src/runtime/x86-64-darwin-os.c

   1 #ifdef LISP_FEATURE_SB_THREAD
   2 #include <mach/mach_init.h>
   3 #endif
   4
   5 #include "thread.h"
   6 #include "validate.h"
   7 #include "runtime.h"
   8 #include "interrupt.h"
   9 #include "x86-64-darwin-os.h"
  10 #include "genesis/fdefn.h"
  11
  12 #include <mach/mach.h>
  13 #include <mach/mach_error.h>
  14 #include <mach/mach_types.h>
  15 #include <mach/sync_policy.h>
  16 #include <mach/machine/thread_state.h>
  17 #include <mach/machine/thread_status.h>
  18 #include <sys/_types.h>
  19 #include <sys/ucontext.h>
  20 #include <pthread.h>
  21 #include <assert.h>
  22 #include <stdlib.h>
  23 #include <stdio.h>
  24
  25 #if __DARWIN_UNIX03
  26 #include <sys/_structs.h>
  27 #endif
  28
  29 #if __DARWIN_UNIX03
  30
  31 typedef struct __darwin_ucontext darwin_ucontext;
  32 typedef struct __darwin_mcontext64 darwin_mcontext;
  33
  34 #define rip __rip
  35 #define rsp __rsp
  36 #define rbp __rbp
  37 #define rax __rax
  38 #define rbx __rbx
  39 #define rcx __rcx
  40 #define rdx __rdx
  41 #define rsi __rsi
  42 #define rdi __rdi
  43 #define r8 __r8
  44 #define r9 __r9
  45 #define faultvaddr __faultvaddr
  46 #define ss __ss
  47 #define es __es
  48 #define fs __fs
  49
  50 #define fpu_fcw __fpu_fcw
  51 #define fpu_mxcsr __fpu_mxcsr
  52
  53 #else
  54
  55 typedef struct ucontext darwin_ucontext;
  56 typedef struct mcontext darwin_mcontext;
  57
  58 #endif
  59
  60 #ifdef LISP_FEATURE_SB_THREAD
  61 pthread_mutex_t mach_exception_lock = PTHREAD_MUTEX_INITIALIZER;
  62 #endif
  63
  64 #ifdef LISP_FEATURE_MACH_EXCEPTION_HANDLER
  65
  66 void sigill_handler(int signal, siginfo_t *siginfo, os_context_t *context);
  67 void sigtrap_handler(int signal, siginfo_t *siginfo, os_context_t *context);
  68 void memory_fault_handler(int signal, siginfo_t *siginfo,
  69                           os_context_t *context);
  70
  71 /* This executes in the faulting thread as part of the signal
  72  * emulation.  It is passed a context with the uc_mcontext field
  73  * pointing to a valid block of memory. */
  74 void build_fake_signal_context(darwin_ucontext *context,
  75                                x86_thread_state64_t *thread_state,
  76                                x86_float_state64_t *float_state) {
  77     pthread_sigmask(0, NULL, &context->uc_sigmask);
  78     context->uc_mcontext->ss = *thread_state;
  79     context->uc_mcontext->fs = *float_state;
  80 }
  81
  82 /* This executes in the faulting thread as part of the signal
  83  * emulation.  It is effectively the inverse operation from above. */
  84 void update_thread_state_from_context(x86_thread_state64_t *thread_state,
  85                                       x86_float_state64_t *float_state,
  86                                       darwin_ucontext  *context) {
  87     *thread_state = context->uc_mcontext->ss;
  88     *float_state = context->uc_mcontext->fs;
  89     pthread_sigmask(SIG_SETMASK, &context->uc_sigmask, NULL);
  90 }
  91
  92 /* Modify a context to push new data on its stack. */
  93 void push_context(u64 data, x86_thread_state64_t *context)
  94 {
  95     u64 *stack_pointer;
  96
  97     stack_pointer = (u64*) context->rsp;
  98     *(--stack_pointer) = data;
  99     context->rsp = (u64) stack_pointer;
 100 }
 101
 102 void align_context_stack(x86_thread_state64_t *context)
 103 {
 104     /* 16byte align the stack (provided that the stack is, as it
 105      * should be, 8byte aligned. */
 106     while (context->rsp & 15) push_context(0, context);
 107 }
 108
 109 /* Stack allocation starts with a context that has a mod-4 ESP value
 110  * and needs to leave a context with a mod-16 ESP that will restore
 111  * the old ESP value and other register state when activated.  The
 112  * first part of this is the recovery trampoline, which loads ESP from
 113  * EBP, pops EBP, and returns. */
