xref: /illumos-gate/usr/src/uts/i86pc/os/trap.c (revision 6523a3aa7f325d64841382707603be7a86e68147)
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 	errcode = 0;
492 	mstate = 0;
493 	rw = S_OTHER;
494 	type = rp->r_trapno;
495 	CPU_STATS_ADDQ(CPU, sys, trap, 1);
496 	ASSERT(ct->t_schedflag & TS_DONT_SWAP);
497 
498 	if (type == T_PGFLT) {
499 
500 		errcode = rp->r_err;
501 		if (errcode & PF_ERR_WRITE)
502 			rw = S_WRITE;
503 		else if ((caddr_t)rp->r_pc == addr ||
504 		    (mmu.pt_nx != 0 && (errcode & PF_ERR_EXEC)))
505 			rw = S_EXEC;
506 		else
507 			rw = S_READ;
508 
509 #if defined(__i386)
510 		/*
511 		 * Pentium Pro work-around
512 		 */
513 		if ((errcode & PF_ERR_PROT) && pentiumpro_bug4046376) {
514 			uint_t	attr;
515 			uint_t	priv_violation;
516 			uint_t	access_violation;
517 
518 			if (hat_getattr(addr < (caddr_t)kernelbase ?
519 			    curproc->p_as->a_hat : kas.a_hat, addr, &attr)
520 			    == -1) {
521 				errcode &= ~PF_ERR_PROT;
522 			} else {
523 				priv_violation = (errcode & PF_ERR_USER) &&
524 				    !(attr & PROT_USER);
525 				access_violation = (errcode & PF_ERR_WRITE) &&
526 				    !(attr & PROT_WRITE);
527 				if (!priv_violation && !access_violation)
528 					goto cleanup;
529 			}
530 		}
531 #endif /* __i386 */
532 
533 	} else if (type == T_SGLSTP && lwp != NULL)
534 		lwp->lwp_pcb.pcb_drstat = (uintptr_t)addr;
535 
536 	if (tdebug)
537 		showregs(type, rp, addr);
538 
539 	if (USERMODE(rp->r_cs)) {
540 		/*
541 		 * Set up the current cred to use during this trap. u_cred
542 		 * no longer exists.  t_cred is used instead.
543 		 * The current process credential applies to the thread for
544 		 * the entire trap.  If trapping from the kernel, this
545 		 * should already be set up.
546 		 */
547 		if (ct->t_cred != p->p_cred) {
548 			cred_t *oldcred = ct->t_cred;
549 			/*
550 			 * DTrace accesses t_cred in probe context.  t_cred
551 			 * must always be either NULL, or point to a valid,
552 			 * allocated cred structure.
553 			 */
554 			ct->t_cred = crgetcred();
555 			crfree(oldcred);
556 		}
557 		ASSERT(lwp != NULL);
558 		type |= USER;
559 		ASSERT(lwptoregs(lwp) == rp);
560 		lwp->lwp_state = LWP_SYS;
561 
562 		switch (type) {
563 		case T_PGFLT + USER:
564 			if ((caddr_t)rp->r_pc == addr)
565 				mstate = LMS_TFAULT;
566 			else
567 				mstate = LMS_DFAULT;
568 			break;
569 		default:
570 			mstate = LMS_TRAP;
571 			break;
572 		}
573 		/* Kernel probe */
574 		TNF_PROBE_1(thread_state, "thread", /* CSTYLED */,
575 		    tnf_microstate, state, mstate);
576 		mstate = new_mstate(ct, mstate);
577 
578 		bzero(&siginfo, sizeof (siginfo));
579 	}
580 
581 	switch (type) {
582 	case T_PGFLT + USER:
583 	case T_SGLSTP:
584 	case T_SGLSTP + USER:
585 	case T_BPTFLT + USER:
586 		break;
587 
588 	default:
589 		FTRACE_2("trap(): type=0x%lx, regs=0x%lx",
590 		    (ulong_t)type, (ulong_t)rp);
591 		break;
592 	}
593 
594 	switch (type) {
595 	case T_SIMDFPE:
596 		/* Make sure we enable interrupts before die()ing */
597 		sti();	/* The SIMD exception comes in via cmninttrap */
598 		/*FALLTHROUGH*/
599 	default:
600 		if (type & USER) {
601 			if (tudebug)
602 				showregs(type, rp, (caddr_t)0);
603 			printf("trap: Unknown trap type %d in user mode\n",
604 			    type & ~USER);
605 			siginfo.si_signo = SIGILL;
606 			siginfo.si_code  = ILL_ILLTRP;
607 			siginfo.si_addr  = (caddr_t)rp->r_pc;
608 			siginfo.si_trapno = type & ~USER;
609 			fault = FLTILL;
610 		} else {
611 			(void) die(type, rp, addr, cpuid);
612 			/*NOTREACHED*/
613 		}
614 		break;
615 
616 	case T_PGFLT:		/* system page fault */
617 		/*
618 		 * If we're under on_trap() protection (see <sys/ontrap.h>),
619 		 * set ot_trap and bounce back to the on_trap() call site
620 		 * via the installed trampoline.
621 		 */
622 		if ((ct->t_ontrap != NULL) &&
623 		    (ct->t_ontrap->ot_prot & OT_DATA_ACCESS)) {
624 			ct->t_ontrap->ot_trap |= OT_DATA_ACCESS;
625 			rp->r_pc = ct->t_ontrap->ot_trampoline;
626 			goto cleanup;
627 		}
628 
629 		/*
630 		 * If we have an Instruction fault in kernel mode, then that
631 		 * means we've tried to execute a user page (SMEP) or both of
632 		 * PAE and NXE are enabled. In either case, given that it's a
633 		 * kernel fault, we should panic immediately and not try to make
634 		 * any more forward progress. This indicates a bug in the
635 		 * kernel, which if execution continued, could be exploited to
636 		 * wreak havoc on the system.
637 		 */
638 		if (errcode & PF_ERR_EXEC) {
639 			(void) die(type, rp, addr, cpuid);
640 		}
641 
642 		/*
643 		 * We need to check if SMAP is in play. If SMAP is in play, then
644 		 * any access to a user page will show up as a protection
645 		 * violation. To see if SMAP is enabled we first check if it's a
646 		 * user address and whether we have the feature flag set. If we
647 		 * do and the interrupted registers do not allow for user
648 		 * accesses (PS_ACHK is not enabled), then we need to die
649 		 * immediately.
650 		 */
651 		if (addr < (caddr_t)kernelbase &&
652 		    is_x86_feature(x86_featureset, X86FSET_SMAP) == B_TRUE &&
653 		    (rp->r_ps & PS_ACHK) == 0) {
654 			(void) die(type, rp, addr, cpuid);
655 		}
656 
657 		/*
658 		 * See if we can handle as pagefault. Save lofault and onfault
659 		 * across this. Here we assume that an address less than
660 		 * KERNELBASE is a user fault.  We can do this as copy.s
661 		 * routines verify that the starting address is less than
662 		 * KERNELBASE before starting and because we know that we
663 		 * always have KERNELBASE mapped as invalid to serve as a
664 		 * "barrier".
665 		 */
666 		lofault = ct->t_lofault;
667 		onfault = ct->t_onfault;
668 		ct->t_lofault = 0;
669 
670 		mstate = new_mstate(ct, LMS_KFAULT);
671 
672 		if (addr < (caddr_t)kernelbase) {
673 			res = pagefault(addr,
674 			    (errcode & PF_ERR_PROT)? F_PROT: F_INVAL, rw, 0);
675 			if (res == FC_NOMAP &&
676 			    addr < p->p_usrstack &&
677 			    grow(addr))
678 				res = 0;
679 		} else {
680 			res = pagefault(addr,
681 			    (errcode & PF_ERR_PROT)? F_PROT: F_INVAL, rw, 1);
682 		}
683 		(void) new_mstate(ct, mstate);
684 
685 		/*
686 		 * Restore lofault and onfault. If we resolved the fault, exit.
687 		 * If we didn't and lofault wasn't set, die.
688 		 */
689 		ct->t_lofault = lofault;
690 		ct->t_onfault = onfault;
691 		if (res == 0)
692 			goto cleanup;
693 
694 #if defined(OPTERON_ERRATUM_93) && defined(_LP64)
695 		if (lofault == 0 && opteron_erratum_93) {
696 			/*
697 			 * Workaround for Opteron Erratum 93. On return from
698 			 * a System Managment Interrupt at a HLT instruction
699 			 * the %rip might be truncated to a 32 bit value.
700 			 * BIOS is supposed to fix this, but some don't.
701 			 * If this occurs we simply restore the high order bits.
702 			 * The HLT instruction is 1 byte of 0xf4.
703 			 */
704 			uintptr_t	rip = rp->r_pc;
705 
706 			if ((rip & 0xfffffffful) == rip) {
707 				rip |= 0xfffffffful << 32;
708 				if (hat_getpfnum(kas.a_hat, (caddr_t)rip) !=
709 				    PFN_INVALID &&
710 				    (*(uchar_t *)rip == 0xf4 ||
711 				    *(uchar_t *)(rip - 1) == 0xf4)) {
712 					rp->r_pc = rip;
713 					goto cleanup;
714 				}
715 			}
716 		}
717 #endif /* OPTERON_ERRATUM_93 && _LP64 */
718 
719 #ifdef OPTERON_ERRATUM_91
720 		if (lofault == 0 && opteron_erratum_91) {
721 			/*
722 			 * Workaround for Opteron Erratum 91. Prefetches may
723 			 * generate a page fault (they're not supposed to do
724 			 * that!). If this occurs we simply return back to the
725 			 * instruction.
726 			 */
727 			caddr_t		pc = (caddr_t)rp->r_pc;
728 
729 			/*
730 			 * If the faulting PC is not mapped, this is a
731 			 * legitimate kernel page fault that must result in a
732 			 * panic. If the faulting PC is mapped, it could contain
733 			 * a prefetch instruction. Check for that here.
