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 #include "cpputil.h"
  48 #include "pseudo-atomic.h"
  49 #include "interrupt.h"
  50 #include "lispregs.h"
  51
  52 #ifdef LISP_FEATURE_WIN32
  53 /*
  54  * Win32 doesn't have SIGSTKSZ, and we're not switching stacks anyway,
  55  * so define it arbitrarily
  56  */
  57 #define SIGSTKSZ 1024
  58 #endif
  59
  60 #if defined(LISP_FEATURE_DARWIN) && defined(LISP_FEATURE_SB_THREAD)
  61 #define DELAY_THREAD_POST_MORTEM 5
  62 #define LOCK_CREATE_THREAD
  63 #endif
  64
  65 #ifdef LISP_FEATURE_FREEBSD
  66 #define CREATE_CLEANUP_THREAD
  67 #define LOCK_CREATE_THREAD
  68 #endif
  69
  70 #ifdef LISP_FEATURE_SB_THREAD
  71 struct thread_post_mortem {
  72 #ifdef DELAY_THREAD_POST_MORTEM
  73     struct thread_post_mortem *next;
  74 #endif
  75     os_thread_t os_thread;
  76     pthread_attr_t *os_attr;
  77     os_vm_address_t os_address;
  78 };
  79
  80 #ifdef DELAY_THREAD_POST_MORTEM
  81 static int pending_thread_post_mortem_count = 0;
  82 pthread_mutex_t thread_post_mortem_lock = PTHREAD_MUTEX_INITIALIZER;
  83 #endif
  84 static struct thread_post_mortem * volatile pending_thread_post_mortem = 0;
  85 #endif
  86
  87 int dynamic_values_bytes=TLS_SIZE*sizeof(lispobj);  /* same for all threads */
  88 struct thread *all_threads;
  89 extern struct interrupt_data * global_interrupt_data;
  90
  91 #ifdef LISP_FEATURE_SB_THREAD
  92 pthread_mutex_t all_threads_lock = PTHREAD_MUTEX_INITIALIZER;
  93 #ifdef LOCK_CREATE_THREAD
  94 static pthread_mutex_t create_thread_lock = PTHREAD_MUTEX_INITIALIZER;
  95 #endif
  96 #ifdef LISP_FEATURE_GCC_TLS
  97 __thread struct thread *current_thread;
  98 #endif
  99 pthread_key_t lisp_thread = 0;
 100 #endif
 101
 102 #if defined(LISP_FEATURE_X86) || defined(LISP_FEATURE_X86_64)
 103 extern lispobj call_into_lisp_first_time(lispobj fun, lispobj *args, int nargs);
 104 #endif
 105
 106 static void
 107 link_thread(struct thread *th)
 108 {
 109     if (all_threads) all_threads->prev=th;
 110     th->next=all_threads;
 111     th->prev=0;
 112     all_threads=th;
 113 }
 114
 115 #ifdef LISP_FEATURE_SB_THREAD
 116 static void
 117 unlink_thread(struct thread *th)
 118 {
 119     if (th->prev)
 120         th->prev->next = th->next;
 121     else
 122         all_threads = th->next;
 123     if (th->next)
 124         th->next->prev = th->prev;
 125 }
 126
 127 /* Only access thread state with blockables blocked. */
 128 lispobj
 129 thread_state(struct thread *thread)
 130 {
 131     lispobj state;
 132     sigset_t old;
 133     block_blockable_signals(NULL, &old);
 134     os_sem_wait(thread->state_sem, "thread_state");
 135     state = thread->state;
 136     os_sem_post(thread->state_sem, "thread_state");
 137     thread_sigmask(SIG_SETMASK, &old, NULL);
 138     return state;
 139 }
 140
 141 void
 142 set_thread_state(struct thread *thread, lispobj state)
 143 {
 144     int i, waitcount = 0;
 145     sigset_t old;
 146     block_blockable_signals(NULL, &old);
 147     os_sem_wait(thread->state_sem, "set_thread_state");
 148     if (thread->state != state) {
 149         if ((STATE_STOPPED==state) ||
 150             (STATE_DEAD==state)) {
 151             waitcount = thread->state_not_running_waitcount;
 152             thread->state_not_running_waitcount = 0;
 153             for (i=0; i<waitcount; i++)
 154                 os_sem_post(thread->state_not_running_sem, "set_thread_state (not running)");
 155         }
 156         if ((STATE_RUNNING==state) ||
 157             (STATE_DEAD==state)) {
