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/pioctl.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  * Exception, fault, and trap interface to the FreeBSD kernel.
 178  * This common code is called from assembly language IDT gate entry
 179  * routines that prepare a suitable stack frame, and restore this
 180  * frame after the exception has been processed.
 181  */
 182
 183 void
 184 trap(struct trapframe *frame)
 185 {
 186         ksiginfo_t ksi;
 187         struct thread *td;
 188         struct proc *p;
 189         register_t addr, dr6;
 190         int pf, signo, ucode;
 191         u_int type;
 192
 193         td = curthread;
 194         p = td->td_proc;
 195         dr6 = 0;
 196
 197         VM_CNT_INC(v_trap);
 198         type = frame->tf_trapno;
 199
 200 #ifdef SMP
 201         /* Handler for NMI IPIs used for stopping CPUs. */
 202         if (type == T_NMI && ipi_nmi_handler() == 0)
 203                 return;
 204 #endif
 205
 206 #ifdef KDB
 207         if (kdb_active) {
 208                 kdb_reenter();
 209                 return;
 210         }
 211 #endif
 212
 213         if (type == T_RESERVED) {
 214                 trap_fatal(frame, 0);
 215                 return;
 216         }
 217
 218         if (type == T_NMI) {
 219 #ifdef HWPMC_HOOKS
 220                 /*
 221                  * CPU PMCs interrupt using an NMI.  If the PMC module is
 222                  * active, pass the 'rip' value to the PMC module's interrupt
 223                  * handler.  A non-zero return value from the handler means that
 224                  * the NMI was consumed by it and we can return immediately.
 225                  */
 226                 if (pmc_intr != NULL &&
 227                     (*pmc_intr)(frame) != 0)
 228                         return;
 229 #endif
 230         }
 231
 232         if ((frame->tf_rflags & PSL_I) == 0) {
 233                 /*
 234                  * Buggy application or kernel code has disabled
 235                  * interrupts and then trapped.  Enabling interrupts
 236                  * now is wrong, but it is better than running with
 237                  * interrupts disabled until they are accidentally
 238                  * enabled later.
 239                  */
 240                 if (TRAPF_USERMODE(frame))
 241                         uprintf(
 242                             "pid %ld (%s): trap %d with interrupts disabled\n",
 243                             (long)curproc->p_pid, curthread->td_name, type);
 244                 else if (type != T_NMI && type != T_BPTFLT &&
 245                     type != T_TRCTRAP) {
 246                         /*
 247                          * XXX not quite right, since this may be for a
 248                          * multiple fault in user mode.
 249                          */
 250                         printf("kernel trap %d with interrupts disabled\n",
 251                             type);
 252
 253                         /*
 254                          * We shouldn't enable interrupts while holding a
 255                          * spin lock.
 256                          */
 257                         if (td->td_md.md_spinlock_count == 0)
 258                                 enable_intr();
 259                 }
 260         }
 261
 262         if (TRAPF_USERMODE(frame)) {
 263                 /* user trap */
 264
 265                 td->td_pticks = 0;
 266                 td->td_frame = frame;
 267                 addr = frame->tf_rip;
 268                 if (td->td_cowgen != p->p_cowgen)
 269                         thread_cow_update(td);
 270
 271                 switch (type) {
 272                 case T_PRIVINFLT:       /* privileged instruction fault */
 273                         signo = SIGILL;
 274                         ucode = ILL_PRVOPC;
 275                         break;
 276
 277                 case T_BPTFLT:          /* bpt instruction fault */
 278 #ifdef KDTRACE_HOOKS
 279                         if (trap_user_dtrace(frame, &dtrace_pid_probe_ptr))
 280                                 return;
 281 #else
 282                         enable_intr();
 283 #endif
 284                         signo = SIGTRAP;
 285                         ucode = TRAP_BRKPT;
 286                         break;
 287
 288                 case T_TRCTRAP:         /* debug exception */
 289                         enable_intr();
 290                         signo = SIGTRAP;
 291                         ucode = TRAP_TRACE;
 292                         dr6 = rdr6();
 293                         if ((dr6 & DBREG_DR6_BS) != 0) {
 294                                 PROC_LOCK(td->td_proc);
 295                                 if ((td->td_dbgflags & TDB_STEP) != 0) {
 296                                         td->td_frame->tf_rflags &= ~PSL_T;
 297                                         td->td_dbgflags &= ~TDB_STEP;
 298                                 }
 299                                 PROC_UNLOCK(td->td_proc);
 300                         }
 301                         break;
 302
 303                 case T_ARITHTRAP:       /* arithmetic trap */
 304                         ucode = fputrap_x87();
 305                         if (ucode == -1)
 306                                 return;
 307                         signo = SIGFPE;
 308                         break;
 309
 310                 case T_PROTFLT:         /* general protection fault */
 311                         signo = SIGBUS;
 312                         ucode = BUS_OBJERR;
 313                         break;
 314                 case T_STKFLT:          /* stack fault */
 315                 case T_SEGNPFLT:        /* segment not present fault */
 316                         signo = SIGBUS;
 317                         ucode = BUS_ADRERR;
 318                         break;
 319                 case T_TSSFLT:          /* invalid TSS fault */
 320                         signo = SIGBUS;
 321                         ucode = BUS_OBJERR;
 322                         break;
 323                 case T_ALIGNFLT:
 324                         signo = SIGBUS;
 325                         ucode = BUS_ADRALN;
 326                         break;
 327                 case T_DOUBLEFLT:       /* double fault */
 328                 default:
 329                         signo = SIGBUS;
 330                         ucode = BUS_OBJERR;
 331                         break;
 332
 333                 case T_PAGEFLT:         /* page fault */
 334                         /*
 335                          * Can emulator handle this trap?