734 			 */
735 			if (hat_getpfnum(kas.a_hat, pc) != PFN_INVALID) {
736 				if (cmp_to_prefetch((uchar_t *)pc)) {
737 #ifdef DEBUG
738 					cmn_err(CE_WARN, "Opteron erratum 91 "
739 					    "occurred: kernel prefetch"
740 					    " at %p generated a page fault!",
741 					    (void *)rp->r_pc);
742 #endif /* DEBUG */
743 					goto cleanup;
744 				}
745 			}
746 			(void) die(type, rp, addr, cpuid);
747 		}
748 #endif /* OPTERON_ERRATUM_91 */
749 
750 		if (lofault == 0)
751 			(void) die(type, rp, addr, cpuid);
752 
753 		/*
754 		 * Cannot resolve fault.  Return to lofault.
755 		 */
756 		if (lodebug) {
757 			showregs(type, rp, addr);
758 			traceregs(rp);
759 		}
760 		if (FC_CODE(res) == FC_OBJERR)
761 			res = FC_ERRNO(res);
762 		else
763 			res = EFAULT;
764 		rp->r_r0 = res;
765 		rp->r_pc = ct->t_lofault;
766 		goto cleanup;
767 
768 	case T_PGFLT + USER:	/* user page fault */
769 		if (faultdebug) {
770 			char *fault_str;
771 
772 			switch (rw) {
773 			case S_READ:
774 				fault_str = "read";
775 				break;
776 			case S_WRITE:
777 				fault_str = "write";
778 				break;
779 			case S_EXEC:
780 				fault_str = "exec";
781 				break;
782 			default:
783 				fault_str = "";
784 				break;
785 			}
786 			printf("user %s fault:  addr=0x%lx errcode=0x%x\n",
787 			    fault_str, (uintptr_t)addr, errcode);
788 		}
789 
790 #if defined(OPTERON_ERRATUM_100) && defined(_LP64)
791 		/*
792 		 * Workaround for AMD erratum 100
793 		 *
794 		 * A 32-bit process may receive a page fault on a non
795 		 * 32-bit address by mistake. The range of the faulting
796 		 * address will be
797 		 *
798 		 *	0xffffffff80000000 .. 0xffffffffffffffff or
799 		 *	0x0000000100000000 .. 0x000000017fffffff
800 		 *
801 		 * The fault is always due to an instruction fetch, however
802 		 * the value of r_pc should be correct (in 32 bit range),
803 		 * so we ignore the page fault on the bogus address.
804 		 */
805 		if (p->p_model == DATAMODEL_ILP32 &&
806 		    (0xffffffff80000000 <= (uintptr_t)addr ||
807 		    (0x100000000 <= (uintptr_t)addr &&
808 		    (uintptr_t)addr <= 0x17fffffff))) {
809 			if (!opteron_erratum_100)
810 				panic("unexpected erratum #100");
811 			if (rp->r_pc <= 0xffffffff)
812 				goto out;
813 		}
814 #endif /* OPTERON_ERRATUM_100 && _LP64 */
815 
816 		ASSERT(!(curthread->t_flag & T_WATCHPT));
817 		watchpage = (pr_watch_active(p) && pr_is_watchpage(addr, rw));
818 #ifdef __i386
819 		/*
820 		 * In 32-bit mode, the lcall (system call) instruction fetches
821 		 * one word from the stack, at the stack pointer, because of the
822 		 * way the call gate is constructed.  This is a bogus
823 		 * read and should not be counted as a read watchpoint.
824 		 * We work around the problem here by testing to see if
825 		 * this situation applies and, if so, simply jumping to
826 		 * the code in locore.s that fields the system call trap.
827 		 * The registers on the stack are already set up properly
828 		 * due to the match between the call gate sequence and the
829 		 * trap gate sequence.  We just have to adjust the pc.
830 		 */
831 		if (watchpage && addr == (caddr_t)rp->r_sp &&
832 		    rw == S_READ && instr_is_lcall_syscall((caddr_t)rp->r_pc)) {
833 			extern void watch_syscall(void);
834 
835 			rp->r_pc += LCALLSIZE;
836 			watch_syscall();	/* never returns */
837 			/* NOTREACHED */
838 		}
839 #endif /* __i386 */
840 		vaddr = addr;
841 		if (!watchpage || (sz = instr_size(rp, &vaddr, rw)) <= 0)
842 			fault_type = (errcode & PF_ERR_PROT)? F_PROT: F_INVAL;
843 		else if ((watchcode = pr_is_watchpoint(&vaddr, &ta,
844 		    sz, NULL, rw)) != 0) {
845 			if (ta) {
846 				do_watch_step(vaddr, sz, rw,
847 				    watchcode, rp->r_pc);
848 				fault_type = F_INVAL;
849 			} else {
850 				bzero(&siginfo, sizeof (siginfo));
851 				siginfo.si_signo = SIGTRAP;
852 				siginfo.si_code = watchcode;
853 				siginfo.si_addr = vaddr;
854 				siginfo.si_trapafter = 0;
855 				siginfo.si_pc = (caddr_t)rp->r_pc;
856 				fault = FLTWATCH;
857 				break;
858 			}
859 		} else {
860 			/* XXX pr_watch_emul() never succeeds (for now) */
861 			if (rw != S_EXEC && pr_watch_emul(rp, vaddr, rw))
862 				goto out;
863 			do_watch_step(vaddr, sz, rw, 0, 0);
864 			fault_type = F_INVAL;
865 		}
866 
867 		res = pagefault(addr, fault_type, rw, 0);
868 
869 		/*
870 		 * If pagefault() succeeded, ok.
871 		 * Otherwise attempt to grow the stack.
872 		 */
873 		if (res == 0 ||
874 		    (res == FC_NOMAP &&
875 		    addr < p->p_usrstack &&
876 		    grow(addr))) {
877 			lwp->lwp_lastfault = FLTPAGE;
878 			lwp->lwp_lastfaddr = addr;
879 			if (prismember(&p->p_fltmask, FLTPAGE)) {
880 				bzero(&siginfo, sizeof (siginfo));
881 				siginfo.si_addr = addr;
882 				(void) stop_on_fault(FLTPAGE, &siginfo);
883 			}
884 			goto out;
885 		} else if (res == FC_PROT && addr < p->p_usrstack &&
886 		    (mmu.pt_nx != 0 && (errcode & PF_ERR_EXEC))) {
887 			report_stack_exec(p, addr);
888 		}
889 
890 #ifdef OPTERON_ERRATUM_91
891 		/*
892 		 * Workaround for Opteron Erratum 91. Prefetches may generate a
893 		 * page fault (they're not supposed to do that!). If this
894 		 * occurs we simply return back to the instruction.
895 		 *
896 		 * We rely on copyin to properly fault in the page with r_pc.
897 		 */
898 		if (opteron_erratum_91 &&
899 		    addr != (caddr_t)rp->r_pc &&
900 		    instr_is_prefetch((caddr_t)rp->r_pc)) {
901 #ifdef DEBUG
902 			cmn_err(CE_WARN, "Opteron erratum 91 occurred: "
903 			    "prefetch at %p in pid %d generated a trap!",
904 			    (void *)rp->r_pc, p->p_pid);
905 #endif /* DEBUG */
906 			goto out;
907 		}
908 #endif /* OPTERON_ERRATUM_91 */
909 
910 		if (tudebug)
911 			showregs(type, rp, addr);
912 		/*
913 		 * In the case where both pagefault and grow fail,
914 		 * set the code to the value provided by pagefault.
915 		 * We map all errors returned from pagefault() to SIGSEGV.
916 		 */
917 		bzero(&siginfo, sizeof (siginfo));
918 		siginfo.si_addr = addr;
919 		switch (FC_CODE(res)) {
920 		case FC_HWERR:
921 		case FC_NOSUPPORT:
922 			siginfo.si_signo = SIGBUS;
923 			siginfo.si_code = BUS_ADRERR;
924 			fault = FLTACCESS;
925 			break;
926 		case FC_ALIGN:
927 			siginfo.si_signo = SIGBUS;
928 			siginfo.si_code = BUS_ADRALN;
929 			fault = FLTACCESS;
930 			break;
931 		case FC_OBJERR:
932 			if ((siginfo.si_errno = FC_ERRNO(res)) != EINTR) {
933 				siginfo.si_signo = SIGBUS;
934 				siginfo.si_code = BUS_OBJERR;
935 				fault = FLTACCESS;
936 			}
937 			break;
938 		default:	/* FC_NOMAP or FC_PROT */
939 			siginfo.si_signo = SIGSEGV;
940 			siginfo.si_code =
941 			    (res == FC_NOMAP)? SEGV_MAPERR : SEGV_ACCERR;
942 			fault = FLTBOUNDS;
943 			break;
944 		}
945 		break;
946 
947 	case T_ILLINST + USER:	/* invalid opcode fault */
948 		/*
949 		 * If the syscall instruction is disabled due to LDT usage, a
950 		 * user program that attempts to execute it will trigger a #ud
951 		 * trap. Check for that case here. If this occurs on a CPU which
952 		 * doesn't even support syscall, the result of all of this will
953 		 * be to emulate that particular instruction.
954 		 */
955 		if (p->p_ldt != NULL &&
956 		    ldt_rewrite_syscall(rp, p, X86FSET_ASYSC))
957 			goto out;
958 
959 #ifdef __amd64
960 		/*
961 		 * Emulate the LAHF and SAHF instructions if needed.
962 		 * See the instr_is_lsahf function for details.
