src/runtime/thread.c

   1 /*
   2  * This software is part of the SBCL system. See the README file for
   3  * more information.
   4  *
   5  * This software is derived from the CMU CL system, which was
   6  * written at Carnegie Mellon University and released into the
   7  * public domain. The software is in the public domain and is
   8  * provided with absolutely no warranty. See the COPYING and CREDITS
   9  * files for more information.
  10  */
  11
  12 #include "sbcl.h"
  13
  14 #include <stdlib.h>
  15 #include <stdio.h>
  16 #include <string.h>
  17 #ifndef LISP_FEATURE_WIN32
  18 #include <sched.h>
  19 #endif
  20 #include <signal.h>
  21 #include <stddef.h>
  22 #include <errno.h>
  23 #include <sys/types.h>
  24 #ifndef LISP_FEATURE_WIN32
  25 #include <sys/wait.h>
  26 #endif
  27
  28 #ifdef LISP_FEATURE_MACH_EXCEPTION_HANDLER
  29 #include <mach/mach.h>
  30 #include <mach/mach_error.h>
  31 #include <mach/mach_types.h>
  32 #endif
  33
  34 #include "runtime.h"
  35 #include "validate.h"           /* for CONTROL_STACK_SIZE etc */
  36 #include "alloc.h"
  37 #include "thread.h"
  38 #include "arch.h"
  39 #include "target-arch-os.h"
  40 #include "os.h"
  41 #include "globals.h"
  42 #include "dynbind.h"
  43 #include "genesis/cons.h"
  44 #include "genesis/fdefn.h"
  45 #include "interr.h"             /* for lose() */
  46 #include "gc-internal.h"
  47
  48 #ifdef LISP_FEATURE_WIN32
  49 /*
  50  * Win32 doesn't have SIGSTKSZ, and we're not switching stacks anyway,
  51  * so define it arbitrarily
  52  */
  53 #define SIGSTKSZ 1024
  54 #endif
  55
  56 #if defined(LISP_FEATURE_DARWIN) && defined(LISP_FEATURE_SB_THREAD)
  57 #define QUEUE_FREEABLE_THREAD_STACKS
  58 #define LOCK_CREATE_THREAD
  59 #endif
  60
  61 #ifdef LISP_FEATURE_FREEBSD
  62 #define CREATE_CLEANUP_THREAD
  63 #define LOCK_CREATE_THREAD
  64 #endif
  65
  66 #define ALIEN_STACK_SIZE (1*1024*1024) /* 1Mb size chosen at random */
  67
  68 struct freeable_stack {
  69 #ifdef QUEUE_FREEABLE_THREAD_STACKS
  70     struct freeable_stack *next;
  71 #endif
  72     os_thread_t os_thread;
  73     os_vm_address_t os_address;
  74 };
  75
  76
  77 #ifdef QUEUE_FREEABLE_THREAD_STACKS
  78 static struct freeable_stack * volatile freeable_stack_queue = 0;
  79 static int freeable_stack_count = 0;
  80 pthread_mutex_t freeable_stack_lock = PTHREAD_MUTEX_INITIALIZER;
  81 #else
  82 static struct freeable_stack * volatile freeable_stack = 0;
  83 #endif
  84
  85 int dynamic_values_bytes=4096*sizeof(lispobj);  /* same for all threads */
  86 struct thread * volatile all_threads;
  87 extern struct interrupt_data * global_interrupt_data;
  88
  89 #ifdef LISP_FEATURE_SB_THREAD
  90 pthread_mutex_t all_threads_lock = PTHREAD_MUTEX_INITIALIZER;
  91 #ifdef LOCK_CREATE_THREAD
  92 static pthread_mutex_t create_thread_lock = PTHREAD_MUTEX_INITIALIZER;
  93 #endif
  94 #ifdef LISP_FEATURE_GCC_TLS
  95 __thread struct thread *current_thread;
  96 #endif
  97 #endif
  98
  99 #if defined(LISP_FEATURE_X86) || defined(LISP_FEATURE_X86_64)
 100 extern lispobj call_into_lisp_first_time(lispobj fun, lispobj *args, int nargs);
 101 #endif
 102
 103 static void
 104 link_thread(struct thread *th)
 105 {
 106     if (all_threads) all_threads->prev=th;
 107     th->next=all_threads;
 108     th->prev=0;
 109     all_threads=th;
 110 }
 111
 112 #ifdef LISP_FEATURE_SB_THREAD
 113 static void
 114 unlink_thread(struct thread *th)
 115 {
