bsd/dev/i386/fbt_x86.c

   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, Version 1.0 only
   6  * (the "License").  You may not use this file except in compliance
   7  * with the License.
   8  *
   9  * You can obtain a copy of the license at usr/src/OPENSOLARIS.LICENSE
  10  * or http://www.opensolaris.org/os/licensing.
  11  * See the License for the specific language governing permissions
  12  * and limitations under the License.
  13  *
  14  * When distributing Covered Code, include this CDDL HEADER in each
  15  * file and include the License file at usr/src/OPENSOLARIS.LICENSE.
  16  * If applicable, add the following below this CDDL HEADER, with the
  17  * fields enclosed by brackets "[]" replaced with your own identifying
  18  * information: Portions Copyright [yyyy] [name of copyright owner]
  19  *
  20  * CDDL HEADER END
  21  */
  22 /*
  23  * Copyright 2005 Sun Microsystems, Inc.  All rights reserved.
  24  * Use is subject to license terms.
  25  */
  26
  27 /* #pragma ident        "@(#)fbt.c      1.15    05/09/19 SMI" */
  28
  29 #ifdef KERNEL
  30 #ifndef _KERNEL
  31 #define _KERNEL /* Solaris vs. Darwin */
  32 #endif
  33 #endif
  34
  35 #define MACH__POSIX_C_SOURCE_PRIVATE 1 /* pulls in suitable savearea from mach/ppc/thread_status.h */
  36 #include <kern/thread.h>
  37 #include <mach/thread_status.h>
  38 #include <mach/vm_param.h>
  39 #include <mach-o/loader.h>
  40 #include <mach-o/nlist.h>
  41
  42 extern struct mach_header _mh_execute_header; /* the kernel's mach header */
  43
  44 #include <sys/param.h>
  45 #include <sys/systm.h>
  46 #include <sys/errno.h>
  47 #include <sys/stat.h>
  48 #include <sys/ioctl.h>
  49 #include <sys/conf.h>
  50 #include <sys/fcntl.h>
  51 #include <miscfs/devfs/devfs.h>
  52
  53 #include <sys/dtrace.h>
  54 #include <sys/dtrace_impl.h>
  55 #include <sys/fbt.h>
  56
  57 #include <sys/dtrace_glue.h>
  58
  59 #define DTRACE_INVOP_NOP_SKIP 1
  60 #define DTRACE_INVOP_MOVL_ESP_EBP 10
  61 #define DTRACE_INVOP_MOVL_ESP_EBP_SKIP 2
  62 #define DTRACE_INVOP_LEAVE_SKIP 1
  63
  64 #define FBT_PUSHL_EBP                   0x55
  65 #define FBT_MOVL_ESP_EBP0_V0    0x8b
  66 #define FBT_MOVL_ESP_EBP1_V0    0xec
  67 #define FBT_MOVL_ESP_EBP0_V1    0x89
  68 #define FBT_MOVL_ESP_EBP1_V1    0xe5
  69 #define FBT_REX_RSP_RBP                 0x48
  70
  71 #define FBT_POPL_EBP                    0x5d
  72 #define FBT_RET                                 0xc3
  73 #define FBT_RET_IMM16                   0xc2
  74 #define FBT_LEAVE                               0xc9
  75 #define FBT_JMP_SHORT_REL               0xeb /* Jump short, relative, displacement relative to next instr. */
  76 #define FBT_JMP_NEAR_REL                0xe9 /* Jump near, relative, displacement relative to next instr. */
  77 #define FBT_JMP_FAR_ABS                 0xea /* Jump far, absolute, address given in operand */
  78 #define FBT_RET_LEN                             1
  79 #define FBT_RET_IMM16_LEN               3
  80 #define FBT_JMP_SHORT_REL_LEN   2
  81 #define FBT_JMP_NEAR_REL_LEN    5
  82 #define FBT_JMP_FAR_ABS_LEN             5
  83
  84 #define FBT_PATCHVAL                    0xf0
  85 #define FBT_AFRAMES_ENTRY               7
  86 #define FBT_AFRAMES_RETURN              6
  87
  88 #define FBT_ENTRY       "entry"
  89 #define FBT_RETURN      "return"
  90 #define FBT_ADDR2NDX(addr)      ((((uintptr_t)(addr)) >> 4) & fbt_probetab_mask)
  91
  92 extern dtrace_provider_id_t     fbt_id;
  93 extern fbt_probe_t              **fbt_probetab;
  94 extern int                      fbt_probetab_mask;
  95
  96 /*
  97  * Critical routines that must not be probed. PR_5221096, PR_5379018.
