xref: /dragonfly/sys/platform/vkernel64/x86_64/trap.c (revision edf2e657)
1 /*-
2  * Copyright (C) 1994, David Greenman
3  * Copyright (c) 1990, 1993
4  *	The Regents of the University of California.  All rights reserved.
5  *
6  * This code is derived from software contributed to Berkeley by
7  * the University of Utah, and William Jolitz.
8  *
9  * Redistribution and use in source and binary forms, with or without
10  * modification, are permitted provided that the following conditions
11  * are met:
12  * 1. Redistributions of source code must retain the above copyright
13  *    notice, this list of conditions and the following disclaimer.
14  * 2. Redistributions in binary form must reproduce the above copyright
15  *    notice, this list of conditions and the following disclaimer in the
16  *    documentation and/or other materials provided with the distribution.
17  * 3. All advertising materials mentioning features or use of this software
18  *    must display the following acknowledgement:
19  *	This product includes software developed by the University of
20  *	California, Berkeley and its contributors.
21  * 4. Neither the name of the University nor the names of its contributors
22  *    may be used to endorse or promote products derived from this software
23  *    without specific prior written permission.
24  *
25  * THIS SOFTWARE IS PROVIDED BY THE REGENTS AND CONTRIBUTORS ``AS IS'' AND
26  * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
27  * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE
28  * ARE DISCLAIMED.  IN NO EVENT SHALL THE REGENTS OR CONTRIBUTORS BE LIABLE
29  * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL
30  * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS
31  * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION)
32  * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT
33  * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY
34  * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF
35  * SUCH DAMAGE.
36  *
37  *	from: @(#)trap.c	7.4 (Berkeley) 5/13/91
38  * $FreeBSD: src/sys/i386/i386/trap.c,v 1.147.2.11 2003/02/27 19:09:59 luoqi Exp $
39  */
40 
41 /*
42  * x86_64 Trap and System call handling
43  */
44 
45 #include "use_isa.h"
46 
47 #include "opt_ddb.h"
48 #include "opt_ktrace.h"
49 
50 #include <sys/param.h>
51 #include <sys/systm.h>
52 #include <sys/proc.h>
53 #include <sys/pioctl.h>
54 #include <sys/kernel.h>
55 #include <sys/resourcevar.h>
56 #include <sys/signalvar.h>
57 #include <sys/signal2.h>
58 #include <sys/syscall.h>
59 #include <sys/sysctl.h>
60 #include <sys/sysent.h>
61 #include <sys/uio.h>
62 #include <sys/vmmeter.h>
63 #include <sys/malloc.h>
64 #ifdef KTRACE
65 #include <sys/ktrace.h>
66 #endif
67 #include <sys/ktr.h>
68 #include <sys/vkernel.h>
69 #include <sys/sysproto.h>
70 #include <sys/sysunion.h>
71 #include <sys/vmspace.h>
72 
73 #include <vm/vm.h>
74 #include <vm/vm_param.h>
75 #include <sys/lock.h>
76 #include <vm/pmap.h>
77 #include <vm/vm_kern.h>
78 #include <vm/vm_map.h>
79 #include <vm/vm_page.h>
80 #include <vm/vm_extern.h>
81 
82 #include <machine/cpu.h>
83 #include <machine/md_var.h>
84 #include <machine/pcb.h>
85 #include <machine/smp.h>
86 #include <machine/tss.h>
87 #include <machine/globaldata.h>
88 
89 #include <ddb/ddb.h>
90 
91 #include <sys/msgport2.h>
92 #include <sys/thread2.h>
93 #include <sys/mplock2.h>
94 
95 int (*pmath_emulate) (struct trapframe *);
96 
97 static int trap_pfault (struct trapframe *, int, vm_offset_t);
98 static void trap_fatal (struct trapframe *, int, vm_offset_t);
99 void dblfault_handler (void);
100 extern int vmm_enabled;
101 
102 static struct krate segfltrate = { 1 };
103 
104 #if 0
105 extern inthand_t IDTVEC(syscall);
106 #endif
107 
108 #define MAX_TRAP_MSG		30
109 static char *trap_msg[] = {
110 	"",					/*  0 unused */
111 	"privileged instruction fault",		/*  1 T_PRIVINFLT */
112 	"",					/*  2 unused */
113 	"breakpoint instruction fault",		/*  3 T_BPTFLT */
114 	"",					/*  4 unused */
115 	"",					/*  5 unused */
116 	"arithmetic trap",			/*  6 T_ARITHTRAP */
117 	"system forced exception",		/*  7 T_ASTFLT */
118 	"",					/*  8 unused */
119 	"general protection fault",		/*  9 T_PROTFLT */
120 	"trace trap",				/* 10 T_TRCTRAP */
121 	"",					/* 11 unused */
122 	"page fault",				/* 12 T_PAGEFLT */
123 	"",					/* 13 unused */
124 	"alignment fault",			/* 14 T_ALIGNFLT */
125 	"",					/* 15 unused */
126 	"",					/* 16 unused */
127 	"",					/* 17 unused */
128 	"integer divide fault",			/* 18 T_DIVIDE */
129 	"non-maskable interrupt trap",		/* 19 T_NMI */
130 	"overflow trap",			/* 20 T_OFLOW */
131 	"FPU bounds check fault",		/* 21 T_BOUND */
132 	"FPU device not available",		/* 22 T_DNA */
133 	"double fault",				/* 23 T_DOUBLEFLT */
134 	"FPU operand fetch fault",		/* 24 T_FPOPFLT */
135 	"invalid TSS fault",			/* 25 T_TSSFLT */
136 	"segment not present fault",		/* 26 T_SEGNPFLT */
137 	"stack fault",				/* 27 T_STKFLT */
138 	"machine check trap",			/* 28 T_MCHK */
139 	"SIMD floating-point exception",	/* 29 T_XMMFLT */
140 	"reserved (unknown) fault",		/* 30 T_RESERVED */
141 };
142 
143 #ifdef DDB
144 static int ddb_on_nmi = 1;
145 SYSCTL_INT(_machdep, OID_AUTO, ddb_on_nmi, CTLFLAG_RW,
146 	&ddb_on_nmi, 0, "Go to DDB on NMI");
147 #endif
148 static int panic_on_nmi = 1;
149 SYSCTL_INT(_machdep, OID_AUTO, panic_on_nmi, CTLFLAG_RW,
150 	&panic_on_nmi, 0, "Panic on NMI");
151 static int fast_release;
152 SYSCTL_INT(_machdep, OID_AUTO, fast_release, CTLFLAG_RW,
153 	&fast_release, 0, "Passive Release was optimal");
154 static int slow_release;
155 SYSCTL_INT(_machdep, OID_AUTO, slow_release, CTLFLAG_RW,
156 	&slow_release, 0, "Passive Release was nonoptimal");
157 
158 /*
159  * Passively intercepts the thread switch function to increase
160  * the thread priority from a user priority to a kernel priority, reducing
161  * syscall and trap overhead for the case where no switch occurs.
162  *
163  * Synchronizes td_ucred with p_ucred.  This is used by system calls,
164  * signal handling, faults, AST traps, and anything else that enters the
165  * kernel from userland and provides the kernel with a stable read-only
166  * copy of the process ucred.
