2 * Copyright (c) 1990, 1993
3 * The Regents of the University of California. All rights reserved.
4 * Copyright (C) 1994, David Greenman
5 * Copyright (c) 2008-2018 The DragonFly Project.
6 * Copyright (c) 2008 Jordan Gordeev.
8 * This code is derived from software contributed to Berkeley by
9 * the University of Utah, and William Jolitz.
11 * Redistribution and use in source and binary forms, with or without
12 * modification, are permitted provided that the following conditions
14 * 1. Redistributions of source code must retain the above copyright
15 * notice, this list of conditions and the following disclaimer.
16 * 2. Redistributions in binary form must reproduce the above copyright
17 * notice, this list of conditions and the following disclaimer in the
18 * documentation and/or other materials provided with the distribution.
19 * 3. All advertising materials mentioning features or use of this software
20 * must display the following acknowledgement:
21 * This product includes software developed by the University of
22 * California, Berkeley and its contributors.
23 * 4. Neither the name of the University nor the names of its contributors
24 * may be used to endorse or promote products derived from this software
25 * without specific prior written permission.
27 * THIS SOFTWARE IS PROVIDED BY THE REGENTS AND CONTRIBUTORS ``AS IS'' AND
28 * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
29 * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE
30 * ARE DISCLAIMED. IN NO EVENT SHALL THE REGENTS OR CONTRIBUTORS BE LIABLE
31 * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL
32 * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS
33 * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION)
34 * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT
35 * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY
36 * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF
39 * from: @(#)trap.c 7.4 (Berkeley) 5/13/91
40 * $FreeBSD: src/sys/i386/i386/trap.c,v 1.147.2.11 2003/02/27 19:09:59 luoqi Exp $
44 * x86_64 Trap and System call handling
50 #include "opt_ktrace.h"
52 #include <machine/frame.h>
53 #include <sys/param.h>
54 #include <sys/systm.h>
55 #include <sys/kernel.h>
56 #include <sys/kerneldump.h>
58 #include <sys/pioctl.h>
59 #include <sys/types.h>
60 #include <sys/signal2.h>
61 #include <sys/syscall.h>
62 #include <sys/sysctl.h>
63 #include <sys/sysent.h>
65 #include <sys/ktrace.h>
68 #include <sys/sysmsg.h>
69 #include <sys/sysproto.h>
70 #include <sys/sysunion.h>
74 #include <vm/vm_extern.h>
75 #include <vm/vm_kern.h>
76 #include <vm/vm_param.h>
77 #include <machine/cpu.h>
78 #include <machine/pcb.h>
79 #include <machine/smp.h>
80 #include <machine/thread.h>
81 #include <machine/clock.h>
82 #include <machine/vmparam.h>
83 #include <machine/md_var.h>
84 #include <machine_base/isa/isa_intr.h>
85 #include <machine_base/apic/lapic.h>
89 #include <sys/thread2.h>
90 #include <sys/spinlock2.h>
93 * These %rip's are used to detect a historical CPU artifact on syscall or
94 * int $3 entry, if not shortcutted in exception.S via
95 * DIRECT_DISALLOW_SS_CPUBUG.
97 extern void Xbpt(void);
98 extern void Xfast_syscall(void);
99 #define IDTVEC(vec) X##vec
101 extern void trap(struct trapframe *frame);
103 static int trap_pfault(struct trapframe *, int);
104 static void trap_fatal(struct trapframe *, vm_offset_t);
105 void dblfault_handler(struct trapframe *frame);
107 #define MAX_TRAP_MSG 30
108 static char *trap_msg[] = {
110 "privileged instruction fault", /* 1 T_PRIVINFLT */
112 "breakpoint instruction fault", /* 3 T_BPTFLT */
115 "arithmetic trap", /* 6 T_ARITHTRAP */
116 "system forced exception", /* 7 T_ASTFLT */
118 "general protection fault", /* 9 T_PROTFLT */
119 "trace trap", /* 10 T_TRCTRAP */
121 "page fault", /* 12 T_PAGEFLT */
123 "alignment fault", /* 14 T_ALIGNFLT */
127 "integer divide fault", /* 18 T_DIVIDE */
128 "non-maskable interrupt trap", /* 19 T_NMI */
129 "overflow trap", /* 20 T_OFLOW */
130 "FPU bounds check fault", /* 21 T_BOUND */
131 "FPU device not available", /* 22 T_DNA */
132 "double fault", /* 23 T_DOUBLEFLT */
133 "FPU operand fetch fault", /* 24 T_FPOPFLT */
134 "invalid TSS fault", /* 25 T_TSSFLT */
135 "segment not present fault", /* 26 T_SEGNPFLT */
136 "stack fault", /* 27 T_STKFLT */
137 "machine check trap", /* 28 T_MCHK */
138 "SIMD floating-point exception", /* 29 T_XMMFLT */
139 "reserved (unknown) fault", /* 30 T_RESERVED */
143 static int ddb_on_nmi = 1;
144 SYSCTL_INT(_machdep, OID_AUTO, ddb_on_nmi, CTLFLAG_RW,
145 &ddb_on_nmi, 0, "Go to DDB on NMI");
146 static int ddb_on_seg_fault = 0;
147 SYSCTL_INT(_machdep, OID_AUTO, ddb_on_seg_fault, CTLFLAG_RW,
148 &ddb_on_seg_fault, 0, "Go to DDB on user seg-fault");
149 static int freeze_on_seg_fault = 0;
150 SYSCTL_INT(_machdep, OID_AUTO, freeze_on_seg_fault, CTLFLAG_RW,
151 &freeze_on_seg_fault, 0, "Go to DDB on user seg-fault");
153 static int panic_on_nmi = 1;
154 SYSCTL_INT(_machdep, OID_AUTO, panic_on_nmi, CTLFLAG_RW,
155 &panic_on_nmi, 0, "Panic on NMI");
156 static int fast_release;
157 SYSCTL_INT(_machdep, OID_AUTO, fast_release, CTLFLAG_RW,
158 &fast_release, 0, "Passive Release was optimal");
159 static int slow_release;
160 SYSCTL_INT(_machdep, OID_AUTO, slow_release, CTLFLAG_RW,
161 &slow_release, 0, "Passive Release was nonoptimal");
164 * System call debugging records the worst-case system call
165 * overhead (inclusive of blocking), but may be inaccurate.