  98  */
  99
 100 static const char * critical_blacklist[] =
 101 {
 102         "bcopy_phys",
 103         "console_cpu_alloc",
 104         "console_cpu_free",
 105         "cpu_IA32e_disable",
 106         "cpu_IA32e_enable",
 107         "cpu_control",
 108         "cpu_data_alloc",
 109         "cpu_desc_init",
 110         "cpu_desc_init64",
 111         "cpu_desc_load64",
 112         "cpu_exit_wait",
 113         "cpu_info",
 114         "cpu_info_count",
 115         "cpu_init",
 116         "cpu_interrupt",
 117         "cpu_machine_init",
 118         "cpu_mode_init",
 119         "cpu_processor_alloc",
 120         "cpu_processor_free",
 121         "cpu_signal_handler",
 122         "cpu_sleep",
 123         "cpu_start",
 124         "cpu_subtype",
 125         "cpu_thread_alloc",
 126         "cpu_thread_halt",
 127         "cpu_thread_init",
 128         "cpu_threadtype",
 129         "cpu_to_processor",
 130         "cpu_topology_start",
 131         "cpu_type",
 132         "cpu_window_init",
 133         "cpuid_cpu_display",
 134         "handle_pending_TLB_flushes",
 135         "hw_compare_and_store",
 136         "machine_idle_cstate",
 137         "mca_cpu_alloc",
 138         "mca_cpu_init",
 139         "ml_nofault_copy",
 140         "pmap_cpu_alloc",
 141         "pmap_cpu_free",
 142         "pmap_cpu_high_map_vaddr",
 143         "pmap_cpu_high_shared_remap",
 144         "pmap_cpu_init",
 145         "register_cpu_setup_func",
 146         "unregister_cpu_setup_func"
 147 };
 148 #define CRITICAL_BLACKLIST_COUNT (sizeof(critical_blacklist)/sizeof(critical_blacklist[0]))
 149
 150 /*
 151  * The transitive closure of entry points that can be reached from probe context.
 152  * (Apart from routines whose names begin with dtrace_ or dtxnu_.)
 153  */
 154 static const char * probe_ctx_closure[] =
 155 {
 156         "Debugger",
 157         "OSCompareAndSwap",
 158         "absolutetime_to_microtime",
 159         "ast_pending",
 160         "clock_get_calendar_nanotime_nowait",
 161         "copyin",
 162         "copyin_user",
 163         "copyinstr",
 164         "copyout",
 165         "copyoutstr",
 166         "cpu_number",
 167         "current_proc",
 168         "current_processor",
 169         "current_task",
 170         "current_thread",
 171         "debug_enter",
 172         "find_user_regs",
 173         "flush_tlb64",
 174         "get_bsdtask_info",
 175         "get_bsdthread_info",
 176         "hw_atomic_and",
 177         "kauth_cred_get",
 178         "kauth_getgid",
 179         "kauth_getuid",
 180         "kernel_preempt_check",
 181         "mach_absolute_time",
 182         "max_valid_stack_address",
 183         "ml_at_interrupt_context",
 184         "ml_phys_write_byte_64",
 185         "ml_phys_write_half_64",
 186         "ml_phys_write_word_64",
 187         "ml_set_interrupts_enabled",
 188         "panic",
 189         "pmap64_pde",
 190         "pmap64_pdpt",
 191         "pmap_find_phys",
 192         "pmap_get_mapwindow",
 193         "pmap_pde",
 194         "pmap_pte",
 195         "pmap_put_mapwindow",
 196         "pmap_valid_page",
 197         "prf",
 198         "proc_is64bit",
 199         "proc_selfname",
 200         "proc_selfpid",
 201         "psignal_lock",
 202         "rtc_nanotime_load",
 203         "rtc_nanotime_read",
 204         "strlcpy",
 205         "sync_iss_to_iks_unconditionally",
 206         "timer_grab"
 207 };
 208 #define PROBE_CTX_CLOSURE_COUNT (sizeof(probe_ctx_closure)/sizeof(probe_ctx_closure[0]))
 209
 210
 211 static int _cmp(const void *a, const void *b)
 212 {
 213         return strcmp((const char *)a, *(const char **)b);
 214 }
 215
 216 static const void * bsearch(
 217         register const void *key,
 218         const void *base0,
 219         size_t nmemb,
 220         register size_t size,
 221         register int (*compar)(const void *, const void *)) {
 222
 223         register const char *base = base0;
 224         register size_t lim;
 225         register int cmp;
 226         register const void *p;
 227
 228         for (lim = nmemb; lim != 0; lim >>= 1) {
 229                 p = base + (lim >> 1) * size;
 230                 cmp = (*compar)(key, p);
 231                 if (cmp == 0)
 232                         return p;
 233                 if (cmp > 0) {  /* key > p: move right */
 234                         base = (const char *)p + size;
 235                         lim--;
 236                 }               /* else move left */
 237         }
 238         return (NULL);
 239 }
 240
 241 int
 242 fbt_invop(uintptr_t addr, uintptr_t *stack, uintptr_t rval)
 243 {
 244         uintptr_t stack0 = 0, stack1 = 0, stack2 = 0, stack3 = 0, stack4 = 0;
