boehm-gc/solaris_threads.c

   1 /*
   2  * Copyright (c) 1994 by Xerox Corporation.  All rights reserved.
   3  *
   4  * THIS MATERIAL IS PROVIDED AS IS, WITH ABSOLUTELY NO WARRANTY EXPRESSED
   5  * OR IMPLIED.  ANY USE IS AT YOUR OWN RISK.
   6  *
   7  * Permission is hereby granted to use or copy this program
   8  * for any purpose,  provided the above notices are retained on all copies.
   9  * Permission to modify the code and to distribute modified code is granted,
  10  * provided the above notices are retained, and a notice that the code was
  11  * modified is included with the above copyright notice.
  12  */
  13 /*
  14  * Support code for Solaris threads.  Provides functionality we wish Sun
  15  * had provided.  Relies on some information we probably shouldn't rely on.
  16  */
  17 /* Boehm, September 14, 1994 4:44 pm PDT */
  18
  19 #include "private/gc_priv.h"
  20
  21 # if defined(GC_SOLARIS_THREADS) || defined(GC_SOLARIS_PTHREADS)
  22 # include "private/solaris_threads.h"
  23 # include <thread.h>
  24 # include <synch.h>
  25 # include <signal.h>
  26 # include <fcntl.h>
  27 # include <sys/types.h>
  28 # include <sys/mman.h>
  29 # include <sys/time.h>
  30 # include <sys/resource.h>
  31 # include <sys/stat.h>
  32 # include <sys/syscall.h>
  33 # include <sys/procfs.h>
  34 # include <sys/lwp.h>
  35 # include <sys/reg.h>
  36 # define _CLASSIC_XOPEN_TYPES
  37 # include <unistd.h>
  38 # include <errno.h>
  39
  40 #ifdef HANDLE_FORK
  41   --> Not yet supported.  Try porting the code from linux_threads.c.
  42 #endif
  43
  44 /*
  45  * This is the default size of the LWP arrays. If there are more LWPs
  46  * than this when a stop-the-world GC happens, set_max_lwps will be
  47  * called to cope.
  48  * This must be higher than the number of LWPs at startup time.
  49  * The threads library creates a thread early on, so the min. is 3
  50  */
  51 # define DEFAULT_MAX_LWPS       4
  52
  53 #undef thr_join
  54 #undef thr_create
  55 #undef thr_suspend
  56 #undef thr_continue
  57
  58 cond_t GC_prom_join_cv;         /* Broadcast when any thread terminates */
  59 cond_t GC_create_cv;            /* Signalled when a new undetached      */
  60                                 /* thread starts.                       */
  61
  62
  63 #ifdef MMAP_STACKS
  64 static int GC_zfd;
  65 #endif /* MMAP_STACKS */
  66
  67 /* We use the allocation lock to protect thread-related data structures. */
  68
  69 /* We stop the world using /proc primitives.  This makes some   */
  70 /* minimal assumptions about the threads implementation.        */
  71 /* We don't play by the rules, since the rules make this        */
  72 /* impossible (as of Solaris 2.3).  Also note that as of        */
  73 /* Solaris 2.3 the various thread and lwp suspension            */
  74 /* primitives failed to stop threads by the time the request    */
  75 /* is completed.                                                */
  76
  77
  78 static sigset_t old_mask;
  79
  80 /* Sleep for n milliseconds, n < 1000   */
  81 void GC_msec_sleep(int n)
  82 {
  83     struct timespec ts;
  84
  85     ts.tv_sec = 0;
  86     ts.tv_nsec = 1000000*n;
  87     if (syscall(SYS_nanosleep, &ts, 0) < 0) {
  88         ABORT("nanosleep failed");
  89     }
  90 }
  91 /* Turn off preemption;  gross but effective.           */
  92 /* Caller has allocation lock.                          */
  93 /* Actually this is not needed under Solaris 2.3 and    */
  94 /* 2.4, but hopefully that'll change.                   */
  95 void preempt_off()
  96 {
  97     sigset_t set;
  98
  99     (void)sigfillset(&set);
 100     sigdelset(&set, SIGABRT);
 101     syscall(SYS_sigprocmask, SIG_SETMASK, &set, &old_mask);
 102 }
 103
 104 void preempt_on()
 105 {
 106     syscall(SYS_sigprocmask, SIG_SETMASK, &old_mask, NULL);
 107 }
 108
 109 int GC_main_proc_fd = -1;
 110
 111
 112 struct lwp_cache_entry {
 113     lwpid_t lc_id;
 114     int lc_descr;       /* /proc file descriptor.       */
 115 }  GC_lwp_cache_default[DEFAULT_MAX_LWPS];
 116
 117 static int max_lwps = DEFAULT_MAX_LWPS;
 118 static struct lwp_cache_entry *GC_lwp_cache = GC_lwp_cache_default;
 119
 120 static prgregset_t GC_lwp_registers_default[DEFAULT_MAX_LWPS];
 121 static prgregset_t *GC_lwp_registers = GC_lwp_registers_default;
 122
 123 /* Return a file descriptor for the /proc entry corresponding   */
 124 /* to the given lwp.  The file descriptor may be stale if the   */
 125 /* lwp exited and a new one was forked.                         */
 126 static int open_lwp(lwpid_t id)
 127 {
 128     int result;
 129     static int next_victim = 0;
 130     register int i;
 131
 132     for (i = 0; i < max_lwps; i++) {
 133         if (GC_lwp_cache[i].lc_id == id) return(GC_lwp_cache[i].lc_descr);
 134     }
 135     result = syscall(SYS_ioctl, GC_main_proc_fd, PIOCOPENLWP, &id);
 136     /*
 137      * If PIOCOPENLWP fails, try closing fds in the cache until it succeeds.
