r_pc == (greg_t)tr_brand_sys_sysenter ||
1081 			    rp->r_pc == (greg_t)tr_sys_sysenter) {
1082 
1083 				rp->r_pc += 0x3; /* sizeof (swapgs) */
1084 
1085 				rp->r_ps &= ~PS_T; /* turn off trace */
1086 				lwp->lwp_pcb.pcb_flags |= DEBUG_PENDING;
1087 				ct->t_post_sys = 1;
1088 				aston(curthread);
1089 				goto cleanup;
1090 			} else {
1091 				if (tudebug && tudebugbpt)
1092 					showregs(type, rp, (caddr_t)0);
1093 			}
1094 		}
1095 #endif /* !__xpv */
1096 
1097 		if (boothowto & RB_DEBUG)
1098 			debug_enter((char *)NULL);
1099 		else
1100 			(void) die(type, rp, addr, cpuid);
1101 		break;
1102 
1103 	case T_NMIFLT:	/* NMI interrupt */
1104 		printf("Unexpected NMI in system mode\n");
1105 		goto cleanup;
1106 
1107 	case T_NMIFLT + USER:	/* NMI interrupt */
1108 		printf("Unexpected NMI in user mode\n");
1109 		break;
1110 
1111 	case T_GPFLT:	/* general protection violation */
1112 		/*
1113 		 * Any #GP that occurs during an on_trap .. no_trap bracket
1114 		 * with OT_DATA_ACCESS or OT_SEGMENT_ACCESS protection,
1115 		 * or in a on_fault .. no_fault bracket, is forgiven
1116 		 * and we trampoline.  This protection is given regardless
1117 		 * of whether we are 32/64 bit etc - if a distinction is
1118 		 * required then define new on_trap protection types.
1119 		 *
1120 		 * On amd64, we can get a #gp from referencing addresses
1121 		 * in the virtual address hole e.g. from a copyin or in
1122 		 * update_sregs while updating user segment registers.
1123 		 *
1124 		 * On the 32-bit hypervisor we could also generate one in
1125 		 * mfn_to_pfn by reaching around or into where the hypervisor
1126 		 * lives which is protected by segmentation.
1127 		 */
1128 
1129 		/*
1130 		 * If we're under on_trap() protection (see <sys/ontrap.h>),
1131 		 * set ot_trap and trampoline back to the on_trap() call site
1132 		 * for OT_DATA_ACCESS or OT_SEGMENT_ACCESS.
1133 		 */
1134 		if (ct->t_ontrap != NULL) {
1135 			int ttype =  ct->t_ontrap->ot_prot &
1136 			    (OT_DATA_ACCESS | OT_SEGMENT_ACCESS);
1137 
1138 			if (ttype != 0) {
1139 				ct->t_ontrap->ot_trap |= ttype;
1140 				if (tudebug)
1141 					showregs(type, rp, (caddr_t)0);
1142 				rp->r_pc = ct->t_ontrap->ot_trampoline;
1143 				goto cleanup;
1144 			}
1145 		}
1146 
1147 		/*
1148 		 * If we're under lofault protection (copyin etc.),
1149 		 * longjmp back to lofault with an EFAULT.
1150 		 */
1151 		if (ct->t_lofault) {
1152 			/*
1153 			 * Fault is not resolvable, so just return to lofault
1154 			 */
1155 			if (lodebug) {
1156 				showregs(type, rp, addr);
1157 				traceregs(rp);
1158 			}
1159 			rp->r_r0 = EFAULT;
1160 			rp->r_pc = ct->t_lofault;
1161 			goto cleanup;
1162 		}
1163 
1164 		/*
1165 		 * We fall through to the next case, which repeats
1166 		 * the OT_SEGMENT_ACCESS check which we've already
1167 		 * done, so we'll always fall through to the
1168 		 * T_STKFLT case.
1169 		 */
1170 		/*FALLTHROUGH*/
1171 	case T_SEGFLT:	/* segment not present fault */
1172 		/*
1173 		 * One example of this is #NP in update_sregs while
1174 		 * attempting to update a user segment register
1175 		 * that points to a descriptor that is marked not
1176 		 * present.
1177 		 */
1178 		if (ct->t_ontrap != NULL &&
1179 		    ct->t_ontrap->ot_prot & OT_SEGMENT_ACCESS) {
1180 			ct->t_ontrap->ot_trap |= OT_SEGMENT_ACCESS;
1181 			if (tudebug)
1182 				showregs(type, rp, (caddr_t)0);
1183 			rp->r_pc = ct->t_ontrap->ot_trampoline;
1184 			goto cleanup;
1185 		}
1186 		/*FALLTHROUGH*/
1187 	case T_STKFLT:	/* stack fault */
1188 	case T_TSSFLT:	/* invalid TSS fault */
1189 		if (tudebug)
1190 			showregs(type, rp, (caddr_t)0);
1191 		if (kern_gpfault(rp))
1192 			(void) die(type, rp, addr, cpuid);
1193 		goto cleanup;
1194 
1195 	/*
1196 	 * ONLY 32-bit PROCESSES can USE a PRIVATE LDT! 64-bit apps
1197 	 * should have no need for them, so we put a stop to it here.
1198 	 *
1199 	 * So: not-present fault is ONLY valid for 32-bit processes with
1200 	 * a private LDT trying to do a system call. Emulate it.
1201 	 *
1202 	 * #gp fault is ONLY valid for 32-bit processes also, which DO NOT
1203 	 * have a private LDT, and are trying to do a system call. Emulate it.
1204 	 */
1205 
1206 	case T_SEGFLT + USER:	/* segment not present fault */
1207 	case T_GPFLT + USER:	/* general protection violation */
1208 #ifdef _SYSCALL32_IMPL
1209 		if (p->p_model != DATAMODEL_NATIVE) {
1210 #endif /* _SYSCALL32_IMPL */
1211 		if (instr_is_lcall_syscall((caddr_t)rp->r_pc)) {
1212 			if (type == T_SEGFLT + USER)
1213 				ASSERT(p->p_ldt != NULL);
1214 
1215 			if ((p->p_ldt == NULL && type == T_GPFLT + USER) ||
1216 			    type == T_SEGFLT + USER) {
1217 
1218 			/*
1219 			 * The user attempted a system call via the obsolete
1220 			 * call gate mechanism. Because the process doesn't have
1221 			 * an LDT (i.e. the ldtr contains 0), a #gp results.
1222 			 * Emulate the syscall here, just as we do above for a
1223 			 * #np trap.
1224 			 */
1225 
1226 			/*
1227 			 * Since this is a not-present trap, rp->r_pc points to
1228 			 * the trapping lcall instruction. We need to bump it
1229 			 * to the next insn so the app can continue on.