 336                          */
 337                         if (*p->p_sysent->sv_trap != NULL &&
 338                             (*p->p_sysent->sv_trap)(td) == 0)
 339                                 return;
 340
 341                         pf = trap_pfault(frame, true, &signo, &ucode);
 342                         if (pf == -1)
 343                                 return;
 344                         if (pf == 0)
 345                                 goto userret;
 346                         addr = frame->tf_addr;
 347                         break;
 348
 349                 case T_DIVIDE:          /* integer divide fault */
 350                         ucode = FPE_INTDIV;
 351                         signo = SIGFPE;
 352                         break;
 353
 354 #ifdef DEV_ISA
 355                 case T_NMI:
 356                         nmi_handle_intr(type, frame);
 357                         return;
 358 #endif
 359
 360                 case T_OFLOW:           /* integer overflow fault */
 361                         ucode = FPE_INTOVF;
 362                         signo = SIGFPE;
 363                         break;
 364
 365                 case T_BOUND:           /* bounds check fault */
 366                         ucode = FPE_FLTSUB;
 367                         signo = SIGFPE;
 368                         break;
 369
 370                 case T_DNA:
 371                         /* transparent fault (due to context switch "late") */
 372                         KASSERT(PCB_USER_FPU(td->td_pcb),
 373                             ("kernel FPU ctx has leaked"));
 374                         fpudna();
 375                         return;
 376
 377                 case T_FPOPFLT:         /* FPU operand fetch fault */
 378                         ucode = ILL_COPROC;
 379                         signo = SIGILL;
 380                         break;
 381
 382                 case T_XMMFLT:          /* SIMD floating-point exception */
 383                         ucode = fputrap_sse();
 384                         if (ucode == -1)
 385                                 return;
 386                         signo = SIGFPE;
 387                         break;
 388 #ifdef KDTRACE_HOOKS
 389                 case T_DTRACE_RET:
 390                         (void)trap_user_dtrace(frame, &dtrace_return_probe_ptr);
 391                         return;
 392 #endif
 393                 }
 394         } else {
 395                 /* kernel trap */
 396
 397                 KASSERT(cold || td->td_ucred != NULL,
 398                     ("kernel trap doesn't have ucred"));
 399                 switch (type) {
 400                 case T_PAGEFLT:                 /* page fault */
 401                         (void)trap_pfault(frame, false, NULL, NULL);
 402                         return;
 403
 404                 case T_DNA:
 405                         if (PCB_USER_FPU(td->td_pcb))
 406                                 panic("Unregistered use of FPU in kernel");
 407                         fpudna();
 408                         return;
 409
 410                 case T_ARITHTRAP:       /* arithmetic trap */
 411                 case T_XMMFLT:          /* SIMD floating-point exception */
 412                 case T_FPOPFLT:         /* FPU operand fetch fault */
 413                         /*
 414                          * For now, supporting kernel handler
 415                          * registration for FPU traps is overkill.
 416                          */
 417                         trap_fatal(frame, 0);
 418                         return;
 419
 420                 case T_STKFLT:          /* stack fault */
 421                 case T_PROTFLT:         /* general protection fault */
 422                 case T_SEGNPFLT:        /* segment not present fault */
 423                         if (td->td_intr_nesting_level != 0)
 424                                 break;
 425
 426                         /*
 427                          * Invalid segment selectors and out of bounds
 428                          * %rip's and %rsp's can be set up in user mode.
 429                          * This causes a fault in kernel mode when the
 430                          * kernel tries to return to user mode.  We want
 431                          * to get this fault so that we can fix the
 432                          * problem here and not have to check all the
 433                          * selectors and pointers when the user changes
 434                          * them.
 435                          *
 436                          * In case of PTI, the IRETQ faulted while the
 437                          * kernel used the pti stack, and exception
 438                          * frame records %rsp value pointing to that
 439                          * stack.  If we return normally to
 440                          * doreti_iret_fault, the trapframe is
 441                          * reconstructed on pti stack, and calltrap()
 442                          * called on it as well.  Due to the very
 443                          * limited pti stack size, kernel does not
 444                          * survive for too long.  Switch to the normal
 445                          * thread stack for the trap handling.
 446                          *
 447                          * Magic '5' is the number of qwords occupied by
 448                          * the hardware trap frame.