167 /*#define SYSCALL_DEBUG*/
169 uint64_t SysCallsWorstCase[SYS_MAXSYSCALL];
173 * Passively intercepts the thread switch function to increase
174 * the thread priority from a user priority to a kernel priority, reducing
175 * syscall and trap overhead for the case where no switch occurs.
177 * Synchronizes td_ucred with p_ucred. This is used by system calls,
178 * signal handling, faults, AST traps, and anything else that enters the
179 * kernel from userland and provides the kernel with a stable read-only
180 * copy of the process ucred.
182 * To avoid races with another thread updating p_ucred we obtain p_spin.
183 * The other thread doing the update will obtain both p_token and p_spin.
184 * In the case where the cached cred pointer matches, we will already have
185 * the ref and we don't have to do one blessed thing.
188 userenter(struct thread *curtd, struct proc *curp)
193 curtd->td_release = lwkt_passive_release;
195 if (curtd->td_ucred != curp->p_ucred) {
196 spin_lock(&curp->p_spin);
197 ncred = crhold(curp->p_ucred);
198 spin_unlock(&curp->p_spin);
199 ocred = curtd->td_ucred;
200 curtd->td_ucred = ncred;
207 * Debugging, remove top two user stack pages to catch kernel faults
209 if (freeze_on_seg_fault > 1 && curtd->td_lwp) {
210 pmap_remove(vmspace_pmap(curtd->td_lwp->lwp_vmspace),
211 0x00007FFFFFFFD000LU,
212 0x0000800000000000LU);
218 * Handle signals, upcalls, profiling, and other AST's and/or tasks that
219 * must be completed before we can return to or try to return to userland.
221 * Note that td_sticks is a 64 bit quantity, but there's no point doing 64
222 * arithmatic on the delta calculation so the absolute tick values are
223 * truncated to an integer.
226 userret(struct lwp *lp, struct trapframe *frame, int sticks)
228 struct proc *p = lp->lwp_proc;
233 * Charge system time if profiling. Note: times are in microseconds.
234 * This may do a copyout and block, so do it first even though it
235 * means some system time will be charged as user time.
237 if (p->p_flags & P_PROFIL) {
238 addupc_task(p, frame->tf_rip,
239 (u_int)((int)lp->lwp_thread->td_sticks - sticks));
244 * Specific on-return-to-usermode checks (LWP_MP_WEXIT,
245 * LWP_MP_VNLRU, etc).
247 if (lp->lwp_mpflags & LWP_MP_URETMASK)
251 * Block here if we are in a stopped state.
253 if (STOPLWP(p, lp)) {
254 lwkt_gettoken(&p->p_token);
256 lwkt_reltoken(&p->p_token);
259 while (dump_stop_usertds) {
260 tsleep(&dump_stop_usertds, 0, "dumpstp", 0);
264 * Post any pending upcalls. If running a virtual kernel be sure
265 * to restore the virtual kernel's vmspace before posting the upcall.
267 if (p->p_flags & (P_SIGVTALRM | P_SIGPROF)) {
268 lwkt_gettoken(&p->p_token);
269 if (p->p_flags & P_SIGVTALRM) {
270 p->p_flags &= ~P_SIGVTALRM;
271 ksignal(p, SIGVTALRM);
273 if (p->p_flags & P_SIGPROF) {
274 p->p_flags &= ~P_SIGPROF;
277 lwkt_reltoken(&p->p_token);
282 * Post any pending signals. If running a virtual kernel be sure
283 * to restore the virtual kernel's vmspace before posting the signal.
285 * WARNING! postsig() can exit and not return.
287 if ((sig = CURSIG_LCK_TRACE(lp, &ptok)) != 0) {
293 * block here if we are swapped out, but still process signals
294 * (such as SIGKILL). proc0 (the swapin scheduler) is already
295 * aware of our situation, we do not have to wake it up.
