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 #include "opt_stack.h"
  56
  57 #include <sys/param.h>
  58 #include <sys/bus.h>
  59 #include <sys/systm.h>
  60 #include <sys/proc.h>
  61 #include <sys/pioctl.h>
  62 #include <sys/ptrace.h>
  63 #include <sys/kdb.h>
  64 #include <sys/kernel.h>
  65 #include <sys/ktr.h>
  66 #include <sys/lock.h>
  67 #include <sys/mutex.h>
  68 #include <sys/resourcevar.h>
  69 #include <sys/signalvar.h>
  70 #include <sys/syscall.h>
  71 #include <sys/sysctl.h>
  72 #include <sys/sysent.h>
  73 #include <sys/uio.h>
  74 #include <sys/vmmeter.h>
  75 #ifdef HWPMC_HOOKS
  76 #include <sys/pmckern.h>
  77 PMC_SOFT_DEFINE( , , page_fault, all);
  78 PMC_SOFT_DEFINE( , , page_fault, read);
  79 PMC_SOFT_DEFINE( , , page_fault, write);
  80 #endif
  81
  82 #include <vm/vm.h>
  83 #include <vm/vm_param.h>
  84 #include <vm/pmap.h>
  85 #include <vm/vm_kern.h>
  86 #include <vm/vm_map.h>
  87 #include <vm/vm_page.h>
  88 #include <vm/vm_extern.h>
  89
  90 #include <machine/cpu.h>
  91 #include <machine/intr_machdep.h>
  92 #include <x86/mca.h>
  93 #include <machine/md_var.h>
  94 #include <machine/pcb.h>
  95 #ifdef SMP
  96 #include <machine/smp.h>
  97 #endif
  98 #include <machine/stack.h>
  99 #include <machine/trap.h>
 100 #include <machine/tss.h>
 101
 102 #ifdef KDTRACE_HOOKS
 103 #include <sys/dtrace_bsd.h>
 104 #endif
 105
 106 extern inthand_t IDTVEC(bpt), IDTVEC(bpt_pti), IDTVEC(dbg),
 107     IDTVEC(fast_syscall), IDTVEC(fast_syscall_pti), IDTVEC(fast_syscall32),
 108     IDTVEC(int0x80_syscall_pti), IDTVEC(int0x80_syscall);
 109
 110 void __noinline trap(struct trapframe *frame);
 111 void trap_check(struct trapframe *frame);
 112 void dblfault_handler(struct trapframe *frame);
 113
 114 static int trap_pfault(struct trapframe *, bool, int *, int *);
 115 static void trap_fatal(struct trapframe *, vm_offset_t);
 116 #ifdef KDTRACE_HOOKS
 117 static bool trap_user_dtrace(struct trapframe *,
 118     int (**hook)(struct trapframe *));
 119 #endif
 120
 121 static const char UNKNOWN[] = "unknown";
 122 static const char *const trap_msg[] = {
 123         [0] =                   UNKNOWN,                        /* unused */
 124         [T_PRIVINFLT] =         "privileged instruction fault",
 125         [2] =                   UNKNOWN,                        /* unused */
 126         [T_BPTFLT] =            "breakpoint instruction fault",
 127         [4] =                   UNKNOWN,                        /* unused */
 128         [5] =                   UNKNOWN,                        /* unused */
 129         [T_ARITHTRAP] =         "arithmetic trap",
 130         [7] =                   UNKNOWN,                        /* unused */
 131         [8] =                   UNKNOWN,                        /* unused */
 132         [T_PROTFLT] =           "general protection fault",
 133         [T_TRCTRAP] =           "debug exception",
 134         [11] =                  UNKNOWN,                        /* unused */
 135         [T_PAGEFLT] =           "page fault",
 136         [13] =                  UNKNOWN,                        /* unused */
 137         [T_ALIGNFLT] =          "alignment fault",
 138         [15] =                  UNKNOWN,                        /* unused */
 139         [16] =                  UNKNOWN,                        /* unused */
 140         [17] =                  UNKNOWN,                        /* unused */
 141         [T_DIVIDE] =            "integer divide fault",
 142         [T_NMI] =               "non-maskable interrupt trap",
 143         [T_OFLOW] =             "overflow trap",
 144         [T_BOUND] =             "FPU bounds check fault",
 145         [T_DNA] =               "FPU device not available",
 146         [T_DOUBLEFLT] =         "double fault",
 147         [T_FPOPFLT] =           "FPU operand fetch fault",
 148         [T_TSSFLT] =            "invalid TSS fault",
 149         [T_SEGNPFLT] =          "segment not present fault",
 150         [T_STKFLT] =            "stack fault",
 151         [T_MCHK] =              "machine check trap",
 152         [T_XMMFLT] =            "SIMD floating-point exception",
 153         [T_RESERVED] =          "reserved (unknown) fault",
 154         [31] =                  UNKNOWN,                        /* reserved */
 155         [T_DTRACE_RET] =        "DTrace pid return trap",
 156 };
 157
 158 static int uprintf_signal;
 159 SYSCTL_INT(_machdep, OID_AUTO, uprintf_signal, CTLFLAG_RWTUN,
 160     &uprintf_signal, 0,
 161     "Print debugging information on trap signal to ctty");
 162
 163 /*
 164  * Control L1D flush on return from NMI.
 165  *
 166  * Tunable  can be set to the following values:
 167  * 0 - only enable flush on return from NMI if required by vmm.ko (default)
 168  * >1 - always flush on return from NMI.
 169  *
 170  * Post-boot, the sysctl indicates if flushing is currently enabled.
 171  */
 172 int nmi_flush_l1d_sw;
 173 SYSCTL_INT(_machdep, OID_AUTO, nmi_flush_l1d_sw, CTLFLAG_RWTUN,
 174     &nmi_flush_l1d_sw, 0,
 175     "Flush L1 Data Cache on NMI exit, software bhyve L1TF mitigation assist");
 176
 177 /*
 178  * Exception, fault, and trap interface to the FreeBSD kernel.