1230 			 */
1231 			rp->r_pc += LCALLSIZE;
1232 			lwp->lwp_regs = rp;
1233 
1234 			/*
1235 			 * Normally the microstate of the LWP is forced back to
1236 			 * LMS_USER by the syscall handlers. Emulate that
1237 			 * behavior here.
1238 			 */
1239 			mstate = LMS_USER;
1240 
1241 			dosyscall();
1242 			goto out;
1243 			}
1244 		}
1245 #ifdef _SYSCALL32_IMPL
1246 		}
1247 #endif /* _SYSCALL32_IMPL */
1248 		/*
1249 		 * If the current process is using a private LDT and the
1250 		 * trapping instruction is sysenter, the sysenter instruction
1251 		 * has been disabled on the CPU because it destroys segment
1252 		 * registers. If this is the case, rewrite the instruction to
1253 		 * be a safe system call and retry it. If this occurs on a CPU
1254 		 * which doesn't even support sysenter, the result of all of
1255 		 * this will be to emulate that particular instruction.
1256 		 */
1257 		if (p->p_ldt != NULL &&
1258 		    ldt_rewrite_syscall(rp, p, X86FSET_SEP))
1259 			goto out;
1260 
1261 		/*FALLTHROUGH*/
1262 
1263 	case T_BOUNDFLT + USER:	/* bound fault */
1264 	case T_STKFLT + USER:	/* stack fault */
1265 	case T_TSSFLT + USER:	/* invalid TSS fault */
1266 		if (tudebug)
1267 			showregs(type, rp, (caddr_t)0);
1268 		siginfo.si_signo = SIGSEGV;
1269 		siginfo.si_code  = SEGV_MAPERR;
1270 		siginfo.si_addr  = (caddr_t)rp->r_pc;
1271 		fault = FLTBOUNDS;
1272 		break;
1273 
1274 	case T_ALIGNMENT + USER:	/* user alignment error (486) */
1275 		if (tudebug)
1276 			showregs(type, rp, (caddr_t)0);
1277 		bzero(&siginfo, sizeof (siginfo));
1278 		siginfo.si_signo = SIGBUS;
1279 		siginfo.si_code = BUS_ADRALN;
1280 		siginfo.si_addr = (caddr_t)rp->r_pc;
1281 		fault = FLTACCESS;
1282 		break;
1283 
1284 	case T_SGLSTP + USER: /* single step/hw breakpoint exception */
1285 		if (tudebug && tudebugbpt)
1286 			showregs(type, rp, (caddr_t)0);
1287 
1288 		/* Was it single-stepping? */
1289 		if (lwp->lwp_pcb.pcb_drstat & DR_SINGLESTEP) {
1290 			pcb_t *pcb = &lwp->lwp_pcb;
1291 
1292 			rp->r_ps &= ~PS_T;
1293 			/*
1294 			 * If both NORMAL_STEP and WATCH_STEP are in effect,
1295 			 * give precedence to WATCH_STEP.  If neither is set,
1296 			 * user must have set the PS_T bit in %efl; treat this
1297 			 * as NORMAL_STEP.
1298 			 */
1299 			if ((fault = undo_watch_step(&siginfo)) == 0 &&
1300 			    ((pcb->pcb_flags & NORMAL_STEP) ||
1301 			    !(pcb->pcb_flags & WATCH_STEP))) {
1302 				siginfo.si_signo = SIGTRAP;
1303 				siginfo.si_code = TRAP_TRACE;
1304 				siginfo.si_addr = (caddr_t)rp->r_pc;
1305 				fault = FLTTRACE;
1306 			}
1307 			pcb->pcb_flags &= ~(NORMAL_STEP|WATCH_STEP);
1308 		}
1309 		break;
1310 
1311 	case T_BPTFLT + USER:	/* breakpoint trap */
1312 		if (tudebug && tudebugbpt)
1313 			showregs(type, rp, (caddr_t)0);
1314 		/*
1315 		 * int 3 (the breakpoint instruction) leaves the pc referring
1316 		 * to the address one byte after the breakpointed address.
1317 		 * If the P_PR_BPTADJ flag has been set via /proc, We adjust
1318 		 * it back so it refers to the breakpointed address.
1319 		 */
1320 		if (p->p_proc_flag & P_PR_BPTADJ)
1321 			rp->r_pc--;
1322 		siginfo.si_signo = SIGTRAP;
1323 		siginfo.si_code  = TRAP_BRKPT;
1324 		siginfo.si_addr  = (caddr_t)rp->r_pc;
1325 		fault = FLTBPT;
1326 		break;
1327 
1328 	case T_AST:
1329 		/*
1330 		 * This occurs only after the cs register has been made to
1331 		 * look like a kernel selector, either through debugging or
1332 		 * possibly by functions like setcontext().  The thread is
1333 		 * about to cause a general protection fault at common_iret()
1334 		 * in locore.  We let that happen immediately instead of
1335 		 * doing the T_AST processing.
1336 		 */
1337 		goto cleanup;
1338 
1339 	case T_AST + USER:	/* profiling, resched, h/w error pseudo trap */
1340 		if (lwp->lwp_pcb.pcb_flags & ASYNC_HWERR) {
1341 			proc_t *p = ttoproc(curthread);
1342 			extern void print_msg_hwerr(ctid_t ct_id, proc_t *p);
1343 
1344 			lwp->lwp_pcb.pcb_flags &= ~ASYNC_HWERR;
1345 			print_msg_hwerr(p->p_ct_process->conp_contract.ct_id,
1346 			    p);
1347 			contract_process_hwerr(p->p_ct_process, p);
1348 			siginfo.si_signo = SIGKILL;
1349 			siginfo.si_code = SI_NOINFO;
1350 		} else if (lwp->lwp_pcb.pcb_flags & CPC_OVERFLOW) {
1351 			lwp->lwp_pcb.pcb_flags &= ~CPC_OVERFLOW;
1352 			if (kcpc_overflow_ast()) {
1353 				/*
1354 				 * Signal performance counter overflow
1355 				 */
1356 				if (tudebug)
1357 					showregs(type, rp, (caddr_t)0);
1358 				bzero(&siginfo, sizeof (siginfo));
1359 				siginfo.si_signo = SIGEMT;
1360 				siginfo.si_code = EMT_CPCOVF;
1361 				siginfo.si_addr = (caddr_t)rp->r_pc;
1362 				fault = FLTCPCOVF;
1363 			}
1364 		}
1365 
1366 		break;
1367 	}
1368 
1369 	/*
1370 	 * We can't get here from a system trap
1371 	 */
1372 	ASSERT(type & USER);
1373 
1374 	if (fault) {
1375 		/* We took a fault so abort single step. */
1376 		lwp->lwp_pcb.pcb_flags &= ~(NORMAL_STEP|WATCH_STEP);
1377 		/*
1378 		 * Remember the fault and fault adddress
1379 		 * for real-time (SIGPROF) profiling.