963 		 */
964 		if (p->p_model == DATAMODEL_LP64 &&
965 		    instr_is_lsahf((caddr_t)rp->r_pc, &instr)) {
966 			emulate_lsahf(rp, instr);
967 			goto out;
968 		}
969 #endif
970 
971 		/*FALLTHROUGH*/
972 
973 		if (tudebug)
974 			showregs(type, rp, (caddr_t)0);
975 		siginfo.si_signo = SIGILL;
976 		siginfo.si_code  = ILL_ILLOPC;
977 		siginfo.si_addr  = (caddr_t)rp->r_pc;
978 		fault = FLTILL;
979 		break;
980 
981 	case T_ZERODIV + USER:		/* integer divide by zero */
982 		if (tudebug && tudebugfpe)
983 			showregs(type, rp, (caddr_t)0);
984 		siginfo.si_signo = SIGFPE;
985 		siginfo.si_code  = FPE_INTDIV;
986 		siginfo.si_addr  = (caddr_t)rp->r_pc;
987 		fault = FLTIZDIV;
988 		break;
989 
990 	case T_OVFLW + USER:	/* integer overflow */
991 		if (tudebug && tudebugfpe)
992 			showregs(type, rp, (caddr_t)0);
993 		siginfo.si_signo = SIGFPE;
994 		siginfo.si_code  = FPE_INTOVF;
995 		siginfo.si_addr  = (caddr_t)rp->r_pc;
996 		fault = FLTIOVF;
997 		break;
998 
999 	/*
1000 	 * When using an eager FPU on x86, the #NM trap is no longer meaningful.
1001 	 * Userland should not be able to trigger it. Anything that does
1002 	 * represents a fatal error in the kernel and likely in the register
1003 	 * state of the system. User FPU state should always be valid.
1004 	 */
1005 	case T_NOEXTFLT + USER:	/* math coprocessor not available */
1006 	case T_NOEXTFLT:
1007 		(void) die(type, rp, addr, cpuid);
1008 		break;
1009 
1010 	/*
1011 	 * Kernel threads leveraging floating point need to mask the exceptions
1012 	 * or ensure that they cannot happen. There is no recovery from this.
1013 	 */
1014 	case T_EXTERRFLT:	/* x87 floating point exception pending */
1015 		sti(); /* T_EXTERRFLT comes in via cmninttrap */
1016 		(void) die(type, rp, addr, cpuid);
1017 		break;
1018 
1019 	case T_EXTERRFLT + USER: /* x87 floating point exception pending */
1020 		if (tudebug && tudebugfpe)
1021 			showregs(type, rp, addr);
1022 		if (sicode = fpexterrflt(rp)) {
1023 			siginfo.si_signo = SIGFPE;
1024 			siginfo.si_code  = sicode;
1025 			siginfo.si_addr  = (caddr_t)rp->r_pc;
1026 			fault = FLTFPE;
1027 		}
1028 		break;
1029 
1030 	case T_SIMDFPE + USER:		/* SSE and SSE2 exceptions */
1031 		if (tudebug && tudebugsse)
1032 			showregs(type, rp, addr);
1033 		if (!is_x86_feature(x86_featureset, X86FSET_SSE) &&
1034 		    !is_x86_feature(x86_featureset, X86FSET_SSE2)) {
1035 			/*
1036 			 * There are rumours that some user instructions
1037 			 * on older CPUs can cause this trap to occur; in
1038 			 * which case send a SIGILL instead of a SIGFPE.
1039 			 */
1040 			siginfo.si_signo = SIGILL;
1041 			siginfo.si_code  = ILL_ILLTRP;
1042 			siginfo.si_addr  = (caddr_t)rp->r_pc;
1043 			siginfo.si_trapno = type & ~USER;
1044 			fault = FLTILL;
1045 		} else if ((sicode = fpsimderrflt(rp)) != 0) {
1046 			siginfo.si_signo = SIGFPE;
1047 			siginfo.si_code = sicode;
1048 			siginfo.si_addr = (caddr_t)rp->r_pc;
1049 			fault = FLTFPE;
1050 		}
1051 
1052 		sti();	/* The SIMD exception comes in via cmninttrap */
1053 		break;
1054 
1055 	case T_BPTFLT:	/* breakpoint trap */
1056 		/*
1057 		 * Kernel breakpoint traps should only happen when kmdb is
1058 		 * active, and even then, it'll have interposed on the IDT, so
1059 		 * control won't get here.  If it does, we've hit a breakpoint
1060 		 * without the debugger, which is very strange, and very
1061 		 * fatal.
1062 		 */
1063 		if (tudebug && tudebugbpt)
1064 			showregs(type, rp, (caddr_t)0);
1065 
1066 		(void) die(type, rp, addr, cpuid);
1067 		break;
1068 
1069 	case T_SGLSTP: /* single step/hw breakpoint exception */
1070 
1071 #if !defined(__xpv)
1072 		/*
1073 		 * We'd never normally get here, as kmdb handles its own single
1074 		 * step traps.  There is one nasty exception though, as
1075 		 * described in more detail in sys_sysenter().  Note that
1076 		 * checking for all four locations covers both the KPTI and the
1077 		 * non-KPTI cases correctly: the former will never be found at
1078 		 * (brand_)sys_sysenter, and vice versa.
1079 		 */
1080 		if (lwp != NULL && (lwp->lwp_pcb.pcb_drstat & DR_SINGLESTEP)) {
1081 			if (rp->r_pc == (greg_t)brand_sys_sysenter ||
1082 			    rp->r_pc == (greg_t)sys_sysenter ||
1083 			    rp->r_pc == (greg_t)tr_brand_sys_sysenter ||
1084 			    rp->r_pc == (greg_t)tr_sys_sysenter) {
1085 
1086 				rp->r_pc += 0x3; /* sizeof (swapgs) */
1087 
1088 				rp->r_ps &= ~PS_T; /* turn off trace */
1089 				lwp->lwp_pcb.pcb_flags |= DEBUG_PENDING;
1090 				ct->t_post_sys = 1;
1091 				aston(curthread);
1092 				goto cleanup;
1093 			} else {
1094 				if (tudebug && tudebugbpt)
1095 					showregs(type, rp, (caddr_t)0);
1096 			}
1097 		}
1098 #endif /* !__xpv */
1099 
1100 		if (boothowto & RB_DEBUG)
1101 			debug_enter((char *)NULL);
1102 		else
1103 			(void) die(type, rp, addr, cpuid);
1104 		break;
1105 
1106 	case T_NMIFLT:	/* NMI interrupt */
1107 		printf("Unexpected NMI in system mode\n");
1108 		goto cleanup;
1109 
1110 	case T_NMIFLT + USER:	/* NMI interrupt */
1111 		printf("Unexpected NMI in user mode\n");
1112 		break;
1113 
1114 	case T_GPFLT:	/* general protection violation */
1115 		/*
1116 		 * Any #GP that occurs during an on_trap .. no_trap bracket
1117 		 * with OT_DATA_ACCESS or OT_SEGMENT_ACCESS protection,
1118 		 * or in a on_fault .. no_fault bracket, is forgiven
1119 		 * and we trampoline.  This protection is given regardless
1120 		 * of whether we are 32/64 bit etc - if a distinction is
1121 		 * required then define new on_trap protection types.
1122 		 *
1123 		 * On amd64, we can get a #gp from referencing addresses
1124 		 * in the virtual address hole e.g. from a copyin or in
1125 		 * update_sregs while updating user segment registers.
1126 		 *
1127 		 * On the 32-bit hypervisor we could also generate one in
1128 		 * mfn_to_pfn by reaching around or into where the hypervisor
1129 		 * lives which is protected by segmentation.
1130 		 */
1131 
1132 		/*
1133 		 * If we're under on_trap() protection (see <sys/ontrap.h>),
1134 		 * set ot_trap and trampoline back to the on_trap() call site
1135 		 * for OT_DATA_ACCESS or OT_SEGMENT_ACCESS.
1136 		 */
1137 		if (ct->t_ontrap != NULL) {
1138 			int ttype =  ct->t_ontrap->ot_prot &
1139 			    (OT_DATA_ACCESS | OT_SEGMENT_ACCESS);
1140 
1141 			if (ttype != 0) {
1142 				ct->t_ontrap->ot_trap |= ttype;
1143 				if (tudebug)
1144 					showregs(type, rp, (caddr_t)0);
1145 				rp->r_pc = ct->t_ontrap->ot_trampoline;
1146 				goto cleanup;
1147 			}
1148 		}
1149 
1150 		/*
1151 		 * If we're under lofault protection (copyin etc.),
1152 		 * longjmp back to lofault with an EFAULT.
1153 		 */
1154 		if (ct->t_lofault) {
1155 			/*
1156 			 * Fault is not resolvable, so just return to lofault
1157 			 */
1158 			if (lodebug) {
1159 				showregs(type, rp, addr);
1160 				traceregs(rp);
1161 			}
1162 			rp->r_r0 = EFAULT;
1163 			rp->r_pc = ct->t_lofault;
1164 			goto cleanup;
1165 		}
1166 
1167 		/*
1168 		 * We fall through to the next case, which repeats
1169 		 * the OT_SEGMENT_ACCESS check which we've already
1170 		 * done, so we'll always fall through to the
1171 		 * T_STKFLT case.
1172 		 */
1173 		/*FALLTHROUGH*/
1174 	case T_SEGFLT:	/* segment not present fault */
1175 		/*
1176 		 * One example of this is #NP in update_sregs while
1177 		 * attempting to update a user segment register
1178 		 * that points to a descriptor that is marked not
1179 		 * present.
