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 BINDING_STACK_SIZE etc */
  36 #include "thread.h"
  37 #include "arch.h"
  38 #include "target-arch-os.h"
  39 #include "os.h"
  40 #include "globals.h"
  41 #include "dynbind.h"
  42 #include "genesis/cons.h"
  43 #include "genesis/fdefn.h"
  44 #include "interr.h"             /* for lose() */
  45 #include "alloc.h"
  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 DELAY_THREAD_POST_MORTEM 5
  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 #ifdef LISP_FEATURE_SB_THREAD
  67 struct thread_post_mortem {
  68 #ifdef DELAY_THREAD_POST_MORTEM
  69     struct thread_post_mortem *next;
  70 #endif
  71     os_thread_t os_thread;
  72     pthread_attr_t *os_attr;
  73     os_vm_address_t os_address;
  74 };
  75
  76 #ifdef DELAY_THREAD_POST_MORTEM
  77 static int pending_thread_post_mortem_count = 0;
  78 pthread_mutex_t thread_post_mortem_lock = PTHREAD_MUTEX_INITIALIZER;
  79 #endif
  80 static struct thread_post_mortem * volatile pending_thread_post_mortem = 0;
  81 #endif
  82
  83 int dynamic_values_bytes=TLS_SIZE*sizeof(lispobj);  /* same for all threads */
  84 struct thread *all_threads;
  85 extern struct interrupt_data * global_interrupt_data;
  86
  87 #ifdef LISP_FEATURE_SB_THREAD
  88 pthread_mutex_t all_threads_lock = PTHREAD_MUTEX_INITIALIZER;
  89 #ifdef LOCK_CREATE_THREAD
  90 static pthread_mutex_t create_thread_lock = PTHREAD_MUTEX_INITIALIZER;
  91 #endif
  92 #ifdef LISP_FEATURE_GCC_TLS
  93 __thread struct thread *current_thread;
  94 #endif
  95 pthread_key_t lisp_thread = 0;
  96 #endif
  97
  98 #if defined(LISP_FEATURE_X86) || defined(LISP_FEATURE_X86_64)
  99 extern lispobj call_into_lisp_first_time(lispobj fun, lispobj *args, int nargs);
 100 #endif
 101
 102 static void
 103 link_thread(struct thread *th)
 104 {
 105     if (all_threads) all_threads->prev=th;
 106     th->next=all_threads;
 107     th->prev=0;
 108     all_threads=th;
 109 }
 110
 111 #ifdef LISP_FEATURE_SB_THREAD
 112 static void
 113 unlink_thread(struct thread *th)
 114 {
 115     if (th->prev)
 116         th->prev->next = th->next;
 117     else
 118         all_threads = th->next;
 119     if (th->next)
 120         th->next->prev = th->prev;
 121 }
 122
 123 /* Only access thread state with blockables blocked. */
 124 lispobj
 125 thread_state(struct thread *thread)
 126 {
 127     lispobj state;
 128     sigset_t old;
 129     block_blockable_signals(NULL, &old);
 130     os_sem_wait(thread->state_sem, "thread_state");
 131     state = thread->state;
 132     os_sem_post(thread->state_sem, "thread_state");
 133     thread_sigmask(SIG_SETMASK, &old, NULL);
 134     return state;
 135 }
 136
 137 void
 138 set_thread_state(struct thread *thread, lispobj state)
 139 {
 140     int i, waitcount = 0;
 141     sigset_t old;
 142     block_blockable_signals(NULL, &old);
 143     os_sem_wait(thread->state_sem, "set_thread_state");
 144     if (thread->state != state) {
 145         if ((STATE_STOPPED==state) ||
 146             (STATE_DEAD==state)) {
 147             waitcount = thread->state_not_running_waitcount;
 148             thread->state_not_running_waitcount = 0;
 149             for (i=0; i<waitcount; i++)
 150                 os_sem_post(thread->state_not_running_sem, "set_thread_state (not running)");
 151         }
 152         if ((STATE_RUNNING==state) ||
 153             (STATE_DEAD==state)) {
 154             waitcount = thread->state_not_stopped_waitcount;
 155             thread->state_not_stopped_waitcount = 0;
 156             for (i=0; i<waitcount; i++)
 157                 os_sem_post(thread->state_not_stopped_sem, "set_thread_state (not stopped)");
 158         }
 159         thread->state = state;
 160     }
 161     os_sem_post(thread->state_sem, "set_thread_state");
 162     thread_sigmask(SIG_SETMASK, &old, NULL);
 163 }
 164
 165 void
 166 wait_for_thread_state_change(struct thread *thread, lispobj state)
 167 {
 168     sigset_t old;
