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