297 if (p->p_flags & P_SWAPPEDOUT) {
298 lwkt_gettoken(&p->p_token);
299 p->p_flags |= P_SWAPWAIT;
301 if (p->p_flags & P_SWAPWAIT)
302 tsleep(p, PCATCH, "SWOUT", 0);
303 p->p_flags &= ~P_SWAPWAIT;
304 lwkt_reltoken(&p->p_token);
309 * In a multi-threaded program it is possible for a thread to change
310 * signal state during a system call which temporarily changes the
311 * signal mask. In this case postsig() might not be run and we
312 * have to restore the mask ourselves.
314 if (lp->lwp_flags & LWP_OLDMASK) {
315 lp->lwp_flags &= ~LWP_OLDMASK;
316 lp->lwp_sigmask = lp->lwp_oldsigmask;
322 * Cleanup from userenter and any passive release that might have occured.
323 * We must reclaim the current-process designation before we can return
324 * to usermode. We also handle both LWKT and USER reschedule requests.
327 userexit(struct lwp *lp)
329 struct thread *td = lp->lwp_thread;
330 /* globaldata_t gd = td->td_gd; */
333 * Handle stop requests at kernel priority. Any requests queued
334 * after this loop will generate another AST.
336 while (STOPLWP(lp->lwp_proc, lp)) {
337 lwkt_gettoken(&lp->lwp_proc->p_token);
339 lwkt_reltoken(&lp->lwp_proc->p_token);
343 * Reduce our priority in preparation for a return to userland. If
344 * our passive release function was still in place, our priority was
345 * never raised and does not need to be reduced.
347 lwkt_passive_recover(td);
349 /* WARNING: we may have migrated cpu's */
350 /* gd = td->td_gd; */
353 * Become the current user scheduled process if we aren't already,
354 * and deal with reschedule requests and other factors.
356 lp->lwp_proc->p_usched->acquire_curproc(lp);
359 #if !defined(KTR_KERNENTRY)
360 #define KTR_KERNENTRY KTR_ALL
362 KTR_INFO_MASTER(kernentry);
363 KTR_INFO(KTR_KERNENTRY, kernentry, trap, 0,
364 "TRAP(pid %d, tid %d, trapno %ld, eva %lu)",
365 pid_t pid, lwpid_t tid, register_t trapno, vm_offset_t eva);
366 KTR_INFO(KTR_KERNENTRY, kernentry, trap_ret, 0, "TRAP_RET(pid %d, tid %d)",
367 pid_t pid, lwpid_t tid);
368 KTR_INFO(KTR_KERNENTRY, kernentry, syscall, 0, "SYSC(pid %d, tid %d, nr %ld)",
369 pid_t pid, lwpid_t tid, register_t trapno);
370 KTR_INFO(KTR_KERNENTRY, kernentry, syscall_ret, 0, "SYSRET(pid %d, tid %d, err %d)",
371 pid_t pid, lwpid_t tid, int err);
372 KTR_INFO(KTR_KERNENTRY, kernentry, fork_ret, 0, "FORKRET(pid %d, tid %d)",
373 pid_t pid, lwpid_t tid);
376 * Exception, fault, and trap interface to the kernel.
377 * This common code is called from assembly language IDT gate entry
378 * routines that prepare a suitable stack frame, and restore this
379 * frame after the exception has been processed.
381 * This function is also called from doreti in an interlock to handle ASTs.
382 * For example: hardwareint->INTROUTINE->(set ast)->doreti->trap
384 * NOTE! We have to retrieve the fault address prior to potentially
385 * blocking, including blocking on any token.
387 * NOTE! NMI and kernel DBG traps remain on their respective pcpu IST
388 * stacks if taken from a kernel RPL. trap() cannot block in this
389 * situation. DDB entry or a direct report-and-return is ok.
391 * XXX gd_trap_nesting_level currently prevents lwkt_switch() from panicing
392 * if an attempt is made to switch from a fast interrupt or IPI.
395 trap(struct trapframe *frame)
397 static struct krate sscpubugrate = { 1 };
398 struct globaldata *gd = mycpu;
399 struct thread *td = gd->gd_curthread;
400 struct lwp *lp = td->td_lwp;
403 int i = 0, ucode = 0, type, code;
405 int crit_count = td->td_critcount;
406 lwkt_tokref_t curstop = td->td_toks_stop;
415 * We need to allow T_DNA faults when the debugger is active since
416 * some dumping paths do large bcopy() which use the floating
417 * point registers for faster copying.
419 if (db_active && frame->tf_trapno != T_DNA) {
420 eva = (frame->tf_trapno == T_PAGEFLT ? frame->tf_addr : 0);
421 ++gd->gd_trap_nesting_level;
422 trap_fatal(frame, eva);
423 --gd->gd_trap_nesting_level;
430 if ((frame->tf_rflags & PSL_I) == 0) {
432 * Buggy application or kernel code has disabled interrupts
433 * and then trapped. Enabling interrupts now is wrong, but
434 * it is better than running with interrupts disabled until
435 * they are accidentally enabled later.
438 type = frame->tf_trapno;
439 if (ISPL(frame->tf_cs) == SEL_UPL) {
440 /* JG curproc can be NULL */
442 "pid %ld (%s): trap %d with interrupts disabled\n",
443 (long)curproc->p_pid, curproc->p_comm, type);
444 } else if ((type == T_STKFLT || type == T_PROTFLT ||
445 type == T_SEGNPFLT) &&
446 frame->tf_rip == (long)doreti_iret) {
448 * iretq fault from kernel mode during return to
451 * This situation is expected, don't complain.