 169     os_sem_t *wait_sem;
 170     block_blockable_signals(NULL, &old);
 171   start:
 172     os_sem_wait(thread->state_sem, "wait_for_thread_state_change");
 173     if (thread->state == state) {
 174         switch (state) {
 175         case STATE_RUNNING:
 176             wait_sem = thread->state_not_running_sem;
 177             thread->state_not_running_waitcount++;
 178             break;
 179         case STATE_STOPPED:
 180             wait_sem = thread->state_not_stopped_sem;
 181             thread->state_not_stopped_waitcount++;
 182             break;
 183         default:
 184             lose("Invalid state in wait_for_thread_state_change: "OBJ_FMTX"\n", state);
 185         }
 186     } else {
 187         wait_sem = NULL;
 188     }
 189     os_sem_post(thread->state_sem, "wait_for_thread_state_change");
 190     if (wait_sem) {
 191         os_sem_wait(wait_sem, "wait_for_thread_state_change");
 192         goto start;
 193     }
 194     thread_sigmask(SIG_SETMASK, &old, NULL);
 195 }
 196 #endif
 197
 198 static int
 199 initial_thread_trampoline(struct thread *th)
 200 {
 201     lispobj function;
 202 #if defined(LISP_FEATURE_X86) || defined(LISP_FEATURE_X86_64)
 203     lispobj *args = NULL;
 204 #endif
 205 #ifdef LISP_FEATURE_SB_THREAD
 206     pthread_setspecific(lisp_thread, (void *)1);
 207 #endif
 208 #if defined(LISP_FEATURE_SB_THREAD) && defined(LISP_FEATURE_PPC)
 209     /* SIG_STOP_FOR_GC defaults to blocked on PPC? */
 210     unblock_gc_signals(0,0);
 211 #endif
 212     function = th->no_tls_value_marker;
 213     th->no_tls_value_marker = NO_TLS_VALUE_MARKER_WIDETAG;
 214     if(arch_os_thread_init(th)==0) return 1;
 215     link_thread(th);
 216     th->os_thread=thread_self();
 217 #ifndef LISP_FEATURE_WIN32
 218     protect_control_stack_hard_guard_page(1, NULL);
 219     protect_binding_stack_hard_guard_page(1, NULL);
 220     protect_alien_stack_hard_guard_page(1, NULL);
 221     protect_control_stack_guard_page(1, NULL);
 222     protect_binding_stack_guard_page(1, NULL);
 223     protect_alien_stack_guard_page(1, NULL);
 224 #endif
 225
 226 #if defined(LISP_FEATURE_X86) || defined(LISP_FEATURE_X86_64)
 227     return call_into_lisp_first_time(function,args,0);
 228 #else
 229     return funcall0(function);
 230 #endif
 231 }
 232
 233 #ifdef LISP_FEATURE_SB_THREAD
 234 #define THREAD_STATE_LOCK_SIZE \
 235     ((sizeof(os_sem_t))+(sizeof(os_sem_t))+(sizeof(os_sem_t)))
 236 #else
 237 #define THREAD_STATE_LOCK_SIZE 0
 238 #endif
 239
 240 #define THREAD_STRUCT_SIZE (thread_control_stack_size + BINDING_STACK_SIZE + \
 241                             ALIEN_STACK_SIZE +                               \
 242                             THREAD_STATE_LOCK_SIZE +                         \
 243                             dynamic_values_bytes +                           \
 244                             32 * SIGSTKSZ +                                  \
 245                             THREAD_ALIGNMENT_BYTES)
 246
 247 #ifdef LISP_FEATURE_SB_THREAD
 248 /* THREAD POST MORTEM CLEANUP
 249  *
 250  * Memory allocated for the thread stacks cannot be reclaimed while
 251  * the thread is still alive, so we need a mechanism for post mortem
 252  * cleanups. FIXME: We actually have three, for historical reasons as
 253  * the saying goes. Do we really need three? Nikodemus guesses that
 254  * not anymore, now that we properly call pthread_attr_destroy before
 255  * freeing the stack. */
 256
 257 static struct thread_post_mortem *
 258 plan_thread_post_mortem(struct thread *corpse)
 259 {
 260     if (corpse) {
 261         struct thread_post_mortem *post_mortem = malloc(sizeof(struct thread_post_mortem));
 262         gc_assert(post_mortem);
 263         post_mortem->os_thread = corpse->os_thread;
 264         post_mortem->os_attr = corpse->os_attr;
 265         post_mortem->os_address = corpse->os_address;
 266 #ifdef DELAY_THREAD_POST_MORTEM
 267         post_mortem->next = NULL;