 158             waitcount = thread->state_not_stopped_waitcount;
 159             thread->state_not_stopped_waitcount = 0;
 160             for (i=0; i<waitcount; i++)
 161                 os_sem_post(thread->state_not_stopped_sem, "set_thread_state (not stopped)");
 162         }
 163         thread->state = state;
 164     }
 165     os_sem_post(thread->state_sem, "set_thread_state");
 166     thread_sigmask(SIG_SETMASK, &old, NULL);
 167 }
 168
 169 void
 170 wait_for_thread_state_change(struct thread *thread, lispobj state)
 171 {
 172     sigset_t old;
 173     os_sem_t *wait_sem;
 174     block_blockable_signals(NULL, &old);
 175   start:
 176     os_sem_wait(thread->state_sem, "wait_for_thread_state_change");
 177     if (thread->state == state) {
 178         switch (state) {
 179         case STATE_RUNNING:
 180             wait_sem = thread->state_not_running_sem;
 181             thread->state_not_running_waitcount++;
 182             break;
 183         case STATE_STOPPED:
 184             wait_sem = thread->state_not_stopped_sem;
 185             thread->state_not_stopped_waitcount++;
 186             break;
 187         default:
 188             lose("Invalid state in wait_for_thread_state_change: "OBJ_FMTX"\n", state);
 189         }
 190     } else {
 191         wait_sem = NULL;
 192     }
 193     os_sem_post(thread->state_sem, "wait_for_thread_state_change");
 194     if (wait_sem) {
 195         os_sem_wait(wait_sem, "wait_for_thread_state_change");
 196         goto start;
 197     }
 198     thread_sigmask(SIG_SETMASK, &old, NULL);
 199 }
 200 #endif
 201
 202 static int
 203 initial_thread_trampoline(struct thread *th)
 204 {
 205     lispobj function;
 206 #if defined(LISP_FEATURE_X86) || defined(LISP_FEATURE_X86_64)
 207     lispobj *args = NULL;
 208 #endif
 209 #ifdef LISP_FEATURE_SB_THREAD
 210     pthread_setspecific(lisp_thread, (void *)1);
 211 #endif
 212 #if defined(THREADS_USING_GCSIGNAL) && defined(LISP_FEATURE_PPC)
 213     /* SIG_STOP_FOR_GC defaults to blocked on PPC? */
 214     unblock_gc_signals(0,0);
 215 #endif
 216     function = th->no_tls_value_marker;
 217     th->no_tls_value_marker = NO_TLS_VALUE_MARKER_WIDETAG;
 218     if(arch_os_thread_init(th)==0) return 1;
 219 #ifdef LISP_FEATURE_SB_SAFEPOINT
 220     pthread_mutex_lock(thread_qrl(th));
 221 #endif
 222     link_thread(th);
 223     th->os_thread=thread_self();
 224 #ifndef LISP_FEATURE_WIN32
 225     protect_control_stack_hard_guard_page(1, NULL);
 226     protect_binding_stack_hard_guard_page(1, NULL);
 227     protect_alien_stack_hard_guard_page(1, NULL);
 228     protect_control_stack_guard_page(1, NULL);
 229     protect_binding_stack_guard_page(1, NULL);
 230     protect_alien_stack_guard_page(1, NULL);
 231 #endif
 232
 233 #if defined(LISP_FEATURE_X86) || defined(LISP_FEATURE_X86_64)
 234     return call_into_lisp_first_time(function,args,0);
 235 #else
 236     return funcall0(function);
 237 #endif
 238 }
 239
 240 #ifdef LISP_FEATURE_SB_THREAD
 241 /* THREAD POST MORTEM CLEANUP
 242  *
 243  * Memory allocated for the thread stacks cannot be reclaimed while
 244  * the thread is still alive, so we need a mechanism for post mortem
 245  * cleanups. FIXME: We actually have three, for historical reasons as
 246  * the saying goes. Do we really need three? Nikodemus guesses that
 247  * not anymore, now that we properly call pthread_attr_destroy before
 248  * freeing the stack. */
 249
 250 static struct thread_post_mortem *
 251 plan_thread_post_mortem(struct thread *corpse)