167  */
168 static __inline void
169 userenter(struct thread *curtd, struct proc *curp)
170 {
171 	struct ucred *ocred;
172 	struct ucred *ncred;
173 
174 	curtd->td_release = lwkt_passive_release;
175 
176 	if (curtd->td_ucred != curp->p_ucred) {
177 		ncred = crhold(curp->p_ucred);
178 		ocred = curtd->td_ucred;
179 		curtd->td_ucred = ncred;
180 		if (ocred)
181 			crfree(ocred);
182 	}
183 }
184 
185 /*
186  * Handle signals, profiling, and other AST's and/or tasks that
187  * must be completed before we can return to or try to return to userland.
188  *
189  * Note that td_sticks is a 64 bit quantity, but there's no point doing 64
190  * arithmatic on the delta calculation so the absolute tick values are
191  * truncated to an integer.
192  */
193 static void
194 userret(struct lwp *lp, struct trapframe *frame, int sticks)
195 {
196 	struct proc *p = lp->lwp_proc;
197 	int sig;
198 	int ptok;
199 
200 	/*
201 	 * Charge system time if profiling.  Note: times are in microseconds.
202 	 * This may do a copyout and block, so do it first even though it
203 	 * means some system time will be charged as user time.
204 	 */
205 	if (p->p_flags & P_PROFIL) {
206 		addupc_task(p, frame->tf_rip,
207 			(u_int)((int)lp->lwp_thread->td_sticks - sticks));
208 	}
209 
210 recheck:
211 	/*
212 	 * Specific on-return-to-usermode checks (LWP_MP_WEXIT,
213 	 * LWP_MP_VNLRU, etc).
214 	 */
215 	if (lp->lwp_mpflags & LWP_MP_URETMASK)
216 		lwpuserret(lp);
217 
218 	/*
219 	 * Block here if we are in a stopped state.
220 	 */
221 	if (STOPLWP(p, lp)) {
222 		lwkt_gettoken(&p->p_token);
223 		tstop();
224 		lwkt_reltoken(&p->p_token);
225 		goto recheck;
226 	}
227 
228 	/*
229 	 * Post any pending upcalls.  If running a virtual kernel be sure
230 	 * to restore the virtual kernel's vmspace before posting the upcall.
231 	 */
232 	if (p->p_flags & (P_SIGVTALRM | P_SIGPROF)) {
233 		lwkt_gettoken(&p->p_token);
234 		if (p->p_flags & P_SIGVTALRM) {
235 			p->p_flags &= ~P_SIGVTALRM;
236 			ksignal(p, SIGVTALRM);
237 		}
238 		if (p->p_flags & P_SIGPROF) {
239 			p->p_flags &= ~P_SIGPROF;
240 			ksignal(p, SIGPROF);
241 		}
242 		lwkt_reltoken(&p->p_token);
243 		goto recheck;
244 	}
245 
246 	/*
247 	 * Post any pending signals
248 	 *
249 	 * WARNING!  postsig() can exit and not return.
250 	 */
251 	if ((sig = CURSIG_LCK_TRACE(lp, &ptok)) != 0) {
252 		postsig(sig, ptok);
253 		goto recheck;
254 	}
255 
256 	/*
257 	 * block here if we are swapped out, but still process signals
258 	 * (such as SIGKILL).  proc0 (the swapin scheduler) is already
259 	 * aware of our situation, we do not have to wake it up.
260 	 */
261 	if (p->p_flags & P_SWAPPEDOUT) {
262 		lwkt_gettoken(&p->p_token);
263 		get_mplock();
264 		p->p_flags |= P_SWAPWAIT;
265 		swapin_request();
266 		if (p->p_flags & P_SWAPWAIT)
267 			tsleep(p, PCATCH, "SWOUT", 0);
268 		p->p_flags &= ~P_SWAPWAIT;
269 		rel_mplock();
270 		lwkt_reltoken(&p->p_token);
271 		goto recheck;
272 	}
273 
274 	/*
275 	 * In a multi-threaded program it is possible for a thread to change
276 	 * signal state during a system call which temporarily changes the
277 	 * signal mask.  In this case postsig() might not be run and we
278 	 * have to restore the mask ourselves.
279 	 */
280 	if (lp->lwp_flags & LWP_OLDMASK) {
281 		lp->lwp_flags &= ~LWP_OLDMASK;
282 		lp->lwp_sigmask = lp->lwp_oldsigmask;
283 		goto recheck;
284 	}
285 }
286 
287 /*
288  * Cleanup from userenter and any passive release that might have occured.
289  * We must reclaim the current-process designation before we can return
290  * to usermode.  We also handle both LWKT and USER reschedule requests.
291  */
292 static __inline void
293 userexit(struct lwp *lp)
294 {
295 	struct thread *td = lp->lwp_thread;
296 	/* globaldata_t gd = td->td_gd; */
297 
298 	/*
299 	 * Handle stop requests at kernel priority.  Any requests queued
300 	 * after this loop will generate another AST.
301 	 */
302 	while (STOPLWP(lp->lwp_proc, lp)) {
303 		lwkt_gettoken(&lp->lwp_proc->p_token);
304 		tstop();
305 		lwkt_reltoken(&lp->lwp_proc->p_token);
306 	}
307 
308 	/*
309 	 * Reduce our priority in preparation for a return to userland.  If
310 	 * our passive release function was still in place, our priority was
311 	 * never raised and does not need to be reduced.
312 	 */
313 	lwkt_passive_recover(td);
314 
315 	/*
316 	 * Become the current user scheduled process if we aren't already,
317 	 * and deal with reschedule requests and other factors.
318 	 */
319 	lp->lwp_proc->p_usched->acquire_curproc(lp);
320 	/* WARNING: we may have migrated cpu's */
321 	/* gd = td->td_gd; */
322 }
323 
324 #if !defined(KTR_KERNENTRY)
325 #define	KTR_KERNENTRY	KTR_ALL
326 #endif
327 KTR_INFO_MASTER(kernentry);
328 KTR_INFO(KTR_KERNENTRY, kernentry, trap, 0,
329 	 "TRAP(pid %hd, tid %hd, trapno %ld, eva %lu)",
330 	 pid_t pid, lwpid_t tid,  register_t trapno, vm_offset_t eva);
331 KTR_INFO(KTR_KERNENTRY, kernentry, trap_ret, 0, "TRAP_RET(pid %hd, tid %hd)",
332 	 pid_t pid, lwpid_t tid);
333 KTR_INFO(KTR_KERNENTRY, kernentry, syscall, 0, "SYSC(pid %hd, tid %hd, nr %ld)",
334 	 pid_t pid, lwpid_t tid,  register_t trapno);
335 KTR_INFO(KTR_KERNENTRY, kernentry, syscall_ret, 0, "SYSRET(pid %hd, tid %hd, err %d)",
336 	 pid_t pid, lwpid_t tid,  int err);
337 KTR_INFO(KTR_KERNENTRY, kernentry, fork_ret, 0, "FORKRET(pid %hd, tid %hd)",
338 	 pid_t pid, lwpid_t tid);
339 
340 /*
341  * Exception, fault, and trap interface to the kernel.
