src/runtime/x86-64-arch.c

   1 /*
   2  * This software is part of the SBCL system. See the README file for
   3  * more information.
   4  *
   5  * This software is derived from the CMU CL system, which was
   6  * written at Carnegie Mellon University and released into the
   7  * public domain. The software is in the public domain and is
   8  * provided with absolutely no warranty. See the COPYING and CREDITS
   9  * files for more information.
  10  */
  11
  12 #include <stdio.h>
  13
  14 #include "sbcl.h"
  15 #include "runtime.h"
  16 #include "globals.h"
  17 #include "validate.h"
  18 #include "os.h"
  19 #include "sbcl.h"
  20 #include "arch.h"
  21 #include "lispregs.h"
  22 #include "signal.h"
  23 #include "alloc.h"
  24 #include "interrupt.h"
  25 #include "interr.h"
  26 #include "breakpoint.h"
  27 #include "thread.h"
  28 #include "pseudo-atomic.h"
  29
  30 #include "genesis/static-symbols.h"
  31 #include "genesis/symbol.h"
  32
  33 #define BREAKPOINT_INST 0xcc    /* INT3 */
  34 #define UD2_INST 0x0b0f         /* UD2 */
  35
  36 #ifndef LISP_FEATURE_UD2_BREAKPOINTS
  37 #define BREAKPOINT_WIDTH 1
  38 #else
  39 #define BREAKPOINT_WIDTH 2
  40 #endif
  41
  42 void arch_init(void)
  43 {}
  44
  45 #ifndef _WIN64
  46 os_vm_address_t
  47 arch_get_bad_addr(int sig, siginfo_t *code, os_context_t *context)
  48 {
  49     return (os_vm_address_t)code->si_addr;
  50 }
  51 #endif
  52
  53 \f
  54 /*
  55  * hacking signal contexts
  56  *
  57  * (This depends both on architecture, which determines what we might
  58  * want to get to, and on OS, which determines how we get to it.)
  59  */
  60
  61 os_context_register_t *
  62 context_eflags_addr(os_context_t *context)
  63 {
  64 #if defined __linux__ || defined __sun
  65     /* KLUDGE: As of kernel 2.2.14 on Red Hat 6.2, there's code in the
  66      * <sys/ucontext.h> file to define symbolic names for offsets into
  67      * gregs[], but it's conditional on __USE_GNU and not defined, so
  68      * we need to do this nasty absolute index magic number thing
  69      * instead. */
  70     return &context->uc_mcontext.gregs[17];
  71 #elif defined __FreeBSD__
  72     return &context->uc_mcontext.mc_rflags;
  73 #elif defined LISP_FEATURE_DARWIN
  74     return CONTEXT_ADDR_FROM_STEM(rflags);
  75 #elif defined __OpenBSD__
  76     return &context->sc_rflags;
  77 #elif defined __NetBSD__
  78     return CONTEXT_ADDR_FROM_STEM(RFLAGS);
  79 #elif defined _WIN64
  80     return (os_context_register_t*)&context->win32_context->EFlags;
  81 #else
  82 #error unsupported OS
  83 #endif
  84 }
  85 \f
  86 void arch_skip_instruction(os_context_t *context)
  87 {
  88     /* Assuming we get here via an INT3 xxx instruction, the PC now
  89      * points to the interrupt code (a Lisp value) so we just move
  90      * past it. Skip the code; after that, if the code is an
  91      * error-trap or cerror-trap then skip the data bytes that follow. */
  92
  93     int vlen;
  94     long code;
  95
  96
  97     /* Get and skip the Lisp interrupt code. */
  98     code = *(char*)(*os_context_pc_addr(context))++;
  99     switch (code)
 100         {
 101         case trap_Error:
 102         case trap_Cerror:
 103             /* Lisp error arg vector length */
 104             vlen = *(char*)(*os_context_pc_addr(context))++;
 105             /* Skip Lisp error arg data bytes. */
 106             while (vlen-- > 0) {
 107                 ++*os_context_pc_addr(context);
 108             }
 109             break;
 110
 111         case trap_Breakpoint:           /* not tested */
 112         case trap_FunEndBreakpoint: /* not tested */
 113             break;
 114
 115 #ifdef LISP_FEATURE_SB_SAFEPOINT
 116         case trap_GlobalSafepoint:
 117         case trap_CspSafepoint:
 118 #endif
 119         case trap_PendingInterrupt:
 120         case trap_Halt:
 121         case trap_SingleStepAround:
 122         case trap_SingleStepBefore:
 123             /* only needed to skip the Code */
 124             break;
 125
 126         default:
 127             fprintf(stderr,"[arch_skip_inst invalid code %ld\n]\n",code);
 128             break;
 129         }
 130
 131     FSHOW((stderr,
 132            "/[arch_skip_inst resuming at %x]\n",
 133            *os_context_pc_addr(context)));
 134 }
 135
 136 unsigned char *
 137 arch_internal_error_arguments(os_context_t *context)
 138 {
 139     return 1 + (unsigned char *)(*os_context_pc_addr(context));
 140 }
 141
 142 boolean
 143 arch_pseudo_atomic_atomic(os_context_t *context)
 144 {
 145     return get_pseudo_atomic_atomic(arch_os_get_current_thread());
 146 }
 147
 148 void
 149 arch_set_pseudo_atomic_interrupted(os_context_t *context)
 150 {
 151     struct thread *thread = arch_os_get_current_thread();
 152     set_pseudo_atomic_interrupted(thread);
 153 }
 154
 155 void
 156 arch_clear_pseudo_atomic_interrupted(os_context_t *context)
 157 {
 158     struct thread *thread = arch_os_get_current_thread();
 159     clear_pseudo_atomic_interrupted(thread);
 160 }
 161 \f
 162 /*
 163  * This stuff seems to get called for TRACE and debug activity.
