1 /*-
2 * SPDX-License-Identifier: BSD-4-Clause
3 *
4 * Copyright (C) 1994, David Greenman
5 * Copyright (c) 1990, 1993
6 * The Regents of the University of California. All rights reserved.
7 *
8 * This code is derived from software contributed to Berkeley by
9 * the University of Utah, and William Jolitz.
10 *
11 * Redistribution and use in source and binary forms, with or without
12 * modification, are permitted provided that the following conditions
13 * are met:
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.
26 *
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
37 * SUCH DAMAGE.
38 */
39
40 #include <sys/cdefs.h>
41 /*
42 * AMD64 Trap and System call handling
43 */
44
45 #include "opt_clock.h"
46 #include "opt_cpu.h"
47 #include "opt_hwpmc_hooks.h"
48 #include "opt_isa.h"
49 #include "opt_kdb.h"
50
51 #include <sys/param.h>
52 #include <sys/asan.h>
53 #include <sys/bus.h>
54 #include <sys/systm.h>
55 #include <sys/proc.h>
56 #include <sys/ptrace.h>
57 #include <sys/kdb.h>
58 #include <sys/kernel.h>
59 #include <sys/ktr.h>
60 #include <sys/lock.h>
61 #include <sys/msan.h>
62 #include <sys/mutex.h>
63 #include <sys/resourcevar.h>
64 #include <sys/signalvar.h>
65 #include <sys/syscall.h>
66 #include <sys/sysctl.h>
67 #include <sys/sysent.h>
68 #include <sys/uio.h>
69 #include <sys/vmmeter.h>
70 #ifdef HWPMC_HOOKS
71 #include <sys/pmckern.h>
72 PMC_SOFT_DEFINE( , , page_fault, all);
73 PMC_SOFT_DEFINE( , , page_fault, read);
74 PMC_SOFT_DEFINE( , , page_fault, write);
75 #endif
76
77 #include <vm/vm.h>
78 #include <vm/vm_param.h>
79 #include <vm/pmap.h>
80 #include <vm/vm_kern.h>
81 #include <vm/vm_map.h>
82 #include <vm/vm_page.h>
83 #include <vm/vm_extern.h>
84
85 #include <machine/cpu.h>
86 #include <machine/intr_machdep.h>
87 #include <x86/mca.h>
88 #include <machine/md_var.h>
89 #include <machine/pcb.h>
90 #ifdef SMP
91 #include <machine/smp.h>
92 #endif
93 #include <machine/stack.h>
94 #include <machine/trap.h>
95 #include <machine/tss.h>
96
97 #ifdef KDTRACE_HOOKS
98 #include <sys/dtrace_bsd.h>
99 #endif
100
101 extern inthand_t IDTVEC(bpt), IDTVEC(bpt_pti), IDTVEC(dbg),
102 IDTVEC(fast_syscall), IDTVEC(fast_syscall_pti), IDTVEC(fast_syscall32),
103 IDTVEC(int0x80_syscall_pti), IDTVEC(int0x80_syscall);
104
105 void __noinline trap(struct trapframe *frame);
106 void trap_check(struct trapframe *frame);
107 void dblfault_handler(struct trapframe *frame);
108
109 static int trap_pfault(struct trapframe *, bool, int *, int *);
110 static void trap_diag(struct trapframe *, vm_offset_t);
111 static void trap_fatal(struct trapframe *, vm_offset_t);
112 #ifdef KDTRACE_HOOKS
113 static bool trap_user_dtrace(struct trapframe *,
114 int (**hook)(struct trapframe *));
115 #endif
116
117 static const char UNKNOWN[] = "unknown";
118 static const char *const trap_msg[] = {
119 [0] = UNKNOWN, /* unused */
120 [T_PRIVINFLT] = "privileged instruction fault",
121 [2] = UNKNOWN, /* unused */
122 [T_BPTFLT] = "breakpoint instruction fault",
123 [4] = UNKNOWN, /* unused */
124 [5] = UNKNOWN, /* unused */
125 [T_ARITHTRAP] = "arithmetic trap",
126 [7] = UNKNOWN, /* unused */
127 [8] = UNKNOWN, /* unused */
128 [T_PROTFLT] = "general protection fault",
129 [T_TRCTRAP] = "debug exception",
130 [11] = UNKNOWN, /* unused */
131 [T_PAGEFLT] = "page fault",
132 [13] = UNKNOWN, /* unused */
133 [T_ALIGNFLT] = "alignment fault",
134 [15] = UNKNOWN, /* unused */
135 [16] = UNKNOWN, /* unused */
136 [17] = UNKNOWN, /* unused */
137 [T_DIVIDE] = "integer divide fault",
138 [T_NMI] = "non-maskable interrupt trap",
139 [T_OFLOW] = "overflow trap",
140 [T_BOUND] = "FPU bounds check fault",
141 [T_DNA] = "FPU device not available",
142 [T_DOUBLEFLT] = "double fault",
143 [T_FPOPFLT] = "FPU operand fetch fault",
144 [T_TSSFLT] = "invalid TSS fault",
145 [T_SEGNPFLT] = "segment not present fault",
146 [T_STKFLT] = "stack fault",
147 [T_MCHK] = "machine check trap",
148 [T_XMMFLT] = "SIMD floating-point exception",
149 [T_RESERVED] = "reserved (unknown) fault",
150 [31] = UNKNOWN, /* reserved */
151 [T_DTRACE_RET] = "DTrace pid return trap",
152 };
153
154 static const char *
traptype_to_msg(u_int type)155 traptype_to_msg(u_int type)
156 {
157 return (type < nitems(trap_msg) ? trap_msg[type] :
158 "unknown/reserved trap");
159 }
160
161 static int uprintf_signal;
162 SYSCTL_INT(_machdep, OID_AUTO, uprintf_signal, CTLFLAG_RWTUN,
163 &uprintf_signal, 0,
164 "Print debugging information on trap signal to ctty");
165
166 u_long cnt_efirt_faults;
167 int print_efirt_faults = 1;
168
169 /*
170 * Control L1D flush on return from NMI.