 449                          */
 450                         if (frame->tf_rip == (long)doreti_iret) {
 451                                 frame->tf_rip = (long)doreti_iret_fault;
 452                                 if ((PCPU_GET(curpmap)->pm_ucr3 !=
 453                                     PMAP_NO_CR3) &&
 454                                     (frame->tf_rsp == (uintptr_t)PCPU_GET(
 455                                     pti_rsp0) - 5 * sizeof(register_t))) {
 456                                         frame->tf_rsp = PCPU_GET(rsp0) - 5 *
 457                                             sizeof(register_t);
 458                                 }
 459                                 return;
 460                         }
 461                         if (frame->tf_rip == (long)ld_ds) {
 462                                 frame->tf_rip = (long)ds_load_fault;
 463                                 return;
 464                         }
 465                         if (frame->tf_rip == (long)ld_es) {
 466                                 frame->tf_rip = (long)es_load_fault;
 467                                 return;
 468                         }
 469                         if (frame->tf_rip == (long)ld_fs) {
 470                                 frame->tf_rip = (long)fs_load_fault;
 471                                 return;
 472                         }
 473                         if (frame->tf_rip == (long)ld_gs) {
 474                                 frame->tf_rip = (long)gs_load_fault;
 475                                 return;
 476                         }
 477                         if (frame->tf_rip == (long)ld_gsbase) {
 478                                 frame->tf_rip = (long)gsbase_load_fault;
 479                                 return;
 480                         }
 481                         if (frame->tf_rip == (long)ld_fsbase) {
 482                                 frame->tf_rip = (long)fsbase_load_fault;
 483                                 return;
 484                         }
 485                         if (curpcb->pcb_onfault != NULL) {
 486                                 frame->tf_rip = (long)curpcb->pcb_onfault;
 487                                 return;
 488                         }
 489                         break;
 490
 491                 case T_TSSFLT:
 492                         /*
 493                          * PSL_NT can be set in user mode and isn't cleared
 494                          * automatically when the kernel is entered.  This
 495                          * causes a TSS fault when the kernel attempts to
 496                          * `iret' because the TSS link is uninitialized.  We
 497                          * want to get this fault so that we can fix the
 498                          * problem here and not every time the kernel is
 499                          * entered.
 500                          */
 501                         if (frame->tf_rflags & PSL_NT) {
 502                                 frame->tf_rflags &= ~PSL_NT;
 503                                 return;
 504                         }
 505                         break;
 506
 507                 case T_TRCTRAP:  /* debug exception */
 508                         /* Clear any pending debug events. */
 509                         dr6 = rdr6();
 510                         load_dr6(0);
 511
 512                         /*
 513                          * Ignore debug register exceptions due to
 514                          * accesses in the user's address space, which
 515                          * can happen under several conditions such as
 516                          * if a user sets a watchpoint on a buffer and
 517                          * then passes that buffer to a system call.
 518                          * We still want to get TRCTRAPS for addresses
 519                          * in kernel space because that is useful when
 520                          * debugging the kernel.
 521                          */
 522                         if (user_dbreg_trap(dr6))
 523                                 return;
 524
 525                         /*
 526                          * Malicious user code can configure a debug
 527                          * register watchpoint to trap on data access
 528                          * to the top of stack and then execute 'pop
 529                          * %ss; int 3'.  Due to exception deferral for
 530                          * 'pop %ss', the CPU will not interrupt 'int
 531                          * 3' to raise the DB# exception for the debug
 532                          * register but will postpone the DB# until
 533                          * execution of the first instruction of the
 534                          * BP# handler (in kernel mode).  Normally the
 535                          * previous check would ignore DB# exceptions
 536                          * for watchpoints on user addresses raised in
 537                          * kernel mode.  However, some CPU errata
 538                          * include cases where DB# exceptions do not
 539                          * properly set bits in %dr6, e.g. Haswell
 540                          * HSD23 and Skylake-X SKZ24.
 541                          *
 542                          * A deferred DB# can also be raised on the
 543                          * first instructions of system call entry
 544                          * points or single-step traps via similar use
 545                          * of 'pop %ss' or 'mov xxx, %ss'.
 546                          */
 547                         if (pti) {
 548                                 if (frame->tf_rip ==
 549                                     (uintptr_t)IDTVEC(fast_syscall_pti) ||
 550 #ifdef COMPAT_FREEBSD32
 551                                     frame->tf_rip ==
 552                                     (uintptr_t)IDTVEC(int0x80_syscall_pti) ||
 553 #endif
 554                                     frame->tf_rip == (uintptr_t)IDTVEC(bpt_pti))
 555                                         return;
 556                         } else {
 557                                 if (frame->tf_rip ==
 558                                     (uintptr_t)IDTVEC(fast_syscall) ||
 559 #ifdef COMPAT_FREEBSD32
 560                                     frame->tf_rip ==
 561                                     (uintptr_t)IDTVEC(int0x80_syscall) ||
 562 #endif
 563                                     frame->tf_rip == (uintptr_t)IDTVEC(bpt))
 564                                         return;
 565                         }
 566                         if (frame->tf_rip == (uintptr_t)IDTVEC(dbg) ||
 567                             /* Needed for AMD. */
 568                             frame->tf_rip == (uintptr_t)IDTVEC(fast_syscall32))
 569                                 return;
 570                         /*
 571                          * FALLTHROUGH (TRCTRAP kernel mode, kernel address)
 572                          */
 573                 case T_BPTFLT:
 574                         /*
 575                          * If KDB is enabled, let it handle the debugger trap.
 576                          * Otherwise, debugger traps "can't happen".