 268 #endif
 269         return post_mortem;
 270     } else {
 271         /* FIXME: When does this happen? */
 272         return NULL;
 273     }
 274 }
 275
 276 static void
 277 perform_thread_post_mortem(struct thread_post_mortem *post_mortem)
 278 {
 279 #ifdef CREATE_POST_MORTEM_THREAD
 280     pthread_detach(pthread_self());
 281 #endif
 282     if (post_mortem) {
 283         gc_assert(!pthread_join(post_mortem->os_thread, NULL));
 284         gc_assert(!pthread_attr_destroy(post_mortem->os_attr));
 285         free(post_mortem->os_attr);
 286         os_invalidate(post_mortem->os_address, THREAD_STRUCT_SIZE);
 287         free(post_mortem);
 288     }
 289 }
 290
 291 static void
 292 schedule_thread_post_mortem(struct thread *corpse)
 293 {
 294     struct thread_post_mortem *post_mortem = NULL;
 295     if (corpse) {
 296         post_mortem = plan_thread_post_mortem(corpse);
 297
 298 #ifdef DELAY_THREAD_POST_MORTEM
 299         pthread_mutex_lock(&thread_post_mortem_lock);
 300         /* First stick the new post mortem to the end of the queue. */
 301         if (pending_thread_post_mortem) {
 302             struct thread_post_mortem *next = pending_thread_post_mortem;
 303             while (next->next) {
 304                 next = next->next;
 305             }
 306             next->next = post_mortem;
 307         } else {
 308             pending_thread_post_mortem = post_mortem;
 309         }
 310         /* Then, if there are enough things in the queue, clean up one
 311          * from the head -- or increment the count, and null out the
 312          * post_mortem we have. */
 313         if (pending_thread_post_mortem_count > DELAY_THREAD_POST_MORTEM) {
 314             post_mortem = pending_thread_post_mortem;
 315             pending_thread_post_mortem = post_mortem->next;
 316         } else {
 317             pending_thread_post_mortem_count++;
 318             post_mortem = NULL;
 319         }
 320         pthread_mutex_unlock(&thread_post_mortem_lock);
 321         /* Finally run, the cleanup, if any. */
 322         perform_thread_post_mortem(post_mortem);
 323 #elif defined(CREATE_POST_MORTEM_THREAD)
 324         gc_assert(!pthread_create(&thread, NULL, perform_thread_post_mortem, post_mortem));
 325 #else
 326         post_mortem = (struct thread_post_mortem *)
 327             swap_lispobjs((lispobj *)(void *)&pending_thread_post_mortem,
 328                           (lispobj)post_mortem);
 329         perform_thread_post_mortem(post_mortem);
 330 #endif
 331     }
 332 }
 333
 334 /* this is the first thing that runs in the child (which is why the
 335  * silly calling convention).  Basically it calls the user's requested
 336  * lisp function after doing arch_os_thread_init and whatever other
 337  * bookkeeping needs to be done
 338  */
 339 int
 340 new_thread_trampoline(struct thread *th)
 341 {
 342     lispobj function;
 343     int result, lock_ret;
 344
 345     FSHOW((stderr,"/creating thread %lu\n", thread_self()));
 346     check_deferrables_blocked_or_lose(0);
 347     check_gc_signals_unblocked_or_lose(0);
 348     pthread_setspecific(lisp_thread, (void *)1);
 349     function = th->no_tls_value_marker;
 350     th->no_tls_value_marker = NO_TLS_VALUE_MARKER_WIDETAG;
 351     if(arch_os_thread_init(th)==0) {
 352         /* FIXME: handle error */
 353         lose("arch_os_thread_init failed\n");
 354     }
 355
 356     th->os_thread=thread_self();
 357     protect_control_stack_guard_page(1, NULL);
 358     protect_binding_stack_guard_page(1, NULL);
 359     protect_alien_stack_guard_page(1, NULL);
 360     /* Since GC can only know about this thread from the all_threads
 361      * list and we're just adding this thread to it, there is no
 362      * danger of deadlocking even with SIG_STOP_FOR_GC blocked (which
 363      * it is not). */
 364     lock_ret = pthread_mutex_lock(&all_threads_lock);
 365     gc_assert(lock_ret == 0);
 366     link_thread(th);