 138      */
 139     if (result < 0 && errno == EMFILE) {
 140             for (i = 0; i < max_lwps; i++) {
 141                 if (GC_lwp_cache[i].lc_id != 0) {
 142                         (void)syscall(SYS_close, GC_lwp_cache[i].lc_descr);
 143                         result = syscall(SYS_ioctl, GC_main_proc_fd, PIOCOPENLWP, &id);
 144                         if (result >= 0 || (result < 0 && errno != EMFILE))
 145                                 break;
 146                 }
 147             }
 148     }
 149     if (result < 0) {
 150         if (errno == EMFILE) {
 151                 ABORT("Too many open files");
 152         }
 153         return(-1) /* exited? */;
 154     }
 155     if (GC_lwp_cache[next_victim].lc_id != 0)
 156         (void)syscall(SYS_close, GC_lwp_cache[next_victim].lc_descr);
 157     GC_lwp_cache[next_victim].lc_id = id;
 158     GC_lwp_cache[next_victim].lc_descr = result;
 159     if (++next_victim >= max_lwps)
 160         next_victim = 0;
 161     return(result);
 162 }
 163
 164 static void uncache_lwp(lwpid_t id)
 165 {
 166     register int i;
 167
 168     for (i = 0; i < max_lwps; i++) {
 169         if (GC_lwp_cache[i].lc_id == id) {
 170             (void)syscall(SYS_close, GC_lwp_cache[id].lc_descr);
 171             GC_lwp_cache[i].lc_id = 0;
 172             break;
 173         }
 174     }
 175 }
 176         /* Sequence of current lwp ids  */
 177 static lwpid_t GC_current_ids_default[DEFAULT_MAX_LWPS + 1];
 178 static lwpid_t *GC_current_ids = GC_current_ids_default;
 179
 180         /* Temporary used below (can be big if large number of LWPs) */
 181 static lwpid_t last_ids_default[DEFAULT_MAX_LWPS + 1];
 182 static lwpid_t *last_ids = last_ids_default;
 183
 184
 185 #define ROUNDUP(n)    WORDS_TO_BYTES(ROUNDED_UP_WORDS(n))
 186
 187 static void set_max_lwps(GC_word n)
 188 {
 189     char *mem;
 190     char *oldmem;
 191     int required_bytes = ROUNDUP(n * sizeof(struct lwp_cache_entry))
 192         + ROUNDUP(n * sizeof(prgregset_t))
 193         + ROUNDUP((n + 1) * sizeof(lwpid_t))
 194         + ROUNDUP((n + 1) * sizeof(lwpid_t));
 195
 196     GC_expand_hp_inner(divHBLKSZ((word)required_bytes));
 197     oldmem = mem = GC_scratch_alloc(required_bytes);
 198     if (0 == mem) ABORT("No space for lwp data structures");
 199
 200     /*
 201      * We can either flush the old lwp cache or copy it over. Do the latter.