 245         fbt_probe_t *fbt = fbt_probetab[FBT_ADDR2NDX(addr)];
 246
 247         for (; fbt != NULL; fbt = fbt->fbtp_hashnext) {
 248                 if ((uintptr_t)fbt->fbtp_patchpoint == addr) {
 249
 250                         if (fbt->fbtp_roffset == 0) {
 251                                 uintptr_t *stacktop;
 252                                 if (CPU_ON_INTR(CPU))
 253                                         stacktop = (uintptr_t *)dtrace_get_cpu_int_stack_top();
 254                                 else
 255                                         stacktop = (uintptr_t *)(dtrace_get_kernel_stack(current_thread()) + KERNEL_STACK_SIZE);
 256
 257                                 stack += 1; /* skip over the target's pushl'd %ebp */
 258
 259                                 if (stack <= stacktop)
 260                                         CPU->cpu_dtrace_caller = *stack++;
 261                                 if (stack <= stacktop)
 262                                         stack0 = *stack++;
 263                                 if (stack <= stacktop)
 264                                         stack1 = *stack++;
 265                                 if (stack <= stacktop)
 266                                         stack2 = *stack++;
 267                                 if (stack <= stacktop)
 268                                         stack3 = *stack++;
 269                                 if (stack <= stacktop)
 270                                         stack4 = *stack++;
 271
 272                                 dtrace_probe(fbt->fbtp_id, stack0, stack1, stack2, stack3, stack4);
 273                                 CPU->cpu_dtrace_caller = 0;
 274                         } else {
 275                                 dtrace_probe(fbt->fbtp_id, fbt->fbtp_roffset, rval, 0, 0, 0);
 276                                 CPU->cpu_dtrace_caller = 0;
 277                         }
 278
 279                         return (fbt->fbtp_rval);
 280                 }
 281         }
 282
 283         return (0);
 284 }
 285
 286 #define IS_USER_TRAP(regs) (regs && (((regs)->cs & 3) != 0))
 287 #define T_INVALID_OPCODE 6
 288 #define FBT_EXCEPTION_CODE T_INVALID_OPCODE
 289
 290 kern_return_t
 291 fbt_perfCallback(
 292                 int                     trapno,
 293                 x86_saved_state_t       *tagged_regs,
 294                 __unused int        unused1,
 295                 __unused int        unused2)
 296 {
 297         kern_return_t retval = KERN_FAILURE;
 298         x86_saved_state32_t *saved_state = saved_state32(tagged_regs);
 299         struct x86_saved_state32_from_kernel *regs = (struct x86_saved_state32_from_kernel *)saved_state;
 300
 301         if (FBT_EXCEPTION_CODE == trapno && !IS_USER_TRAP(saved_state)) {
 302                 boolean_t oldlevel, cpu_64bit;
 303                 uint32_t esp_probe, *ebp, edi, fp, *pDst, delta = 0;
 304                 int emul;
 305
 306                 cpu_64bit = ml_is64bit();
 307                 oldlevel = ml_set_interrupts_enabled(FALSE);
 308
 309                 /* Calculate where the stack pointer was when the probe instruction "fired." */
 310                 if (cpu_64bit) {
 311                         esp_probe = saved_state->uesp; /* Easy, x86_64 establishes this value in idt64.s */
 312                 } else {
 313                         esp_probe = (uint32_t)&(regs[1]); /* Nasty, infer the location above the save area */
 314                 }
 315
 316                 emul = dtrace_invop( saved_state->eip, (uintptr_t *)esp_probe, saved_state->eax );
 317                 __asm__ volatile(".globl _dtrace_invop_callsite");
 318                 __asm__ volatile("_dtrace_invop_callsite:");
 319
 320                 switch (emul) {
 321                 case DTRACE_INVOP_NOP:
 322                         saved_state->eip += DTRACE_INVOP_NOP_SKIP;      /* Skip over the patched NOP */
 323                         retval = KERN_SUCCESS;
 324                         break;
 325
 326                 case DTRACE_INVOP_MOVL_ESP_EBP:
 327                         saved_state->ebp = esp_probe;                                           /* Emulate patched movl %esp,%ebp */
 328                         saved_state->eip += DTRACE_INVOP_MOVL_ESP_EBP_SKIP;     /* Skip over the bytes of the patched movl %esp,%ebp */
 329                         retval = KERN_SUCCESS;
 330                         break;
 331
 332                 case DTRACE_INVOP_POPL_EBP:
 333                 case DTRACE_INVOP_LEAVE:
 334 /*
 335  * Emulate first micro-op of patched leave: movl %ebp,%esp
 336  * fp points just below the return address slot for target's ret
 337  * and at the slot holding the frame pointer saved by the target's prologue.