1380 		 */
1381 		lwp->lwp_lastfault = fault;
1382 		lwp->lwp_lastfaddr = siginfo.si_addr;
1383 
1384 		DTRACE_PROC2(fault, int, fault, ksiginfo_t *, &siginfo);
1385 
1386 		/*
1387 		 * If a debugger has declared this fault to be an
1388 		 * event of interest, stop the lwp.  Otherwise just
1389 		 * deliver the associated signal.
1390 		 */
1391 		if (siginfo.si_signo != SIGKILL &&
1392 		    prismember(&p->p_fltmask, fault) &&
1393 		    stop_on_fault(fault, &siginfo) == 0)
1394 			siginfo.si_signo = 0;
1395 	}
1396 
1397 	if (siginfo.si_signo)
1398 		trapsig(&siginfo, (fault != FLTFPE && fault != FLTCPCOVF));
1399 
1400 	if (lwp->lwp_oweupc)
1401 		profil_tick(rp->r_pc);
1402 
1403 	if (ct->t_astflag | ct->t_sig_check) {
1404 		/*
1405 		 * Turn off the AST flag before checking all the conditions that
1406 		 * may have caused an AST.  This flag is on whenever a signal or
1407 		 * unusual condition should be handled after the next trap or
1408 		 * syscall.
1409 		 */
1410 		astoff(ct);
1411 		/*
1412 		 * If a single-step trap occurred on a syscall (see above)
1413 		 * recognize it now.  Do this before checking for signals
1414 		 * because deferred_singlestep_trap() may generate a SIGTRAP to
1415 		 * the LWP or may otherwise mark the LWP to call issig(FORREAL).
1416 		 */
1417 		if (lwp->lwp_pcb.pcb_flags & DEBUG_PENDING)
1418 			deferred_singlestep_trap((caddr_t)rp->r_pc);
1419 
1420 		ct->t_sig_check = 0;
1421 
1422 		/*
1423 		 * As in other code paths that check against TP_CHANGEBIND,
1424 		 * we perform the check first without p_lock held -- only
1425 		 * acquiring p_lock in the unlikely event that it is indeed
1426 		 * set.  This is safe because we are doing this after the
1427 		 * astoff(); if we are racing another thread setting
1428 		 * TP_CHANGEBIND on us, we will pick it up on a subsequent
1429 		 * lap through.
1430 		 */
1431 		if (curthread->t_proc_flag & TP_CHANGEBIND) {
1432 			mutex_enter(&p->p_lock);
1433 			if (curthread->t_proc_flag & TP_CHANGEBIND) {
1434 				timer_lwpbind();
1435 				curthread->t_proc_flag &= ~TP_CHANGEBIND;
1436 			}
1437 			mutex_exit(&p->p_lock);
1438 		}
1439 
1440 		/*
1441 		 * for kaio requests that are on the per-process poll queue,
1442 		 * aiop->aio_pollq, they're AIO_POLL bit is set, the kernel
1443 		 * should copyout their result_t to user memory. by copying
1444 		 * out the result_t, the user can poll on memory waiting
1445 		 * for the kaio request to complete.
1446 		 */
1447 		if (p->p_aio)
1448 			aio_cleanup(0);
1449 		/*
1450 		 * If this LWP was asked to hold, call holdlwp(), which will
1451 		 * stop.  holdlwps() sets this up and calls pokelwps() which
1452 		 * sets the AST flag.
1453 		 *
1454 		 * Also check TP_EXITLWP, since this is used by fresh new LWPs
1455 		 * through lwp_rtt().  That flag is set if the lwp_create(2)
1456 		 * syscall failed after creating the LWP.
1457 		 */
1458 		if (ISHOLD(p))
1459 			holdlwp();
1460 
1461 		/*
1462 		 * All code that sets signals and makes ISSIG evaluate true must
1463 		 * set t_astflag afterwards.
1464 		 */
1465 		if (ISSIG_PENDING(ct, lwp, p)) {
1466 			if (issig(FORREAL))
1467 				psig();
1468 			ct->t_sig_check = 1;
1469 		}
1470 
1471 		if (ct->t_rprof != NULL) {
1472 			realsigprof(0, 0, 0);
1473 			ct->t_sig_check = 1;
1474 		}
1475 
1476 		/*
1477 		 * /proc can't enable/disable the trace bit itself
1478 		 * because that could race with the call gate used by
1479 		 * system calls via "lcall". If that happened, an
1480 		 * invalid EFLAGS would result. prstep()/prnostep()
1481 		 * therefore schedule an AST for the purpose.
1482 		 */
1483 		if (lwp->lwp_pcb.pcb_flags & REQUEST_STEP) {
1484 			lwp->lwp_pcb.pcb_flags &= ~REQUEST_STEP;
1485 			rp->r_ps |= PS_T;
1486 		}
1487 		if (lwp->lwp_pcb.pcb_flags & REQUEST_NOSTEP) {
1488 			lwp->lwp_pcb.pcb_flags &= ~REQUEST_NOSTEP;
1489 			rp->r_ps &= ~PS_T;
1490 		}
1491 	}
1492 
1493 out:	/* We can't get here from a system trap */
1494 	ASSERT(type & USER);
1495 
1496 	if (ISHOLD(p))
1497 		holdlwp();
1498 
1499 	/*
1500 	 * Set state to LWP_USER here so preempt won't give us a kernel
1501 	 * priority if it occurs after this point.  Call CL_TRAPRET() to
1502 	 * restore the user-level priority.
1503 	 *
1504 	 * It is important that no locks (other than spinlocks) be entered
1505 	 * after this point before returning to user mode (unless lwp_state
1506 	 * is set back to LWP_SYS).
1507 	 */
1508 	lwp->lwp_state = LWP_USER;
1509 
1510 	if (ct->t_trapret) {
1511 		ct->t_trapret = 0;
1512 		thread_lock(ct);
1513 		CL_TRAPRET(ct);
1514 		thread_unlock(ct);
1515 	}
1516 	if (CPU->cpu_runrun || curthread->t_schedflag & TS_ANYWAITQ)
1517 		preempt();
1518 	prunstop();
1519 	(void) new_mstate(ct, mstate);
1520 
1521 	/* Kernel probe */
1522 	TNF_PROBE_1(thread_state, "thread", /* CSTYLED */,
1523 	    tnf_microstate, state, LMS_USER);
1524 
1525 	return;
1526 
1527 cleanup:	/* system traps end up here */
1528 	ASSERT(!(type & USER));
1529 }
1530 
1531 /*
1532  * Patch non-zero to disable preemption of threads in the kernel.