 179  * This common code is called from assembly language IDT gate entry
 180  * routines that prepare a suitable stack frame, and restore this
 181  * frame after the exception has been processed.
 182  */
 183
 184 void
 185 trap(struct trapframe *frame)
 186 {
 187         ksiginfo_t ksi;
 188         struct thread *td;
 189         struct proc *p;
 190         register_t addr, dr6;
 191         int pf, signo, ucode;
 192         u_int type;
 193
 194         td = curthread;
 195         p = td->td_proc;
 196         dr6 = 0;
 197
 198         VM_CNT_INC(v_trap);
 199         type = frame->tf_trapno;
 200
 201 #ifdef SMP
 202         /* Handler for NMI IPIs used for stopping CPUs. */
 203         if (type == T_NMI && ipi_nmi_handler() == 0)
 204                 return;
 205 #endif
 206
 207 #ifdef KDB
 208         if (kdb_active) {
 209                 kdb_reenter();
 210                 return;
 211         }
 212 #endif
 213
 214         if (type == T_RESERVED) {
 215                 trap_fatal(frame, 0);
 216                 return;
 217         }
 218
 219         if (type == T_NMI) {
 220 #ifdef HWPMC_HOOKS
 221                 /*
 222                  * CPU PMCs interrupt using an NMI.  If the PMC module is
 223                  * active, pass the 'rip' value to the PMC module's interrupt
 224                  * handler.  A non-zero return value from the handler means that
 225                  * the NMI was consumed by it and we can return immediately.
 226                  */
 227                 if (pmc_intr != NULL &&
 228                     (*pmc_intr)(frame) != 0)
 229                         return;
 230 #endif
 231
 232 #ifdef STACK
 233                 if (stack_nmi_handler(frame) != 0)
 234                         return;
 235 #endif
 236         }
 237
 238         if ((frame->tf_rflags & PSL_I) == 0) {
 239                 /*
 240                  * Buggy application or kernel code has disabled
 241                  * interrupts and then trapped.  Enabling interrupts
 242                  * now is wrong, but it is better than running with
 243                  * interrupts disabled until they are accidentally
 244                  * enabled later.
 245                  */
 246                 if (TRAPF_USERMODE(frame))
 247                         uprintf(
 248                             "pid %ld (%s): trap %d with interrupts disabled\n",
 249                             (long)curproc->p_pid, curthread->td_name, type);
 250                 else if (type != T_NMI && type != T_BPTFLT &&
 251                     type != T_TRCTRAP) {
 252                         /*
 253                          * XXX not quite right, since this may be for a
 254                          * multiple fault in user mode.
 255                          */
 256                         printf("kernel trap %d with interrupts disabled\n",
 257                             type);
 258
 259                         /*
 260                          * We shouldn't enable interrupts while holding a
 261                          * spin lock.
 262                          */
 263                         if (td->td_md.md_spinlock_count == 0)
 264                                 enable_intr();
 265                 }
 266         }
 267
 268         if (TRAPF_USERMODE(frame)) {
 269                 /* user trap */
 270
 271                 td->td_pticks = 0;
 272                 td->td_frame = frame;
 273                 addr = frame->tf_rip;
 274                 if (td->td_cowgen != p->p_cowgen)
 275                         thread_cow_update(td);
 276
 277                 switch (type) {
 278                 case T_PRIVINFLT:       /* privileged instruction fault */
 279                         signo = SIGILL;
 280                         ucode = ILL_PRVOPC;
 281                         break;
 282
 283                 case T_BPTFLT:          /* bpt instruction fault */
 284 #ifdef KDTRACE_HOOKS
 285                         if (trap_user_dtrace(frame, &dtrace_pid_probe_ptr))
 286                                 return;
 287 #else
 288                         enable_intr();
 289 #endif
 290                         signo = SIGTRAP;
 291                         ucode = TRAP_BRKPT;
 292                         break;
 293
 294                 case T_TRCTRAP:         /* debug exception */
 295                         enable_intr();
 296                         signo = SIGTRAP;
 297                         ucode = TRAP_TRACE;
 298                         dr6 = rdr6();
 299                         if ((dr6 & DBREG_DR6_BS) != 0) {
 300                                 PROC_LOCK(td->td_proc);
 301                                 if ((td->td_dbgflags & TDB_STEP) != 0) {
 302                                         td->td_frame->tf_rflags &= ~PSL_T;
 303                                         td->td_dbgflags &= ~TDB_STEP;
 304                                 }
 305                                 PROC_UNLOCK(td->td_proc);
 306                         }
 307                         break;
 308
 309                 case T_ARITHTRAP:       /* arithmetic trap */
 310                         ucode = fputrap_x87();
 311                         if (ucode == -1)
 312                                 return;
 313                         signo = SIGFPE;
 314                         break;
 315
 316                 case T_PROTFLT:         /* general protection fault */
 317                         signo = SIGBUS;
 318                         ucode = BUS_OBJERR;
 319                         break;
 320                 case T_STKFLT:          /* stack fault */
 321                 case T_SEGNPFLT:        /* segment not present fault */
 322                         signo = SIGBUS;
 323                         ucode = BUS_ADRERR;
 324                         break;
 325                 case T_TSSFLT:          /* invalid TSS fault */
 326                         signo = SIGBUS;
 327                         ucode = BUS_OBJERR;
 328                         break;
 329                 case T_ALIGNFLT:
 330                         signo = SIGBUS;
 331                         ucode = BUS_ADRALN;
 332                         break;
 333                 case T_DOUBLEFLT:       /* double fault */
 334                 default:
 335                         signo = SIGBUS;
 336                         ucode = BUS_OBJERR;
 337                         break;
 338
 339                 case T_PAGEFLT:         /* page fault */
 340                         /*
 341                          * Can emulator handle this trap?