1180 		 */
1181 		if (ct->t_ontrap != NULL &&
1182 		    ct->t_ontrap->ot_prot & OT_SEGMENT_ACCESS) {
1183 			ct->t_ontrap->ot_trap |= OT_SEGMENT_ACCESS;
1184 			if (tudebug)
1185 				showregs(type, rp, (caddr_t)0);
1186 			rp->r_pc = ct->t_ontrap->ot_trampoline;
1187 			goto cleanup;
1188 		}
1189 		/*FALLTHROUGH*/
1190 	case T_STKFLT:	/* stack fault */
1191 	case T_TSSFLT:	/* invalid TSS fault */
1192 		if (tudebug)
1193 			showregs(type, rp, (caddr_t)0);
1194 		if (kern_gpfault(rp))
1195 			(void) die(type, rp, addr, cpuid);
1196 		goto cleanup;
1197 
1198 	/*
1199 	 * ONLY 32-bit PROCESSES can USE a PRIVATE LDT! 64-bit apps
1200 	 * should have no need for them, so we put a stop to it here.
1201 	 *
1202 	 * So: not-present fault is ONLY valid for 32-bit processes with
1203 	 * a private LDT trying to do a system call. Emulate it.
1204 	 *
1205 	 * #gp fault is ONLY valid for 32-bit processes also, which DO NOT
1206 	 * have a private LDT, and are trying to do a system call. Emulate it.
1207 	 */
1208 
1209 	case T_SEGFLT + USER:	/* segment not present fault */
1210 	case T_GPFLT + USER:	/* general protection violation */
1211 #ifdef _SYSCALL32_IMPL
1212 		if (p->p_model != DATAMODEL_NATIVE) {
1213 #endif /* _SYSCALL32_IMPL */
1214 		if (instr_is_lcall_syscall((caddr_t)rp->r_pc)) {
1215 			if (type == T_SEGFLT + USER)
1216 				ASSERT(p->p_ldt != NULL);
1217 
1218 			if ((p->p_ldt == NULL && type == T_GPFLT + USER) ||
1219 			    type == T_SEGFLT + USER) {
1220 
1221 			/*
1222 			 * The user attempted a system call via the obsolete
1223 			 * call gate mechanism. Because the process doesn't have
1224 			 * an LDT (i.e. the ldtr contains 0), a #gp results.
1225 			 * Emulate the syscall here, just as we do above for a
1226 			 * #np trap.
1227 			 */
1228 
1229 			/*
1230 			 * Since this is a not-present trap, rp->r_pc points to
1231 			 * the trapping lcall instruction. We need to bump it
1232 			 * to the next insn so the app can continue on.
1233 			 */
1234 			rp->r_pc += LCALLSIZE;
1235 			lwp->lwp_regs = rp;
1236 
1237 			/*
1238 			 * Normally the microstate of the LWP is forced back to
1239 			 * LMS_USER by the syscall handlers. Emulate that
1240 			 * behavior here.
1241 			 */
1242 			mstate = LMS_USER;
1243 
1244 			dosyscall();
1245 			goto out;
1246 			}
1247 		}
1248 #ifdef _SYSCALL32_IMPL
1249 		}
1250 #endif /* _SYSCALL32_IMPL */
1251 		/*
1252 		 * If the current process is using a private LDT and the
1253 		 * trapping instruction is sysenter, the sysenter instruction
1254 		 * has been disabled on the CPU because it destroys segment
1255 		 * registers. If this is the case, rewrite the instruction to
1256 		 * be a safe system call and retry it. If this occurs on a CPU
1257 		 * which doesn't even support sysenter, the result of all of
1258 		 * this will be to emulate that particular instruction.
1259 		 */
1260 		if (p->p_ldt != NULL &&
1261 		    ldt_rewrite_syscall(rp, p, X86FSET_SEP))
1262 			goto out;
1263 
1264 		/*FALLTHROUGH*/
1265 
1266 	case T_BOUNDFLT + USER:	/* bound fault */
1267 	case T_STKFLT + USER:	/* stack fault */
1268 	case T_TSSFLT + USER:	/* invalid TSS fault */
1269 		if (tudebug)
1270 			showregs(type, rp, (caddr_t)0);
1271 		siginfo.si_signo = SIGSEGV;
1272 		siginfo.si_code  = SEGV_MAPERR;
1273 		siginfo.si_addr  = (caddr_t)rp->r_pc;
1274 		fault = FLTBOUNDS;
1275 		break;
1276 
1277 	case T_ALIGNMENT + USER:	/* user alignment error (486) */
1278 		if (tudebug)
1279 			showregs(type, rp, (caddr_t)0);
1280 		bzero(&siginfo, sizeof (siginfo));
1281 		siginfo.si_signo = SIGBUS;
1282 		siginfo.si_code = BUS_ADRALN;
1283 		siginfo.si_addr = (caddr_t)rp->r_pc;
1284 		fault = FLTACCESS;
1285 		break;
1286 
1287 	case T_SGLSTP + USER: /* single step/hw breakpoint exception */
1288 		if (tudebug && tudebugbpt)
1289 			showregs(type, rp, (caddr_t)0);
1290 
1291 		/* Was it single-stepping? */
1292 		if (lwp->lwp_pcb.pcb_drstat & DR_SINGLESTEP) {
1293 			pcb_t *pcb = &lwp->lwp_pcb;
1294 
1295 			rp->r_ps &= ~PS_T;
1296 			/*
1297 			 * If both NORMAL_STEP and WATCH_STEP are in effect,
1298 			 * give precedence to WATCH_STEP.  If neither is set,
1299 			 * user must have set the PS_T bit in %efl; treat this
1300 			 * as NORMAL_STEP.
1301 			 */
1302 			if ((fault = undo_watch_step(&siginfo)) == 0 &&
1303 			    ((pcb->pcb_flags & NORMAL_STEP) ||
1304 			    !(pcb->pcb_flags & WATCH_STEP))) {
1305 				siginfo.si_signo = SIGTRAP;
1306 				siginfo.si_code = TRAP_TRACE;
1307 				siginfo.si_addr = (caddr_t)rp->r_pc;
1308 				fault = FLTTRACE;
1309 			}
1310 			pcb->pcb_flags &= ~(NORMAL_STEP|WATCH_STEP);
1311 		}
1312 		break;
1313 
1314 	case T_BPTFLT + USER:	/* breakpoint trap */
1315 		if (tudebug && tudebugbpt)
1316 			showregs(type, rp, (caddr_t)0);
1317 		/*
1318 		 * int 3 (the breakpoint instruction) leaves the pc referring
1319 		 * to the address one byte after the breakpointed address.
1320 		 * If the P_PR_BPTADJ flag has been set via /proc, We adjust
1321 		 * it back so it refers to the breakpointed address.
1322 		 */
1323 		if (p->p_proc_flag & P_PR_BPTADJ)
1324 			rp->r_pc--;
1325 		siginfo.si_signo = SIGTRAP;
1326 		siginfo.si_code  = TRAP_BRKPT;
1327 		siginfo.si_addr  = (caddr_t)rp->r_pc;
1328 		fault = FLTBPT;
1329 		break;
1330 
1331 	case T_AST:
1332 		/*
1333 		 * This occurs only after the cs register has been made to
1334 		 * look like a kernel selector, either through debugging or
1335 		 * possibly by functions like setcontext().  The thread is
1336 		 * about to cause a general protection fault at common_iret()
1337 		 * in locore.  We let that happen immediately instead of
1338 		 * doing the T_AST processing.
1339 		 */
1340 		goto cleanup;
1341 
1342 	case T_AST + USER:	/* profiling, resched, h/w error pseudo trap */
1343 		if (lwp->lwp_pcb.pcb_flags & ASYNC_HWERR) {
1344 			proc_t *p = ttoproc(curthread);
1345 			extern void print_msg_hwerr(ctid_t ct_id, proc_t *p);
1346 
1347 			lwp->lwp_pcb.pcb_flags &= ~ASYNC_HWERR;
1348 			print_msg_hwerr(p->p_ct_process->conp_contract.ct_id,
1349 			    p);
1350 			contract_process_hwerr(p->p_ct_process, p);
1351 			siginfo.si_signo = SIGKILL;
1352 			siginfo.si_code = SI_NOINFO;
1353 		} else if (lwp->lwp_pcb.pcb_flags & CPC_OVERFLOW) {
1354 			lwp->lwp_pcb.pcb_flags &= ~CPC_OVERFLOW;
1355 			if (kcpc_overflow_ast()) {
1356 				/*
1357 				 * Signal performance counter overflow
1358 				 */
1359 				if (tudebug)
1360 					showregs(type, rp, (caddr_t)0);
1361 				bzero(&siginfo, sizeof (siginfo));
1362 				siginfo.si_signo = SIGEMT;
1363 				siginfo.si_code = EMT_CPCOVF;
1364 				siginfo.si_addr = (caddr_t)rp->r_pc;
1365 				fault = FLTCPCOVF;
1366 			}
1367 		}
1368 
1369 		break;
1370 	}
1371 
1372 	/*
1373 	 * We can't get here from a system trap
1374 	 */
1375 	ASSERT(type & USER);
1376 
1377 	if (fault) {
1378 		/* We took a fault so abort single step. */
1379 		lwp->lwp_pcb.pcb_flags &= ~(NORMAL_STEP|WATCH_STEP);
1380 		/*
1381 		 * Remember the fault and fault adddress
1382 		 * for real-time (SIGPROF) profiling.
1383 		 */
1384 		lwp->lwp_lastfault = fault;
1385 		lwp->lwp_lastfaddr = siginfo.si_addr;
1386 
1387 		DTRACE_PROC2(fault, int, fault, ksiginfo_t *, &siginfo);
1388 
1389 		/*
1390 		 * If a debugger has declared this fault to be an
1391 		 * event of interest, stop the lwp.  Otherwise just
1392 		 * deliver the associated signal.
