libgc/win32_threads.c

   1 #include "private/gc_priv.h"
   2
   3 #if defined(GC_WIN32_THREADS)
   4
   5 #include <windows.h>
   6
   7 #ifdef CYGWIN32
   8 # include <errno.h>
   9
  10  /* Cygwin-specific forward decls */
  11 # undef pthread_create
  12 # undef pthread_sigmask
  13 # undef pthread_join
  14 # undef pthread_detach
  15 # undef dlopen
  16
  17 # define DEBUG_CYGWIN_THREADS 0
  18
  19   void * GC_start_routine(void * arg);
  20   void GC_thread_exit_proc(void *arg);
  21
  22 #endif
  23
  24 /* The type of the first argument to InterlockedExchange.       */
  25 /* Documented to be LONG volatile *, but at least gcc likes     */
  26 /* this better.                                                 */
  27 typedef LONG * IE_t;
  28
  29 #ifndef MAX_THREADS
  30 # define MAX_THREADS 256
  31     /* FIXME:                                                   */
  32     /* Things may get quite slow for large numbers of threads,  */
  33     /* since we look them up with sequential search.            */
  34 #endif
  35
  36 GC_bool GC_thr_initialized = FALSE;
  37
  38 DWORD GC_main_thread = 0;
  39
  40 struct GC_thread_Rep {
  41   LONG in_use; /* Updated without lock. */
  42                         /* We assert that unused        */
  43                         /* entries have invalid ids of  */
  44                         /* zero and zero stack fields.  */
  45   DWORD id;
  46   HANDLE handle;
  47   ptr_t stack_base;     /* The cold end of the stack.   */
  48                         /* 0 ==> entry not valid.       */
  49                         /* !in_use ==> stack_base == 0  */
  50   GC_bool suspended;
  51
  52 # ifdef CYGWIN32
  53     void *status; /* hold exit value until join in case it's a pointer */
  54     pthread_t pthread_id;
  55     short flags;                /* Protected by GC lock.        */
  56 #       define FINISHED 1       /* Thread has exited.   */
  57 #       define DETACHED 2       /* Thread is intended to be detached.   */
  58 # endif
  59 };
  60
  61 typedef volatile struct GC_thread_Rep * GC_thread;
  62
  63 /*
  64  * We generally assume that volatile ==> memory ordering, at least among
  65  * volatiles.
  66  */
  67
  68 volatile GC_bool GC_please_stop = FALSE;
  69
  70 volatile struct GC_thread_Rep thread_table[MAX_THREADS];
  71
  72 volatile LONG GC_max_thread_index = 0; /* Largest index in thread_table */
  73                                        /* that was ever used.           */
  74
  75 extern LONG WINAPI GC_write_fault_handler(struct _EXCEPTION_POINTERS *exc_info);
  76
  77 int GC_thread_is_registered (void)
  78 {
  79 #if defined(GC_DLL) || defined(GC_INSIDE_DLL)
  80         /* Registered by DllMain */
  81         return 1;
  82 #else
  83         /* FIXME: */
  84         return 0;
  85 #endif
  86 }
  87
  88 /*
  89  * This may be called from DllMain, and hence operates under unusual
  90  * constraints.