 342                          */
 343                         if (*p->p_sysent->sv_trap != NULL &&
 344                             (*p->p_sysent->sv_trap)(td) == 0)
 345                                 return;
 346
 347                         pf = trap_pfault(frame, true, &signo, &ucode);
 348                         if (pf == -1)
 349                                 return;
 350                         if (pf == 0)
 351                                 goto userret;
 352                         addr = frame->tf_addr;
 353                         break;
 354
 355                 case T_DIVIDE:          /* integer divide fault */
 356                         ucode = FPE_INTDIV;
 357                         signo = SIGFPE;
 358                         break;
 359
 360 #ifdef DEV_ISA
 361                 case T_NMI:
 362                         nmi_handle_intr(type, frame);
 363                         return;
 364 #endif
 365
 366                 case T_OFLOW:           /* integer overflow fault */
 367                         ucode = FPE_INTOVF;
 368                         signo = SIGFPE;
 369                         break;
 370
 371                 case T_BOUND:           /* bounds check fault */
 372                         ucode = FPE_FLTSUB;
 373                         signo = SIGFPE;
 374                         break;
 375
 376                 case T_DNA:
 377                         /* transparent fault (due to context switch "late") */
 378                         KASSERT(PCB_USER_FPU(td->td_pcb),
 379                             ("kernel FPU ctx has leaked"));
 380                         fpudna();
 381                         return;
 382
 383                 case T_FPOPFLT:         /* FPU operand fetch fault */
 384                         ucode = ILL_COPROC;
 385                         signo = SIGILL;
 386                         break;
 387
 388                 case T_XMMFLT:          /* SIMD floating-point exception */
 389                         ucode = fputrap_sse();
 390                         if (ucode == -1)
 391                                 return;
 392                         signo = SIGFPE;
 393                         break;
 394 #ifdef KDTRACE_HOOKS
 395                 case T_DTRACE_RET:
 396                         (void)trap_user_dtrace(frame, &dtrace_return_probe_ptr);
 397                         return;
 398 #endif
 399                 }
 400         } else {
 401                 /* kernel trap */
 402
 403                 KASSERT(cold || td->td_ucred != NULL,
 404                     ("kernel trap doesn't have ucred"));
 405                 switch (type) {
 406                 case T_PAGEFLT:                 /* page fault */
 407                         (void)trap_pfault(frame, false, NULL, NULL);
 408                         return;
 409
 410                 case T_DNA:
 411                         if (PCB_USER_FPU(td->td_pcb))
 412                                 panic("Unregistered use of FPU in kernel");
 413                         fpudna();
 414                         return;
 415
 416                 case T_ARITHTRAP:       /* arithmetic trap */
 417                 case T_XMMFLT:          /* SIMD floating-point exception */
 418                 case T_FPOPFLT:         /* FPU operand fetch fault */
 419                         /*
 420                          * For now, supporting kernel handler
 421                          * registration for FPU traps is overkill.
 422                          */
 423                         trap_fatal(frame, 0);
 424                         return;
 425
 426                 case T_STKFLT:          /* stack fault */
 427                 case T_PROTFLT:         /* general protection fault */
 428                 case T_SEGNPFLT:        /* segment not present fault */
 429                         if (td->td_intr_nesting_level != 0)
 430                                 break;
 431
 432                         /*
 433                          * Invalid segment selectors and out of bounds
 434                          * %rip's and %rsp's can be set up in user mode.
 435                          * This causes a fault in kernel mode when the
 436                          * kernel tries to return to user mode.  We want
 437                          * to get this fault so that we can fix the
 438                          * problem here and not have to check all the
 439                          * selectors and pointers when the user changes
 440                          * them.
 441                          *
 442                          * In case of PTI, the IRETQ faulted while the
 443                          * kernel used the pti stack, and exception
 444                          * frame records %rsp value pointing to that
 445                          * stack.  If we return normally to
 446                          * doreti_iret_fault, the trapframe is
 447                          * reconstructed on pti stack, and calltrap()
 448                          * called on it as well.  Due to the very
 449                          * limited pti stack size, kernel does not
 450                          * survive for too long.  Switch to the normal
 451                          * thread stack for the trap handling.
 452                          *
 453                          * Magic '5' is the number of qwords occupied by
 454                          * the hardware trap frame.
 455                          */
 456                         if (frame->tf_rip == (long)doreti_iret) {
 457                                 frame->tf_rip = (long)doreti_iret_fault;
 458                                 if ((PCPU_GET(curpmap)->pm_ucr3 !=
 459                                     PMAP_NO_CR3) &&
 460                                     (frame->tf_rsp == (uintptr_t)PCPU_GET(
 461                                     pti_rsp0) - 5 * sizeof(register_t))) {
 462                                         frame->tf_rsp = PCPU_GET(rsp0) - 5 *
 463                                             sizeof(register_t);
 464                                 }
 465                                 return;
 466                         }
 467                         if (frame->tf_rip == (long)ld_ds) {
 468                                 frame->tf_rip = (long)ds_load_fault;
 469                                 return;
 470                         }
 471                         if (frame->tf_rip == (long)ld_es) {
 472                                 frame->tf_rip = (long)es_load_fault;
 473                                 return;
 474                         }
 475                         if (frame->tf_rip == (long)ld_fs) {
 476                                 frame->tf_rip = (long)fs_load_fault;
 477                                 return;
 478                         }
 479                         if (frame->tf_rip == (long)ld_gs) {
 480                                 frame->tf_rip = (long)gs_load_fault;
 481                                 return;
 482                         }
 483                         if (frame->tf_rip == (long)ld_gsbase) {
 484                                 frame->tf_rip = (long)gsbase_load_fault;
 485                                 return;
 486                         }
 487                         if (frame->tf_rip == (long)ld_fsbase) {
 488                                 frame->tf_rip = (long)fsbase_load_fault;
 489                                 return;
 490                         }
 491                         if (curpcb->pcb_onfault != NULL) {
 492                                 frame->tf_rip = (long)curpcb->pcb_onfault;
 493                                 return;
 494                         }
 495                         break;
 496
 497                 case T_TSSFLT:
 498                         /*
 499                          * PSL_NT can be set in user mode and isn't cleared
 500                          * automatically when the kernel is entered.  This
 501                          * causes a TSS fault when the kernel attempts to
 502                          * `iret' because the TSS link is uninitialized.  We
 503                          * want to get this fault so that we can fix the
 504                          * problem here and not every time the kernel is
 505                          * entered.