453 } else if (type != T_NMI && type != T_BPTFLT &&
456 * XXX not quite right, since this may be for a
457 * multiple fault in user mode.
459 kprintf("kernel trap %d (%s @ 0x%016jx) with "
460 "interrupts disabled\n",
468 type = frame->tf_trapno;
469 code = frame->tf_err;
471 if (ISPL(frame->tf_cs) == SEL_UPL) {
474 KTR_LOG(kernentry_trap, p->p_pid, lp->lwp_tid,
475 frame->tf_trapno, eva);
479 sticks = (int)td->td_sticks;
480 KASSERT(lp->lwp_md.md_regs == frame,
481 ("Frame mismatch %p %p", lp->lwp_md.md_regs, frame));
484 case T_PRIVINFLT: /* privileged instruction fault */
489 case T_BPTFLT: /* bpt instruction fault */
490 case T_TRCTRAP: /* trace trap */
491 frame->tf_rflags &= ~PSL_T;
493 ucode = (type == T_TRCTRAP ? TRAP_TRACE : TRAP_BRKPT);
496 case T_ARITHTRAP: /* arithmetic trap */
501 case T_ASTFLT: /* Allow process switch */
502 mycpu->gd_cnt.v_soft++;
503 if (mycpu->gd_reqflags & RQF_AST_OWEUPC) {
504 atomic_clear_int(&mycpu->gd_reqflags,
506 addupc_task(p, p->p_prof.pr_addr,
511 case T_PROTFLT: /* general protection fault */
515 case T_STKFLT: /* stack fault */
516 case T_SEGNPFLT: /* segment not present fault */
520 case T_TSSFLT: /* invalid TSS fault */
521 case T_DOUBLEFLT: /* double fault */
527 case T_PAGEFLT: /* page fault */
528 i = trap_pfault(frame, TRUE);
530 if (frame->tf_rip == 0) {
531 /* used for kernel debugging only */
532 while (freeze_on_seg_fault)
533 tsleep(p, 0, "freeze", hz * 20);
536 if (i == -1 || i == 0)
546 case T_DIVIDE: /* integer divide fault */
553 /* machine/parity/power fail/"kitchen sink" faults */
554 if (isa_nmi(code) == 0) {
557 * NMI can be hooked up to a pushbutton
561 kprintf ("NMI ... going to debugger\n");
562 kdb_trap(type, 0, frame);
566 } else if (panic_on_nmi)
567 panic("NMI indicates hardware failure");
569 #endif /* NISA > 0 */
571 case T_OFLOW: /* integer overflow fault */
576 case T_BOUND: /* bounds check fault */
583 * Virtual kernel intercept - pass the DNA exception
584 * to the virtual kernel if it asked to handle it.
585 * This occurs when the virtual kernel is holding
586 * onto the FP context for a different emulated
587 * process then the one currently running.
589 * We must still call npxdna() since we may have
590 * saved FP state that the virtual kernel needs
591 * to hand over to a different emulated process.
593 if (lp->lwp_vkernel && lp->lwp_vkernel->ve &&
594 (td->td_pcb->pcb_flags & FP_VIRTFP)
601 * The kernel may have switched out the FP unit's
602 * state, causing the user process to take a fault
603 * when it tries to use the FP unit. Restore the
611 ucode = FPE_FPU_NP_TRAP;
614 case T_FPOPFLT: /* FPU operand fetch fault */
619 case T_XMMFLT: /* SIMD floating-point exception */
628 case T_PAGEFLT: /* page fault */
629 trap_pfault(frame, FALSE);
634 * The kernel is apparently using fpu for copying.
635 * XXX this should be fatal unless the kernel has
636 * registered such use.
644 case T_STKFLT: /* stack fault */
645 case T_PROTFLT: /* general protection fault */
646 case T_SEGNPFLT: /* segment not present fault */
648 * Invalid segment selectors and out of bounds
649 * %rip's and %rsp's can be set up in user mode.
650 * This causes a fault in kernel mode when the
651 * kernel tries to return to user mode. We want
652 * to get this fault so that we can fix the
653 * problem here and not have to check all the
654 * selectors and pointers when the user changes
657 if (mycpu->gd_intr_nesting_level == 0) {
659 * NOTE: in 64-bit mode traps push rsp/ss
660 * even if no ring change occurs.
662 if (td->td_pcb->pcb_onfault &&
663 td->td_pcb->pcb_onfault_sp ==
665 frame->tf_rip = (register_t)
666 td->td_pcb->pcb_onfault;
671 * If the iretq in doreti faults during
672 * return to user, it will be special-cased
673 * in IDTVEC(prot) to get here. We want
674 * to 'return' to doreti_iret_fault in
675 * ipl.s in approximately the same state we
676 * were in at the iretq.
678 if (frame->tf_rip == (long)doreti_iret) {
679 frame->tf_rip = (long)doreti_iret_fault;
687 * PSL_NT can be set in user mode and isn't cleared
688 * automatically when the kernel is entered. This
689 * causes a TSS fault when the kernel attempts to
690 * `iret' because the TSS link is uninitialized. We
691 * want to get this fault so that we can fix the
692 * problem here and not every time the kernel is
695 if (frame->tf_rflags & PSL_NT) {
696 frame->tf_rflags &= ~PSL_NT;
698 /* do we need this? */
699 if (frame->tf_rip == (long)doreti_iret)
700 frame->tf_rip = (long)doreti_iret_fault;
706 case T_TRCTRAP: /* trace trap */
708 * Detect historical CPU artifact on syscall or int $3
709 * entry (if not shortcutted in exception.s via
710 * DIRECT_DISALLOW_SS_CPUBUG).
