6583 remove whole-process swapping

   1 /*
   2  * CDDL HEADER START
   3  *
   4  * The contents of this file are subject to the terms of the
   5  * Common Development and Distribution License (the "License").
   6  * You may not use this file except in compliance with the License.
   7  *
   8  * You can obtain a copy of the license at usr/src/OPENSOLARIS.LICENSE
   9  * or http://www.opensolaris.org/os/licensing.
  10  * See the License for the specific language governing permissions
  11  * and limitations under the License.
  12  *
  13  * When distributing Covered Code, include this CDDL HEADER in each
  14  * file and include the License file at usr/src/OPENSOLARIS.LICENSE.
  15  * If applicable, add the following below this CDDL HEADER, with the
  16  * fields enclosed by brackets "[]" replaced with your own identifying
  17  * information: Portions Copyright [yyyy] [name of copyright owner]
  18  *
  19  * CDDL HEADER END
  20  */
  21 
  22 /*
  23  * Copyright (c) 1992, 2010, Oracle and/or its affiliates. All rights reserved.
  24  */
  25 
  26 /*      Copyright (c) 1990, 1991 UNIX System Laboratories, Inc. */
  27 /*      Copyright (c) 1984, 1986, 1987, 1988, 1989, 1990 AT&T   */
  28 /*              All Rights Reserved                             */
  29 /*                                                              */
  30 /*      Copyright (c) 1987, 1988 Microsoft Corporation          */
  31 /*              All Rights Reserved                             */
  32 /*                                                              */
  33 
  34 /*
  35  * Copyright 2012 Joyent, Inc. All rights reserved.
  36  */
  37 
  38 #include <sys/types.h>
  39 #include <sys/sysmacros.h>
  40 #include <sys/param.h>
  41 #include <sys/signal.h>
  42 #include <sys/systm.h>
  43 #include <sys/user.h>
  44 #include <sys/proc.h>
  45 #include <sys/disp.h>
  46 #include <sys/class.h>
  47 #include <sys/core.h>
  48 #include <sys/syscall.h>
  49 #include <sys/cpuvar.h>
  50 #include <sys/vm.h>
  51 #include <sys/sysinfo.h>
  52 #include <sys/fault.h>
  53 #include <sys/stack.h>
  54 #include <sys/psw.h>
  55 #include <sys/regset.h>
  56 #include <sys/fp.h>
  57 #include <sys/trap.h>
  58 #include <sys/kmem.h>
  59 #include <sys/vtrace.h>
  60 #include <sys/cmn_err.h>
  61 #include <sys/prsystm.h>
  62 #include <sys/mutex_impl.h>
  63 #include <sys/machsystm.h>
  64 #include <sys/archsystm.h>
  65 #include <sys/sdt.h>
  66 #include <sys/avintr.h>
  67 #include <sys/kobj.h>
  68 
  69 #include <vm/hat.h>
  70 
  71 #include <vm/seg_kmem.h>
  72 #include <vm/as.h>
  73 #include <vm/seg.h>
  74 #include <vm/hat_pte.h>
  75 #include <vm/hat_i86.h>
  76 
  77 #include <sys/procfs.h>
  78 
  79 #include <sys/reboot.h>
  80 #include <sys/debug.h>
  81 #include <sys/debugreg.h>
  82 #include <sys/modctl.h>
  83 #include <sys/aio_impl.h>
  84 #include <sys/tnf.h>
  85 #include <sys/tnf_probe.h>
  86 #include <sys/cred.h>
  87 #include <sys/mman.h>
  88 #include <sys/x86_archext.h>
  89 #include <sys/copyops.h>
  90 #include <c2/audit.h>
  91 #include <sys/ftrace.h>
  92 #include <sys/panic.h>
  93 #include <sys/traptrace.h>
  94 #include <sys/ontrap.h>
  95 #include <sys/cpc_impl.h>
  96 #include <sys/bootconf.h>
  97 #include <sys/bootinfo.h>
  98 #include <sys/promif.h>
  99 #include <sys/mach_mmu.h>
 100 #if defined(__xpv)
 101 #include <sys/hypervisor.h>
 102 #endif
 103 #include <sys/contract/process_impl.h>
 104 
 105 #define USER    0x10000         /* user-mode flag added to trap type */
 106 
 107 static const char *trap_type_mnemonic[] = {
 108         "de",   "db",   "2",    "bp",
 109         "of",   "br",   "ud",   "nm",
 110         "df",   "9",    "ts",   "np",
 111         "ss",   "gp",   "pf",   "15",
 112         "mf",   "ac",   "mc",   "xf"
 113 };
 114 
 115 static const char *trap_type[] = {
 116         "Divide error",                         /* trap id 0    */
 117         "Debug",                                /* trap id 1    */
 118         "NMI interrupt",                        /* trap id 2    */
 119         "Breakpoint",                           /* trap id 3    */
 120         "Overflow",                             /* trap id 4    */
 121         "BOUND range exceeded",                 /* trap id 5    */
 122         "Invalid opcode",                       /* trap id 6    */
 123         "Device not available",                 /* trap id 7    */
 124         "Double fault",                         /* trap id 8    */
 125         "Coprocessor segment overrun",          /* trap id 9    */
 126         "Invalid TSS",                          /* trap id 10   */
 127         "Segment not present",                  /* trap id 11   */
 128         "Stack segment fault",                  /* trap id 12   */
 129         "General protection",                   /* trap id 13   */
 130         "Page fault",                           /* trap id 14   */
 131         "Reserved",                             /* trap id 15   */
 132         "x87 floating point error",             /* trap id 16   */
 133         "Alignment check",                      /* trap id 17   */
 134         "Machine check",                        /* trap id 18   */
 135         "SIMD floating point exception",        /* trap id 19   */
 136 };
 137 
 138 #define TRAP_TYPES      (sizeof (trap_type) / sizeof (trap_type[0]))
 139 
 140 #define SLOW_SCALL_SIZE 2
 141 #define FAST_SCALL_SIZE 2
 142 
 143 int tudebug = 0;
 144 int tudebugbpt = 0;
 145 int tudebugfpe = 0;
 146 int tudebugsse = 0;
 147 
 148 #if defined(TRAPDEBUG) || defined(lint)
 149 int tdebug = 0;
 150 int lodebug = 0;
 151 int faultdebug = 0;
 152 #else
 153 #define tdebug  0
 154 #define lodebug 0
 155 #define faultdebug      0
 156 #endif /* defined(TRAPDEBUG) || defined(lint) */
 157 
 158 #if defined(TRAPTRACE)
 159 /*
 160  * trap trace record for cpu0 is allocated here.
 161  * trap trace records for non-boot cpus are allocated in mp_startup_init().
 162  */
 163 static trap_trace_rec_t trap_tr0[TRAPTR_NENT];
 164 trap_trace_ctl_t trap_trace_ctl[NCPU] = {
 165         {
 166             (uintptr_t)trap_tr0,                        /* next record */
 167             (uintptr_t)trap_tr0,                        /* first record */
 168             (uintptr_t)(trap_tr0 + TRAPTR_NENT),        /* limit */
 169             (uintptr_t)0                                /* current */
 170         },
 171 };
 172 
 173 /*
 174  * default trap buffer size
 175  */
 176 size_t trap_trace_bufsize = TRAPTR_NENT * sizeof (trap_trace_rec_t);
 177 int trap_trace_freeze = 0;
 178 int trap_trace_off = 0;
 179 
 180 /*
 181  * A dummy TRAPTRACE entry to use after death.
 182  */
 183 trap_trace_rec_t trap_trace_postmort;
 184 
 185 static void dump_ttrace(void);
 186 #endif  /* TRAPTRACE */
 187 static void dumpregs(struct regs *);
 188 static void showregs(uint_t, struct regs *, caddr_t);
 189 static int kern_gpfault(struct regs *);
 190 
 191 /*ARGSUSED*/
 192 static int
 193 die(uint_t type, struct regs *rp, caddr_t addr, processorid_t cpuid)
 194 {
 195         struct panic_trap_info ti;
 196         const char *trap_name, *trap_mnemonic;
 197 
 198         if (type < TRAP_TYPES) {
 199                 trap_name = trap_type[type];
 200                 trap_mnemonic = trap_type_mnemonic[type];
 201         } else {
 202                 trap_name = "trap";
 203                 trap_mnemonic = "-";
 204         }
 205 
 206 #ifdef TRAPTRACE
 207         TRAPTRACE_FREEZE;
 208 #endif
 209 
 210         ti.trap_regs = rp;
 211         ti.trap_type = type & ~USER;
 212         ti.trap_addr = addr;
 213 
 214         curthread->t_panic_trap = &ti;
 215 
 216         if (type == T_PGFLT && addr < (caddr_t)KERNELBASE) {
 217                 panic("BAD TRAP: type=%x (#%s %s) rp=%p addr=%p "
 218                     "occurred in module \"%s\" due to %s",
 219                     type, trap_mnemonic, trap_name, (void *)rp, (void *)addr,
 220                     mod_containing_pc((caddr_t)rp->r_pc),
 221                     addr < (caddr_t)PAGESIZE ?
 222                     "a NULL pointer dereference" :
 223                     "an illegal access to a user address");
 224         } else
 225                 panic("BAD TRAP: type=%x (#%s %s) rp=%p addr=%p",
 226                     type, trap_mnemonic, trap_name, (void *)rp, (void *)addr);
 227         return (0);
 228 }
 229 
 230 /*
 231  * Rewrite the instruction at pc to be an int $T_SYSCALLINT instruction.
 232  *
 233  * int <vector> is two bytes: 0xCD <vector>
 234  */
 235 
 236 static int
 237 rewrite_syscall(caddr_t pc)
 238 {
 239         uchar_t instr[SLOW_SCALL_SIZE] = { 0xCD, T_SYSCALLINT };
 240 
 241         if (uwrite(curthread->t_procp, instr, SLOW_SCALL_SIZE,
 242             (uintptr_t)pc) != 0)
 243                 return (1);
 244 
 245         return (0);
 246 }
 247 
 248 /*
 249  * Test to see if the instruction at pc is sysenter or syscall. The second
 250  * argument should be the x86 feature flag corresponding to the expected
 251  * instruction.
 252  *
 253  * sysenter is two bytes: 0x0F 0x34
 254  * syscall is two bytes:  0x0F 0x05
 255  * int $T_SYSCALLINT is two bytes: 0xCD 0x91
 256  */
 257 
 258 static int
 259 instr_is_other_syscall(caddr_t pc, int which)
 260 {
 261         uchar_t instr[FAST_SCALL_SIZE];
 262 
 263         ASSERT(which == X86FSET_SEP || which == X86FSET_ASYSC || which == 0xCD);
 264 
 265         if (copyin_nowatch(pc, (caddr_t)instr, FAST_SCALL_SIZE) != 0)
 266                 return (0);
 267 
 268         switch (which) {
 269         case X86FSET_SEP:
 270                 if (instr[0] == 0x0F && instr[1] == 0x34)
 271                         return (1);
 272                 break;
 273         case X86FSET_ASYSC:
 274                 if (instr[0] == 0x0F && instr[1] == 0x05)
 275                         return (1);
 276                 break;
 277         case 0xCD:
 278                 if (instr[0] == 0xCD && instr[1] == T_SYSCALLINT)
 279                         return (1);
 280                 break;
 281         }
 282 
 283         return (0);
 284 }
 285 
 286 static const char *
 287 syscall_insn_string(int syscall_insn)
 288 {
 289         switch (syscall_insn) {
 290         case X86FSET_SEP:
 291                 return ("sysenter");
 292         case X86FSET_ASYSC:
 293                 return ("syscall");
 294         case 0xCD:
 295                 return ("int");
 296         default:
 297                 return ("Unknown");
 298         }
 299 }
 300 
 301 static int
 302 ldt_rewrite_syscall(struct regs *rp, proc_t *p, int syscall_insn)
 303 {
 304         caddr_t linearpc;
 305         int return_code = 0;
 306 
 307         mutex_enter(&p->p_ldtlock);      /* Must be held across linear_pc() */
 308 
 309         if (linear_pc(rp, p, &linearpc) == 0) {
 310 
 311                 /*
 312                  * If another thread beat us here, it already changed
 313                  * this site to the slower (int) syscall instruction.