 506                          */
 507                         if (frame->tf_rflags & PSL_NT) {
 508                                 frame->tf_rflags &= ~PSL_NT;
 509                                 return;
 510                         }
 511                         break;
 512
 513                 case T_TRCTRAP:  /* debug exception */
 514                         /* Clear any pending debug events. */
 515                         dr6 = rdr6();
 516                         load_dr6(0);
 517
 518                         /*
 519                          * Ignore debug register exceptions due to
 520                          * accesses in the user's address space, which
 521                          * can happen under several conditions such as
 522                          * if a user sets a watchpoint on a buffer and
 523                          * then passes that buffer to a system call.
 524                          * We still want to get TRCTRAPS for addresses
 525                          * in kernel space because that is useful when
 526                          * debugging the kernel.
 527                          */
 528                         if (user_dbreg_trap(dr6))
 529                                 return;
 530
 531                         /*
 532                          * Malicious user code can configure a debug
 533                          * register watchpoint to trap on data access
 534                          * to the top of stack and then execute 'pop
 535                          * %ss; int 3'.  Due to exception deferral for
 536                          * 'pop %ss', the CPU will not interrupt 'int
 537                          * 3' to raise the DB# exception for the debug
 538                          * register but will postpone the DB# until
 539                          * execution of the first instruction of the
 540                          * BP# handler (in kernel mode).  Normally the
 541                          * previous check would ignore DB# exceptions
 542                          * for watchpoints on user addresses raised in
 543                          * kernel mode.  However, some CPU errata
 544                          * include cases where DB# exceptions do not
 545                          * properly set bits in %dr6, e.g. Haswell
 546                          * HSD23 and Skylake-X SKZ24.
 547                          *
 548                          * A deferred DB# can also be raised on the
 549                          * first instructions of system call entry
 550                          * points or single-step traps via similar use
 551                          * of 'pop %ss' or 'mov xxx, %ss'.
 552                          */
 553                         if (pti) {
 554                                 if (frame->tf_rip ==
 555                                     (uintptr_t)IDTVEC(fast_syscall_pti) ||
 556 #ifdef COMPAT_FREEBSD32
 557                                     frame->tf_rip ==
 558                                     (uintptr_t)IDTVEC(int0x80_syscall_pti) ||
 559 #endif
 560                                     frame->tf_rip == (uintptr_t)IDTVEC(bpt_pti))
 561                                         return;
 562                         } else {
 563                                 if (frame->tf_rip ==
 564                                     (uintptr_t)IDTVEC(fast_syscall) ||
 565 #ifdef COMPAT_FREEBSD32
 566                                     frame->tf_rip ==
 567                                     (uintptr_t)IDTVEC(int0x80_syscall) ||
 568 #endif
 569                                     frame->tf_rip == (uintptr_t)IDTVEC(bpt))
 570                                         return;
 571                         }
 572                         if (frame->tf_rip == (uintptr_t)IDTVEC(dbg) ||
 573                             /* Needed for AMD. */
 574                             frame->tf_rip == (uintptr_t)IDTVEC(fast_syscall32))
 575                                 return;
 576                         /*
 577                          * FALLTHROUGH (TRCTRAP kernel mode, kernel address)
 578                          */
 579                 case T_BPTFLT:
 580                         /*
 581                          * If KDB is enabled, let it handle the debugger trap.
 582                          * Otherwise, debugger traps "can't happen".
 583                          */
 584 #ifdef KDB
 585                         if (kdb_trap(type, dr6, frame))
 586                                 return;
 587 #endif
 588                         break;
 589
 590 #ifdef DEV_ISA
 591                 case T_NMI:
 592                         nmi_handle_intr(type, frame);
 593                         return;
 594 #endif
 595                 }
 596
 597                 trap_fatal(frame, 0);
 598                 return;
 599         }
 600
 601         /* Translate fault for emulators (e.g. Linux) */
 602         if (*p->p_sysent->sv_transtrap != NULL)
 603                 signo = (*p->p_sysent->sv_transtrap)(signo, type);
 604
 605         ksiginfo_init_trap(&ksi);
 606         ksi.ksi_signo = signo;
 607         ksi.ksi_code = ucode;
 608         ksi.ksi_trapno = type;
 609         ksi.ksi_addr = (void *)addr;
 610         if (uprintf_signal) {
 611                 uprintf("pid %d comm %s: signal %d err %lx code %d type %d "
 612                     "addr 0x%lx rsp 0x%lx rip 0x%lx "
 613                     "<%02x %02x %02x %02x %02x %02x %02x %02x>\n",
 614                     p->p_pid, p->p_comm, signo, frame->tf_err, ucode, type,
 615                     addr, frame->tf_rsp, frame->tf_rip,
 616                     fubyte((void *)(frame->tf_rip + 0)),
 617                     fubyte((void *)(frame->tf_rip + 1)),
 618                     fubyte((void *)(frame->tf_rip + 2)),
 619                     fubyte((void *)(frame->tf_rip + 3)),
 620                     fubyte((void *)(frame->tf_rip + 4)),
 621                     fubyte((void *)(frame->tf_rip + 5)),
 622                     fubyte((void *)(frame->tf_rip + 6)),
 623                     fubyte((void *)(frame->tf_rip + 7)));
 624         }
 625         KASSERT((read_rflags() & PSL_I) != 0, ("interrupts disabled"));
 626         trapsignal(td, &ksi);
 627
 628 userret:
 629         userret(td, frame);
 630         KASSERT(PCB_USER_FPU(td->td_pcb),
 631             ("Return from trap with kernel FPU ctx leaked"));
 632 }
 633
 634 /*
 635  * Ensure that we ignore any DTrace-induced faults. This function cannot
 636  * be instrumented, so it cannot generate such faults itself.