713 if (frame->tf_rip == (register_t)IDTVEC(fast_syscall)) {
714 krateprintf(&sscpubugrate,
715 "Caught #DB at syscall cpu artifact\n");
718 if (frame->tf_rip == (register_t)IDTVEC(bpt)) {
719 krateprintf(&sscpubugrate,
720 "Caught #DB at int $N cpu artifact\n");
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.
734 if (user_dbreg_trap()) {
736 * Reset breakpoint bits because the
739 load_dr6(rdr6() & ~0xf);
743 * FALLTHROUGH (TRCTRAP kernel mode, kernel address)
747 * If DDB is enabled, let it handle the debugger trap.
748 * Otherwise, debugger traps "can't happen".
752 if (kdb_trap(type, 0, frame))
759 /* machine/parity/power fail/"kitchen sink" faults */
760 if (isa_nmi(code) == 0) {
763 * NMI can be hooked up to a pushbutton
767 kprintf ("NMI ... going to debugger\n");
768 kdb_trap(type, 0, frame);
772 } else if (panic_on_nmi == 0)
775 #endif /* NISA > 0 */
777 trap_fatal(frame, 0);
782 * Fault from user mode, virtual kernel interecept.
784 * If the fault is directly related to a VM context managed by a
785 * virtual kernel then let the virtual kernel handle it.
787 if (lp->lwp_vkernel && lp->lwp_vkernel->ve) {
788 vkernel_trap(lp, frame);
792 /* Translate fault for emulators (e.g. Linux) */
793 if (*p->p_sysent->sv_transtrap)
794 i = (*p->p_sysent->sv_transtrap)(i, type);
797 trapsignal(lp, i, ucode);
800 if (type <= MAX_TRAP_MSG) {
801 uprintf("fatal process exception: %s",
803 if ((type == T_PAGEFLT) || (type == T_PROTFLT))
804 uprintf(", fault VA = 0x%lx", frame->tf_addr);
810 userret(lp, frame, sticks);
813 if (p != NULL && lp != NULL)
814 KTR_LOG(kernentry_trap_ret, p->p_pid, lp->lwp_tid);
816 KASSERT(crit_count == td->td_critcount,
817 ("trap: critical section count mismatch! %d/%d",
818 crit_count, td->td_pri));
819 KASSERT(curstop == td->td_toks_stop,
820 ("trap: extra tokens held after trap! %ld/%ld",
821 curstop - &td->td_toks_base,
822 td->td_toks_stop - &td->td_toks_base));
827 trap_handle_userenter(struct thread *td)
829 userenter(td, td->td_proc);
833 trap_handle_userexit(struct trapframe *frame, int sticks)
835 struct lwp *lp = curthread->td_lwp;
838 userret(lp, frame, sticks);
844 trap_pfault(struct trapframe *frame, int usermode)
847 struct vmspace *vm = NULL;
852 thread_t td = curthread;
853 struct lwp *lp = td->td_lwp;
856 va = trunc_page(frame->tf_addr);
857 if (va >= VM_MIN_KERNEL_ADDRESS) {
859 * Don't allow user-mode faults in kernel address space.
870 * This is a fault on non-kernel virtual memory.
871 * vm is initialized above to NULL. If curproc is NULL
872 * or curproc->p_vmspace is NULL the fault is fatal.
875 vm = lp->lwp_vmspace;
884 * Debugging, try to catch kernel faults on the user address
885 * space when not inside on onfault (e.g. copyin/copyout)
888 if (usermode == 0 && (td->td_pcb == NULL ||
889 td->td_pcb->pcb_onfault == NULL)) {
891 if (freeze_on_seg_fault) {
892 kprintf("trap_pfault: user address fault from kernel mode "
893 "%016lx\n", (long)frame->tf_addr);
894 while (freeze_on_seg_fault)
895 tsleep(&freeze_on_seg_fault, 0, "frzseg", hz * 20);
903 * PGEX_I is defined only if the execute disable bit capability is
904 * supported and enabled.
906 if (frame->tf_err & PGEX_W)
907 ftype = VM_PROT_WRITE;
908 else if (frame->tf_err & PGEX_I)
909 ftype = VM_PROT_EXECUTE;
911 ftype = VM_PROT_READ;
913 if (map != &kernel_map) {
915 * Keep swapout from messing with us during this
925 fault_flags |= VM_FAULT_BURST | VM_FAULT_USERMODE;
926 if (ftype & VM_PROT_WRITE)
927 fault_flags |= VM_FAULT_DIRTY;
929 fault_flags |= VM_FAULT_NORMAL;
930 rv = vm_fault(map, va, ftype, fault_flags);
935 * Don't have to worry about process locking or stacks in the
938 fault_flags = VM_FAULT_NORMAL;
939 rv = vm_fault(map, va, ftype, VM_FAULT_NORMAL);
941 if (rv == KERN_SUCCESS)
946 * NOTE: in 64-bit mode traps push rsp/ss
947 * even if no ring change occurs.