 577                          */
 578 #ifdef KDB
 579                         if (kdb_trap(type, dr6, frame))
 580                                 return;
 581 #endif
 582                         break;
 583
 584 #ifdef DEV_ISA
 585                 case T_NMI:
 586                         nmi_handle_intr(type, frame);
 587                         return;
 588 #endif
 589                 }
 590
 591                 trap_fatal(frame, 0);
 592                 return;
 593         }
 594
 595         /* Translate fault for emulators (e.g. Linux) */
 596         if (*p->p_sysent->sv_transtrap != NULL)
 597                 signo = (*p->p_sysent->sv_transtrap)(signo, type);
 598
 599         ksiginfo_init_trap(&ksi);
 600         ksi.ksi_signo = signo;
 601         ksi.ksi_code = ucode;
 602         ksi.ksi_trapno = type;
 603         ksi.ksi_addr = (void *)addr;
 604         if (uprintf_signal) {
 605                 uprintf("pid %d comm %s: signal %d err %lx code %d type %d "
 606                     "addr 0x%lx rsp 0x%lx rip 0x%lx "
 607                     "<%02x %02x %02x %02x %02x %02x %02x %02x>\n",
 608                     p->p_pid, p->p_comm, signo, frame->tf_err, ucode, type,
 609                     addr, frame->tf_rsp, frame->tf_rip,
 610                     fubyte((void *)(frame->tf_rip + 0)),
 611                     fubyte((void *)(frame->tf_rip + 1)),
 612                     fubyte((void *)(frame->tf_rip + 2)),
 613                     fubyte((void *)(frame->tf_rip + 3)),
 614                     fubyte((void *)(frame->tf_rip + 4)),
 615                     fubyte((void *)(frame->tf_rip + 5)),
 616                     fubyte((void *)(frame->tf_rip + 6)),
 617                     fubyte((void *)(frame->tf_rip + 7)));
 618         }
 619         KASSERT((read_rflags() & PSL_I) != 0, ("interrupts disabled"));
 620         trapsignal(td, &ksi);
 621
 622 userret:
 623         userret(td, frame);
 624         KASSERT(PCB_USER_FPU(td->td_pcb),
 625             ("Return from trap with kernel FPU ctx leaked"));
 626 }
 627
 628 /*
 629  * Ensure that we ignore any DTrace-induced faults. This function cannot
 630  * be instrumented, so it cannot generate such faults itself.
 631  */
 632 void
 633 trap_check(struct trapframe *frame)
 634 {
 635
 636 #ifdef KDTRACE_HOOKS
 637         if (dtrace_trap_func != NULL &&
 638             (*dtrace_trap_func)(frame, frame->tf_trapno) != 0)
 639                 return;
 640 #endif
 641         trap(frame);
 642 }
 643
 644 static bool
 645 trap_is_smap(struct trapframe *frame)
 646 {
 647
 648         /*
 649          * A page fault on a userspace address is classified as
 650          * SMAP-induced if:
 651          * - SMAP is supported;
 652          * - kernel mode accessed present data page;
 653          * - rflags.AC was cleared.
 654          * Kernel must never access user space with rflags.AC cleared
 655          * if SMAP is enabled.
 656          */
 657         return ((cpu_stdext_feature & CPUID_STDEXT_SMAP) != 0 &&
 658             (frame->tf_err & (PGEX_P | PGEX_U | PGEX_I | PGEX_RSV)) ==
 659             PGEX_P && (frame->tf_rflags & PSL_AC) == 0);
 660 }
 661
 662 static bool
 663 trap_is_pti(struct trapframe *frame)
 664 {
 665
 666         return (PCPU_GET(curpmap)->pm_ucr3 != PMAP_NO_CR3 &&
 667             pg_nx != 0 && (frame->tf_err & (PGEX_P | PGEX_W |
 668             PGEX_U | PGEX_I)) == (PGEX_P | PGEX_U | PGEX_I) &&
 669             (curpcb->pcb_saved_ucr3 & ~CR3_PCID_MASK) ==
 670             (PCPU_GET(curpmap)->pm_cr3 & ~CR3_PCID_MASK));
 671 }
 672
 673 /*
 674  * Handle all details of a page fault.
 675  * Returns:
 676  * -1 if this fault was fatal, typically from kernel mode
 677  *    (cannot happen, but we need to return something).
 678  * 0  if this fault was handled by updating either the user or kernel
 679  *    page table, execution can continue.
 680  * 1  if this fault was from usermode and it was not handled, a synchronous
 681  *    signal should be delivered to the thread.  *signo returns the signal
 682  *    number, *ucode gives si_code.
 683  */
 684 static int
 685 trap_pfault(struct trapframe *frame, bool usermode, int *signo, int *ucode)
 686 {
 687         struct thread *td;
 688         struct proc *p;
 689         vm_map_t map;
 690         vm_offset_t eva;
 691         int rv;
 692         vm_prot_t ftype;
 693
 694         MPASS(!usermode || (signo != NULL && ucode != NULL));
 695
 696         td = curthread;
 697         p = td->td_proc;
 698         eva = frame->tf_addr;
 699
 700         if (__predict_false((td->td_pflags & TDP_NOFAULTING) != 0)) {
 701                 /*
 702                  * Due to both processor errata and lazy TLB invalidation when
 703                  * access restrictions are removed from virtual pages, memory
 704                  * accesses that are allowed by the physical mapping layer may
 705                  * nonetheless cause one spurious page fault per virtual page.