 202      */
 203     memcpy(mem, GC_lwp_cache, max_lwps * sizeof(struct lwp_cache_entry));
 204     GC_lwp_cache = (struct lwp_cache_entry*)mem;
 205     mem += ROUNDUP(n * sizeof(struct lwp_cache_entry));
 206
 207     BZERO(GC_lwp_registers, max_lwps * sizeof(GC_lwp_registers[0]));
 208     GC_lwp_registers = (prgregset_t *)mem;
 209     mem += ROUNDUP(n * sizeof(prgregset_t));
 210
 211
 212     GC_current_ids = (lwpid_t *)mem;
 213     mem += ROUNDUP((n + 1) * sizeof(lwpid_t));
 214
 215     last_ids = (lwpid_t *)mem;
 216     mem += ROUNDUP((n + 1)* sizeof(lwpid_t));
 217
 218     if (mem > oldmem + required_bytes)
 219         ABORT("set_max_lwps buffer overflow");
 220
 221     max_lwps = n;
 222 }
 223
 224
 225 /* Stop all lwps in process.  Assumes preemption is off.        */
 226 /* Caller has allocation lock (and any other locks he may       */
 227 /* need).                                                       */
 228 static void stop_all_lwps()
 229 {
 230     int lwp_fd;
 231     char buf[30];
 232     prstatus_t status;
 233     register int i;
 234     GC_bool changed;
 235     lwpid_t me = _lwp_self();
 236
 237     if (GC_main_proc_fd == -1) {
 238         sprintf(buf, "/proc/%d", getpid());
 239         GC_main_proc_fd = syscall(SYS_open, buf, O_RDONLY);
 240         if (GC_main_proc_fd < 0) {
 241                 if (errno == EMFILE)
 242                         ABORT("/proc open failed: too many open files");
 243                 GC_printf1("/proc open failed: errno %d", errno);
 244                 abort();
 245         }
 246     }
 247     BZERO(GC_lwp_registers, sizeof (prgregset_t) * max_lwps);
 248     for (i = 0; i < max_lwps; i++)
 249         last_ids[i] = 0;
 250     for (;;) {
 251     if (syscall(SYS_ioctl, GC_main_proc_fd, PIOCSTATUS, &status) < 0)
 252         ABORT("Main PIOCSTATUS failed");
 253         if (status.pr_nlwp < 1)
 254                 ABORT("Invalid number of lwps returned by PIOCSTATUS");
 255         if (status.pr_nlwp >= max_lwps) {
 256                 set_max_lwps(status.pr_nlwp*2 + 10);
 257                 /*
 258                  * The data in the old GC_current_ids and
 259                  * GC_lwp_registers has been trashed. Cleaning out last_ids
 260                  * will make sure every LWP gets re-examined.
 261                  */
 262                 for (i = 0; i < max_lwps; i++)
 263                         last_ids[i] = 0;
 264                 continue;
 265     }
 266         if (syscall(SYS_ioctl, GC_main_proc_fd, PIOCLWPIDS, GC_current_ids) < 0)
 267             ABORT("PIOCLWPIDS failed");
 268         changed = FALSE;
 269         for (i = 0; GC_current_ids[i] != 0 && i < max_lwps; i++) {
 270             if (GC_current_ids[i] != last_ids[i]) {
 271                 changed = TRUE;
 272                 if (GC_current_ids[i] != me) {
 273                     /* PIOCSTOP doesn't work without a writable         */
 274                     /* descriptor.  And that makes the process          */
 275                     /* undebuggable.                                    */
 276                     if (_lwp_suspend(GC_current_ids[i]) < 0) {
 277                         /* Could happen if the lwp exited */
 278                         uncache_lwp(GC_current_ids[i]);
 279                         GC_current_ids[i] = me; /* ignore */
 280                     }
 281                 }
 282             }
 283         }
 284         /*
 285          * In the unlikely event something does a fork between the
 286          * PIOCSTATUS and the PIOCLWPIDS.
 287          */
 288         if (i >= max_lwps)
 289                 continue;
 290         /* All lwps in GC_current_ids != me have been suspended.  Note  */
 291         /* that _lwp_suspend is idempotent.                             */
 292         for (i = 0; GC_current_ids[i] != 0; i++) {
 293             if (GC_current_ids[i] != last_ids[i]) {
 294                 if (GC_current_ids[i] != me) {
 295                     lwp_fd = open_lwp(GC_current_ids[i]);
 296                     if (lwp_fd == -1)
 297                     {
 298                             GC_current_ids[i] = me;
 299                             continue;
 300                     }
 301                     /* LWP should be stopped.  Empirically it sometimes */
 302                     /* isn't, and more frequently the PR_STOPPED flag   */
 303                     /* is not set.  Wait for PR_STOPPED.                */
 304                     if (syscall(SYS_ioctl, lwp_fd,
 305                                 PIOCSTATUS, &status) < 0) {
 306                         /* Possible if the descriptor was stale, or */
 307                         /* we encountered the 2.3 _lwp_suspend bug. */