 116     if (th->prev)
 117         th->prev->next = th->next;
 118     else
 119         all_threads = th->next;
 120     if (th->next)
 121         th->next->prev = th->prev;
 122 }
 123 #endif
 124
 125 static int
 126 initial_thread_trampoline(struct thread *th)
 127 {
 128     lispobj function;
 129 #if defined(LISP_FEATURE_X86) || defined(LISP_FEATURE_X86_64)
 130     lispobj *args = NULL;
 131 #endif
 132     function = th->no_tls_value_marker;
 133     th->no_tls_value_marker = NO_TLS_VALUE_MARKER_WIDETAG;
 134     if(arch_os_thread_init(th)==0) return 1;
 135     link_thread(th);
 136     th->os_thread=thread_self();
 137 #ifndef LISP_FEATURE_WIN32
 138     protect_control_stack_guard_page(1);
 139 #endif
 140
 141 #if defined(LISP_FEATURE_X86) || defined(LISP_FEATURE_X86_64)
 142     return call_into_lisp_first_time(function,args,0);
 143 #else
 144     return funcall0(function);
 145 #endif
 146 }
 147
 148 #define THREAD_STRUCT_SIZE (THREAD_CONTROL_STACK_SIZE + BINDING_STACK_SIZE + \
 149                             ALIEN_STACK_SIZE + dynamic_values_bytes + \
 150                             32 * SIGSTKSZ + \
 151                             BACKEND_PAGE_SIZE)
 152
 153 #ifdef LISP_FEATURE_SB_THREAD
 154
 155 #ifdef QUEUE_FREEABLE_THREAD_STACKS
 156
 157 static void
 158 queue_freeable_thread_stack(struct thread *thread_to_be_cleaned_up)
 159 {
 160      struct freeable_stack *new_freeable_stack = 0;
 161      if (thread_to_be_cleaned_up) {
 162         /* FIXME: os_validate is mmap -- for small things like these
 163          * malloc would probably perform better. */
 164         new_freeable_stack = (struct freeable_stack *)
 165             os_validate(0, sizeof(struct freeable_stack));
 166         new_freeable_stack->next = NULL;
 167         new_freeable_stack->os_thread = thread_to_be_cleaned_up->os_thread;
 168         new_freeable_stack->os_address = thread_to_be_cleaned_up->os_address;
 169         pthread_mutex_lock(&freeable_stack_lock);
 170         if (freeable_stack_queue) {
 171             struct freeable_stack *next;
 172             next = freeable_stack_queue;
 173             while (next->next) {
 174                 next = next->next;
 175             }
 176             next->next = new_freeable_stack;
 177         } else {
 178             freeable_stack_queue = new_freeable_stack;
 179         }
 180         freeable_stack_count++;
 181         pthread_mutex_unlock(&freeable_stack_lock);
 182     }
 183 }
 184
 185 #define FREEABLE_STACK_QUEUE_SIZE 4
 186
 187 static void
 188 free_freeable_stacks() {
 189     if (freeable_stack_queue && (freeable_stack_count > FREEABLE_STACK_QUEUE_SIZE)) {
 190         struct freeable_stack* old;
 191         pthread_mutex_lock(&freeable_stack_lock);
 192         old = freeable_stack_queue;
 193         freeable_stack_queue = old->next;
 194         freeable_stack_count--;
 195         gc_assert(pthread_join(old->os_thread, NULL) == 0);
 196         FSHOW((stderr, "freeing thread %x stack\n", old->os_thread));
 197         os_invalidate(old->os_address, THREAD_STRUCT_SIZE);
 198         os_invalidate((os_vm_address_t)old, sizeof(struct freeable_stack));
 199         pthread_mutex_unlock(&freeable_stack_lock);
 200     }
 201 }
 202
 203 #elif defined(CREATE_CLEANUP_THREAD)
 204 static void *
 205 cleanup_thread(void *arg)
 206 {
 207     struct freeable_stack *freeable = arg;
 208     pthread_t self = pthread_self();
 209