 706                  * When the thread is executing a "no faulting" section that
 707                  * is bracketed by vm_fault_{disable,enable}_pagefaults(),
 708                  * every page fault is treated as a spurious page fault,
 709                  * unless it accesses the same virtual address as the most
 710                  * recent page fault within the same "no faulting" section.
 711                  */
 712                 if (td->td_md.md_spurflt_addr != eva ||
 713                     (td->td_pflags & TDP_RESETSPUR) != 0) {
 714                         /*
 715                          * Do nothing to the TLB.  A stale TLB entry is
 716                          * flushed automatically by a page fault.
 717                          */
 718                         td->td_md.md_spurflt_addr = eva;
 719                         td->td_pflags &= ~TDP_RESETSPUR;
 720                         return (0);
 721                 }
 722         } else {
 723                 /*
 724                  * If we get a page fault while in a critical section, then
 725                  * it is most likely a fatal kernel page fault.  The kernel
 726                  * is already going to panic trying to get a sleep lock to
 727                  * do the VM lookup, so just consider it a fatal trap so the
 728                  * kernel can print out a useful trap message and even get
 729                  * to the debugger.
 730                  *
 731                  * If we get a page fault while holding a non-sleepable
 732                  * lock, then it is most likely a fatal kernel page fault.
 733                  * If WITNESS is enabled, then it's going to whine about
 734                  * bogus LORs with various VM locks, so just skip to the
 735                  * fatal trap handling directly.
 736                  */
 737                 if (td->td_critnest != 0 ||
 738                     WITNESS_CHECK(WARN_SLEEPOK | WARN_GIANTOK, NULL,
 739                     "Kernel page fault") != 0) {
 740                         trap_fatal(frame, eva);
 741                         return (-1);
 742                 }
 743         }
 744         if (eva >= VM_MIN_KERNEL_ADDRESS) {
 745                 /*
 746                  * Don't allow user-mode faults in kernel address space.
 747                  */
 748                 if (usermode) {
 749                         *signo = SIGSEGV;
 750                         *ucode = SEGV_MAPERR;
 751                         return (1);
 752                 }
 753
 754                 map = kernel_map;
 755         } else {
 756                 map = &p->p_vmspace->vm_map;
 757
 758                 /*
 759                  * When accessing a usermode address, kernel must be
 760                  * ready to accept the page fault, and provide a
 761                  * handling routine.  Since accessing the address
 762                  * without the handler is a bug, do not try to handle
 763                  * it normally, and panic immediately.
 764                  *
 765                  * If SMAP is enabled, filter SMAP faults also,
 766                  * because illegal access might occur to the mapped
 767                  * user address, causing infinite loop.
 768                  */
 769                 if (!usermode && (td->td_intr_nesting_level != 0 ||
 770                     trap_is_smap(frame) || curpcb->pcb_onfault == NULL)) {
 771                         trap_fatal(frame, eva);
 772                         return (-1);
 773                 }
 774         }
 775
 776         /*
 777          * If the trap was caused by errant bits in the PTE then panic.
 778          */
 779         if (frame->tf_err & PGEX_RSV) {
 780                 trap_fatal(frame, eva);
 781                 return (-1);
 782         }
 783
 784         /*
 785          * User-mode protection key violation (PKU).  May happen
 786          * either from usermode or from kernel if copyin accessed
 787          * key-protected mapping.
 788          */
 789         if ((frame->tf_err & PGEX_PK) != 0) {
 790                 if (eva > VM_MAXUSER_ADDRESS) {
 791                         trap_fatal(frame, eva);
 792                         return (-1);
 793                 }
 794                 if (usermode) {
 795                         *signo = SIGSEGV;
 796                         *ucode = SEGV_PKUERR;
 797                         return (1);
 798                 }
 799                 goto after_vmfault;
 800         }
 801
 802         /*
 803          * If nx protection of the usermode portion of kernel page
 804          * tables caused trap, panic.
 805          */
 806         if (usermode && trap_is_pti(frame))
 807                 panic("PTI: pid %d comm %s tf_err %#lx", p->p_pid,
 808                     p->p_comm, frame->tf_err);
 809
 810         /*
 811          * PGEX_I is defined only if the execute disable bit capability is
 812          * supported and enabled.