 637  */
 638 void
 639 trap_check(struct trapframe *frame)
 640 {
 641
 642 #ifdef KDTRACE_HOOKS
 643         if (dtrace_trap_func != NULL &&
 644             (*dtrace_trap_func)(frame, frame->tf_trapno) != 0)
 645                 return;
 646 #endif
 647         trap(frame);
 648 }
 649
 650 static bool
 651 trap_is_smap(struct trapframe *frame)
 652 {
 653
 654         /*
 655          * A page fault on a userspace address is classified as
 656          * SMAP-induced if:
 657          * - SMAP is supported;
 658          * - kernel mode accessed present data page;
 659          * - rflags.AC was cleared.
 660          * Kernel must never access user space with rflags.AC cleared
 661          * if SMAP is enabled.
 662          */
 663         return ((cpu_stdext_feature & CPUID_STDEXT_SMAP) != 0 &&
 664             (frame->tf_err & (PGEX_P | PGEX_U | PGEX_I | PGEX_RSV)) ==
 665             PGEX_P && (frame->tf_rflags & PSL_AC) == 0);
 666 }
 667
 668 static bool
 669 trap_is_pti(struct trapframe *frame)
 670 {
 671
 672         return (PCPU_GET(curpmap)->pm_ucr3 != PMAP_NO_CR3 &&
 673             pg_nx != 0 && (frame->tf_err & (PGEX_P | PGEX_W |
 674             PGEX_U | PGEX_I)) == (PGEX_P | PGEX_U | PGEX_I) &&
 675             (curpcb->pcb_saved_ucr3 & ~CR3_PCID_MASK) ==
 676             (PCPU_GET(curpmap)->pm_cr3 & ~CR3_PCID_MASK));
 677 }
 678
 679 /*
 680  * Handle all details of a page fault.
 681  * Returns:
 682  * -1 if this fault was fatal, typically from kernel mode
 683  *    (cannot happen, but we need to return something).
 684  * 0  if this fault was handled by updating either the user or kernel
 685  *    page table, execution can continue.
 686  * 1  if this fault was from usermode and it was not handled, a synchronous
 687  *    signal should be delivered to the thread.  *signo returns the signal
 688  *    number, *ucode gives si_code.
 689  */
 690 static int
 691 trap_pfault(struct trapframe *frame, bool usermode, int *signo, int *ucode)
 692 {
 693         struct thread *td;
 694         struct proc *p;
 695         vm_map_t map;
 696         vm_offset_t eva;
 697         int rv;
 698         vm_prot_t ftype;
 699
 700         MPASS(!usermode || (signo != NULL && ucode != NULL));
 701
 702         td = curthread;
 703         p = td->td_proc;
 704         eva = frame->tf_addr;
 705
 706         if (__predict_false((td->td_pflags & TDP_NOFAULTING) != 0)) {
 707                 /*
 708                  * Due to both processor errata and lazy TLB invalidation when
 709                  * access restrictions are removed from virtual pages, memory
 710                  * accesses that are allowed by the physical mapping layer may
 711                  * nonetheless cause one spurious page fault per virtual page.
 712                  * When the thread is executing a "no faulting" section that
 713                  * is bracketed by vm_fault_{disable,enable}_pagefaults(),
 714                  * every page fault is treated as a spurious page fault,
 715                  * unless it accesses the same virtual address as the most
 716                  * recent page fault within the same "no faulting" section.
 717                  */
 718                 if (td->td_md.md_spurflt_addr != eva ||
 719                     (td->td_pflags & TDP_RESETSPUR) != 0) {
 720                         /*
 721                          * Do nothing to the TLB.  A stale TLB entry is
 722                          * flushed automatically by a page fault.
 723                          */
 724                         td->td_md.md_spurflt_addr = eva;
 725                         td->td_pflags &= ~TDP_RESETSPUR;
 726                         return (0);
 727                 }
 728         } else {
 729                 /*
 730                  * If we get a page fault while in a critical section, then
 731                  * it is most likely a fatal kernel page fault.  The kernel
 732                  * is already going to panic trying to get a sleep lock to
 733                  * do the VM lookup, so just consider it a fatal trap so the
 734                  * kernel can print out a useful trap message and even get
 735                  * to the debugger.
 736                  *
 737                  * If we get a page fault while holding a non-sleepable
 738                  * lock, then it is most likely a fatal kernel page fault.
 739                  * If WITNESS is enabled, then it's going to whine about
 740                  * bogus LORs with various VM locks, so just skip to the
 741                  * fatal trap handling directly.
 742                  */
 743                 if (td->td_critnest != 0 ||
 744                     WITNESS_CHECK(WARN_SLEEPOK | WARN_GIANTOK, NULL,
 745                     "Kernel page fault") != 0) {
 746                         trap_fatal(frame, eva);
 747                         return (-1);
 748                 }
 749         }
 750         if (eva >= VM_MIN_KERNEL_ADDRESS) {
 751                 /*
 752                  * Don't allow user-mode faults in kernel address space.
 753                  */
 754                 if (usermode) {
 755                         *signo = SIGSEGV;
 756                         *ucode = SEGV_MAPERR;
 757                         return (1);
 758                 }
 759
 760                 map = kernel_map;
 761         } else {
 762                 map = &p->p_vmspace->vm_map;
 763
 764                 /*
 765                  * When accessing a usermode address, kernel must be
 766                  * ready to accept the page fault, and provide a
 767                  * handling routine.  Since accessing the address
 768                  * without the handler is a bug, do not try to handle
 769                  * it normally, and panic immediately.
 770                  *
 771                  * If SMAP is enabled, filter SMAP faults also,
 772                  * because illegal access might occur to the mapped
 773                  * user address, causing infinite loop.
 774                  */
 775                 if (!usermode && (td->td_intr_nesting_level != 0 ||
 776                     trap_is_smap(frame) || curpcb->pcb_onfault == NULL)) {
 777                         trap_fatal(frame, eva);
 778                         return (-1);
 779                 }
 780         }
 781
 782         /*
 783          * If the trap was caused by errant bits in the PTE then panic.