171 *
172 * Tunable can be set to the following values:
173 * 0 - only enable flush on return from NMI if required by vmm.ko (default)
174 * >1 - always flush on return from NMI.
175 *
176 * Post-boot, the sysctl indicates if flushing is currently enabled.
177 */
178 int nmi_flush_l1d_sw;
179 SYSCTL_INT(_machdep, OID_AUTO, nmi_flush_l1d_sw, CTLFLAG_RWTUN,
180 &nmi_flush_l1d_sw, 0,
181 "Flush L1 Data Cache on NMI exit, software bhyve L1TF mitigation assist");
182
183 /*
184 * Table of handlers for various segment load faults.
185 */
186 static const struct {
187 uintptr_t faddr;
188 uintptr_t fhandler;
189 } sfhandlers[] = {
190 {
191 .faddr = (uintptr_t)ld_ds,
192 .fhandler = (uintptr_t)ds_load_fault,
193 },
194 {
195 .faddr = (uintptr_t)ld_es,
196 .fhandler = (uintptr_t)es_load_fault,
197 },
198 {
199 .faddr = (uintptr_t)ld_fs,
200 .fhandler = (uintptr_t)fs_load_fault,
201 },
202 {
203 .faddr = (uintptr_t)ld_gs,
204 .fhandler = (uintptr_t)gs_load_fault,
205 },
206 {
207 .faddr = (uintptr_t)ld_gsbase,
208 .fhandler = (uintptr_t)gsbase_load_fault
209 },
210 {
211 .faddr = (uintptr_t)ld_fsbase,
212 .fhandler = (uintptr_t)fsbase_load_fault,
213 },
214 };
215
216 /*
217 * Exception, fault, and trap interface to the FreeBSD kernel.
218 * This common code is called from assembly language IDT gate entry
219 * routines that prepare a suitable stack frame, and restore this
220 * frame after the exception has been processed.
221 */
222
223 void
trap(struct trapframe * frame)224 trap(struct trapframe *frame)
225 {
226 ksiginfo_t ksi;
227 struct thread *td;
228 struct proc *p;
229 register_t addr, dr6;
230 size_t i;
231 int pf, signo, ucode;
232 u_int type;
233
234 td = curthread;
235 p = td->td_proc;
236 dr6 = 0;
237
238 kasan_mark(frame, sizeof(*frame), sizeof(*frame), 0);
239 kmsan_mark(frame, sizeof(*frame), KMSAN_STATE_INITED);
240
241 VM_CNT_INC(v_trap);
242 type = frame->tf_trapno;
243
244 #ifdef KDB
245 if (kdb_active) {
246 kdb_reenter();
247 return;
248 }
249 #endif
250 if (type == T_NMI) {
251 nmi_handle_intr(frame);
252 return;
253 }
254
255 if (type == T_RESERVED) {
256 trap_fatal(frame, 0);
257 return;
258 }
259
260 if ((frame->tf_rflags & PSL_I) == 0) {
261 /*
262 * Buggy application or kernel code has disabled
263 * interrupts and then trapped. Enabling interrupts
264 * now is wrong, but it is better than running with
265 * interrupts disabled until they are accidentally
266 * enabled later.
267 */
268 if (TRAPF_USERMODE(frame)) {
269 uprintf(
270 "pid %ld (%s): trap %d (%s) "
271 "with interrupts disabled\n",
272 (long)curproc->p_pid, curthread->td_name, type,
273 trap_msg[type]);
274 } else {
275 switch (type) {
276 case T_NMI:
277 case T_BPTFLT:
278 case T_TRCTRAP:
279 case T_PROTFLT:
280 case T_SEGNPFLT:
281 case T_STKFLT:
282 break;
283 default:
284 printf(
285 "kernel trap %d with interrupts disabled\n",
286 type);
287
288 /*
289 * We shouldn't enable interrupts while holding a
290 * spin lock.
291 */
292 if (td->td_md.md_spinlock_count == 0)
293 enable_intr();
294 }
295 }
296 }
297
298 if (TRAPF_USERMODE(frame)) {
299 /* user trap */
300
301 td->td_pticks = 0;
302 td->td_frame = frame;
303 addr = frame->tf_rip;
304 if (td->td_cowgen != atomic_load_int(&p->p_cowgen))
305 thread_cow_update(td);
306
307 switch (type) {
308 case T_PRIVINFLT: /* privileged instruction fault */
309 signo = SIGILL;
310 ucode = ILL_PRVOPC;
311 break;
312
313 case T_BPTFLT: /* bpt instruction fault */
314 #ifdef KDTRACE_HOOKS
315 if (trap_user_dtrace(frame, &dtrace_pid_probe_ptr))
316 return;
317 #else
318 enable_intr();
319 #endif
320 signo = SIGTRAP;
321 ucode = TRAP_BRKPT;
322 break;
323
324 case T_TRCTRAP: /* debug exception */
325 enable_intr();
326 signo = SIGTRAP;
327 ucode = TRAP_TRACE;
328 dr6 = rdr6();
329 if ((dr6 & DBREG_DR6_BS) != 0) {
330 PROC_LOCK(td->td_proc);
331 if ((td->td_dbgflags & TDB_STEP) != 0) {
332 td->td_frame->tf_rflags &= ~PSL_T;
333 td->td_dbgflags &= ~TDB_STEP;
334 }
335 PROC_UNLOCK(td->td_proc);
336 }
337 break;
338
339 case T_ARITHTRAP: /* arithmetic trap */
340 ucode = fputrap_x87();
341 if (ucode == -1)
342 return;
343 signo = SIGFPE;
344 break;
345
346 case T_PROTFLT: /* general protection fault */
347 signo = SIGBUS;
348 ucode = BUS_OBJERR;
349 break;
350 case T_STKFLT: /* stack fault */
351 case T_SEGNPFLT: /* segment not present fault */
352 signo = SIGBUS;
353 ucode = BUS_ADRERR;
354 break;
355 case T_TSSFLT: /* invalid TSS fault */
356 signo = SIGBUS;
357 ucode = BUS_OBJERR;
358 break;
359 case T_ALIGNFLT:
360 signo = SIGBUS;
361 ucode = BUS_ADRALN;
362 break;
363 case T_DOUBLEFLT: /* double fault */
364 default:
365 signo = SIGBUS;
366 ucode = BUS_OBJERR;
367 break;
368
369 case T_PAGEFLT: /* page fault */
370 /*
371 * Can emulator handle this trap?
