sys/amd64/amd64/trap.c

   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  *      from: @(#)trap.c        7.4 (Berkeley) 5/13/91
  40  */
  41
  42 #include <sys/cdefs.h>
  43 __FBSDID("$FreeBSD$");
  44
  45 /*
  46  * AMD64 Trap and System call handling
  47  */
  48
  49 #include "opt_clock.h"
  50 #include "opt_compat.h"
  51 #include "opt_cpu.h"
  52 #include "opt_hwpmc_hooks.h"
  53 #include "opt_isa.h"
  54 #include "opt_kdb.h"
  55
  56 #include <sys/param.h>
  57 #include <sys/asan.h>
  58 #include <sys/bus.h>
  59 #include <sys/systm.h>
  60 #include <sys/proc.h>
  61 #include <sys/ptrace.h>
  62 #include <sys/kdb.h>
  63 #include <sys/kernel.h>
  64 #include <sys/ktr.h>
  65 #include <sys/lock.h>
  66 #include <sys/mutex.h>
  67 #include <sys/resourcevar.h>
  68 #include <sys/signalvar.h>
  69 #include <sys/syscall.h>
  70 #include <sys/sysctl.h>
  71 #include <sys/sysent.h>
  72 #include <sys/uio.h>
  73 #include <sys/vmmeter.h>
  74 #ifdef HWPMC_HOOKS
  75 #include <sys/pmckern.h>
  76 PMC_SOFT_DEFINE( , , page_fault, all);
  77 PMC_SOFT_DEFINE( , , page_fault, read);
  78 PMC_SOFT_DEFINE( , , page_fault, write);
  79 #endif
  80
  81 #include <vm/vm.h>
  82 #include <vm/vm_param.h>
  83 #include <vm/pmap.h>
  84 #include <vm/vm_kern.h>
  85 #include <vm/vm_map.h>
  86 #include <vm/vm_page.h>
  87 #include <vm/vm_extern.h>
  88
  89 #include <machine/cpu.h>
  90 #include <machine/intr_machdep.h>
  91 #include <x86/mca.h>
  92 #include <machine/md_var.h>
  93 #include <machine/pcb.h>
  94 #ifdef SMP
  95 #include <machine/smp.h>
  96 #endif
  97 #include <machine/stack.h>
  98 #include <machine/trap.h>
  99 #include <machine/tss.h>
 100
 101 #ifdef KDTRACE_HOOKS
 102 #include <sys/dtrace_bsd.h>
 103 #endif
 104
 105 extern inthand_t IDTVEC(bpt), IDTVEC(bpt_pti), IDTVEC(dbg),
 106     IDTVEC(fast_syscall), IDTVEC(fast_syscall_pti), IDTVEC(fast_syscall32),
 107     IDTVEC(int0x80_syscall_pti), IDTVEC(int0x80_syscall);
 108
 109 void __noinline trap(struct trapframe *frame);
 110 void trap_check(struct trapframe *frame);
 111 void dblfault_handler(struct trapframe *frame);
 112
 113 static int trap_pfault(struct trapframe *, bool, int *, int *);
 114 static void trap_fatal(struct trapframe *, vm_offset_t);
 115 #ifdef KDTRACE_HOOKS
 116 static bool trap_user_dtrace(struct trapframe *,
 117     int (**hook)(struct trapframe *));
 118 #endif
 119
 120 static const char UNKNOWN[] = "unknown";
 121 static const char *const trap_msg[] = {
 122         [0] =                   UNKNOWN,                        /* unused */
 123         [T_PRIVINFLT] =         "privileged instruction fault",
 124         [2] =                   UNKNOWN,                        /* unused */
 125         [T_BPTFLT] =            "breakpoint instruction fault",
 126         [4] =                   UNKNOWN,                        /* unused */
 127         [5] =                   UNKNOWN,                        /* unused */
 128         [T_ARITHTRAP] =         "arithmetic trap",
 129         [7] =                   UNKNOWN,                        /* unused */
 130         [8] =                   UNKNOWN,                        /* unused */
 131         [T_PROTFLT] =           "general protection fault",
 132         [T_TRCTRAP] =           "debug exception",
 133         [11] =                  UNKNOWN,                        /* unused */
 134         [T_PAGEFLT] =           "page fault",
 135         [13] =                  UNKNOWN,                        /* unused */
 136         [T_ALIGNFLT] =          "alignment fault",
 137         [15] =                  UNKNOWN,                        /* unused */
 138         [16] =                  UNKNOWN,                        /* unused */
 139         [17] =                  UNKNOWN,                        /* unused */
 140         [T_DIVIDE] =            "integer divide fault",
 141         [T_NMI] =               "non-maskable interrupt trap",
 142         [T_OFLOW] =             "overflow trap",
 143         [T_BOUND] =             "FPU bounds check fault",
 144         [T_DNA] =               "FPU device not available",
 145         [T_DOUBLEFLT] =         "double fault",
 146         [T_FPOPFLT] =           "FPU operand fetch fault",
 147         [T_TSSFLT] =            "invalid TSS fault",
 148         [T_SEGNPFLT] =          "segment not present fault",
 149         [T_STKFLT] =            "stack fault",
 150         [T_MCHK] =              "machine check trap",
 151         [T_XMMFLT] =            "SIMD floating-point exception",
 152         [T_RESERVED] =          "reserved (unknown) fault",
 153         [31] =                  UNKNOWN,                        /* reserved */
 154         [T_DTRACE_RET] =        "DTrace pid return trap",
 155 };
 156
 157 static int uprintf_signal;
 158 SYSCTL_INT(_machdep, OID_AUTO, uprintf_signal, CTLFLAG_RWTUN,
 159     &uprintf_signal, 0,
 160     "Print debugging information on trap signal to ctty");
 161
 162 /*
 163  * Control L1D flush on return from NMI.
 164  *
 165  * Tunable  can be set to the following values:
 166  * 0 - only enable flush on return from NMI if required by vmm.ko (default)
 167  * >1 - always flush on return from NMI.
 168  *
 169  * Post-boot, the sysctl indicates if flushing is currently enabled.