 338  */
 339                         fp = saved_state->ebp;
 340 /* Emulate second micro-op of patched leave: patched popl %ebp
 341  * savearea ebp is set for the frame of the caller to target
 342  * The *live* %esp will be adjusted below for pop increment(s)
 343  */
 344                         saved_state->ebp = *(uint32_t *)fp;
 345 /* Skip over the patched leave */
 346                         saved_state->eip += DTRACE_INVOP_LEAVE_SKIP;
 347 /*
 348  * Lift the stack to account for the emulated leave
 349  * Account for words local in this frame
 350  * (in "case DTRACE_INVOP_POPL_EBP:" this is zero.)
 351  */
 352                         delta = ((uint32_t *)fp) - ((uint32_t *)esp_probe);
 353 /* Account for popping off the ebp (just accomplished by the emulation
 354  * above...)
 355  */
 356                         delta += 1;
 357
 358                         if (cpu_64bit)
 359                                 saved_state->uesp += (delta << 2);
 360
 361 /* XXX Fragile in the extreme. Obtain the value of %edi that our caller pushed
 362  * (on behalf of its caller -- trap_from_kernel()). Ultimately,
 363  * trap_from_kernel's stack pointer is restored from this slot.
 364  * This is sensitive to the manner in which the compiler preserves %edi,
 365  * and trap_from_kernel()'s internals.
 366  */
 367                         ebp = (uint32_t *)__builtin_frame_address(0);
 368                         ebp = (uint32_t *)*ebp;
 369                         edi = *(ebp - 1);
 370 /* Shift contents of stack */
 371                         for (pDst = (uint32_t *)fp;
 372                              pDst > (((uint32_t *)edi));
 373                                  pDst--)
 374                                 *pDst = pDst[-delta];
 375 /* Now adjust the value of %edi in our caller (kernel_trap)'s frame */
 376                         *(ebp - 1) = edi + (delta << 2);
 377
 378                         retval = KERN_SUCCESS;
 379                         break;
 380
 381                 default:
 382                         retval = KERN_FAILURE;
 383                         break;
 384                 }
 385                 ml_set_interrupts_enabled(oldlevel);
 386         }
 387
 388         return retval;
 389 }
 390
 391 /*ARGSUSED*/
 392 static void
 393 __fbt_provide_module(void *arg, struct modctl *ctl)
 394 {
 395 #pragma unused(arg)
 396         struct mach_header                      *mh;
 397         struct load_command         *cmd;
 398         struct segment_command      *orig_ts = NULL, *orig_le = NULL;
 399         struct symtab_command       *orig_st = NULL;
 400         struct nlist                *sym = NULL;
 401         char                                            *strings;
 402         uintptr_t                                       instrLow, instrHigh;
 403         char                                            *modname;
 404         unsigned int i, j;
 405
 406         int gIgnoreFBTBlacklist = 0;
 407         PE_parse_boot_argn("IgnoreFBTBlacklist", &gIgnoreFBTBlacklist, sizeof (gIgnoreFBTBlacklist));
 408
 409         mh = (struct mach_header *)(ctl->address);
 410         modname = ctl->mod_modname;
 411
 412         if (0 == ctl->address || 0 == ctl->size) /* Has the linker been jettisoned? */
 413                 return;
 414
 415         /*
 416          * Employees of dtrace and their families are ineligible.  Void
 417          * where prohibited.