 314                  */
 315                 if (instr_is_other_syscall(linearpc, 0xCD)) {
 316                         return_code = 1;
 317                 } else if (instr_is_other_syscall(linearpc, syscall_insn)) {
 318 
 319                         if (rewrite_syscall(linearpc) == 0) {
 320                                 return_code = 1;
 321                         }
 322 #ifdef DEBUG
 323                         else
 324                                 cmn_err(CE_WARN, "failed to rewrite %s "
 325                                     "instruction in process %d",
 326                                     syscall_insn_string(syscall_insn),
 327                                     p->p_pid);
 328 #endif /* DEBUG */
 329                 }
 330         }
 331 
 332         mutex_exit(&p->p_ldtlock);       /* Must be held across linear_pc() */
 333 
 334         return (return_code);
 335 }
 336 
 337 /*
 338  * Test to see if the instruction at pc is a system call instruction.
 339  *
 340  * The bytes of an lcall instruction used for the syscall trap.
 341  * static uchar_t lcall[7] = { 0x9a, 0, 0, 0, 0, 0x7, 0 };
 342  * static uchar_t lcallalt[7] = { 0x9a, 0, 0, 0, 0, 0x27, 0 };
 343  */
 344 
 345 #define LCALLSIZE       7
 346 
 347 static int
 348 instr_is_lcall_syscall(caddr_t pc)
 349 {
 350         uchar_t instr[LCALLSIZE];
 351 
 352         if (copyin_nowatch(pc, (caddr_t)instr, LCALLSIZE) == 0 &&
 353             instr[0] == 0x9a &&
 354             instr[1] == 0 &&
 355             instr[2] == 0 &&
 356             instr[3] == 0 &&
 357             instr[4] == 0 &&
 358             (instr[5] == 0x7 || instr[5] == 0x27) &&
 359             instr[6] == 0)
 360                 return (1);
 361 
 362         return (0);
 363 }
 364 
 365 #ifdef __amd64
 366 
 367 /*
 368  * In the first revisions of amd64 CPUs produced by AMD, the LAHF and
 369  * SAHF instructions were not implemented in 64-bit mode. Later revisions
 370  * did implement these instructions. An extension to the cpuid instruction
 371  * was added to check for the capability of executing these instructions
 372  * in 64-bit mode.
 373  *
 374  * Intel originally did not implement these instructions in EM64T either,
 375  * but added them in later revisions.
 376  *
 377  * So, there are different chip revisions by both vendors out there that
 378  * may or may not implement these instructions. The easy solution is to
 379  * just always emulate these instructions on demand.
 380  *
 381  * SAHF == store %ah in the lower 8 bits of %rflags (opcode 0x9e)
 382  * LAHF == load the lower 8 bits of %rflags into %ah (opcode 0x9f)
 383  */
 384 
 385 #define LSAHFSIZE 1
 386 
 387 static int
 388 instr_is_lsahf(caddr_t pc, uchar_t *instr)
 389 {
 390         if (copyin_nowatch(pc, (caddr_t)instr, LSAHFSIZE) == 0 &&
 391             (*instr == 0x9e || *instr == 0x9f))
 392                 return (1);
 393         return (0);
 394 }
 395 
 396 /*
 397  * Emulate the LAHF and SAHF instructions. The reference manuals define
 398  * these instructions to always load/store bit 1 as a 1, and bits 3 and 5
 399  * as a 0. The other, defined, bits are copied (the PS_ICC bits and PS_P).
 400  *
 401  * Note that %ah is bits 8-15 of %rax.
 402  */
 403 static void
 404 emulate_lsahf(struct regs *rp, uchar_t instr)
 405 {
 406         if (instr == 0x9e) {
 407                 /* sahf. Copy bits from %ah to flags. */
 408                 rp->r_ps = (rp->r_ps & ~0xff) |
 409                     ((rp->r_rax >> 8) & PSL_LSAHFMASK) | PS_MB1;
 410         } else {
 411                 /* lahf. Copy bits from flags to %ah. */
 412                 rp->r_rax = (rp->r_rax & ~0xff00) |
 413                     (((rp->r_ps & PSL_LSAHFMASK) | PS_MB1) << 8);
 414         }
 415         rp->r_pc += LSAHFSIZE;
 416 }
 417 #endif /* __amd64 */
 418 
 419 #ifdef OPTERON_ERRATUM_91
 420 
 421 /*
 422  * Test to see if the instruction at pc is a prefetch instruction.
 423  *
 424  * The first byte of prefetch instructions is always 0x0F.
 425  * The second byte is 0x18 for regular prefetch or 0x0D for AMD 3dnow prefetch.
 426  * The third byte (ModRM) contains the register field bits (bits 3-5).
 427  * These bits must be between 0 and 3 inclusive for regular prefetch and
 428  * 0 and 1 inclusive for AMD 3dnow prefetch.
 429  *
 430  * In 64-bit mode, there may be a one-byte REX prefex (0x40-0x4F).
 431  */
 432 
 433 static int
 434 cmp_to_prefetch(uchar_t *p)
 435 {
 436 #ifdef _LP64
 437         if ((p[0] & 0xF0) == 0x40)  /* 64-bit REX prefix */
 438                 p++;
 439 #endif
 440         return ((p[0] == 0x0F && p[1] == 0x18 && ((p[2] >> 3) & 7) <= 3) ||
 441             (p[0] == 0x0F && p[1] == 0x0D && ((p[2] >> 3) & 7) <= 1));
 442 }
 443 
 444 static int
 445 instr_is_prefetch(caddr_t pc)
 446 {
 447         uchar_t instr[4];       /* optional REX prefix plus 3-byte opcode */
 448 
 449         return (copyin_nowatch(pc, instr, sizeof (instr)) == 0 &&
 450             cmp_to_prefetch(instr));
 451 }
 452 
 453 #endif /* OPTERON_ERRATUM_91 */
 454 
 455 /*
 456  * Called from the trap handler when a processor trap occurs.
 457  *
 458  * Note: All user-level traps that might call stop() must exit
 459  * trap() by 'goto out' or by falling through.
 460  * Note Also: trap() is usually called with interrupts enabled, (PS_IE == 1)
 461  * however, there are paths that arrive here with PS_IE == 0 so special care
 462  * must be taken in those cases.
 463  */
 464 void
 465 trap(struct regs *rp, caddr_t addr, processorid_t cpuid)
 466 {
 467         kthread_t *ct = curthread;
 468         enum seg_rw rw;
 469         unsigned type;
 470         proc_t *p = ttoproc(ct);
 471         klwp_t *lwp = ttolwp(ct);
 472         uintptr_t lofault;
 473         label_t *onfault;
 474         faultcode_t pagefault(), res, errcode;
 475         enum fault_type fault_type;
 476         k_siginfo_t siginfo;
 477         uint_t fault = 0;
 478         int mstate;
 479         int sicode = 0;
 480         int watchcode;
 481         int watchpage;
 482         caddr_t vaddr;
 483         int singlestep_twiddle;
 484         size_t sz;
 485         int ta;
 486 #ifdef __amd64
 487         uchar_t instr;
 488 #endif
 489 
 490         ASSERT_STACK_ALIGNED();
 491 
 492         type = rp->r_trapno;
 493         CPU_STATS_ADDQ(CPU, sys, trap, 1);
 494         ASSERT(ct->t_schedflag & TS_DONT_SWAP);
 495 
 496         if (type == T_PGFLT) {
 497 
 498                 errcode = rp->r_err;
 499                 if (errcode & PF_ERR_WRITE)
 500                         rw = S_WRITE;
 501                 else if ((caddr_t)rp->r_pc == addr ||
 502                     (mmu.pt_nx != 0 && (errcode & PF_ERR_EXEC)))
 503                         rw = S_EXEC;
 504                 else
 505                         rw = S_READ;
 506 
 507 #if defined(__i386)
 508                 /*
 509                  * Pentium Pro work-around
 510                  */
 511                 if ((errcode & PF_ERR_PROT) && pentiumpro_bug4046376) {
 512                         uint_t  attr;
 513                         uint_t  priv_violation;
 514                         uint_t  access_violation;
 515 
 516                         if (hat_getattr(addr < (caddr_t)kernelbase ?
 517                             curproc->p_as->a_hat : kas.a_hat, addr, &attr)
 518                             == -1) {
 519                                 errcode &= ~PF_ERR_PROT;
 520                         } else {
 521                                 priv_violation = (errcode & PF_ERR_USER) &&
 522                                     !(attr & PROT_USER);
 523                                 access_violation = (errcode & PF_ERR_WRITE) &&
 524                                     !(attr & PROT_WRITE);
 525                                 if (!priv_violation && !access_violation)
 526                                         goto cleanup;
 527                         }
 528                 }
 529 #endif /* __i386 */
 530 
 531         } else if (type == T_SGLSTP && lwp != NULL)
 532                 lwp->lwp_pcb.pcb_drstat = (uintptr_t)addr;
 533 
 534         if (tdebug)
 535                 showregs(type, rp, addr);
 536 
 537         if (USERMODE(rp->r_cs)) {
 538                 /*
 539                  * Set up the current cred to use during this trap. u_cred
 540                  * no longer exists.  t_cred is used instead.
 541                  * The current process credential applies to the thread for
 542                  * the entire trap.  If trapping from the kernel, this
 543                  * should already be set up.
 544                  */
 545                 if (ct->t_cred != p->p_cred) {
 546                         cred_t *oldcred = ct->t_cred;
 547                         /*
 548                          * DTrace accesses t_cred in probe context.  t_cred
 549                          * must always be either NULL, or point to a valid,
 550                          * allocated cred structure.
 551                          */
 552                         ct->t_cred = crgetcred();
 553                         crfree(oldcred);
 554                 }
 555                 ASSERT(lwp != NULL);
 556                 type |= USER;
 557                 ASSERT(lwptoregs(lwp) == rp);
 558                 lwp->lwp_state = LWP_SYS;
 559 
 560                 switch (type) {
 561                 case T_PGFLT + USER:
 562                         if ((caddr_t)rp->r_pc == addr)
 563                                 mstate = LMS_TFAULT;
 564                         else
 565                                 mstate = LMS_DFAULT;
 566                         break;
 567                 default:
 568                         mstate = LMS_TRAP;
 569                         break;
 570                 }
 571                 /* Kernel probe */
 572                 TNF_PROBE_1(thread_state, "thread", /* CSTYLED */,
 573                     tnf_microstate, state, mstate);
 574                 mstate = new_mstate(ct, mstate);
 575 
 576                 bzero(&siginfo, sizeof (siginfo));
 577         }
 578 
 579         switch (type) {
 580         case T_PGFLT + USER:
 581         case T_SGLSTP:
 582         case T_SGLSTP + USER:
 583         case T_BPTFLT + USER:
 584                 break;
 585 
 586         default:
 587                 FTRACE_2("trap(): type=0x%lx, regs=0x%lx",
 588                     (ulong_t)type, (ulong_t)rp);
 589                 break;
 590         }
 591 
 592         switch (type) {
 593         case T_SIMDFPE:
 594                 /* Make sure we enable interrupts before die()ing */
 595                 sti();  /* The SIMD exception comes in via cmninttrap */
 596                 /*FALLTHROUGH*/
 597         default:
 598                 if (type & USER) {
 599                         if (tudebug)
 600                                 showregs(type, rp, (caddr_t)0);
 601                         printf("trap: Unknown trap type %d in user mode\n",
 602                             type & ~USER);
 603                         siginfo.si_signo = SIGILL;
 604                         siginfo.si_code  = ILL_ILLTRP;
 605                         siginfo.si_addr  = (caddr_t)rp->r_pc;
 606                         siginfo.si_trapno = type & ~USER;
 607                         fault = FLTILL;
 608                         break;
 609                 } else {
 610                         (void) die(type, rp, addr, cpuid);
 611                         /*NOTREACHED*/
 612                 }
 613 
 614         case T_PGFLT:           /* system page fault */
 615                 /*
 616                  * If we're under on_trap() protection (see <sys/ontrap.h>),
 617                  * set ot_trap and bounce back to the on_trap() call site
 618                  * via the installed trampoline.