1393 		 */
1394 		if (siginfo.si_signo != SIGKILL &&
1395 		    prismember(&p->p_fltmask, fault) &&
1396 		    stop_on_fault(fault, &siginfo) == 0)
1397 			siginfo.si_signo = 0;
1398 	}
1399 
1400 	if (siginfo.si_signo)
1401 		trapsig(&siginfo, (fault != FLTFPE && fault != FLTCPCOVF));
1402 
1403 	if (lwp->lwp_oweupc)
1404 		profil_tick(rp->r_pc);
1405 
1406 	if (ct->t_astflag | ct->t_sig_check) {
1407 		/*
1408 		 * Turn off the AST flag before checking all the conditions that
1409 		 * may have caused an AST.  This flag is on whenever a signal or
1410 		 * unusual condition should be handled after the next trap or
1411 		 * syscall.
1412 		 */
1413 		astoff(ct);
1414 		/*
1415 		 * If a single-step trap occurred on a syscall (see above)
1416 		 * recognize it now.  Do this before checking for signals
1417 		 * because deferred_singlestep_trap() may generate a SIGTRAP to
1418 		 * the LWP or may otherwise mark the LWP to call issig(FORREAL).
1419 		 */
1420 		if (lwp->lwp_pcb.pcb_flags & DEBUG_PENDING)
1421 			deferred_singlestep_trap((caddr_t)rp->r_pc);
1422 
1423 		ct->t_sig_check = 0;
1424 
1425 		/*
1426 		 * As in other code paths that check against TP_CHANGEBIND,
1427 		 * we perform the check first without p_lock held -- only
1428 		 * acquiring p_lock in the unlikely event that it is indeed
1429 		 * set.  This is safe because we are doing this after the
1430 		 * astoff(); if we are racing another thread setting
1431 		 * TP_CHANGEBIND on us, we will pick it up on a subsequent
1432 		 * lap through.
1433 		 */
1434 		if (curthread->t_proc_flag & TP_CHANGEBIND) {
1435 			mutex_enter(&p->p_lock);
1436 			if (curthread->t_proc_flag & TP_CHANGEBIND) {
1437 				timer_lwpbind();
1438 				curthread->t_proc_flag &= ~TP_CHANGEBIND;
1439 			}
1440 			mutex_exit(&p->p_lock);
1441 		}
1442 
1443 		/*
1444 		 * for kaio requests that are on the per-process poll queue,
1445 		 * aiop->aio_pollq, they're AIO_POLL bit is set, the kernel
1446 		 * should copyout their result_t to user memory. by copying
1447 		 * out the result_t, the user can poll on memory waiting
1448 		 * for the kaio request to complete.
1449 		 */
1450 		if (p->p_aio)
1451 			aio_cleanup(0);
1452 		/*
1453 		 * If this LWP was asked to hold, call holdlwp(), which will
1454 		 * stop.  holdlwps() sets this up and calls pokelwps() which
1455 		 * sets the AST flag.
1456 		 *
1457 		 * Also check TP_EXITLWP, since this is used by fresh new LWPs
1458 		 * through lwp_rtt().  That flag is set if the lwp_create(2)
1459 		 * syscall failed after creating the LWP.
1460 		 */
1461 		if (ISHOLD(p))
1462 			holdlwp();
1463 
1464 		/*
1465 		 * All code that sets signals and makes ISSIG evaluate true must
1466 		 * set t_astflag afterwards.
1467 		 */
1468 		if (ISSIG_PENDING(ct, lwp, p)) {
1469 			if (issig(FORREAL))
1470 				psig();
1471 			ct->t_sig_check = 1;
1472 		}
1473 
1474 		if (ct->t_rprof != NULL) {
1475 			realsigprof(0, 0, 0);
1476 			ct->t_sig_check = 1;
1477 		}
1478 
1479 		/*
1480 		 * /proc can't enable/disable the trace bit itself
1481 		 * because that could race with the call gate used by
1482 		 * system calls via "lcall". If that happened, an
1483 		 * invalid EFLAGS would result. prstep()/prnostep()
1484 		 * therefore schedule an AST for the purpose.
1485 		 */
1486 		if (lwp->lwp_pcb.pcb_flags & REQUEST_STEP) {
1487 			lwp->lwp_pcb.pcb_flags &= ~REQUEST_STEP;
1488 			rp->r_ps |= PS_T;
1489 		}
1490 		if (lwp->lwp_pcb.pcb_flags & REQUEST_NOSTEP) {
1491 			lwp->lwp_pcb.pcb_flags &= ~REQUEST_NOSTEP;
1492 			rp->r_ps &= ~PS_T;
1493 		}
1494 	}
1495 
1496 out:	/* We can't get here from a system trap */
1497 	ASSERT(type & USER);
1498 
1499 	if (ISHOLD(p))
1500 		holdlwp();
1501 
1502 	/*
1503 	 * Set state to LWP_USER here so preempt won't give us a kernel
1504 	 * priority if it occurs after this point.  Call CL_TRAPRET() to
1505 	 * restore the user-level priority.
1506 	 *
1507 	 * It is important that no locks (other than spinlocks) be entered
1508 	 * after this point before returning to user mode (unless lwp_state
1509 	 * is set back to LWP_SYS).
1510 	 */
1511 	lwp->lwp_state = LWP_USER;
1512 
1513 	if (ct->t_trapret) {
1514 		ct->t_trapret = 0;
1515 		thread_lock(ct);
1516 		CL_TRAPRET(ct);
1517 		thread_unlock(ct);
1518 	}
1519 	if (CPU->cpu_runrun || curthread->t_schedflag & TS_ANYWAITQ)
1520 		preempt();
1521 	prunstop();
1522 	(void) new_mstate(ct, mstate);
1523 
1524 	/* Kernel probe */
1525 	TNF_PROBE_1(thread_state, "thread", /* CSTYLED */,
1526 	    tnf_microstate, state, LMS_USER);
1527 
1528 	return;
1529 
1530 cleanup:	/* system traps end up here */
1531 	ASSERT(!(type & USER));
1532 }
1533 
1534 /*
1535  * Patch non-zero to disable preemption of threads in the kernel.
1536  */
1537 int IGNORE_KERNEL_PREEMPTION = 0;	/* XXX - delete this someday */
1538 
1539 struct kpreempt_cnts {		/* kernel preemption statistics */
1540 	int	kpc_idle;	/* executing idle thread */
1541 	int	kpc_intr;	/* executing interrupt thread */
1542 	int	kpc_clock;	/* executing clock thread */
1543 	int	kpc_blocked;	/* thread has blocked preemption (t_preempt) */
1544 	int	kpc_notonproc;	/* thread is surrendering processor */
1545 	int	kpc_inswtch;	/* thread has ratified scheduling decision */
1546 	int	kpc_prilevel;	/* processor interrupt level is too high */
1547 	int	kpc_apreempt;	/* asynchronous preemption */
1548 	int	kpc_spreempt;	/* synchronous preemption */
1549 } kpreempt_cnts;
1550 
1551 /*
1552  * kernel preemption: forced rescheduling, preempt the running kernel thread.
1553  *	the argument is old PIL for an interrupt,
1554  *	or the distingished value KPREEMPT_SYNC.
1555  */
1556 void
1557 kpreempt(int asyncspl)
1558 {
1559 	kthread_t *ct = curthread;
1560 
1561 	if (IGNORE_KERNEL_PREEMPTION) {
1562 		aston(CPU->cpu_dispthread);
1563 		return;
1564 	}
1565 
1566 	/*
1567 	 * Check that conditions are right for kernel preemption
1568 	 */
1569 	do {
1570 		if (ct->t_preempt) {
1571 			/*
1572 			 * either a privileged thread (idle, panic, interrupt)
1573 			 * or will check when t_preempt is lowered
1574 			 * We need to specifically handle the case where
1575 			 * the thread is in the middle of swtch (resume has
1576 			 * been called) and has its t_preempt set
1577 			 * [idle thread and a thread which is in kpreempt
1578 			 * already] and then a high priority thread is
1579 			 * available in the local dispatch queue.
1580 			 * In this case the resumed thread needs to take a
1581 			 * trap so that it can call kpreempt. We achieve
1582 			 * this by using siron().
1583 			 * How do we detect this condition:
1584 			 * idle thread is running and is in the midst of
1585 			 * resume: curthread->t_pri == -1 && CPU->dispthread
1586 			 * != CPU->thread
1587 			 * Need to ensure that this happens only at high pil
1588 			 * resume is called at high pil
1589 			 * Only in resume_from_idle is the pil changed.
1590 			 */
1591 			if (ct->t_pri < 0) {
1592 				kpreempt_cnts.kpc_idle++;
1593 				if (CPU->cpu_dispthread != CPU->cpu_thread)
1594 					siron();
1595 			} else if (ct->t_flag & T_INTR_THREAD) {
1596 				kpreempt_cnts.kpc_intr++;
1597 				if (ct->t_pil == CLOCK_LEVEL)
1598 					kpreempt_cnts.kpc_clock++;
1599 			} else {
1600 				kpreempt_cnts.kpc_blocked++;
1601 				if (CPU->cpu_dispthread != CPU->cpu_thread)
1602 					siron();
1603 			}
1604 			aston(CPU->cpu_dispthread);
1605 			return;
1606 		}
1607 		if (ct->t_state != TS_ONPROC ||
1608 		    ct->t_disp_queue != CPU->cpu_disp) {
1609 			/* this thread will be calling swtch() shortly */
1610 			kpreempt_cnts.kpc_notonproc++;
1611 			if (CPU->cpu_thread != CPU->cpu_dispthread) {
1612 				/* already in swtch(), force another */
1613 				kpreempt_cnts.kpc_inswtch++;
1614 				siron();
1615 			}
1616 			return;
1617 		}
1618 		if (getpil() >= DISP_LEVEL) {
1619 			/*
1620 			 * We can't preempt this thread if it is at
1621 			 * a PIL >= DISP_LEVEL since it may be holding
1622 			 * a spin lock (like sched_lock).