 784          */
 785         if (frame->tf_err & PGEX_RSV) {
 786                 trap_fatal(frame, eva);
 787                 return (-1);
 788         }
 789
 790         /*
 791          * User-mode protection key violation (PKU).  May happen
 792          * either from usermode or from kernel if copyin accessed
 793          * key-protected mapping.
 794          */
 795         if ((frame->tf_err & PGEX_PK) != 0) {
 796                 if (eva > VM_MAXUSER_ADDRESS) {
 797                         trap_fatal(frame, eva);
 798                         return (-1);
 799                 }
 800                 if (usermode) {
 801                         *signo = SIGSEGV;
 802                         *ucode = SEGV_PKUERR;
 803                         return (1);
 804                 }
 805                 goto after_vmfault;
 806         }
 807
 808         /*
 809          * If nx protection of the usermode portion of kernel page
 810          * tables caused trap, panic.
 811          */
 812         if (usermode && trap_is_pti(frame))
 813                 panic("PTI: pid %d comm %s tf_err %#lx", p->p_pid,
 814                     p->p_comm, frame->tf_err);
 815
 816         /*
 817          * PGEX_I is defined only if the execute disable bit capability is
 818          * supported and enabled.
 819          */
 820         if (frame->tf_err & PGEX_W)
 821                 ftype = VM_PROT_WRITE;
 822         else if ((frame->tf_err & PGEX_I) && pg_nx != 0)
 823                 ftype = VM_PROT_EXECUTE;
 824         else
 825                 ftype = VM_PROT_READ;
 826
 827         /* Fault in the page. */
 828         rv = vm_fault_trap(map, eva, ftype, VM_FAULT_NORMAL, signo, ucode);
 829         if (rv == KERN_SUCCESS) {
 830 #ifdef HWPMC_HOOKS
 831                 if (ftype == VM_PROT_READ || ftype == VM_PROT_WRITE) {
 832                         PMC_SOFT_CALL_TF( , , page_fault, all, frame);
 833                         if (ftype == VM_PROT_READ)
 834                                 PMC_SOFT_CALL_TF( , , page_fault, read,
 835                                     frame);
 836                         else
 837                                 PMC_SOFT_CALL_TF( , , page_fault, write,
 838                                     frame);
 839                 }
 840 #endif
 841                 return (0);
 842         }
 843
 844         if (usermode)
 845                 return (1);
 846 after_vmfault:
 847         if (td->td_intr_nesting_level == 0 &&
 848             curpcb->pcb_onfault != NULL) {
 849                 frame->tf_rip = (long)curpcb->pcb_onfault;
 850                 return (0);
 851         }
 852         trap_fatal(frame, eva);
 853         return (-1);
 854 }
 855
 856 static void
 857 trap_fatal(frame, eva)
 858         struct trapframe *frame;
 859         vm_offset_t eva;
 860 {
 861         int code, ss;
 862         u_int type;
 863         struct soft_segment_descriptor softseg;
 864         struct user_segment_descriptor *gdt;
 865 #ifdef KDB
 866         bool handled;
 867 #endif
 868
 869         code = frame->tf_err;
 870         type = frame->tf_trapno;
 871         gdt = *PCPU_PTR(gdt);
 872         sdtossd(&gdt[IDXSEL(frame->tf_cs & 0xffff)], &softseg);
 873
 874         printf("\n\nFatal trap %d: %s while in %s mode\n", type,
 875             type < nitems(trap_msg) ? trap_msg[type] : UNKNOWN,
 876             TRAPF_USERMODE(frame) ? "user" : "kernel");
 877 #ifdef SMP
 878         /* two separate prints in case of a trap on an unmapped page */
 879         printf("cpuid = %d; ", PCPU_GET(cpuid));
 880         printf("apic id = %02x\n", PCPU_GET(apic_id));
 881 #endif
 882         if (type == T_PAGEFLT) {
 883                 printf("fault virtual address   = 0x%lx\n", eva);
 884                 printf("fault code              = %s %s %s%s%s, %s\n",
 885                         code & PGEX_U ? "user" : "supervisor",
 886                         code & PGEX_W ? "write" : "read",
 887                         code & PGEX_I ? "instruction" : "data",
 888                         code & PGEX_PK ? " prot key" : "",
 889                         code & PGEX_SGX ? " SGX" : "",
 890                         code & PGEX_RSV ? "reserved bits in PTE" :
 891                         code & PGEX_P ? "protection violation" : "page not present");
 892         }
 893         printf("instruction pointer     = 0x%lx:0x%lx\n",
 894                frame->tf_cs & 0xffff, frame->tf_rip);
 895         ss = frame->tf_ss & 0xffff;
 896         printf("stack pointer           = 0x%x:0x%lx\n", ss, frame->tf_rsp);
 897         printf("frame pointer           = 0x%x:0x%lx\n", ss, frame->tf_rbp);
 898         printf("code segment            = base 0x%lx, limit 0x%lx, type 0x%x\n",
 899                softseg.ssd_base, softseg.ssd_limit, softseg.ssd_type);
 900         printf("                        = DPL %d, pres %d, long %d, def32 %d, gran %d\n",
 901                softseg.ssd_dpl, softseg.ssd_p, softseg.ssd_long, softseg.ssd_def32,
 902                softseg.ssd_gran);
 903         printf("processor eflags        = ");
 904         if (frame->tf_rflags & PSL_T)
 905                 printf("trace trap, ");
 906         if (frame->tf_rflags & PSL_I)
 907                 printf("interrupt enabled, ");
 908         if (frame->tf_rflags & PSL_NT)
 909                 printf("nested task, ");
 910         if (frame->tf_rflags & PSL_RF)
 911                 printf("resume, ");
 912         printf("IOPL = %ld\n", (frame->tf_rflags & PSL_IOPL) >> 12);
 913         printf("current process         = %d (%s)\n",
 914             curproc->p_pid, curthread->td_name);
 915
 916 #ifdef KDB
 917         if (debugger_on_trap) {
 918                 kdb_why = KDB_WHY_TRAP;
 919                 handled = kdb_trap(type, 0, frame);
 920                 kdb_why = KDB_WHY_UNSET;
 921                 if (handled)
 922                         return;
 923         }
 924 #endif
 925         printf("trap number             = %d\n", type);
 926         panic("%s", type < nitems(trap_msg) ? trap_msg[type] :
 927             "unknown/reserved trap");
 928 }
 929
 930 #ifdef KDTRACE_HOOKS
 931 /*
 932  * Invoke a userspace DTrace hook.  The hook pointer is cleared when no
 933  * userspace probes are enabled, so we must synchronize with DTrace to ensure
 934  * that a trapping thread is able to call the hook before it is cleared.