 619                  */
 620                 if ((ct->t_ontrap != NULL) &&
 621                     (ct->t_ontrap->ot_prot & OT_DATA_ACCESS)) {
 622                         ct->t_ontrap->ot_trap |= OT_DATA_ACCESS;
 623                         rp->r_pc = ct->t_ontrap->ot_trampoline;
 624                         goto cleanup;
 625                 }
 626 
 627                 /*
 628                  * If we have an Instruction fault in kernel mode, then that
 629                  * means we've tried to execute a user page (SMEP) or both of
 630                  * PAE and NXE are enabled. In either case, given that it's a
 631                  * kernel fault, we should panic immediately and not try to make
 632                  * any more forward progress. This indicates a bug in the
 633                  * kernel, which if execution continued, could be exploited to
 634                  * wreak havoc on the system.
 635                  */
 636                 if (errcode & PF_ERR_EXEC) {
 637                         (void) die(type, rp, addr, cpuid);
 638                 }
 639 
 640                 /*
 641                  * See if we can handle as pagefault. Save lofault and onfault
 642                  * across this. Here we assume that an address less than
 643                  * KERNELBASE is a user fault.  We can do this as copy.s
 644                  * routines verify that the starting address is less than
 645                  * KERNELBASE before starting and because we know that we
 646                  * always have KERNELBASE mapped as invalid to serve as a
 647                  * "barrier".
 648                  */
 649                 lofault = ct->t_lofault;
 650                 onfault = ct->t_onfault;
 651                 ct->t_lofault = 0;
 652 
 653                 mstate = new_mstate(ct, LMS_KFAULT);
 654 
 655                 if (addr < (caddr_t)kernelbase) {
 656                         res = pagefault(addr,
 657                             (errcode & PF_ERR_PROT)? F_PROT: F_INVAL, rw, 0);
 658                         if (res == FC_NOMAP &&
 659                             addr < p->p_usrstack &&
 660                             grow(addr))
 661                                 res = 0;
 662                 } else {
 663                         res = pagefault(addr,
 664                             (errcode & PF_ERR_PROT)? F_PROT: F_INVAL, rw, 1);
 665                 }
 666                 (void) new_mstate(ct, mstate);
 667 
 668                 /*
 669                  * Restore lofault and onfault. If we resolved the fault, exit.
 670                  * If we didn't and lofault wasn't set, die.
 671                  */
 672                 ct->t_lofault = lofault;
 673                 ct->t_onfault = onfault;
 674                 if (res == 0)
 675                         goto cleanup;
 676 
 677 #if defined(OPTERON_ERRATUM_93) && defined(_LP64)
 678                 if (lofault == 0 && opteron_erratum_93) {
 679                         /*
 680                          * Workaround for Opteron Erratum 93. On return from
 681                          * a System Managment Interrupt at a HLT instruction
 682                          * the %rip might be truncated to a 32 bit value.
 683                          * BIOS is supposed to fix this, but some don't.
 684                          * If this occurs we simply restore the high order bits.
 685                          * The HLT instruction is 1 byte of 0xf4.
 686                          */
 687                         uintptr_t       rip = rp->r_pc;
 688 
 689                         if ((rip & 0xfffffffful) == rip) {
 690                                 rip |= 0xfffffffful << 32;
 691                                 if (hat_getpfnum(kas.a_hat, (caddr_t)rip) !=
 692                                     PFN_INVALID &&
 693                                     (*(uchar_t *)rip == 0xf4 ||
 694                                     *(uchar_t *)(rip - 1) == 0xf4)) {
 695                                         rp->r_pc = rip;
 696                                         goto cleanup;
 697                                 }
 698                         }
 699                 }
 700 #endif /* OPTERON_ERRATUM_93 && _LP64 */
 701 
 702 #ifdef OPTERON_ERRATUM_91
 703                 if (lofault == 0 && opteron_erratum_91) {
 704                         /*
 705                          * Workaround for Opteron Erratum 91. Prefetches may
 706                          * generate a page fault (they're not supposed to do
 707                          * that!). If this occurs we simply return back to the
 708                          * instruction.
 709                          */
 710                         caddr_t         pc = (caddr_t)rp->r_pc;
 711 
 712                         /*
 713                          * If the faulting PC is not mapped, this is a
 714                          * legitimate kernel page fault that must result in a
 715                          * panic. If the faulting PC is mapped, it could contain
 716                          * a prefetch instruction. Check for that here.
 717                          */
 718                         if (hat_getpfnum(kas.a_hat, pc) != PFN_INVALID) {
 719                                 if (cmp_to_prefetch((uchar_t *)pc)) {
 720 #ifdef DEBUG
 721                                         cmn_err(CE_WARN, "Opteron erratum 91 "
 722                                             "occurred: kernel prefetch"
 723                                             " at %p generated a page fault!",
 724                                             (void *)rp->r_pc);
 725 #endif /* DEBUG */
 726                                         goto cleanup;
 727                                 }
 728                         }
 729                         (void) die(type, rp, addr, cpuid);
 730                 }
 731 #endif /* OPTERON_ERRATUM_91 */
 732 
 733                 if (lofault == 0)
 734                         (void) die(type, rp, addr, cpuid);
 735 
 736                 /*
 737                  * Cannot resolve fault.  Return to lofault.
 738                  */
 739                 if (lodebug) {
 740                         showregs(type, rp, addr);
 741                         traceregs(rp);
 742                 }
 743                 if (FC_CODE(res) == FC_OBJERR)
 744                         res = FC_ERRNO(res);
 745                 else
 746                         res = EFAULT;
 747                 rp->r_r0 = res;
 748                 rp->r_pc = ct->t_lofault;
 749                 goto cleanup;
 750 
 751         case T_PGFLT + USER:    /* user page fault */
 752                 if (faultdebug) {
 753                         char *fault_str;
 754 
 755                         switch (rw) {
 756                         case S_READ:
 757                                 fault_str = "read";
 758                                 break;
 759                         case S_WRITE:
 760                                 fault_str = "write";
 761                                 break;
 762                         case S_EXEC:
 763                                 fault_str = "exec";
 764                                 break;
 765                         default:
 766                                 fault_str = "";
 767                                 break;
 768                         }
 769                         printf("user %s fault:  addr=0x%lx errcode=0x%x\n",
 770                             fault_str, (uintptr_t)addr, errcode);
 771                 }
 772 
 773 #if defined(OPTERON_ERRATUM_100) && defined(_LP64)
 774                 /*
 775                  * Workaround for AMD erratum 100
 776                  *
 777                  * A 32-bit process may receive a page fault on a non
 778                  * 32-bit address by mistake. The range of the faulting
 779                  * address will be
 780                  *
 781                  *      0xffffffff80000000 .. 0xffffffffffffffff or
 782                  *      0x0000000100000000 .. 0x000000017fffffff
 783                  *
 784                  * The fault is always due to an instruction fetch, however
 785                  * the value of r_pc should be correct (in 32 bit range),
 786                  * so we ignore the page fault on the bogus address.
 787                  */
 788                 if (p->p_model == DATAMODEL_ILP32 &&
 789                     (0xffffffff80000000 <= (uintptr_t)addr ||
 790                     (0x100000000 <= (uintptr_t)addr &&
 791                     (uintptr_t)addr <= 0x17fffffff))) {
 792                         if (!opteron_erratum_100)
 793                                 panic("unexpected erratum #100");
 794                         if (rp->r_pc <= 0xffffffff)
 795                                 goto out;
 796                 }
 797 #endif /* OPTERON_ERRATUM_100 && _LP64 */
 798 
 799                 ASSERT(!(curthread->t_flag & T_WATCHPT));
 800                 watchpage = (pr_watch_active(p) && pr_is_watchpage(addr, rw));
 801 #ifdef __i386
 802                 /*
 803                  * In 32-bit mode, the lcall (system call) instruction fetches
 804                  * one word from the stack, at the stack pointer, because of the
 805                  * way the call gate is constructed.  This is a bogus
 806                  * read and should not be counted as a read watchpoint.
 807                  * We work around the problem here by testing to see if
 808                  * this situation applies and, if so, simply jumping to
 809                  * the code in locore.s that fields the system call trap.
 810                  * The registers on the stack are already set up properly
 811                  * due to the match between the call gate sequence and the
 812                  * trap gate sequence.  We just have to adjust the pc.
 813                  */
 814                 if (watchpage && addr == (caddr_t)rp->r_sp &&
 815                     rw == S_READ && instr_is_lcall_syscall((caddr_t)rp->r_pc)) {
 816                         extern void watch_syscall(void);
 817 
 818                         rp->r_pc += LCALLSIZE;
 819                         watch_syscall();        /* never returns */
 820                         /* NOTREACHED */
 821                 }
 822 #endif /* __i386 */
 823                 vaddr = addr;
 824                 if (!watchpage || (sz = instr_size(rp, &vaddr, rw)) <= 0)
 825                         fault_type = (errcode & PF_ERR_PROT)? F_PROT: F_INVAL;
 826                 else if ((watchcode = pr_is_watchpoint(&vaddr, &ta,
 827                     sz, NULL, rw)) != 0) {
 828                         if (ta) {
 829                                 do_watch_step(vaddr, sz, rw,
 830                                     watchcode, rp->r_pc);
 831                                 fault_type = F_INVAL;
 832                         } else {
 833                                 bzero(&siginfo, sizeof (siginfo));
 834                                 siginfo.si_signo = SIGTRAP;
 835                                 siginfo.si_code = watchcode;
 836                                 siginfo.si_addr = vaddr;
 837                                 siginfo.si_trapafter = 0;
 838                                 siginfo.si_pc = (caddr_t)rp->r_pc;
 839                                 fault = FLTWATCH;
 840                                 break;
 841                         }
 842                 } else {
 843                         /* XXX pr_watch_emul() never succeeds (for now) */
 844                         if (rw != S_EXEC && pr_watch_emul(rp, vaddr, rw))
 845                                 goto out;
 846                         do_watch_step(vaddr, sz, rw, 0, 0);
 847                         fault_type = F_INVAL;
 848                 }
 849 
 850                 res = pagefault(addr, fault_type, rw, 0);
 851 
 852                 /*
 853                  * If pagefault() succeeded, ok.
 854                  * Otherwise attempt to grow the stack.
 855                  */
 856                 if (res == 0 ||
 857                     (res == FC_NOMAP &&
 858                     addr < p->p_usrstack &&
 859                     grow(addr))) {
 860                         lwp->lwp_lastfault = FLTPAGE;
 861                         lwp->lwp_lastfaddr = addr;
 862                         if (prismember(&p->p_fltmask, FLTPAGE)) {
 863                                 bzero(&siginfo, sizeof (siginfo));
 864                                 siginfo.si_addr = addr;
 865                                 (void) stop_on_fault(FLTPAGE, &siginfo);
 866                         }
 867                         goto out;
 868                 } else if (res == FC_PROT && addr < p->p_usrstack &&
 869                     (mmu.pt_nx != 0 && (errcode & PF_ERR_EXEC))) {
 870                         report_stack_exec(p, addr);
 871                 }
 872 
 873 #ifdef OPTERON_ERRATUM_91
 874                 /*
 875                  * Workaround for Opteron Erratum 91. Prefetches may generate a
 876                  * page fault (they're not supposed to do that!). If this
 877                  * occurs we simply return back to the instruction.
 878                  *
 879                  * We rely on copyin to properly fault in the page with r_pc.