342  * This common code is called from assembly language IDT gate entry
343  * routines that prepare a suitable stack frame, and restore this
344  * frame after the exception has been processed.
345  *
346  * This function is also called from doreti in an interlock to handle ASTs.
347  * For example:  hardwareint->INTROUTINE->(set ast)->doreti->trap
348  *
349  * NOTE!  We have to retrieve the fault address prior to obtaining the
350  * MP lock because get_mplock() may switch out.  YYY cr2 really ought
351  * to be retrieved by the assembly code, not here.
352  *
353  * XXX gd_trap_nesting_level currently prevents lwkt_switch() from panicing
354  * if an attempt is made to switch from a fast interrupt or IPI.  This is
355  * necessary to properly take fatal kernel traps on SMP machines if
356  * get_mplock() has to block.
357  */
358 
359 void
360 user_trap(struct trapframe *frame)
361 {
362 	struct globaldata *gd = mycpu;
363 	struct thread *td = gd->gd_curthread;
364 	struct lwp *lp = td->td_lwp;
365 	struct proc *p;
366 	int sticks = 0;
367 	int i = 0, ucode = 0, type, code;
368 #ifdef INVARIANTS
369 	int crit_count = td->td_critcount;
370 	lwkt_tokref_t curstop = td->td_toks_stop;
371 #endif
372 	vm_offset_t eva;
373 
374 	p = td->td_proc;
375 
376 	if (frame->tf_trapno == T_PAGEFLT)
377 		eva = frame->tf_addr;
378 	else
379 		eva = 0;
380 #if 0
381 	kprintf("USER_TRAP AT %08lx xflags %ld trapno %ld eva %08lx\n",
382 		frame->tf_rip, frame->tf_xflags, frame->tf_trapno, eva);
383 #endif
384 
385 	/*
386 	 * Everything coming from user mode runs through user_trap,
387 	 * including system calls.
388 	 */
389 	if (frame->tf_trapno == T_FAST_SYSCALL) {
390 		syscall2(frame);
391 		return;
392 	}
393 
394 	KTR_LOG(kernentry_trap, lp->lwp_proc->p_pid, lp->lwp_tid,
395 		frame->tf_trapno, eva);
396 
397 #ifdef DDB
398 	if (db_active) {
399 		eva = (frame->tf_trapno == T_PAGEFLT ? rcr2() : 0);
400 		++gd->gd_trap_nesting_level;
401 		trap_fatal(frame, TRUE, eva);
402 		--gd->gd_trap_nesting_level;
403 		goto out2;
404 	}
405 #endif
406 
407 	type = frame->tf_trapno;
408 	code = frame->tf_err;
409 
410 	userenter(td, p);
411 
412 	sticks = (int)td->td_sticks;
413 	lp->lwp_md.md_regs = frame;
414 
415 	switch (type) {
416 	case T_PRIVINFLT:	/* privileged instruction fault */
417 		i = SIGILL;
418 		ucode = ILL_PRVOPC;
419 		break;
420 
421 	case T_BPTFLT:		/* bpt instruction fault */
422 	case T_TRCTRAP:		/* trace trap */
423 		frame->tf_rflags &= ~PSL_T;
424 		i = SIGTRAP;
425 		ucode = (type == T_TRCTRAP ? TRAP_TRACE : TRAP_BRKPT);
426 		break;
427 
428 	case T_ARITHTRAP:	/* arithmetic trap */
429 		ucode = code;
430 		i = SIGFPE;
431 		break;
432 
433 	case T_ASTFLT:		/* Allow process switch */
434 		mycpu->gd_cnt.v_soft++;
435 		if (mycpu->gd_reqflags & RQF_AST_OWEUPC) {
436 			atomic_clear_int(&mycpu->gd_reqflags, RQF_AST_OWEUPC);
437 			addupc_task(p, p->p_prof.pr_addr, p->p_prof.pr_ticks);
438 		}
439 		goto out;
440 
441 		/*
442 		 * The following two traps can happen in
443 		 * vm86 mode, and, if so, we want to handle
444 		 * them specially.
445 		 */
446 	case T_PROTFLT:		/* general protection fault */
447 	case T_STKFLT:		/* stack fault */
448 #if 0
449 		if (frame->tf_eflags & PSL_VM) {
450 			i = vm86_emulate((struct vm86frame *)frame);
451 			if (i == 0)
452 				goto out;
453 			break;
454 		}
455 #endif
456 		/* FALL THROUGH */
457 
458 	case T_SEGNPFLT:	/* segment not present fault */
459 	case T_TSSFLT:		/* invalid TSS fault */
460 	case T_DOUBLEFLT:	/* double fault */
461 	default:
462 		i = SIGBUS;
463 		ucode = code + BUS_SEGM_FAULT ;
464 		break;
465 
466 	case T_PAGEFLT:		/* page fault */
467 		i = trap_pfault(frame, TRUE, eva);
468 		if (i == -1 || i == 0)
469 			goto out;
470 
471 
472 		if (i == SIGSEGV)
473 			ucode = SEGV_MAPERR;
474 		else {
475 			i = SIGSEGV;
476 			ucode = SEGV_ACCERR;
477 		}
478 		break;
479 
480 	case T_DIVIDE:		/* integer divide fault */
481 		ucode = FPE_INTDIV;
482 		i = SIGFPE;
483 		break;
484 
485 #if NISA > 0
486 	case T_NMI:
487 		/* machine/parity/power fail/"kitchen sink" faults */
488 		if (isa_nmi(code) == 0) {
489 #ifdef DDB
490 			/*
491 			 * NMI can be hooked up to a pushbutton
492 			 * for debugging.
493 			 */
494 			if (ddb_on_nmi) {
495 				kprintf ("NMI ... going to debugger\n");
496 				kdb_trap(type, 0, frame);
497 			}
498 #endif /* DDB */
499 			goto out2;
500 		} else if (panic_on_nmi)
501 			panic("NMI indicates hardware failure");
502 		break;
503 #endif /* NISA > 0 */
504 
505 	case T_OFLOW:		/* integer overflow fault */
506 		ucode = FPE_INTOVF;
507 		i = SIGFPE;
508 		break;
509 
510 	case T_BOUND:		/* bounds check fault */
511 		ucode = FPE_FLTSUB;
512 		i = SIGFPE;
513 		break;
514 
515 	case T_DNA:
516 		/*
517 		 * Virtual kernel intercept - pass the DNA exception
518 		 * to the (emulated) virtual kernel if it asked to handle
519 		 * it.  This occurs when the virtual kernel is holding
520 		 * onto the FP context for a different emulated
521 		 * process then the one currently running.
522 		 *
523 		 * We must still call npxdna() since we may have
524 		 * saved FP state that the (emulated) virtual kernel
525 		 * needs to hand over to a different emulated process.
