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