 880                  */
 881                 if (opteron_erratum_91 &&
 882                     addr != (caddr_t)rp->r_pc &&
 883                     instr_is_prefetch((caddr_t)rp->r_pc)) {
 884 #ifdef DEBUG
 885                         cmn_err(CE_WARN, "Opteron erratum 91 occurred: "
 886                             "prefetch at %p in pid %d generated a trap!",
 887                             (void *)rp->r_pc, p->p_pid);
 888 #endif /* DEBUG */
 889                         goto out;
 890                 }
 891 #endif /* OPTERON_ERRATUM_91 */
 892 
 893                 if (tudebug)
 894                         showregs(type, rp, addr);
 895                 /*
 896                  * In the case where both pagefault and grow fail,
 897                  * set the code to the value provided by pagefault.
 898                  * We map all errors returned from pagefault() to SIGSEGV.
 899                  */
 900                 bzero(&siginfo, sizeof (siginfo));
 901                 siginfo.si_addr = addr;
 902                 switch (FC_CODE(res)) {
 903                 case FC_HWERR:
 904                 case FC_NOSUPPORT:
 905                         siginfo.si_signo = SIGBUS;
 906                         siginfo.si_code = BUS_ADRERR;
 907                         fault = FLTACCESS;
 908                         break;
 909                 case FC_ALIGN:
 910                         siginfo.si_signo = SIGBUS;
 911                         siginfo.si_code = BUS_ADRALN;
 912                         fault = FLTACCESS;
 913                         break;
 914                 case FC_OBJERR:
 915                         if ((siginfo.si_errno = FC_ERRNO(res)) != EINTR) {
 916                                 siginfo.si_signo = SIGBUS;
 917                                 siginfo.si_code = BUS_OBJERR;
 918                                 fault = FLTACCESS;
 919                         }
 920                         break;
 921                 default:        /* FC_NOMAP or FC_PROT */
 922                         siginfo.si_signo = SIGSEGV;
 923                         siginfo.si_code =
 924                             (res == FC_NOMAP)? SEGV_MAPERR : SEGV_ACCERR;
 925                         fault = FLTBOUNDS;
 926                         break;
 927                 }
 928                 break;
 929 
 930         case T_ILLINST + USER:  /* invalid opcode fault */
 931                 /*
 932                  * If the syscall instruction is disabled due to LDT usage, a
 933                  * user program that attempts to execute it will trigger a #ud
 934                  * trap. Check for that case here. If this occurs on a CPU which
 935                  * doesn't even support syscall, the result of all of this will
 936                  * be to emulate that particular instruction.
 937                  */
 938                 if (p->p_ldt != NULL &&
 939                     ldt_rewrite_syscall(rp, p, X86FSET_ASYSC))
 940                         goto out;
 941 
 942 #ifdef __amd64
 943                 /*
 944                  * Emulate the LAHF and SAHF instructions if needed.
 945                  * See the instr_is_lsahf function for details.
 946                  */
 947                 if (p->p_model == DATAMODEL_LP64 &&
 948                     instr_is_lsahf((caddr_t)rp->r_pc, &instr)) {
 949                         emulate_lsahf(rp, instr);
 950                         goto out;
 951                 }
 952 #endif
 953 
 954                 /*FALLTHROUGH*/
 955 
 956                 if (tudebug)
 957                         showregs(type, rp, (caddr_t)0);
 958                 siginfo.si_signo = SIGILL;
 959                 siginfo.si_code  = ILL_ILLOPC;
 960                 siginfo.si_addr  = (caddr_t)rp->r_pc;
 961                 fault = FLTILL;
 962                 break;
 963 
 964         case T_ZERODIV + USER:          /* integer divide by zero */
 965                 if (tudebug && tudebugfpe)
 966                         showregs(type, rp, (caddr_t)0);
 967                 siginfo.si_signo = SIGFPE;
 968                 siginfo.si_code  = FPE_INTDIV;
 969                 siginfo.si_addr  = (caddr_t)rp->r_pc;
 970                 fault = FLTIZDIV;
 971                 break;
 972 
 973         case T_OVFLW + USER:    /* integer overflow */
 974                 if (tudebug && tudebugfpe)
 975                         showregs(type, rp, (caddr_t)0);
 976                 siginfo.si_signo = SIGFPE;
 977                 siginfo.si_code  = FPE_INTOVF;
 978                 siginfo.si_addr  = (caddr_t)rp->r_pc;
 979                 fault = FLTIOVF;
 980                 break;
 981 
 982         case T_NOEXTFLT + USER: /* math coprocessor not available */
 983                 if (tudebug && tudebugfpe)
 984                         showregs(type, rp, addr);
 985                 if (fpnoextflt(rp)) {
 986                         siginfo.si_signo = SIGILL;
 987                         siginfo.si_code  = ILL_ILLOPC;
 988                         siginfo.si_addr  = (caddr_t)rp->r_pc;
 989                         fault = FLTILL;
 990                 }
 991                 break;
 992 
 993         case T_EXTOVRFLT:       /* extension overrun fault */
 994                 /* check if we took a kernel trap on behalf of user */
 995                 {
 996                         extern  void ndptrap_frstor(void);
 997                         if (rp->r_pc != (uintptr_t)ndptrap_frstor) {
 998                                 sti(); /* T_EXTOVRFLT comes in via cmninttrap */
 999                                 (void) die(type, rp, addr, cpuid);
1000                         }
1001                         type |= USER;
1002                 }
1003                 /*FALLTHROUGH*/
1004         case T_EXTOVRFLT + USER:        /* extension overrun fault */
1005                 if (tudebug && tudebugfpe)
1006                         showregs(type, rp, addr);
1007                 if (fpextovrflt(rp)) {
1008                         siginfo.si_signo = SIGSEGV;
1009                         siginfo.si_code  = SEGV_MAPERR;
1010                         siginfo.si_addr  = (caddr_t)rp->r_pc;
1011                         fault = FLTBOUNDS;
1012                 }
1013                 break;
1014 
1015         case T_EXTERRFLT:       /* x87 floating point exception pending */
1016                 /* check if we took a kernel trap on behalf of user */
1017                 {
1018                         extern  void ndptrap_frstor(void);
1019                         if (rp->r_pc != (uintptr_t)ndptrap_frstor) {
1020                                 sti(); /* T_EXTERRFLT comes in via cmninttrap */
1021                                 (void) die(type, rp, addr, cpuid);
1022                         }
1023                         type |= USER;
1024                 }
1025                 /*FALLTHROUGH*/
1026 
1027         case T_EXTERRFLT + USER: /* x87 floating point exception pending */
1028                 if (tudebug && tudebugfpe)
1029                         showregs(type, rp, addr);
1030                 if (sicode = fpexterrflt(rp)) {
1031                         siginfo.si_signo = SIGFPE;
1032                         siginfo.si_code  = sicode;
1033                         siginfo.si_addr  = (caddr_t)rp->r_pc;
1034                         fault = FLTFPE;
1035                 }
1036                 break;
1037 
1038         case T_SIMDFPE + USER:          /* SSE and SSE2 exceptions */
1039                 if (tudebug && tudebugsse)
1040                         showregs(type, rp, addr);
1041                 if (!is_x86_feature(x86_featureset, X86FSET_SSE) &&
1042                     !is_x86_feature(x86_featureset, X86FSET_SSE2)) {
1043                         /*
1044                          * There are rumours that some user instructions
1045                          * on older CPUs can cause this trap to occur; in
1046                          * which case send a SIGILL instead of a SIGFPE.
1047                          */
1048                         siginfo.si_signo = SIGILL;
1049                         siginfo.si_code  = ILL_ILLTRP;
1050                         siginfo.si_addr  = (caddr_t)rp->r_pc;
1051                         siginfo.si_trapno = type & ~USER;
1052                         fault = FLTILL;
1053                 } else if ((sicode = fpsimderrflt(rp)) != 0) {
1054                         siginfo.si_signo = SIGFPE;
1055                         siginfo.si_code = sicode;
1056                         siginfo.si_addr = (caddr_t)rp->r_pc;
1057                         fault = FLTFPE;
1058                 }
1059 
1060                 sti();  /* The SIMD exception comes in via cmninttrap */
1061                 break;
1062 
1063         case T_BPTFLT:  /* breakpoint trap */
1064                 /*
1065                  * Kernel breakpoint traps should only happen when kmdb is
1066                  * active, and even then, it'll have interposed on the IDT, so
1067                  * control won't get here.  If it does, we've hit a breakpoint
1068                  * without the debugger, which is very strange, and very
1069                  * fatal.
1070                  */
1071                 if (tudebug && tudebugbpt)
1072                         showregs(type, rp, (caddr_t)0);
1073 
1074                 (void) die(type, rp, addr, cpuid);
1075                 break;
1076 
1077         case T_SGLSTP: /* single step/hw breakpoint exception */
1078 
1079                 /* Now evaluate how we got here */
1080                 if (lwp != NULL && (lwp->lwp_pcb.pcb_drstat & DR_SINGLESTEP)) {
1081                         /*
1082                          * i386 single-steps even through lcalls which
1083                          * change the privilege level. So we take a trap at
1084                          * the first instruction in privileged mode.
1085                          *
1086                          * Set a flag to indicate that upon completion of
1087                          * the system call, deal with the single-step trap.
1088                          *
1089                          * The same thing happens for sysenter, too.
1090                          */
1091                         singlestep_twiddle = 0;
1092                         if (rp->r_pc == (uintptr_t)sys_sysenter ||
1093                             rp->r_pc == (uintptr_t)brand_sys_sysenter) {
1094                                 singlestep_twiddle = 1;
1095 #if defined(__amd64)
1096                                 /*
1097                                  * Since we are already on the kernel's
1098                                  * %gs, on 64-bit systems the sysenter case
1099                                  * needs to adjust the pc to avoid
1100                                  * executing the swapgs instruction at the
1101                                  * top of the handler.
1102                                  */
1103                                 if (rp->r_pc == (uintptr_t)sys_sysenter)
1104                                         rp->r_pc = (uintptr_t)
1105                                             _sys_sysenter_post_swapgs;
1106                                 else
1107                                         rp->r_pc = (uintptr_t)
1108                                             _brand_sys_sysenter_post_swapgs;
1109 #endif
1110                         }
1111 #if defined(__i386)
1112                         else if (rp->r_pc == (uintptr_t)sys_call ||
1113                             rp->r_pc == (uintptr_t)brand_sys_call) {
1114                                 singlestep_twiddle = 1;
1115                         }
1116 #endif
1117                         else {
1118                                 /* not on sysenter/syscall; uregs available */
1119                                 if (tudebug && tudebugbpt)
1120                                         showregs(type, rp, (caddr_t)0);
1121                         }
1122                         if (singlestep_twiddle) {
1123                                 rp->r_ps &= ~PS_T; /* turn off trace */
1124                                 lwp->lwp_pcb.pcb_flags |= DEBUG_PENDING;
1125                                 ct->t_post_sys = 1;
1126                                 aston(curthread);
1127                                 goto cleanup;
1128                         }
1129                 }
1130                 /* XXX - needs review on debugger interface? */
1131                 if (boothowto & RB_DEBUG)
1132                         debug_enter((char *)NULL);
1133                 else
1134                         (void) die(type, rp, addr, cpuid);
1135                 break;
1136 
1137         case T_NMIFLT:  /* NMI interrupt */
1138                 printf("Unexpected NMI in system mode\n");
1139                 goto cleanup;
1140 
1141         case T_NMIFLT + USER:   /* NMI interrupt */
1142                 printf("Unexpected NMI in user mode\n");
1143                 break;
1144 
1145         case T_GPFLT:   /* general protection violation */
1146                 /*
1147                  * Any #GP that occurs during an on_trap .. no_trap bracket
1148                  * with OT_DATA_ACCESS or OT_SEGMENT_ACCESS protection,
1149                  * or in a on_fault .. no_fault bracket, is forgiven
1150                  * and we trampoline.  This protection is given regardless
1151                  * of whether we are 32/64 bit etc - if a distinction is
1152                  * required then define new on_trap protection types.