  91  */
  92 static GC_thread GC_new_thread(void) {
  93   int i;
  94   /* It appears to be unsafe to acquire a lock here, since this */
  95   /* code is apparently not preeemptible on some systems.       */
  96   /* (This is based on complaints, not on Microsoft's official  */
  97   /* documentation, which says this should perform "only simple */
  98   /* initialization tasks".)                                    */
  99   /* Hence we make do with nonblocking synchronization.         */
 100
 101   /* The following should be a noop according to the win32      */
 102   /* documentation.  There is empirical evidence that it        */
 103   /* isn't.             - HB                                    */
 104 # if defined(MPROTECT_VDB)
 105    if (GC_incremental) SetUnhandledExceptionFilter(GC_write_fault_handler);
 106 # endif
 107                 /* cast away volatile qualifier */
 108   for (i = 0; InterlockedExchange((IE_t)&thread_table[i].in_use,1) != 0; i++) {
 109     /* Compare-and-swap would make this cleaner, but that's not         */
 110     /* supported before Windows 98 and NT 4.0.  In Windows 2000,        */
 111     /* InterlockedExchange is supposed to be replaced by                */
 112     /* InterlockedExchangePointer, but that's not really what I         */
 113     /* want here.                                                       */
 114     if (i == MAX_THREADS - 1)
 115       ABORT("too many threads");
 116   }
 117   /* Update GC_max_thread_index if necessary.  The following is safe,   */
 118   /* and unlike CompareExchange-based solutions seems to work on all    */
 119   /* Windows95 and later platforms.                                     */
 120   /* Unfortunately, GC_max_thread_index may be temporarily out of       */
 121   /* bounds, so readers have to compensate.                             */
 122   while (i > GC_max_thread_index) {
 123     InterlockedIncrement((IE_t)&GC_max_thread_index);
 124   }
 125   if (GC_max_thread_index >= MAX_THREADS) {
 126     /* We overshot due to simultaneous increments.      */
 127     /* Setting it to MAX_THREADS-1 is always safe.      */
 128     GC_max_thread_index = MAX_THREADS - 1;
 129   }
 130
 131 # ifdef CYGWIN32
 132     thread_table[i].pthread_id = pthread_self();
 133 # endif
 134   if (!DuplicateHandle(GetCurrentProcess(),
 135                        GetCurrentThread(),
 136                        GetCurrentProcess(),
 137                        (HANDLE*)&thread_table[i].handle,
 138                        0,
 139                        0,
 140                        DUPLICATE_SAME_ACCESS)) {
 141         DWORD last_error = GetLastError();
 142         GC_printf1("Last error code: %lx\n", last_error);
 143         ABORT("DuplicateHandle failed");
 144   }
 145   thread_table[i].stack_base = GC_get_stack_base();
 146   /* Up until this point, GC_push_all_stacks considers this thread      */
 147   /* invalid.                                                           */
 148   if (thread_table[i].stack_base == NULL)
 149     ABORT("Failed to find stack base in GC_new_thread");
 150   /* Up until this point, this entry is viewed as reserved but invalid  */
 151   /* by GC_delete_thread.                                               */
 152   thread_table[i].id = GetCurrentThreadId();
 153   /* If this thread is being created while we are trying to stop        */
 154   /* the world, wait here.  Hopefully this can't happen on any  */
 155   /* systems that don't allow us to block here.                 */
 156   while (GC_please_stop) Sleep(20);
 157   return thread_table + i;
 158 }
 159
 160 /*
 161  * GC_max_thread_index may temporarily be larger than MAX_THREADS.
 162  * To avoid subscript errors, we check on access.
 163  */
 164 #ifdef __GNUC__
 165 __inline__