1623 			 */
1624 			siron();	/* check back later */
1625 			kpreempt_cnts.kpc_prilevel++;
1626 			return;
1627 		}
1628 		if (!interrupts_enabled()) {
1629 			/*
1630 			 * Can't preempt while running with ints disabled
1631 			 */
1632 			kpreempt_cnts.kpc_prilevel++;
1633 			return;
1634 		}
1635 		if (asyncspl != KPREEMPT_SYNC)
1636 			kpreempt_cnts.kpc_apreempt++;
1637 		else
1638 			kpreempt_cnts.kpc_spreempt++;
1639 
1640 		ct->t_preempt++;
1641 		preempt();
1642 		ct->t_preempt--;
1643 	} while (CPU->cpu_kprunrun);
1644 }
1645 
1646 /*
1647  * Print out debugging info.
1648  */
1649 static void
1650 showregs(uint_t type, struct regs *rp, caddr_t addr)
1651 {
1652 	int s;
1653 
1654 	s = spl7();
1655 	type &= ~USER;
1656 	if (PTOU(curproc)->u_comm[0])
1657 		printf("%s: ", PTOU(curproc)->u_comm);
1658 	if (type < TRAP_TYPES)
1659 		printf("#%s %s\n", trap_type_mnemonic[type], trap_type[type]);
1660 	else
1661 		switch (type) {
1662 		case T_SYSCALL:
1663 			printf("Syscall Trap:\n");
1664 			break;
1665 		case T_AST:
1666 			printf("AST\n");
1667 			break;
1668 		default:
1669 			printf("Bad Trap = %d\n", type);
1670 			break;
1671 		}
1672 	if (type == T_PGFLT) {
1673 		printf("Bad %s fault at addr=0x%lx\n",
1674 		    USERMODE(rp->r_cs) ? "user": "kernel", (uintptr_t)addr);
1675 	} else if (addr) {
1676 		printf("addr=0x%lx\n", (uintptr_t)addr);
1677 	}
1678 
1679 	printf("pid=%d, pc=0x%lx, sp=0x%lx, eflags=0x%lx\n",
1680 	    (ttoproc(curthread) && ttoproc(curthread)->p_pidp) ?
1681 	    ttoproc(curthread)->p_pid : 0, rp->r_pc, rp->r_sp, rp->r_ps);
1682 
1683 #if defined(__lint)
1684 	/*
1685 	 * this clause can be deleted when lint bug 4870403 is fixed
1686 	 * (lint thinks that bit 32 is illegal in a %b format string)
1687 	 */
1688 	printf("cr0: %x  cr4: %b\n",
1689 	    (uint_t)getcr0(), (uint_t)getcr4(), FMT_CR4);
1690 #else
1691 	printf("cr0: %b  cr4: %b\n",
1692 	    (uint_t)getcr0(), FMT_CR0, (uint_t)getcr4(), FMT_CR4);
1693 #endif	/* __lint */
1694 
1695 	printf("cr2: %lx  ", getcr2());
1696 #if !defined(__xpv)
1697 	printf("cr3: %lx  ", getcr3());
1698 #if defined(__amd64)
1699 	printf("cr8: %lx\n", getcr8());
1700 #endif
1701 #endif
1702 	printf("\n");
1703 
1704 	dumpregs(rp);
1705 	splx(s);
1706 }
1707 
1708 static void
1709 dumpregs(struct regs *rp)
1710 {
1711 #if defined(__amd64)
1712 	const char fmt[] = "\t%3s: %16lx %3s: %16lx %3s: %16lx\n";
1713 
1714 	printf(fmt, "rdi", rp->r_rdi, "rsi", rp->r_rsi, "rdx", rp->r_rdx);
1715 	printf(fmt, "rcx", rp->r_rcx, " r8", rp->r_r8, " r9", rp->r_r9);
1716 	printf(fmt, "rax", rp->r_rax, "rbx", rp->r_rbx, "rbp", rp->r_rbp);
1717 	printf(fmt, "r10", rp->r_r10, "r11", rp->r_r11, "r12", rp->r_r12);
1718 	printf(fmt, "r13", rp->r_r13, "r14", rp->r_r14, "r15", rp->r_r15);
1719 
1720 	printf(fmt, "fsb", rdmsr(MSR_AMD_FSBASE), "gsb", rdmsr(MSR_AMD_GSBASE),
1721 	    " ds", rp->r_ds);
1722 	printf(fmt, " es", rp->r_es, " fs", rp->r_fs, " gs", rp->r_gs);
1723 
1724 	printf(fmt, "trp", rp->r_trapno, "err", rp->r_err, "rip", rp->r_rip);
1725 	printf(fmt, " cs", rp->r_cs, "rfl", rp->r_rfl, "rsp", rp->r_rsp);
1726 
1727 	printf("\t%3s: %16lx\n", " ss", rp->r_ss);
1728 
1729 #elif defined(__i386)
1730 	const char fmt[] = "\t%3s: %8lx %3s: %8lx %3s: %8lx %3s: %8lx\n";
1731 
1732 	printf(fmt, " gs", rp->r_gs, " fs", rp->r_fs,
1733 	    " es", rp->r_es, " ds", rp->r_ds);
1734 	printf(fmt, "edi", rp->r_edi, "esi", rp->r_esi,
1735 	    "ebp", rp->r_ebp, "esp", rp->r_esp);
1736 	printf(fmt, "ebx", rp->r_ebx, "edx", rp->r_edx,
1737 	    "ecx", rp->r_ecx, "eax", rp->r_eax);
1738 	printf(fmt, "trp", rp->r_trapno, "err", rp->r_err,
1739 	    "eip", rp->r_eip, " cs", rp->r_cs);
1740 	printf("\t%3s: %8lx %3s: %8lx %3s: %8lx\n",
1741 	    "efl", rp->r_efl, "usp", rp->r_uesp, " ss", rp->r_ss);
1742 
1743 #endif	/* __i386 */
1744 }
1745 
1746 /*
1747  * Test to see if the instruction is iret on i386 or iretq on amd64.
1748  *
1749  * On the hypervisor we can only test for nopop_sys_rtt_syscall. If true
1750  * then we are in the context of hypervisor's failsafe handler because it
1751  * tried to iret and failed due to a bad selector. See xen_failsafe_callback.
1752  */
1753 static int
1754 instr_is_iret(caddr_t pc)
1755 {
1756 
1757 #if defined(__xpv)
1758 	extern void nopop_sys_rtt_syscall(void);
1759 	return ((pc == (caddr_t)nopop_sys_rtt_syscall) ? 1 : 0);
1760 
1761 #else
1762 
1763 #if defined(__amd64)
1764 	static const uint8_t iret_insn[2] = { 0x48, 0xcf };	/* iretq */
1765 
1766 #elif defined(__i386)
1767 	static const uint8_t iret_insn[1] = { 0xcf };		/* iret */
1768 #endif	/* __i386 */
1769 	return (bcmp(pc, iret_insn, sizeof (iret_insn)) == 0);
1770 
1771 #endif	/* __xpv */
1772 }
1773 
1774 #if defined(__i386)
1775 
1776 /*
1777  * Test to see if the instruction is part of __SEGREGS_POP
1778  *
1779  * Note carefully the appallingly awful dependency between
1780  * the instruction sequence used in __SEGREGS_POP and these
1781  * instructions encoded here.
1782  */
1783 static int
1784 instr_is_segregs_pop(caddr_t pc)
1785 {
1786 	static const uint8_t movw_0_esp_gs[4] = { 0x8e, 0x6c, 0x24, 0x0 };
1787 	static const uint8_t movw_4_esp_fs[4] = { 0x8e, 0x64, 0x24, 0x4 };
1788 	static const uint8_t movw_8_esp_es[4] = { 0x8e, 0x44, 0x24, 0x8 };
1789 	static const uint8_t movw_c_esp_ds[4] = { 0x8e, 0x5c, 0x24, 0xc };
1790 
1791 	if (bcmp(pc, movw_0_esp_gs, sizeof (movw_0_esp_gs)) == 0 ||
1792 	    bcmp(pc, movw_4_esp_fs, sizeof (movw_4_esp_fs)) == 0 ||
1793 	    bcmp(pc, movw_8_esp_es, sizeof (movw_8_esp_es)) == 0 ||
1794 	    bcmp(pc, movw_c_esp_ds, sizeof (movw_c_esp_ds)) == 0)
1795 		return (1);
1796 
1797 	return (0);
1798 }
1799 
1800 #endif	/* __i386 */
1801 
1802 /*
1803  * Test to see if the instruction is part of _sys_rtt (or the KPTI trampolines
1804  * which are used by _sys_rtt).
1805  *
1806  * Again on the hypervisor if we try to IRET to user land with a bad code
1807  * or stack selector we will get vectored through xen_failsafe_callback.
1808  * In which case we assume we got here via _sys_rtt since we only allow
1809  * IRET to user land to take place in _sys_rtt.
1810  */
1811 static int
1812 instr_is_sys_rtt(caddr_t pc)
1813 {
1814 	extern void _sys_rtt(), _sys_rtt_end();
1815 
1816 #if !defined(__xpv)
1817 	extern void tr_sysc_ret_start(), tr_sysc_ret_end();
1818 	extern void tr_intr_ret_start(), tr_intr_ret_end();
1819 
1820 	if ((uintptr_t)pc >= (uintptr_t)tr_sysc_ret_start &&
1821 	    (uintptr_t)pc <= (uintptr_t)tr_sysc_ret_end)
1822 		return (1);
1823 
1824 	if ((uintptr_t)pc >= (uintptr_t)tr_intr_ret_start &&
1825 	    (uintptr_t)pc <= (uintptr_t)tr_intr_ret_end)
1826 		return (1);
1827 #endif
1828 
1829 	if ((uintptr_t)pc < (uintptr_t)_sys_rtt ||
1830 	    (uintptr_t)pc > (uintptr_t)_sys_rtt_end)
1831 		return (0);
1832 
1833 	return (1);
1834 }
1835 
1836 /*
1837  * Handle #gp faults in kernel mode.