 418          */
 419
 420         if (strcmp(modname, "com.apple.driver.dtrace") == 0)
 421                 return;
 422
 423         if (strstr(modname, "CHUD") != NULL)
 424                 return;
 425
 426         if (mh->magic != MH_MAGIC)
 427                 return;
 428
 429         cmd = (struct load_command *) &mh[1];
 430         for (i = 0; i < mh->ncmds; i++) {
 431                 if (cmd->cmd == LC_SEGMENT) {
 432                         struct segment_command *orig_sg = (struct segment_command *) cmd;
 433
 434                         if (strcmp(SEG_TEXT, orig_sg->segname) == 0)
 435                                 orig_ts = orig_sg;
 436                         else if (strcmp(SEG_LINKEDIT, orig_sg->segname) == 0)
 437                                 orig_le = orig_sg;
 438                         else if (strcmp("", orig_sg->segname) == 0)
 439                                 orig_ts = orig_sg; /* kexts have a single unnamed segment */
 440                 }
 441                 else if (cmd->cmd == LC_SYMTAB)
 442                         orig_st = (struct symtab_command *) cmd;
 443
 444                 cmd = (struct load_command *) ((caddr_t) cmd + cmd->cmdsize);
 445         }
 446
 447         if ((orig_ts == NULL) || (orig_st == NULL) || (orig_le == NULL))
 448                 return;
 449
 450         sym = (struct nlist *)orig_le->vmaddr;
 451         strings = ((char *)sym) + orig_st->nsyms * sizeof(struct nlist);
 452
 453         /* Find extent of the TEXT section */
 454         instrLow = (uintptr_t)orig_ts->vmaddr;
 455         instrHigh = (uintptr_t)(orig_ts->vmaddr + orig_ts->vmsize);
 456
 457         for (i = 0; i < orig_st->nsyms; i++) {
 458                 fbt_probe_t *fbt, *retfbt;
 459                 machine_inst_t *instr, *limit, theInstr, i1, i2;
 460                 uint8_t n_type = sym[i].n_type & (N_TYPE | N_EXT);
 461                 char *name = strings + sym[i].n_un.n_strx;
 462                 int size;
 463
 464                 /* Check that the symbol is a global and that it has a name. */
 465                 if (((N_SECT | N_EXT) != n_type && (N_ABS | N_EXT) != n_type))
 466                         continue;
 467
 468                 if (0 == sym[i].n_un.n_strx) /* iff a null, "", name. */
 469                         continue;
 470
 471                 /* Lop off omnipresent leading underscore. */
 472                 if (*name == '_')
 473                         name += 1;
 474
 475                 if (strstr(name, "dtrace_") == name &&
 476                     strstr(name, "dtrace_safe_") != name) {
 477                         /*
 478                          * Anything beginning with "dtrace_" may be called
 479                          * from probe context unless it explitly indicates
 480                          * that it won't be called from probe context by
 481                          * using the prefix "dtrace_safe_".
 482                          */
 483                         continue;
 484                 }
 485
 486                 if (strstr(name, "dsmos_") == name)
 487                         continue; /* Don't Steal Mac OS X! */
 488
 489         if (strstr(name, "dtxnu_") == name ||
 490                         strstr(name, "_dtrace") == name)
 491                         continue; /* Shims in dtrace.c */
 492
 493                 if (strstr(name, "chud") == name)
 494                         continue; /* Professional courtesy. */
 495
 496                 if (strstr(name, "hibernate_") == name)
 497                         continue; /* Let sleeping dogs lie. */
 498
 499                 if (0 == strcmp(name, "ZN9IOService14newTemperatureElPS_") || /* IOService::newTemperature */
 500                         0 == strcmp(name, "ZN9IOService26temperatureCriticalForZoneEPS_")) /* IOService::temperatureCriticalForZone */
 501                         continue; /* Per the fire code */
 502
 503                 /*
 504                  * Place no probes (illegal instructions) in the exception handling path!
 505                  */
 506                 if (0 == strcmp(name, "t_invop") ||
 507                         0 == strcmp(name, "enter_lohandler") ||
 508                         0 == strcmp(name, "lo_alltraps") ||
 509                         0 == strcmp(name, "kernel_trap") ||
 510                         0 == strcmp(name, "i386_astintr"))
 511                         continue;
 512
 513                 if (0 == strcmp(name, "current_thread") ||
 514                         0 == strcmp(name, "ast_pending") ||
 515                         0 == strcmp(name, "fbt_perfCallback") ||
 516                         0 == strcmp(name, "machine_thread_get_kern_state") ||
 517                         0 == strcmp(name, "ml_set_interrupts_enabled") ||
 518                         0 == strcmp(name, "dtrace_invop") ||
 519                         0 == strcmp(name, "fbt_invop") ||
 520                         0 == strcmp(name, "sdt_invop") ||
 521                         0 == strcmp(name, "max_valid_stack_address"))
 522                         continue;
 523
 524                 /*
 525                  * Voodoo.