949 if (td->td_pcb->pcb_onfault &&
950 td->td_pcb->pcb_onfault_sp == frame->tf_rsp &&
951 td->td_gd->gd_intr_nesting_level == 0) {
952 frame->tf_rip = (register_t)td->td_pcb->pcb_onfault;
955 trap_fatal(frame, frame->tf_addr);
960 * NOTE: on x86_64 we have a tf_addr field in the trapframe, no
961 * kludge is needed to pass the fault address to signal handlers.
965 if (td->td_lwp->lwp_vkernel == NULL) {
966 while (freeze_on_seg_fault) {
967 tsleep(p, 0, "freeze", hz * 20);
969 if (ddb_on_seg_fault)
970 Debugger("ddb_on_seg_fault");
974 return((rv == KERN_PROTECTION_FAILURE) ? SIGBUS : SIGSEGV);
978 trap_fatal(struct trapframe *frame, vm_offset_t eva)
983 struct soft_segment_descriptor softseg;
986 code = frame->tf_err;
987 type = frame->tf_trapno;
988 sdtossd(&gdt[IDXSEL(frame->tf_cs & 0xffff)], &softseg);
990 if (type <= MAX_TRAP_MSG)
991 msg = trap_msg[type];
994 kprintf("\n\nFatal trap %d: %s while in %s mode\n", type, msg,
995 ISPL(frame->tf_cs) == SEL_UPL ? "user" : "kernel");
996 /* three separate prints in case of a trap on an unmapped page */
997 kprintf("cpuid = %d; ", mycpu->gd_cpuid);
999 kprintf("lapic id = %u\n", LAPIC_READID);
1000 if (type == T_PAGEFLT) {
1001 kprintf("fault virtual address = 0x%lx\n", eva);
1002 kprintf("fault code = %s %s %s, %s\n",
1003 code & PGEX_U ? "user" : "supervisor",
1004 code & PGEX_W ? "write" : "read",
1005 code & PGEX_I ? "instruction" : "data",
1006 code & PGEX_P ? "protection violation" : "page not present");
1008 kprintf("instruction pointer = 0x%lx:0x%lx\n",
1009 frame->tf_cs & 0xffff, frame->tf_rip);
1010 if (ISPL(frame->tf_cs) == SEL_UPL) {
1011 ss = frame->tf_ss & 0xffff;
1012 rsp = frame->tf_rsp;
1015 * NOTE: in 64-bit mode traps push rsp/ss even if no ring
1018 ss = GSEL(GDATA_SEL, SEL_KPL);
1019 rsp = frame->tf_rsp;
1021 kprintf("stack pointer = 0x%x:0x%lx\n", ss, rsp);
1022 kprintf("frame pointer = 0x%x:0x%lx\n", ss, frame->tf_rbp);
1023 kprintf("code segment = base 0x%lx, limit 0x%lx, type 0x%x\n",
1024 softseg.ssd_base, softseg.ssd_limit, softseg.ssd_type);
1025 kprintf(" = DPL %d, pres %d, long %d, def32 %d, gran %d\n",
1026 softseg.ssd_dpl, softseg.ssd_p, softseg.ssd_long, softseg.ssd_def32,
1028 kprintf("processor eflags = ");
1029 if (frame->tf_rflags & PSL_T)
1030 kprintf("trace trap, ");
1031 if (frame->tf_rflags & PSL_I)
1032 kprintf("interrupt enabled, ");
1033 if (frame->tf_rflags & PSL_NT)
1034 kprintf("nested task, ");
1035 if (frame->tf_rflags & PSL_RF)
1036 kprintf("resume, ");
1037 kprintf("IOPL = %ld\n", (frame->tf_rflags & PSL_IOPL) >> 12);
1038 kprintf("current process = ");
1041 (u_long)curproc->p_pid);
1045 kprintf("current thread = pri %d ", curthread->td_pri);
1046 if (curthread->td_critcount)
1051 if ((debugger_on_panic || db_active) && kdb_trap(type, code, frame))
1054 kprintf("trap number = %d\n", type);
1055 if (type <= MAX_TRAP_MSG)
1056 panic("%s", trap_msg[type]);
1058 panic("unknown/reserved trap");
1062 * Double fault handler. Called when a fault occurs while writing
1063 * a frame for a trap/exception onto the stack. This usually occurs
1064 * when the stack overflows (such is the case with infinite recursion,
1069 in_kstack_guard(register_t rptr)
1071 thread_t td = curthread;
1073 if ((char *)rptr >= td->td_kstack &&
1074 (char *)rptr < td->td_kstack + PAGE_SIZE) {
1081 dblfault_handler(struct trapframe *frame)
1083 thread_t td = curthread;
1085 if (in_kstack_guard(frame->tf_rsp) || in_kstack_guard(frame->tf_rbp)) {
1086 kprintf("DOUBLE FAULT - KERNEL STACK GUARD HIT!\n");
1087 if (in_kstack_guard(frame->tf_rsp))
1088 frame->tf_rsp = (register_t)(td->td_kstack + PAGE_SIZE);
1089 if (in_kstack_guard(frame->tf_rbp))
1090 frame->tf_rbp = (register_t)(td->td_kstack + PAGE_SIZE);
1092 kprintf("DOUBLE FAULT\n");
1094 kprintf("\nFatal double fault\n");
1095 kprintf("rip = 0x%lx\n", frame->tf_rip);
1096 kprintf("rsp = 0x%lx\n", frame->tf_rsp);
1097 kprintf("rbp = 0x%lx\n", frame->tf_rbp);
1098 /* three separate prints in case of a trap on an unmapped page */
1099 kprintf("cpuid = %d; ", mycpu->gd_cpuid);
1101 kprintf("lapic id = %u\n", LAPIC_READID);
1102 panic("double fault");
1106 * syscall2 - MP aware system call request C handler
1108 * A system call is essentially treated as a trap except that the
1109 * MP lock is not held on entry or return. We are responsible for
1110 * obtaining the MP lock if necessary and for handling ASTs
1111 * (e.g. a task switch) prior to return.