 210     FSHOW((stderr, "/cleaner thread(%p): joining %p\n",
 211            self, freeable->os_thread));
 212     gc_assert(pthread_join(freeable->os_thread, NULL) == 0);
 213     FSHOW((stderr, "/cleaner thread(%p): free stack %p\n",
 214            self, freeable->stack));
 215     os_invalidate(freeable->os_address, THREAD_STRUCT_SIZE);
 216     free(freeable);
 217
 218     pthread_detach(self);
 219
 220     return NULL;
 221 }
 222
 223 static void
 224 create_cleanup_thread(struct thread *thread_to_be_cleaned_up)
 225 {
 226     pthread_t thread;
 227     int result;
 228
 229     if (thread_to_be_cleaned_up) {
 230         struct freeable_stack *freeable =
 231             malloc(sizeof(struct freeable_stack));
 232         gc_assert(freeable != NULL);
 233         freeable->os_thread = thread_to_be_cleaned_up->os_thread;
 234         freeable->os_address =
 235             (os_vm_address_t) thread_to_be_cleaned_up->os_address;
 236         result = pthread_create(&thread, NULL, cleanup_thread, freeable);
 237         gc_assert(result == 0);
 238     }
 239 }
 240
 241 #else
 242 static void
 243 free_thread_stack_later(struct thread *thread_to_be_cleaned_up)
 244 {
 245     struct freeable_stack *new_freeable_stack = 0;
 246     if (thread_to_be_cleaned_up) {
 247         new_freeable_stack = (struct freeable_stack *)
 248             os_validate(0, sizeof(struct freeable_stack));
 249         new_freeable_stack->os_thread = thread_to_be_cleaned_up->os_thread;
 250         new_freeable_stack->os_address = (os_vm_address_t)
 251             thread_to_be_cleaned_up->os_address;
 252     }
 253     new_freeable_stack = (struct freeable_stack *)
 254         swap_lispobjs((lispobj *)(void *)&freeable_stack,
 255                       (lispobj)new_freeable_stack);
 256     if (new_freeable_stack) {
 257         FSHOW((stderr,"/reaping %p\n", (void*) new_freeable_stack->os_thread));
 258         /* Under NPTL pthread_join really waits until the thread
 259          * exists and the stack can be safely freed. This is sadly not
 260          * mandated by the pthread spec. */
 261         gc_assert(pthread_join(new_freeable_stack->os_thread, NULL) == 0);
 262         os_invalidate(new_freeable_stack->os_address, THREAD_STRUCT_SIZE);
 263         os_invalidate((os_vm_address_t) new_freeable_stack,
 264                       sizeof(struct freeable_stack));
 265     }
 266 }
 267 #endif
 268
 269 /* this is the first thing that runs in the child (which is why the
 270  * silly calling convention).  Basically it calls the user's requested
 271  * lisp function after doing arch_os_thread_init and whatever other
 272  * bookkeeping needs to be done
 273  */
 274 int
 275 new_thread_trampoline(struct thread *th)
 276 {
 277     lispobj function;
 278     int result, lock_ret;
 279
 280     FSHOW((stderr,"/creating thread %lu\n", thread_self()));
 281     function = th->no_tls_value_marker;
 282     th->no_tls_value_marker = NO_TLS_VALUE_MARKER_WIDETAG;
 283     if(arch_os_thread_init(th)==0) {
 284         /* FIXME: handle error */
 285         lose("arch_os_thread_init failed\n");
 286     }
 287
 288     th->os_thread=thread_self();
 289     protect_control_stack_guard_page(1);
 290     /* Since GC can only know about this thread from the all_threads
 291      * list and we're just adding this thread to it there is no danger
 292      * of deadlocking even with SIG_STOP_FOR_GC blocked (which it is
 293      * not). */