 308                         uncache_lwp(GC_current_ids[i]);
 309                         GC_current_ids[i] = me; /* handle next time. */
 310                     } else {
 311                         while (!(status.pr_flags & PR_STOPPED)) {
 312                             GC_msec_sleep(1);
 313                             if (syscall(SYS_ioctl, lwp_fd,
 314                                         PIOCSTATUS, &status) < 0) {
 315                                 ABORT("Repeated PIOCSTATUS failed");
 316                             }
 317                             if (status.pr_flags & PR_STOPPED) break;
 318
 319                             GC_msec_sleep(20);
 320                             if (syscall(SYS_ioctl, lwp_fd,
 321                                         PIOCSTATUS, &status) < 0) {
 322                                 ABORT("Repeated PIOCSTATUS failed");
 323                             }
 324                         }
 325                         if (status.pr_who !=  GC_current_ids[i]) {
 326                                 /* can happen if thread was on death row */
 327                                 uncache_lwp(GC_current_ids[i]);
 328                                 GC_current_ids[i] = me; /* handle next time. */
 329                                 continue;
 330                         }
 331                         /* Save registers where collector can */
 332                         /* find them.                     */
 333                             BCOPY(status.pr_reg, GC_lwp_registers[i],
 334                                   sizeof (prgregset_t));
 335                     }
 336                 }
 337             }
 338         }
 339         if (!changed) break;
 340         for (i = 0; i < max_lwps; i++) last_ids[i] = GC_current_ids[i];
 341     }
 342 }
 343
 344 /* Restart all lwps in process.  Assumes preemption is off.     */
 345 static void restart_all_lwps()
 346 {
 347     int lwp_fd;
 348     register int i;
 349     GC_bool changed;
 350     lwpid_t me = _lwp_self();
 351 #   define PARANOID
 352
 353     for (i = 0; GC_current_ids[i] != 0; i++) {
 354 #       ifdef PARANOID
 355           if (GC_current_ids[i] != me) {
 356             int lwp_fd = open_lwp(GC_current_ids[i]);
 357             prstatus_t status;
 358
 359             if (lwp_fd < 0) ABORT("open_lwp failed");
 360             if (syscall(SYS_ioctl, lwp_fd,
 361                         PIOCSTATUS, &status) < 0) {
 362                 ABORT("PIOCSTATUS failed in restart_all_lwps");
 363             }
 364             if (memcmp(status.pr_reg, GC_lwp_registers[i],
 365                        sizeof (prgregset_t)) != 0) {
 366                     int j;
 367
 368                     for(j = 0; j < NPRGREG; j++)
 369                     {
 370                             GC_printf3("%i: %x -> %x\n", j,
 371                                        GC_lwp_registers[i][j],
 372                                        status.pr_reg[j]);
 373                     }
 374                 ABORT("Register contents changed");
 375             }
 376             if (!status.pr_flags & PR_STOPPED) {
 377                 ABORT("lwp no longer stopped");
 378             }
 379 #ifdef SPARC
 380             {
 381                     gwindows_t windows;
 382               if (syscall(SYS_ioctl, lwp_fd,
 383                         PIOCGWIN, &windows) < 0) {
 384                 ABORT("PIOCSTATUS failed in restart_all_lwps");
 385               }
 386               if (windows.wbcnt > 0) ABORT("unsaved register windows");
 387             }
 388 #endif
 389           }
 390 #       endif /* PARANOID */
 391         if (GC_current_ids[i] == me) continue;
 392         if (_lwp_continue(GC_current_ids[i]) < 0) {
 393             ABORT("Failed to restart lwp");
 394         }
 395     }
 396     if (i >= max_lwps) ABORT("Too many lwps");
 397 }
 398
 399 GC_bool GC_multithreaded = 0;
 400
 401 void GC_stop_world()
 402 {
 403     preempt_off();
 404     if (GC_multithreaded)
 405         stop_all_lwps();
 406 }
 407
 408 void GC_start_world()
 409 {
 410     if (GC_multithreaded)
 411         restart_all_lwps();
 412     preempt_on();
 413 }
 414
 415 void GC_thr_init(void);
 416
 417 GC_bool GC_thr_initialized = FALSE;
 418
 419 size_t GC_min_stack_sz;
 420
 421
 422 /*
 423  * stack_head is stored at the top of free stacks
 424  */
 425 struct stack_head {
 426         struct stack_head       *next;
 427         ptr_t                   base;
 428         thread_t                owner;
 429 };
 430
 431 # define N_FREE_LISTS 25
 432 struct stack_head *GC_stack_free_lists[N_FREE_LISTS] = { 0 };
 433                 /* GC_stack_free_lists[i] is free list for stacks of    */