 164  */
 165
 166 unsigned int
 167 arch_install_breakpoint(void *pc)
 168 {
 169     unsigned int result = *(unsigned int*)pc;
 170
 171 #ifndef LISP_FEATURE_UD2_BREAKPOINTS
 172     *(char*)pc = BREAKPOINT_INST;               /* x86 INT3       */
 173     *((char*)pc+1) = trap_Breakpoint;           /* Lisp trap code */
 174 #else
 175     *(char*)pc = UD2_INST & 0xff;
 176     *((char*)pc+1) = UD2_INST >> 8;
 177     *((char*)pc+2) = trap_Breakpoint;
 178 #endif
 179
 180     return result;
 181 }
 182
 183 void
 184 arch_remove_breakpoint(void *pc, unsigned int orig_inst)
 185 {
 186     *((char *)pc) = orig_inst & 0xff;
 187     *((char *)pc + 1) = (orig_inst & 0xff00) >> 8;
 188 #if BREAKPOINT_WIDTH > 1
 189     *((char *)pc + 2) = (orig_inst & 0xff0000) >> 16;
 190 #endif
 191 }
 192 \f
 193 /* When single stepping, single_stepping holds the original instruction
 194  * PC location. */
 195 unsigned int *single_stepping = NULL;
 196 #ifdef CANNOT_GET_TO_SINGLE_STEP_FLAG
 197 unsigned int  single_step_save1;
 198 unsigned int  single_step_save2;
 199 unsigned int  single_step_save3;
 200 #endif
 201
 202 void
 203 arch_do_displaced_inst(os_context_t *context, unsigned int orig_inst)
 204 {
 205     unsigned int *pc = (unsigned int*)(*os_context_pc_addr(context));
 206
 207     /* Put the original instruction back. */
 208     arch_remove_breakpoint(pc, orig_inst);
 209
 210 #ifdef CANNOT_GET_TO_SINGLE_STEP_FLAG
 211     /* Install helper instructions for the single step:
 212      * pushf; or [esp],0x100; popf. */
 213     single_step_save1 = *(pc-3);
 214     single_step_save2 = *(pc-2);
 215     single_step_save3 = *(pc-1);
 216     *(pc-3) = 0x9c909090;
 217     *(pc-2) = 0x00240c81;
 218     *(pc-1) = 0x9d000001;
 219 #else
 220     *context_eflags_addr(context) |= 0x100;
 221 #endif
 222
 223     single_stepping = pc;
 224
 225 #ifdef CANNOT_GET_TO_SINGLE_STEP_FLAG
 226     *os_context_pc_addr(context) = (os_context_register_t)((char *)pc - 9);
 227 #endif
 228 }
 229
 230 void
 231 arch_handle_breakpoint(os_context_t *context)
 232 {
 233     *os_context_pc_addr(context) -= BREAKPOINT_WIDTH;
 234     handle_breakpoint(context);
 235 }
 236
 237 void
 238 arch_handle_fun_end_breakpoint(os_context_t *context)
 239 {
 240     *os_context_pc_addr(context) -= BREAKPOINT_WIDTH;
 241     *os_context_pc_addr(context) =
 242         (uword_t)handle_fun_end_breakpoint(context);
 243 }
 244
 245 void
 246 arch_handle_single_step_trap(os_context_t *context, int trap)
 247 {
 248     arch_skip_instruction(context);
 249     /* On x86-64 the fdefn / function is always in RAX, so we pass
 250      * 0 as the register_offset. */
 251     handle_single_step_trap(context, trap, 0);
 252 }
 253
 254 \f
 255 void
 256 restore_breakpoint_from_single_step(os_context_t * context)
 257 {
 258 #ifdef CANNOT_GET_TO_SINGLE_STEP_FLAG
 259     /* Un-install single step helper instructions. */
 260     *(single_stepping-3) = single_step_save1;
 261     *(single_stepping-2) = single_step_save2;
 262     *(single_stepping-1) = single_step_save3;
 263 #else
 264     *context_eflags_addr(context) &= ~0x100;
 265 #endif
 266     /* Re-install the breakpoint if possible. */
 267     if (((char *)*os_context_pc_addr(context) >
 268          (char *)single_stepping) &&
 269         ((char *)*os_context_pc_addr(context) <=
 270          (char *)single_stepping + BREAKPOINT_WIDTH)) {
 271         fprintf(stderr, "warning: couldn't reinstall breakpoint\n");
 272     } else {
 273         arch_install_breakpoint(single_stepping);