 294     lock_ret = pthread_mutex_lock(&all_threads_lock);
 295     gc_assert(lock_ret == 0);
 296     link_thread(th);
 297     lock_ret = pthread_mutex_unlock(&all_threads_lock);
 298     gc_assert(lock_ret == 0);
 299
 300     result = funcall0(function);
 301
 302     /* Block GC */
 303     block_blockable_signals();
 304     th->state=STATE_DEAD;
 305
 306     /* SIG_STOP_FOR_GC is blocked and GC might be waiting for this
 307      * thread, but since we are already dead it won't wait long. */
 308     lock_ret = pthread_mutex_lock(&all_threads_lock);
 309     gc_assert(lock_ret == 0);
 310
 311     gc_alloc_update_page_tables(0, &th->alloc_region);
 312     unlink_thread(th);
 313     pthread_mutex_unlock(&all_threads_lock);
 314     gc_assert(lock_ret == 0);
 315
 316     if(th->tls_cookie>=0) arch_os_thread_cleanup(th);
 317     os_invalidate((os_vm_address_t)th->interrupt_data,
 318                   (sizeof (struct interrupt_data)));
 319
 320 #ifdef LISP_FEATURE_MACH_EXCEPTION_HANDLER
 321     FSHOW((stderr, "Deallocating mach port %x\n", THREAD_STRUCT_TO_EXCEPTION_PORT(th)));
 322     mach_port_move_member(mach_task_self(),
 323                           THREAD_STRUCT_TO_EXCEPTION_PORT(th),
 324                           MACH_PORT_NULL);
 325     mach_port_deallocate(mach_task_self(),
 326                          THREAD_STRUCT_TO_EXCEPTION_PORT(th));
 327     mach_port_destroy(mach_task_self(),
 328                       THREAD_STRUCT_TO_EXCEPTION_PORT(th));
 329 #endif
 330
 331 #ifdef QUEUE_FREEABLE_THREAD_STACKS
 332     queue_freeable_thread_stack(th);
 333 #elif defined(CREATE_CLEANUP_THREAD)
 334     create_cleanup_thread(th);
 335 #else
 336     free_thread_stack_later(th);
 337 #endif
 338
 339     FSHOW((stderr,"/exiting thread %p\n", thread_self()));
 340     return result;
 341 }
 342
 343 #endif /* LISP_FEATURE_SB_THREAD */
 344
 345 static void
 346 free_thread_struct(struct thread *th)
 347 {
 348     if (th->interrupt_data)
 349         os_invalidate((os_vm_address_t) th->interrupt_data,
 350                       (sizeof (struct interrupt_data)));
 351     os_invalidate((os_vm_address_t) th->os_address,
 352                   THREAD_STRUCT_SIZE);
 353 }
 354
 355 /* this is called from any other thread to create the new one, and
 356  * initialize all parts of it that can be initialized from another
 357  * thread
 358  */
 359
 360 static struct thread *
 361 create_thread_struct(lispobj initial_function) {
 362     union per_thread_data *per_thread;
 363     struct thread *th=0;        /*  subdue gcc */
 364     void *spaces=0;
 365     void *aligned_spaces=0;
 366 #ifdef LISP_FEATURE_SB_THREAD
 367     int i;
 368 #endif
 369
 370     /* May as well allocate all the spaces at once: it saves us from
 371      * having to decide what to do if only some of the allocations
 372      * succeed.  SPACES must be page-aligned, since the GC expects the
 373      * control stack to start at a page boundary.  We can't rely on the
 374      * alignment passed from os_validate, since that might assume the
 375      * current (e.g. 4k) pagesize, while we calculate with the biggest
 376      * (e.g. 64k) pagesize allowed by the ABI.  */
 377     spaces=os_validate(0, THREAD_STRUCT_SIZE);
 378     if(!spaces)
 379         return NULL;
 380     /* Aligning up is safe as THREAD_STRUCT_SIZE has BACKEND_PAGE_SIZE
 381      * padding. */
 382     aligned_spaces = (void *)((((unsigned long)(char *)spaces)