1533  */
1534 int IGNORE_KERNEL_PREEMPTION = 0;	/* XXX - delete this someday */
1535 
1536 struct kpreempt_cnts {		/* kernel preemption statistics */
1537 	int	kpc_idle;	/* executing idle thread */
1538 	int	kpc_intr;	/* executing interrupt thread */
1539 	int	kpc_clock;	/* executing clock thread */
1540 	int	kpc_blocked;	/* thread has blocked preemption (t_preempt) */
1541 	int	kpc_notonproc;	/* thread is surrendering processor */
1542 	int	kpc_inswtch;	/* thread has ratified scheduling decision */
1543 	int	kpc_prilevel;	/* processor interrupt level is too high */
1544 	int	kpc_apreempt;	/* asynchronous preemption */
1545 	int	kpc_spreempt;	/* synchronous preemption */
1546 } kpreempt_cnts;
1547 
1548 /*
1549  * kernel preemption: forced rescheduling, preempt the running kernel thread.
1550  *	the argument is old PIL for an interrupt,
1551  *	or the distingished value KPREEMPT_SYNC.
1552  */
1553 void
1554 kpreempt(int asyncspl)
1555 {
1556 	kthread_t *ct = curthread;
1557 
1558 	if (IGNORE_KERNEL_PREEMPTION) {
1559 		aston(CPU->cpu_dispthread);
1560 		return;
1561 	}
1562 
1563 	/*
1564 	 * Check that conditions are right for kernel preemption
1565 	 */
1566 	do {
1567 		if (ct->t_preempt) {
1568 			/*
1569 			 * either a privileged thread (idle, panic, interrupt)
1570 			 * or will check when t_preempt is lowered
1571 			 * We need to specifically handle the case where
1572 			 * the thread is in the middle of swtch (resume has
1573 			 * been called) and has its t_preempt set
1574 			 * [idle thread and a thread which is in kpreempt
1575 			 * already] and then a high priority thread is
1576 			 * available in the local dispatch queue.
1577 			 * In this case the resumed thread needs to take a
1578 			 * trap so that it can call kpreempt. We achieve
1579 			 * this by using siron().
1580 			 * How do we detect this condition:
1581 			 * idle thread is running and is in the midst of
1582 			 * resume: curthread->t_pri == -1 && CPU->dispthread
1583 			 * != CPU->thread
1584 			 * Need to ensure that this happens only at high pil
1585 			 * resume is called at high pil
1586 			 * Only in resume_from_idle is the pil changed.
1587 			 */
1588 			if (ct->t_pri < 0) {
1589 				kpreempt_cnts.kpc_idle++;
1590 				if (CPU->cpu_dispthread != CPU->cpu_thread)
1591 					siron();
1592 			} else if (ct->t_flag & T_INTR_THREAD) {
1593 				kpreempt_cnts.kpc_intr++;
1594 				if (ct->t_pil == CLOCK_LEVEL)
1595 					kpreempt_cnts.kpc_clock++;
1596 			} else {
1597 				kpreempt_cnts.kpc_blocked++;
1598 				if (CPU->cpu_dispthread != CPU->cpu_thread)
1599 					siron();
1600 			}
1601 			aston(CPU->cpu_dispthread);
1602 			return;
1603 		}
1604 		if (ct->t_state != TS_ONPROC ||
1605 		    ct->t_disp_queue != CPU->cpu_disp) {
1606 			/* this thread will be calling swtch() shortly */
1607 			kpreempt_cnts.kpc_notonproc++;
1608 			if (CPU->cpu_thread != CPU->cpu_dispthread) {
1609 				/* already in swtch(), force another */
1610 				kpreempt_cnts.kpc_inswtch++;
1611 				siron();
1612 			}
1613 			return;
1614 		}
1615 		if (getpil() >= DISP_LEVEL) {
1616 			/*
1617 			 * We can't preempt this thread if it is at
1618 			 * a PIL >= DISP_LEVEL since it may be holding
1619 			 * a spin lock (like sched_lock).
1620 			 */
1621 			siron();	/* check back later */
1622 			kpreempt_cnts.kpc_prilevel++;
1623 			return;
1624 		}
1625 		if (!interrupts_enabled()) {
1626 			/*
1627 			 * Can't preempt while running with ints disabled
1628 			 */
1629 			kpreempt_cnts.kpc_prilevel++;
1630 			return;
1631 		}
1632 		if (asyncspl != KPREEMPT_SYNC)
1633 			kpreempt_cnts.kpc_apreempt++;
1634 		else
1635 			kpreempt_cnts.kpc_spreempt++;
1636 
1637 		ct->t_preempt++;
1638 		preempt();
1639 		ct->t_preempt--;
1640 	} while (CPU->cpu_kprunrun);
1641 }
1642 
1643 /*
1644  * Print out debugging info.
1645  */
1646 static void
1647 showregs(uint_t type, struct regs *rp, caddr_t addr)
1648 {
1649 	int s;
1650 
1651 	s = spl7();
1652 	type &= ~USER;
1653 	if (PTOU(curproc)->u_comm[0])
1654 		printf("%s: ", PTOU(curproc)->u_comm);
1655 	if (type < TRAP_TYPES)
1656 		printf("#%s %s\n", trap_type_mnemonic[type], trap_type[type]);
1657 	else
1658 		switch (type) {
1659 		case T_SYSCALL:
1660 			printf("Syscall Trap:\n");
1661 			break;
1662 		case T_AST:
1663 			printf("AST\n");
1664 			break;
1665 		default:
1666 			printf("Bad Trap = %d\n", type);
1667 			break;
1668 		}
1669 	if (type == T_PGFLT) {
1670 		printf("Bad %s fault at addr=0x%lx\n",
1671 		    USERMODE(rp->r_cs) ? "user": "kernel", (uintptr_t)addr);
1672 	} else if (addr) {
1673 		printf("addr=0x%lx\n", (uintptr_t)addr);
1674 	}
1675 
1676 	printf("pid=%d, pc=0x%lx, sp=0x%lx, eflags=0x%lx\n",
1677 	    (ttoproc(curthread) && ttoproc(curthread)->p_pidp) ?