 935  */
 936 static bool
 937 trap_user_dtrace(struct trapframe *frame, int (**hookp)(struct trapframe *))
 938 {
 939         int (*hook)(struct trapframe *);
 940
 941         hook = (int (*)(struct trapframe *))atomic_load_ptr(hookp);
 942         enable_intr();
 943         if (hook != NULL)
 944                 return ((hook)(frame) == 0);
 945         return (false);
 946 }
 947 #endif
 948
 949 /*
 950  * Double fault handler. Called when a fault occurs while writing
 951  * a frame for a trap/exception onto the stack. This usually occurs
 952  * when the stack overflows (such is the case with infinite recursion,
 953  * for example).
 954  */
 955 void
 956 dblfault_handler(struct trapframe *frame)
 957 {
 958 #ifdef KDTRACE_HOOKS
 959         if (dtrace_doubletrap_func != NULL)
 960                 (*dtrace_doubletrap_func)();
 961 #endif
 962         printf("\nFatal double fault\n"
 963             "rip %#lx rsp %#lx rbp %#lx\n"
 964             "rax %#lx rdx %#lx rbx %#lx\n"
 965             "rcx %#lx rsi %#lx rdi %#lx\n"
 966             "r8 %#lx r9 %#lx r10 %#lx\n"
 967             "r11 %#lx r12 %#lx r13 %#lx\n"
 968             "r14 %#lx r15 %#lx rflags %#lx\n"
 969             "cs %#lx ss %#lx ds %#hx es %#hx fs %#hx gs %#hx\n"
 970             "fsbase %#lx gsbase %#lx kgsbase %#lx\n",
 971             frame->tf_rip, frame->tf_rsp, frame->tf_rbp,
 972             frame->tf_rax, frame->tf_rdx, frame->tf_rbx,
 973             frame->tf_rcx, frame->tf_rdi, frame->tf_rsi,
 974             frame->tf_r8, frame->tf_r9, frame->tf_r10,
 975             frame->tf_r11, frame->tf_r12, frame->tf_r13,
 976             frame->tf_r14, frame->tf_r15, frame->tf_rflags,
 977             frame->tf_cs, frame->tf_ss, frame->tf_ds, frame->tf_es,
 978             frame->tf_fs, frame->tf_gs,
 979             rdmsr(MSR_FSBASE), rdmsr(MSR_GSBASE), rdmsr(MSR_KGSBASE));
 980 #ifdef SMP
 981         /* two separate prints in case of a trap on an unmapped page */
 982         printf("cpuid = %d; ", PCPU_GET(cpuid));
 983         printf("apic id = %02x\n", PCPU_GET(apic_id));
 984 #endif
 985         panic("double fault");
 986 }
 987
 988 static int __noinline
 989 cpu_fetch_syscall_args_fallback(struct thread *td, struct syscall_args *sa)
 990 {
 991         struct proc *p;
 992         struct trapframe *frame;
 993         register_t *argp;
 994         caddr_t params;
 995         int reg, regcnt, error;
 996
 997         p = td->td_proc;
 998         frame = td->td_frame;
 999         reg = 0;
1000         regcnt = NARGREGS;
1001
1002         sa->code = frame->tf_rax;
1003
1004         if (sa->code == SYS_syscall || sa->code == SYS___syscall) {
1005                 sa->code = frame->tf_rdi;
1006                 reg++;
1007                 regcnt--;
1008         }
1009
1010         if (sa->code >= p->p_sysent->sv_size)
1011                 sa->callp = &p->p_sysent->sv_table[0];
1012         else
1013                 sa->callp = &p->p_sysent->sv_table[sa->code];
1014
1015         sa->narg = sa->callp->sy_narg;
1016         KASSERT(sa->narg <= nitems(sa->args), ("Too many syscall arguments!"));
1017         argp = &frame->tf_rdi;
1018         argp += reg;
1019         memcpy(sa->args, argp, sizeof(sa->args[0]) * NARGREGS);
1020         if (sa->narg > regcnt) {
1021                 params = (caddr_t)frame->tf_rsp + sizeof(register_t);
1022                 error = copyin(params, &sa->args[regcnt],
1023                     (sa->narg - regcnt) * sizeof(sa->args[0]));
1024                 if (__predict_false(error != 0))
1025                         return (error);
1026         }
1027
1028         td->td_retval[0] = 0;
1029         td->td_retval[1] = frame->tf_rdx;
1030
1031         return (0);
1032 }
1033
1034 int
1035 cpu_fetch_syscall_args(struct thread *td)
1036 {
1037         struct proc *p;
1038         struct trapframe *frame;
1039         struct syscall_args *sa;
1040
1041         p = td->td_proc;
1042         frame = td->td_frame;
1043         sa = &td->td_sa;
1044
1045         sa->code = frame->tf_rax;
1046
1047         if (__predict_false(sa->code == SYS_syscall ||
1048             sa->code == SYS___syscall ||
1049             sa->code >= p->p_sysent->sv_size))
1050                 return (cpu_fetch_syscall_args_fallback(td, sa));
1051
1052         sa->callp = &p->p_sysent->sv_table[sa->code];
1053         sa->narg = sa->callp->sy_narg;
1054         KASSERT(sa->narg <= nitems(sa->args), ("Too many syscall arguments!"));
1055
1056         if (__predict_false(sa->narg > NARGREGS))
1057                 return (cpu_fetch_syscall_args_fallback(td, sa));
1058
1059         memcpy(sa->args, &frame->tf_rdi, sizeof(sa->args[0]) * NARGREGS);
1060
1061         td->td_retval[0] = 0;