 383                                + BACKEND_PAGE_SIZE - 1)
 384                               & ~(unsigned long)(BACKEND_PAGE_SIZE - 1));
 385     per_thread=(union per_thread_data *)
 386         (aligned_spaces+
 387          THREAD_CONTROL_STACK_SIZE+
 388          BINDING_STACK_SIZE+
 389          ALIEN_STACK_SIZE);
 390
 391 #ifdef LISP_FEATURE_SB_THREAD
 392     for(i = 0; i < (dynamic_values_bytes / sizeof(lispobj)); i++)
 393         per_thread->dynamic_values[i] = NO_TLS_VALUE_MARKER_WIDETAG;
 394     if (all_threads == 0) {
 395         if(SymbolValue(FREE_TLS_INDEX,0)==UNBOUND_MARKER_WIDETAG) {
 396             SetSymbolValue
 397                 (FREE_TLS_INDEX,
 398                  /* FIXME: should be MAX_INTERRUPTS -1 ? */
 399                  make_fixnum(MAX_INTERRUPTS+
 400                              sizeof(struct thread)/sizeof(lispobj)),
 401                  0);
 402             SetSymbolValue(TLS_INDEX_LOCK,make_fixnum(0),0);
 403         }
 404 #define STATIC_TLS_INIT(sym,field) \
 405   ((struct symbol *)(sym-OTHER_POINTER_LOWTAG))->tls_index= \
 406   make_fixnum(THREAD_SLOT_OFFSET_WORDS(field))
 407
 408         STATIC_TLS_INIT(BINDING_STACK_START,binding_stack_start);
 409         STATIC_TLS_INIT(BINDING_STACK_POINTER,binding_stack_pointer);
 410         STATIC_TLS_INIT(CONTROL_STACK_START,control_stack_start);
 411         STATIC_TLS_INIT(CONTROL_STACK_END,control_stack_end);
 412         STATIC_TLS_INIT(ALIEN_STACK,alien_stack_pointer);
 413 #if defined(LISP_FEATURE_X86) || defined (LISP_FEATURE_X86_64)
 414         STATIC_TLS_INIT(PSEUDO_ATOMIC_BITS,pseudo_atomic_bits);
 415 #endif
 416 #undef STATIC_TLS_INIT
 417     }
 418 #endif
 419
 420     th=&per_thread->thread;
 421     th->os_address = spaces;
 422     th->control_stack_start = aligned_spaces;
 423     th->binding_stack_start=
 424         (lispobj*)((void*)th->control_stack_start+THREAD_CONTROL_STACK_SIZE);
 425     th->control_stack_end = th->binding_stack_start;
 426     th->alien_stack_start=
 427         (lispobj*)((void*)th->binding_stack_start+BINDING_STACK_SIZE);
 428     th->binding_stack_pointer=th->binding_stack_start;
 429     th->this=th;
 430     th->os_thread=0;
 431     th->state=STATE_RUNNING;
 432 #ifdef LISP_FEATURE_STACK_GROWS_DOWNWARD_NOT_UPWARD
 433     th->alien_stack_pointer=((void *)th->alien_stack_start
 434                              + ALIEN_STACK_SIZE-N_WORD_BYTES);
 435 #else
 436     th->alien_stack_pointer=((void *)th->alien_stack_start);
 437 #endif
 438 #if defined(LISP_FEATURE_X86) || defined (LISP_FEATURE_X86_64)
 439     th->pseudo_atomic_bits=0;
 440 #endif
 441 #ifdef LISP_FEATURE_GENCGC
 442     gc_set_region_empty(&th->alloc_region);
 443 #endif
 444
 445 #ifndef LISP_FEATURE_SB_THREAD
 446     /* the tls-points-into-struct-thread trick is only good for threaded
 447      * sbcl, because unithread sbcl doesn't have tls.  So, we copy the
 448      * appropriate values from struct thread here, and make sure that
 449      * we use the appropriate SymbolValue macros to access any of the
 450      * variable quantities from the C runtime.  It's not quite OAOOM,
 451      * it just feels like it */
 452     SetSymbolValue(BINDING_STACK_START,(lispobj)th->binding_stack_start,th);
 453     SetSymbolValue(CONTROL_STACK_START,(lispobj)th->control_stack_start,th);
 454     SetSymbolValue(CONTROL_STACK_END,(lispobj)th->control_stack_end,th);