 170  */
 171 int nmi_flush_l1d_sw;
 172 SYSCTL_INT(_machdep, OID_AUTO, nmi_flush_l1d_sw, CTLFLAG_RWTUN,
 173     &nmi_flush_l1d_sw, 0,
 174     "Flush L1 Data Cache on NMI exit, software bhyve L1TF mitigation assist");
 175
 176 /*
 177  * Table of handlers for various segment load faults.
 178  */
 179 static const struct {
 180         uintptr_t       faddr;
 181         uintptr_t       fhandler;
 182 } sfhandlers[] = {
 183         {
 184                 .faddr = (uintptr_t)ld_ds,
 185                 .fhandler = (uintptr_t)ds_load_fault,
 186         },
 187         {
 188                 .faddr = (uintptr_t)ld_es,
 189                 .fhandler = (uintptr_t)es_load_fault,
 190         },
 191         {
 192                 .faddr = (uintptr_t)ld_fs,
 193                 .fhandler = (uintptr_t)fs_load_fault,
 194         },
 195         {
 196                 .faddr = (uintptr_t)ld_gs,
 197                 .fhandler = (uintptr_t)gs_load_fault,
 198         },
 199         {
 200                 .faddr = (uintptr_t)ld_gsbase,
 201                 .fhandler = (uintptr_t)gsbase_load_fault
 202         },
 203         {
 204                 .faddr = (uintptr_t)ld_fsbase,
 205                 .fhandler = (uintptr_t)fsbase_load_fault,
 206         },
 207 };
 208
 209 /*
 210  * Exception, fault, and trap interface to the FreeBSD kernel.
 211  * This common code is called from assembly language IDT gate entry
 212  * routines that prepare a suitable stack frame, and restore this
 213  * frame after the exception has been processed.
 214  */
 215
 216 void
 217 trap(struct trapframe *frame)
 218 {
 219         ksiginfo_t ksi;
 220         struct thread *td;
 221         struct proc *p;
 222         register_t addr, dr6;
 223         size_t i;
 224         int pf, signo, ucode;
 225         u_int type;
 226
 227         td = curthread;
 228         p = td->td_proc;
 229         dr6 = 0;
 230
 231         kasan_mark(frame, sizeof(*frame), sizeof(*frame), 0);
 232
 233         VM_CNT_INC(v_trap);
 234         type = frame->tf_trapno;
 235
 236 #ifdef SMP
 237         /* Handler for NMI IPIs used for stopping CPUs. */
 238         if (type == T_NMI && ipi_nmi_handler() == 0)
 239                 return;
 240 #endif
 241
 242 #ifdef KDB
 243         if (kdb_active) {
 244                 kdb_reenter();
 245                 return;
 246         }
 247 #endif
 248
 249         if (type == T_RESERVED) {
 250                 trap_fatal(frame, 0);
 251                 return;
 252         }
 253
 254         if (type == T_NMI) {
 255 #ifdef HWPMC_HOOKS
 256                 /*
 257                  * CPU PMCs interrupt using an NMI.  If the PMC module is
 258                  * active, pass the 'rip' value to the PMC module's interrupt
 259                  * handler.  A non-zero return value from the handler means that
 260                  * the NMI was consumed by it and we can return immediately.
 261                  */
 262                 if (pmc_intr != NULL &&
 263                     (*pmc_intr)(frame) != 0)
 264                         return;
 265 #endif
 266         }
 267
 268         if ((frame->tf_rflags & PSL_I) == 0) {
 269                 /*
 270                  * Buggy application or kernel code has disabled
 271                  * interrupts and then trapped.  Enabling interrupts
 272                  * now is wrong, but it is better than running with
 273                  * interrupts disabled until they are accidentally
 274                  * enabled later.
 275                  */
 276                 if (TRAPF_USERMODE(frame)) {
 277                         uprintf(
 278                             "pid %ld (%s): trap %d (%s) "
 279                             "with interrupts disabled\n",
 280                             (long)curproc->p_pid, curthread->td_name, type,
 281                             trap_msg[type]);
 282                 } else {
 283                         switch (type) {
 284                         case T_NMI:
 285                         case T_BPTFLT:
 286                         case T_TRCTRAP:
 287                         case T_PROTFLT:
 288                         case T_SEGNPFLT:
 289                         case T_STKFLT:
 290                                 break;
 291                         default:
 292                                 printf(
 293                                     "kernel trap %d with interrupts disabled\n",
 294                                     type);
 295
 296                                 /*
 297                                  * We shouldn't enable interrupts while holding a
 298                                  * spin lock.
 299                                  */
 300                                 if (td->td_md.md_spinlock_count == 0)
 301                                         enable_intr();
 302                         }
 303                 }
 304         }
 305
 306         if (TRAPF_USERMODE(frame)) {
 307                 /* user trap */
 308
 309                 td->td_pticks = 0;
 310                 td->td_frame = frame;
 311                 addr = frame->tf_rip;
 312                 if (td->td_cowgen != p->p_cowgen)
 313                         thread_cow_update(td);
 314
 315                 switch (type) {
 316                 case T_PRIVINFLT:       /* privileged instruction fault */
 317                         signo = SIGILL;
 318                         ucode = ILL_PRVOPC;
 319                         break;
 320
 321                 case T_BPTFLT:          /* bpt instruction fault */
 322 #ifdef KDTRACE_HOOKS
 323                         if (trap_user_dtrace(frame, &dtrace_pid_probe_ptr))
 324                                 return;
 325 #else
 326                         enable_intr();
 327 #endif
 328                         signo = SIGTRAP;
 329                         ucode = TRAP_BRKPT;
 330                         break;
 331
 332                 case T_TRCTRAP:         /* debug exception */
 333                         enable_intr();
 334                         signo = SIGTRAP;
 335                         ucode = TRAP_TRACE;
 336                         dr6 = rdr6();
 337                         if ((dr6 & DBREG_DR6_BS) != 0) {
 338                                 PROC_LOCK(td->td_proc);
 339                                 if ((td->td_dbgflags & TDB_STEP) != 0) {
 340                                         td->td_frame->tf_rflags &= ~PSL_T;
 341                                         td->td_dbgflags &= ~TDB_STEP;
 342                                 }
 343                                 PROC_UNLOCK(td->td_proc);
 344                         }
 345                         break;
 346
 347                 case T_ARITHTRAP:       /* arithmetic trap */
 348                         ucode = fputrap_x87();
 349                         if (ucode == -1)
 350                                 return;
 351                         signo = SIGFPE;
 352                         break;
 353
 354                 case T_PROTFLT:         /* general protection fault */
 355                         signo = SIGBUS;
 356                         ucode = BUS_OBJERR;
 357                         break;
 358                 case T_STKFLT:          /* stack fault */
 359                 case T_SEGNPFLT:        /* segment not present fault */
 360                         signo = SIGBUS;
 361                         ucode = BUS_ADRERR;
 362                         break;
 363                 case T_TSSFLT:          /* invalid TSS fault */
 364                         signo = SIGBUS;
 365                         ucode = BUS_OBJERR;
 366                         break;
 367                 case T_ALIGNFLT:
 368                         signo = SIGBUS;
 369                         ucode = BUS_ADRALN;
 370                         break;
 371                 case T_DOUBLEFLT:       /* double fault */
 372                 default:
 373                         signo = SIGBUS;
 374                         ucode = BUS_OBJERR;
 375                         break;
 376
 377                 case T_PAGEFLT:         /* page fault */
 378                         /*
 379                          * Can emulator handle this trap?