 526                  */
 527                 if (strstr(name, "machine_stack_") == name ||
 528                         strstr(name, "mapping_") == name ||
 529                         0 == strcmp(name, "tmrCvt") ||
 530
 531                         strstr(name, "tsc_") == name ||
 532
 533                         strstr(name, "pmCPU") == name ||
 534                         0 == strcmp(name, "Cstate_table_set") ||
 535                         0 == strcmp(name, "pmKextRegister") ||
 536                         0 == strcmp(name, "pmSafeMode") ||
 537                         0 == strcmp(name, "pmUnregister") ||
 538                         strstr(name, "pms") == name ||
 539                         0 == strcmp(name, "power_management_init") ||
 540                         strstr(name, "usimple_") == name ||
 541
 542                         strstr(name, "rtc_") == name ||
 543                         strstr(name, "_rtc_") == name ||
 544                         strstr(name, "rtclock_") == name ||
 545                         strstr(name, "clock_") == name ||
 546                         strstr(name, "absolutetime_to_") == name ||
 547                         0 == strcmp(name, "setPop") ||
 548                         0 == strcmp(name, "nanoseconds_to_absolutetime") ||
 549                         0 == strcmp(name, "nanotime_to_absolutetime") ||
 550
 551                         strstr(name, "etimer_") == name ||
 552
 553                         strstr(name, "commpage_") == name ||
 554                         strstr(name, "pmap_") == name ||
 555                         strstr(name, "ml_") == name ||
 556                         strstr(name, "PE_") == name ||
 557                         strstr(name, "lapic_") == name ||
 558                         strstr(name, "acpi_") == name)
 559                         continue;
 560
 561         /*
 562          * Avoid machine_ routines. PR_5346750.
 563          */
 564         if (strstr(name, "machine_") == name)
 565             continue;
 566
 567                 if (0 == strcmp(name, "handle_pending_TLB_flushes"))
 568                         continue;
 569
 570         /*
 571          * Place no probes on critical routines. PR_5221096
 572          */
 573         if (!gIgnoreFBTBlacklist &&
 574             bsearch( name, critical_blacklist, CRITICAL_BLACKLIST_COUNT, sizeof(name), _cmp ) != NULL)
 575                 continue;
 576
 577         /*
 578                  * Place no probes that could be hit in probe context.
 579                  */
 580                 if (!gIgnoreFBTBlacklist &&
 581                         bsearch( name, probe_ctx_closure, PROBE_CTX_CLOSURE_COUNT, sizeof(name), _cmp ) != NULL)
 582                         continue;
 583
 584                 /*
 585                  * Place no probes that could be hit on the way to the debugger.
 586                  */
 587                 if (strstr(name, "kdp_") == name ||
 588                         strstr(name, "kdb_") == name ||
 589                         strstr(name, "kdbg_") == name ||
 590                         strstr(name, "kdebug_") == name ||
 591                         0 == strcmp(name, "kernel_debug") ||
 592                         0 == strcmp(name, "Debugger") ||
 593                         0 == strcmp(name, "Call_DebuggerC") ||
 594                         0 == strcmp(name, "lock_debugger") ||
 595                         0 == strcmp(name, "unlock_debugger") ||
 596                         0 == strcmp(name, "SysChoked"))
 597                         continue;
 598
 599                 /*
 600                  * Place no probes that could be hit on the way to a panic.