 813          */
 814         if (frame->tf_err & PGEX_W)
 815                 ftype = VM_PROT_WRITE;
 816         else if ((frame->tf_err & PGEX_I) && pg_nx != 0)
 817                 ftype = VM_PROT_EXECUTE;
 818         else
 819                 ftype = VM_PROT_READ;
 820
 821         /* Fault in the page. */
 822         rv = vm_fault_trap(map, eva, ftype, VM_FAULT_NORMAL, signo, ucode);
 823         if (rv == KERN_SUCCESS) {
 824 #ifdef HWPMC_HOOKS
 825                 if (ftype == VM_PROT_READ || ftype == VM_PROT_WRITE) {
 826                         PMC_SOFT_CALL_TF( , , page_fault, all, frame);
 827                         if (ftype == VM_PROT_READ)
 828                                 PMC_SOFT_CALL_TF( , , page_fault, read,
 829                                     frame);
 830                         else
 831                                 PMC_SOFT_CALL_TF( , , page_fault, write,
 832                                     frame);
 833                 }
 834 #endif
 835                 return (0);
 836         }
 837
 838         if (usermode)
 839                 return (1);
 840 after_vmfault:
 841         if (td->td_intr_nesting_level == 0 &&
 842             curpcb->pcb_onfault != NULL) {
 843                 frame->tf_rip = (long)curpcb->pcb_onfault;
 844                 return (0);
 845         }
 846         trap_fatal(frame, eva);
 847         return (-1);
 848 }
 849
 850 static void
 851 trap_fatal(frame, eva)
 852         struct trapframe *frame;
 853         vm_offset_t eva;
 854 {
 855         int code, ss;
 856         u_int type;
 857         struct soft_segment_descriptor softseg;
 858         struct user_segment_descriptor *gdt;
 859 #ifdef KDB
 860         bool handled;
 861 #endif
 862
 863         code = frame->tf_err;
 864         type = frame->tf_trapno;
 865         gdt = *PCPU_PTR(gdt);
 866         sdtossd(&gdt[IDXSEL(frame->tf_cs & 0xffff)], &softseg);
 867
 868         printf("\n\nFatal trap %d: %s while in %s mode\n", type,
 869             type < nitems(trap_msg) ? trap_msg[type] : UNKNOWN,
 870             TRAPF_USERMODE(frame) ? "user" : "kernel");
 871 #ifdef SMP
 872         /* two separate prints in case of a trap on an unmapped page */
 873         printf("cpuid = %d; ", PCPU_GET(cpuid));
 874         printf("apic id = %02x\n", PCPU_GET(apic_id));
 875 #endif
 876         if (type == T_PAGEFLT) {
 877                 printf("fault virtual address   = 0x%lx\n", eva);
 878                 printf("fault code              = %s %s %s%s%s, %s\n",
 879                         code & PGEX_U ? "user" : "supervisor",
 880                         code & PGEX_W ? "write" : "read",
 881                         code & PGEX_I ? "instruction" : "data",
 882                         code & PGEX_PK ? " prot key" : "",
 883                         code & PGEX_SGX ? " SGX" : "",
 884                         code & PGEX_RSV ? "reserved bits in PTE" :
 885                         code & PGEX_P ? "protection violation" : "page not present");
 886         }
 887         printf("instruction pointer     = 0x%lx:0x%lx\n",
 888                frame->tf_cs & 0xffff, frame->tf_rip);
 889         ss = frame->tf_ss & 0xffff;
 890         printf("stack pointer           = 0x%x:0x%lx\n", ss, frame->tf_rsp);
 891         printf("frame pointer           = 0x%x:0x%lx\n", ss, frame->tf_rbp);
 892         printf("code segment            = base 0x%lx, limit 0x%lx, type 0x%x\n",
 893                softseg.ssd_base, softseg.ssd_limit, softseg.ssd_type);
 894         printf("                        = DPL %d, pres %d, long %d, def32 %d, gran %d\n",
 895                softseg.ssd_dpl, softseg.ssd_p, softseg.ssd_long, softseg.ssd_def32,
 896                softseg.ssd_gran);
 897         printf("processor eflags        = ");
 898         if (frame->tf_rflags & PSL_T)
 899                 printf("trace trap, ");
 900         if (frame->tf_rflags & PSL_I)
 901                 printf("interrupt enabled, ");
 902         if (frame->tf_rflags & PSL_NT)
 903                 printf("nested task, ");
 904         if (frame->tf_rflags & PSL_RF)
 905                 printf("resume, ");
 906         printf("IOPL = %ld\n", (frame->tf_rflags & PSL_IOPL) >> 12);
 907         printf("current process         = %d (%s)\n",
 908             curproc->p_pid, curthread->td_name);
 909
 910 #ifdef KDB
 911         if (debugger_on_trap) {
 912                 kdb_why = KDB_WHY_TRAP;
 913                 handled = kdb_trap(type, 0, frame);
 914                 kdb_why = KDB_WHY_UNSET;
 915                 if (handled)
 916                         return;
 917         }
 918 #endif
 919         printf("trap number             = %d\n", type);
 920         panic("%s", type < nitems(trap_msg) ? trap_msg[type] :
 921             "unknown/reserved trap");
 922 }
 923
 924 #ifdef KDTRACE_HOOKS
 925 /*
 926  * Invoke a userspace DTrace hook.  The hook pointer is cleared when no
 927  * userspace probes are enabled, so we must synchronize with DTrace to ensure
 928  * that a trapping thread is able to call the hook before it is cleared.
 929  */
 930 static bool
 931 trap_user_dtrace(struct trapframe *frame, int (**hookp)(struct trapframe *))
 932 {
 933         int (*hook)(struct trapframe *);
 934
 935         hook = (int (*)(struct trapframe *))atomic_load_ptr(hookp);
 936         enable_intr();
 937         if (hook != NULL)
 938                 return ((hook)(frame) == 0);
 939         return (false);
 940 }
 941 #endif
 942
 943 /*
 944  * Double fault handler. Called when a fault occurs while writing
 945  * a frame for a trap/exception onto the stack. This usually occurs
 946  * when the stack overflows (such is the case with infinite recursion,
 947  * for example).