 380                          */
 381                         if (*p->p_sysent->sv_trap != NULL &&
 382                             (*p->p_sysent->sv_trap)(td) == 0)
 383                                 return;
 384
 385                         pf = trap_pfault(frame, true, &signo, &ucode);
 386                         if (pf == -1)
 387                                 return;
 388                         if (pf == 0)
 389                                 goto userret;
 390                         addr = frame->tf_addr;
 391                         break;
 392
 393                 case T_DIVIDE:          /* integer divide fault */
 394                         ucode = FPE_INTDIV;
 395                         signo = SIGFPE;
 396                         break;
 397
 398                 case T_NMI:
 399                         nmi_handle_intr(type, frame);
 400                         return;
 401
 402                 case T_OFLOW:           /* integer overflow fault */
 403                         ucode = FPE_INTOVF;
 404                         signo = SIGFPE;
 405                         break;
 406
 407                 case T_BOUND:           /* bounds check fault */
 408                         ucode = FPE_FLTSUB;
 409                         signo = SIGFPE;
 410                         break;
 411
 412                 case T_DNA:
 413                         /* transparent fault (due to context switch "late") */
 414                         KASSERT(PCB_USER_FPU(td->td_pcb),
 415                             ("kernel FPU ctx has leaked"));
 416                         fpudna();
 417                         return;
 418
 419                 case T_FPOPFLT:         /* FPU operand fetch fault */
 420                         ucode = ILL_COPROC;
 421                         signo = SIGILL;
 422                         break;
 423
 424                 case T_XMMFLT:          /* SIMD floating-point exception */
 425                         ucode = fputrap_sse();
 426                         if (ucode == -1)
 427                                 return;
 428                         signo = SIGFPE;
 429                         break;
 430 #ifdef KDTRACE_HOOKS
 431                 case T_DTRACE_RET:
 432                         (void)trap_user_dtrace(frame, &dtrace_return_probe_ptr);
 433                         return;
 434 #endif
 435                 }
 436         } else {
 437                 /* kernel trap */
 438
 439                 KASSERT(cold || td->td_ucred != NULL,
 440                     ("kernel trap doesn't have ucred"));
 441                 switch (type) {
 442                 case T_PAGEFLT:                 /* page fault */
 443                         (void)trap_pfault(frame, false, NULL, NULL);
 444                         return;
 445
 446                 case T_DNA:
 447                         if (PCB_USER_FPU(td->td_pcb))
 448                                 panic("Unregistered use of FPU in kernel");
 449                         fpudna();
 450                         return;
 451
 452                 case T_ARITHTRAP:       /* arithmetic trap */
 453                 case T_XMMFLT:          /* SIMD floating-point exception */
 454                 case T_FPOPFLT:         /* FPU operand fetch fault */
 455                         /*
 456                          * For now, supporting kernel handler
 457                          * registration for FPU traps is overkill.
 458                          */
 459                         trap_fatal(frame, 0);
 460                         return;
 461
 462                 case T_STKFLT:          /* stack fault */
 463                 case T_PROTFLT:         /* general protection fault */
 464                 case T_SEGNPFLT:        /* segment not present fault */
 465                         if (td->td_intr_nesting_level != 0)
 466                                 break;
 467
 468                         /*
 469                          * Invalid segment selectors and out of bounds
 470                          * %rip's and %rsp's can be set up in user mode.
 471                          * This causes a fault in kernel mode when the
 472                          * kernel tries to return to user mode.  We want
 473                          * to get this fault so that we can fix the
 474                          * problem here and not have to check all the
 475                          * selectors and pointers when the user changes
 476                          * them.
 477                          *
 478                          * In case of PTI, the IRETQ faulted while the
 479                          * kernel used the pti stack, and exception
 480                          * frame records %rsp value pointing to that
 481                          * stack.  If we return normally to
 482                          * doreti_iret_fault, the trapframe is
 483                          * reconstructed on pti stack, and calltrap()
 484                          * called on it as well.  Due to the very
 485                          * limited pti stack size, kernel does not
 486                          * survive for too long.  Switch to the normal
 487                          * thread stack for the trap handling.
 488                          *
 489                          * Magic '5' is the number of qwords occupied by
 490                          * the hardware trap frame.