372 */
373 if (*p->p_sysent->sv_trap != NULL &&
374 (*p->p_sysent->sv_trap)(td) == 0)
375 return;
376
377 pf = trap_pfault(frame, true, &signo, &ucode);
378 if (pf == -1)
379 return;
380 if (pf == 0)
381 goto userret;
382 addr = frame->tf_addr;
383 break;
384
385 case T_DIVIDE: /* integer divide fault */
386 ucode = FPE_INTDIV;
387 signo = SIGFPE;
388 break;
389
390 case T_OFLOW: /* integer overflow fault */
391 ucode = FPE_INTOVF;
392 signo = SIGFPE;
393 break;
394
395 case T_BOUND: /* bounds check fault */
396 ucode = FPE_FLTSUB;
397 signo = SIGFPE;
398 break;
399
400 case T_DNA:
401 /* transparent fault (due to context switch "late") */
402 KASSERT(PCB_USER_FPU(td->td_pcb),
403 ("kernel FPU ctx has leaked"));
404 fpudna();
405 return;
406
407 case T_FPOPFLT: /* FPU operand fetch fault */
408 ucode = ILL_COPROC;
409 signo = SIGILL;
410 break;
411
412 case T_XMMFLT: /* SIMD floating-point exception */
413 ucode = fputrap_sse();
414 if (ucode == -1)
415 return;
416 signo = SIGFPE;
417 break;
418 #ifdef KDTRACE_HOOKS
419 case T_DTRACE_RET:
420 (void)trap_user_dtrace(frame, &dtrace_return_probe_ptr);
421 return;
422 #endif
423 }
424 } else {
425 /* kernel trap */
426
427 KASSERT(cold || td->td_ucred != NULL,
428 ("kernel trap doesn't have ucred"));
429
430 /*
431 * Most likely, EFI RT faulted. This check prevents
432 * kdb from handling breakpoints set on the BIOS text,
433 * if such option is ever needed.
434 */
435 if ((td->td_pflags & TDP_EFIRT) != 0 &&
436 curpcb->pcb_onfault != NULL && type != T_PAGEFLT) {
437 u_long cnt = atomic_fetchadd_long(&cnt_efirt_faults, 1);
438
439 if ((print_efirt_faults == 1 && cnt == 0) ||
440 print_efirt_faults == 2) {
441 trap_diag(frame, 0);
442 printf("EFI RT fault %s\n",
443 traptype_to_msg(type));
444 }
445 frame->tf_rip = (long)curpcb->pcb_onfault;
446 return;
447 }
448
449 switch (type) {
450 case T_PAGEFLT: /* page fault */
451 (void)trap_pfault(frame, false, NULL, NULL);
452 return;
453
454 case T_DNA:
455 if (PCB_USER_FPU(td->td_pcb))
456 panic("Unregistered use of FPU in kernel");
457 fpudna();
458 return;
459
460 case T_ARITHTRAP: /* arithmetic trap */
461 case T_XMMFLT: /* SIMD floating-point exception */
462 case T_FPOPFLT: /* FPU operand fetch fault */
463 /*
464 * For now, supporting kernel handler
465 * registration for FPU traps is overkill.
466 */
467 trap_fatal(frame, 0);
468 return;
469
470 case T_STKFLT: /* stack fault */
471 case T_PROTFLT: /* general protection fault */
472 case T_SEGNPFLT: /* segment not present fault */
473 if (td->td_intr_nesting_level != 0)
474 break;
475
476 /*
477 * Invalid segment selectors and out of bounds
478 * %rip's and %rsp's can be set up in user mode.
479 * This causes a fault in kernel mode when the
480 * kernel tries to return to user mode. We want
481 * to get this fault so that we can fix the
482 * problem here and not have to check all the
483 * selectors and pointers when the user changes
484 * them.
485 *
486 * In case of PTI, the IRETQ faulted while the
487 * kernel used the pti stack, and exception
488 * frame records %rsp value pointing to that
489 * stack. If we return normally to
490 * doreti_iret_fault, the trapframe is
491 * reconstructed on pti stack, and calltrap()
492 * called on it as well. Due to the very
493 * limited pti stack size, kernel does not
494 * survive for too long. Switch to the normal
495 * thread stack for the trap handling.
496 *
497 * Magic '5' is the number of qwords occupied by
498 * the hardware trap frame.
499 */
500 if (frame->tf_rip == (long)doreti_iret) {
501 KASSERT((read_rflags() & PSL_I) == 0,
502 ("interrupts enabled"));
503 frame->tf_rip = (long)doreti_iret_fault;
504 if ((PCPU_GET(curpmap)->pm_ucr3 !=
505 PMAP_NO_CR3) &&
506 (frame->tf_rsp == (uintptr_t)PCPU_GET(
507 pti_rsp0) - 5 * sizeof(register_t))) {
508 frame->tf_rsp = PCPU_GET(rsp0) - 5 *
509 sizeof(register_t);
510 }
511 return;
512 }
513
514 for (i = 0; i < nitems(sfhandlers); i++) {
515 if (frame->tf_rip == sfhandlers[i].faddr) {
516 KASSERT((read_rflags() & PSL_I) == 0,
517 ("interrupts enabled"));
518 frame->tf_rip = sfhandlers[i].fhandler;
519 return;
520 }
521 }
522
523 if (curpcb->pcb_onfault != NULL) {
524 frame->tf_rip = (long)curpcb->pcb_onfault;
525 return;
526 }
527 break;
528
529 case T_TSSFLT:
530 /*
531 * PSL_NT can be set in user mode and isn't cleared
532 * automatically when the kernel is entered. This
533 * causes a TSS fault when the kernel attempts to
534 * `iret' because the TSS link is uninitialized. We
535 * want to get this fault so that we can fix the
536 * problem here and not every time the kernel is
537 * entered.