1116 syscall2(struct trapframe *frame)
1118 struct thread *td = curthread;
1119 struct proc *p = td->td_proc;
1120 struct lwp *lp = td->td_lwp;
1121 struct sysent *callp;
1122 register_t orig_tf_rflags;
1127 int crit_count = td->td_critcount;
1131 int regcnt, optimized_regcnt;
1132 union sysunion args;
1133 register_t *argsdst;
1135 mycpu->gd_cnt.v_syscall++;
1138 if (ISPL(frame->tf_cs) != SEL_UPL) {
1144 KTR_LOG(kernentry_syscall, p->p_pid, lp->lwp_tid,
1147 userenter(td, p); /* lazy raise our priority */
1150 optimized_regcnt = 6;
1155 sticks = (int)td->td_sticks;
1156 orig_tf_rflags = frame->tf_rflags;
1159 * Virtual kernel intercept - if a VM context managed by a virtual
1160 * kernel issues a system call the virtual kernel handles it, not us.
1161 * Restore the virtual kernel context and return from its system
1162 * call. The current frame is copied out to the virtual kernel.
1164 if (lp->lwp_vkernel && lp->lwp_vkernel->ve) {
1165 vkernel_trap(lp, frame);
1166 error = EJUSTRETURN;
1173 * Get the system call parameters and account for time
1175 KASSERT(lp->lwp_md.md_regs == frame,
1176 ("Frame mismatch %p %p", lp->lwp_md.md_regs, frame));
1177 code = (u_int)frame->tf_rax;
1179 if (code == SYS_syscall || code == SYS___syscall) {
1180 code = frame->tf_rdi;
1182 argp = &frame->tf_rdi + 1;
1184 argp = &frame->tf_rdi;
1187 if (code >= p->p_sysent->sv_size)
1188 callp = &p->p_sysent->sv_table[0];
1190 callp = &p->p_sysent->sv_table[code];
1192 narg = callp->sy_narg & SYF_ARGMASK;
1195 * On x86_64 we get up to six arguments in registers. The rest are
1196 * on the stack. The first six members of 'struct trapframe' happen
1197 * to be the registers used to pass arguments, in exactly the right
1200 argsdst = (register_t *)(&args.nosys.sysmsg + 1);
1203 * Its easier to copy up to the highest number of syscall arguments
1204 * passed in registers, which is 6, than to conditionalize it.
1206 __builtin_memcpy(argsdst, argp, sizeof(register_t) * optimized_regcnt);
1209 * Any arguments beyond available argument-passing registers must
1210 * be copyin()'d from the user stack.
1212 if (narg > regcnt) {
1215 params = (caddr_t)frame->tf_rsp + sizeof(register_t);
1216 error = copyin(params, &argsdst[regcnt],
1217 (narg - regcnt) * sizeof(register_t));
1220 if (KTRPOINT(td, KTR_SYSCALL)) {
1221 ktrsyscall(lp, code, narg,
1222 (void *)(&args.nosys.sysmsg + 1));
1230 if (KTRPOINT(td, KTR_SYSCALL)) {
1231 ktrsyscall(lp, code, narg, (void *)(&args.nosys.sysmsg + 1));
1236 * Default return value is 0 (will be copied to %rax). Double-value
1237 * returns use %rax and %rdx. %rdx is left unchanged for system
1238 * calls which return only one result.
1240 args.sysmsg_fds[0] = 0;
1241 args.sysmsg_fds[1] = frame->tf_rdx;
1244 * The syscall might manipulate the trap frame. If it does it
1245 * will probably return EJUSTRETURN.
1247 args.sysmsg_frame = frame;
1249 STOPEVENT(p, S_SCE, narg); /* MP aware */
1252 * NOTE: All system calls run MPSAFE now. The system call itself
1253 * is responsible for getting the MP lock.
1255 #ifdef SYSCALL_DEBUG
1256 tsc_uclock_t tscval = rdtsc();
1258 error = (*callp->sy_call)(&args);
1259 #ifdef SYSCALL_DEBUG
1260 tscval = rdtsc() - tscval;
1261 tscval = tscval * 1000000 / tsc_frequency;
1262 if (SysCallsWorstCase[code] < tscval)
1263 SysCallsWorstCase[code] = tscval;
1268 * MP SAFE (we may or may not have the MP lock at this point)
1270 //kprintf("SYSMSG %d ", error);
1274 * Reinitialize proc pointer `p' as it may be different
1275 * if this is a child returning from fork syscall.