 491                          */
 492                         if (frame->tf_rip == (long)doreti_iret) {
 493                                 KASSERT((read_rflags() & PSL_I) == 0,
 494                                     ("interrupts enabled"));
 495                                 frame->tf_rip = (long)doreti_iret_fault;
 496                                 if ((PCPU_GET(curpmap)->pm_ucr3 !=
 497                                     PMAP_NO_CR3) &&
 498                                     (frame->tf_rsp == (uintptr_t)PCPU_GET(
 499                                     pti_rsp0) - 5 * sizeof(register_t))) {
 500                                         frame->tf_rsp = PCPU_GET(rsp0) - 5 *
 501                                             sizeof(register_t);
 502                                 }
 503                                 return;
 504                         }
 505
 506                         for (i = 0; i < nitems(sfhandlers); i++) {
 507                                 if (frame->tf_rip == sfhandlers[i].faddr) {
 508                                         KASSERT((read_rflags() & PSL_I) == 0,
 509                                             ("interrupts enabled"));
 510                                         frame->tf_rip = sfhandlers[i].fhandler;
 511                                         return;
 512                                 }
 513                         }
 514
 515                         if (curpcb->pcb_onfault != NULL) {
 516                                 frame->tf_rip = (long)curpcb->pcb_onfault;
 517                                 return;
 518                         }
 519                         break;
 520
 521                 case T_TSSFLT:
 522                         /*
 523                          * PSL_NT can be set in user mode and isn't cleared
 524                          * automatically when the kernel is entered.  This
 525                          * causes a TSS fault when the kernel attempts to
 526                          * `iret' because the TSS link is uninitialized.  We
 527                          * want to get this fault so that we can fix the
 528                          * problem here and not every time the kernel is
 529                          * entered.
 530                          */
 531                         if (frame->tf_rflags & PSL_NT) {
 532                                 frame->tf_rflags &= ~PSL_NT;
 533                                 return;
 534                         }
 535                         break;
 536
 537                 case T_TRCTRAP:  /* debug exception */
 538                         /* Clear any pending debug events. */
 539                         dr6 = rdr6();
 540                         load_dr6(0);
 541
 542                         /*
 543                          * Ignore debug register exceptions due to
 544                          * accesses in the user's address space, which
 545                          * can happen under several conditions such as
 546                          * if a user sets a watchpoint on a buffer and
 547                          * then passes that buffer to a system call.
 548                          * We still want to get TRCTRAPS for addresses
 549                          * in kernel space because that is useful when
 550                          * debugging the kernel.
 551                          */
 552                         if (user_dbreg_trap(dr6))
 553                                 return;
 554
 555                         /*
 556                          * Malicious user code can configure a debug
 557                          * register watchpoint to trap on data access
 558                          * to the top of stack and then execute 'pop
 559                          * %ss; int 3'.  Due to exception deferral for
 560                          * 'pop %ss', the CPU will not interrupt 'int
 561                          * 3' to raise the DB# exception for the debug
 562                          * register but will postpone the DB# until
 563                          * execution of the first instruction of the
 564                          * BP# handler (in kernel mode).  Normally the
 565                          * previous check would ignore DB# exceptions
 566                          * for watchpoints on user addresses raised in
 567                          * kernel mode.  However, some CPU errata
 568                          * include cases where DB# exceptions do not
 569                          * properly set bits in %dr6, e.g. Haswell
 570                          * HSD23 and Skylake-X SKZ24.
 571                          *
 572                          * A deferred DB# can also be raised on the
 573                          * first instructions of system call entry
 574                          * points or single-step traps via similar use
 575                          * of 'pop %ss' or 'mov xxx, %ss'.
 576                          */
 577                         if (pti) {
 578                                 if (frame->tf_rip ==
 579                                     (uintptr_t)IDTVEC(fast_syscall_pti) ||
 580 #ifdef COMPAT_FREEBSD32
 581                                     frame->tf_rip ==
 582                                     (uintptr_t)IDTVEC(int0x80_syscall_pti) ||
 583 #endif
 584                                     frame->tf_rip == (uintptr_t)IDTVEC(bpt_pti))
 585                                         return;
 586                         } else {
 587                                 if (frame->tf_rip ==
 588                                     (uintptr_t)IDTVEC(fast_syscall) ||
 589 #ifdef COMPAT_FREEBSD32
 590                                     frame->tf_rip ==
 591                                     (uintptr_t)IDTVEC(int0x80_syscall) ||
 592 #endif
 593                                     frame->tf_rip == (uintptr_t)IDTVEC(bpt))
 594                                         return;
 595                         }
 596                         if (frame->tf_rip == (uintptr_t)IDTVEC(dbg) ||
 597                             /* Needed for AMD. */
 598                             frame->tf_rip == (uintptr_t)IDTVEC(fast_syscall32))
 599                                 return;
 600                         /*
 601                          * FALLTHROUGH (TRCTRAP kernel mode, kernel address)
 602                          */
 603                 case T_BPTFLT:
 604                         /*
 605                          * If KDB is enabled, let it handle the debugger trap.
 606                          * Otherwise, debugger traps "can't happen".
 607                          */
 608 #ifdef KDB
 609                         if (kdb_trap(type, dr6, frame))
 610                                 return;
 611 #endif
 612                         break;
 613
 614                 case T_NMI:
 615                         nmi_handle_intr(type, frame);
 616                         return;
 617                 }
 618
 619                 trap_fatal(frame, 0);
 620                 return;
 621         }
 622
 623         ksiginfo_init_trap(&ksi);
 624         ksi.ksi_signo = signo;
 625         ksi.ksi_code = ucode;
 626         ksi.ksi_trapno = type;
 627         ksi.ksi_addr = (void *)addr;
 628         if (uprintf_signal) {
 629                 uprintf("pid %d comm %s: signal %d err %#lx code %d type %d "
 630                     "addr %#lx rsp %#lx rip %#lx rax %#lx"
 631                     "<%02x %02x %02x %02x %02x %02x %02x %02x>\n",
 632                     p->p_pid, p->p_comm, signo, frame->tf_err, ucode, type,
 633                     addr, frame->tf_rsp, frame->tf_rip, frame->tf_rax,
 634                     fubyte((void *)(frame->tf_rip + 0)),
 635                     fubyte((void *)(frame->tf_rip + 1)),
 636                     fubyte((void *)(frame->tf_rip + 2)),
 637                     fubyte((void *)(frame->tf_rip + 3)),
 638                     fubyte((void *)(frame->tf_rip + 4)),
 639                     fubyte((void *)(frame->tf_rip + 5)),
 640                     fubyte((void *)(frame->tf_rip + 6)),
 641                     fubyte((void *)(frame->tf_rip + 7)));
 642         }
 643         KASSERT((read_rflags() & PSL_I) != 0, ("interrupts disabled"));
 644         trapsignal(td, &ksi);
 645
 646 userret:
 647         userret(td, frame);
 648         KASSERT(PCB_USER_FPU(td->td_pcb),
 649             ("Return from trap with kernel FPU ctx leaked"));
 650 }
 651
 652 /*
 653  * Ensure that we ignore any DTrace-induced faults. This function cannot
 654  * be instrumented, so it cannot generate such faults itself.