1838  *
1839  * One legitimate way this can happen is if we attempt to update segment
1840  * registers to naughty values on the way out of the kernel.
1841  *
1842  * This can happen in a couple of ways: someone - either accidentally or
1843  * on purpose - creates (setcontext(2), lwp_create(2)) or modifies
1844  * (signal(2)) a ucontext that contains silly segment register values.
1845  * Or someone - either accidentally or on purpose - modifies the prgregset_t
1846  * of a subject process via /proc to contain silly segment register values.
1847  *
1848  * (The unfortunate part is that we can end up discovering the bad segment
1849  * register value in the middle of an 'iret' after we've popped most of the
1850  * stack.  So it becomes quite difficult to associate an accurate ucontext
1851  * with the lwp, because the act of taking the #gp trap overwrites most of
1852  * what we were going to send the lwp.)
1853  *
1854  * OTOH if it turns out that's -not- the problem, and we're -not- an lwp
1855  * trying to return to user mode and we get a #gp fault, then we need
1856  * to die() -- which will happen if we return non-zero from this routine.
1857  */
1858 static int
1859 kern_gpfault(struct regs *rp)
1860 {
1861 	kthread_t *t = curthread;
1862 	proc_t *p = ttoproc(t);
1863 	klwp_t *lwp = ttolwp(t);
1864 	struct regs tmpregs, *trp = NULL;
1865 	caddr_t pc = (caddr_t)rp->r_pc;
1866 	int v;
1867 	uint32_t auditing = AU_AUDITING();
1868 
1869 	/*
1870 	 * if we're not an lwp, or in the case of running native the
1871 	 * pc range is outside _sys_rtt, then we should immediately
1872 	 * be die()ing horribly.
1873 	 */
1874 	if (lwp == NULL || !instr_is_sys_rtt(pc))
1875 		return (1);
1876 
1877 	/*
1878 	 * So at least we're in the right part of the kernel.
1879 	 *
1880 	 * Disassemble the instruction at the faulting pc.
1881 	 * Once we know what it is, we carefully reconstruct the stack
1882 	 * based on the order in which the stack is deconstructed in
1883 	 * _sys_rtt. Ew.
1884 	 */
1885 	if (instr_is_iret(pc)) {
1886 		/*
1887 		 * We took the #gp while trying to perform the IRET.
1888 		 * This means that either %cs or %ss are bad.
1889 		 * All we know for sure is that most of the general
1890 		 * registers have been restored, including the
1891 		 * segment registers, and all we have left on the
1892 		 * topmost part of the lwp's stack are the
1893 		 * registers that the iretq was unable to consume.
1894 		 *
1895 		 * All the rest of the state was crushed by the #gp
1896 		 * which pushed -its- registers atop our old save area
1897 		 * (because we had to decrement the stack pointer, sigh) so
1898 		 * all that we can try and do is to reconstruct the
1899 		 * crushed frame from the #gp trap frame itself.
1900 		 */
1901 		trp = &tmpregs;
1902 		trp->r_ss = lwptoregs(lwp)->r_ss;
1903 		trp->r_sp = lwptoregs(lwp)->r_sp;
1904 		trp->r_ps = lwptoregs(lwp)->r_ps;
1905 		trp->r_cs = lwptoregs(lwp)->r_cs;
1906 		trp->r_pc = lwptoregs(lwp)->r_pc;
1907 		bcopy(rp, trp, offsetof(struct regs, r_pc));
1908 
1909 		/*
1910 		 * Validate simple math
1911 		 */
1912 		ASSERT(trp->r_pc == lwptoregs(lwp)->r_pc);
1913 		ASSERT(trp->r_err == rp->r_err);
1914 
1915 
1916 
1917 	}
1918 
1919 #if defined(__amd64)
1920 	if (trp == NULL && PCB_NEED_UPDATE_SEGS(&lwp->lwp_pcb)) {
1921 
1922 		/*
1923 		 * This is the common case -- we're trying to load
1924 		 * a bad segment register value in the only section
1925 		 * of kernel code that ever loads segment registers.
1926 		 *
1927 		 * We don't need to do anything at this point because
1928 		 * the pcb contains all the pending segment register
1929 		 * state, and the regs are still intact because we
1930 		 * didn't adjust the stack pointer yet.  Given the fidelity
1931 		 * of all this, we could conceivably send a signal
1932 		 * to the lwp, rather than core-ing.
1933 		 */
1934 		trp = lwptoregs(lwp);
1935 		ASSERT((caddr_t)trp == (caddr_t)rp->r_sp);
1936 	}
1937 
1938 #elif defined(__i386)
1939 
1940 	if (trp == NULL && instr_is_segregs_pop(pc))
1941 		trp = lwptoregs(lwp);
1942 
1943 #endif	/* __i386 */
1944 
1945 	if (trp == NULL)
1946 		return (1);
1947 
1948 	/*
1949 	 * If we get to here, we're reasonably confident that we've
1950 	 * correctly decoded what happened on the way out of the kernel.
1951 	 * Rewrite the lwp's registers so that we can create a core dump
1952 	 * the (at least vaguely) represents the mcontext we were
1953 	 * being asked to restore when things went so terribly wrong.
1954 	 */
1955 
1956 	/*
1957 	 * Make sure that we have a meaningful %trapno and %err.
1958 	 */
1959 	trp->r_trapno = rp->r_trapno;
1960 	trp->r_err = rp->r_err;
1961 
1962 	if ((caddr_t)trp != (caddr_t)lwptoregs(lwp))
1963 		bcopy(trp, lwptoregs(lwp), sizeof (*trp));
1964 
1965 
1966 	mutex_enter(&p->p_lock);
1967 	lwp->lwp_cursig = SIGSEGV;
1968 	mutex_exit(&p->p_lock);
1969 
1970 	/*
1971 	 * Terminate all LWPs but don't discard them.  If another lwp beat
1972 	 * us to the punch by calling exit(), evaporate now.
1973 	 */
1974 	proc_is_exiting(p);
1975 	if (exitlwps(1) != 0) {
1976 		mutex_enter(&p->p_lock);
1977 		lwp_exit();
1978 	}
1979 
1980 	if (auditing)		/* audit core dump */
1981 		audit_core_start(SIGSEGV);
1982 	v = core(SIGSEGV, B_FALSE);
1983 	if (auditing)		/* audit core dump */
1984 		audit_core_finish(v ? CLD_KILLED : CLD_DUMPED);
1985 	exit(v ? CLD_KILLED : CLD_DUMPED, SIGSEGV);
1986 	return (0);
1987 }
1988 
1989 /*
1990  * dump_tss() - Display the TSS structure
1991  */
1992 
1993 #if !defined(__xpv)
1994 #if defined(__amd64)
1995 
1996 static void
1997 dump_tss(void)
1998 {
1999 	const char tss_fmt[] = "tss.%s:\t0x%p\n";  /* Format string */
2000 	tss_t *tss = CPU->cpu_tss;
2001 
2002 	printf(tss_fmt, "tss_rsp0", (void *)tss->tss_rsp0);
2003 	printf(tss_fmt, "tss_rsp1", (void *)tss->tss_rsp1);
2004 	printf(tss_fmt, "tss_rsp2", (void *)tss->tss_rsp2);
2005 
2006 	printf(tss_fmt, "tss_ist1", (void *)tss->tss_ist1);
2007 	printf(tss_fmt, "tss_ist2", (void *)tss->tss_ist2);
2008 	printf(tss_fmt, "tss_ist3", (void *)tss->tss_ist3);
2009 	printf(tss_fmt, "tss_ist4", (void *)tss->tss_ist4);
2010 	printf(tss_fmt, "tss_ist5", (void *)tss->tss_ist5);
2011 	printf(tss_fmt, "tss_ist6", (void *)tss->tss_ist6);
2012 	printf(tss_fmt, "tss_ist7", (void *)tss->tss_ist7);
2013 }
2014 
2015 #elif defined(__i386)
2016 
2017 static void
2018 dump_tss(void)
2019 {
2020 	const char tss_fmt[] = "tss.%s:\t0x%p\n";  /* Format string */
2021 	tss_t *tss = CPU->cpu_tss;
2022 
2023 	printf(tss_fmt, "tss_link", (void *)(uintptr_t)tss->tss_link);
2024 	printf(tss_fmt, "tss_esp0", (void *)(uintptr_t)tss->tss_esp0);
2025 	printf(tss_fmt, "tss_ss0", (void *)(uintptr_t)tss->tss_ss0);
2026 	printf(tss_fmt, "tss_esp1", (void *)(uintptr_t)tss->tss_esp1);
2027 	printf(tss_fmt, "tss_ss1", (void *)(uintptr_t)tss->tss_ss1);
2028 	printf(tss_fmt, "tss_esp2", (void *)(uintptr_t)tss->tss_esp2);
2029 	printf(tss_fmt, "tss_ss2", (void *)(uintptr_t)tss->tss_ss2);
2030 	printf(tss_fmt, "tss_cr3", (void *)(uintptr_t)tss->tss_cr3);
2031 	printf(tss_fmt, "tss_eip", (void *)(uintptr_t)tss->tss_eip);
2032 	printf(tss_fmt, "tss_eflags", (void *)(uintptr_t)tss->tss_eflags);
2033 	printf(tss_fmt, "tss_eax", (void *)(uintptr_t)tss->tss_eax);
2034 	printf(tss_fmt, "tss_ebx", (void *)(uintptr_t)tss->tss_ebx);
2035 	printf(tss_fmt, "tss_ecx", (void *)(uintptr_t)tss->tss_ecx);
2036 	printf(tss_fmt, "tss_edx", (void *)(uintptr_t)tss->tss_edx);
2037 	printf(tss_fmt, "tss_esp", (void *)(uintptr_t)tss->tss_esp);
2038 }
2039 
2040 #endif	/* __amd64 */
2041 #endif	/* !__xpv */
2042 
2043 #if defined(TRAPTRACE)
2044 
2045 int ttrace_nrec = 10;		/* number of records to dump out */
2046 int ttrace_dump_nregs = 0;	/* dump out this many records with regs too */
2047 
2048 /*
2049  * Dump out the last ttrace_nrec traptrace records on each CPU
2050  */
2051 static void
2052 dump_ttrace(void)
2053 {
2054 	trap_trace_ctl_t *ttc;
2055 	trap_trace_rec_t *rec;
2056 	uintptr_t current;
2057 	int i, j, k;
2058 	int n = NCPU;
2059 #if defined(__amd64)
2060 	const char banner[] =
2061 	    "CPU          ADDRESS    TIMESTAMP TYPE  VC HANDLER          PC\n";
2062 	/* Define format for the CPU, ADDRESS, and TIMESTAMP fields */
2063 	const char fmt1[] = "%3d %016lx %12llx";
2064 	char data1[34];	/* length of string formatted by fmt1 + 1 */
2065 #elif defined(__i386)
2066 	const char banner[] =
2067 	    "CPU  ADDRESS     TIMESTAMP TYPE  VC HANDLER          PC\n";
2068 	/* Define format for the CPU, ADDRESS, and TIMESTAMP fields */
2069 	const char fmt1[] = "%3d %08lx %12llx";
2070 	char data1[26];	/* length of string formatted by fmt1 + 1 */
2071 #endif
2072 	/* Define format for the TYPE and VC fields */
2073 	const char fmt2[] = "%4s %3x";
2074 	const char fmt2s[] = "%4s %3s";
2075 	char data2[9];	/* length of string formatted by fmt2 + 1 */
2076 	/*
2077 	 * Define format for the HANDLER field. Width is arbitrary, but should
2078 	 * be enough for common handler's names, and leave enough space for
2079 	 * the PC field, especially when we are in kmdb.