 252 {
 253     if (corpse) {
 254         struct thread_post_mortem *post_mortem = malloc(sizeof(struct thread_post_mortem));
 255         gc_assert(post_mortem);
 256         post_mortem->os_thread = corpse->os_thread;
 257         post_mortem->os_attr = corpse->os_attr;
 258         post_mortem->os_address = corpse->os_address;
 259 #ifdef DELAY_THREAD_POST_MORTEM
 260         post_mortem->next = NULL;
 261 #endif
 262         return post_mortem;
 263     } else {
 264         /* FIXME: When does this happen? */
 265         return NULL;
 266     }
 267 }
 268
 269 static void
 270 perform_thread_post_mortem(struct thread_post_mortem *post_mortem)
 271 {
 272 #ifdef CREATE_POST_MORTEM_THREAD
 273     pthread_detach(pthread_self());
 274 #endif
 275     if (post_mortem) {
 276         gc_assert(!pthread_join(post_mortem->os_thread, NULL));
 277         gc_assert(!pthread_attr_destroy(post_mortem->os_attr));
 278         free(post_mortem->os_attr);
 279         os_invalidate(post_mortem->os_address, THREAD_STRUCT_SIZE);
 280         free(post_mortem);
 281     }
 282 }
 283
 284 static void
 285 schedule_thread_post_mortem(struct thread *corpse)
 286 {
 287     struct thread_post_mortem *post_mortem = NULL;
 288     if (corpse) {
 289         post_mortem = plan_thread_post_mortem(corpse);
 290
 291 #ifdef DELAY_THREAD_POST_MORTEM
 292         pthread_mutex_lock(&thread_post_mortem_lock);
 293         /* First stick the new post mortem to the end of the queue. */
 294         if (pending_thread_post_mortem) {
 295             struct thread_post_mortem *next = pending_thread_post_mortem;
 296             while (next->next) {
 297                 next = next->next;
 298             }
 299             next->next = post_mortem;
 300         } else {
 301             pending_thread_post_mortem = post_mortem;
 302         }
 303         /* Then, if there are enough things in the queue, clean up one
 304          * from the head -- or increment the count, and null out the
 305          * post_mortem we have. */
 306         if (pending_thread_post_mortem_count > DELAY_THREAD_POST_MORTEM) {
 307             post_mortem = pending_thread_post_mortem;
 308             pending_thread_post_mortem = post_mortem->next;
 309         } else {
 310             pending_thread_post_mortem_count++;
 311             post_mortem = NULL;
 312         }
 313         pthread_mutex_unlock(&thread_post_mortem_lock);
 314         /* Finally run, the cleanup, if any. */
 315         perform_thread_post_mortem(post_mortem);
 316 #elif defined(CREATE_POST_MORTEM_THREAD)
 317         gc_assert(!pthread_create(&thread, NULL, perform_thread_post_mortem, post_mortem));
 318 #else
 319         post_mortem = (struct thread_post_mortem *)
 320             swap_lispobjs((lispobj *)(void *)&pending_thread_post_mortem,
 321                           (lispobj)post_mortem);
 322         perform_thread_post_mortem(post_mortem);
 323 #endif
 324     }
 325 }
 326
 327 /* this is the first thing that runs in the child (which is why the
 328  * silly calling convention).  Basically it calls the user's requested
 329  * lisp function after doing arch_os_thread_init and whatever other
 330  * bookkeeping needs to be done
 331  */
 332 int
 333 new_thread_trampoline(struct thread *th)
 334 {
 335     lispobj function;
 336     int result, lock_ret;
 337
 338     FSHOW((stderr,"/creating thread %lu\n", thread_self()));
 339     check_deferrables_blocked_or_lose(0);
 340 #ifndef LISP_FEATURE_SB_SAFEPOINT