1062         td->td_retval[1] = frame->tf_rdx;
1063
1064         return (0);
1065 }
1066
1067 #include "../../kern/subr_syscall.c"
1068
1069 static void (*syscall_ret_l1d_flush)(void);
1070 int syscall_ret_l1d_flush_mode;
1071
1072 static void
1073 flush_l1d_hw(void)
1074 {
1075
1076         wrmsr(MSR_IA32_FLUSH_CMD, IA32_FLUSH_CMD_L1D);
1077 }
1078
1079 static void __inline
1080 amd64_syscall_ret_flush_l1d_inline(int error)
1081 {
1082         void (*p)(void);
1083
1084         if (error != 0 && error != EEXIST && error != EAGAIN &&
1085             error != EXDEV && error != ENOENT && error != ENOTCONN &&
1086             error != EINPROGRESS) {
1087                 p = syscall_ret_l1d_flush;
1088                 if (p != NULL)
1089                         p();
1090         }
1091 }
1092
1093 void
1094 amd64_syscall_ret_flush_l1d(int error)
1095 {
1096
1097         amd64_syscall_ret_flush_l1d_inline(error);
1098 }
1099
1100 void
1101 amd64_syscall_ret_flush_l1d_recalc(void)
1102 {
1103         bool l1d_hw;
1104
1105         l1d_hw = (cpu_stdext_feature3 & CPUID_STDEXT3_L1D_FLUSH) != 0;
1106 again:
1107         switch (syscall_ret_l1d_flush_mode) {
1108         case 0:
1109                 syscall_ret_l1d_flush = NULL;
1110                 break;
1111         case 1:
1112                 syscall_ret_l1d_flush = l1d_hw ? flush_l1d_hw :
1113                     flush_l1d_sw_abi;
1114                 break;
1115         case 2:
1116                 syscall_ret_l1d_flush = l1d_hw ? flush_l1d_hw : NULL;
1117                 break;
1118         case 3:
1119                 syscall_ret_l1d_flush = flush_l1d_sw_abi;
1120                 break;
1121         default:
1122                 syscall_ret_l1d_flush_mode = 1;
1123                 goto again;
1124         }
1125 }
1126
1127 static int
1128 machdep_syscall_ret_flush_l1d(SYSCTL_HANDLER_ARGS)
1129 {
1130         int error, val;
1131
1132         val = syscall_ret_l1d_flush_mode;
1133         error = sysctl_handle_int(oidp, &val, 0, req);
1134         if (error != 0 || req->newptr == NULL)
1135                 return (error);
1136         syscall_ret_l1d_flush_mode = val;
1137         amd64_syscall_ret_flush_l1d_recalc();
1138         return (0);
1139 }
1140 SYSCTL_PROC(_machdep, OID_AUTO, syscall_ret_flush_l1d, CTLTYPE_INT |
1141     CTLFLAG_RWTUN | CTLFLAG_NOFETCH | CTLFLAG_MPSAFE, NULL, 0,
1142     machdep_syscall_ret_flush_l1d, "I",
1143     "Flush L1D on syscall return with error (0 - off, 1 - on, "
1144     "2 - use hw only, 3 - use sw only");
1145
1146
1147 /*
1148  * System call handler for native binaries.  The trap frame is already
1149  * set up by the assembler trampoline and a pointer to it is saved in
1150  * td_frame.
1151  */
1152 void
1153 amd64_syscall(struct thread *td, int traced)
1154 {
1155         ksiginfo_t ksi;
1156
1157 #ifdef DIAGNOSTIC
1158         if (!TRAPF_USERMODE(td->td_frame)) {
1159                 panic("syscall");
1160                 /* NOT REACHED */
1161         }
1162 #endif
1163         syscallenter(td);
1164
1165         /*
1166          * Traced syscall.
1167          */
1168         if (__predict_false(traced)) {
1169                 td->td_frame->tf_rflags &= ~PSL_T;
1170                 ksiginfo_init_trap(&ksi);
1171                 ksi.ksi_signo = SIGTRAP;
1172                 ksi.ksi_code = TRAP_TRACE;
1173                 ksi.ksi_addr = (void *)td->td_frame->tf_rip;
1174                 trapsignal(td, &ksi);
1175         }
1176
1177         KASSERT(PCB_USER_FPU(td->td_pcb),
1178             ("System call %s returning with kernel FPU ctx leaked",
1179              syscallname(td->td_proc, td->td_sa.code)));
1180         KASSERT(td->td_pcb->pcb_save == get_pcb_user_save_td(td),
1181             ("System call %s returning with mangled pcb_save",
1182              syscallname(td->td_proc, td->td_sa.code)));
1183         KASSERT(pmap_not_in_di(),
1184             ("System call %s returning with leaked invl_gen %lu",
1185             syscallname(td->td_proc, td->td_sa.code),
1186             td->td_md.md_invl_gen.gen));
1187
1188         syscallret(td);
1189
1190         /*
1191          * If the user-supplied value of %rip is not a canonical
1192          * address, then some CPUs will trigger a ring 0 #GP during
1193          * the sysret instruction.  However, the fault handler would
1194          * execute in ring 0 with the user's %gs and %rsp which would
1195          * not be safe.  Instead, use the full return path which
1196          * catches the problem safely.
1197          */
1198         if (__predict_false(td->td_frame->tf_rip >= VM_MAXUSER_ADDRESS))
1199                 set_pcb_flags(td->td_pcb, PCB_FULL_IRET);
1200
1201         amd64_syscall_ret_flush_l1d_inline(td->td_errno);
1202 }