 166 #endif
 167 LONG GC_get_max_thread_index()
 168 {
 169   LONG my_max = GC_max_thread_index;
 170
 171   if (my_max >= MAX_THREADS) return MAX_THREADS-1;
 172   return my_max;
 173 }
 174
 175 /* This is intended to be lock-free, though that                        */
 176 /* assumes that the CloseHandle becomes visible before the              */
 177 /* in_use assignment.                                                   */
 178 static void GC_delete_gc_thread(GC_thread thr)
 179 {
 180     CloseHandle(thr->handle);
 181       /* cast away volatile qualifier */
 182     thr->stack_base = 0;
 183     thr->id = 0;
 184 #   ifdef CYGWIN32
 185       thr->pthread_id = 0;
 186 #   endif /* CYGWIN32 */
 187     thr->in_use = FALSE;
 188 }
 189
 190 static void GC_delete_thread(DWORD thread_id) {
 191   int i;
 192   LONG my_max = GC_get_max_thread_index();
 193
 194   for (i = 0;
 195        i <= my_max &&
 196        (!thread_table[i].in_use || thread_table[i].id != thread_id);
 197        /* Must still be in_use, since nobody else can store our thread_id. */
 198        i++) {}
 199   if (i > my_max) {
 200     WARN("Removing nonexistent thread %ld\n", (GC_word)thread_id);
 201   } else {
 202     GC_delete_gc_thread(thread_table+i);
 203   }
 204 }
 205
 206
 207 #ifdef CYGWIN32
 208
 209 /* Return a GC_thread corresponding to a given pthread_t.       */
 210 /* Returns 0 if it's not there.                                 */
 211 /* We assume that this is only called for pthread ids that      */
 212 /* have not yet terminated or are still joinable.               */
 213 static GC_thread GC_lookup_thread(pthread_t id)
 214 {
 215   int i;
 216   LONG my_max = GC_get_max_thread_index();
 217
 218   for (i = 0;
 219        i <= my_max &&
 220        (!thread_table[i].in_use || thread_table[i].pthread_id != id
 221         || !thread_table[i].in_use);
 222        /* Must still be in_use, since nobody else can store our thread_id. */
 223        i++);
 224   if (i > my_max) return 0;
 225   return thread_table + i;
 226 }
 227
 228 #endif /* CYGWIN32 */
 229
 230 void GC_push_thread_structures GC_PROTO((void))
 231 {
 232     /* Unlike the other threads implementations, the thread table here  */
 233     /* contains no pointers to the collectable heap.  Thus we have      */
 234     /* no private structures we need to preserve.                       */
 235 # ifdef CYGWIN32
 236   { int i; /* pthreads may keep a pointer in the thread exit value */
 237     LONG my_max = GC_get_max_thread_index();
 238
 239     for (i = 0; i <= my_max; i++)
 240       if (thread_table[i].in_use)
 241         GC_push_all((ptr_t)&(thread_table[i].status),
 242                     (ptr_t)(&(thread_table[i].status)+1));
 243   }
 244 # endif
 245 }
 246
 247 /* Defined in misc.c */
 248 extern CRITICAL_SECTION GC_write_cs;
 249
 250 void GC_stop_world()
 251 {
 252   DWORD thread_id = GetCurrentThreadId();
 253   int i;
 254
 255   if (!GC_thr_initialized) ABORT("GC_stop_world() called before GC_thr_init()");
 256
 257   GC_please_stop = TRUE;
 258 # ifndef CYGWIN32
 259     EnterCriticalSection(&GC_write_cs);
 260 # endif /* !CYGWIN32 */
 261   for (i = 0; i <= GC_get_max_thread_index(); i++)
 262     if (thread_table[i].stack_base != 0
 263         && thread_table[i].id != thread_id) {
 264 #     ifdef MSWINCE
 265         /* SuspendThread will fail if thread is running kernel code */
 266         while (SuspendThread(thread_table[i].handle) == (DWORD)-1)
 267           Sleep(10);
 268 #     else
 269         /* Apparently the Windows 95 GetOpenFileName call creates       */
 270         /* a thread that does not properly get cleaned up, and          */
 271         /* SuspendThread on its descriptor may provoke a crash.         */
 272         /* This reduces the probability of that event, though it still  */