 948  */
 949 void
 950 dblfault_handler(struct trapframe *frame)
 951 {
 952 #ifdef KDTRACE_HOOKS
 953         if (dtrace_doubletrap_func != NULL)
 954                 (*dtrace_doubletrap_func)();
 955 #endif
 956         printf("\nFatal double fault\n"
 957             "rip %#lx rsp %#lx rbp %#lx\n"
 958             "rax %#lx rdx %#lx rbx %#lx\n"
 959             "rcx %#lx rsi %#lx rdi %#lx\n"
 960             "r8 %#lx r9 %#lx r10 %#lx\n"
 961             "r11 %#lx r12 %#lx r13 %#lx\n"
 962             "r14 %#lx r15 %#lx rflags %#lx\n"
 963             "cs %#lx ss %#lx ds %#hx es %#hx fs %#hx gs %#hx\n"
 964             "fsbase %#lx gsbase %#lx kgsbase %#lx\n",
 965             frame->tf_rip, frame->tf_rsp, frame->tf_rbp,
 966             frame->tf_rax, frame->tf_rdx, frame->tf_rbx,
 967             frame->tf_rcx, frame->tf_rdi, frame->tf_rsi,
 968             frame->tf_r8, frame->tf_r9, frame->tf_r10,
 969             frame->tf_r11, frame->tf_r12, frame->tf_r13,
 970             frame->tf_r14, frame->tf_r15, frame->tf_rflags,
 971             frame->tf_cs, frame->tf_ss, frame->tf_ds, frame->tf_es,
 972             frame->tf_fs, frame->tf_gs,
 973             rdmsr(MSR_FSBASE), rdmsr(MSR_GSBASE), rdmsr(MSR_KGSBASE));
 974 #ifdef SMP
 975         /* two separate prints in case of a trap on an unmapped page */
 976         printf("cpuid = %d; ", PCPU_GET(cpuid));
 977         printf("apic id = %02x\n", PCPU_GET(apic_id));
 978 #endif
 979         panic("double fault");
 980 }
 981
 982 static int __noinline
 983 cpu_fetch_syscall_args_fallback(struct thread *td, struct syscall_args *sa)
 984 {
 985         struct proc *p;
 986         struct trapframe *frame;
 987         register_t *argp;
 988         caddr_t params;
 989         int reg, regcnt, error;
 990
 991         p = td->td_proc;
 992         frame = td->td_frame;
 993         reg = 0;
 994         regcnt = NARGREGS;
 995
 996         sa->code = frame->tf_rax;
 997
 998         if (sa->code == SYS_syscall || sa->code == SYS___syscall) {
 999                 sa->code = frame->tf_rdi;
1000                 reg++;
1001                 regcnt--;
1002         }
1003
1004         if (sa->code >= p->p_sysent->sv_size)
1005                 sa->callp = &p->p_sysent->sv_table[0];
1006         else
1007                 sa->callp = &p->p_sysent->sv_table[sa->code];
1008
1009         sa->narg = sa->callp->sy_narg;
1010         KASSERT(sa->narg <= nitems(sa->args), ("Too many syscall arguments!"));
1011         argp = &frame->tf_rdi;
1012         argp += reg;
1013         memcpy(sa->args, argp, sizeof(sa->args[0]) * NARGREGS);
1014         if (sa->narg > regcnt) {
1015                 params = (caddr_t)frame->tf_rsp + sizeof(register_t);
1016                 error = copyin(params, &sa->args[regcnt],
1017                     (sa->narg - regcnt) * sizeof(sa->args[0]));
1018                 if (__predict_false(error != 0))
1019                         return (error);
1020         }
1021
1022         td->td_retval[0] = 0;
1023         td->td_retval[1] = frame->tf_rdx;
1024
1025         return (0);
1026 }
1027
1028 int
1029 cpu_fetch_syscall_args(struct thread *td)
1030 {
1031         struct proc *p;
1032         struct trapframe *frame;
1033         struct syscall_args *sa;
1034
1035         p = td->td_proc;
1036         frame = td->td_frame;
1037         sa = &td->td_sa;
1038
1039         sa->code = frame->tf_rax;
1040
1041         if (__predict_false(sa->code == SYS_syscall ||
1042             sa->code == SYS___syscall ||
1043             sa->code >= p->p_sysent->sv_size))
1044                 return (cpu_fetch_syscall_args_fallback(td, sa));
1045
1046         sa->callp = &p->p_sysent->sv_table[sa->code];
1047         sa->narg = sa->callp->sy_narg;
1048         KASSERT(sa->narg <= nitems(sa->args), ("Too many syscall arguments!"));
1049
1050         if (__predict_false(sa->narg > NARGREGS))
1051                 return (cpu_fetch_syscall_args_fallback(td, sa));
1052
1053         memcpy(sa->args, &frame->tf_rdi, sizeof(sa->args[0]) * NARGREGS);
1054
1055         td->td_retval[0] = 0;
1056         td->td_retval[1] = frame->tf_rdx;
1057
1058         return (0);
1059 }
1060
1061 #include "../../kern/subr_syscall.c"
1062
1063 static void (*syscall_ret_l1d_flush)(void);
1064 int syscall_ret_l1d_flush_mode;
1065
1066 static void
1067 flush_l1d_hw(void)
1068 {
1069
1070         wrmsr(MSR_IA32_FLUSH_CMD, IA32_FLUSH_CMD_L1D);