538 */
539 if (frame->tf_rflags & PSL_NT) {
540 frame->tf_rflags &= ~PSL_NT;
541 return;
542 }
543 break;
544
545 case T_TRCTRAP: /* debug exception */
546 /* Clear any pending debug events. */
547 dr6 = rdr6();
548 load_dr6(0);
549
550 /*
551 * Ignore debug register exceptions due to
552 * accesses in the user's address space, which
553 * can happen under several conditions such as
554 * if a user sets a watchpoint on a buffer and
555 * then passes that buffer to a system call.
556 * We still want to get TRCTRAPS for addresses
557 * in kernel space because that is useful when
558 * debugging the kernel.
559 */
560 if (user_dbreg_trap(dr6))
561 return;
562
563 /*
564 * Malicious user code can configure a debug
565 * register watchpoint to trap on data access
566 * to the top of stack and then execute 'pop
567 * %ss; int 3'. Due to exception deferral for
568 * 'pop %ss', the CPU will not interrupt 'int
569 * 3' to raise the DB# exception for the debug
570 * register but will postpone the DB# until
571 * execution of the first instruction of the
572 * BP# handler (in kernel mode). Normally the
573 * previous check would ignore DB# exceptions
574 * for watchpoints on user addresses raised in
575 * kernel mode. However, some CPU errata
576 * include cases where DB# exceptions do not
577 * properly set bits in %dr6, e.g. Haswell
578 * HSD23 and Skylake-X SKZ24.
579 *
580 * A deferred DB# can also be raised on the
581 * first instructions of system call entry
582 * points or single-step traps via similar use
583 * of 'pop %ss' or 'mov xxx, %ss'.
584 */
585 if (pti) {
586 if (frame->tf_rip ==
587 (uintptr_t)IDTVEC(fast_syscall_pti) ||
588 #ifdef COMPAT_FREEBSD32
589 frame->tf_rip ==
590 (uintptr_t)IDTVEC(int0x80_syscall_pti) ||
591 #endif
592 frame->tf_rip == (uintptr_t)IDTVEC(bpt_pti))
593 return;
594 } else {
595 if (frame->tf_rip ==
596 (uintptr_t)IDTVEC(fast_syscall) ||
597 #ifdef COMPAT_FREEBSD32
598 frame->tf_rip ==
599 (uintptr_t)IDTVEC(int0x80_syscall) ||
600 #endif
601 frame->tf_rip == (uintptr_t)IDTVEC(bpt))
602 return;
603 }
604 if (frame->tf_rip == (uintptr_t)IDTVEC(dbg) ||
605 /* Needed for AMD. */
606 frame->tf_rip == (uintptr_t)IDTVEC(fast_syscall32))
607 return;
608 /*
609 * FALLTHROUGH (TRCTRAP kernel mode, kernel address)
610 */
611 case T_BPTFLT:
612 /*
613 * If KDB is enabled, let it handle the debugger trap.
614 * Otherwise, debugger traps "can't happen".
615 */
616 #ifdef KDB
617 if (kdb_trap(type, dr6, frame))
618 return;
619 #endif
620 break;
621 }
622
623 trap_fatal(frame, 0);
624 return;
625 }
626
627 ksiginfo_init_trap(&ksi);
628 ksi.ksi_signo = signo;
629 ksi.ksi_code = ucode;
630 ksi.ksi_trapno = type;
631 ksi.ksi_addr = (void *)addr;
632 if (uprintf_signal) {
633 uprintf("pid %d comm %s: signal %d err %#lx code %d type %d "
634 "addr %#lx rsp %#lx rip %#lx rax %#lx "
635 "<%02x %02x %02x %02x %02x %02x %02x %02x>\n",
636 p->p_pid, p->p_comm, signo, frame->tf_err, ucode, type,
637 addr, frame->tf_rsp, frame->tf_rip, frame->tf_rax,
638 fubyte((void *)(frame->tf_rip + 0)),
639 fubyte((void *)(frame->tf_rip + 1)),
640 fubyte((void *)(frame->tf_rip + 2)),
641 fubyte((void *)(frame->tf_rip + 3)),
642 fubyte((void *)(frame->tf_rip + 4)),
643 fubyte((void *)(frame->tf_rip + 5)),
644 fubyte((void *)(frame->tf_rip + 6)),
645 fubyte((void *)(frame->tf_rip + 7)));
646 }
647 KASSERT((read_rflags() & PSL_I) != 0, ("interrupts disabled"));
648 trapsignal(td, &ksi);
649
650 userret:
651 userret(td, frame);
652 KASSERT(PCB_USER_FPU(td->td_pcb),
653 ("Return from trap with kernel FPU ctx leaked"));
654 }
655
656 /*
657 * Ensure that we ignore any DTrace-induced faults. This function cannot
658 * be instrumented, so it cannot generate such faults itself.
659 */
660 void
trap_check(struct trapframe * frame)661 trap_check(struct trapframe *frame)
662 {
663
664 #ifdef KDTRACE_HOOKS
665 if (dtrace_trap_func != NULL &&
666 (*dtrace_trap_func)(frame, frame->tf_trapno) != 0)
667 return;
668 #endif
669 trap(frame);
670 }
671
672 static bool
trap_is_smap(struct trapframe * frame)673 trap_is_smap(struct trapframe *frame)
674 {
675
676 /*
677 * A page fault on a userspace address is classified as
678 * SMAP-induced if:
679 * - SMAP is supported;
680 * - kernel mode accessed present data page;
681 * - rflags.AC was cleared.
682 * Kernel must never access user space with rflags.AC cleared
683 * if SMAP is enabled.