1153                  *
1154                  * On amd64, we can get a #gp from referencing addresses
1155                  * in the virtual address hole e.g. from a copyin or in
1156                  * update_sregs while updating user segment registers.
1157                  *
1158                  * On the 32-bit hypervisor we could also generate one in
1159                  * mfn_to_pfn by reaching around or into where the hypervisor
1160                  * lives which is protected by segmentation.
1161                  */
1162 
1163                 /*
1164                  * If we're under on_trap() protection (see <sys/ontrap.h>),
1165                  * set ot_trap and trampoline back to the on_trap() call site
1166                  * for OT_DATA_ACCESS or OT_SEGMENT_ACCESS.
1167                  */
1168                 if (ct->t_ontrap != NULL) {
1169                         int ttype =  ct->t_ontrap->ot_prot &
1170                             (OT_DATA_ACCESS | OT_SEGMENT_ACCESS);
1171 
1172                         if (ttype != 0) {
1173                                 ct->t_ontrap->ot_trap |= ttype;
1174                                 if (tudebug)
1175                                         showregs(type, rp, (caddr_t)0);
1176                                 rp->r_pc = ct->t_ontrap->ot_trampoline;
1177                                 goto cleanup;
1178                         }
1179                 }
1180 
1181                 /*
1182                  * If we're under lofault protection (copyin etc.),
1183                  * longjmp back to lofault with an EFAULT.
1184                  */
1185                 if (ct->t_lofault) {
1186                         /*
1187                          * Fault is not resolvable, so just return to lofault
1188                          */
1189                         if (lodebug) {
1190                                 showregs(type, rp, addr);
1191                                 traceregs(rp);
1192                         }
1193                         rp->r_r0 = EFAULT;
1194                         rp->r_pc = ct->t_lofault;
1195                         goto cleanup;
1196                 }
1197 
1198                 /*
1199                  * We fall through to the next case, which repeats
1200                  * the OT_SEGMENT_ACCESS check which we've already
1201                  * done, so we'll always fall through to the
1202                  * T_STKFLT case.
1203                  */
1204                 /*FALLTHROUGH*/
1205         case T_SEGFLT:  /* segment not present fault */
1206                 /*
1207                  * One example of this is #NP in update_sregs while
1208                  * attempting to update a user segment register
1209                  * that points to a descriptor that is marked not
1210                  * present.
1211                  */
1212                 if (ct->t_ontrap != NULL &&
1213                     ct->t_ontrap->ot_prot & OT_SEGMENT_ACCESS) {
1214                         ct->t_ontrap->ot_trap |= OT_SEGMENT_ACCESS;
1215                         if (tudebug)
1216                                 showregs(type, rp, (caddr_t)0);
1217                         rp->r_pc = ct->t_ontrap->ot_trampoline;
1218                         goto cleanup;
1219                 }
1220                 /*FALLTHROUGH*/
1221         case T_STKFLT:  /* stack fault */
1222         case T_TSSFLT:  /* invalid TSS fault */
1223                 if (tudebug)
1224                         showregs(type, rp, (caddr_t)0);
1225                 if (kern_gpfault(rp))
1226                         (void) die(type, rp, addr, cpuid);
1227                 goto cleanup;
1228 
1229         /*
1230          * ONLY 32-bit PROCESSES can USE a PRIVATE LDT! 64-bit apps
1231          * should have no need for them, so we put a stop to it here.
1232          *
1233          * So: not-present fault is ONLY valid for 32-bit processes with
1234          * a private LDT trying to do a system call. Emulate it.
1235          *
1236          * #gp fault is ONLY valid for 32-bit processes also, which DO NOT
1237          * have a private LDT, and are trying to do a system call. Emulate it.
1238          */
1239 
1240         case T_SEGFLT + USER:   /* segment not present fault */
1241         case T_GPFLT + USER:    /* general protection violation */
1242 #ifdef _SYSCALL32_IMPL
1243                 if (p->p_model != DATAMODEL_NATIVE) {
1244 #endif /* _SYSCALL32_IMPL */
1245                 if (instr_is_lcall_syscall((caddr_t)rp->r_pc)) {
1246                         if (type == T_SEGFLT + USER)
1247                                 ASSERT(p->p_ldt != NULL);
1248 
1249                         if ((p->p_ldt == NULL && type == T_GPFLT + USER) ||
1250                             type == T_SEGFLT + USER) {
1251 
1252                         /*
1253                          * The user attempted a system call via the obsolete
1254                          * call gate mechanism. Because the process doesn't have
1255                          * an LDT (i.e. the ldtr contains 0), a #gp results.
1256                          * Emulate the syscall here, just as we do above for a
1257                          * #np trap.
1258                          */
1259 
1260                         /*
1261                          * Since this is a not-present trap, rp->r_pc points to
1262                          * the trapping lcall instruction. We need to bump it
1263                          * to the next insn so the app can continue on.
1264                          */
1265                         rp->r_pc += LCALLSIZE;
1266                         lwp->lwp_regs = rp;
1267 
1268                         /*
1269                          * Normally the microstate of the LWP is forced back to
1270                          * LMS_USER by the syscall handlers. Emulate that
1271                          * behavior here.
1272                          */
1273                         mstate = LMS_USER;
1274 
1275                         dosyscall();
1276                         goto out;
1277                         }
1278                 }
1279 #ifdef _SYSCALL32_IMPL
1280                 }
1281 #endif /* _SYSCALL32_IMPL */
1282                 /*
1283                  * If the current process is using a private LDT and the
1284                  * trapping instruction is sysenter, the sysenter instruction
1285                  * has been disabled on the CPU because it destroys segment
1286                  * registers. If this is the case, rewrite the instruction to
1287                  * be a safe system call and retry it. If this occurs on a CPU
1288                  * which doesn't even support sysenter, the result of all of
1289                  * this will be to emulate that particular instruction.
1290                  */
1291                 if (p->p_ldt != NULL &&
1292                     ldt_rewrite_syscall(rp, p, X86FSET_SEP))
1293                         goto out;
1294 
1295                 /*FALLTHROUGH*/
1296 
1297         case T_BOUNDFLT + USER: /* bound fault */
1298         case T_STKFLT + USER:   /* stack fault */
1299         case T_TSSFLT + USER:   /* invalid TSS fault */
1300                 if (tudebug)
1301                         showregs(type, rp, (caddr_t)0);
1302                 siginfo.si_signo = SIGSEGV;
1303                 siginfo.si_code  = SEGV_MAPERR;
1304                 siginfo.si_addr  = (caddr_t)rp->r_pc;
1305                 fault = FLTBOUNDS;
1306                 break;
1307 
1308         case T_ALIGNMENT + USER:        /* user alignment error (486) */
1309                 if (tudebug)
1310                         showregs(type, rp, (caddr_t)0);
1311                 bzero(&siginfo, sizeof (siginfo));
1312                 siginfo.si_signo = SIGBUS;
1313                 siginfo.si_code = BUS_ADRALN;
1314                 siginfo.si_addr = (caddr_t)rp->r_pc;
1315                 fault = FLTACCESS;
1316                 break;
1317 
1318         case T_SGLSTP + USER: /* single step/hw breakpoint exception */
1319                 if (tudebug && tudebugbpt)
1320                         showregs(type, rp, (caddr_t)0);
1321 
1322                 /* Was it single-stepping? */
1323                 if (lwp->lwp_pcb.pcb_drstat & DR_SINGLESTEP) {
1324                         pcb_t *pcb = &lwp->lwp_pcb;
1325 
1326                         rp->r_ps &= ~PS_T;
1327                         /*
1328                          * If both NORMAL_STEP and WATCH_STEP are in effect,
1329                          * give precedence to WATCH_STEP.  If neither is set,
1330                          * user must have set the PS_T bit in %efl; treat this
1331                          * as NORMAL_STEP.
1332                          */
1333                         if ((fault = undo_watch_step(&siginfo)) == 0 &&
1334                             ((pcb->pcb_flags & NORMAL_STEP) ||
1335                             !(pcb->pcb_flags & WATCH_STEP))) {
1336                                 siginfo.si_signo = SIGTRAP;
1337                                 siginfo.si_code = TRAP_TRACE;
1338                                 siginfo.si_addr = (caddr_t)rp->r_pc;
1339                                 fault = FLTTRACE;
1340                         }
1341                         pcb->pcb_flags &= ~(NORMAL_STEP|WATCH_STEP);
1342                 }
1343                 break;
1344 
1345         case T_BPTFLT + USER:   /* breakpoint trap */
1346                 if (tudebug && tudebugbpt)
1347                         showregs(type, rp, (caddr_t)0);
1348                 /*
1349                  * int 3 (the breakpoint instruction) leaves the pc referring
1350                  * to the address one byte after the breakpointed address.
1351                  * If the P_PR_BPTADJ flag has been set via /proc, We adjust
1352                  * it back so it refers to the breakpointed address.
1353                  */
1354                 if (p->p_proc_flag & P_PR_BPTADJ)
1355                         rp->r_pc--;
1356                 siginfo.si_signo = SIGTRAP;
1357                 siginfo.si_code  = TRAP_BRKPT;
1358                 siginfo.si_addr  = (caddr_t)rp->r_pc;
1359                 fault = FLTBPT;
1360                 break;
1361 
1362         case T_AST:
1363                 /*
1364                  * This occurs only after the cs register has been made to
1365                  * look like a kernel selector, either through debugging or
1366                  * possibly by functions like setcontext().  The thread is
1367                  * about to cause a general protection fault at common_iret()
1368                  * in locore.  We let that happen immediately instead of
1369                  * doing the T_AST processing.
1370                  */
1371                 goto cleanup;
1372 
1373         case T_AST + USER:      /* profiling, resched, h/w error pseudo trap */
1374                 if (lwp->lwp_pcb.pcb_flags & ASYNC_HWERR) {
1375                         proc_t *p = ttoproc(curthread);
1376                         extern void print_msg_hwerr(ctid_t ct_id, proc_t *p);
1377 
1378                         lwp->lwp_pcb.pcb_flags &= ~ASYNC_HWERR;
1379                         print_msg_hwerr(p->p_ct_process->conp_contract.ct_id,
1380                             p);
1381                         contract_process_hwerr(p->p_ct_process, p);
1382                         siginfo.si_signo = SIGKILL;
1383                         siginfo.si_code = SI_NOINFO;
1384                 } else if (lwp->lwp_pcb.pcb_flags & CPC_OVERFLOW) {
1385                         lwp->lwp_pcb.pcb_flags &= ~CPC_OVERFLOW;
1386                         if (kcpc_overflow_ast()) {
1387                                 /*
1388                                  * Signal performance counter overflow
1389                                  */
1390                                 if (tudebug)
1391                                         showregs(type, rp, (caddr_t)0);
1392                                 bzero(&siginfo, sizeof (siginfo));
1393                                 siginfo.si_signo = SIGEMT;
1394                                 siginfo.si_code = EMT_CPCOVF;
1395                                 siginfo.si_addr = (caddr_t)rp->r_pc;
1396                                 fault = FLTCPCOVF;
1397                         }
1398                 }
1399 
1400                 break;
1401         }
1402 
1403         /*
1404          * We can't get here from a system trap
1405          */
1406         ASSERT(type & USER);
1407 
1408         if (fault) {
1409                 /* We took a fault so abort single step. */
1410                 lwp->lwp_pcb.pcb_flags &= ~(NORMAL_STEP|WATCH_STEP);
1411                 /*
1412                  * Remember the fault and fault adddress
1413                  * for real-time (SIGPROF) profiling.