 273         /* appears there's a race here.                                 */
 274         DWORD exitCode;
 275         if (GetExitCodeThread(thread_table[i].handle,&exitCode) &&
 276             exitCode != STILL_ACTIVE) {
 277           thread_table[i].stack_base = 0; /* prevent stack from being pushed */
 278 #         ifndef CYGWIN32
 279             /* this breaks pthread_join on Cygwin, which is guaranteed to  */
 280             /* only see user pthreads                                      */
 281             thread_table[i].in_use = FALSE;
 282             CloseHandle(thread_table[i].handle);
 283 #         endif
 284           continue;
 285         }
 286         if (SuspendThread(thread_table[i].handle) == (DWORD)-1) {
 287           thread_table[i].stack_base = 0; /* prevent stack from being pushed */
 288 #         ifndef CYGWIN32
 289             /* this breaks pthread_join on Cygwin, which is guaranteed to  */
 290             /* only see user pthreads                                      */
 291             thread_table[i].in_use = FALSE;
 292             CloseHandle(thread_table[i].handle);
 293 #         endif
 294         }
 295 #     endif
 296       thread_table[i].suspended = TRUE;
 297     }
 298 # ifndef CYGWIN32
 299     LeaveCriticalSection(&GC_write_cs);
 300 # endif /* !CYGWIN32 */
 301 }
 302
 303 void GC_start_world()
 304 {
 305   DWORD thread_id = GetCurrentThreadId();
 306   int i;
 307   LONG my_max = GC_get_max_thread_index();
 308
 309   for (i = 0; i <= my_max; i++)
 310     if (thread_table[i].stack_base != 0 && thread_table[i].suspended
 311         && thread_table[i].id != thread_id) {
 312       if (ResumeThread(thread_table[i].handle) == (DWORD)-1)
 313         ABORT("ResumeThread failed");
 314       thread_table[i].suspended = FALSE;
 315     }
 316   GC_please_stop = FALSE;
 317 }
 318
 319 # ifdef _MSC_VER
 320 #   pragma warning(disable:4715)
 321 # endif
 322 ptr_t GC_current_stackbottom()
 323 {
 324   DWORD thread_id = GetCurrentThreadId();
 325   int i;
 326   LONG my_max = GC_get_max_thread_index();
 327
 328   for (i = 0; i <= my_max; i++)
 329     if (thread_table[i].stack_base && thread_table[i].id == thread_id)
 330       return thread_table[i].stack_base;
 331   ABORT("no thread table entry for current thread");
 332 }
 333 # ifdef _MSC_VER
 334 #   pragma warning(default:4715)
 335 # endif
 336
 337 # ifdef MSWINCE
 338     /* The VirtualQuery calls below won't work properly on WinCE, but   */
 339     /* since each stack is restricted to an aligned 64K region of       */
 340     /* virtual memory we can just take the next lowest multiple of 64K. */
 341 #   define GC_get_stack_min(s) \
 342         ((ptr_t)(((DWORD)(s) - 1) & 0xFFFF0000))
 343 # else
 344     static ptr_t GC_get_stack_min(ptr_t s)
 345     {
 346         ptr_t bottom;
 347         MEMORY_BASIC_INFORMATION info;
 348         VirtualQuery(s, &info, sizeof(info));
 349         do {
 350             bottom = info.BaseAddress;
 351             VirtualQuery(bottom - 1, &info, sizeof(info));
 352         } while ((info.Protect & PAGE_READWRITE)
 353                  && !(info.Protect & PAGE_GUARD));
 354         return(bottom);
 355     }
 356 # endif
 357
 358 void GC_push_all_stacks()
 359 {
 360   DWORD thread_id = GetCurrentThreadId();
 361   GC_bool found_me = FALSE;
 362   int i;
 363   int dummy;
 364   ptr_t sp, stack_min;
 365   GC_thread thread;
 366   LONG my_max = GC_get_max_thread_index();
 367
 368   for (i = 0; i <= my_max; i++) {
 369     thread = thread_table + i;
 370     if (thread -> in_use && thread -> stack_base) {
 371       if (thread -> id == thread_id) {
 372         sp = (ptr_t) &dummy;
 373         found_me = TRUE;
 374       } else {
 375         CONTEXT context;