 601                  */
 602                 if (NULL != strstr(name, "panic_") ||
 603                         0 == strcmp(name, "panic") ||
 604                         0 == strcmp(name, "handleMck") ||
 605                         0 == strcmp(name, "unresolved_kernel_trap"))
 606                         continue;
 607
 608                 if (dtrace_probe_lookup(fbt_id, modname, name, NULL) != 0)
 609                         continue;
 610
 611                 for (j = 0, instr = (machine_inst_t *)sym[i].n_value, theInstr = 0;
 612                          (j < 4) && ((uintptr_t)instr >= instrLow) && (instrHigh > (uintptr_t)(instr + 2));
 613                          j++) {
 614                         theInstr = instr[0];
 615                         if (theInstr == FBT_PUSHL_EBP || theInstr == FBT_RET || theInstr == FBT_RET_IMM16)
 616                                 break;
 617
 618                         if ((size = dtrace_instr_size(instr)) <= 0)
 619                                 break;
 620
 621                         instr += size;
 622                 }
 623
 624                 if (theInstr != FBT_PUSHL_EBP)
 625                         continue;
 626
 627                 i1 = instr[1];
 628                 i2 = instr[2];
 629
 630                 limit = (machine_inst_t *)instrHigh;
 631
 632                 if ((i1 == FBT_MOVL_ESP_EBP0_V0 && i2 == FBT_MOVL_ESP_EBP1_V0) ||
 633                         (i1 == FBT_MOVL_ESP_EBP0_V1 && i2 == FBT_MOVL_ESP_EBP1_V1)) {
 634                                 instr += 1; /* Advance to the movl %esp,%ebp */
 635                                 theInstr = i1;
 636                 } else {
 637                         /*
 638                          * Sometimes, the compiler will schedule an intervening instruction
 639                          * in the function prologue. Example:
 640                          *
 641                          * _mach_vm_read:
 642                          * 000006d8        pushl   %ebp
 643                          * 000006d9        movl    $0x00000004,%edx
 644                          * 000006de        movl    %esp,%ebp
 645                          *
 646                          * Try the next instruction, to see if it is a movl %esp,%ebp
 647                          */
 648
 649                         instr += 1; /* Advance past the pushl %ebp */
 650                         if ((size = dtrace_instr_size(instr)) <= 0)
 651                                 continue;
 652
 653                         instr += size;
 654
 655                         if ((instr + 1) >= limit)
 656                                 continue;
 657
 658                         i1 = instr[0];
 659                         i2 = instr[1];
 660
 661                         if (!(i1 == FBT_MOVL_ESP_EBP0_V0 && i2 == FBT_MOVL_ESP_EBP1_V0) &&
 662                                 !(i1 == FBT_MOVL_ESP_EBP0_V1 && i2 == FBT_MOVL_ESP_EBP1_V1))
 663                                 continue;
 664
 665                         /* instr already points at the movl %esp,%ebp */
 666                         theInstr = i1;
 667                 }
 668
 669                 fbt = kmem_zalloc(sizeof (fbt_probe_t), KM_SLEEP);
 670                 strlcpy( (char *)&(fbt->fbtp_name), name, MAX_FBTP_NAME_CHARS );
 671                 fbt->fbtp_id = dtrace_probe_create(fbt_id, modname, name, FBT_ENTRY, FBT_AFRAMES_ENTRY, fbt);
 672                 fbt->fbtp_patchpoint = instr;
 673                 fbt->fbtp_ctl = ctl;
 674                 fbt->fbtp_loadcnt = ctl->mod_loadcnt;
 675                 fbt->fbtp_rval = DTRACE_INVOP_MOVL_ESP_EBP;
 676                 fbt->fbtp_savedval = theInstr;
 677                 fbt->fbtp_patchval = FBT_PATCHVAL;
 678
 679                 fbt->fbtp_hashnext = fbt_probetab[FBT_ADDR2NDX(instr)];
 680                 fbt->fbtp_symndx = i;
 681                 fbt_probetab[FBT_ADDR2NDX(instr)] = fbt;
 682
 683                 retfbt = NULL;
 684 again:
 685                 if (instr >= limit)
 686                         continue;
 687
 688                 /*
 689                  * If this disassembly fails, then we've likely walked off into
 690                  * a jump table or some other unsuitable area.  Bail out of the
 691                  * disassembly now.
 692                  */
 693                 if ((size = dtrace_instr_size(instr)) <= 0)
 694                         continue;
 695
 696                 /*
 697                  * We (desperately) want to avoid erroneously instrumenting a
 698                  * jump table, especially given that our markers are pretty
 699                  * short:  two bytes on x86, and just one byte on amd64.  To
 700                  * determine if we're looking at a true instruction sequence
 701                  * or an inline jump table that happens to contain the same
 702                  * byte sequences, we resort to some heuristic sleeze:  we
 703                  * treat this instruction as being contained within a pointer,
 704                  * and see if that pointer points to within the body of the
 705                  * function.  If it does, we refuse to instrument it.
 706                  */
 707                 for (j = 0; j < sizeof (uintptr_t); j++) {
 708                         uintptr_t check = (uintptr_t)instr - j;
 709                         uint8_t *ptr;
 710
 711                         if (check < sym[i].n_value)
 712                                 break;
 713
 714                         if (check + sizeof (uintptr_t) > (uintptr_t)limit)
 715                                 continue;
 716
 717                         ptr = *(uint8_t **)check;
 718
 719                         if (ptr >= (uint8_t *)sym[i].n_value && ptr < limit) {
 720                                 instr += size;
 721                                 goto again;
 722                         }
 723                 }
 724
 725                 /*
 726                  * OK, it's an instruction.