 341     check_gc_signals_unblocked_or_lose(0);
 342 #endif
 343     pthread_setspecific(lisp_thread, (void *)1);
 344     function = th->no_tls_value_marker;
 345     th->no_tls_value_marker = NO_TLS_VALUE_MARKER_WIDETAG;
 346     if(arch_os_thread_init(th)==0) {
 347         /* FIXME: handle error */
 348         lose("arch_os_thread_init failed\n");
 349     }
 350
 351     th->os_thread=thread_self();
 352     protect_control_stack_guard_page(1, NULL);
 353     protect_binding_stack_guard_page(1, NULL);
 354     protect_alien_stack_guard_page(1, NULL);
 355     /* Since GC can only know about this thread from the all_threads
 356      * list and we're just adding this thread to it, there is no
 357      * danger of deadlocking even with SIG_STOP_FOR_GC blocked (which
 358      * it is not). */
 359 #ifdef LISP_FEATURE_SB_SAFEPOINT
 360     *th->csp_around_foreign_call = (lispobj)&function;
 361     pthread_mutex_lock(thread_qrl(th));
 362 #endif
 363     lock_ret = pthread_mutex_lock(&all_threads_lock);
 364     gc_assert(lock_ret == 0);
 365     link_thread(th);
 366     lock_ret = pthread_mutex_unlock(&all_threads_lock);
 367     gc_assert(lock_ret == 0);
 368
 369     /* Kludge: Changed the order of some steps between the safepoint/
 370      * non-safepoint versions of this code.  Can we unify this more?
 371      */
 372 #ifdef LISP_FEATURE_SB_SAFEPOINT
 373     WITH_GC_AT_SAFEPOINTS_ONLY() {
 374         result = funcall0(function);
 375         block_blockable_signals(0, 0);
 376         gc_alloc_update_page_tables(BOXED_PAGE_FLAG, &th->alloc_region);
 377     }
 378     lock_ret = pthread_mutex_lock(&all_threads_lock);
 379     gc_assert(lock_ret == 0);
 380     unlink_thread(th);
 381     lock_ret = pthread_mutex_unlock(&all_threads_lock);
 382     gc_assert(lock_ret == 0);
 383     pthread_mutex_unlock(thread_qrl(th));
 384     set_thread_state(th,STATE_DEAD);
 385 #else
 386     result = funcall0(function);
 387
 388     /* Block GC */
 389     block_blockable_signals(0, 0);
 390     set_thread_state(th, STATE_DEAD);
 391
 392     /* SIG_STOP_FOR_GC is blocked and GC might be waiting for this
 393      * thread, but since we are already dead it won't wait long. */
 394     lock_ret = pthread_mutex_lock(&all_threads_lock);
 395     gc_assert(lock_ret == 0);
 396
 397     gc_alloc_update_page_tables(BOXED_PAGE_FLAG, &th->alloc_region);
 398     unlink_thread(th);
 399     pthread_mutex_unlock(&all_threads_lock);
 400     gc_assert(lock_ret == 0);
 401 #endif
 402
 403     if(th->tls_cookie>=0) arch_os_thread_cleanup(th);
 404     os_sem_destroy(th->state_sem);
 405     os_sem_destroy(th->state_not_running_sem);
 406     os_sem_destroy(th->state_not_stopped_sem);
 407
 408     os_invalidate((os_vm_address_t)th->interrupt_data,
 409                   (sizeof (struct interrupt_data)));
 410
 411 #ifdef LISP_FEATURE_MACH_EXCEPTION_HANDLER
 412     mach_lisp_thread_destroy(th);
 413 #endif
 414
 415     schedule_thread_post_mortem(th);
 416     FSHOW((stderr,"/exiting thread %lu\n", thread_self()));
 417     return result;
 418 }
 419
 420 #endif /* LISP_FEATURE_SB_THREAD */
 421
 422 static void
 423 free_thread_struct(struct thread *th)
 424 {
 425     if (th->interrupt_data)
 426         os_invalidate((os_vm_address_t) th->interrupt_data,
 427                       (sizeof (struct interrupt_data)));
 428     os_invalidate((os_vm_address_t) th->os_address,
 429                   THREAD_STRUCT_SIZE);
 430 }
 431
 432 #ifdef LISP_FEATURE_SB_THREAD
 433 /* FIXME: should be MAX_INTERRUPTS -1 ? */