 376         context.ContextFlags = CONTEXT_INTEGER|CONTEXT_CONTROL;
 377         if (!GetThreadContext(thread_table[i].handle, &context))
 378           ABORT("GetThreadContext failed");
 379
 380         /* Push all registers that might point into the heap.  Frame    */
 381         /* pointer registers are included in case client code was       */
 382         /* compiled with the 'omit frame pointer' optimisation.         */
 383 #       define PUSH1(reg) GC_push_one((word)context.reg)
 384 #       define PUSH2(r1,r2) PUSH1(r1), PUSH1(r2)
 385 #       define PUSH4(r1,r2,r3,r4) PUSH2(r1,r2), PUSH2(r3,r4)
 386 #       if defined(I386)
 387           PUSH4(Edi,Esi,Ebx,Edx), PUSH2(Ecx,Eax), PUSH1(Ebp);
 388           sp = (ptr_t)context.Esp;
 389 #       elif defined(X86_64)
 390           PUSH4(Rax,Rcx,Rdx,Rbx); PUSH2(Rbp, Rsi); PUSH1(Rdi);
 391           PUSH4(R8, R9, R10, R11); PUSH4(R12, R13, R14, R15);
 392           sp = (ptr_t)context.Rsp;
 393 #       elif defined(ARM32)
 394           PUSH4(R0,R1,R2,R3),PUSH4(R4,R5,R6,R7),PUSH4(R8,R9,R10,R11),PUSH1(R12);
 395           sp = (ptr_t)context.Sp;
 396 #       elif defined(SHx)
 397           PUSH4(R0,R1,R2,R3), PUSH4(R4,R5,R6,R7), PUSH4(R8,R9,R10,R11);
 398           PUSH2(R12,R13), PUSH1(R14);
 399           sp = (ptr_t)context.R15;
 400 #       elif defined(MIPS)
 401           PUSH4(IntAt,IntV0,IntV1,IntA0), PUSH4(IntA1,IntA2,IntA3,IntT0);
 402           PUSH4(IntT1,IntT2,IntT3,IntT4), PUSH4(IntT5,IntT6,IntT7,IntS0);
 403           PUSH4(IntS1,IntS2,IntS3,IntS4), PUSH4(IntS5,IntS6,IntS7,IntT8);
 404           PUSH4(IntT9,IntK0,IntK1,IntS8);
 405           sp = (ptr_t)context.IntSp;
 406 #       elif defined(PPC)
 407           PUSH4(Gpr0, Gpr3, Gpr4, Gpr5),  PUSH4(Gpr6, Gpr7, Gpr8, Gpr9);
 408           PUSH4(Gpr10,Gpr11,Gpr12,Gpr14), PUSH4(Gpr15,Gpr16,Gpr17,Gpr18);
 409           PUSH4(Gpr19,Gpr20,Gpr21,Gpr22), PUSH4(Gpr23,Gpr24,Gpr25,Gpr26);
 410           PUSH4(Gpr27,Gpr28,Gpr29,Gpr30), PUSH1(Gpr31);
 411           sp = (ptr_t)context.Gpr1;
 412 #       elif defined(ALPHA)
 413           PUSH4(IntV0,IntT0,IntT1,IntT2), PUSH4(IntT3,IntT4,IntT5,IntT6);
 414           PUSH4(IntT7,IntS0,IntS1,IntS2), PUSH4(IntS3,IntS4,IntS5,IntFp);
 415           PUSH4(IntA0,IntA1,IntA2,IntA3), PUSH4(IntA4,IntA5,IntT8,IntT9);
 416           PUSH4(IntT10,IntT11,IntT12,IntAt);
 417           sp = (ptr_t)context.IntSp;
 418 #       else
 419 #         error "architecture is not supported"
 420 #       endif
 421       }
 422
 423       stack_min = GC_get_stack_min(thread->stack_base);
 424
 425       if (sp >= stack_min && sp < thread->stack_base)
 426         GC_push_all_stack(sp, thread->stack_base);
 427       else {
 428         WARN("Thread stack pointer 0x%lx out of range, pushing everything\n",
 429              (unsigned long)sp);
 430         GC_push_all_stack(stack_min, thread->stack_base);
 431       }
 432     }
 433   }
 434   if (!found_me) ABORT("Collecting from unknown thread.");
 435 }
 436
 437 void GC_get_next_stack(char *start, char **lo, char **hi)
 438 {
 439     int i;
 440 #   define ADDR_LIMIT (char *)(-1L)
 441     char * current_min = ADDR_LIMIT;
 442     LONG my_max = GC_get_max_thread_index();
 443
 444     for (i = 0; i <= my_max; i++) {
 445         char * s = (char *)thread_table[i].stack_base;
 446
 447         if (0 != s && s > start && s < current_min) {
 448             current_min = s;
 449         }
 450     }
 451     *hi = current_min;
 452     if (current_min == ADDR_LIMIT) {
 453         *lo = ADDR_LIMIT;
 454         return;
 455     }
 456     *lo = GC_get_stack_min(current_min);
 457     if (*lo < start) *lo = start;
 458 }
 459
 460 #if !defined(CYGWIN32)
 461
 462 #if !defined(MSWINCE) && defined(GC_DLL)
 463
 464 /* We register threads from DllMain */
 465
 466 GC_API HANDLE WINAPI GC_CreateThread(