684 */
685 return ((cpu_stdext_feature & CPUID_STDEXT_SMAP) != 0 &&
686 (frame->tf_err & (PGEX_P | PGEX_U | PGEX_I | PGEX_RSV)) ==
687 PGEX_P && (frame->tf_rflags & PSL_AC) == 0);
688 }
689
690 static bool
trap_is_pti(struct trapframe * frame)691 trap_is_pti(struct trapframe *frame)
692 {
693
694 return (PCPU_GET(curpmap)->pm_ucr3 != PMAP_NO_CR3 &&
695 pg_nx != 0 && (frame->tf_err & (PGEX_P | PGEX_W |
696 PGEX_U | PGEX_I)) == (PGEX_P | PGEX_U | PGEX_I) &&
697 (curpcb->pcb_saved_ucr3 & ~CR3_PCID_MASK) ==
698 (PCPU_GET(curpmap)->pm_cr3 & ~CR3_PCID_MASK));
699 }
700
701 /*
702 * Handle all details of a page fault.
703 * Returns:
704 * -1 if this fault was fatal, typically from kernel mode
705 * (cannot happen, but we need to return something).
706 * 0 if this fault was handled by updating either the user or kernel
707 * page table, execution can continue.
708 * 1 if this fault was from usermode and it was not handled, a synchronous
709 * signal should be delivered to the thread. *signo returns the signal
710 * number, *ucode gives si_code.
711 */
712 static int
trap_pfault(struct trapframe * frame,bool usermode,int * signo,int * ucode)713 trap_pfault(struct trapframe *frame, bool usermode, int *signo, int *ucode)
714 {
715 struct thread *td;
716 struct proc *p;
717 vm_map_t map;
718 vm_offset_t eva;
719 int rv;
720 vm_prot_t ftype;
721
722 MPASS(!usermode || (signo != NULL && ucode != NULL));
723
724 td = curthread;
725 p = td->td_proc;
726 eva = frame->tf_addr;
727
728 if (__predict_false((td->td_pflags & TDP_NOFAULTING) != 0)) {
729 /*
730 * Due to both processor errata and lazy TLB invalidation when
731 * access restrictions are removed from virtual pages, memory
732 * accesses that are allowed by the physical mapping layer may
733 * nonetheless cause one spurious page fault per virtual page.
734 * When the thread is executing a "no faulting" section that
735 * is bracketed by vm_fault_{disable,enable}_pagefaults(),
736 * every page fault is treated as a spurious page fault,
737 * unless it accesses the same virtual address as the most
738 * recent page fault within the same "no faulting" section.
739 */
740 if (td->td_md.md_spurflt_addr != eva ||
741 (td->td_pflags & TDP_RESETSPUR) != 0) {
742 /*
743 * Do nothing to the TLB. A stale TLB entry is
744 * flushed automatically by a page fault.
745 */
746 td->td_md.md_spurflt_addr = eva;
747 td->td_pflags &= ~TDP_RESETSPUR;
748 return (0);
749 }
750 } else {
751 /*
752 * If we get a page fault while in a critical section, then
753 * it is most likely a fatal kernel page fault. The kernel
754 * is already going to panic trying to get a sleep lock to
755 * do the VM lookup, so just consider it a fatal trap so the
756 * kernel can print out a useful trap message and even get
757 * to the debugger.
758 *
759 * If we get a page fault while holding a non-sleepable
760 * lock, then it is most likely a fatal kernel page fault.
761 * If WITNESS is enabled, then it's going to whine about
762 * bogus LORs with various VM locks, so just skip to the
763 * fatal trap handling directly.
764 */
765 if (td->td_critnest != 0 ||
766 WITNESS_CHECK(WARN_SLEEPOK | WARN_GIANTOK, NULL,
767 "Kernel page fault") != 0) {
768 trap_fatal(frame, eva);
769 return (-1);
770 }
771 }
772 if (eva >= VM_MIN_KERNEL_ADDRESS) {
773 /*
774 * Don't allow user-mode faults in kernel address space.
775 */
776 if (usermode) {
777 *signo = SIGSEGV;
778 *ucode = SEGV_MAPERR;
779 return (1);
780 }
781
782 map = kernel_map;
783 } else {
784 map = &p->p_vmspace->vm_map;
785
786 /*
787 * When accessing a usermode address, kernel must be
788 * ready to accept the page fault, and provide a
789 * handling routine. Since accessing the address
790 * without the handler is a bug, do not try to handle
791 * it normally, and panic immediately.
792 *
793 * If SMAP is enabled, filter SMAP faults also,
794 * because illegal access might occur to the mapped
795 * user address, causing infinite loop.
796 */
797 if (!usermode && (td->td_intr_nesting_level != 0 ||
798 trap_is_smap(frame) || curpcb->pcb_onfault == NULL)) {
799 trap_fatal(frame, eva);
800 return (-1);
801 }
802 }
803
804 /*
805 * If the trap was caused by errant bits in the PTE then panic.
806 */
807 if (frame->tf_err & PGEX_RSV) {
808 trap_fatal(frame, eva);
809 return (-1);
810 }
811
812 /*
813 * User-mode protection key violation (PKU). May happen
814 * either from usermode or from kernel if copyin accessed
815 * key-protected mapping.
816 */
817 if ((frame->tf_err & PGEX_PK) != 0) {
818 if (eva > VM_MAXUSER_ADDRESS) {
819 trap_fatal(frame, eva);
820 return (-1);
821 }
822 if (usermode) {
823 *signo = SIGSEGV;
824 *ucode = SEGV_PKUERR;
825 return (1);
826 }
827 goto after_vmfault;
828 }
829
830 /*
831 * If nx protection of the usermode portion of kernel page
832 * tables caused trap, panic.
833 */
834 if (usermode && trap_is_pti(frame))
835 panic("PTI: pid %d comm %s tf_err %#lx", p->p_pid,
836 p->p_comm, frame->tf_err);
837
838 /*
839 * PGEX_I is defined only if the execute disable bit capability is
840 * supported and enabled.