 367     lock_ret = pthread_mutex_unlock(&all_threads_lock);
 368     gc_assert(lock_ret == 0);
 369
 370     result = funcall0(function);
 371
 372     /* Block GC */
 373     block_blockable_signals(0, 0);
 374     set_thread_state(th, STATE_DEAD);
 375
 376     /* SIG_STOP_FOR_GC is blocked and GC might be waiting for this
 377      * thread, but since we are already dead it won't wait long. */
 378     lock_ret = pthread_mutex_lock(&all_threads_lock);
 379     gc_assert(lock_ret == 0);
 380
 381     gc_alloc_update_page_tables(BOXED_PAGE_FLAG, &th->alloc_region);
 382     unlink_thread(th);
 383     pthread_mutex_unlock(&all_threads_lock);
 384     gc_assert(lock_ret == 0);
 385
 386     if(th->tls_cookie>=0) arch_os_thread_cleanup(th);
 387     os_sem_destroy(th->state_sem);
 388     os_sem_destroy(th->state_not_running_sem);
 389     os_sem_destroy(th->state_not_stopped_sem);
 390
 391     os_invalidate((os_vm_address_t)th->interrupt_data,
 392                   (sizeof (struct interrupt_data)));
 393
 394 #ifdef LISP_FEATURE_MACH_EXCEPTION_HANDLER
 395     mach_lisp_thread_destroy(th);
 396 #endif
 397
 398     schedule_thread_post_mortem(th);
 399     FSHOW((stderr,"/exiting thread %lu\n", thread_self()));
 400     return result;
 401 }
 402
 403 #endif /* LISP_FEATURE_SB_THREAD */
 404
 405 static void
 406 free_thread_struct(struct thread *th)
 407 {
 408     if (th->interrupt_data)
 409         os_invalidate((os_vm_address_t) th->interrupt_data,
 410                       (sizeof (struct interrupt_data)));
 411     os_invalidate((os_vm_address_t) th->os_address,
 412                   THREAD_STRUCT_SIZE);
 413 }
 414
 415 #ifdef LISP_FEATURE_SB_THREAD
 416 /* FIXME: should be MAX_INTERRUPTS -1 ? */
 417 const unsigned int tls_index_start =
 418   MAX_INTERRUPTS + sizeof(struct thread)/sizeof(lispobj);
 419 #endif
 420
 421 /* this is called from any other thread to create the new one, and
 422  * initialize all parts of it that can be initialized from another
 423  * thread
 424  */
 425
 426 static struct thread *
 427 create_thread_struct(lispobj initial_function) {
 428     union per_thread_data *per_thread;
 429     struct thread *th=0;        /*  subdue gcc */
 430     void *spaces=0;
 431     void *aligned_spaces=0;
 432 #ifdef LISP_FEATURE_SB_THREAD
 433     unsigned int i;
 434 #endif
 435
 436     /* May as well allocate all the spaces at once: it saves us from
 437      * having to decide what to do if only some of the allocations
 438      * succeed. SPACES must be appropriately aligned, since the GC
 439      * expects the control stack to start at a page boundary -- and
 440      * the OS may have even more rigorous requirements. We can't rely
 441      * on the alignment passed from os_validate, since that might
 442      * assume the current (e.g. 4k) pagesize, while we calculate with
 443      * the biggest (e.g. 64k) pagesize allowed by the ABI. */
 444     spaces=os_validate(0, THREAD_STRUCT_SIZE);
 445     if(!spaces)
 446         return NULL;
 447     /* Aligning up is safe as THREAD_STRUCT_SIZE has
 448      * THREAD_ALIGNMENT_BYTES padding. */
 449     aligned_spaces = (void *)((((unsigned long)(char *)spaces)
 450                                + THREAD_ALIGNMENT_BYTES-1)
 451                               &~(unsigned long)(THREAD_ALIGNMENT_BYTES-1));
 452     per_thread=(union per_thread_data *)
 453         (aligned_spaces+
 454          thread_control_stack_size+
 455          BINDING_STACK_SIZE+
 456          ALIEN_STACK_SIZE +
 457          THREAD_STATE_LOCK_SIZE);
 458
 459 #ifdef LISP_FEATURE_SB_THREAD
 460     for(i = 0; i < (dynamic_values_bytes / sizeof(lispobj)); i++)
 461         per_thread->dynamic_values[i] = NO_TLS_VALUE_MARKER_WIDETAG;
 462     if (all_threads == 0) {
 463         if(SymbolValue(FREE_TLS_INDEX,0)==UNBOUND_MARKER_WIDETAG) {
 464             SetSymbolValue(FREE_TLS_INDEX,tls_index_start << WORD_SHIFT,0);
 465             SetSymbolValue(TLS_INDEX_LOCK,make_fixnum(0),0);