 655  */
 656 void
 657 trap_check(struct trapframe *frame)
 658 {
 659
 660 #ifdef KDTRACE_HOOKS
 661         if (dtrace_trap_func != NULL &&
 662             (*dtrace_trap_func)(frame, frame->tf_trapno) != 0)
 663                 return;
 664 #endif
 665         trap(frame);
 666 }
 667
 668 static bool
 669 trap_is_smap(struct trapframe *frame)
 670 {
 671
 672         /*
 673          * A page fault on a userspace address is classified as
 674          * SMAP-induced if:
 675          * - SMAP is supported;
 676          * - kernel mode accessed present data page;
 677          * - rflags.AC was cleared.
 678          * Kernel must never access user space with rflags.AC cleared
 679          * if SMAP is enabled.
 680          */
 681         return ((cpu_stdext_feature & CPUID_STDEXT_SMAP) != 0 &&
 682             (frame->tf_err & (PGEX_P | PGEX_U | PGEX_I | PGEX_RSV)) ==
 683             PGEX_P && (frame->tf_rflags & PSL_AC) == 0);
 684 }
 685
 686 static bool
 687 trap_is_pti(struct trapframe *frame)
 688 {
 689
 690         return (PCPU_GET(curpmap)->pm_ucr3 != PMAP_NO_CR3 &&
 691             pg_nx != 0 && (frame->tf_err & (PGEX_P | PGEX_W |
 692             PGEX_U | PGEX_I)) == (PGEX_P | PGEX_U | PGEX_I) &&
 693             (curpcb->pcb_saved_ucr3 & ~CR3_PCID_MASK) ==
 694             (PCPU_GET(curpmap)->pm_cr3 & ~CR3_PCID_MASK));
 695 }
 696
 697 /*
 698  * Handle all details of a page fault.
 699  * Returns:
 700  * -1 if this fault was fatal, typically from kernel mode
 701  *    (cannot happen, but we need to return something).
 702  * 0  if this fault was handled by updating either the user or kernel
 703  *    page table, execution can continue.
 704  * 1  if this fault was from usermode and it was not handled, a synchronous
 705  *    signal should be delivered to the thread.  *signo returns the signal
 706  *    number, *ucode gives si_code.
 707  */
 708 static int
 709 trap_pfault(struct trapframe *frame, bool usermode, int *signo, int *ucode)
 710 {
 711         struct thread *td;
 712         struct proc *p;
 713         vm_map_t map;
 714         vm_offset_t eva;
 715         int rv;
 716         vm_prot_t ftype;
 717
 718         MPASS(!usermode || (signo != NULL && ucode != NULL));
 719
 720         td = curthread;
 721         p = td->td_proc;
 722         eva = frame->tf_addr;
 723
 724         if (__predict_false((td->td_pflags & TDP_NOFAULTING) != 0)) {
 725                 /*
 726                  * Due to both processor errata and lazy TLB invalidation when
 727                  * access restrictions are removed from virtual pages, memory
 728                  * accesses that are allowed by the physical mapping layer may
 729                  * nonetheless cause one spurious page fault per virtual page.
 730                  * When the thread is executing a "no faulting" section that
 731                  * is bracketed by vm_fault_{disable,enable}_pagefaults(),
 732                  * every page fault is treated as a spurious page fault,
 733                  * unless it accesses the same virtual address as the most
 734                  * recent page fault within the same "no faulting" section.
 735                  */
 736                 if (td->td_md.md_spurflt_addr != eva ||
 737                     (td->td_pflags & TDP_RESETSPUR) != 0) {
 738                         /*
 739                          * Do nothing to the TLB.  A stale TLB entry is
 740                          * flushed automatically by a page fault.
 741                          */
 742                         td->td_md.md_spurflt_addr = eva;
 743                         td->td_pflags &= ~TDP_RESETSPUR;
 744                         return (0);
 745                 }
 746         } else {
 747                 /*
 748                  * If we get a page fault while in a critical section, then
 749                  * it is most likely a fatal kernel page fault.  The kernel
 750                  * is already going to panic trying to get a sleep lock to
 751                  * do the VM lookup, so just consider it a fatal trap so the
 752                  * kernel can print out a useful trap message and even get
 753                  * to the debugger.
 754                  *
 755                  * If we get a page fault while holding a non-sleepable
 756                  * lock, then it is most likely a fatal kernel page fault.
 757                  * If WITNESS is enabled, then it's going to whine about
 758                  * bogus LORs with various VM locks, so just skip to the
 759                  * fatal trap handling directly.
 760                  */
 761                 if (td->td_critnest != 0 ||
 762                     WITNESS_CHECK(WARN_SLEEPOK | WARN_GIANTOK, NULL,
 763                     "Kernel page fault") != 0) {
 764                         trap_fatal(frame, eva);
 765                         return (-1);
 766                 }
 767         }
 768         if (eva >= VM_MIN_KERNEL_ADDRESS) {
 769                 /*
 770                  * Don't allow user-mode faults in kernel address space.
 771                  */
 772                 if (usermode) {
 773                         *signo = SIGSEGV;
 774                         *ucode = SEGV_MAPERR;
 775                         return (1);
 776                 }
 777
 778                 map = kernel_map;
 779         } else {
 780                 map = &p->p_vmspace->vm_map;
 781
 782                 /*
 783                  * When accessing a usermode address, kernel must be
 784                  * ready to accept the page fault, and provide a
 785                  * handling routine.  Since accessing the address
 786                  * without the handler is a bug, do not try to handle
 787                  * it normally, and panic immediately.
 788                  *
 789                  * If SMAP is enabled, filter SMAP faults also,
 790                  * because illegal access might occur to the mapped
 791                  * user address, causing infinite loop.