 114 asm(".globl _stack_allocation_recover; \
 115     .align 4; \
 116  _stack_allocation_recover: \
 117     lea -48(%rbp), %rsp; \
 118     pop %rsi; \
 119     pop %rdi; \
 120     pop %rdx; \
 121     pop %rcx; \
 122     pop %r8; \
 123     pop %r9; \
 124     pop %rbp; \
 125     ret;");
 126
 127 void open_stack_allocation(x86_thread_state64_t *context)
 128 {
 129     void stack_allocation_recover(void);
 130
 131     push_context(context->rip, context);
 132     push_context(context->rbp, context);
 133     context->rbp = context->rsp;
 134
 135     push_context(context->r9, context);
 136     push_context(context->r8, context);
 137     push_context(context->rcx, context);
 138     push_context(context->rdx, context);
 139     push_context(context->rsi, context);
 140     push_context(context->rdi, context);
 141
 142     context->rip = (u64) stack_allocation_recover;
 143
 144     align_context_stack(context);
 145 }
 146
 147 /* Stack allocation of data starts with a context with a mod-16 ESP
 148  * value and reserves some space on it by manipulating the ESP
 149  * register. */
 150 void *stack_allocate(x86_thread_state64_t *context, size_t size)
 151 {
 152     /* round up size to 16byte multiple */
 153     size = (size + 15) & -16;
 154
 155     context->rsp = ((u64)context->rsp) - size;
 156
 157     return (void *)context->rsp;
 158 }
 159
 160 /* Arranging to invoke a C function is tricky, as we have to assume
 161  * cdecl calling conventions (caller removes args) and x86/darwin
 162  * alignment requirements.  The simplest way to arrange this,
 163  * actually, is to open a new stack allocation.
 164  * WARNING!!! THIS DOES NOT PRESERVE REGISTERS! */
 165 void call_c_function_in_context(x86_thread_state64_t *context,
 166                                 void *function,
 167                                 int nargs,
 168                                 ...)
 169 {
 170     va_list ap;
 171     int i;
 172     u64 *stack_pointer;
 173
 174     /* Set up to restore stack on exit. */
 175     open_stack_allocation(context);
 176
 177     /* Have to keep stack 16byte aligned on x86/darwin. */
 178     for (i = (1 & -nargs); i; i--) {
 179         push_context(0, context);
 180     }
 181
 182     context->rsp = ((u64)context->rsp) - nargs * 8;
 183     stack_pointer = (u64 *)context->rsp;
 184
 185     va_start(ap, nargs);
 186     if (nargs > 0) context->rdi = va_arg(ap, u64);
 187     if (nargs > 1) context->rsi = va_arg(ap, u64);
 188     if (nargs > 2) context->rdx = va_arg(ap, u64);
 189     if (nargs > 3) context->rcx = va_arg(ap, u64);
 190     if (nargs > 4) context->r8 = va_arg(ap, u64);
 191     if (nargs > 5) context->r9 = va_arg(ap, u64);
 192     for (i = 6; i < nargs; i++) {
 193         stack_pointer[i] = va_arg(ap, u64);
 194     }
 195     va_end(ap);
 196
 197     push_context(context->rip, context);
 198     context->rip = (u64) function;
 199 }
 200
 201 void signal_emulation_wrapper(x86_thread_state64_t *thread_state,
 202                               x86_float_state64_t *float_state,
 203                               int signal,
 204                               siginfo_t *siginfo,
 205                               void (*handler)(int, siginfo_t *, void *))
 206 {
 207
 208     /* CLH: FIXME **NOTE: HACK ALERT!** Ideally, we would allocate
 209      * context and regs on the stack as local variables, but this
 210      * causes problems for the lisp debugger. When it walks the stack
 211      * for a back trace, it sees the 1) address of the local variable
 212      * on the stack and thinks that is a frame pointer to a lisp
 213      * frame, and, 2) the address of the sap that we alloc'ed in
 214      * dynamic space and thinks that is a return address, so it,
 215      * heuristicly (and wrongly), chooses that this should be
 216      * interpreted as a lisp frame instead of as a C frame.