 466         }
 467 #define STATIC_TLS_INIT(sym,field) \
 468   ((struct symbol *)(sym-OTHER_POINTER_LOWTAG))->tls_index= \
 469   (THREAD_SLOT_OFFSET_WORDS(field) << WORD_SHIFT)
 470
 471         STATIC_TLS_INIT(BINDING_STACK_START,binding_stack_start);
 472 #ifdef BINDING_STACK_POINTER
 473         STATIC_TLS_INIT(BINDING_STACK_POINTER,binding_stack_pointer);
 474 #endif
 475         STATIC_TLS_INIT(CONTROL_STACK_START,control_stack_start);
 476         STATIC_TLS_INIT(CONTROL_STACK_END,control_stack_end);
 477 #ifdef ALIEN_STACK
 478         STATIC_TLS_INIT(ALIEN_STACK,alien_stack_pointer);
 479 #endif
 480 #if defined(LISP_FEATURE_X86) || defined (LISP_FEATURE_X86_64)
 481         STATIC_TLS_INIT(PSEUDO_ATOMIC_BITS,pseudo_atomic_bits);
 482 #endif
 483 #undef STATIC_TLS_INIT
 484     }
 485 #endif
 486
 487     th=&per_thread->thread;
 488     th->os_address = spaces;
 489     th->control_stack_start = aligned_spaces;
 490     th->binding_stack_start=
 491         (lispobj*)((void*)th->control_stack_start+thread_control_stack_size);
 492     th->control_stack_end = th->binding_stack_start;
 493     th->control_stack_guard_page_protected = T;
 494     th->alien_stack_start=
 495         (lispobj*)((void*)th->binding_stack_start+BINDING_STACK_SIZE);
 496     set_binding_stack_pointer(th,th->binding_stack_start);
 497     th->this=th;
 498     th->os_thread=0;
 499 #ifdef LISP_FEATURE_SB_THREAD
 500     th->os_attr=malloc(sizeof(pthread_attr_t));
 501     th->state_sem=(os_sem_t *)((void *)th->alien_stack_start + ALIEN_STACK_SIZE);
 502     th->state_not_running_sem=(os_sem_t *)
 503         ((void *)th->state_sem + (sizeof(os_sem_t)));
 504     th->state_not_stopped_sem=(os_sem_t *)
 505         ((void *)th->state_not_running_sem + (sizeof(os_sem_t)));
 506     th->state_not_running_waitcount = 0;
 507     th->state_not_stopped_waitcount = 0;
 508     os_sem_init(th->state_sem, 1);
 509     os_sem_init(th->state_not_running_sem, 0);
 510     os_sem_init(th->state_not_stopped_sem, 0);
 511 #endif
 512     th->state=STATE_RUNNING;
 513 #ifdef LISP_FEATURE_STACK_GROWS_DOWNWARD_NOT_UPWARD
 514     th->alien_stack_pointer=((void *)th->alien_stack_start
 515                              + ALIEN_STACK_SIZE-N_WORD_BYTES);
 516 #else
 517     th->alien_stack_pointer=((void *)th->alien_stack_start);
 518 #endif
 519 #if defined(LISP_FEATURE_X86) || defined (LISP_FEATURE_X86_64) || defined(LISP_FEATURE_SB_THREAD)
 520     th->pseudo_atomic_bits=0;
 521 #endif
 522 #ifdef LISP_FEATURE_GENCGC
 523     gc_set_region_empty(&th->alloc_region);
 524 #endif
 525 #ifdef LISP_FEATURE_SB_THREAD
 526     /* This parallels the same logic in globals.c for the
 527      * single-threaded foreign_function_call_active, KLUDGE and
 528      * all. */
 529 #if defined(LISP_FEATURE_X86) || defined(LISP_FEATURE_X86_64)
 530     th->foreign_function_call_active = 0;
 531 #else
 532     th->foreign_function_call_active = 1;
 533 #endif
 534 #endif
 535
 536 #ifndef LISP_FEATURE_SB_THREAD
 537     /* the tls-points-into-struct-thread trick is only good for threaded
 538      * sbcl, because unithread sbcl doesn't have tls.  So, we copy the
 539      * appropriate values from struct thread here, and make sure that
 540      * we use the appropriate SymbolValue macros to access any of the
 541      * variable quantities from the C runtime.  It's not quite OAOOM,
 542      * it just feels like it */
 543     SetSymbolValue(BINDING_STACK_START,(lispobj)th->binding_stack_start,th);
 544     SetSymbolValue(CONTROL_STACK_START,(lispobj)th->control_stack_start,th);
 545     SetSymbolValue(CONTROL_STACK_END,(lispobj)th->control_stack_end,th);
 546 #if defined(LISP_FEATURE_X86) || defined (LISP_FEATURE_X86_64)
 547     SetSymbolValue(ALIEN_STACK,(lispobj)th->alien_stack_pointer,th);
 548     SetSymbolValue(PSEUDO_ATOMIC_BITS,(lispobj)th->pseudo_atomic_bits,th);
 549 #endif
 550 #endif
 551     bind_variable(CURRENT_CATCH_BLOCK,make_fixnum(0),th);