841 */
842 if (frame->tf_err & PGEX_W)
843 ftype = VM_PROT_WRITE;
844 else if ((frame->tf_err & PGEX_I) && pg_nx != 0)
845 ftype = VM_PROT_EXECUTE;
846 else
847 ftype = VM_PROT_READ;
848
849 /* Fault in the page. */
850 rv = vm_fault_trap(map, eva, ftype, VM_FAULT_NORMAL, signo, ucode);
851 if (rv == KERN_SUCCESS) {
852 #ifdef HWPMC_HOOKS
853 if (ftype == VM_PROT_READ || ftype == VM_PROT_WRITE) {
854 PMC_SOFT_CALL_TF( , , page_fault, all, frame);
855 if (ftype == VM_PROT_READ)
856 PMC_SOFT_CALL_TF( , , page_fault, read,
857 frame);
858 else
859 PMC_SOFT_CALL_TF( , , page_fault, write,
860 frame);
861 }
862 #endif
863 return (0);
864 }
865
866 if (usermode)
867 return (1);
868 after_vmfault:
869 if (td->td_intr_nesting_level == 0 &&
870 curpcb->pcb_onfault != NULL) {
871 if ((td->td_pflags & TDP_EFIRT) != 0) {
872 u_long cnt = atomic_fetchadd_long(&cnt_efirt_faults, 1);
873
874 if ((print_efirt_faults == 1 && cnt == 0) ||
875 print_efirt_faults == 2) {
876 trap_diag(frame, eva);
877 printf("EFI RT page fault\n");
878 }
879 }
880 frame->tf_rip = (long)curpcb->pcb_onfault;
881 return (0);
882 }
883 trap_fatal(frame, eva);
884 return (-1);
885 }
886
887 static void
trap_diag(struct trapframe * frame,vm_offset_t eva)888 trap_diag(struct trapframe *frame, vm_offset_t eva)
889 {
890 int code, ss;
891 u_int type;
892 struct soft_segment_descriptor softseg;
893 struct user_segment_descriptor *gdt;
894
895 code = frame->tf_err;
896 type = frame->tf_trapno;
897 gdt = *PCPU_PTR(gdt);
898 sdtossd(&gdt[IDXSEL(frame->tf_cs & 0xffff)], &softseg);
899
900 printf("\n\nFatal trap %d: %s while in %s mode\n", type,
901 type < nitems(trap_msg) ? trap_msg[type] : UNKNOWN,
902 TRAPF_USERMODE(frame) ? "user" : "kernel");
903 #ifdef SMP
904 /* two separate prints in case of a trap on an unmapped page */
905 printf("cpuid = %d; ", PCPU_GET(cpuid));
906 printf("apic id = %02x\n", PCPU_GET(apic_id));
907 #endif
908 if (type == T_PAGEFLT) {
909 printf("fault virtual address = 0x%lx\n", eva);
910 printf("fault code = %s %s %s%s%s, %s\n",
911 code & PGEX_U ? "user" : "supervisor",
912 code & PGEX_W ? "write" : "read",
913 code & PGEX_I ? "instruction" : "data",
914 code & PGEX_PK ? " prot key" : "",
915 code & PGEX_SGX ? " SGX" : "",
916 code & PGEX_RSV ? "reserved bits in PTE" :
917 code & PGEX_P ? "protection violation" : "page not present");
918 }
919 printf("instruction pointer = 0x%lx:0x%lx\n",
920 frame->tf_cs & 0xffff, frame->tf_rip);
921 ss = frame->tf_ss & 0xffff;
922 printf("stack pointer = 0x%x:0x%lx\n", ss, frame->tf_rsp);
923 printf("frame pointer = 0x%x:0x%lx\n", ss, frame->tf_rbp);
924 printf("code segment = base 0x%lx, limit 0x%lx, type 0x%x\n",
925 softseg.ssd_base, softseg.ssd_limit, softseg.ssd_type);
926 printf(" = DPL %d, pres %d, long %d, def32 %d, gran %d\n",
927 softseg.ssd_dpl, softseg.ssd_p, softseg.ssd_long, softseg.ssd_def32,
928 softseg.ssd_gran);
929 printf("processor eflags = ");
930 if (frame->tf_rflags & PSL_T)
931 printf("trace trap, ");
932 if (frame->tf_rflags & PSL_I)
933 printf("interrupt enabled, ");
934 if (frame->tf_rflags & PSL_NT)
935 printf("nested task, ");
936 if (frame->tf_rflags & PSL_RF)
937 printf("resume, ");
938 printf("IOPL = %ld\n", (frame->tf_rflags & PSL_IOPL) >> 12);
939 printf("current process = %d (%s)\n",
940 curproc->p_pid, curthread->td_name);
941
942 printf("rdi: %016lx rsi: %016lx rdx: %016lx\n", frame->tf_rdi,
943 frame->tf_rsi, frame->tf_rdx);
944 printf("rcx: %016lx r8: %016lx r9: %016lx\n", frame->tf_rcx,
945 frame->tf_r8, frame->tf_r9);
946 printf("rax: %016lx rbx: %016lx rbp: %016lx\n", frame->tf_rax,
947 frame->tf_rbx, frame->tf_rbp);
948 printf("r10: %016lx r11: %016lx r12: %016lx\n", frame->tf_r10,
949 frame->tf_r11, frame->tf_r12);
950 printf("r13: %016lx r14: %016lx r15: %016lx\n", frame->tf_r13,
951 frame->tf_r14, frame->tf_r15);
952
953 printf("trap number = %d\n", type);
954 }
955
956 static void
trap_fatal(struct trapframe * frame,vm_offset_t eva)957 trap_fatal(struct trapframe *frame, vm_offset_t eva)
958 {
959 u_int type;
960
961 type = frame->tf_trapno;
962 trap_diag(frame, eva);
963 #ifdef KDB
964 if (debugger_on_trap) {
965 bool handled;
966
967 kdb_why = KDB_WHY_TRAP;
968 handled = kdb_trap(type, 0, frame);
969 kdb_why = KDB_WHY_UNSET;
970 if (handled)
971 return;
972 }
973 #endif
974 panic("%s", traptype_to_msg(type));
975 }
976
977 #ifdef KDTRACE_HOOKS
978 /*
979 * Invoke a userspace DTrace hook. The hook pointer is cleared when no
980 * userspace probes are enabled, so we must synchronize with DTrace to ensure
981 * that a trapping thread is able to call the hook before it is cleared.