526 		 */
527 		if (lp->lwp_vkernel && lp->lwp_vkernel->ve &&
528 		    (td->td_pcb->pcb_flags & FP_VIRTFP)
529 		) {
530 			npxdna(frame);
531 			break;
532 		}
533 
534 		/*
535 		 * The kernel may have switched out the FP unit's
536 		 * state, causing the user process to take a fault
537 		 * when it tries to use the FP unit.  Restore the
538 		 * state here
539 		 */
540 		if (npxdna(frame))
541 			goto out;
542 		if (!pmath_emulate) {
543 			i = SIGFPE;
544 			ucode = FPE_FPU_NP_TRAP;
545 			break;
546 		}
547 		i = (*pmath_emulate)(frame);
548 		if (i == 0) {
549 			if (!(frame->tf_rflags & PSL_T))
550 				goto out2;
551 			frame->tf_rflags &= ~PSL_T;
552 			i = SIGTRAP;
553 		}
554 		/* else ucode = emulator_only_knows() XXX */
555 		break;
556 
557 	case T_FPOPFLT:		/* FPU operand fetch fault */
558 		ucode = T_FPOPFLT;
559 		i = SIGILL;
560 		break;
561 
562 	case T_XMMFLT:		/* SIMD floating-point exception */
563 		ucode = 0; /* XXX */
564 		i = SIGFPE;
565 		break;
566 	}
567 
568 	/*
569 	 * Virtual kernel intercept - if the fault is directly related to a
570 	 * VM context managed by a virtual kernel then let the virtual kernel
571 	 * handle it.
572 	 */
573 	if (lp->lwp_vkernel && lp->lwp_vkernel->ve) {
574 		vkernel_trap(lp, frame);
575 		goto out;
576 	}
577 
578 	/*
579 	 * Translate fault for emulators (e.g. Linux)
580 	 */
581 	if (*p->p_sysent->sv_transtrap)
582 		i = (*p->p_sysent->sv_transtrap)(i, type);
583 
584 	trapsignal(lp, i, ucode);
585 
586 #ifdef DEBUG
587 	if (type <= MAX_TRAP_MSG) {
588 		uprintf("fatal process exception: %s",
589 			trap_msg[type]);
590 		if ((type == T_PAGEFLT) || (type == T_PROTFLT))
591 			uprintf(", fault VA = 0x%lx", (u_long)eva);
592 		uprintf("\n");
593 	}
594 #endif
595 
596 out:
597 	userret(lp, frame, sticks);
598 	userexit(lp);
599 out2:	;
600 	KTR_LOG(kernentry_trap_ret, lp->lwp_proc->p_pid, lp->lwp_tid);
601 #ifdef INVARIANTS
602 	KASSERT(crit_count == td->td_critcount,
603 		("trap: critical section count mismatch! %d/%d",
604 		crit_count, td->td_pri));
605 	KASSERT(curstop == td->td_toks_stop,
606 		("trap: extra tokens held after trap! %ld/%ld",
607 		curstop - &td->td_toks_base,
608 		td->td_toks_stop - &td->td_toks_base));
609 #endif
610 }
611 
612 void
613 kern_trap(struct trapframe *frame)
614 {
615 	struct globaldata *gd = mycpu;
616 	struct thread *td = gd->gd_curthread;
617 	struct lwp *lp;
618 	struct proc *p;
619 	int i = 0, ucode = 0, type, code;
620 #ifdef INVARIANTS
621 	int crit_count = td->td_critcount;
622 	lwkt_tokref_t curstop = td->td_toks_stop;
623 #endif
624 	vm_offset_t eva;
625 
626 	lp = td->td_lwp;
627 	p = td->td_proc;
628 
629 	if (frame->tf_trapno == T_PAGEFLT)
630 		eva = frame->tf_addr;
631 	else
632 		eva = 0;
633 
634 #ifdef DDB
635 	if (db_active) {
636 		++gd->gd_trap_nesting_level;
637 		trap_fatal(frame, FALSE, eva);
638 		--gd->gd_trap_nesting_level;
639 		goto out2;
640 	}
641 #endif
642 
643 	type = frame->tf_trapno;
644 	code = frame->tf_err;
645 
646 #if 0
647 kernel_trap:
648 #endif
649 	/* kernel trap */
650 
651 	switch (type) {
652 	case T_PAGEFLT:			/* page fault */
653 		trap_pfault(frame, FALSE, eva);
654 		goto out2;
655 
656 	case T_DNA:
657 		/*
658 		 * The kernel may be using npx for copying or other
659 		 * purposes.
660 		 */
661 		panic("kernel NPX should not happen");
662 		if (npxdna(frame))
663 			goto out2;
664 		break;
665 
666 	case T_PROTFLT:		/* general protection fault */
667 	case T_SEGNPFLT:	/* segment not present fault */
668 		/*
669 		 * Invalid segment selectors and out of bounds
670 		 * %eip's and %esp's can be set up in user mode.
671 		 * This causes a fault in kernel mode when the
672 		 * kernel tries to return to user mode.  We want
673 		 * to get this fault so that we can fix the
674 		 * problem here and not have to check all the
675 		 * selectors and pointers when the user changes
676 		 * them.
677 		 */
678 		if (mycpu->gd_intr_nesting_level == 0) {
679 			if (td->td_pcb->pcb_onfault) {
680 				frame->tf_rip =
681 				    (register_t)td->td_pcb->pcb_onfault;
682 				goto out2;
683 			}
684 		}
685 		break;
686 
687 	case T_TSSFLT:
688 		/*
689 		 * PSL_NT can be set in user mode and isn't cleared
690 		 * automatically when the kernel is entered.  This
691 		 * causes a TSS fault when the kernel attempts to
692 		 * `iret' because the TSS link is uninitialized.  We
693 		 * want to get this fault so that we can fix the
694 		 * problem here and not every time the kernel is
695 		 * entered.
696 		 */
697 		if (frame->tf_rflags & PSL_NT) {
698 			frame->tf_rflags &= ~PSL_NT;
699 			goto out2;
700 		}
701 		break;
702 
703 	case T_TRCTRAP:	 /* trace trap */
704 #if 0
705 		if (frame->tf_eip == (int)IDTVEC(syscall)) {
706 			/*
707 			 * We've just entered system mode via the
708 			 * syscall lcall.  Continue single stepping
709 			 * silently until the syscall handler has
710 			 * saved the flags.
711 			 */
712 			goto out2;
713 		}
714 		if (frame->tf_eip == (int)IDTVEC(syscall) + 1) {
715 			/*
716 			 * The syscall handler has now saved the
717 			 * flags.  Stop single stepping it.
718 			 */
719 			frame->tf_eflags &= ~PSL_T;
720 			goto out2;
721 		}
722 #endif
723 #if 0
724 		/*
725 		 * Ignore debug register trace traps due to
726 		 * accesses in the user's address space, which
727 		 * can happen under several conditions such as
728 		 * if a user sets a watchpoint on a buffer and
729 		 * then passes that buffer to a system call.
730 		 * We still want to get TRCTRAPS for addresses
731 		 * in kernel space because that is useful when
732 		 * debugging the kernel.