 434                 /* size GC_min_stack_sz*2**i.                           */
 435                 /* Free lists are linked through stack_head stored      */                      /* at top of stack.                                     */
 436
 437 /* Return a stack of size at least *stack_size.  *stack_size is */
 438 /* replaced by the actual stack size.                           */
 439 /* Caller holds allocation lock.                                */
 440 ptr_t GC_stack_alloc(size_t * stack_size)
 441 {
 442     register size_t requested_sz = *stack_size;
 443     register size_t search_sz = GC_min_stack_sz;
 444     register int index = 0;     /* = log2(search_sz/GC_min_stack_sz) */
 445     register ptr_t base;
 446     register struct stack_head *result;
 447
 448     while (search_sz < requested_sz) {
 449         search_sz *= 2;
 450         index++;
 451     }
 452     if ((result = GC_stack_free_lists[index]) == 0
 453         && (result = GC_stack_free_lists[index+1]) != 0) {
 454         /* Try next size up. */
 455         search_sz *= 2; index++;
 456     }
 457     if (result != 0) {
 458         base =  GC_stack_free_lists[index]->base;
 459         GC_stack_free_lists[index] = GC_stack_free_lists[index]->next;
 460     } else {
 461 #ifdef MMAP_STACKS
 462         base = (ptr_t)mmap(0, search_sz + GC_page_size,
 463                              PROT_READ|PROT_WRITE, MAP_PRIVATE |MAP_NORESERVE,
 464                              GC_zfd, 0);
 465         if (base == (ptr_t)-1)
 466         {
 467                 *stack_size = 0;
 468                 return NULL;
 469         }
 470
 471         mprotect(base, GC_page_size, PROT_NONE);
 472         /* Should this use divHBLKSZ(search_sz + GC_page_size) ? -- cf */
 473         GC_is_fresh((struct hblk *)base, divHBLKSZ(search_sz));
 474         base += GC_page_size;
 475
 476 #else
 477         base = (ptr_t) GC_scratch_alloc(search_sz + 2*GC_page_size);
 478         if (base == NULL)
 479         {
 480                 *stack_size = 0;
 481                 return NULL;
 482         }
 483
 484         base = (ptr_t)(((word)base + GC_page_size) & ~(GC_page_size - 1));
 485         /* Protect hottest page to detect overflow. */
 486 #       ifdef SOLARIS23_MPROTECT_BUG_FIXED
 487             mprotect(base, GC_page_size, PROT_NONE);
 488 #       endif
 489         GC_is_fresh((struct hblk *)base, divHBLKSZ(search_sz));
 490
 491         base += GC_page_size;
 492 #endif
 493     }
 494     *stack_size = search_sz;
 495     return(base);
 496 }
 497
 498 /* Caller holds  allocationlock.                                        */
 499 void GC_stack_free(ptr_t stack, size_t size)
 500 {
 501     register int index = 0;
 502     register size_t search_sz = GC_min_stack_sz;
 503     register struct stack_head *head;
 504
 505 #ifdef MMAP_STACKS
 506     /* Zero pointers */
 507     mmap(stack, size, PROT_READ|PROT_WRITE, MAP_PRIVATE|MAP_NORESERVE|MAP_FIXED,
 508          GC_zfd, 0);
 509 #endif
 510     while (search_sz < size) {
 511         search_sz *= 2;
 512         index++;
 513     }
 514     if (search_sz != size) ABORT("Bad stack size");
 515
 516     head = (struct stack_head *)(stack + search_sz - sizeof(struct stack_head));
 517     head->next = GC_stack_free_lists[index];
 518     head->base = stack;
 519     GC_stack_free_lists[index] = head;
 520 }
 521
 522 void GC_my_stack_limits();
 523
 524 /* Notify virtual dirty bit implementation that known empty parts of    */
 525 /* stacks do not contain useful data.                                   */
 526 /* Caller holds allocation lock.                                        */
 527 void GC_old_stacks_are_fresh()
 528 {
 529 /* No point in doing this for MMAP stacks - and pointers are zero'd out */
 530 /* by the mmap in GC_stack_free */
 531 #ifndef MMAP_STACKS
 532     register int i;
 533     register struct stack_head *s;
 534     register ptr_t p;
 535     register size_t sz;
 536     register struct hblk * h;
 537     int dummy;
 538
 539     for (i = 0, sz= GC_min_stack_sz; i < N_FREE_LISTS;
 540          i++, sz *= 2) {
 541          for (s = GC_stack_free_lists[i]; s != 0; s = s->next) {
 542              p = s->base;
 543              h = (struct hblk *)(((word)p + HBLKSIZE-1) & ~(HBLKSIZE-1));
 544              if ((ptr_t)h == p) {
 545                  GC_is_fresh((struct hblk *)p, divHBLKSZ(sz));
 546              } else {
 547                  GC_is_fresh((struct hblk *)p, divHBLKSZ(sz) - 1);
 548                  BZERO(p, (ptr_t)h - p);
 549              }
 550          }
 551     }
 552 #endif /* MMAP_STACKS */
 553     GC_my_stack_limits();
 554 }
 555
 556 /* The set of all known threads.  We intercept thread creation and      */