982 */
983 static bool
trap_user_dtrace(struct trapframe * frame,int (** hookp)(struct trapframe *))984 trap_user_dtrace(struct trapframe *frame, int (**hookp)(struct trapframe *))
985 {
986 int (*hook)(struct trapframe *);
987
988 hook = atomic_load_ptr(hookp);
989 enable_intr();
990 if (hook != NULL)
991 return ((hook)(frame) == 0);
992 return (false);
993 }
994 #endif
995
996 /*
997 * Double fault handler. Called when a fault occurs while writing
998 * a frame for a trap/exception onto the stack. This usually occurs
999 * when the stack overflows (such is the case with infinite recursion,
1000 * for example).
1001 */
1002 void
dblfault_handler(struct trapframe * frame)1003 dblfault_handler(struct trapframe *frame)
1004 {
1005 kmsan_mark(frame, sizeof(*frame), KMSAN_STATE_INITED);
1006 #ifdef KDTRACE_HOOKS
1007 if (dtrace_doubletrap_func != NULL)
1008 (*dtrace_doubletrap_func)();
1009 #endif
1010 printf("\nFatal double fault\n"
1011 "rip %#lx rsp %#lx rbp %#lx\n"
1012 "rax %#lx rdx %#lx rbx %#lx\n"
1013 "rcx %#lx rsi %#lx rdi %#lx\n"
1014 "r8 %#lx r9 %#lx r10 %#lx\n"
1015 "r11 %#lx r12 %#lx r13 %#lx\n"
1016 "r14 %#lx r15 %#lx rflags %#lx\n"
1017 "cs %#lx ss %#lx ds %#hx es %#hx fs %#hx gs %#hx\n"
1018 "fsbase %#lx gsbase %#lx kgsbase %#lx\n",
1019 frame->tf_rip, frame->tf_rsp, frame->tf_rbp,
1020 frame->tf_rax, frame->tf_rdx, frame->tf_rbx,
1021 frame->tf_rcx, frame->tf_rdi, frame->tf_rsi,
1022 frame->tf_r8, frame->tf_r9, frame->tf_r10,
1023 frame->tf_r11, frame->tf_r12, frame->tf_r13,
1024 frame->tf_r14, frame->tf_r15, frame->tf_rflags,
1025 frame->tf_cs, frame->tf_ss, frame->tf_ds, frame->tf_es,
1026 frame->tf_fs, frame->tf_gs,
1027 rdmsr(MSR_FSBASE), rdmsr(MSR_GSBASE), rdmsr(MSR_KGSBASE));
1028 #ifdef SMP
1029 /* two separate prints in case of a trap on an unmapped page */
1030 printf("cpuid = %d; ", PCPU_GET(cpuid));
1031 printf("apic id = %02x\n", PCPU_GET(apic_id));
1032 #endif
1033 panic("double fault");
1034 }
1035
1036 static int __noinline
cpu_fetch_syscall_args_fallback(struct thread * td,struct syscall_args * sa)1037 cpu_fetch_syscall_args_fallback(struct thread *td, struct syscall_args *sa)
1038 {
1039 struct proc *p;
1040 struct trapframe *frame;
1041 syscallarg_t *argp;
1042 caddr_t params;
1043 int reg, regcnt, error;
1044
1045 p = td->td_proc;
1046 frame = td->td_frame;
1047 reg = 0;
1048 regcnt = NARGREGS;
1049
1050 if (sa->code == SYS_syscall || sa->code == SYS___syscall) {
1051 sa->code = frame->tf_rdi;
1052 reg++;
1053 regcnt--;
1054 }
1055
1056 if (sa->code >= p->p_sysent->sv_size)
1057 sa->callp = &nosys_sysent;
1058 else
1059 sa->callp = &p->p_sysent->sv_table[sa->code];
1060
1061 KASSERT(sa->callp->sy_narg <= nitems(sa->args),
1062 ("Too many syscall arguments!"));
1063 argp = &frame->tf_rdi;
1064 argp += reg;
1065 memcpy(sa->args, argp, sizeof(sa->args[0]) * NARGREGS);
1066 if (sa->callp->sy_narg > regcnt) {
1067 params = (caddr_t)frame->tf_rsp + sizeof(register_t);
1068 error = copyin(params, &sa->args[regcnt],
1069 (sa->callp->sy_narg - regcnt) * sizeof(sa->args[0]));
1070 if (__predict_false(error != 0))
1071 return (error);
1072 }
1073
1074 td->td_retval[0] = 0;
1075 td->td_retval[1] = frame->tf_rdx;
1076
1077 return (0);
1078 }
1079
1080 int
cpu_fetch_syscall_args(struct thread * td)1081 cpu_fetch_syscall_args(struct thread *td)
1082 {
1083 struct proc *p;
1084 struct trapframe *frame;
1085 struct syscall_args *sa;
1086
1087 p = td->td_proc;
1088 frame = td->td_frame;
1089 sa = &td->td_sa;
1090
1091 sa->code = frame->tf_rax;
1092 sa->original_code = sa->code;
1093
1094 if (__predict_false(sa->code == SYS_syscall ||
1095 sa->code == SYS___syscall ||
1096 sa->code >= p->p_sysent->sv_size))
1097 return (cpu_fetch_syscall_args_fallback(td, sa));
1098
1099 sa->callp = &p->p_sysent->sv_table[sa->code];
1100 KASSERT(sa->callp->sy_narg <= nitems(sa->args),
1101 ("Too many syscall arguments!"));
1102
1103 if (__predict_false(sa->callp->sy_narg > NARGREGS))
1104 return (cpu_fetch_syscall_args_fallback(td, sa));
1105
1106 memcpy(sa->args, &frame->tf_rdi, sizeof(sa->args[0]) * NARGREGS);
1107
1108 td->td_retval[0] = 0;
1109 td->td_retval[1] = frame->tf_rdx;
1110
1111 return (0);
1112 }
1113
1114 #include "../../kern/subr_syscall.c"
1115
1116 static void (*syscall_ret_l1d_flush)(void);