 552     bind_variable(CURRENT_UNWIND_PROTECT_BLOCK,make_fixnum(0),th);
 553     bind_variable(FREE_INTERRUPT_CONTEXT_INDEX,make_fixnum(0),th);
 554     bind_variable(INTERRUPT_PENDING, NIL,th);
 555     bind_variable(INTERRUPTS_ENABLED,T,th);
 556     bind_variable(ALLOW_WITH_INTERRUPTS,T,th);
 557     bind_variable(GC_PENDING,NIL,th);
 558     bind_variable(ALLOC_SIGNAL,NIL,th);
 559 #ifdef PINNED_OBJECTS
 560     bind_variable(PINNED_OBJECTS,NIL,th);
 561 #endif
 562 #ifdef LISP_FEATURE_SB_THREAD
 563     bind_variable(STOP_FOR_GC_PENDING,NIL,th);
 564 #endif
 565 #ifndef LISP_FEATURE_C_STACK_IS_CONTROL_STACK
 566     access_control_stack_pointer(th)=th->control_stack_start;
 567 #endif
 568
 569     th->interrupt_data = (struct interrupt_data *)
 570         os_validate(0,(sizeof (struct interrupt_data)));
 571     if (!th->interrupt_data) {
 572         free_thread_struct(th);
 573         return 0;
 574     }
 575     th->interrupt_data->pending_handler = 0;
 576     th->interrupt_data->gc_blocked_deferrables = 0;
 577 #ifdef LISP_FEATURE_PPC
 578     th->interrupt_data->allocation_trap_context = 0;
 579 #endif
 580     th->no_tls_value_marker=initial_function;
 581
 582     th->stepping = NIL;
 583     return th;
 584 }
 585
 586 void create_initial_thread(lispobj initial_function) {
 587     struct thread *th=create_thread_struct(initial_function);
 588 #ifdef LISP_FEATURE_SB_THREAD
 589     pthread_key_create(&lisp_thread, 0);
 590 #endif
 591     if(th) {
 592         initial_thread_trampoline(th); /* no return */
 593     } else lose("can't create initial thread\n");
 594 }
 595
 596 #ifdef LISP_FEATURE_SB_THREAD
 597
 598 #ifndef __USE_XOPEN2K
 599 extern int pthread_attr_setstack (pthread_attr_t *__attr, void *__stackaddr,
 600                                   size_t __stacksize);
 601 #endif
 602
 603 boolean create_os_thread(struct thread *th,os_thread_t *kid_tid)
 604 {
 605     /* The new thread inherits the restrictive signal mask set here,
 606      * and enables signals again when it is set up properly. */
 607     sigset_t oldset;
 608     boolean r=1;
 609     int retcode = 0, initcode;
 610
 611     FSHOW_SIGNAL((stderr,"/create_os_thread: creating new thread\n"));
 612
 613     /* Blocking deferrable signals is enough, no need to block
 614      * SIG_STOP_FOR_GC because the child process is not linked onto
 615      * all_threads until it's ready. */
 616     block_deferrable_signals(0, &oldset);
 617
 618 #ifdef LOCK_CREATE_THREAD
 619     retcode = pthread_mutex_lock(&create_thread_lock);
 620     gc_assert(retcode == 0);
 621     FSHOW_SIGNAL((stderr,"/create_os_thread: got lock\n"));
 622 #endif
 623
 624     if((initcode = pthread_attr_init(th->os_attr)) ||
 625        /* call_into_lisp_first_time switches the stack for the initial
 626         * thread. For the others, we use this. */
 627        (pthread_attr_setstack(th->os_attr,th->control_stack_start,
 628                               thread_control_stack_size)) ||
 629        (retcode = pthread_create
 630         (kid_tid,th->os_attr,(void *(*)(void *))new_thread_trampoline,th))) {
 631         FSHOW_SIGNAL((stderr, "init = %d\n", initcode));
 632         FSHOW_SIGNAL((stderr, "pthread_create returned %d, errno %d\n",
 633                       retcode, errno));
 634         if(retcode < 0) {
 635             perror("create_os_thread");
 636         }
 637         r=0;
 638     }
 639
 640 #ifdef LOCK_CREATE_THREAD
 641     retcode = pthread_mutex_unlock(&create_thread_lock);
 642     gc_assert(retcode == 0);
 643     FSHOW_SIGNAL((stderr,"/create_os_thread: released lock\n"));
 644 #endif
 645     thread_sigmask(SIG_SETMASK,&oldset,0);
 646     return r;
 647 }
 648
 649 os_thread_t create_thread(lispobj initial_function) {
 650     struct thread *th, *thread = arch_os_get_current_thread();
 651     os_thread_t kid_tid = 0;
 652
 653     /* Must defend against async unwinds. */
 654     if (SymbolValue(INTERRUPTS_ENABLED, thread) != NIL)