 557 /* joins.  We never actually create detached threads.  We allocate all  */
 558 /* new thread stacks ourselves.  These allow us to maintain this        */
 559 /* data structure.                                                      */
 560
 561 # define THREAD_TABLE_SZ 128    /* Must be power of 2   */
 562 volatile GC_thread GC_threads[THREAD_TABLE_SZ];
 563
 564 void GC_push_thread_structures GC_PROTO((void))
 565 {
 566     GC_push_all((ptr_t)(GC_threads), (ptr_t)(GC_threads)+sizeof(GC_threads));
 567 }
 568
 569 /* Add a thread to GC_threads.  We assume it wasn't already there.      */
 570 /* Caller holds allocation lock.                                        */
 571 GC_thread GC_new_thread(thread_t id)
 572 {
 573     int hv = ((word)id) % THREAD_TABLE_SZ;
 574     GC_thread result;
 575     static struct GC_Thread_Rep first_thread;
 576     static GC_bool first_thread_used = FALSE;
 577
 578     if (!first_thread_used) {
 579         result = &first_thread;
 580         first_thread_used = TRUE;
 581         /* Dont acquire allocation lock, since we may already hold it. */
 582     } else {
 583         result = (struct GC_Thread_Rep *)
 584                  GC_INTERNAL_MALLOC(sizeof(struct GC_Thread_Rep), NORMAL);
 585     }
 586     if (result == 0) return(0);
 587     result -> id = id;
 588     result -> next = GC_threads[hv];
 589     GC_threads[hv] = result;
 590     /* result -> finished = 0; */
 591     (void) cond_init(&(result->join_cv), USYNC_THREAD, 0);
 592     return(result);
 593 }
 594
 595 /* Delete a thread from GC_threads.  We assume it is there.     */
 596 /* (The code intentionally traps if it wasn't.)                 */
 597 /* Caller holds allocation lock.                                */
 598 void GC_delete_thread(thread_t id)
 599 {
 600     int hv = ((word)id) % THREAD_TABLE_SZ;
 601     register GC_thread p = GC_threads[hv];
 602     register GC_thread prev = 0;
 603
 604     while (p -> id != id) {
 605         prev = p;
 606         p = p -> next;
 607     }
 608     if (prev == 0) {
 609         GC_threads[hv] = p -> next;
 610     } else {
 611         prev -> next = p -> next;
 612     }
 613 }
 614
 615 /* Return the GC_thread correpsonding to a given thread_t.      */
 616 /* Returns 0 if it's not there.                                 */
 617 /* Caller holds  allocation lock.                               */
 618 GC_thread GC_lookup_thread(thread_t id)
 619 {
 620     int hv = ((word)id) % THREAD_TABLE_SZ;
 621     register GC_thread p = GC_threads[hv];
 622
 623     while (p != 0 && p -> id != id) p = p -> next;
 624     return(p);
 625 }
 626
 627 /* Solaris 2/Intel uses an initial stack size limit slightly bigger than the
 628    SPARC default of 8 MB.  Account for this to warn only if the user has
 629    raised the limit beyond the default.
 630
 631    This is identical to DFLSSIZ defined in <sys/vm_machparam.h>.  This file
 632    is installed in /usr/platform/`uname -m`/include, which is not in the
 633    default include directory list, so copy the definition here.  */
 634 #ifdef I386
 635 # define MAX_ORIG_STACK_SIZE (8 * 1024 * 1024 + ((USRSTACK) & 0x3FFFFF))
 636 #else
 637 # define MAX_ORIG_STACK_SIZE (8 * 1024 * 1024)
 638 #endif
 639
 640 word GC_get_orig_stack_size() {
 641     struct rlimit rl;
 642     static int warned = 0;
 643     int result;
 644
 645     if (getrlimit(RLIMIT_STACK, &rl) != 0) ABORT("getrlimit failed");
 646     result = (word)rl.rlim_cur & ~(HBLKSIZE-1);
 647     if (result > MAX_ORIG_STACK_SIZE) {
 648         if (!warned) {
 649             WARN("Large stack limit(%ld): only scanning 8 MB\n", result);
 650             warned = 1;
 651         }
 652         result = MAX_ORIG_STACK_SIZE;
 653     }
 654     return result;
 655 }
 656
 657 /* Notify dirty bit implementation of unused parts of my stack. */
 658 /* Caller holds allocation lock.                                */
 659 void GC_my_stack_limits()
 660 {
 661     int dummy;
 662     register ptr_t hottest = (ptr_t)((word)(&dummy) & ~(HBLKSIZE-1));
 663     register GC_thread me = GC_lookup_thread(thr_self());
 664     register size_t stack_size = me -> stack_size;
 665     register ptr_t stack;
 666
 667     if (stack_size == 0) {
 668       /* original thread */
 669         /* Empirically, what should be the stack page with lowest       */
 670         /* address is actually inaccessible.                            */
 671         stack_size = GC_get_orig_stack_size() - GC_page_size;
 672         stack = GC_stackbottom - stack_size + GC_page_size;
 673     } else {
 674         stack = me -> stack;
 675     }
 676     if (stack > hottest || stack + stack_size < hottest) {
 677         ABORT("sp out of bounds");