733 		 */
734 		if (user_dbreg_trap()) {
735 			/*
736 			 * Reset breakpoint bits because the
737 			 * processor doesn't
738 			 */
739 			load_dr6(rdr6() & 0xfffffff0);
740 			goto out2;
741 		}
742 #endif
743 		/*
744 		 * Fall through (TRCTRAP kernel mode, kernel address)
745 		 */
746 	case T_BPTFLT:
747 		/*
748 		 * If DDB is enabled, let it handle the debugger trap.
749 		 * Otherwise, debugger traps "can't happen".
750 		 */
751 #ifdef DDB
752 		if (kdb_trap (type, 0, frame))
753 			goto out2;
754 #endif
755 		break;
756 	case T_DIVIDE:
757 		trap_fatal(frame, FALSE, eva);
758 		goto out2;
759 	case T_NMI:
760 		trap_fatal(frame, FALSE, eva);
761 		goto out2;
762 	case T_SYSCALL80:
763 	case T_FAST_SYSCALL:
764 		/*
765 		 * Ignore this trap generated from a spurious SIGTRAP.
766 		 *
767 		 * single stepping in / syscalls leads to spurious / SIGTRAP
768 		 * so ignore
769 		 *
770 		 * Haiku (c) 2007 Simon 'corecode' Schubert
771 		 */
772 		goto out2;
773 	}
774 
775 	/*
776 	 * Translate fault for emulators (e.g. Linux)
777 	 */
778 	if (*p->p_sysent->sv_transtrap)
779 		i = (*p->p_sysent->sv_transtrap)(i, type);
780 
781 	trapsignal(lp, i, ucode);
782 
783 #ifdef DEBUG
784 	if (type <= MAX_TRAP_MSG) {
785 		uprintf("fatal process exception: %s",
786 			trap_msg[type]);
787 		if ((type == T_PAGEFLT) || (type == T_PROTFLT))
788 			uprintf(", fault VA = 0x%lx", (u_long)eva);
789 		uprintf("\n");
790 	}
791 #endif
792 
793 out2:
794 	;
795 #ifdef INVARIANTS
796 	KASSERT(crit_count == td->td_critcount,
797 		("trap: critical section count mismatch! %d/%d",
798 		crit_count, td->td_pri));
799 	KASSERT(curstop == td->td_toks_stop,
800 		("trap: extra tokens held after trap! %ld/%ld",
801 		curstop - &td->td_toks_base,
802 		td->td_toks_stop - &td->td_toks_base));
803 #endif
804 }
805 
806 int
807 trap_pfault(struct trapframe *frame, int usermode, vm_offset_t eva)
808 {
809 	vm_offset_t va;
810 	struct vmspace *vm = NULL;
811 	vm_map_t map = 0;
812 	int rv = 0;
813 	vm_prot_t ftype;
814 	thread_t td = curthread;
815 	struct lwp *lp = td->td_lwp;
816 	int fault_flags;
817 
818 	va = trunc_page(eva);
819 	if (usermode == FALSE) {
820 		/*
821 		 * This is a fault on kernel virtual memory.
822 		 */
823 		map = &kernel_map;
824 	} else {
825 		/*
826 		 * This is a fault on non-kernel virtual memory.
827 		 * vm is initialized above to NULL. If curproc is NULL
828 		 * or curproc->p_vmspace is NULL the fault is fatal.
829 		 */
830 		if (lp != NULL)
831 			vm = lp->lwp_vmspace;
832 
833 		if (vm == NULL)
834 			goto nogo;
835 
836 		map = &vm->vm_map;
837 	}
838 
839 	if (frame->tf_err & PGEX_W)
840 		ftype = VM_PROT_READ | VM_PROT_WRITE;
841 	else if (frame->tf_err & PGEX_I)
842 		ftype = VM_PROT_EXECUTE;
843 	else
844 		ftype = VM_PROT_READ;
845 
846 	if (map != &kernel_map) {
847 		/*
848 		 * Keep swapout from messing with us during this
849 		 *	critical time.
850 		 */
851 		PHOLD(lp->lwp_proc);
852 
853 #if 0
854 		/*
855 		 * Grow the stack if necessary
856 		 */
857 		/* grow_stack returns false only if va falls into
858 		 * a growable stack region and the stack growth
859 		 * fails.  It returns true if va was not within
860 		 * a growable stack region, or if the stack
861 		 * growth succeeded.
862 		 */
863 		if (!grow_stack (map, va)) {
864 			rv = KERN_FAILURE;
865 			PRELE(lp->lwp_proc);
866 			goto nogo;
867 		}
868 #endif
869 
870 		fault_flags = 0;
871 		if (usermode)
872 			fault_flags |= VM_FAULT_BURST | VM_FAULT_USERMODE;
873 		if (ftype & VM_PROT_WRITE)
874 			fault_flags |= VM_FAULT_DIRTY;
875 		else
876 			fault_flags |= VM_FAULT_NORMAL;
877 		rv = vm_fault(map, va, ftype, fault_flags);
878 
879 		PRELE(lp->lwp_proc);
880 	} else {
881 		/*
882 		 * Don't have to worry about process locking or stacks in the kernel.
883 		 */
884 		rv = vm_fault(map, va, ftype, VM_FAULT_NORMAL);
885 	}
886 
887 	if (rv == KERN_SUCCESS)
888 		return (0);
889 nogo:
890 	if (!usermode) {
891 		if (td->td_gd->gd_intr_nesting_level == 0 &&
892 		    td->td_pcb->pcb_onfault) {
893 			frame->tf_rip = (register_t)td->td_pcb->pcb_onfault;
894 			return (0);
895 		}
896 		trap_fatal(frame, usermode, eva);
897 		return (-1);
898 	}
899 
900 	/*
901 	 * NOTE: on x86_64 we have a tf_addr field in the trapframe, no
902 	 * kludge is needed to pass the fault address to signal handlers.
903 	 */
904 	struct proc *p = td->td_proc;
905 	krateprintf(&segfltrate,
906 		    "seg-fault accessing address %p "
907 		    "rip=%p pid=%d p_comm=%s\n",
908 		    (void *)va,
909 		    (void *)frame->tf_rip, p->p_pid, p->p_comm);
910 	/* Debugger("seg-fault"); */
911 
912 	return((rv == KERN_PROTECTION_FAILURE) ? SIGBUS : SIGSEGV);
913 }
914 
915 static void
916 trap_fatal(struct trapframe *frame, int usermode, vm_offset_t eva)
917 {
918 	int code, type, ss;
919 	long rsp;
920 
921 	code = frame->tf_xflags;
922 	type = frame->tf_trapno;
923 
924 	if (type <= MAX_TRAP_MSG) {
925 		kprintf("\n\nFatal trap %d: %s while in %s mode\n",
926 			type, trap_msg[type],
927 			(usermode ? "user" : "kernel"));
928 	}
929 	/* two separate prints in case of a trap on an unmapped page */
930 	kprintf("cpuid = %d\n", mycpu->gd_cpuid);
931 	if (type == T_PAGEFLT) {
932 		kprintf("fault virtual address	= %p\n", (void *)eva);
933 		kprintf("fault code		= %s %s, %s\n",
934 			usermode ? "user" : "supervisor",
935 			code & PGEX_W ? "write" : "read",
936 			code & PGEX_P ? "protection violation" : "page not present");
937 	}
938 	kprintf("instruction pointer	= 0x%lx:0x%lx\n",
939 	       frame->tf_cs & 0xffff, frame->tf_rip);
940 	if (usermode) {
941 		ss = frame->tf_ss & 0xffff;
942 		rsp = frame->tf_rsp;
943 	} else {
944 		ss = GSEL(GDATA_SEL, SEL_KPL);
945 		rsp = (long)&frame->tf_rsp;
946 	}
947 	kprintf("stack pointer	        = 0x%x:0x%lx\n", ss, rsp);
948 	kprintf("frame pointer	        = 0x%x:0x%lx\n", ss, frame->tf_rbp);
949 	kprintf("processor eflags	= ");
950 	if (frame->tf_rflags & PSL_T)
951 		kprintf("trace trap, ");
952 	if (frame->tf_rflags & PSL_I)
953 		kprintf("interrupt enabled, ");
954 	if (frame->tf_rflags & PSL_NT)
955 		kprintf("nested task, ");
956 	if (frame->tf_rflags & PSL_RF)
957 		kprintf("resume, ");
958 #if 0
959 	if (frame->tf_eflags & PSL_VM)
960 		kprintf("vm86, ");
961 #endif
962 	kprintf("IOPL = %jd\n", (intmax_t)((frame->tf_rflags & PSL_IOPL) >> 12));
963 	kprintf("current process		= ");
964 	if (curproc) {
965 		kprintf("%lu (%s)\n",
966 		    (u_long)curproc->p_pid, curproc->p_comm ?