 467     LPSECURITY_ATTRIBUTES lpThreadAttributes,
 468     DWORD dwStackSize, LPTHREAD_START_ROUTINE lpStartAddress,
 469     LPVOID lpParameter, DWORD dwCreationFlags, LPDWORD lpThreadId )
 470 {
 471     return CreateThread(lpThreadAttributes, dwStackSize, lpStartAddress,
 472                         lpParameter, dwCreationFlags, lpThreadId);
 473 }
 474
 475 #else /* defined(MSWINCE) || !defined(GC_DLL))  */
 476
 477 /* We have no DllMain to take care of new threads.  Thus we     */
 478 /* must properly intercept thread creation.                     */
 479
 480 typedef struct {
 481     LPTHREAD_START_ROUTINE start;
 482     LPVOID param;
 483 } thread_args;
 484
 485 static DWORD WINAPI thread_start(LPVOID arg);
 486
 487 GC_API HANDLE WINAPI GC_CreateThread(
 488     LPSECURITY_ATTRIBUTES lpThreadAttributes,
 489     DWORD dwStackSize, LPTHREAD_START_ROUTINE lpStartAddress,
 490     LPVOID lpParameter, DWORD dwCreationFlags, LPDWORD lpThreadId )
 491 {
 492     HANDLE thread_h = NULL;
 493
 494     thread_args *args;
 495
 496     if (!GC_is_initialized) GC_init();
 497                 /* make sure GC is initialized (i.e. main thread is attached) */
 498
 499     args = GC_malloc_uncollectable(sizeof(thread_args));
 500         /* Handed off to and deallocated by child thread.       */
 501     if (0 == args) {
 502         SetLastError(ERROR_NOT_ENOUGH_MEMORY);
 503         return NULL;
 504     }
 505
 506     /* set up thread arguments */
 507         args -> start = lpStartAddress;
 508         args -> param = lpParameter;
 509
 510     thread_h = CreateThread(lpThreadAttributes,
 511                             dwStackSize, thread_start,
 512                             args, dwCreationFlags,
 513                             lpThreadId);
 514
 515     return thread_h;
 516 }
 517
 518 static DWORD WINAPI thread_start(LPVOID arg)
 519 {
 520     DWORD ret = 0;
 521     thread_args *args = (thread_args *)arg;
 522
 523     GC_new_thread();
 524
 525     /* Clear the thread entry even if we exit with an exception.        */
 526     /* This is probably pointless, since an uncaught exception is       */
 527     /* supposed to result in the process being killed.                  */
 528 #ifndef __GNUC__
 529     __try {
 530 #endif /* __GNUC__ */
 531         ret = args->start (args->param);
 532 #ifndef __GNUC__
 533     } __finally {
 534 #endif /* __GNUC__ */
 535         GC_free(args);
 536         GC_delete_thread(GetCurrentThreadId());
 537 #ifndef __GNUC__
 538     }
 539 #endif /* __GNUC__ */
 540
 541     return ret;
 542 }
 543 #endif /* !defined(MSWINCE) && !(defined(__MINGW32__) && !defined(_DLL))  */
 544
 545 #endif /* !CYGWIN32 */
 546
 547 #ifdef MSWINCE
 548
 549 typedef struct {
 550     HINSTANCE hInstance;
 551     HINSTANCE hPrevInstance;
 552     LPWSTR lpCmdLine;
 553     int nShowCmd;
 554 } main_thread_args;
 555
 556 DWORD WINAPI main_thread_start(LPVOID arg);
 557
 558 int WINAPI WinMain(HINSTANCE hInstance, HINSTANCE hPrevInstance,
 559                    LPWSTR lpCmdLine, int nShowCmd)
 560 {
 561     DWORD exit_code = 1;
 562
 563     main_thread_args args = {
 564         hInstance, hPrevInstance, lpCmdLine, nShowCmd
 565     };
 566     HANDLE thread_h;
 567     DWORD thread_id;
 568
 569     /* initialize everything */
 570     GC_init();
 571
 572     /* start the main thread */
 573     thread_h = GC_CreateThread(
 574         NULL, 0, main_thread_start, &args, 0, &thread_id);
 575
 576     if (thread_h != NULL)
 577     {
 578         WaitForSingleObject (thread_h, INFINITE);
 579         GetExitCodeThread (thread_h, &exit_code);
 580         CloseHandle (thread_h);
 581     }
 582
 583     GC_deinit();
 584     DeleteCriticalSection(&GC_allocate_ml);
 585
 586     return (int) exit_code;
 587 }