1414                  */
1415                 lwp->lwp_lastfault = fault;
1416                 lwp->lwp_lastfaddr = siginfo.si_addr;
1417 
1418                 DTRACE_PROC2(fault, int, fault, ksiginfo_t *, &siginfo);
1419 
1420                 /*
1421                  * If a debugger has declared this fault to be an
1422                  * event of interest, stop the lwp.  Otherwise just
1423                  * deliver the associated signal.
1424                  */
1425                 if (siginfo.si_signo != SIGKILL &&
1426                     prismember(&p->p_fltmask, fault) &&
1427                     stop_on_fault(fault, &siginfo) == 0)
1428                         siginfo.si_signo = 0;
1429         }
1430 
1431         if (siginfo.si_signo)
1432                 trapsig(&siginfo, (fault != FLTFPE && fault != FLTCPCOVF));
1433 
1434         if (lwp->lwp_oweupc)
1435                 profil_tick(rp->r_pc);
1436 
1437         if (ct->t_astflag | ct->t_sig_check) {
1438                 /*
1439                  * Turn off the AST flag before checking all the conditions that
1440                  * may have caused an AST.  This flag is on whenever a signal or
1441                  * unusual condition should be handled after the next trap or
1442                  * syscall.
1443                  */
1444                 astoff(ct);
1445                 /*
1446                  * If a single-step trap occurred on a syscall (see above)
1447                  * recognize it now.  Do this before checking for signals
1448                  * because deferred_singlestep_trap() may generate a SIGTRAP to
1449                  * the LWP or may otherwise mark the LWP to call issig(FORREAL).
1450                  */
1451                 if (lwp->lwp_pcb.pcb_flags & DEBUG_PENDING)
1452                         deferred_singlestep_trap((caddr_t)rp->r_pc);
1453 
1454                 ct->t_sig_check = 0;
1455 
1456                 mutex_enter(&p->p_lock);
1457                 if (curthread->t_proc_flag & TP_CHANGEBIND) {
1458                         timer_lwpbind();
1459                         curthread->t_proc_flag &= ~TP_CHANGEBIND;
1460                 }
1461                 mutex_exit(&p->p_lock);
1462 
1463                 /*
1464                  * for kaio requests that are on the per-process poll queue,
1465                  * aiop->aio_pollq, they're AIO_POLL bit is set, the kernel
1466                  * should copyout their result_t to user memory. by copying
1467                  * out the result_t, the user can poll on memory waiting
1468                  * for the kaio request to complete.
1469                  */
1470                 if (p->p_aio)
1471                         aio_cleanup(0);
1472                 /*
1473                  * If this LWP was asked to hold, call holdlwp(), which will
1474                  * stop.  holdlwps() sets this up and calls pokelwps() which
1475                  * sets the AST flag.
1476                  *
1477                  * Also check TP_EXITLWP, since this is used by fresh new LWPs
1478                  * through lwp_rtt().  That flag is set if the lwp_create(2)
1479                  * syscall failed after creating the LWP.
1480                  */
1481                 if (ISHOLD(p))
1482                         holdlwp();
1483 
1484                 /*
1485                  * All code that sets signals and makes ISSIG evaluate true must
1486                  * set t_astflag afterwards.
1487                  */
1488                 if (ISSIG_PENDING(ct, lwp, p)) {
1489                         if (issig(FORREAL))
1490                                 psig();
1491                         ct->t_sig_check = 1;
1492                 }
1493 
1494                 if (ct->t_rprof != NULL) {
1495                         realsigprof(0, 0, 0);
1496                         ct->t_sig_check = 1;
1497                 }
1498 
1499                 /*
1500                  * /proc can't enable/disable the trace bit itself
1501                  * because that could race with the call gate used by
1502                  * system calls via "lcall". If that happened, an
1503                  * invalid EFLAGS would result. prstep()/prnostep()
1504                  * therefore schedule an AST for the purpose.
1505                  */
1506                 if (lwp->lwp_pcb.pcb_flags & REQUEST_STEP) {
1507                         lwp->lwp_pcb.pcb_flags &= ~REQUEST_STEP;
1508                         rp->r_ps |= PS_T;
1509                 }
1510                 if (lwp->lwp_pcb.pcb_flags & REQUEST_NOSTEP) {
1511                         lwp->lwp_pcb.pcb_flags &= ~REQUEST_NOSTEP;
1512                         rp->r_ps &= ~PS_T;
1513                 }
1514         }
1515 
1516 out:    /* We can't get here from a system trap */
1517         ASSERT(type & USER);
1518 
1519         if (ISHOLD(p))
1520                 holdlwp();
1521 
1522         /*
1523          * Set state to LWP_USER here so preempt won't give us a kernel
1524          * priority if it occurs after this point.  Call CL_TRAPRET() to
1525          * restore the user-level priority.
1526          *
1527          * It is important that no locks (other than spinlocks) be entered
1528          * after this point before returning to user mode (unless lwp_state
1529          * is set back to LWP_SYS).
1530          */
1531         lwp->lwp_state = LWP_USER;
1532 
1533         if (ct->t_trapret) {
1534                 ct->t_trapret = 0;
1535                 thread_lock(ct);
1536                 CL_TRAPRET(ct);
1537                 thread_unlock(ct);
1538         }
1539         if (CPU->cpu_runrun || curthread->t_schedflag & TS_ANYWAITQ)
1540                 preempt();
1541         prunstop();
1542         (void) new_mstate(ct, mstate);
1543 
1544         /* Kernel probe */
1545         TNF_PROBE_1(thread_state, "thread", /* CSTYLED */,
1546             tnf_microstate, state, LMS_USER);
1547 
1548         return;
1549 
1550 cleanup:        /* system traps end up here */
1551         ASSERT(!(type & USER));
1552 }
1553 
1554 /*
1555  * Patch non-zero to disable preemption of threads in the kernel.
1556  */
1557 int IGNORE_KERNEL_PREEMPTION = 0;       /* XXX - delete this someday */
1558 
1559 struct kpreempt_cnts {          /* kernel preemption statistics */
1560         int     kpc_idle;       /* executing idle thread */
1561         int     kpc_intr;       /* executing interrupt thread */
1562         int     kpc_clock;      /* executing clock thread */
1563         int     kpc_blocked;    /* thread has blocked preemption (t_preempt) */
1564         int     kpc_notonproc;  /* thread is surrendering processor */
1565         int     kpc_inswtch;    /* thread has ratified scheduling decision */
1566         int     kpc_prilevel;   /* processor interrupt level is too high */
1567         int     kpc_apreempt;   /* asynchronous preemption */
1568         int     kpc_spreempt;   /* synchronous preemption */
1569 } kpreempt_cnts;
1570 
1571 /*
1572  * kernel preemption: forced rescheduling, preempt the running kernel thread.
1573  *      the argument is old PIL for an interrupt,
1574  *      or the distingished value KPREEMPT_SYNC.
1575  */
1576 void
1577 kpreempt(int asyncspl)
1578 {
1579         kthread_t *ct = curthread;
1580 
1581         if (IGNORE_KERNEL_PREEMPTION) {
1582                 aston(CPU->cpu_dispthread);
1583                 return;
1584         }
1585 
1586         /*
1587          * Check that conditions are right for kernel preemption
1588          */
1589         do {
1590                 if (ct->t_preempt) {
1591                         /*
1592                          * either a privileged thread (idle, panic, interrupt)
1593                          * or will check when t_preempt is lowered
1594                          * We need to specifically handle the case where
1595                          * the thread is in the middle of swtch (resume has
1596                          * been called) and has its t_preempt set
1597                          * [idle thread and a thread which is in kpreempt
1598                          * already] and then a high priority thread is
1599                          * available in the local dispatch queue.
1600                          * In this case the resumed thread needs to take a
1601                          * trap so that it can call kpreempt. We achieve
1602                          * this by using siron().
1603                          * How do we detect this condition:
1604                          * idle thread is running and is in the midst of
1605                          * resume: curthread->t_pri == -1 && CPU->dispthread
1606                          * != CPU->thread
1607                          * Need to ensure that this happens only at high pil
1608                          * resume is called at high pil
1609                          * Only in resume_from_idle is the pil changed.
1610                          */
1611                         if (ct->t_pri < 0) {
1612                                 kpreempt_cnts.kpc_idle++;
1613                                 if (CPU->cpu_dispthread != CPU->cpu_thread)
1614                                         siron();
1615                         } else if (ct->t_flag & T_INTR_THREAD) {
1616                                 kpreempt_cnts.kpc_intr++;
1617                                 if (ct->t_pil == CLOCK_LEVEL)
1618                                         kpreempt_cnts.kpc_clock++;
1619                         } else {
1620                                 kpreempt_cnts.kpc_blocked++;
1621                                 if (CPU->cpu_dispthread != CPU->cpu_thread)
1622                                         siron();
1623                         }
1624                         aston(CPU->cpu_dispthread);
1625                         return;
1626                 }
1627                 if (ct->t_state != TS_ONPROC ||
1628                     ct->t_disp_queue != CPU->cpu_disp) {
1629                         /* this thread will be calling swtch() shortly */
1630                         kpreempt_cnts.kpc_notonproc++;
1631                         if (CPU->cpu_thread != CPU->cpu_dispthread) {
1632                                 /* already in swtch(), force another */
1633                                 kpreempt_cnts.kpc_inswtch++;
1634                                 siron();
1635                         }
1636                         return;
1637                 }
1638                 if (getpil() >= DISP_LEVEL) {
1639                         /*
1640                          * We can't preempt this thread if it is at
1641                          * a PIL >= DISP_LEVEL since it may be holding
1642                          * a spin lock (like sched_lock).
1643                          */
1644                         siron();        /* check back later */
1645                         kpreempt_cnts.kpc_prilevel++;
1646                         return;
1647                 }
1648                 if (!interrupts_enabled()) {
1649                         /*
1650                          * Can't preempt while running with ints disabled
1651                          */
1652                         kpreempt_cnts.kpc_prilevel++;
1653                         return;
1654                 }
1655                 if (asyncspl != KPREEMPT_SYNC)
1656                         kpreempt_cnts.kpc_apreempt++;
1657                 else
1658                         kpreempt_cnts.kpc_spreempt++;
1659 
1660                 ct->t_preempt++;
1661                 preempt();
1662                 ct->t_preempt--;
1663         } while (CPU->cpu_kprunrun);
1664 }
1665 
1666 /*
1667  * Print out debugging info.
1668  */
1669 static void
1670 showregs(uint_t type, struct regs *rp, caddr_t addr)
1671 {
1672         int s;
1673 
1674         s = spl7();
1675         type &= ~USER;
1676         if (PTOU(curproc)->u_comm[0])
1677                 printf("%s: ", PTOU(curproc)->u_comm);
1678         if (type < TRAP_TYPES)
1679                 printf("#%s %s\n", trap_type_mnemonic[type], trap_type[type]);
1680         else
1681                 switch (type) {
1682                 case T_SYSCALL:
1683                         printf("Syscall Trap:\n");
1684                         break;
1685                 case T_AST:
1686                         printf("AST\n");
1687                         break;
1688                 default:
1689                         printf("Bad Trap = %d\n", type);
1690                         break;
1691                 }
1692         if (type == T_PGFLT) {
1693                 printf("Bad %s fault at addr=0x%lx\n",
1694                     USERMODE(rp->r_cs) ? "user": "kernel", (uintptr_t)addr);
1695         } else if (addr) {
1696                 printf("addr=0x%lx\n", (uintptr_t)addr);
1697         }
1698 
1699         printf("pid=%d, pc=0x%lx, sp=0x%lx, eflags=0x%lx\n",
1700             (ttoproc(curthread) && ttoproc(curthread)->p_pidp) ?