967 		    curproc->p_comm : "");
968 	} else {
969 		kprintf("Idle\n");
970 	}
971 	kprintf("current thread          = pri %d ", curthread->td_pri);
972 	if (curthread->td_critcount)
973 		kprintf("(CRIT)");
974 	kprintf("\n");
975 /**
976  *  XXX FIXME:
977  *	we probably SHOULD have stopped the other CPUs before now!
978  *	another CPU COULD have been touching cpl at this moment...
979  */
980 	kprintf(" <- SMP: XXX");
981 	kprintf("\n");
982 
983 #ifdef KDB
984 	if (kdb_trap(&psl))
985 		return;
986 #endif
987 #ifdef DDB
988 	if ((debugger_on_panic || db_active) && kdb_trap(type, code, frame))
989 		return;
990 #endif
991 	kprintf("trap number		= %d\n", type);
992 	if (type <= MAX_TRAP_MSG)
993 		panic("%s", trap_msg[type]);
994 	else
995 		panic("unknown/reserved trap");
996 }
997 
998 /*
999  * Double fault handler. Called when a fault occurs while writing
1000  * a frame for a trap/exception onto the stack. This usually occurs
1001  * when the stack overflows (such is the case with infinite recursion,
1002  * for example).
1003  *
1004  * XXX Note that the current PTD gets replaced by IdlePTD when the
1005  * task switch occurs. This means that the stack that was active at
1006  * the time of the double fault is not available at <kstack> unless
1007  * the machine was idle when the double fault occurred. The downside
1008  * of this is that "trace <ebp>" in ddb won't work.
1009  */
1010 void
1011 dblfault_handler(void)
1012 {
1013 #if JG
1014 	struct mdglobaldata *gd = mdcpu;
1015 #endif
1016 
1017 	kprintf("\nFatal double fault:\n");
1018 #if JG
1019 	kprintf("rip = 0x%lx\n", gd->gd_common_tss.tss_rip);
1020 	kprintf("rsp = 0x%lx\n", gd->gd_common_tss.tss_rsp);
1021 	kprintf("rbp = 0x%lx\n", gd->gd_common_tss.tss_rbp);
1022 #endif
1023 	/* two separate prints in case of a trap on an unmapped page */
1024 	kprintf("cpuid = %d\n", mycpu->gd_cpuid);
1025 	panic("double fault");
1026 }
1027 
1028 /*
1029  *	syscall2 -	MP aware system call request C handler
1030  *
1031  *	A system call is essentially treated as a trap except that the
1032  *	MP lock is not held on entry or return.  We are responsible for
1033  *	obtaining the MP lock if necessary and for handling ASTs
1034  *	(e.g. a task switch) prior to return.
1035  *
1036  *	In general, only simple access and manipulation of curproc and
1037  *	the current stack is allowed without having to hold MP lock.
1038  *
1039  *	MPSAFE - note that large sections of this routine are run without
1040  *		 the MP lock.
1041  */
1042 void
1043 syscall2(struct trapframe *frame)
1044 {
1045 	struct thread *td = curthread;
1046 	struct proc *p = td->td_proc;
1047 	struct lwp *lp = td->td_lwp;
1048 	caddr_t params;
1049 	struct sysent *callp;
1050 	register_t orig_tf_rflags;
1051 	int sticks;
1052 	int error;
1053 	int narg;
1054 #ifdef INVARIANTS
1055 	int crit_count = td->td_critcount;
1056 	lwkt_tokref_t curstop = td->td_toks_stop;
1057 #endif
1058 	register_t *argp;
1059 	u_int code;
1060 	int reg, regcnt;
1061 	union sysunion args;
1062 	register_t *argsdst;
1063 
1064 	mycpu->gd_cnt.v_syscall++;
1065 
1066 	KTR_LOG(kernentry_syscall, lp->lwp_proc->p_pid, lp->lwp_tid,
1067 		frame->tf_rax);
1068 
1069 	userenter(td, p);	/* lazy raise our priority */
1070 
1071 	reg = 0;
1072 	regcnt = 6;
1073 	/*
1074 	 * Misc
1075 	 */
1076 	sticks = (int)td->td_sticks;
1077 	orig_tf_rflags = frame->tf_rflags;
1078 
1079 	/*
1080 	 * Virtual kernel intercept - if a VM context managed by a virtual
1081 	 * kernel issues a system call the virtual kernel handles it, not us.
1082 	 * Restore the virtual kernel context and return from its system
1083 	 * call.  The current frame is copied out to the virtual kernel.
1084 	 */
1085 	if (lp->lwp_vkernel && lp->lwp_vkernel->ve) {
1086 		vkernel_trap(lp, frame);
1087 		error = EJUSTRETURN;
1088 		goto out;
1089 	}
1090 
1091 	/*
1092 	 * Get the system call parameters and account for time
1093 	 */
1094 	lp->lwp_md.md_regs = frame;
1095 	params = (caddr_t)frame->tf_rsp + sizeof(register_t);
1096 	code = frame->tf_rax;
1097 
1098 	if (p->p_sysent->sv_prepsyscall) {
1099 		(*p->p_sysent->sv_prepsyscall)(
1100 			frame, (int *)(&args.nosys.sysmsg + 1),
1101 			&code, &params);
1102 	} else {
1103 		if (code == SYS_syscall || code == SYS___syscall) {
1104 			code = frame->tf_rdi;
1105 			reg++;
1106 			regcnt--;
1107 		}
1108 	}
1109 
1110 	if (p->p_sysent->sv_mask)
1111 		code &= p->p_sysent->sv_mask;
1112 
1113 	if (code >= p->p_sysent->sv_size)
1114 		callp = &p->p_sysent->sv_table[0];
1115 	else
1116 		callp = &p->p_sysent->sv_table[code];
1117 
1118 	narg = callp->sy_narg & SYF_ARGMASK;
1119 
1120 	/*
1121 	 * On x86_64 we get up to six arguments in registers. The rest are
1122 	 * on the stack. The first six members of 'struct trapframe' happen
1123 	 * to be the registers used to pass arguments, in exactly the right
1124 	 * order.