2080 	 */
2081 	const char fmt3h[] = "#%-15s";
2082 	const char fmt3p[] = "%-16p";
2083 	const char fmt3s[] = "%-16s";
2084 	char data3[17];	/* length of string formatted by fmt3* + 1 */
2085 
2086 	if (ttrace_nrec == 0)
2087 		return;
2088 
2089 	printf("\n");
2090 	printf(banner);
2091 
2092 	for (i = 0; i < n; i++) {
2093 		ttc = &trap_trace_ctl[i];
2094 		if (ttc->ttc_first == (uintptr_t)NULL)
2095 			continue;
2096 
2097 		current = ttc->ttc_next - sizeof (trap_trace_rec_t);
2098 		for (j = 0; j < ttrace_nrec; j++) {
2099 			struct sysent	*sys;
2100 			struct autovec	*vec;
2101 			extern struct av_head autovect[];
2102 			int type;
2103 			ulong_t	off;
2104 			char *sym, *stype;
2105 
2106 			if (current < ttc->ttc_first)
2107 				current =
2108 				    ttc->ttc_limit - sizeof (trap_trace_rec_t);
2109 
2110 			if (current == (uintptr_t)NULL)
2111 				continue;
2112 
2113 			rec = (trap_trace_rec_t *)current;
2114 
2115 			if (rec->ttr_stamp == 0)
2116 				break;
2117 
2118 			(void) snprintf(data1, sizeof (data1), fmt1, i,
2119 			    (uintptr_t)rec, rec->ttr_stamp);
2120 
2121 			switch (rec->ttr_marker) {
2122 			case TT_SYSCALL:
2123 			case TT_SYSENTER:
2124 			case TT_SYSC:
2125 			case TT_SYSC64:
2126 				sys = &sysent32[rec->ttr_sysnum];
2127 				switch (rec->ttr_marker) {
2128 				case TT_SYSC64:
2129 					sys = &sysent[rec->ttr_sysnum];
2130 					/* FALLTHROUGH */
2131 				case TT_SYSC:
2132 					stype = "sysc";	/* syscall */
2133 					break;
2134 				case TT_SYSCALL:
2135 					stype = "lcal";	/* lcall */
2136 					break;
2137 				case TT_SYSENTER:
2138 					stype = "syse";	/* sysenter */
2139 					break;
2140 				default:
2141 					break;
2142 				}
2143 				(void) snprintf(data2, sizeof (data2), fmt2,
2144 				    stype, rec->ttr_sysnum);
2145 				if (sys != NULL) {
2146 					sym = kobj_getsymname(
2147 					    (uintptr_t)sys->sy_callc,
2148 					    &off);
2149 					if (sym != NULL) {
2150 						(void) snprintf(data3,
2151 						    sizeof (data3), fmt3s, sym);
2152 					} else {
2153 						(void) snprintf(data3,
2154 						    sizeof (data3), fmt3p,
2155 						    sys->sy_callc);
2156 					}
2157 				} else {
2158 					(void) snprintf(data3, sizeof (data3),
2159 					    fmt3s, "unknown");
2160 				}
2161 				break;
2162 
2163 			case TT_INTERRUPT:
2164 				if (rec->ttr_regs.r_trapno == T_SOFTINT) {
2165 					(void) snprintf(data2, sizeof (data2),
2166 					    fmt2s, "intr", "-");
2167 					(void) snprintf(data3, sizeof (data3),
2168 					    fmt3s, "(fakesoftint)");
2169 					break;
2170 				}
2171 
2172 				(void) snprintf(data2, sizeof (data2), fmt2,
2173 				    "intr", rec->ttr_vector);
2174 				if (get_intr_handler != NULL)
2175 					vec = (struct autovec *)
2176 					    (*get_intr_handler)
2177 					    (rec->ttr_cpuid, rec->ttr_vector);
2178 				else
2179 					vec =
2180 					    autovect[rec->ttr_vector].avh_link;
2181 
2182 				if (vec != NULL) {
2183 					sym = kobj_getsymname(
2184 					    (uintptr_t)vec->av_vector, &off);
2185 					if (sym != NULL) {
2186 						(void) snprintf(data3,
2187 						    sizeof (data3), fmt3s, sym);
2188 					} else {
2189 						(void) snprintf(data3,
2190 						    sizeof (data3), fmt3p,
2191 						    vec->av_vector);
2192 					}
2193 				} else {
2194 					(void) snprintf(data3, sizeof (data3),
2195 					    fmt3s, "unknown");
2196 				}
2197 				break;
2198 
2199 			case TT_TRAP:
2200 			case TT_EVENT:
2201 				type = rec->ttr_regs.r_trapno;
2202 				(void) snprintf(data2, sizeof (data2), fmt2,
2203 				    "trap", type);
2204 				if (type < TRAP_TYPES) {
2205 					(void) snprintf(data3, sizeof (data3),
2206 					    fmt3h, trap_type_mnemonic[type]);
2207 				} else {
2208 					switch (type) {
2209 					case T_AST:
2210 						(void) snprintf(data3,
2211 						    sizeof (data3), fmt3s,
2212 						    "ast");
2213 						break;
2214 					default:
2215 						(void) snprintf(data3,
2216 						    sizeof (data3), fmt3s, "");
2217 						break;
2218 					}
2219 				}
2220 				break;
2221 
2222 			default:
2223 				break;
2224 			}
2225 
2226 			sym = kobj_getsymname(rec->ttr_regs.r_pc, &off);
2227 			if (sym != NULL) {
2228 				printf("%s %s %s %s+%lx\n", data1, data2, data3,
2229 				    sym, off);
2230 			} else {
2231 				printf("%s %s %s %lx\n", data1, data2, data3,
2232 				    rec->ttr_regs.r_pc);
2233 			}
2234 
2235 			if (ttrace_dump_nregs-- > 0) {
2236 				int s;
2237 
2238 				if (rec->ttr_marker == TT_INTERRUPT)
2239 					printf(
2240 					    "\t\tipl %x spl %x pri %x\n",
2241 					    rec->ttr_ipl,
2242 					    rec->ttr_spl,
2243 					    rec->ttr_pri);
2244 
2245 				dumpregs(&rec->ttr_regs);
2246 
2247 				printf("\t%3s: %p\n\n", " ct",
2248 				    (void *)rec->ttr_curthread);
2249 
2250 				/*
2251 				 * print out the pc stack that we recorded
2252 				 * at trap time (if any)
2253 				 */
2254 				for (s = 0; s < rec->ttr_sdepth; s++) {
2255 					uintptr_t fullpc;
2256 
2257 					if (s >= TTR_STACK_DEPTH) {
2258 						printf("ttr_sdepth corrupt\n");
2259 						break;
2260 					}
2261 
2262 					fullpc = (uintptr_t)rec->ttr_stack[s];
2263 
2264 					sym = kobj_getsymname(fullpc, &off);
2265 					if (sym != NULL)
2266 						printf("-> %s+0x%lx()\n",
2267 						    sym, off);
2268 					else
2269 						printf("-> 0x%lx()\n", fullpc);
2270 				}
2271 				printf("\n");
2272 			}
2273 			current -= sizeof (trap_trace_rec_t);
2274 		}
2275 	}
2276 }
2277 
2278 #endif	/* TRAPTRACE */
2279 
2280 void
2281 panic_showtrap(struct panic_trap_info *tip)
2282 {
2283 	showregs(tip->trap_type, tip->trap_regs, tip->trap_addr);
2284 
2285 #if defined(TRAPTRACE)
2286 	dump_ttrace();
2287 #endif
2288 
2289 #if !defined(__xpv)
2290 	if (tip->trap_type == T_DBLFLT)
2291 		dump_tss();
2292 #endif
2293 }
2294 
2295 void
2296 panic_savetrap(panic_data_t *pdp, struct panic_trap_info *tip)
2297 {
2298 	panic_saveregs(pdp, tip->trap_regs);
2299 }
2300