1701             ttoproc(curthread)->p_pid : 0, rp->r_pc, rp->r_sp, rp->r_ps);
1702 
1703 #if defined(__lint)
1704         /*
1705          * this clause can be deleted when lint bug 4870403 is fixed
1706          * (lint thinks that bit 32 is illegal in a %b format string)
1707          */
1708         printf("cr0: %x cr4: %b\n",
1709             (uint_t)getcr0(), (uint_t)getcr4(), FMT_CR4);
1710 #else
1711         printf("cr0: %b cr4: %b\n",
1712             (uint_t)getcr0(), FMT_CR0, (uint_t)getcr4(), FMT_CR4);
1713 #endif  /* __lint */
1714 
1715         printf("cr2: %lx", getcr2());
1716 #if !defined(__xpv)
1717         printf("cr3: %lx", getcr3());
1718 #if defined(__amd64)
1719         printf("cr8: %lx\n", getcr8());
1720 #endif
1721 #endif
1722         printf("\n");
1723 
1724         dumpregs(rp);
1725         splx(s);
1726 }
1727 
1728 static void
1729 dumpregs(struct regs *rp)
1730 {
1731 #if defined(__amd64)
1732         const char fmt[] = "\t%3s: %16lx %3s: %16lx %3s: %16lx\n";
1733 
1734         printf(fmt, "rdi", rp->r_rdi, "rsi", rp->r_rsi, "rdx", rp->r_rdx);
1735         printf(fmt, "rcx", rp->r_rcx, " r8", rp->r_r8, " r9", rp->r_r9);
1736         printf(fmt, "rax", rp->r_rax, "rbx", rp->r_rbx, "rbp", rp->r_rbp);
1737         printf(fmt, "r10", rp->r_r10, "r11", rp->r_r11, "r12", rp->r_r12);
1738         printf(fmt, "r13", rp->r_r13, "r14", rp->r_r14, "r15", rp->r_r15);
1739 
1740         printf(fmt, "fsb", rdmsr(MSR_AMD_FSBASE), "gsb", rdmsr(MSR_AMD_GSBASE),
1741             " ds", rp->r_ds);
1742         printf(fmt, " es", rp->r_es, " fs", rp->r_fs, " gs", rp->r_gs);
1743 
1744         printf(fmt, "trp", rp->r_trapno, "err", rp->r_err, "rip", rp->r_rip);
1745         printf(fmt, " cs", rp->r_cs, "rfl", rp->r_rfl, "rsp", rp->r_rsp);
1746 
1747         printf("\t%3s: %16lx\n", " ss", rp->r_ss);
1748 
1749 #elif defined(__i386)
1750         const char fmt[] = "\t%3s: %8lx %3s: %8lx %3s: %8lx %3s: %8lx\n";
1751 
1752         printf(fmt, " gs", rp->r_gs, " fs", rp->r_fs,
1753             " es", rp->r_es, " ds", rp->r_ds);
1754         printf(fmt, "edi", rp->r_edi, "esi", rp->r_esi,
1755             "ebp", rp->r_ebp, "esp", rp->r_esp);
1756         printf(fmt, "ebx", rp->r_ebx, "edx", rp->r_edx,
1757             "ecx", rp->r_ecx, "eax", rp->r_eax);
1758         printf(fmt, "trp", rp->r_trapno, "err", rp->r_err,
1759             "eip", rp->r_eip, " cs", rp->r_cs);
1760         printf("\t%3s: %8lx %3s: %8lx %3s: %8lx\n",
1761             "efl", rp->r_efl, "usp", rp->r_uesp, " ss", rp->r_ss);
1762 
1763 #endif  /* __i386 */
1764 }
1765 
1766 /*
1767  * Test to see if the instruction is iret on i386 or iretq on amd64.
1768  *
1769  * On the hypervisor we can only test for nopop_sys_rtt_syscall. If true
1770  * then we are in the context of hypervisor's failsafe handler because it
1771  * tried to iret and failed due to a bad selector. See xen_failsafe_callback.
1772  */
1773 static int
1774 instr_is_iret(caddr_t pc)
1775 {
1776 
1777 #if defined(__xpv)
1778         extern void nopop_sys_rtt_syscall(void);
1779         return ((pc == (caddr_t)nopop_sys_rtt_syscall) ? 1 : 0);
1780 
1781 #else
1782 
1783 #if defined(__amd64)
1784         static const uint8_t iret_insn[2] = { 0x48, 0xcf };     /* iretq */
1785 
1786 #elif defined(__i386)
1787         static const uint8_t iret_insn[1] = { 0xcf };           /* iret */
1788 #endif  /* __i386 */
1789         return (bcmp(pc, iret_insn, sizeof (iret_insn)) == 0);
1790 
1791 #endif  /* __xpv */
1792 }
1793 
1794 #if defined(__i386)
1795 
1796 /*
1797  * Test to see if the instruction is part of __SEGREGS_POP
1798  *
1799  * Note carefully the appallingly awful dependency between
1800  * the instruction sequence used in __SEGREGS_POP and these
1801  * instructions encoded here.
1802  */
1803 static int
1804 instr_is_segregs_pop(caddr_t pc)
1805 {
1806         static const uint8_t movw_0_esp_gs[4] = { 0x8e, 0x6c, 0x24, 0x0 };
1807         static const uint8_t movw_4_esp_fs[4] = { 0x8e, 0x64, 0x24, 0x4 };
1808         static const uint8_t movw_8_esp_es[4] = { 0x8e, 0x44, 0x24, 0x8 };
1809         static const uint8_t movw_c_esp_ds[4] = { 0x8e, 0x5c, 0x24, 0xc };
1810 
1811         if (bcmp(pc, movw_0_esp_gs, sizeof (movw_0_esp_gs)) == 0 ||
1812             bcmp(pc, movw_4_esp_fs, sizeof (movw_4_esp_fs)) == 0 ||
1813             bcmp(pc, movw_8_esp_es, sizeof (movw_8_esp_es)) == 0 ||
1814             bcmp(pc, movw_c_esp_ds, sizeof (movw_c_esp_ds)) == 0)
1815                 return (1);
1816 
1817         return (0);
1818 }
1819 
1820 #endif  /* __i386 */
1821 
1822 /*
1823  * Test to see if the instruction is part of _sys_rtt.
1824  *
1825  * Again on the hypervisor if we try to IRET to user land with a bad code
1826  * or stack selector we will get vectored through xen_failsafe_callback.
1827  * In which case we assume we got here via _sys_rtt since we only allow
1828  * IRET to user land to take place in _sys_rtt.
1829  */
1830 static int
1831 instr_is_sys_rtt(caddr_t pc)
1832 {
1833         extern void _sys_rtt(), _sys_rtt_end();
1834 
1835         if ((uintptr_t)pc < (uintptr_t)_sys_rtt ||
1836             (uintptr_t)pc > (uintptr_t)_sys_rtt_end)
1837                 return (0);
1838 
1839         return (1);
1840 }
1841 
1842 /*
1843  * Handle #gp faults in kernel mode.
1844  *
1845  * One legitimate way this can happen is if we attempt to update segment
1846  * registers to naughty values on the way out of the kernel.
1847  *
1848  * This can happen in a couple of ways: someone - either accidentally or
1849  * on purpose - creates (setcontext(2), lwp_create(2)) or modifies
1850  * (signal(2)) a ucontext that contains silly segment register values.
1851  * Or someone - either accidentally or on purpose - modifies the prgregset_t
1852  * of a subject process via /proc to contain silly segment register values.
1853  *
1854  * (The unfortunate part is that we can end up discovering the bad segment
1855  * register value in the middle of an 'iret' after we've popped most of the
1856  * stack.  So it becomes quite difficult to associate an accurate ucontext
1857  * with the lwp, because the act of taking the #gp trap overwrites most of
1858  * what we were going to send the lwp.)
1859  *
1860  * OTOH if it turns out that's -not- the problem, and we're -not- an lwp
1861  * trying to return to user mode and we get a #gp fault, then we need
1862  * to die() -- which will happen if we return non-zero from this routine.
1863  */
1864 static int
1865 kern_gpfault(struct regs *rp)
1866 {
1867         kthread_t *t = curthread;
1868         proc_t *p = ttoproc(t);
1869         klwp_t *lwp = ttolwp(t);
1870         struct regs tmpregs, *trp = NULL;
1871         caddr_t pc = (caddr_t)rp->r_pc;
1872         int v;
1873         uint32_t auditing = AU_AUDITING();
1874 
1875         /*
1876          * if we're not an lwp, or in the case of running native the
1877          * pc range is outside _sys_rtt, then we should immediately
1878          * be die()ing horribly.
1879          */
1880         if (lwp == NULL || !instr_is_sys_rtt(pc))
1881                 return (1);
1882 
1883         /*
1884          * So at least we're in the right part of the kernel.
1885          *
1886          * Disassemble the instruction at the faulting pc.
1887          * Once we know what it is, we carefully reconstruct the stack
1888          * based on the order in which the stack is deconstructed in
1889          * _sys_rtt. Ew.
1890          */
1891         if (instr_is_iret(pc)) {
1892                 /*
1893                  * We took the #gp while trying to perform the IRET.
1894                  * This means that either %cs or %ss are bad.
1895                  * All we know for sure is that most of the general
1896                  * registers have been restored, including the
1897                  * segment registers, and all we have left on the
1898                  * topmost part of the lwp's stack are the
1899                  * registers that the iretq was unable to consume.
1900                  *
1901                  * All the rest of the state was crushed by the #gp
1902                  * which pushed -its- registers atop our old save area
1903                  * (because we had to decrement the stack pointer, sigh) so
1904                  * all that we can try and do is to reconstruct the
1905                  * crushed frame from the #gp trap frame itself.
1906                  */
1907                 trp = &tmpregs;
1908                 trp->r_ss = lwptoregs(lwp)->r_ss;
1909                 trp->r_sp = lwptoregs(lwp)->r_sp;
1910                 trp->r_ps = lwptoregs(lwp)->r_ps;
1911                 trp->r_cs = lwptoregs(lwp)->r_cs;
1912                 trp->r_pc = lwptoregs(lwp)->r_pc;
1913                 bcopy(rp, trp, offsetof(struct regs, r_pc));
1914 
1915                 /*
1916                  * Validate simple math
1917                  */
1918                 ASSERT(trp->r_pc == lwptoregs(lwp)->r_pc);
1919                 ASSERT(trp->r_err == rp->r_err);
1920 
1921 
1922 
1923         }
1924 
1925 #if defined(__amd64)
1926         if (trp == NULL && lwp->lwp_pcb.pcb_rupdate != 0) {
1927 
1928                 /*
1929                  * This is the common case -- we're trying to load
1930                  * a bad segment register value in the only section
1931                  * of kernel code that ever loads segment registers.
1932                  *
1933                  * We don't need to do anything at this point because
1934                  * the pcb contains all the pending segment register
1935                  * state, and the regs are still intact because we
1936                  * didn't adjust the stack pointer yet.  Given the fidelity
1937                  * of all this, we could conceivably send a signal
1938                  * to the lwp, rather than core-ing.
1939                  */
1940                 trp = lwptoregs(lwp);
1941                 ASSERT((caddr_t)trp == (caddr_t)rp->r_sp);
1942         }
1943 
1944 #elif defined(__i386)
1945 
1946         if (trp == NULL && instr_is_segregs_pop(pc))
1947                 trp = lwptoregs(lwp);
1948 
1949 #endif  /* __i386 */
1950 
1951         if (trp == NULL)
1952                 return (1);
1953 
1954         /*
1955          * If we get to here, we're reasonably confident that we've
1956          * correctly decoded what happened on the way out of the kernel.
1957          * Rewrite the lwp's registers so that we can create a core dump
1958          * the (at least vaguely) represents the mcontext we were
1959          * being asked to restore when things went so terribly wrong.
1960          */
1961 
1962         /*
1963          * Make sure that we have a meaningful %trapno and %err.
1964          */
1965         trp->r_trapno = rp->r_trapno;
1966         trp->r_err = rp->r_err;
1967 
1968         if ((caddr_t)trp != (caddr_t)lwptoregs(lwp))
1969                 bcopy(trp, lwptoregs(lwp), sizeof (*trp));
1970 
1971 
1972         mutex_enter(&p->p_lock);
1973         lwp->lwp_cursig = SIGSEGV;
1974         mutex_exit(&p->p_lock);
1975 
1976         /*
1977          * Terminate all LWPs but don't discard them.  If another lwp beat
1978          * us to the punch by calling exit(), evaporate now.