1678 	    ttoproc(curthread)->p_pid : 0, rp->r_pc, rp->r_sp, rp->r_ps);
1679 
1680 #if defined(__lint)
1681 	/*
1682 	 * this clause can be deleted when lint bug 4870403 is fixed
1683 	 * (lint thinks that bit 32 is illegal in a %b format string)
1684 	 */
1685 	printf("cr0: %x  cr4: %b\n",
1686 	    (uint_t)getcr0(), (uint_t)getcr4(), FMT_CR4);
1687 #else
1688 	printf("cr0: %b  cr4: %b\n",
1689 	    (uint_t)getcr0(), FMT_CR0, (uint_t)getcr4(), FMT_CR4);
1690 #endif	/* __lint */
1691 
1692 	printf("cr2: %lx  ", getcr2());
1693 #if !defined(__xpv)
1694 	printf("cr3: %lx  ", getcr3());
1695 #if defined(__amd64)
1696 	printf("cr8: %lx\n", getcr8());
1697 #endif
1698 #endif
1699 	printf("\n");
1700 
1701 	dumpregs(rp);
1702 	splx(s);
1703 }
1704 
1705 static void
1706 dumpregs(struct regs *rp)
1707 {
1708 #if defined(__amd64)
1709 	const char fmt[] = "\t%3s: %16lx %3s: %16lx %3s: %16lx\n";
1710 
1711 	printf(fmt, "rdi", rp->r_rdi, "rsi", rp->r_rsi, "rdx", rp->r_rdx);
1712 	printf(fmt, "rcx", rp->r_rcx, " r8", rp->r_r8, " r9", rp->r_r9);
1713 	printf(fmt, "rax", rp->r_rax, "rbx", rp->r_rbx, "rbp", rp->r_rbp);
1714 	printf(fmt, "r10", rp->r_r10, "r11", rp->r_r11, "r12", rp->r_r12);
1715 	printf(fmt, "r13", rp->r_r13, "r14", rp->r_r14, "r15", rp->r_r15);
1716 
1717 	printf(fmt, "fsb", rdmsr(MSR_AMD_FSBASE), "gsb", rdmsr(MSR_AMD_GSBASE),
1718 	    " ds", rp->r_ds);
1719 	printf(fmt, " es", rp->r_es, " fs", rp->r_fs, " gs", rp->r_gs);
1720 
1721 	printf(fmt, "trp", rp->r_trapno, "err", rp->r_err, "rip", rp->r_rip);
1722 	printf(fmt, " cs", rp->r_cs, "rfl", rp->r_rfl, "rsp", rp->r_rsp);
1723 
1724 	printf("\t%3s: %16lx\n", " ss", rp->r_ss);
1725 
1726 #elif defined(__i386)
1727 	const char fmt[] = "\t%3s: %8lx %3s: %8lx %3s: %8lx %3s: %8lx\n";
1728 
1729 	printf(fmt, " gs", rp->r_gs, " fs", rp->r_fs,
1730 	    " es", rp->r_es, " ds", rp->r_ds);
1731 	printf(fmt, "edi", rp->r_edi, "esi", rp->r_esi,
1732 	    "ebp", rp->r_ebp, "esp", rp->r_esp);
1733 	printf(fmt, "ebx", rp->r_ebx, "edx", rp->r_edx,
1734 	    "ecx", rp->r_ecx, "eax", rp->r_eax);
1735 	printf(fmt, "trp", rp->r_trapno, "err", rp->r_err,
1736 	    "eip", rp->r_eip, " cs", rp->r_cs);
1737 	printf("\t%3s: %8lx %3s: %8lx %3s: %8lx\n",
1738 	    "efl", rp->r_efl, "usp", rp->r_uesp, " ss", rp->r_ss);
1739 
1740 #endif	/* __i386 */
1741 }
1742 
1743 /*
1744  * Test to see if the instruction is iret on i386 or iretq on amd64.
1745  *
1746  * On the hypervisor we can only test for nopop_sys_rtt_syscall. If true
1747  * then we are in the context of hypervisor's failsafe handler because it
1748  * tried to iret and failed due to a bad selector. See xen_failsafe_callback.
1749  */
1750 static int
1751 instr_is_iret(caddr_t pc)
1752 {
1753 
1754 #if defined(__xpv)
1755 	extern void nopop_sys_rtt_syscall(void);
1756 	return ((pc == (caddr_t)nopop_sys_rtt_syscall) ? 1 : 0);
1757 
1758 #else
1759 
1760 #if defined(__amd64)
1761 	static const uint8_t iret_insn[2] = { 0x48, 0xcf };	/* iretq */
1762 
1763 #elif defined(__i386)
1764 	static const uint8_t iret_insn[1] = { 0xcf };		/* iret */
1765 #endif	/* __i386 */
1766 	return (bcmp(pc, iret_insn, sizeof (iret_insn)) == 0);
1767 
1768 #endif	/* __xpv */
1769 }
1770 
1771 #if defined(__i386)
1772 
1773 /*
1774  * Test to see if the instruction is part of __SEGREGS_POP
1775  *
1776  * Note carefully the appallingly awful dependency between
1777  * the instruction sequence used in __SEGREGS_POP and these
1778  * instructions encoded here.
1779  */
1780 static int
1781 instr_is_segregs_pop(caddr_t pc)
1782 {
1783 	static const uint8_t movw_0_esp_gs[4] = { 0x8e, 0x6c, 0x24, 0x0 };
1784 	static const uint8_t movw_4_esp_fs[4] = { 0x8e, 0x64, 0x24, 0x4 };
1785 	static const uint8_t movw_8_esp_es[4] = { 0x8e, 0x44, 0x24, 0x8 };
1786 	static const uint8_t movw_c_esp_ds[4] = { 0x8e, 0x5c, 0x24, 0xc };
1787 
1788 	if (bcmp(pc, movw_0_esp_gs, sizeof (movw_0_esp_gs)) == 0 ||
1789 	    bcmp(pc, movw_4_esp_fs, sizeof (movw_4_esp_fs)) == 0 ||
1790 	    bcmp(pc, movw_8_esp_es, sizeof (movw_8_esp_es)) == 0 ||
1791 	    bcmp(pc, movw_c_esp_ds, sizeof (movw_c_esp_ds)) == 0)
1792 		return (1);
1793 
1794 	return (0);
1795 }
1796 
1797 #endif	/* __i386 */
1798 
1799 /*
1800  * Test to see if the instruction is part of _sys_rtt (or the KPTI trampolines
1801  * which are used by _sys_rtt).
1802  *
1803  * Again on the hypervisor if we try to IRET to user land with a bad code
1804  * or stack selector we will get vectored through xen_failsafe_callback.
1805  * In which case we assume we got here via _sys_rtt since we only allow
1806  * IRET to user land to take place in _sys_rtt.