 678     }
 679     GC_is_fresh((struct hblk *)stack, divHBLKSZ(hottest - stack));
 680 }
 681
 682
 683 /* We hold allocation lock.  Should do exactly the right thing if the   */
 684 /* world is stopped.  Should not fail if it isn't.                      */
 685 void GC_push_all_stacks()
 686 {
 687     register int i;
 688     register GC_thread p;
 689     register ptr_t sp = GC_approx_sp();
 690     register ptr_t bottom, top;
 691     struct rlimit rl;
 692
 693 #   define PUSH(bottom,top) \
 694       if (GC_dirty_maintained) { \
 695         GC_push_selected((bottom), (top), GC_page_was_ever_dirty, \
 696                       GC_push_all_stack); \
 697       } else { \
 698         GC_push_all_stack((bottom), (top)); \
 699       }
 700     GC_push_all_stack((ptr_t)GC_lwp_registers,
 701                       (ptr_t)GC_lwp_registers
 702                       + max_lwps * sizeof(GC_lwp_registers[0]));
 703     for (i = 0; i < THREAD_TABLE_SZ; i++) {
 704       for (p = GC_threads[i]; p != 0; p = p -> next) {
 705         if (p -> stack_size != 0) {
 706             bottom = p -> stack;
 707             top = p -> stack + p -> stack_size;
 708         } else {
 709             /* The original stack. */
 710             bottom = GC_stackbottom - GC_get_orig_stack_size() + GC_page_size;
 711             top = GC_stackbottom;
 712         }
 713         if ((word)sp > (word)bottom && (word)sp < (word)top) bottom = sp;
 714         PUSH(bottom, top);
 715       }
 716     }
 717 }
 718
 719
 720 int GC_is_thread_stack(ptr_t addr)
 721 {
 722     register int i;
 723     register GC_thread p;
 724     register ptr_t bottom, top;
 725
 726     for (i = 0; i < THREAD_TABLE_SZ; i++) {
 727       for (p = GC_threads[i]; p != 0; p = p -> next) {
 728         if (p -> stack_size != 0) {
 729             if (p -> stack <= addr &&
 730                 addr < p -> stack + p -> stack_size)
 731                     return 1;
 732         }
 733       }
 734     }
 735     return 0;
 736 }
 737
 738 /* The only thread that ever really performs a thr_join.        */
 739 void * GC_thr_daemon(void * dummy)
 740 {
 741     void *status;
 742     thread_t departed;
 743     register GC_thread t;
 744     register int i;
 745     register int result;
 746
 747     for(;;) {
 748       start:
 749         result = thr_join((thread_t)0, &departed, &status);
 750         LOCK();
 751         if (result != 0) {
 752             /* No more threads; wait for create. */
 753             for (i = 0; i < THREAD_TABLE_SZ; i++) {
 754                 for (t = GC_threads[i]; t != 0; t = t -> next) {
 755                     if (!(t -> flags & (DETACHED | FINISHED))) {
 756                       UNLOCK();
 757                       goto start; /* Thread started just before we */
 758                                   /* acquired the lock.            */
 759                     }
 760                 }
 761             }
 762             cond_wait(&GC_create_cv, &GC_allocate_ml);
 763             UNLOCK();
 764         } else {
 765             t = GC_lookup_thread(departed);
 766             GC_multithreaded--;
 767             if (!(t -> flags & CLIENT_OWNS_STACK)) {
 768                 GC_stack_free(t -> stack, t -> stack_size);
 769             }
 770             if (t -> flags & DETACHED) {
 771                 GC_delete_thread(departed);
 772             } else {
 773                 t -> status = status;
 774                 t -> flags |= FINISHED;
 775                 cond_signal(&(t -> join_cv));
 776                 cond_broadcast(&GC_prom_join_cv);
 777             }
 778             UNLOCK();
 779         }
 780     }
 781 }
 782
 783 /* We hold the allocation lock, or caller ensures that 2 instances      */
 784 /* cannot be invoked concurrently.                                      */
 785 void GC_thr_init(void)
 786 {
 787     GC_thread t;
 788     thread_t tid;
 789     int ret;
 790
 791     if (GC_thr_initialized)
 792             return;
 793     GC_thr_initialized = TRUE;
 794     GC_min_stack_sz = ((thr_min_stack() + 32*1024 + HBLKSIZE-1)
 795                        & ~(HBLKSIZE - 1));
 796 #ifdef MMAP_STACKS
 797     GC_zfd = open("/dev/zero", O_RDONLY);
 798     if (GC_zfd == -1)
 799             ABORT("Can't open /dev/zero");
 800 #endif /* MMAP_STACKS */
 801     cond_init(&GC_prom_join_cv, USYNC_THREAD, 0);
 802     cond_init(&GC_create_cv, USYNC_THREAD, 0);
 803     /* Add the initial thread, so we can stop it.       */
 804       t = GC_new_thread(thr_self());
 805       t -> stack_size = 0;
 806       t -> flags = DETACHED | CLIENT_OWNS_STACK;
 807     ret = thr_create(0 /* stack */, 0 /* stack_size */, GC_thr_daemon,