 792                  */
 793                 if (!usermode && (td->td_intr_nesting_level != 0 ||
 794                     trap_is_smap(frame) || curpcb->pcb_onfault == NULL)) {
 795                         trap_fatal(frame, eva);
 796                         return (-1);
 797                 }
 798         }
 799
 800         /*
 801          * If the trap was caused by errant bits in the PTE then panic.
 802          */
 803         if (frame->tf_err & PGEX_RSV) {
 804                 trap_fatal(frame, eva);
 805                 return (-1);
 806         }
 807
 808         /*
 809          * User-mode protection key violation (PKU).  May happen
 810          * either from usermode or from kernel if copyin accessed
 811          * key-protected mapping.
 812          */
 813         if ((frame->tf_err & PGEX_PK) != 0) {
 814                 if (eva > VM_MAXUSER_ADDRESS) {
 815                         trap_fatal(frame, eva);
 816                         return (-1);
 817                 }
 818                 if (usermode) {
 819                         *signo = SIGSEGV;
 820                         *ucode = SEGV_PKUERR;
 821                         return (1);
 822                 }
 823                 goto after_vmfault;
 824         }
 825
 826         /*
 827          * If nx protection of the usermode portion of kernel page
 828          * tables caused trap, panic.
 829          */
 830         if (usermode && trap_is_pti(frame))
 831                 panic("PTI: pid %d comm %s tf_err %#lx", p->p_pid,
 832                     p->p_comm, frame->tf_err);
 833
 834         /*
 835          * PGEX_I is defined only if the execute disable bit capability is
 836          * supported and enabled.
 837          */
 838         if (frame->tf_err & PGEX_W)
 839                 ftype = VM_PROT_WRITE;
 840         else if ((frame->tf_err & PGEX_I) && pg_nx != 0)
 841                 ftype = VM_PROT_EXECUTE;
 842         else
 843                 ftype = VM_PROT_READ;
 844
 845         /* Fault in the page. */
 846         rv = vm_fault_trap(map, eva, ftype, VM_FAULT_NORMAL, signo, ucode);
 847         if (rv == KERN_SUCCESS) {
 848 #ifdef HWPMC_HOOKS
 849                 if (ftype == VM_PROT_READ || ftype == VM_PROT_WRITE) {
 850                         PMC_SOFT_CALL_TF( , , page_fault, all, frame);
 851                         if (ftype == VM_PROT_READ)
 852                                 PMC_SOFT_CALL_TF( , , page_fault, read,
 853                                     frame);
 854                         else
 855                                 PMC_SOFT_CALL_TF( , , page_fault, write,
 856                                     frame);
 857                 }
 858 #endif
 859                 return (0);
 860         }
 861
 862         if (usermode)
 863                 return (1);
 864 after_vmfault:
 865         if (td->td_intr_nesting_level == 0 &&
 866             curpcb->pcb_onfault != NULL) {
 867                 frame->tf_rip = (long)curpcb->pcb_onfault;
 868                 return (0);
 869         }
 870         trap_fatal(frame, eva);
 871         return (-1);
 872 }
 873
 874 static void
 875 trap_fatal(struct trapframe *frame, vm_offset_t eva)
 876 {
 877         int code, ss;
 878         u_int type;
 879         struct soft_segment_descriptor softseg;
 880         struct user_segment_descriptor *gdt;
 881 #ifdef KDB
 882         bool handled;
 883 #endif
 884
 885         code = frame->tf_err;
 886         type = frame->tf_trapno;
 887         gdt = *PCPU_PTR(gdt);
 888         sdtossd(&gdt[IDXSEL(frame->tf_cs & 0xffff)], &softseg);
 889
 890         printf("\n\nFatal trap %d: %s while in %s mode\n", type,
 891             type < nitems(trap_msg) ? trap_msg[type] : UNKNOWN,
 892             TRAPF_USERMODE(frame) ? "user" : "kernel");
 893 #ifdef SMP
 894         /* two separate prints in case of a trap on an unmapped page */
 895         printf("cpuid = %d; ", PCPU_GET(cpuid));
 896         printf("apic id = %02x\n", PCPU_GET(apic_id));
 897 #endif
 898         if (type == T_PAGEFLT) {
 899                 printf("fault virtual address   = 0x%lx\n", eva);
 900                 printf("fault code              = %s %s %s%s%s, %s\n",
 901                         code & PGEX_U ? "user" : "supervisor",
 902                         code & PGEX_W ? "write" : "read",
 903                         code & PGEX_I ? "instruction" : "data",
 904                         code & PGEX_PK ? " prot key" : "",
 905                         code & PGEX_SGX ? " SGX" : "",
 906                         code & PGEX_RSV ? "reserved bits in PTE" :
 907                         code & PGEX_P ? "protection violation" : "page not present");
 908         }
 909         printf("instruction pointer     = 0x%lx:0x%lx\n",
 910                frame->tf_cs & 0xffff, frame->tf_rip);
 911         ss = frame->tf_ss & 0xffff;
 912         printf("stack pointer           = 0x%x:0x%lx\n", ss, frame->tf_rsp);
 913         printf("frame pointer           = 0x%x:0x%lx\n", ss, frame->tf_rbp);
 914         printf("code segment            = base 0x%lx, limit 0x%lx, type 0x%x\n",
 915                softseg.ssd_base, softseg.ssd_limit, softseg.ssd_type);
 916         printf("                        = DPL %d, pres %d, long %d, def32 %d, gran %d\n",
 917                softseg.ssd_dpl, softseg.ssd_p, softseg.ssd_long, softseg.ssd_def32,
 918                softseg.ssd_gran);
 919         printf("processor eflags        = ");
 920         if (frame->tf_rflags & PSL_T)
 921                 printf("trace trap, ");
 922         if (frame->tf_rflags & PSL_I)
 923                 printf("interrupt enabled, ");
 924         if (frame->tf_rflags & PSL_NT)
 925                 printf("nested task, ");
 926         if (frame->tf_rflags & PSL_RF)
 927                 printf("resume, ");
 928         printf("IOPL = %ld\n", (frame->tf_rflags & PSL_IOPL) >> 12);
 929         printf("current process         = %d (%s)\n",
 930             curproc->p_pid, curthread->td_name);
 931
 932 #ifdef KDB
 933         if (debugger_on_trap) {
 934                 kdb_why = KDB_WHY_TRAP;
 935                 handled = kdb_trap(type, 0, frame);
 936                 kdb_why = KDB_WHY_UNSET;
 937                 if (handled)
 938                         return;
 939         }
 940 #endif
 941         printf("trap number             = %d\n", type);
 942         panic("%s", type < nitems(trap_msg) ? trap_msg[type] :
 943             "unknown/reserved trap");
 944 }
 945
 946 #ifdef KDTRACE_HOOKS
 947 /*
 948  * Invoke a userspace DTrace hook.  The hook pointer is cleared when no
 949  * userspace probes are enabled, so we must synchronize with DTrace to ensure
 950  * that a trapping thread is able to call the hook before it is cleared.