1807  */
1808 static int
1809 instr_is_sys_rtt(caddr_t pc)
1810 {
1811 	extern void _sys_rtt(), _sys_rtt_end();
1812 
1813 #if !defined(__xpv)
1814 	extern void tr_sysc_ret_start(), tr_sysc_ret_end();
1815 	extern void tr_intr_ret_start(), tr_intr_ret_end();
1816 
1817 	if ((uintptr_t)pc >= (uintptr_t)tr_sysc_ret_start &&
1818 	    (uintptr_t)pc <= (uintptr_t)tr_sysc_ret_end)
1819 		return (1);
1820 
1821 	if ((uintptr_t)pc >= (uintptr_t)tr_intr_ret_start &&
1822 	    (uintptr_t)pc <= (uintptr_t)tr_intr_ret_end)
1823 		return (1);
1824 #endif
1825 
1826 	if ((uintptr_t)pc < (uintptr_t)_sys_rtt ||
1827 	    (uintptr_t)pc > (uintptr_t)_sys_rtt_end)
1828 		return (0);
1829 
1830 	return (1);
1831 }
1832 
1833 /*
1834  * Handle #gp faults in kernel mode.
1835  *
1836  * One legitimate way this can happen is if we attempt to update segment
1837  * registers to naughty values on the way out of the kernel.
1838  *
1839  * This can happen in a couple of ways: someone - either accidentally or
1840  * on purpose - creates (setcontext(2), lwp_create(2)) or modifies
1841  * (signal(2)) a ucontext that contains silly segment register values.
1842  * Or someone - either accidentally or on purpose - modifies the prgregset_t
1843  * of a subject process via /proc to contain silly segment register values.
1844  *
1845  * (The unfortunate part is that we can end up discovering the bad segment
1846  * register value in the middle of an 'iret' after we've popped most of the
1847  * stack.  So it becomes quite difficult to associate an accurate ucontext
1848  * with the lwp, because the act of taking the #gp trap overwrites most of
1849  * what we were going to send the lwp.)
1850  *
1851  * OTOH if it turns out that's -not- the problem, and we're -not- an lwp
1852  * trying to return to user mode and we get a #gp fault, then we need
1853  * to die() -- which will happen if we return non-zero from this routine.
1854  */
1855 static int
1856 kern_gpfault(struct regs *rp)
1857 {
1858 	kthread_t *t = curthread;
1859 	proc_t *p = ttoproc(t);
1860 	klwp_t *lwp = ttolwp(t);
1861 	struct regs tmpregs, *trp = NULL;
1862 	caddr_t pc = (caddr_t)rp->r_pc;
1863 	int v;
1864 	uint32_t auditing = AU_AUDITING();
1865 
1866 	/*
1867 	 * if we're not an lwp, or in the case of running native the
1868 	 * pc range is outside _sys_rtt, then we should immediately
1869 	 * be die()ing horribly.
1870 	 */
1871 	if (lwp == NULL || !instr_is_sys_rtt(pc))
1872 		return (1);
1873 
1874 	/*
1875 	 * So at least we're in the right part of the kernel.
1876 	 *
1877 	 * Disassemble the instruction at the faulting pc.
1878 	 * Once we know what it is, we carefully reconstruct the stack
1879 	 * based on the order in which the stack is deconstructed in
1880 	 * _sys_rtt. Ew.
1881 	 */
1882 	if (instr_is_iret(pc)) {
1883 		/*
1884 		 * We took the #gp while trying to perform the IRET.
1885 		 * This means that either %cs or %ss are bad.
1886 		 * All we know for sure is that most of the general
1887 		 * registers have been restored, including the
1888 		 * segment registers, and all we have left on the
1889 		 * topmost part of the lwp's stack are the
1890 		 * registers that the iretq was unable to consume.
1891 		 *
1892 		 * All the rest of the state was crushed by the #gp
1893 		 * which pushed -its- registers atop our old save area
1894 		 * (because we had to decrement the stack pointer, sigh) so
1895 		 * all that we can try and do is to reconstruct the
1896 		 * crushed frame from the #gp trap frame itself.
1897 		 */
1898 		trp = &tmpregs;
1899 		trp->r_ss = lwptoregs(lwp)->r_ss;
1900 		trp->r_sp = lwptoregs(lwp)->r_sp;
1901 		trp->r_ps = lwptoregs(lwp)->r_ps;
1902 		trp->r_cs = lwptoregs(lwp)->r_cs;
1903 		trp->r_pc = lwptoregs(lwp)->r_pc;
1904 		bcopy(rp, trp, offsetof(struct regs, r_pc));
1905 
1906 		/*
1907 		 * Validate simple math
1908 		 */
1909 		ASSERT(trp->r_pc == lwptoregs(lwp)->r_pc);
1910 		ASSERT(trp->r_err == rp->r_err);
1911 
1912 
1913 
1914 	}
1915 
1916 #if defined(__amd64)
1917 	if (trp == NULL && PCB_NEED_UPDATE_SEGS(&lwp->lwp_pcb)) {
1918 
1919 		/*
1920 		 * This is the common case -- we're trying to load
1921 		 * a bad segment register value in the only section
1922 		 * of kernel code that ever loads segment registers.
1923 		 *
1924 		 * We don't need to do anything at this point because
1925 		 * the pcb contains all the pending segment register
1926 		 * state, and the regs are still intact because we
1927 		 * didn't adjust the stack pointer yet.  Given the fidelity
1928 		 * of all this, we could conceivably send a signal
1929 		 * to the lwp, rather than core-ing.
1930 		 */
1931 		trp = lwptoregs(lwp);
1932 		ASSERT((caddr_t)trp == (caddr_t)rp->r_sp);
1933 	}
1934 
1935 #elif defined(__i386)
1936 
1937 	if (trp == NULL && instr_is_segregs_pop(pc))
1938 		trp = lwptoregs(lwp);
1939 
1940 #endif	/* __i386 */
1941 
1942 	if (trp == NULL)
1943 		return (1);
1944 
1945 	/*
1946 	 * If we get to here, we're reasonably confident that we've
1947 	 * correctly decoded what happened on the way out of the kernel.
1948 	 * Rewrite the lwp's registers so that we can create a core dump
1949 	 * the (at least vaguely) represents the mcontext we were
1950 	 * being asked to restore when things went so terribly wrong.
1951 	 */
1952 
1953 	/*
1954 	 * Make sure that we have a meaningful %trapno and %err.
1955 	 */
1956 	trp->r_trapno = rp->r_trapno;
1957 	trp->r_err = rp->r_err;
1958 
1959 	if ((caddr_t)trp != (caddr_t)lwptoregs(lwp))
1960 		bcopy(trp, lwptoregs(lwp), sizeof (*trp));
1961 
1962 
1963 	mutex_enter(&p->p_lock);
1964 	lwp->lwp_cursig = SIGSEGV;
1965 	mutex_exit(&p->p_lock);
1966 
1967 	/*
1968 	 * Terminate all LWPs but don't discard them.  If another lwp beat
1969 	 * us to the punch by calling exit(), evaporate now.