1117 int syscall_ret_l1d_flush_mode;
1118
1119 static void
flush_l1d_hw(void)1120 flush_l1d_hw(void)
1121 {
1122
1123 wrmsr(MSR_IA32_FLUSH_CMD, IA32_FLUSH_CMD_L1D);
1124 }
1125
1126 static void __noinline
amd64_syscall_ret_flush_l1d_check(int error)1127 amd64_syscall_ret_flush_l1d_check(int error)
1128 {
1129 void (*p)(void);
1130
1131 if (error != EEXIST && error != EAGAIN && error != EXDEV &&
1132 error != ENOENT && error != ENOTCONN && error != EINPROGRESS) {
1133 p = atomic_load_ptr(&syscall_ret_l1d_flush);
1134 if (p != NULL)
1135 p();
1136 }
1137 }
1138
1139 static void __inline
amd64_syscall_ret_flush_l1d_check_inline(int error)1140 amd64_syscall_ret_flush_l1d_check_inline(int error)
1141 {
1142
1143 if (__predict_false(error != 0))
1144 amd64_syscall_ret_flush_l1d_check(error);
1145 }
1146
1147 void
amd64_syscall_ret_flush_l1d(int error)1148 amd64_syscall_ret_flush_l1d(int error)
1149 {
1150
1151 amd64_syscall_ret_flush_l1d_check_inline(error);
1152 }
1153
1154 void
amd64_syscall_ret_flush_l1d_recalc(void)1155 amd64_syscall_ret_flush_l1d_recalc(void)
1156 {
1157 bool l1d_hw;
1158
1159 l1d_hw = (cpu_stdext_feature3 & CPUID_STDEXT3_L1D_FLUSH) != 0;
1160 again:
1161 switch (syscall_ret_l1d_flush_mode) {
1162 case 0:
1163 syscall_ret_l1d_flush = NULL;
1164 break;
1165 case 1:
1166 syscall_ret_l1d_flush = l1d_hw ? flush_l1d_hw :
1167 flush_l1d_sw_abi;
1168 break;
1169 case 2:
1170 syscall_ret_l1d_flush = l1d_hw ? flush_l1d_hw : NULL;
1171 break;
1172 case 3:
1173 syscall_ret_l1d_flush = flush_l1d_sw_abi;
1174 break;
1175 default:
1176 syscall_ret_l1d_flush_mode = 1;
1177 goto again;
1178 }
1179 }
1180
1181 static int
machdep_syscall_ret_flush_l1d(SYSCTL_HANDLER_ARGS)1182 machdep_syscall_ret_flush_l1d(SYSCTL_HANDLER_ARGS)
1183 {
1184 int error, val;
1185
1186 val = syscall_ret_l1d_flush_mode;
1187 error = sysctl_handle_int(oidp, &val, 0, req);
1188 if (error != 0 || req->newptr == NULL)
1189 return (error);
1190 syscall_ret_l1d_flush_mode = val;
1191 amd64_syscall_ret_flush_l1d_recalc();
1192 return (0);
1193 }
1194 SYSCTL_PROC(_machdep, OID_AUTO, syscall_ret_flush_l1d, CTLTYPE_INT |
1195 CTLFLAG_RWTUN | CTLFLAG_NOFETCH | CTLFLAG_MPSAFE, NULL, 0,
1196 machdep_syscall_ret_flush_l1d, "I",
1197 "Flush L1D on syscall return with error (0 - off, 1 - on, "
1198 "2 - use hw only, 3 - use sw only)");
1199
1200 /*
1201 * System call handler for native binaries. The trap frame is already
1202 * set up by the assembler trampoline and a pointer to it is saved in
1203 * td_frame.
1204 */
1205 void
amd64_syscall(struct thread * td,int traced)1206 amd64_syscall(struct thread *td, int traced)
1207 {
1208 ksiginfo_t ksi;
1209
1210 kmsan_mark(td->td_frame, sizeof(*td->td_frame), KMSAN_STATE_INITED);
1211
1212 KASSERT(TRAPF_USERMODE(td->td_frame),
1213 ("%s: not from user mode", __func__));
1214
1215 syscallenter(td);
1216
1217 /*
1218 * Traced syscall.
1219 */
1220 if (__predict_false(traced)) {
1221 td->td_frame->tf_rflags &= ~PSL_T;
1222 ksiginfo_init_trap(&ksi);
1223 ksi.ksi_signo = SIGTRAP;
1224 ksi.ksi_code = TRAP_TRACE;
1225 ksi.ksi_addr = (void *)td->td_frame->tf_rip;
1226 trapsignal(td, &ksi);
1227 }
1228
1229 KASSERT(PCB_USER_FPU(td->td_pcb),
1230 ("System call %s returning with kernel FPU ctx leaked",
1231 syscallname(td->td_proc, td->td_sa.code)));
1232 KASSERT(td->td_pcb->pcb_save == get_pcb_user_save_td(td),
1233 ("System call %s returning with mangled pcb_save",
1234 syscallname(td->td_proc, td->td_sa.code)));
1235 KASSERT(pmap_not_in_di(),
1236 ("System call %s returning with leaked invl_gen %lu",
1237 syscallname(td->td_proc, td->td_sa.code),
1238 td->td_md.md_invl_gen.gen));
1239
1240 syscallret(td);
1241
1242 /*
1243 * If the user-supplied value of %rip is not a canonical
1244 * address, then some CPUs will trigger a ring 0 #GP during
1245 * the sysret instruction. However, the fault handler would
1246 * execute in ring 0 with the user's %gs and %rsp which would
1247 * not be safe. Instead, use the full return path which
1248 * catches the problem safely.
1249 */
1250 if (__predict_false(td->td_frame->tf_rip >= (la57 ?
1251 VM_MAXUSER_ADDRESS_LA57 : VM_MAXUSER_ADDRESS_LA48)))
1252 set_pcb_flags(td->td_pcb, PCB_FULL_IRET);
1253
1254 amd64_syscall_ret_flush_l1d_check_inline(td->td_errno);
1255 }
1256