 951  */
 952 static bool
 953 trap_user_dtrace(struct trapframe *frame, int (**hookp)(struct trapframe *))
 954 {
 955         int (*hook)(struct trapframe *);
 956
 957         hook = atomic_load_ptr(hookp);
 958         enable_intr();
 959         if (hook != NULL)
 960                 return ((hook)(frame) == 0);
 961         return (false);
 962 }
 963 #endif
 964
 965 /*
 966  * Double fault handler. Called when a fault occurs while writing
 967  * a frame for a trap/exception onto the stack. This usually occurs
 968  * when the stack overflows (such is the case with infinite recursion,
 969  * for example).
 970  */
 971 void
 972 dblfault_handler(struct trapframe *frame)
 973 {
 974 #ifdef KDTRACE_HOOKS
 975         if (dtrace_doubletrap_func != NULL)
 976                 (*dtrace_doubletrap_func)();
 977 #endif
 978         printf("\nFatal double fault\n"
 979             "rip %#lx rsp %#lx rbp %#lx\n"
 980             "rax %#lx rdx %#lx rbx %#lx\n"
 981             "rcx %#lx rsi %#lx rdi %#lx\n"
 982             "r8 %#lx r9 %#lx r10 %#lx\n"
 983             "r11 %#lx r12 %#lx r13 %#lx\n"
 984             "r14 %#lx r15 %#lx rflags %#lx\n"
 985             "cs %#lx ss %#lx ds %#hx es %#hx fs %#hx gs %#hx\n"
 986             "fsbase %#lx gsbase %#lx kgsbase %#lx\n",
 987             frame->tf_rip, frame->tf_rsp, frame->tf_rbp,
 988             frame->tf_rax, frame->tf_rdx, frame->tf_rbx,
 989             frame->tf_rcx, frame->tf_rdi, frame->tf_rsi,
 990             frame->tf_r8, frame->tf_r9, frame->tf_r10,
 991             frame->tf_r11, frame->tf_r12, frame->tf_r13,
 992             frame->tf_r14, frame->tf_r15, frame->tf_rflags,
 993             frame->tf_cs, frame->tf_ss, frame->tf_ds, frame->tf_es,
 994             frame->tf_fs, frame->tf_gs,
 995             rdmsr(MSR_FSBASE), rdmsr(MSR_GSBASE), rdmsr(MSR_KGSBASE));
 996 #ifdef SMP
 997         /* two separate prints in case of a trap on an unmapped page */
 998         printf("cpuid = %d; ", PCPU_GET(cpuid));
 999         printf("apic id = %02x\n", PCPU_GET(apic_id));
1000 #endif
1001         panic("double fault");
1002 }
1003
1004 static int __noinline
1005 cpu_fetch_syscall_args_fallback(struct thread *td, struct syscall_args *sa)
1006 {
1007         struct proc *p;
1008         struct trapframe *frame;
1009         register_t *argp;
1010         caddr_t params;
1011         int reg, regcnt, error;
1012
1013         p = td->td_proc;
1014         frame = td->td_frame;
1015         reg = 0;
1016         regcnt = NARGREGS;
1017
1018         if (sa->code == SYS_syscall || sa->code == SYS___syscall) {
1019                 sa->code = frame->tf_rdi;
1020                 reg++;
1021                 regcnt--;
1022         }
1023
1024         if (sa->code >= p->p_sysent->sv_size)
1025                 sa->callp = &p->p_sysent->sv_table[0];
1026         else
1027                 sa->callp = &p->p_sysent->sv_table[sa->code];
1028
1029         KASSERT(sa->callp->sy_narg <= nitems(sa->args),
1030             ("Too many syscall arguments!"));
1031         argp = &frame->tf_rdi;
1032         argp += reg;
1033         memcpy(sa->args, argp, sizeof(sa->args[0]) * NARGREGS);
1034         if (sa->callp->sy_narg > regcnt) {
1035                 params = (caddr_t)frame->tf_rsp + sizeof(register_t);
1036                 error = copyin(params, &sa->args[regcnt],
1037                     (sa->callp->sy_narg - regcnt) * sizeof(sa->args[0]));
1038                 if (__predict_false(error != 0))
1039                         return (error);
1040         }
1041
1042         td->td_retval[0] = 0;
1043         td->td_retval[1] = frame->tf_rdx;
1044
1045         return (0);
1046 }
1047
1048 int
1049 cpu_fetch_syscall_args(struct thread *td)
1050 {
1051         struct proc *p;
1052         struct trapframe *frame;
1053         struct syscall_args *sa;
1054
1055         p = td->td_proc;
1056         frame = td->td_frame;
1057         sa = &td->td_sa;
1058
1059         sa->code = frame->tf_rax;
1060
1061         if (__predict_false(sa->code == SYS_syscall ||
1062             sa->code == SYS___syscall ||
1063             sa->code >= p->p_sysent->sv_size))
1064                 return (cpu_fetch_syscall_args_fallback(td, sa));
1065
1066         sa->callp = &p->p_sysent->sv_table[sa->code];
1067         KASSERT(sa->callp->sy_narg <= nitems(sa->args),
1068             ("Too many syscall arguments!"));
1069
1070         if (__predict_false(sa->callp->sy_narg > NARGREGS))
1071                 return (cpu_fetch_syscall_args_fallback(td, sa));
1072
1073         memcpy(sa->args, &frame->tf_rdi, sizeof(sa->args[0]) * NARGREGS);
1074
1075         td->td_retval[0] = 0;
1076         td->td_retval[1] = frame->tf_rdx;
1077
1078         return (0);
1079 }
1080
1081 #include "../../kern/subr_syscall.c"
1082
1083 static void (*syscall_ret_l1d_flush)(void);