 455 #if defined(LISP_FEATURE_X86) || defined (LISP_FEATURE_X86_64)
 456     SetSymbolValue(BINDING_STACK_POINTER,(lispobj)th->binding_stack_pointer,th);
 457     SetSymbolValue(ALIEN_STACK,(lispobj)th->alien_stack_pointer,th);
 458     SetSymbolValue(PSEUDO_ATOMIC_BITS,(lispobj)th->pseudo_atomic_bits,th);
 459 #else
 460     current_binding_stack_pointer=th->binding_stack_pointer;
 461     current_control_stack_pointer=th->control_stack_start;
 462 #endif
 463 #endif
 464     bind_variable(CURRENT_CATCH_BLOCK,make_fixnum(0),th);
 465     bind_variable(CURRENT_UNWIND_PROTECT_BLOCK,make_fixnum(0),th);
 466     bind_variable(FREE_INTERRUPT_CONTEXT_INDEX,make_fixnum(0),th);
 467     bind_variable(INTERRUPT_PENDING, NIL,th);
 468     bind_variable(INTERRUPTS_ENABLED,T,th);
 469     bind_variable(ALLOW_WITH_INTERRUPTS,T,th);
 470     bind_variable(GC_PENDING,NIL,th);
 471 #ifdef LISP_FEATURE_SB_THREAD
 472     bind_variable(STOP_FOR_GC_PENDING,NIL,th);
 473 #endif
 474
 475     th->interrupt_data = (struct interrupt_data *)
 476         os_validate(0,(sizeof (struct interrupt_data)));
 477     if (!th->interrupt_data) {
 478         free_thread_struct(th);
 479         return 0;
 480     }
 481     th->interrupt_data->pending_handler = 0;
 482     th->no_tls_value_marker=initial_function;
 483
 484     th->stepping = NIL;
 485     return th;
 486 }
 487
 488 #ifdef LISP_FEATURE_MACH_EXCEPTION_HANDLER
 489 mach_port_t setup_mach_exception_handling_thread();
 490 kern_return_t mach_thread_init(mach_port_t thread_exception_port);
 491
 492 #endif
 493
 494 void create_initial_thread(lispobj initial_function) {
 495     struct thread *th=create_thread_struct(initial_function);
 496     if(th) {
 497 #ifdef LISP_FEATURE_MACH_EXCEPTION_HANDLER
 498         kern_return_t ret;
 499
 500         setup_mach_exception_handling_thread();
 501 #endif
 502         initial_thread_trampoline(th); /* no return */
 503     } else lose("can't create initial thread\n");
 504 }
 505
 506 #ifdef LISP_FEATURE_SB_THREAD
 507
 508 #ifndef __USE_XOPEN2K
 509 extern int pthread_attr_setstack (pthread_attr_t *__attr, void *__stackaddr,
 510                                   size_t __stacksize);
 511 #endif
 512
 513 boolean create_os_thread(struct thread *th,os_thread_t *kid_tid)
 514 {
 515     /* The new thread inherits the restrictive signal mask set here,
 516      * and enables signals again when it is set up properly. */
 517     pthread_attr_t attr;
 518     sigset_t newset,oldset;
 519     boolean r=1;
 520     int retcode = 0, initcode;
 521
 522     FSHOW_SIGNAL((stderr,"/create_os_thread: creating new thread\n"));
 523
 524 #ifdef LOCK_CREATE_THREAD
 525     retcode = pthread_mutex_lock(&create_thread_lock);
 526     gc_assert(retcode == 0);
 527     FSHOW_SIGNAL((stderr,"/create_os_thread: got lock\n"));
 528 #endif
 529     sigemptyset(&newset);
 530     /* Blocking deferrable signals is enough, no need to block
 531      * SIG_STOP_FOR_GC because the child process is not linked onto
 532      * all_threads until it's ready. */
 533     sigaddset_deferrable(&newset);
 534     thread_sigmask(SIG_BLOCK, &newset, &oldset);
 535
 536 #if defined(LISP_FEATURE_DARWIN)
 537 #define CONTROL_STACK_ADJUST 8192 /* darwin wants page-aligned stacks */
 538 #else
 539 #define CONTROL_STACK_ADJUST 16
 540 #endif
 541
 542     if((initcode = pthread_attr_init(&attr)) ||
 543        /* FIXME: why do we even have this in the first place? */