 808                      0 /* arg */, THR_DETACHED | THR_DAEMON,
 809                      &tid /* thread_id */);
 810     if (ret != 0) {
 811         GC_err_printf1("Thr_create returned %ld\n", ret);
 812         ABORT("Cant fork daemon");
 813     }
 814     thr_setprio(tid, 126);
 815 }
 816
 817 /* We acquire the allocation lock to prevent races with         */
 818 /* stopping/starting world.                                     */
 819 /* This is no more correct than the underlying Solaris 2.X      */
 820 /* implementation.  Under 2.3 THIS IS BROKEN.                   */
 821 int GC_thr_suspend(thread_t target_thread)
 822 {
 823     GC_thread t;
 824     int result;
 825
 826     LOCK();
 827     result = thr_suspend(target_thread);
 828     if (result == 0) {
 829         t = GC_lookup_thread(target_thread);
 830         if (t == 0) ABORT("thread unknown to GC");
 831         t -> flags |= SUSPNDED;
 832     }
 833     UNLOCK();
 834     return(result);
 835 }
 836
 837 int GC_thr_continue(thread_t target_thread)
 838 {
 839     GC_thread t;
 840     int result;
 841
 842     LOCK();
 843     result = thr_continue(target_thread);
 844     if (result == 0) {
 845         t = GC_lookup_thread(target_thread);
 846         if (t == 0) ABORT("thread unknown to GC");
 847         t -> flags &= ~SUSPNDED;
 848     }
 849     UNLOCK();
 850     return(result);
 851 }
 852
 853 int GC_thr_join(thread_t wait_for, thread_t *departed, void **status)
 854 {
 855     register GC_thread t;
 856     int result = 0;
 857
 858     LOCK();
 859     if (wait_for == 0) {
 860         register int i;
 861         register GC_bool thread_exists;
 862
 863         for (;;) {
 864           thread_exists = FALSE;
 865           for (i = 0; i < THREAD_TABLE_SZ; i++) {
 866             for (t = GC_threads[i]; t != 0; t = t -> next) {
 867               if (!(t -> flags & DETACHED)) {
 868                 if (t -> flags & FINISHED) {
 869                   goto found;
 870                 }
 871                 thread_exists = TRUE;
 872               }
 873             }
 874           }
 875           if (!thread_exists) {
 876               result = ESRCH;
 877               goto out;
 878           }
 879           cond_wait(&GC_prom_join_cv, &GC_allocate_ml);
 880         }
 881     } else {
 882         t = GC_lookup_thread(wait_for);
 883         if (t == 0 || t -> flags & DETACHED) {
 884             result = ESRCH;
 885             goto out;
 886         }
 887         if (wait_for == thr_self()) {
 888             result = EDEADLK;
 889             goto out;
 890         }
 891         while (!(t -> flags & FINISHED)) {
 892             cond_wait(&(t -> join_cv), &GC_allocate_ml);
 893         }
 894
 895     }
 896   found:
 897     if (status) *status = t -> status;
 898     if (departed) *departed = t -> id;
 899     cond_destroy(&(t -> join_cv));
 900     GC_delete_thread(t -> id);
 901   out:
 902     UNLOCK();
 903     return(result);
 904 }
 905
 906
 907 int
 908 GC_thr_create(void *stack_base, size_t stack_size,
 909               void *(*start_routine)(void *), void *arg, long flags,
 910               thread_t *new_thread)
 911 {
 912     int result;
 913     GC_thread t;
 914     thread_t my_new_thread;
 915     word my_flags = 0;
 916     void * stack = stack_base;
 917
 918     LOCK();
 919     if (!GC_is_initialized) GC_init_inner();
 920     GC_multithreaded++;
 921     if (stack == 0) {
 922         if (stack_size == 0) stack_size = 1024*1024;
 923         stack = (void *)GC_stack_alloc(&stack_size);
 924         if (stack == 0) {
 925             GC_multithreaded--;
 926             UNLOCK();
 927             return(ENOMEM);
 928         }
 929     } else {
 930         my_flags |= CLIENT_OWNS_STACK;
 931     }
 932     if (flags & THR_DETACHED) my_flags |= DETACHED;
 933     if (flags & THR_SUSPENDED) my_flags |= SUSPNDED;
 934     result = thr_create(stack, stack_size, start_routine,
 935                         arg, flags & ~THR_DETACHED, &my_new_thread);
 936     if (result == 0) {
 937         t = GC_new_thread(my_new_thread);
 938         t -> flags = my_flags;
 939         if (!(my_flags & DETACHED)) cond_init(&(t -> join_cv), USYNC_THREAD, 0);
 940         t -> stack = stack;
 941         t -> stack_size = stack_size;
 942         if (new_thread != 0) *new_thread = my_new_thread;
 943         cond_signal(&GC_create_cv);
 944     } else {
 945         GC_multithreaded--;
 946         if (!(my_flags & CLIENT_OWNS_STACK)) {
 947             GC_stack_free(stack, stack_size);
 948         }
 949     }
 950     UNLOCK();
 951     return(result);
 952 }
 953
 954 # else /* !GC_SOLARIS_THREADS */
 955
 956 #ifndef LINT
 957   int GC_no_sunOS_threads;
 958 #endif
 959 #endif