1970 	 */
1971 	proc_is_exiting(p);
1972 	if (exitlwps(1) != 0) {
1973 		mutex_enter(&p->p_lock);
1974 		lwp_exit();
1975 	}
1976 
1977 	if (auditing)		/* audit core dump */
1978 		audit_core_start(SIGSEGV);
1979 	v = core(SIGSEGV, B_FALSE);
1980 	if (auditing)		/* audit core dump */
1981 		audit_core_finish(v ? CLD_KILLED : CLD_DUMPED);
1982 	exit(v ? CLD_KILLED : CLD_DUMPED, SIGSEGV);
1983 	return (0);
1984 }
1985 
1986 /*
1987  * dump_tss() - Display the TSS structure
1988  */
1989 
1990 #if !defined(__xpv)
1991 #if defined(__amd64)
1992 
1993 static void
1994 dump_tss(void)
1995 {
1996 	const char tss_fmt[] = "tss.%s:\t0x%p\n";  /* Format string */
1997 	tss_t *tss = CPU->cpu_tss;
1998 
1999 	printf(tss_fmt, "tss_rsp0", (void *)tss->tss_rsp0);
2000 	printf(tss_fmt, "tss_rsp1", (void *)tss->tss_rsp1);
2001 	printf(tss_fmt, "tss_rsp2", (void *)tss->tss_rsp2);
2002 
2003 	printf(tss_fmt, "tss_ist1", (void *)tss->tss_ist1);
2004 	printf(tss_fmt, "tss_ist2", (void *)tss->tss_ist2);
2005 	printf(tss_fmt, "tss_ist3", (void *)tss->tss_ist3);
2006 	printf(tss_fmt, "tss_ist4", (void *)tss->tss_ist4);
2007 	printf(tss_fmt, "tss_ist5", (void *)tss->tss_ist5);
2008 	printf(tss_fmt, "tss_ist6", (void *)tss->tss_ist6);
2009 	printf(tss_fmt, "tss_ist7", (void *)tss->tss_ist7);
2010 }
2011 
2012 #elif defined(__i386)
2013 
2014 static void
2015 dump_tss(void)
2016 {
2017 	const char tss_fmt[] = "tss.%s:\t0x%p\n";  /* Format string */
2018 	tss_t *tss = CPU->cpu_tss;
2019 
2020 	printf(tss_fmt, "tss_link", (void *)(uintptr_t)tss->tss_link);
2021 	printf(tss_fmt, "tss_esp0", (void *)(uintptr_t)tss->tss_esp0);
2022 	printf(tss_fmt, "tss_ss0", (void *)(uintptr_t)tss->tss_ss0);
2023 	printf(tss_fmt, "tss_esp1", (void *)(uintptr_t)tss->tss_esp1);
2024 	printf(tss_fmt, "tss_ss1", (void *)(uintptr_t)tss->tss_ss1);
2025 	printf(tss_fmt, "tss_esp2", (void *)(uintptr_t)tss->tss_esp2);
2026 	printf(tss_fmt, "tss_ss2", (void *)(uintptr_t)tss->tss_ss2);
2027 	printf(tss_fmt, "tss_cr3", (void *)(uintptr_t)tss->tss_cr3);
2028 	printf(tss_fmt, "tss_eip", (void *)(uintptr_t)tss->tss_eip);
2029 	printf(tss_fmt, "tss_eflags", (void *)(uintptr_t)tss->tss_eflags);
2030 	printf(tss_fmt, "tss_eax", (void *)(uintptr_t)tss->tss_eax);
2031 	printf(tss_fmt, "tss_ebx", (void *)(uintptr_t)tss->tss_ebx);
2032 	printf(tss_fmt, "tss_ecx", (void *)(uintptr_t)tss->tss_ecx);
2033 	printf(tss_fmt, "tss_edx", (void *)(uintptr_t)tss->tss_edx);
2034 	printf(tss_fmt, "tss_esp", (void *)(uintptr_t)tss->tss_esp);
2035 }
2036 
2037 #endif	/* __amd64 */
2038 #endif	/* !__xpv */
2039 
2040 #if defined(TRAPTRACE)
2041 
2042 int ttrace_nrec = 10;		/* number of records to dump out */
2043 int ttrace_dump_nregs = 0;	/* dump out this many records with regs too */
2044 
2045 /*
2046  * Dump out the last ttrace_nrec traptrace records on each CPU
2047  */
2048 static void
2049 dump_ttrace(void)
2050 {
2051 	trap_trace_ctl_t *ttc;
2052 	trap_trace_rec_t *rec;
2053 	uintptr_t current;
2054 	int i, j, k;
2055 	int n = NCPU;
2056 #if defined(__amd64)
2057 	const char banner[] =
2058 	    "CPU          ADDRESS    TIMESTAMP TYPE  VC HANDLER          PC\n";
2059 	/* Define format for the CPU, ADDRESS, and TIMESTAMP fields */
2060 	const char fmt1[] = "%3d %016lx %12llx";
2061 	char data1[34];	/* length of string formatted by fmt1 + 1 */
2062 #elif defined(__i386)
2063 	const char banner[] =
2064 	    "CPU  ADDRESS     TIMESTAMP TYPE  VC HANDLER          PC\n";
2065 	/* Define format for the CPU, ADDRESS, and TIMESTAMP fields */
2066 	const char fmt1[] = "%3d %08lx %12llx";
2067 	char data1[26];	/* length of string formatted by fmt1 + 1 */
2068 #endif
2069 	/* Define format for the TYPE and VC fields */
2070 	const char fmt2[] = "%4s %3x";
2071 	char data2[9];	/* length of string formatted by fmt2 + 1 */
2072 	/*
2073 	 * Define format for the HANDLER field. Width is arbitrary, but should
2074 	 * be enough for common handler's names, and leave enough space for
2075 	 * the PC field, especially when we are in kmdb.