 274     }
 275
 276     single_stepping = NULL;
 277     return;
 278 }
 279
 280 void
 281 sigtrap_handler(int signal, siginfo_t *info, os_context_t *context)
 282 {
 283     unsigned int trap;
 284
 285     if (single_stepping) {
 286         restore_breakpoint_from_single_step(context);
 287         return;
 288     }
 289
 290     /* This is just for info in case the monitor wants to print an
 291      * approximation. */
 292     access_control_stack_pointer(arch_os_get_current_thread()) =
 293         (lispobj *)*os_context_sp_addr(context);
 294
 295     /* On entry %eip points just after the INT3 byte and aims at the
 296      * 'kind' value (eg trap_Cerror). For error-trap and Cerror-trap a
 297      * number of bytes will follow, the first is the length of the byte
 298      * arguments to follow. */
 299     trap = *(unsigned char *)(*os_context_pc_addr(context));
 300
 301     handle_trap(context, trap);
 302 }
 303
 304 void
 305 sigill_handler(int signal, siginfo_t *siginfo, os_context_t *context) {
 306     /* Triggering SIGTRAP using int3 is unreliable on OS X/x86, so
 307      * we need to use illegal instructions for traps.
 308      */
 309 #if defined(LISP_FEATURE_UD2_BREAKPOINTS) && !defined(LISP_FEATURE_MACH_EXCEPTION_HANDLER)
 310     if (*((unsigned short *)*os_context_pc_addr(context)) == UD2_INST) {
 311         *os_context_pc_addr(context) += 2;
 312         return sigtrap_handler(signal, siginfo, context);
 313     }
 314 #endif
 315
 316     fake_foreign_function_call(context);
 317     lose("Unhandled SIGILL.");
 318 }
 319
 320 #ifdef X86_64_SIGFPE_FIXUP
 321 #define MXCSR_IE (0x01)         /* Invalid Operation */
 322 #define MXCSR_DE (0x02)         /* Denormal */
 323 #define MXCSR_ZE (0x04)         /* Devide-by-Zero */
 324 #define MXCSR_OE (0x08)         /* Overflow */
 325 #define MXCSR_UE (0x10)         /* Underflow */
 326 #define MXCSR_PE (0x20)         /* Precision */
 327
 328 static inline int
 329 mxcsr_to_code(unsigned int mxcsr)
 330 {
 331     /* Extract unmasked exception bits. */
 332     mxcsr &= ~(mxcsr >> 7) & 0x3F;
 333
 334     /* This order is defined at "Intel 64 and IA-32 Architectures
 335      * Software Developerfs Manual" Volume 1: "Basic Architecture",
 336      * 4.9.2 "Floating-Point Exception Priority". */
 337     if (mxcsr & MXCSR_IE)
 338         return FPE_FLTINV;
 339     else if (mxcsr & MXCSR_ZE)
 340         return FPE_FLTDIV;
 341     else if (mxcsr & MXCSR_DE)
 342         return FPE_FLTUND;
 343     else if (mxcsr & MXCSR_OE)
 344         return FPE_FLTOVF;
 345     else if (mxcsr & MXCSR_UE)
 346         return FPE_FLTUND;
 347     else if (mxcsr & MXCSR_PE)
 348         return FPE_FLTRES;
 349
 350     return 0;
 351 }
 352
 353 static void
 354 sigfpe_handler(int signal, siginfo_t *siginfo, os_context_t *context)
 355 {
 356     unsigned int *mxcsr = arch_os_context_mxcsr_addr(context);
 357
 358     if (siginfo->si_code == 0) { /* XMM exception */
 359         siginfo->si_code = mxcsr_to_code(*mxcsr);
 360
 361         /* Clear sticky exception flag. */
 362         *mxcsr &= ~0x3F;
 363     }
 364
 365     interrupt_handle_now(signal, siginfo, context);
 366 }
 367 #endif
 368
 369 void
 370 arch_install_interrupt_handlers()
 371 {
 372     SHOW("entering arch_install_interrupt_handlers()");
 373
 374     /* Note: The old CMU CL code here used sigtrap_handler() to handle
 375      * SIGILL as well as SIGTRAP. I couldn't see any reason to do
 376      * things that way. So, I changed to separate handlers when