1125 	 */
1126 	argp = &frame->tf_rdi;
1127 	argp += reg;
1128 	argsdst = (register_t *)(&args.nosys.sysmsg + 1);
1129 	/*
1130 	 * JG can we overflow the space pointed to by 'argsdst'
1131 	 * either with 'bcopy' or with 'copyin'?
1132 	 */
1133 	bcopy(argp, argsdst, sizeof(register_t) * regcnt);
1134 	/*
1135 	 * copyin is MP aware, but the tracing code is not
1136 	 */
1137 	if (narg > regcnt) {
1138 		KASSERT(params != NULL, ("copyin args with no params!"));
1139 		error = copyin(params, &argsdst[regcnt],
1140 			(narg - regcnt) * sizeof(register_t));
1141 		if (error) {
1142 #ifdef KTRACE
1143 			if (KTRPOINT(td, KTR_SYSCALL)) {
1144 				ktrsyscall(lp, code, narg,
1145 					(void *)(&args.nosys.sysmsg + 1));
1146 			}
1147 #endif
1148 			goto bad;
1149 		}
1150 	}
1151 
1152 #ifdef KTRACE
1153 	if (KTRPOINT(td, KTR_SYSCALL)) {
1154 		ktrsyscall(lp, code, narg, (void *)(&args.nosys.sysmsg + 1));
1155 	}
1156 #endif
1157 
1158 	/*
1159 	 * Default return value is 0 (will be copied to %rax).  Double-value
1160 	 * returns use %rax and %rdx.  %rdx is left unchanged for system
1161 	 * calls which return only one result.
1162 	 */
1163 	args.sysmsg_fds[0] = 0;
1164 	args.sysmsg_fds[1] = frame->tf_rdx;
1165 
1166 	/*
1167 	 * The syscall might manipulate the trap frame. If it does it
1168 	 * will probably return EJUSTRETURN.
1169 	 */
1170 	args.sysmsg_frame = frame;
1171 
1172 	STOPEVENT(p, S_SCE, narg);	/* MP aware */
1173 
1174 	/*
1175 	 * NOTE: All system calls run MPSAFE now.  The system call itself
1176 	 *	 is responsible for getting the MP lock.
1177 	 */
1178 	error = (*callp->sy_call)(&args);
1179 
1180 #if 0
1181 	kprintf("system call %d returned %d\n", code, error);
1182 #endif
1183 
1184 out:
1185 	/*
1186 	 * MP SAFE (we may or may not have the MP lock at this point)
1187 	 */
1188 	switch (error) {
1189 	case 0:
1190 		/*
1191 		 * Reinitialize proc pointer `p' as it may be different
1192 		 * if this is a child returning from fork syscall.
1193 		 */
1194 		p = curproc;
1195 		lp = curthread->td_lwp;
1196 		frame->tf_rax = args.sysmsg_fds[0];
1197 		frame->tf_rdx = args.sysmsg_fds[1];
1198 		frame->tf_rflags &= ~PSL_C;
1199 		break;
1200 	case ERESTART:
1201 		/*
1202 		 * Reconstruct pc, we know that 'syscall' is 2 bytes.
1203 		 * We have to do a full context restore so that %r10
1204 		 * (which was holding the value of %rcx) is restored for
1205 		 * the next iteration.
1206 		 */
1207 		frame->tf_rip -= frame->tf_err;
1208 		frame->tf_r10 = frame->tf_rcx;
1209 		break;
1210 	case EJUSTRETURN:
1211 		break;
1212 	case EASYNC:
1213 		panic("Unexpected EASYNC return value (for now)");
1214 	default:
1215 bad:
1216 		if (p->p_sysent->sv_errsize) {
1217 			if (error >= p->p_sysent->sv_errsize)
1218 				error = -1;	/* XXX */
1219 			else
1220 				error = p->p_sysent->sv_errtbl[error];
1221 		}
1222 		frame->tf_rax = error;
1223 		frame->tf_rflags |= PSL_C;
1224 		break;
1225 	}
1226 
1227 	/*
1228 	 * Traced syscall.  trapsignal() is not MP aware.
1229 	 */
1230 	if (orig_tf_rflags & PSL_T) {
1231 		frame->tf_rflags &= ~PSL_T;
1232 		trapsignal(lp, SIGTRAP, 0);
1233 	}
1234 
1235 	/*
1236 	 * Handle reschedule and other end-of-syscall issues
1237 	 */
1238 	userret(lp, frame, sticks);
1239 
1240 #ifdef KTRACE
1241 	if (KTRPOINT(td, KTR_SYSRET)) {
1242 		ktrsysret(lp, code, error, args.sysmsg_result);
1243 	}
1244 #endif
1245 
1246 	/*
1247 	 * This works because errno is findable through the
1248 	 * register set.  If we ever support an emulation where this
1249 	 * is not the case, this code will need to be revisited.
1250 	 */
1251 	STOPEVENT(p, S_SCX, code);
1252 
1253 	userexit(lp);
1254 	KTR_LOG(kernentry_syscall_ret, lp->lwp_proc->p_pid, lp->lwp_tid, error);
1255 #ifdef INVARIANTS
1256 	KASSERT(&td->td_toks_base == td->td_toks_stop,
1257 		("syscall: critical section count mismatch! %d/%d",
1258 		crit_count, td->td_pri));
1259 	KASSERT(curstop == td->td_toks_stop,
1260 		("syscall: extra tokens held after trap! %ld",
1261 		td->td_toks_stop - &td->td_toks_base));
1262 #endif
1263 }
1264 
1265 /*
1266  * NOTE: mplock not held at any point
1267  */
1268 void
1269 fork_return(struct lwp *lp, struct trapframe *frame)
1270 {
1271 	frame->tf_rax = 0;		/* Child returns zero */
1272 	frame->tf_rflags &= ~PSL_C;	/* success */
1273 	frame->tf_rdx = 1;
1274 
1275 	generic_lwp_return(lp, frame);
1276 	KTR_LOG(kernentry_fork_ret, lp->lwp_proc->p_pid, lp->lwp_tid);
1277 }
1278 
1279 /*
1280  * Simplified back end of syscall(), used when returning from fork()
1281  * directly into user mode.
1282  *
1283  * This code will return back into the fork trampoline code which then
1284  * runs doreti.
1285  *
1286  * NOTE: The mplock is not held at any point.
1287  */
1288 void
1289 generic_lwp_return(struct lwp *lp, struct trapframe *frame)
1290 {
1291 	struct proc *p = lp->lwp_proc;
1292 
1293 	/*
1294 	 * Check for exit-race.  If one lwp exits the process concurrent with
1295 	 * another lwp creating a new thread, the two operations may cross
1296 	 * each other resulting in the newly-created lwp not receiving a
1297 	 * KILL signal.