 727                  */
 728                 theInstr = instr[0];
 729
 730                 /* Walked onto the start of the next routine? If so, bail out of this function. */
 731                 if (theInstr == FBT_PUSHL_EBP)
 732                         continue;
 733
 734                 if (!(size == 1 && (theInstr == FBT_POPL_EBP || theInstr == FBT_LEAVE))) {
 735                         instr += size;
 736                         goto again;
 737                 }
 738
 739                 /*
 740                  * Found the popl %ebp; or leave.
 741                  */
 742                 machine_inst_t *patch_instr = instr;
 743
 744                 /*
 745                  * Scan forward for a "ret", or "jmp".
 746                  */
 747                 instr += size;
 748                 if (instr >= limit)
 749                         continue;
 750
 751                 size = dtrace_instr_size(instr);
 752                 if (size <= 0) /* Failed instruction decode? */
 753                         continue;
 754
 755                 theInstr = instr[0];
 756
 757                 if (!(size == FBT_RET_LEN && (theInstr == FBT_RET)) &&
 758                         !(size == FBT_RET_IMM16_LEN && (theInstr == FBT_RET_IMM16)) &&
 759                         !(size == FBT_JMP_SHORT_REL_LEN && (theInstr == FBT_JMP_SHORT_REL)) &&
 760                         !(size == FBT_JMP_NEAR_REL_LEN && (theInstr == FBT_JMP_NEAR_REL)) &&
 761                         !(size == FBT_JMP_FAR_ABS_LEN && (theInstr == FBT_JMP_FAR_ABS)))
 762                         continue;
 763
 764                 /*
 765                  * popl %ebp; ret; or leave; ret; or leave; jmp tailCalledFun; -- We have a winner!
 766                  */
 767                 fbt = kmem_zalloc(sizeof (fbt_probe_t), KM_SLEEP);
 768                 strlcpy( (char *)&(fbt->fbtp_name), name, MAX_FBTP_NAME_CHARS );
 769
 770                 if (retfbt == NULL) {
 771                         fbt->fbtp_id = dtrace_probe_create(fbt_id, modname,
 772                             name, FBT_RETURN, FBT_AFRAMES_RETURN, fbt);
 773                 } else {
 774                         retfbt->fbtp_next = fbt;
 775                         fbt->fbtp_id = retfbt->fbtp_id;
 776                 }
 777
 778                 retfbt = fbt;
 779                 fbt->fbtp_patchpoint = patch_instr;
 780                 fbt->fbtp_ctl = ctl;
 781                 fbt->fbtp_loadcnt = ctl->mod_loadcnt;
 782
 783                 if (*patch_instr == FBT_POPL_EBP) {
 784                         fbt->fbtp_rval = DTRACE_INVOP_POPL_EBP;
 785                 } else {
 786                         ASSERT(*patch_instr == FBT_LEAVE);
 787                         fbt->fbtp_rval = DTRACE_INVOP_LEAVE;
 788                 }
 789                 fbt->fbtp_roffset =
 790                     (uintptr_t)(patch_instr - (uint8_t *)sym[i].n_value);
 791
 792                 fbt->fbtp_savedval = *patch_instr;
 793                 fbt->fbtp_patchval = FBT_PATCHVAL;
 794                 fbt->fbtp_hashnext = fbt_probetab[FBT_ADDR2NDX(patch_instr)];
 795                 fbt->fbtp_symndx = i;
 796                 fbt_probetab[FBT_ADDR2NDX(patch_instr)] = fbt;
 797
 798                 instr += size;
 799                 goto again;
 800         }
 801 }
 802
 803 extern struct modctl g_fbt_kernctl;
 804 #undef kmem_alloc /* from its binding to dt_kmem_alloc glue */
 805 #undef kmem_free /* from its binding to dt_kmem_free glue */
 806 #include <vm/vm_kern.h>
 807
 808 /*ARGSUSED*/
 809 void
 810 fbt_provide_module(void *arg, struct modctl *ctl)
 811 {
 812 #pragma unused(ctl)
 813         __fbt_provide_module(arg, &g_fbt_kernctl);
 814
 815         kmem_free(kernel_map, (vm_offset_t)g_fbt_kernctl.address, round_page_32(g_fbt_kernctl.size));
 816         g_fbt_kernctl.address = 0;
 817         g_fbt_kernctl.size = 0;
 818 }