 588
 589 DWORD WINAPI main_thread_start(LPVOID arg)
 590 {
 591     main_thread_args * args = (main_thread_args *) arg;
 592
 593     return (DWORD) GC_WinMain (args->hInstance, args->hPrevInstance,
 594                                args->lpCmdLine, args->nShowCmd);
 595 }
 596
 597 # else /* !MSWINCE */
 598
 599 /* Called by GC_init() - we hold the allocation lock.   */
 600 void GC_thr_init() {
 601     if (GC_thr_initialized) return;
 602     GC_main_thread = GetCurrentThreadId();
 603     GC_thr_initialized = TRUE;
 604
 605     /* Add the initial thread, so we can stop it.       */
 606     GC_new_thread();
 607 }
 608
 609 #ifdef CYGWIN32
 610
 611 struct start_info {
 612     void *(*start_routine)(void *);
 613     void *arg;
 614     GC_bool detached;
 615 };
 616
 617 int GC_pthread_join(pthread_t pthread_id, void **retval) {
 618     int result;
 619     int i;
 620     GC_thread me;
 621
 622 #   if DEBUG_CYGWIN_THREADS
 623       GC_printf3("thread 0x%x(0x%x) is joining thread 0x%x.\n",
 624                  (int)pthread_self(), GetCurrentThreadId(), (int)pthread_id);
 625 #   endif
 626
 627     /* Thread being joined might not have registered itself yet. */
 628     /* After the join,thread id may have been recycled.          */
 629     /* FIXME: It would be better if this worked more like        */
 630     /* pthread_support.c.                                        */
 631
 632     while ((me = GC_lookup_thread(pthread_id)) == 0) Sleep(10);
 633
 634     result = pthread_join(pthread_id, retval);
 635
 636     GC_delete_gc_thread(me);
 637
 638 #   if DEBUG_CYGWIN_THREADS
 639       GC_printf3("thread 0x%x(0x%x) completed join with thread 0x%x.\n",
 640                  (int)pthread_self(), GetCurrentThreadId(), (int)pthread_id);
 641 #   endif
 642
 643     return result;
 644 }
 645
 646 /* Cygwin-pthreads calls CreateThread internally, but it's not
 647  * easily interceptible by us..
 648  *   so intercept pthread_create instead
 649  */
 650 int
 651 GC_pthread_create(pthread_t *new_thread,
 652                   const pthread_attr_t *attr,
 653                   void *(*start_routine)(void *), void *arg) {
 654     int result;
 655     struct start_info * si;
 656
 657     if (!GC_is_initialized) GC_init();
 658                 /* make sure GC is initialized (i.e. main thread is attached) */
 659
 660     /* This is otherwise saved only in an area mmapped by the thread */
 661     /* library, which isn't visible to the collector.            */
 662     si = GC_malloc_uncollectable(sizeof(struct start_info));
 663     if (0 == si) return(EAGAIN);
 664
 665     si -> start_routine = start_routine;
 666     si -> arg = arg;
 667     if (attr != 0 &&
 668         pthread_attr_getdetachstate(attr, &si->detached)
 669         == PTHREAD_CREATE_DETACHED) {
 670       si->detached = TRUE;
 671     }
 672
 673 #   if DEBUG_CYGWIN_THREADS
 674       GC_printf2("About to create a thread from 0x%x(0x%x)\n",
 675                  (int)pthread_self(), GetCurrentThreadId);
 676 #   endif
 677     result = pthread_create(new_thread, attr, GC_start_routine, si);
 678
 679     if (result) { /* failure */
 680         GC_free(si);
 681     }
 682
 683     return(result);
 684 }
 685
 686 void * GC_start_routine(void * arg)
 687 {
 688     struct start_info * si = arg;
 689     void * result;
 690     void *(*start)(void *);
 691     void *start_arg;
 692     pthread_t pthread_id;
 693     GC_thread me;
 694     GC_bool detached;
 695     int i;
 696
 697 #   if DEBUG_CYGWIN_THREADS
 698       GC_printf2("thread 0x%x(0x%x) starting...\n",(int)pthread_self(),
 699                                                    GetCurrentThreadId());
 700 #   endif
 701
 702     /* If a GC occurs before the thread is registered, that GC will     */