 544        (pthread_attr_setstack(&attr,th->control_stack_start,
 545                               THREAD_CONTROL_STACK_SIZE-CONTROL_STACK_ADJUST)) ||
 546 #undef CONTROL_STACK_ADJUST
 547        (retcode = pthread_create
 548         (kid_tid,&attr,(void *(*)(void *))new_thread_trampoline,th))) {
 549         FSHOW_SIGNAL((stderr, "init = %d\n", initcode));
 550         FSHOW_SIGNAL((stderr, printf("pthread_create returned %d, errno %d\n", retcode, errno)));
 551         FSHOW_SIGNAL((stderr, "wanted stack size %d, min stack size %d\n",
 552                       THREAD_CONTROL_STACK_SIZE-16, PTHREAD_STACK_MIN));
 553         if(retcode < 0) {
 554             perror("create_os_thread");
 555         }
 556         r=0;
 557     }
 558
 559 #ifdef QUEUE_FREEABLE_THREAD_STACKS
 560     free_freeable_stacks();
 561 #endif
 562     thread_sigmask(SIG_SETMASK,&oldset,0);
 563 #ifdef LOCK_CREATE_THREAD
 564     retcode = pthread_mutex_unlock(&create_thread_lock);
 565     gc_assert(retcode == 0);
 566     FSHOW_SIGNAL((stderr,"/create_os_thread: released lock\n"));
 567 #endif
 568     return r;
 569 }
 570
 571 os_thread_t create_thread(lispobj initial_function) {
 572     struct thread *th;
 573     os_thread_t kid_tid;
 574
 575     /* Assuming that a fresh thread struct has no lisp objects in it,
 576      * linking it to all_threads can be left to the thread itself
 577      * without fear of gc lossage. initial_function violates this
 578      * assumption and must stay pinned until the child starts up. */
 579     th = create_thread_struct(initial_function);
 580     if(th==0) return 0;
 581
 582     if (create_os_thread(th,&kid_tid)) {
 583         return kid_tid;
 584     } else {
 585         free_thread_struct(th);
 586         return 0;
 587     }
 588 }
 589
 590 /* Send the signo to os_thread, retry if the rt signal queue is
 591  * full. */
 592 int
 593 kill_thread_safely(os_thread_t os_thread, int signo)
 594 {
 595     int r;
 596     /* The man page does not mention EAGAIN as a valid return value
 597      * for either pthread_kill or kill. But that's theory, this is
 598      * practice. By waiting here we assume that the delivery of this
 599      * signal is not necessary for the delivery of the signals in the
 600      * queue. In other words, we _assume_ there are no deadlocks. */
 601     while ((r=pthread_kill(os_thread,signo))==EAGAIN) {
 602         /* wait a bit then try again in the hope of the rt signal
 603          * queue not being full */
 604         FSHOW_SIGNAL((stderr,"/rt signal queue full\n"));
 605         /* FIXME: some kind of backoff (random, exponential) would be
 606          * nice. */
 607         sleep(1);
 608     }
 609     return r;
 610 }
 611
 612 int signal_interrupt_thread(os_thread_t os_thread)
 613 {
 614     int status = kill_thread_safely(os_thread, SIG_INTERRUPT_THREAD);
 615     if (status == 0) {
 616         return 0;
 617     } else if (status == ESRCH) {
 618         return -1;
 619     } else {
 620         lose("cannot send SIG_INTERRUPT_THREAD to thread=%lu: %d, %s\n",
 621              os_thread, status, strerror(status));
 622     }
 623 }
 624
 625 /* stopping the world is a two-stage process.  From this thread we signal
 626  * all the others with SIG_STOP_FOR_GC.  The handler for this signal does
 627  * the usual pseudo-atomic checks (we don't want to stop a thread while
 628  * it's in the middle of allocation) then waits for another SIG_STOP_FOR_GC.