1979          */
1980         proc_is_exiting(p);
1981         if (exitlwps(1) != 0) {
1982                 mutex_enter(&p->p_lock);
1983                 lwp_exit();
1984         }
1985 
1986         if (auditing)           /* audit core dump */
1987                 audit_core_start(SIGSEGV);
1988         v = core(SIGSEGV, B_FALSE);
1989         if (auditing)           /* audit core dump */
1990                 audit_core_finish(v ? CLD_KILLED : CLD_DUMPED);
1991         exit(v ? CLD_KILLED : CLD_DUMPED, SIGSEGV);
1992         return (0);
1993 }
1994 
1995 /*
1996  * dump_tss() - Display the TSS structure
1997  */
1998 
1999 #if !defined(__xpv)
2000 #if defined(__amd64)
2001 
2002 static void
2003 dump_tss(void)
2004 {
2005         const char tss_fmt[] = "tss.%s:\t0x%p\n";  /* Format string */
2006         tss_t *tss = CPU->cpu_tss;
2007 
2008         printf(tss_fmt, "tss_rsp0", (void *)tss->tss_rsp0);
2009         printf(tss_fmt, "tss_rsp1", (void *)tss->tss_rsp1);
2010         printf(tss_fmt, "tss_rsp2", (void *)tss->tss_rsp2);
2011 
2012         printf(tss_fmt, "tss_ist1", (void *)tss->tss_ist1);
2013         printf(tss_fmt, "tss_ist2", (void *)tss->tss_ist2);
2014         printf(tss_fmt, "tss_ist3", (void *)tss->tss_ist3);
2015         printf(tss_fmt, "tss_ist4", (void *)tss->tss_ist4);
2016         printf(tss_fmt, "tss_ist5", (void *)tss->tss_ist5);
2017         printf(tss_fmt, "tss_ist6", (void *)tss->tss_ist6);
2018         printf(tss_fmt, "tss_ist7", (void *)tss->tss_ist7);
2019 }
2020 
2021 #elif defined(__i386)
2022 
2023 static void
2024 dump_tss(void)
2025 {
2026         const char tss_fmt[] = "tss.%s:\t0x%p\n";  /* Format string */
2027         tss_t *tss = CPU->cpu_tss;
2028 
2029         printf(tss_fmt, "tss_link", (void *)(uintptr_t)tss->tss_link);
2030         printf(tss_fmt, "tss_esp0", (void *)(uintptr_t)tss->tss_esp0);
2031         printf(tss_fmt, "tss_ss0", (void *)(uintptr_t)tss->tss_ss0);
2032         printf(tss_fmt, "tss_esp1", (void *)(uintptr_t)tss->tss_esp1);
2033         printf(tss_fmt, "tss_ss1", (void *)(uintptr_t)tss->tss_ss1);
2034         printf(tss_fmt, "tss_esp2", (void *)(uintptr_t)tss->tss_esp2);
2035         printf(tss_fmt, "tss_ss2", (void *)(uintptr_t)tss->tss_ss2);
2036         printf(tss_fmt, "tss_cr3", (void *)(uintptr_t)tss->tss_cr3);
2037         printf(tss_fmt, "tss_eip", (void *)(uintptr_t)tss->tss_eip);
2038         printf(tss_fmt, "tss_eflags", (void *)(uintptr_t)tss->tss_eflags);
2039         printf(tss_fmt, "tss_eax", (void *)(uintptr_t)tss->tss_eax);
2040         printf(tss_fmt, "tss_ebx", (void *)(uintptr_t)tss->tss_ebx);
2041         printf(tss_fmt, "tss_ecx", (void *)(uintptr_t)tss->tss_ecx);
2042         printf(tss_fmt, "tss_edx", (void *)(uintptr_t)tss->tss_edx);
2043         printf(tss_fmt, "tss_esp", (void *)(uintptr_t)tss->tss_esp);
2044 }
2045 
2046 #endif  /* __amd64 */
2047 #endif  /* !__xpv */
2048 
2049 #if defined(TRAPTRACE)
2050 
2051 int ttrace_nrec = 10;           /* number of records to dump out */
2052 int ttrace_dump_nregs = 0;      /* dump out this many records with regs too */
2053 
2054 /*
2055  * Dump out the last ttrace_nrec traptrace records on each CPU
2056  */
2057 static void
2058 dump_ttrace(void)
2059 {
2060         trap_trace_ctl_t *ttc;
2061         trap_trace_rec_t *rec;
2062         uintptr_t current;
2063         int i, j, k;
2064         int n = NCPU;
2065 #if defined(__amd64)
2066         const char banner[] =
2067             "\ncpu          address    timestamp "
2068             "type  vc  handler   pc\n";
2069         const char fmt1[] = "%3d %016lx %12llx ";
2070 #elif defined(__i386)
2071         const char banner[] =
2072             "\ncpu  address     timestamp type  vc  handler   pc\n";
2073         const char fmt1[] = "%3d %08lx %12llx ";
2074 #endif
2075         const char fmt2[] = "%4s %3x ";
2076         const char fmt3[] = "%8s ";
2077 
2078         if (ttrace_nrec == 0)
2079                 return;
2080 
2081         printf(banner);
2082 
2083         for (i = 0; i < n; i++) {
2084                 ttc = &trap_trace_ctl[i];
2085                 if (ttc->ttc_first == NULL)
2086                         continue;
2087 
2088                 current = ttc->ttc_next - sizeof (trap_trace_rec_t);
2089                 for (j = 0; j < ttrace_nrec; j++) {
2090                         struct sysent   *sys;
2091                         struct autovec  *vec;
2092                         extern struct av_head autovect[];
2093                         int type;
2094                         ulong_t off;
2095                         char *sym, *stype;
2096 
2097                         if (current < ttc->ttc_first)
2098                                 current =
2099                                     ttc->ttc_limit - sizeof (trap_trace_rec_t);
2100 
2101                         if (current == NULL)
2102                                 continue;
2103 
2104                         rec = (trap_trace_rec_t *)current;
2105 
2106                         if (rec->ttr_stamp == 0)
2107                                 break;
2108 
2109                         printf(fmt1, i, (uintptr_t)rec, rec->ttr_stamp);
2110 
2111                         switch (rec->ttr_marker) {
2112                         case TT_SYSCALL:
2113                         case TT_SYSENTER:
2114                         case TT_SYSC:
2115                         case TT_SYSC64:
2116 #if defined(__amd64)
2117                                 sys = &sysent32[rec->ttr_sysnum];
2118                                 switch (rec->ttr_marker) {
2119                                 case TT_SYSC64:
2120                                         sys = &sysent[rec->ttr_sysnum];
2121                                         /*FALLTHROUGH*/
2122 #elif defined(__i386)
2123                                 sys = &sysent[rec->ttr_sysnum];
2124                                 switch (rec->ttr_marker) {
2125                                 case TT_SYSC64:
2126 #endif
2127                                 case TT_SYSC:
2128                                         stype = "sysc"; /* syscall */
2129                                         break;
2130                                 case TT_SYSCALL:
2131                                         stype = "lcal"; /* lcall */
2132                                         break;
2133                                 case TT_SYSENTER:
2134                                         stype = "syse"; /* sysenter */
2135                                         break;
2136                                 default:
2137                                         break;
2138                                 }
2139                                 printf(fmt2, "sysc", rec->ttr_sysnum);
2140                                 if (sys != NULL) {
2141                                         sym = kobj_getsymname(
2142                                             (uintptr_t)sys->sy_callc,
2143                                             &off);
2144                                         if (sym != NULL)
2145                                                 printf(fmt3, sym);
2146                                         else
2147                                                 printf("%p ", sys->sy_callc);
2148                                 } else {
2149                                         printf(fmt3, "unknown");
2150                                 }
2151                                 break;
2152 
2153                         case TT_INTERRUPT:
2154                                 printf(fmt2, "intr", rec->ttr_vector);
2155                                 if (get_intr_handler != NULL)
2156                                         vec = (struct autovec *)
2157                                             (*get_intr_handler)
2158                                             (rec->ttr_cpuid, rec->ttr_vector);
2159                                 else
2160                                         vec =
2161                                             autovect[rec->ttr_vector].avh_link;
2162 
2163                                 if (vec != NULL) {
2164                                         sym = kobj_getsymname(
2165                                             (uintptr_t)vec->av_vector, &off);
2166                                         if (sym != NULL)
2167                                                 printf(fmt3, sym);
2168                                         else
2169                                                 printf("%p ", vec->av_vector);
2170                                 } else {
2171                                         printf(fmt3, "unknown ");
2172                                 }
2173                                 break;
2174 
2175                         case TT_TRAP:
2176                         case TT_EVENT:
2177                                 type = rec->ttr_regs.r_trapno;
2178                                 printf(fmt2, "trap", type);
2179                                 if (type < TRAP_TYPES)
2180                                         printf("     #%s ",
2181                                             trap_type_mnemonic[type]);
2182                                 else
2183                                         switch (type) {
2184                                         case T_AST:
2185                                                 printf(fmt3, "ast");
2186                                                 break;
2187                                         default:
2188                                                 printf(fmt3, "");
2189                                                 break;
2190                                         }
2191                                 break;
2192 
2193                         default:
2194                                 break;
2195                         }
2196 
2197                         sym = kobj_getsymname(rec->ttr_regs.r_pc, &off);
2198                         if (sym != NULL)
2199                                 printf("%s+%lx\n", sym, off);
2200                         else
2201                                 printf("%lx\n", rec->ttr_regs.r_pc);
2202 
2203                         if (ttrace_dump_nregs-- > 0) {
2204                                 int s;
2205 
2206                                 if (rec->ttr_marker == TT_INTERRUPT)
2207                                         printf(
2208                                             "\t\tipl %x spl %x pri %x\n",
2209                                             rec->ttr_ipl,
2210                                             rec->ttr_spl,
2211                                             rec->ttr_pri);
2212 
2213                                 dumpregs(&rec->ttr_regs);
2214 
2215                                 printf("\t%3s: %p\n\n", " ct",
2216                                     (void *)rec->ttr_curthread);
2217 
2218                                 /*
2219                                  * print out the pc stack that we recorded
2220                                  * at trap time (if any)
2221                                  */
2222                                 for (s = 0; s < rec->ttr_sdepth; s++) {
2223                                         uintptr_t fullpc;
2224 
2225                                         if (s >= TTR_STACK_DEPTH) {
2226                                                 printf("ttr_sdepth corrupt\n");
2227                                                 break;
2228                                         }
2229 
2230                                         fullpc = (uintptr_t)rec->ttr_stack[s];
2231 
2232                                         sym = kobj_getsymname(fullpc, &off);
2233                                         if (sym != NULL)
2234                                                 printf("-> %s+0x%lx()\n",
2235                                                     sym, off);
2236                                         else
2237                                                 printf("-> 0x%lx()\n", fullpc);
2238                                 }
2239                                 printf("\n");
2240                         }
2241                         current -= sizeof (trap_trace_rec_t);
2242                 }
2243         }
2244 }
2245 
2246 #endif  /* TRAPTRACE */
2247 
2248 void
2249 panic_showtrap(struct panic_trap_info *tip)
2250 {
2251         showregs(tip->trap_type, tip->trap_regs, tip->trap_addr);
2252 
2253 #if defined(TRAPTRACE)
2254         dump_ttrace();
2255 #endif
2256 
2257 #if !defined(__xpv)
2258         if (tip->trap_type == T_DBLFLT)
2259                 dump_tss();
2260 #endif
2261 }
2262 
2263 void
2264 panic_savetrap(panic_data_t *pdp, struct panic_trap_info *tip)
2265 {
2266         panic_saveregs(pdp, tip->trap_regs);
2267 }
--- EOF ---