1298 	 */
1299 	if (p->p_flags & P_WEXIT) {
1300 		lwpsignal(p, lp, SIGKILL);
1301 	}
1302 
1303 	/*
1304 	 * Newly forked processes are given a kernel priority.  We have to
1305 	 * adjust the priority to a normal user priority and fake entry
1306 	 * into the kernel (call userenter()) to install a passive release
1307 	 * function just in case userret() decides to stop the process.  This
1308 	 * can occur when ^Z races a fork.  If we do not install the passive
1309 	 * release function the current process designation will not be
1310 	 * released when the thread goes to sleep.
1311 	 */
1312 	lwkt_setpri_self(TDPRI_USER_NORM);
1313 	userenter(lp->lwp_thread, p);
1314 	userret(lp, frame, 0);
1315 #ifdef KTRACE
1316 	if (KTRPOINT(lp->lwp_thread, KTR_SYSRET))
1317 		ktrsysret(lp, SYS_fork, 0, 0);
1318 #endif
1319 	lp->lwp_flags |= LWP_PASSIVE_ACQ;
1320 	userexit(lp);
1321 	lp->lwp_flags &= ~LWP_PASSIVE_ACQ;
1322 }
1323 
1324 /*
1325  * doreti has turned into this.  The frame is directly on the stack.  We
1326  * pull everything else we need (fpu and tls context) from the current
1327  * thread.
1328  *
1329  * Note on fpu interactions: In a virtual kernel, the fpu context for
1330  * an emulated user mode process is not shared with the virtual kernel's
1331  * fpu context, so we only have to 'stack' fpu contexts within the virtual
1332  * kernel itself, and not even then since the signal() contexts that we care
1333  * about save and restore the FPU state (I think anyhow).
1334  *
1335  * vmspace_ctl() returns an error only if it had problems instaling the
1336  * context we supplied or problems copying data to/from our VM space.
1337  */
1338 void
1339 go_user(struct intrframe *frame)
1340 {
1341 	struct trapframe *tf = (void *)&frame->if_rdi;
1342 	globaldata_t gd;
1343 	int r;
1344 	void *id;
1345 
1346 	/*
1347 	 * Interrupts may be disabled on entry, make sure all signals
1348 	 * can be received before beginning our loop.
1349 	 */
1350 	sigsetmask(0);
1351 
1352 	/*
1353 	 * Switch to the current simulated user process, then call
1354 	 * user_trap() when we break out of it (usually due to a signal).
1355 	 */
1356 	for (;;) {
1357 #if 1
1358 		/*
1359 		 * Always make the FPU state correct.  This should generally
1360 		 * be faster because the cost of taking a #NM fault through
1361 		 * the vkernel to the real kernel is astronomical.
1362 		 */
1363 		crit_enter();
1364 		tf->tf_xflags &= ~PGEX_FPFAULT;
1365 		if (mdcpu->gd_npxthread != curthread) {
1366 			if (mdcpu->gd_npxthread)
1367 				npxsave(mdcpu->gd_npxthread->td_savefpu);
1368 			npxdna(tf);
1369 		}
1370 #else
1371 		/*
1372 		 * Tell the real kernel whether it is ok to use the FP
1373 		 * unit or not, allowing us to take a T_DNA exception
1374 		 * if the context tries to use the FP.
1375 		 */
1376 		if (mdcpu->gd_npxthread == curthread) {
1377 			tf->tf_xflags &= ~PGEX_FPFAULT;
1378 		} else {
1379 			tf->tf_xflags |= PGEX_FPFAULT;
1380 		}
1381 #endif
1382 
1383 		/*
1384 		 * Run emulated user process context.  This call interlocks
1385 		 * with new mailbox signals.
1386 		 *
1387 		 * Set PGEX_U unconditionally, indicating a user frame (the
1388 		 * bit is normally set only by T_PAGEFLT).
1389 		 */
1390 		if (vmm_enabled)
1391 			id = (void *)vtophys(curproc->p_vmspace->vm_pmap.pm_pml4);
1392 		else
1393 			id = &curproc->p_vmspace->vm_pmap;
1394 
1395 		/*
1396 		 * The GDF_VIRTUSER hack helps statclock() figure out who
1397 		 * the tick belongs to.
1398 		 */
1399 		gd = mycpu;
1400 		gd->gd_flags |= GDF_VIRTUSER;
1401 		r = vmspace_ctl(id, VMSPACE_CTL_RUN, tf,
1402 				&curthread->td_savevext);
1403 		gd->gd_flags &= ~GDF_VIRTUSER;
1404 
1405 		frame->if_xflags |= PGEX_U;
1406 
1407 		/*
1408 		 * Immediately save the user FPU state.  The vkernel is a
1409 		 * user program and libraries like libc will use the FP
1410 		 * unit.
1411 		 */
1412 		if (mdcpu->gd_npxthread == curthread) {
1413 			npxsave(mdcpu->gd_npxthread->td_savefpu);
1414 		}
1415 		crit_exit();
1416 #if 0
1417 		kprintf("GO USER %d trap %ld EVA %08lx RIP %08lx RSP %08lx XFLAGS %02lx/%02lx\n",
1418 			r, tf->tf_trapno, tf->tf_addr, tf->tf_rip, tf->tf_rsp,
1419 			tf->tf_xflags, frame->if_xflags);
1420 #endif
1421 		if (r < 0) {
1422 			if (errno != EINTR)
1423 				panic("vmspace_ctl failed error %d", errno);
1424 		} else {
1425 			if (tf->tf_trapno) {
1426 				user_trap(tf);
1427 			}
1428 		}
1429 		if (mycpu->gd_reqflags & RQF_AST_MASK) {
1430 			tf->tf_trapno = T_ASTFLT;
1431 			user_trap(tf);
1432 		}
1433 		tf->tf_trapno = 0;
1434 	}
1435 }
1436 
1437 /*
1438  * If PGEX_FPFAULT is set then set FP_VIRTFP in the PCB to force a T_DNA
1439  * fault (which is then passed back to the virtual kernel) if an attempt is
1440  * made to use the FP unit.
1441  *
1442  * XXX this is a fairly big hack.
1443  */
1444 void
1445 set_vkernel_fp(struct trapframe *frame)
1446 {
1447 	struct thread *td = curthread;
1448 
1449 	if (frame->tf_xflags & PGEX_FPFAULT) {
1450 		td->td_pcb->pcb_flags |= FP_VIRTFP;
1451 		if (mdcpu->gd_npxthread == td)
1452 			npxexit();
1453 	} else {
1454 		td->td_pcb->pcb_flags &= ~FP_VIRTFP;
1455 	}
1456 }
1457 
1458 /*
1459  * Called from vkernel_trap() to fixup the vkernel's syscall
1460  * frame for vmspace_ctl() return.
1461  */
1462 void
1463 cpu_vkernel_trap(struct trapframe *frame, int error)
1464 {
1465 	frame->tf_rax = error;
1466 	if (error)
1467 		frame->tf_rflags |= PSL_C;
1468 	else
1469 		frame->tf_rflags &= ~PSL_C;
1470 }
1471