 629  */
 630
 631 /* To avoid deadlocks when gc stops the world all clients of each
 632  * mutex must enable or disable SIG_STOP_FOR_GC for the duration of
 633  * holding the lock, but they must agree on which. */
 634 void gc_stop_the_world()
 635 {
 636     struct thread *p,*th=arch_os_get_current_thread();
 637     int status, lock_ret;
 638 #ifdef LOCK_CREATE_THREAD
 639     /* KLUDGE: Stopping the thread during pthread_create() causes deadlock
 640      * on FreeBSD. */
 641     FSHOW_SIGNAL((stderr,"/gc_stop_the_world:waiting on create_thread_lock, thread=%lu\n",
 642                   th->os_thread));
 643     lock_ret = pthread_mutex_lock(&create_thread_lock);
 644     gc_assert(lock_ret == 0);
 645     FSHOW_SIGNAL((stderr,"/gc_stop_the_world:got create_thread_lock, thread=%lu\n",
 646                   th->os_thread));
 647 #endif
 648     FSHOW_SIGNAL((stderr,"/gc_stop_the_world:waiting on lock, thread=%lu\n",
 649                   th->os_thread));
 650     /* keep threads from starting while the world is stopped. */
 651     lock_ret = pthread_mutex_lock(&all_threads_lock);      \
 652     gc_assert(lock_ret == 0);
 653
 654     FSHOW_SIGNAL((stderr,"/gc_stop_the_world:got lock, thread=%lu\n",
 655                   th->os_thread));
 656     /* stop all other threads by sending them SIG_STOP_FOR_GC */
 657     for(p=all_threads; p; p=p->next) {
 658         gc_assert(p->os_thread != 0);
 659         FSHOW_SIGNAL((stderr,"/gc_stop_the_world: p->state: %x\n", p->state));
 660         if((p!=th) && ((p->state==STATE_RUNNING))) {
 661             FSHOW_SIGNAL((stderr,"/gc_stop_the_world: suspending %x, os_thread %x\n",
 662                           p, p->os_thread));
 663             status=kill_thread_safely(p->os_thread,SIG_STOP_FOR_GC);
 664             if (status==ESRCH) {
 665                 /* This thread has exited. */
 666                 gc_assert(p->state==STATE_DEAD);
 667             } else if (status) {
 668                 lose("cannot send suspend thread=%lu: %d, %s\n",
 669                      p->os_thread,status,strerror(status));
 670             }
 671         }
 672     }
 673     FSHOW_SIGNAL((stderr,"/gc_stop_the_world:signals sent\n"));
 674     /* wait for the running threads to stop or finish */
 675     for(p=all_threads;p;) {
 676         FSHOW_SIGNAL((stderr,"/gc_stop_the_world: th: %p, p: %p\n", th, p));
 677         if((p!=th) && (p->state==STATE_RUNNING)) {
 678             sched_yield();
 679         } else {
 680             p=p->next;
 681         }
 682     }
 683     FSHOW_SIGNAL((stderr,"/gc_stop_the_world:end\n"));
 684 }
 685
 686 void gc_start_the_world()
 687 {
 688     struct thread *p,*th=arch_os_get_current_thread();
 689     int status, lock_ret;
 690     /* if a resumed thread creates a new thread before we're done with
 691      * this loop, the new thread will get consed on the front of
 692      * all_threads, but it won't have been stopped so won't need
 693      * restarting */
 694     FSHOW_SIGNAL((stderr,"/gc_start_the_world:begin\n"));
 695     for(p=all_threads;p;p=p->next) {
 696         gc_assert(p->os_thread!=0);
 697         if((p!=th) && (p->state!=STATE_DEAD)) {
 698             if(p->state!=STATE_SUSPENDED) {
 699                 lose("gc_start_the_world: wrong thread state is %d\n",
 700                      fixnum_value(p->state));
 701             }
 702             FSHOW_SIGNAL((stderr, "/gc_start_the_world: resuming %lu\n",
 703                           p->os_thread));
 704             p->state=STATE_RUNNING;
 705
 706 #if defined(SIG_RESUME_FROM_GC)
 707             status=kill_thread_safely(p->os_thread,SIG_RESUME_FROM_GC);
 708 #else
 709             status=kill_thread_safely(p->os_thread,SIG_STOP_FOR_GC);
 710 #endif
 711             if (status) {
 712                 lose("cannot resume thread=%lu: %d, %s\n",
 713                      p->os_thread,status,strerror(status));
 714             }
 715         }
 716     }
 717     /* If we waited here until all threads leave STATE_SUSPENDED, then
 718      * SIG_STOP_FOR_GC wouldn't need to be a rt signal. That has some
 719      * performance implications, but does away with the 'rt signal
 720      * queue full' problem. */
 721
 722     lock_ret = pthread_mutex_unlock(&all_threads_lock);
 723     gc_assert(lock_ret == 0);
 724 #ifdef LOCK_CREATE_THREAD
 725     lock_ret = pthread_mutex_unlock(&create_thread_lock);
 726     gc_assert(lock_ret == 0);
 727 #endif
 728
 729     FSHOW_SIGNAL((stderr,"/gc_start_the_world:end\n"));
 730 }
 731 #endif
 732
 733 int
 734 thread_yield()
 735 {
 736 #ifdef LISP_FEATURE_SB_THREAD
 737     return sched_yield();
 738 #else
 739     return 0;
 740 #endif
 741 }