 655         lose("create_thread is not safe when interrupts are enabled.\n");
 656
 657     /* Assuming that a fresh thread struct has no lisp objects in it,
 658      * linking it to all_threads can be left to the thread itself
 659      * without fear of gc lossage. initial_function violates this
 660      * assumption and must stay pinned until the child starts up. */
 661     th = create_thread_struct(initial_function);
 662     if (th && !create_os_thread(th,&kid_tid)) {
 663         free_thread_struct(th);
 664         kid_tid = 0;
 665     }
 666     return kid_tid;
 667 }
 668
 669 /* stopping the world is a two-stage process.  From this thread we signal
 670  * all the others with SIG_STOP_FOR_GC.  The handler for this signal does
 671  * the usual pseudo-atomic checks (we don't want to stop a thread while
 672  * it's in the middle of allocation) then waits for another SIG_STOP_FOR_GC.
 673  */
 674
 675 /* To avoid deadlocks when gc stops the world all clients of each
 676  * mutex must enable or disable SIG_STOP_FOR_GC for the duration of
 677  * holding the lock, but they must agree on which. */
 678 void gc_stop_the_world()
 679 {
 680     struct thread *p,*th=arch_os_get_current_thread();
 681     int status, lock_ret;
 682 #ifdef LOCK_CREATE_THREAD
 683     /* KLUDGE: Stopping the thread during pthread_create() causes deadlock
 684      * on FreeBSD. */
 685     FSHOW_SIGNAL((stderr,"/gc_stop_the_world:waiting on create_thread_lock\n"));
 686     lock_ret = pthread_mutex_lock(&create_thread_lock);
 687     gc_assert(lock_ret == 0);
 688     FSHOW_SIGNAL((stderr,"/gc_stop_the_world:got create_thread_lock\n"));
 689 #endif
 690     FSHOW_SIGNAL((stderr,"/gc_stop_the_world:waiting on lock\n"));
 691     /* keep threads from starting while the world is stopped. */
 692     lock_ret = pthread_mutex_lock(&all_threads_lock);      \
 693     gc_assert(lock_ret == 0);
 694
 695     FSHOW_SIGNAL((stderr,"/gc_stop_the_world:got lock\n"));
 696     /* stop all other threads by sending them SIG_STOP_FOR_GC */
 697     for(p=all_threads; p; p=p->next) {
 698         gc_assert(p->os_thread != 0);
 699         FSHOW_SIGNAL((stderr,"/gc_stop_the_world: thread=%lu, state=%x\n",
 700                       p->os_thread, thread_state(p)));
 701         if((p!=th) && ((thread_state(p)==STATE_RUNNING))) {
 702             FSHOW_SIGNAL((stderr,"/gc_stop_the_world: suspending thread %lu\n",
 703                           p->os_thread));
 704             /* We already hold all_thread_lock, P can become DEAD but
 705              * cannot exit, ergo it's safe to use pthread_kill. */
 706             status=pthread_kill(p->os_thread,SIG_STOP_FOR_GC);
 707             if (status==ESRCH) {
 708                 /* This thread has exited. */
 709                 gc_assert(thread_state(p)==STATE_DEAD);
 710             } else if (status) {
 711                 lose("cannot send suspend thread=%lu: %d, %s\n",
 712                      p->os_thread,status,strerror(status));
 713             }
 714         }
 715     }
 716     FSHOW_SIGNAL((stderr,"/gc_stop_the_world:signals sent\n"));
 717     for(p=all_threads;p;p=p->next) {
 718         if (p!=th) {
 719             FSHOW_SIGNAL
 720                 ((stderr,
 721                   "/gc_stop_the_world: waiting for thread=%lu: state=%x\n",
 722                   p->os_thread, thread_state(p)));
 723             wait_for_thread_state_change(p, STATE_RUNNING);
 724             if (p->state == STATE_RUNNING)
 725                 lose("/gc_stop_the_world: unexpected state");
 726         }
 727     }
 728     FSHOW_SIGNAL((stderr,"/gc_stop_the_world:end\n"));
 729 }
 730
 731 void gc_start_the_world()
 732 {
 733     struct thread *p,*th=arch_os_get_current_thread();
 734     int lock_ret;
 735     /* if a resumed thread creates a new thread before we're done with
 736      * this loop, the new thread will get consed on the front of
 737      * all_threads, but it won't have been stopped so won't need