2076 	 */
2077 	const char fmt3h[] = "#%-15s";
2078 	const char fmt3p[] = "%-16p";
2079 	const char fmt3s[] = "%-16s";
2080 	char data3[17];	/* length of string formatted by fmt3* + 1 */
2081 
2082 	if (ttrace_nrec == 0)
2083 		return;
2084 
2085 	printf("\n");
2086 	printf(banner);
2087 
2088 	for (i = 0; i < n; i++) {
2089 		ttc = &trap_trace_ctl[i];
2090 		if (ttc->ttc_first == NULL)
2091 			continue;
2092 
2093 		current = ttc->ttc_next - sizeof (trap_trace_rec_t);
2094 		for (j = 0; j < ttrace_nrec; j++) {
2095 			struct sysent	*sys;
2096 			struct autovec	*vec;
2097 			extern struct av_head autovect[];
2098 			int type;
2099 			ulong_t	off;
2100 			char *sym, *stype;
2101 
2102 			if (current < ttc->ttc_first)
2103 				current =
2104 				    ttc->ttc_limit - sizeof (trap_trace_rec_t);
2105 
2106 			if (current == NULL)
2107 				continue;
2108 
2109 			rec = (trap_trace_rec_t *)current;
2110 
2111 			if (rec->ttr_stamp == 0)
2112 				break;
2113 
2114 			(void) snprintf(data1, sizeof (data1), fmt1, i,
2115 			    (uintptr_t)rec, rec->ttr_stamp);
2116 
2117 			switch (rec->ttr_marker) {
2118 			case TT_SYSCALL:
2119 			case TT_SYSENTER:
2120 			case TT_SYSC:
2121 			case TT_SYSC64:
2122 				sys = &sysent32[rec->ttr_sysnum];
2123 				switch (rec->ttr_marker) {
2124 				case TT_SYSC64:
2125 					sys = &sysent[rec->ttr_sysnum];
2126 					/* FALLTHROUGH */
2127 				case TT_SYSC:
2128 					stype = "sysc";	/* syscall */
2129 					break;
2130 				case TT_SYSCALL:
2131 					stype = "lcal";	/* lcall */
2132 					break;
2133 				case TT_SYSENTER:
2134 					stype = "syse";	/* sysenter */
2135 					break;
2136 				default:
2137 					break;
2138 				}
2139 				(void) snprintf(data2, sizeof (data2), fmt2,
2140 				    stype, rec->ttr_sysnum);
2141 				if (sys != NULL) {
2142 					sym = kobj_getsymname(
2143 					    (uintptr_t)sys->sy_callc,
2144 					    &off);
2145 					if (sym != NULL) {
2146 						(void) snprintf(data3,
2147 						    sizeof (data3), fmt3s, sym);
2148 					} else {
2149 						(void) snprintf(data3,
2150 						    sizeof (data3), fmt3p,
2151 						    sys->sy_callc);
2152 					}
2153 				} else {
2154 					(void) snprintf(data3, sizeof (data3),
2155 					    fmt3s, "unknown");
2156 				}
2157 				break;
2158 
2159 			case TT_INTERRUPT:
2160 				(void) snprintf(data2, sizeof (data2), fmt2,
2161 				    "intr", rec->ttr_vector);
2162 				if (get_intr_handler != NULL)
2163 					vec = (struct autovec *)
2164 					    (*get_intr_handler)
2165 					    (rec->ttr_cpuid, rec->ttr_vector);
2166 				else
2167 					vec =
2168 					    autovect[rec->ttr_vector].avh_link;
2169 
2170 				if (vec != NULL) {
2171 					sym = kobj_getsymname(
2172 					    (uintptr_t)vec->av_vector, &off);
2173 					if (sym != NULL) {
2174 						(void) snprintf(data3,
2175 						    sizeof (data3), fmt3s, sym);
2176 					} else {
2177 						(void) snprintf(data3,
2178 						    sizeof (data3), fmt3p,
2179 						    vec->av_vector);
2180 					}
2181 				} else {
2182 					(void) snprintf(data3, sizeof (data3),
2183 					    fmt3s, "unknown");
2184 				}
2185 				break;
2186 
2187 			case TT_TRAP:
2188 			case TT_EVENT:
2189 				type = rec->ttr_regs.r_trapno;
2190 				(void) snprintf(data2, sizeof (data2), fmt2,
2191 				    "trap", type);
2192 				if (type < TRAP_TYPES) {
2193 					(void) snprintf(data3, sizeof (data3),
2194 					    fmt3h, trap_type_mnemonic[type]);
2195 				} else {
2196 					switch (type) {
2197 					case T_AST:
2198 						(void) snprintf(data3,
2199 						    sizeof (data3), fmt3s,
2200 						    "ast");
2201 						break;
2202 					default:
2203 						(void) snprintf(data3,
2204 						    sizeof (data3), fmt3s, "");
2205 						break;
2206 					}
2207 				}
2208 				break;
2209 
2210 			default:
2211 				break;
2212 			}
2213 
2214 			sym = kobj_getsymname(rec->ttr_regs.r_pc, &off);
2215 			if (sym != NULL) {
2216 				printf("%s %s %s %s+%lx\n", data1, data2, data3,
2217 				    sym, off);
2218 			} else {
2219 				printf("%s %s %s %lx\n", data1, data2, data3,
2220 				    rec->ttr_regs.r_pc);
2221 			}
2222 
2223 			if (ttrace_dump_nregs-- > 0) {
2224 				int s;
2225 
2226 				if (rec->ttr_marker == TT_INTERRUPT)
2227 					printf(
2228 					    "\t\tipl %x spl %x pri %x\n",
2229 					    rec->ttr_ipl,
2230 					    rec->ttr_spl,
2231 					    rec->ttr_pri);
2232 
2233 				dumpregs(&rec->ttr_regs);
2234 
2235 				printf("\t%3s: %p\n\n", " ct",
2236 				    (void *)rec->ttr_curthread);
2237 
2238 				/*
2239 				 * print out the pc stack that we recorded
2240 				 * at trap time (if any)
2241 				 */
2242 				for (s = 0; s < rec->ttr_sdepth; s++) {
2243 					uintptr_t fullpc;
2244 
2245 					if (s >= TTR_STACK_DEPTH) {
2246 						printf("ttr_sdepth corrupt\n");
2247 						break;
2248 					}
2249 
2250 					fullpc = (uintptr_t)rec->ttr_stack[s];
2251 
2252 					sym = kobj_getsymname(fullpc, &off);
2253 					if (sym != NULL)
2254 						printf("-> %s+0x%lx()\n",
2255 						    sym, off);
2256 					else
2257 						printf("-> 0x%lx()\n", fullpc);
2258 				}
2259 				printf("\n");
2260 			}
2261 			current -= sizeof (trap_trace_rec_t);
2262 		}
2263 	}
2264 }
2265 
2266 #endif	/* TRAPTRACE */
2267 
2268 void
2269 panic_showtrap(struct panic_trap_info *tip)
2270 {
2271 	showregs(tip->trap_type, tip->trap_regs, tip->trap_addr);
2272 
2273 #if defined(TRAPTRACE)
2274 	dump_ttrace();
2275 #endif
2276 
2277 #if !defined(__xpv)
2278 	if (tip->trap_type == T_DBLFLT)
2279 		dump_tss();
2280 #endif
2281 }
2282 
2283 void
2284 panic_savetrap(panic_data_t *pdp, struct panic_trap_info *tip)
2285 {
2286 	panic_saveregs(pdp, tip->trap_regs);
2287 }
2288