 377      * debugging a problem on OpenBSD, where SBCL wasn't catching
 378      * SIGILL properly, but was instead letting the process be
 379      * terminated with an "Illegal instruction" output. If this change
 380      * turns out to break something (maybe breakpoint handling on some
 381      * OS I haven't tested on?) and we have to go back to the old CMU
 382      * CL way, I hope there will at least be a comment to explain
 383      * why.. -- WHN 2001-06-07 */
 384 #if !defined(LISP_FEATURE_MACH_EXCEPTION_HANDLER) && !defined(LISP_FEATURE_WIN32)
 385     undoably_install_low_level_interrupt_handler(SIGILL , sigill_handler);
 386     undoably_install_low_level_interrupt_handler(SIGTRAP, sigtrap_handler);
 387 #endif
 388
 389 #if defined(X86_64_SIGFPE_FIXUP) && !defined(LISP_FEATURE_WIN32)
 390     undoably_install_low_level_interrupt_handler(SIGFPE, sigfpe_handler);
 391 #endif
 392
 393     SHOW("returning from arch_install_interrupt_handlers()");
 394 }
 395 \f
 396 #ifdef LISP_FEATURE_LINKAGE_TABLE
 397 /* FIXME: It might be cleaner to generate these from the lisp side of
 398  * things.
 399  */
 400
 401 void
 402 arch_write_linkage_table_jmp(char * reloc, void * fun)
 403 {
 404     uword_t addr = (uword_t) fun;
 405     int i;
 406
 407     *reloc++ = 0xFF; /* Opcode for near jump to absolute reg/mem64. */
 408     *reloc++ = 0x25; /* ModRM #b00 100 101, i.e. RIP-relative. */
 409     *reloc++ = 0x00; /* 32-bit displacement field = 0 */
 410     *reloc++ = 0x00; /* ... */
 411     *reloc++ = 0x00; /* ... */
 412     *reloc++ = 0x00; /* ... */
 413
 414     for (i = 0; i < 8; i++) {
 415         *reloc++ = addr & 0xff;
 416         addr >>= 8;
 417     }
 418
 419     /* write a nop for good measure. */
 420     *reloc = 0x90;
 421 }
 422
 423 void
 424 arch_write_linkage_table_ref(void * reloc, void * data)
 425 {
 426     *(uword_t *)reloc = (uword_t)data;
 427 }
 428
 429 #endif
 430
 431 /* These setup and check *both* the sse2 and x87 FPUs. While lisp code
 432    only uses the sse2 FPU, other code (such as libc) may use the x87 FPU.
 433  */
 434
 435 unsigned int
 436 arch_get_fp_modes()
 437 {
 438     unsigned int temp;
 439     unsigned int result;
 440     /* return the x87 exception flags ored in with the sse2
 441      * control+status flags */
 442     asm ("fnstsw %0" : "=m" (temp));
 443     result = temp;
 444     result &= 0x3F;
 445     asm ("stmxcsr %0" : "=m" (temp));
 446     result |= temp;
 447     /* flip exception mask bits */
 448     return result ^ (0x3F << 7);
 449 }
 450
 451 struct fpenv
 452 {
 453     unsigned short cw;
 454     unsigned short unused1;
 455     unsigned short sw;
 456     unsigned short unused2;
 457     unsigned int other_regs[5];
 458 };
 459
 460 void
 461 arch_set_fp_modes(unsigned int mxcsr)
 462 {
 463     struct fpenv f_env;
 464     unsigned int temp;
 465
 466     /* turn trap enable bits into exception mask */
 467     mxcsr ^= 0x3F << 7;
 468
 469     /* set x87 modes */
 470     asm ("fnstenv %0" : "=m" (f_env));
 471     /* set control word: always long double precision
 472      * get traps and rounding from mxcsr word */
 473     f_env.cw = 0x300 | ((mxcsr >> 7) & 0x3F) | (((mxcsr >> 13) & 0x3) << 10);
 474     /* set status word: only override exception flags, from mxcsr */
 475     f_env.sw &= ~0x3F;
 476     f_env.sw |= (mxcsr & 0x3F);
 477
 478     asm ("fldenv %0" : : "m" (f_env));
 479
 480     /* now, simply, load up the mxcsr register */
 481     temp = mxcsr;
 482     asm ("ldmxcsr %0" : : "m" (temp));
 483 }