1084 int syscall_ret_l1d_flush_mode;
1085
1086 static void
1087 flush_l1d_hw(void)
1088 {
1089
1090         wrmsr(MSR_IA32_FLUSH_CMD, IA32_FLUSH_CMD_L1D);
1091 }
1092
1093 static void __noinline
1094 amd64_syscall_ret_flush_l1d_check(int error)
1095 {
1096         void (*p)(void);
1097
1098         if (error != EEXIST && error != EAGAIN && error != EXDEV &&
1099             error != ENOENT && error != ENOTCONN && error != EINPROGRESS) {
1100                 p = atomic_load_ptr(&syscall_ret_l1d_flush);
1101                 if (p != NULL)
1102                         p();
1103         }
1104 }
1105
1106 static void __inline
1107 amd64_syscall_ret_flush_l1d_check_inline(int error)
1108 {
1109
1110         if (__predict_false(error != 0))
1111                 amd64_syscall_ret_flush_l1d_check(error);
1112 }
1113
1114 void
1115 amd64_syscall_ret_flush_l1d(int error)
1116 {
1117
1118         amd64_syscall_ret_flush_l1d_check_inline(error);
1119 }
1120
1121 void
1122 amd64_syscall_ret_flush_l1d_recalc(void)
1123 {
1124         bool l1d_hw;
1125
1126         l1d_hw = (cpu_stdext_feature3 & CPUID_STDEXT3_L1D_FLUSH) != 0;
1127 again:
1128         switch (syscall_ret_l1d_flush_mode) {
1129         case 0:
1130                 syscall_ret_l1d_flush = NULL;
1131                 break;
1132         case 1:
1133                 syscall_ret_l1d_flush = l1d_hw ? flush_l1d_hw :
1134                     flush_l1d_sw_abi;
1135                 break;
1136         case 2:
1137                 syscall_ret_l1d_flush = l1d_hw ? flush_l1d_hw : NULL;
1138                 break;
1139         case 3:
1140                 syscall_ret_l1d_flush = flush_l1d_sw_abi;
1141                 break;
1142         default:
1143                 syscall_ret_l1d_flush_mode = 1;
1144                 goto again;
1145         }
1146 }
1147
1148 static int
1149 machdep_syscall_ret_flush_l1d(SYSCTL_HANDLER_ARGS)
1150 {
1151         int error, val;
1152
1153         val = syscall_ret_l1d_flush_mode;
1154         error = sysctl_handle_int(oidp, &val, 0, req);
1155         if (error != 0 || req->newptr == NULL)
1156                 return (error);
1157         syscall_ret_l1d_flush_mode = val;
1158         amd64_syscall_ret_flush_l1d_recalc();
1159         return (0);
1160 }
1161 SYSCTL_PROC(_machdep, OID_AUTO, syscall_ret_flush_l1d, CTLTYPE_INT |
1162     CTLFLAG_RWTUN | CTLFLAG_NOFETCH | CTLFLAG_MPSAFE, NULL, 0,
1163     machdep_syscall_ret_flush_l1d, "I",
1164     "Flush L1D on syscall return with error (0 - off, 1 - on, "
1165     "2 - use hw only, 3 - use sw only)");
1166
1167 /*
1168  * System call handler for native binaries.  The trap frame is already
1169  * set up by the assembler trampoline and a pointer to it is saved in
1170  * td_frame.
1171  */
1172 void
1173 amd64_syscall(struct thread *td, int traced)
1174 {
1175         ksiginfo_t ksi;
1176
1177 #ifdef DIAGNOSTIC
1178         if (!TRAPF_USERMODE(td->td_frame)) {
1179                 panic("syscall");
1180                 /* NOT REACHED */
1181         }
1182 #endif
1183         syscallenter(td);
1184
1185         /*
1186          * Traced syscall.
1187          */
1188         if (__predict_false(traced)) {
1189                 td->td_frame->tf_rflags &= ~PSL_T;
1190                 ksiginfo_init_trap(&ksi);
1191                 ksi.ksi_signo = SIGTRAP;
1192                 ksi.ksi_code = TRAP_TRACE;
1193                 ksi.ksi_addr = (void *)td->td_frame->tf_rip;
1194                 trapsignal(td, &ksi);
1195         }
1196
1197         KASSERT(PCB_USER_FPU(td->td_pcb),
1198             ("System call %s returning with kernel FPU ctx leaked",
1199              syscallname(td->td_proc, td->td_sa.code)));
1200         KASSERT(td->td_pcb->pcb_save == get_pcb_user_save_td(td),
1201             ("System call %s returning with mangled pcb_save",
1202              syscallname(td->td_proc, td->td_sa.code)));
1203         KASSERT(pmap_not_in_di(),
1204             ("System call %s returning with leaked invl_gen %lu",
1205             syscallname(td->td_proc, td->td_sa.code),
1206             td->td_md.md_invl_gen.gen));
1207
1208         syscallret(td);
1209
1210         /*
1211          * If the user-supplied value of %rip is not a canonical
1212          * address, then some CPUs will trigger a ring 0 #GP during
1213          * the sysret instruction.  However, the fault handler would
1214          * execute in ring 0 with the user's %gs and %rsp which would
1215          * not be safe.  Instead, use the full return path which
1216          * catches the problem safely.
1217          */
1218         if (__predict_false(td->td_frame->tf_rip >= (la57 ?
1219             VM_MAXUSER_ADDRESS_LA57 : VM_MAXUSER_ADDRESS_LA48)))
1220                 set_pcb_flags(td->td_pcb, PCB_FULL_IRET);
1221
1222         amd64_syscall_ret_flush_l1d_check_inline(td->td_errno);
1223 }