 738      * restarting */
 739     FSHOW_SIGNAL((stderr,"/gc_start_the_world:begin\n"));
 740     for(p=all_threads;p;p=p->next) {
 741         gc_assert(p->os_thread!=0);
 742         if (p!=th) {
 743             lispobj state = thread_state(p);
 744             if (state != STATE_DEAD) {
 745                 if(state != STATE_STOPPED) {
 746                     lose("gc_start_the_world: wrong thread state is %d\n",
 747                          fixnum_value(state));
 748                 }
 749                 FSHOW_SIGNAL((stderr, "/gc_start_the_world: resuming %lu\n",
 750                               p->os_thread));
 751                 set_thread_state(p, STATE_RUNNING);
 752             }
 753         }
 754     }
 755
 756     lock_ret = pthread_mutex_unlock(&all_threads_lock);
 757     gc_assert(lock_ret == 0);
 758 #ifdef LOCK_CREATE_THREAD
 759     lock_ret = pthread_mutex_unlock(&create_thread_lock);
 760     gc_assert(lock_ret == 0);
 761 #endif
 762
 763     FSHOW_SIGNAL((stderr,"/gc_start_the_world:end\n"));
 764 }
 765 #endif
 766
 767 int
 768 thread_yield()
 769 {
 770 #ifdef LISP_FEATURE_SB_THREAD
 771     return sched_yield();
 772 #else
 773     return 0;
 774 #endif
 775 }
 776
 777 /* If the thread id given does not belong to a running thread (it has
 778  * exited or never even existed) pthread_kill _may_ fail with ESRCH,
 779  * but it is also allowed to just segfault, see
 780  * <http://udrepper.livejournal.com/16844.html>.
 781  *
 782  * Relying on thread ids can easily backfire since ids are recycled
 783  * (NPTL recycles them extremely fast) so a signal can be sent to
 784  * another process if the one it was sent to exited.
 785  *
 786  * We send signals in two places: signal_interrupt_thread sends a
 787  * signal that's harmless if delivered to another thread, but
 788  * SIG_STOP_FOR_GC is fatal.
 789  *
 790  * For these reasons, we must make sure that the thread is still alive
 791  * when the pthread_kill is called and return if the thread is
 792  * exiting. */
 793 int
 794 kill_safely(os_thread_t os_thread, int signal)
 795 {
 796     FSHOW_SIGNAL((stderr,"/kill_safely: %lu, %d\n", os_thread, signal));
 797     {
 798 #ifdef LISP_FEATURE_SB_THREAD
 799         sigset_t oldset;
 800         struct thread *thread;
 801         /* pthread_kill is not async signal safe and we don't want to be
 802          * interrupted while holding the lock. */
 803         block_deferrable_signals(0, &oldset);
 804         pthread_mutex_lock(&all_threads_lock);
 805         for (thread = all_threads; thread; thread = thread->next) {
 806             if (thread->os_thread == os_thread) {
 807                 int status = pthread_kill(os_thread, signal);
 808                 if (status)
 809                     lose("kill_safely: pthread_kill failed with %d\n", status);
 810                 break;
 811             }
 812         }
 813         pthread_mutex_unlock(&all_threads_lock);
 814         thread_sigmask(SIG_SETMASK,&oldset,0);
 815         if (thread)
 816             return 0;
 817         else
 818             return -1;
 819 #else
 820         int status;
 821         if (os_thread != 0)
 822             lose("kill_safely: who do you want to kill? %d?\n", os_thread);
 823         /* Dubious (as in don't know why it works) workaround for the
 824          * signal sometimes not being generated on darwin. */
 825 #ifdef LISP_FEATURE_DARWIN
 826         {
 827             sigset_t oldset;
 828             sigprocmask(SIG_BLOCK, &deferrable_sigset, &oldset);
 829             status = raise(signal);
 830             sigprocmask(SIG_SETMASK,&oldset,0);
 831         }
 832 #else
 833         status = raise(signal);
 834 #endif
 835         if (status == 0) {
 836             return 0;
 837         } else {
 838             lose("cannot raise signal %d, %d %s\n",
 839                  signal, status, strerror(errno));
 840         }
 841 #endif
 842     }
 843 }