 217      * We can work around this in this case by os_validating the
 218      * context (and regs just for symmetry).
 219      */
 220
 221     darwin_ucontext  *context;
 222     darwin_mcontext *regs;
 223
 224     context = (darwin_ucontext *) os_validate(0, sizeof(darwin_ucontext));
 225     regs = (darwin_mcontext*) os_validate(0, sizeof(darwin_mcontext));
 226     context->uc_mcontext = regs;
 227
 228     /* when BSD signals are fired, they mask they signals in sa_mask
 229        which always seem to be the blockable_sigset, for us, so we
 230        need to:
 231        1) save the current sigmask
 232        2) block blockable signals
 233        3) call the signal handler
 234        4) restore the sigmask */
 235
 236     build_fake_signal_context(context, thread_state, float_state);
 237
 238     block_blockable_signals(0, 0);
 239
 240     handler(signal, siginfo, context);
 241
 242     update_thread_state_from_context(thread_state, float_state, context);
 243
 244     os_invalidate((os_vm_address_t)context, sizeof(darwin_ucontext));
 245     os_invalidate((os_vm_address_t)regs, sizeof(darwin_mcontext));
 246
 247     /* Trap to restore the signal context. */
 248     asm volatile (".quad 0xffffffffffff0b0f"
 249                   : : "a" (thread_state), "b" (float_state));
 250 }
 251
 252 #if defined DUMP_CONTEXT
 253 void dump_context(x86_thread_state64_t *context)
 254 {
 255     int i;
 256     u64 *stack_pointer;
 257
 258     printf("rax: %08lx  rcx: %08lx  rdx: %08lx  rbx: %08lx\n",
 259            context->rax, context->rcx, context->rdx, context->rbx);
 260     printf("rsp: %08lx  rbp: %08lx  rsi: %08lx  rdi: %08lx\n",
 261            context->rsp, context->rbp, context->rsi, context->rdi);
 262     printf("rip: %08lx  eflags: %08lx\n",
 263            context->rip, context->rflags);
 264     printf("cs: %04hx  ds: %04hx  es: %04hx  "
 265            "ss: %04hx  fs: %04hx  gs: %04hx\n",
 266            context->cs, context->ds, context->rs,
 267            context->ss, context->fs, context->gs);
 268
 269     stack_pointer = (u64 *)context->rsp;
 270     for (i = 0; i < 48; i+=4) {
 271         printf("%08x:  %08x %08x %08x %08x\n",
 272                context->rsp + (i * 4),
 273                stack_pointer[i],
 274                stack_pointer[i+1],
 275                stack_pointer[i+2],
 276                stack_pointer[i+3]);
 277     }
 278 }
 279 #endif
 280
 281 void
 282 control_stack_exhausted_handler(int signal, siginfo_t *siginfo,
 283                                 os_context_t *context) {
 284     unblock_signals_in_context_and_maybe_warn(context);
 285     arrange_return_to_lisp_function
 286         (context, StaticSymbolFunction(CONTROL_STACK_EXHAUSTED_ERROR));
 287 }
 288
 289 void
 290 undefined_alien_handler(int signal, siginfo_t *siginfo, os_context_t *context) {
 291     arrange_return_to_lisp_function
 292         (context, StaticSymbolFunction(UNDEFINED_ALIEN_VARIABLE_ERROR));
 293 }
 294
 295 kern_return_t
 296 catch_exception_raise(mach_port_t exception_port,
 297                       mach_port_t thread,
 298                       mach_port_t task,
 299                       exception_type_t exception,
 300                       exception_data_t code_vector,
 301                       mach_msg_type_number_t code_count)
 302 {
 303     kern_return_t ret, dealloc_ret;
 304     int signal;
 305     siginfo_t* siginfo;
 306
 307 #ifdef LISP_FEATURE_SB_THREAD
 308     thread_mutex_lock(&mach_exception_lock);
 309 #endif
 310
 311     x86_thread_state64_t thread_state;
 312     mach_msg_type_number_t thread_state_count = x86_THREAD_STATE64_COUNT;
 313
 314     x86_float_state64_t float_state;
 315     mach_msg_type_number_t float_state_count = x86_FLOAT_STATE64_COUNT;
 316
 317     x86_exception_state64_t exception_state;
 318     mach_msg_type_number_t exception_state_count = x86_EXCEPTION_STATE64_COUNT;
 319
 320     x86_thread_state64_t backup_thread_state;
 321     x86_thread_state64_t *target_thread_state;
 322     x86_float_state64_t *target_float_state;
 323
 324     os_vm_address_t addr;
 325
 326     struct thread *th;
 327
 328     FSHOW((stderr,"/entering catch_exception_raise with exception: %d\n", exception));
 329     th = *(struct thread**)exception_port;
 330
 331     switch (exception) {
 332
 333     case EXC_BAD_ACCESS:
 334         signal = SIGBUS;