1278 lp = curthread->td_lwp;
1279 frame->tf_rax = args.sysmsg_fds[0];
1280 frame->tf_rdx = args.sysmsg_fds[1];
1281 frame->tf_rflags &= ~PSL_C;
1285 * Reconstruct pc, we know that 'syscall' is 2 bytes.
1286 * We have to do a full context restore so that %r10
1287 * (which was holding the value of %rcx) is restored for
1288 * the next iteration.
1290 if (frame->tf_err != 0 && frame->tf_err != 2)
1291 kprintf("lp %s:%d frame->tf_err is weird %ld\n",
1292 td->td_comm, lp->lwp_proc->p_pid, frame->tf_err);
1293 frame->tf_rip -= frame->tf_err;
1294 frame->tf_r10 = frame->tf_rcx;
1299 panic("Unexpected EASYNC return value (for now)");
1302 if (p->p_sysent->sv_errsize) {
1303 if (error >= p->p_sysent->sv_errsize)
1304 error = -1; /* XXX */
1306 error = p->p_sysent->sv_errtbl[error];
1308 frame->tf_rax = error;
1309 frame->tf_rflags |= PSL_C;
1314 * Traced syscall. trapsignal() should now be MP aware
1316 if (orig_tf_rflags & PSL_T) {
1317 frame->tf_rflags &= ~PSL_T;
1318 trapsignal(lp, SIGTRAP, TRAP_TRACE);
1322 * Handle reschedule and other end-of-syscall issues
1324 userret(lp, frame, sticks);
1327 if (KTRPOINT(td, KTR_SYSRET)) {
1328 ktrsysret(lp, code, error, args.sysmsg_result);
1333 * This works because errno is findable through the
1334 * register set. If we ever support an emulation where this
1335 * is not the case, this code will need to be revisited.
1337 STOPEVENT(p, S_SCX, code);
1340 KTR_LOG(kernentry_syscall_ret, p->p_pid, lp->lwp_tid, error);
1342 KASSERT(crit_count == td->td_critcount,
1343 ("syscall: critical section count mismatch! %d/%d",
1344 crit_count, td->td_pri));
1345 KASSERT(&td->td_toks_base == td->td_toks_stop,
1346 ("syscall: %ld extra tokens held after trap! syscall %p",
1347 td->td_toks_stop - &td->td_toks_base,
1353 fork_return(struct lwp *lp, struct trapframe *frame)
1355 frame->tf_rax = 0; /* Child returns zero */
1356 frame->tf_rflags &= ~PSL_C; /* success */
1359 generic_lwp_return(lp, frame);
1360 KTR_LOG(kernentry_fork_ret, lp->lwp_proc->p_pid, lp->lwp_tid);
1364 * Simplified back end of syscall(), used when returning from fork()
1365 * directly into user mode.
1367 * This code will return back into the fork trampoline code which then
1371 generic_lwp_return(struct lwp *lp, struct trapframe *frame)
1373 struct proc *p = lp->lwp_proc;
1376 * Check for exit-race. If one lwp exits the process concurrent with
1377 * another lwp creating a new thread, the two operations may cross
1378 * each other resulting in the newly-created lwp not receiving a
1381 if (p->p_flags & P_WEXIT) {
1382 lwpsignal(p, lp, SIGKILL);
1386 * Newly forked processes are given a kernel priority. We have to
1387 * adjust the priority to a normal user priority and fake entry
1388 * into the kernel (call userenter()) to install a passive release
1389 * function just in case userret() decides to stop the process. This
1390 * can occur when ^Z races a fork. If we do not install the passive
1391 * release function the current process designation will not be
1392 * released when the thread goes to sleep.
1394 lwkt_setpri_self(TDPRI_USER_NORM);
1395 userenter(lp->lwp_thread, p);
1396 userret(lp, frame, 0);
1398 if (KTRPOINT(lp->lwp_thread, KTR_SYSRET))
1399 ktrsysret(lp, SYS_fork, 0, 0);
1401 lp->lwp_flags |= LWP_PASSIVE_ACQ;
1403 lp->lwp_flags &= ~LWP_PASSIVE_ACQ;
1407 * If PGEX_FPFAULT is set then set FP_VIRTFP in the PCB to force a T_DNA
1408 * fault (which is then passed back to the virtual kernel) if an attempt is
1409 * made to use the FP unit.
1411 * XXX this is a fairly big hack.
1414 set_vkernel_fp(struct trapframe *frame)
1416 struct thread *td = curthread;
1418 if (frame->tf_xflags & PGEX_FPFAULT) {
1419 td->td_pcb->pcb_flags |= FP_VIRTFP;
1420 if (mdcpu->gd_npxthread == td)
1423 td->td_pcb->pcb_flags &= ~FP_VIRTFP;
1428 * Called from vkernel_trap() to fixup the vkernel's syscall
1429 * frame for vmspace_ctl() return.
1432 cpu_vkernel_trap(struct trapframe *frame, int error)
1434 frame->tf_rax = error;
1436 frame->tf_rflags |= PSL_C;
1438 frame->tf_rflags &= ~PSL_C;