1071 }
1072
1073 static void __inline
1074 amd64_syscall_ret_flush_l1d_inline(int error)
1075 {
1076         void (*p)(void);
1077
1078         if (error != 0 && error != EEXIST && error != EAGAIN &&
1079             error != EXDEV && error != ENOENT && error != ENOTCONN &&
1080             error != EINPROGRESS) {
1081                 p = syscall_ret_l1d_flush;
1082                 if (p != NULL)
1083                         p();
1084         }
1085 }
1086
1087 void
1088 amd64_syscall_ret_flush_l1d(int error)
1089 {
1090
1091         amd64_syscall_ret_flush_l1d_inline(error);
1092 }
1093
1094 void
1095 amd64_syscall_ret_flush_l1d_recalc(void)
1096 {
1097         bool l1d_hw;
1098
1099         l1d_hw = (cpu_stdext_feature3 & CPUID_STDEXT3_L1D_FLUSH) != 0;
1100 again:
1101         switch (syscall_ret_l1d_flush_mode) {
1102         case 0:
1103                 syscall_ret_l1d_flush = NULL;
1104                 break;
1105         case 1:
1106                 syscall_ret_l1d_flush = l1d_hw ? flush_l1d_hw :
1107                     flush_l1d_sw_abi;
1108                 break;
1109         case 2:
1110                 syscall_ret_l1d_flush = l1d_hw ? flush_l1d_hw : NULL;
1111                 break;
1112         case 3:
1113                 syscall_ret_l1d_flush = flush_l1d_sw_abi;
1114                 break;
1115         default:
1116                 syscall_ret_l1d_flush_mode = 1;
1117                 goto again;
1118         }
1119 }
1120
1121 static int
1122 machdep_syscall_ret_flush_l1d(SYSCTL_HANDLER_ARGS)
1123 {
1124         int error, val;
1125
1126         val = syscall_ret_l1d_flush_mode;
1127         error = sysctl_handle_int(oidp, &val, 0, req);
1128         if (error != 0 || req->newptr == NULL)
1129                 return (error);
1130         syscall_ret_l1d_flush_mode = val;
1131         amd64_syscall_ret_flush_l1d_recalc();
1132         return (0);
1133 }
1134 SYSCTL_PROC(_machdep, OID_AUTO, syscall_ret_flush_l1d, CTLTYPE_INT |
1135     CTLFLAG_RWTUN | CTLFLAG_NOFETCH | CTLFLAG_MPSAFE, NULL, 0,
1136     machdep_syscall_ret_flush_l1d, "I",
1137     "Flush L1D on syscall return with error (0 - off, 1 - on, "
1138     "2 - use hw only, 3 - use sw only");
1139
1140
1141 /*
1142  * System call handler for native binaries.  The trap frame is already
1143  * set up by the assembler trampoline and a pointer to it is saved in
1144  * td_frame.
1145  */
1146 void
1147 amd64_syscall(struct thread *td, int traced)
1148 {
1149         ksiginfo_t ksi;
1150
1151 #ifdef DIAGNOSTIC
1152         if (!TRAPF_USERMODE(td->td_frame)) {
1153                 panic("syscall");
1154                 /* NOT REACHED */
1155         }
1156 #endif
1157         syscallenter(td);
1158
1159         /*
1160          * Traced syscall.
1161          */
1162         if (__predict_false(traced)) {
1163                 td->td_frame->tf_rflags &= ~PSL_T;
1164                 ksiginfo_init_trap(&ksi);
1165                 ksi.ksi_signo = SIGTRAP;
1166                 ksi.ksi_code = TRAP_TRACE;
1167                 ksi.ksi_addr = (void *)td->td_frame->tf_rip;
1168                 trapsignal(td, &ksi);
1169         }
1170
1171         KASSERT(PCB_USER_FPU(td->td_pcb),
1172             ("System call %s returning with kernel FPU ctx leaked",
1173              syscallname(td->td_proc, td->td_sa.code)));
1174         KASSERT(td->td_pcb->pcb_save == get_pcb_user_save_td(td),
1175             ("System call %s returning with mangled pcb_save",
1176              syscallname(td->td_proc, td->td_sa.code)));
1177         KASSERT(pmap_not_in_di(),
1178             ("System call %s returning with leaked invl_gen %lu",
1179             syscallname(td->td_proc, td->td_sa.code),
1180             td->td_md.md_invl_gen.gen));
1181
1182         syscallret(td);
1183
1184         /*
1185          * If the user-supplied value of %rip is not a canonical
1186          * address, then some CPUs will trigger a ring 0 #GP during
1187          * the sysret instruction.  However, the fault handler would
1188          * execute in ring 0 with the user's %gs and %rsp which would
1189          * not be safe.  Instead, use the full return path which
1190          * catches the problem safely.
1191          */
1192         if (__predict_false(td->td_frame->tf_rip >= VM_MAXUSER_ADDRESS))
1193                 set_pcb_flags(td->td_pcb, PCB_FULL_IRET);
1194
1195         amd64_syscall_ret_flush_l1d_inline(td->td_errno);
1196 }