 335         ret = thread_get_state(thread,
 336                                x86_THREAD_STATE64,
 337                                (thread_state_t)&thread_state,
 338                                &thread_state_count);
 339         ret = thread_get_state(thread,
 340                                x86_FLOAT_STATE64,
 341                                (thread_state_t)&float_state,
 342                                &float_state_count);
 343         ret = thread_get_state(thread,
 344                                x86_EXCEPTION_STATE64,
 345                                (thread_state_t)&exception_state,
 346                                &exception_state_count);
 347         addr = (void*)exception_state.faultvaddr;
 348         /* note the os_context hackery here.  When the signal handler returns,
 349          * it won't go back to what it was doing ... */
 350         if(addr >= CONTROL_STACK_GUARD_PAGE(th) &&
 351            addr < CONTROL_STACK_GUARD_PAGE(th) + os_vm_page_size) {
 352             /* We hit the end of the control stack: disable guard page
 353              * protection so the error handler has some headroom, protect the
 354              * previous page so that we can catch returns from the guard page
 355              * and restore it. */
 356             lower_thread_control_stack_guard_page(th);
 357
 358             backup_thread_state = thread_state;
 359             open_stack_allocation(&thread_state);
 360             /* Reserve a 256 byte zone for signal handlers
 361              * to use on the interrupted thread stack.
 362              */
 363             stack_allocate(&thread_state, 256);
 364
 365             /* Save thread state */
 366             target_thread_state =
 367                 stack_allocate(&thread_state, sizeof(*target_thread_state));
 368             (*target_thread_state) = backup_thread_state;
 369
 370             /* Save float state */
 371             target_float_state =
 372                 stack_allocate(&thread_state, sizeof(*target_float_state));
 373             (*target_float_state) = float_state;
 374
 375             /* Set up siginfo */
 376             siginfo = stack_allocate(&thread_state, sizeof(*siginfo));
 377             /* what do we need to put in our fake siginfo?  It looks like
 378              * the x86 code only uses si_signo and si_adrr. */
 379             siginfo->si_signo = signal;
 380             siginfo->si_addr = (void*)exception_state.faultvaddr;
 381
 382             call_c_function_in_context(&thread_state,
 383                                        signal_emulation_wrapper,
 384                                        5,
 385                                        target_thread_state,
 386                                        target_float_state,
 387                                        signal,
 388                                        siginfo,
 389                                        control_stack_exhausted_handler);
 390         }
 391         else if(addr >= CONTROL_STACK_RETURN_GUARD_PAGE(th) &&
 392                 addr < CONTROL_STACK_RETURN_GUARD_PAGE(th) + os_vm_page_size) {
 393             /* We're returning from the guard page: reprotect it, and
 394              * unprotect this one. This works even if we somehow missed
 395              * the return-guard-page, and hit it on our way to new
 396              * exhaustion instead. */
 397             reset_thread_control_stack_guard_page(th);
 398         }
 399         else if (addr >= undefined_alien_address &&
 400                  addr < undefined_alien_address + os_vm_page_size) {
 401             backup_thread_state = thread_state;
 402             open_stack_allocation(&thread_state);
 403             stack_allocate(&thread_state, 256);
 404
 405             /* Save thread state */
 406             target_thread_state =
 407                 stack_allocate(&thread_state, sizeof(*target_thread_state));
 408             (*target_thread_state) = backup_thread_state;
 409
 410             target_float_state =
 411                 stack_allocate(&thread_state, sizeof(*target_float_state));
 412             (*target_float_state) = float_state;
 413
 414             /* Set up siginfo */
 415             siginfo = stack_allocate(&thread_state, sizeof(*siginfo));
 416             /* what do we need to put in our fake siginfo?  It looks like
