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