 703     /* ignore this thread.  That's fine, since it will block trying to  */
 704     /* acquire the allocation lock, and won't yet hold interesting      */
 705     /* pointers.                                                        */
 706     LOCK();
 707     /* We register the thread here instead of in the parent, so that    */
 708     /* we don't need to hold the allocation lock during pthread_create. */
 709     me = GC_new_thread();
 710     UNLOCK();
 711
 712     start = si -> start_routine;
 713     start_arg = si -> arg;
 714     if (si-> detached) me -> flags |= DETACHED;
 715     me -> pthread_id = pthread_id = pthread_self();
 716
 717     GC_free(si); /* was allocated uncollectable */
 718
 719     pthread_cleanup_push(GC_thread_exit_proc, (void *)me);
 720     result = (*start)(start_arg);
 721     me -> status = result;
 722     pthread_cleanup_pop(0);
 723
 724 #   if DEBUG_CYGWIN_THREADS
 725       GC_printf2("thread 0x%x(0x%x) returned from start routine.\n",
 726                  (int)pthread_self(),GetCurrentThreadId());
 727 #   endif
 728
 729     return(result);
 730 }
 731
 732 void GC_thread_exit_proc(void *arg)
 733 {
 734     GC_thread me = (GC_thread)arg;
 735     int i;
 736
 737 #   if DEBUG_CYGWIN_THREADS
 738       GC_printf2("thread 0x%x(0x%x) called pthread_exit().\n",
 739                  (int)pthread_self(),GetCurrentThreadId());
 740 #   endif
 741
 742     LOCK();
 743     if (me -> flags & DETACHED) {
 744       GC_delete_thread(GetCurrentThreadId());
 745     } else {
 746       /* deallocate it as part of join */
 747       me -> flags |= FINISHED;
 748     }
 749     UNLOCK();
 750 }
 751
 752 /* nothing required here... */
 753 int GC_pthread_sigmask(int how, const sigset_t *set, sigset_t *oset) {
 754   return pthread_sigmask(how, set, oset);
 755 }
 756
 757 int GC_pthread_detach(pthread_t thread)
 758 {
 759     int result;
 760     GC_thread thread_gc_id;
 761
 762     LOCK();
 763     thread_gc_id = GC_lookup_thread(thread);
 764     UNLOCK();
 765     result = pthread_detach(thread);
 766     if (result == 0) {
 767       LOCK();
 768       thread_gc_id -> flags |= DETACHED;
 769       /* Here the pthread thread id may have been recycled. */
 770       if (thread_gc_id -> flags & FINISHED) {
 771         GC_delete_gc_thread(thread_gc_id);
 772       }
 773       UNLOCK();
 774     }
 775     return result;
 776 }
 777
 778 #else /* !CYGWIN32 */
 779
 780 /*
 781  * We avoid acquiring locks here, since this doesn't seem to be preemptable.
 782  * Pontus Rydin suggests wrapping the thread start routine instead.
 783  */
 784 #if defined(GC_DLL) || defined(GC_INSIDE_DLL)
 785 BOOL WINAPI GC_DllMain(HINSTANCE inst, ULONG reason, LPVOID reserved)
 786 {
 787   switch (reason) {
 788   case DLL_PROCESS_ATTACH:
 789     GC_init();  /* Force initialization before thread attach.   */
 790     /* fall through */
 791   case DLL_THREAD_ATTACH:
 792     GC_ASSERT(GC_thr_initialized);
 793     if (GC_main_thread != GetCurrentThreadId()) {
 794         GC_new_thread();
 795     } /* o.w. we already did it during GC_thr_init(), called by GC_init() */
 796     break;
 797
 798   case DLL_THREAD_DETACH:
 799     GC_delete_thread(GetCurrentThreadId());
 800     break;
 801
 802   case DLL_PROCESS_DETACH:
 803     {
 804       int i;
 805
 806       LOCK();
 807       for (i = 0; i <= GC_get_max_thread_index(); ++i)
 808       {
 809           if (thread_table[i].in_use)
 810             GC_delete_gc_thread(thread_table + i);
 811       }
 812       UNLOCK();
 813
 814       GC_deinit();
 815       DeleteCriticalSection(&GC_allocate_ml);
 816     }
 817     break;
 818
 819   }
 820   return TRUE;
 821 }
 822 #endif /* GC_DLL */
 823 #endif /* !CYGWIN32 */
 824
 825 # endif /* !MSWINCE */
 826
 827 #endif /* GC_WIN32_THREADS */