 417              * the x86 code only uses si_signo and si_adrr. */
 418             siginfo->si_signo = signal;
 419             siginfo->si_addr = (void*)exception_state.faultvaddr;
 420
 421             call_c_function_in_context(&thread_state,
 422                                        signal_emulation_wrapper,
 423                                        5,
 424                                        target_thread_state,
 425                                        target_float_state,
 426                                        signal,
 427                                        siginfo,
 428                                        undefined_alien_handler);
 429         } else {
 430
 431             backup_thread_state = thread_state;
 432             open_stack_allocation(&thread_state);
 433             stack_allocate(&thread_state, 256);
 434
 435             /* Save thread state */
 436             target_thread_state =
 437                 stack_allocate(&thread_state, sizeof(*target_thread_state));
 438             (*target_thread_state) = backup_thread_state;
 439
 440             target_float_state =
 441                 stack_allocate(&thread_state, sizeof(*target_float_state));
 442             (*target_float_state) = float_state;
 443
 444             /* Set up siginfo */
 445             siginfo = stack_allocate(&thread_state, sizeof(*siginfo));
 446             /* what do we need to put in our fake siginfo?  It looks like
 447              * the x86 code only uses si_signo and si_adrr. */
 448             siginfo->si_signo = signal;
 449             siginfo->si_addr = (void*)exception_state.faultvaddr;
 450
 451             call_c_function_in_context(&thread_state,
 452                                        signal_emulation_wrapper,
 453                                        5,
 454                                        target_thread_state,
 455                                        target_float_state,
 456                                        signal,
 457                                        siginfo,
 458                                        memory_fault_handler);
 459         }
 460         ret = thread_set_state(thread,
 461                                x86_THREAD_STATE64,
 462                                (thread_state_t)&thread_state,
 463                                thread_state_count);
 464
 465         ret = thread_set_state(thread,
 466                                x86_FLOAT_STATE64,
 467                                (thread_state_t)&float_state,
 468                                float_state_count);
 469 #ifdef LISP_FEATURE_SB_THREAD
 470         thread_mutex_unlock(&mach_exception_lock);
 471 #endif
 472         ret = KERN_SUCCESS;
 473         break;
 474
 475     case EXC_BAD_INSTRUCTION:
 476
 477         ret = thread_get_state(thread,
 478                                x86_THREAD_STATE64,
 479                                (thread_state_t)&thread_state,
 480                                &thread_state_count);
 481         ret = thread_get_state(thread,
 482                                x86_FLOAT_STATE64,
 483                                (thread_state_t)&float_state,
 484                                &float_state_count);
 485         ret = thread_get_state(thread,
 486                                x86_EXCEPTION_STATE64,
 487                                (thread_state_t)&exception_state,
 488                                &exception_state_count);
 489         if (0xffffffffffff0b0f == *((u64 *)thread_state.rip)) {
 490             /* fake sigreturn. */
 491
 492             /* When we get here, thread_state.rax is a pointer to a
 493              * thread_state to restore. */
 494             /* thread_state = *((thread_state_t *)thread_state.rax); */
 495
 496             ret = thread_set_state(thread,
 497                                    x86_THREAD_STATE64,
 498                                    (thread_state_t) thread_state.rax,
 499                                    /* &thread_state, */
 500                                    thread_state_count);
 501
 502             ret = thread_set_state(thread,
 503                                    x86_FLOAT_STATE64,
 504                                    (thread_state_t) thread_state.rbx,
 505                                    /* &thread_state, */
 506                                    float_state_count);
 507         } else {
 508
 509             backup_thread_state = thread_state;
 510             open_stack_allocation(&thread_state);
 511             stack_allocate(&thread_state, 256);
 512
 513             /* Save thread state */
 514             target_thread_state =
 515                 stack_allocate(&thread_state, sizeof(*target_thread_state));
 516             (*target_thread_state) = backup_thread_state;
 517
 518             target_float_state =
 519                 stack_allocate(&thread_state, sizeof(*target_float_state));
 520             (*target_float_state) = float_state;
 521
 522             /* Set up siginfo */
 523             siginfo = stack_allocate(&thread_state, sizeof(*siginfo));
 524             /* what do we need to put in our fake siginfo?  It looks like
 525              * the x86 code only uses si_signo and si_adrr. */
 526             if (*((unsigned short *)target_thread_state->rip) == 0x0b0f) {
 527                 signal = SIGTRAP;
 528                 siginfo->si_signo = signal;
 529                 siginfo->si_addr = (void*)exception_state.faultvaddr;
 530                 target_thread_state->rip += 2;
 531                 call_c_function_in_context(&thread_state,
 532                                            signal_emulation_wrapper,
 533                                            5,
 534                                            target_thread_state,
 535                                            target_float_state,
 536                                            signal,
 537                                            siginfo,
 538                                            sigtrap_handler);
 539             } else {
 540                 signal = SIGILL;
 541                 siginfo->si_signo = signal;
 542                 siginfo->si_addr = (void*)exception_state.faultvaddr;
 543
 544                 call_c_function_in_context(&thread_state,
 545                                            signal_emulation_wrapper,
 546                                            5,
 547                                            target_thread_state,
 548                                            target_float_state,
 549                                            signal,
 550                                            siginfo,
 551                                            sigill_handler);
 552             }
 553             ret = thread_set_state(thread,
 554                                    x86_THREAD_STATE64,
 555                                    (thread_state_t)&thread_state,
 556                                    thread_state_count);
 557             ret = thread_set_state(thread,
 558                                    x86_FLOAT_STATE64,
 559                                    (thread_state_t)&float_state,
 560                                    float_state_count);
 561         }
 562 #ifdef LISP_FEATURE_SB_THREAD
 563         thread_mutex_unlock(&mach_exception_lock);
 564 #endif
 565         ret = KERN_SUCCESS;
 566         break;
 567
 568     default:
 569 #ifdef LISP_FEATURE_SB_THREAD
 570         thread_mutex_unlock(&mach_exception_lock);
 571 #endif
 572         ret = KERN_INVALID_RIGHT;
 573     }
 574
 575     if (current_mach_task == MACH_PORT_NULL)
 576         current_mach_task = mach_task_self();
 577
 578     dealloc_ret = mach_port_deallocate (current_mach_task, thread);
 579     if (dealloc_ret) {
 580       lose("mach_port_deallocate (thread) failed with return_code %d\n", dealloc_ret);
 581     }
 582
 583     dealloc_ret = mach_port_deallocate (current_mach_task, task);
 584     if (dealloc_ret) {
 585       lose("mach_port_deallocate (task) failed with return_code %d\n", dealloc_ret);
 586     }
 587
 588     return ret;
 589 }
 590
 591 void
 592 os_restore_fp_control(os_context_t *context)
 593 {
 594     /* KLUDGE: The x87 FPU control word is some nasty bitfield struct
 595      * thing.  Rather than deal with that, just grab it as a 16-bit
 596      * integer. */
 597     unsigned short fpu_control_word =
 598         *((unsigned short *)&context->uc_mcontext->fs.fpu_fcw);
 599     /* reset exception flags and restore control flags on SSE2 FPU */
 600     unsigned int temp = (context->uc_mcontext->fs.fpu_mxcsr) & ~0x3F;
 601     asm ("ldmxcsr %0" : : "m" (temp));
 602     /* same for x87 FPU. */
 603     asm ("fldcw %0" : : "m" (fpu_control_word));
 604 }
 605
 606 #endif