 434 const unsigned int tls_index_start =
 435   MAX_INTERRUPTS + sizeof(struct thread)/sizeof(lispobj);
 436 #endif
 437
 438 /* this is called from any other thread to create the new one, and
 439  * initialize all parts of it that can be initialized from another
 440  * thread
 441  */
 442
 443 static struct thread *
 444 create_thread_struct(lispobj initial_function) {
 445     union per_thread_data *per_thread;
 446     struct thread *th=0;        /*  subdue gcc */
 447     void *spaces=0;
 448     void *aligned_spaces=0;
 449 #ifdef LISP_FEATURE_SB_THREAD
 450     unsigned int i;
 451 #endif
 452
 453     /* May as well allocate all the spaces at once: it saves us from
 454      * having to decide what to do if only some of the allocations
 455      * succeed. SPACES must be appropriately aligned, since the GC
 456      * expects the control stack to start at a page boundary -- and
 457      * the OS may have even more rigorous requirements. We can't rely
 458      * on the alignment passed from os_validate, since that might
 459      * assume the current (e.g. 4k) pagesize, while we calculate with
 460      * the biggest (e.g. 64k) pagesize allowed by the ABI. */
 461     spaces=os_validate(0, THREAD_STRUCT_SIZE);
 462     if(!spaces)
 463         return NULL;
 464     /* Aligning up is safe as THREAD_STRUCT_SIZE has
 465      * THREAD_ALIGNMENT_BYTES padding. */
 466     aligned_spaces = (void *)((((unsigned long)(char *)spaces)
 467                                + THREAD_ALIGNMENT_BYTES-1)
 468                               &~(unsigned long)(THREAD_ALIGNMENT_BYTES-1));
 469     void* csp_page=
 470         (aligned_spaces+
 471          thread_control_stack_size+
 472          BINDING_STACK_SIZE+
 473          ALIEN_STACK_SIZE);
 474     per_thread=(union per_thread_data *)
 475         (csp_page + THREAD_CSP_PAGE_SIZE);
 476     struct nonpointer_thread_data *nonpointer_data
 477         = (void *) &per_thread->dynamic_values[TLS_SIZE];
 478
 479 #ifdef LISP_FEATURE_SB_THREAD
 480     for(i = 0; i < (dynamic_values_bytes / sizeof(lispobj)); i++)
 481         per_thread->dynamic_values[i] = NO_TLS_VALUE_MARKER_WIDETAG;
 482     if (all_threads == 0) {
 483         if(SymbolValue(FREE_TLS_INDEX,0)==UNBOUND_MARKER_WIDETAG) {
 484             SetSymbolValue(FREE_TLS_INDEX,tls_index_start << WORD_SHIFT,0);
 485             SetSymbolValue(TLS_INDEX_LOCK,make_fixnum(0),0);
 486         }
 487 #define STATIC_TLS_INIT(sym,field) \
 488   ((struct symbol *)(sym-OTHER_POINTER_LOWTAG))->tls_index= \
 489   (THREAD_SLOT_OFFSET_WORDS(field) << WORD_SHIFT)
 490
 491         STATIC_TLS_INIT(BINDING_STACK_START,binding_stack_start);
 492 #ifdef BINDING_STACK_POINTER
 493         STATIC_TLS_INIT(BINDING_STACK_POINTER,binding_stack_pointer);
 494 #endif
 495         STATIC_TLS_INIT(CONTROL_STACK_START,control_stack_start);
 496         STATIC_TLS_INIT(CONTROL_STACK_END,control_stack_end);
 497 #ifdef ALIEN_STACK
 498         STATIC_TLS_INIT(ALIEN_STACK,alien_stack_pointer);
 499 #endif
 500 #if defined(LISP_FEATURE_X86) || defined (LISP_FEATURE_X86_64)
 501         STATIC_TLS_INIT(PSEUDO_ATOMIC_BITS,pseudo_atomic_bits);
 502 #endif
 503 #undef STATIC_TLS_INIT
 504     }
 505 #endif
 506
 507     th=&per_thread->thread;
 508     th->os_address = spaces;
 509     th->control_stack_start = aligned_spaces;
 510     th->binding_stack_start=
 511         (lispobj*)((void*)th->control_stack_start+thread_control_stack_size);
 512     th->control_stack_end = th->binding_stack_start;
 513     th->control_stack_guard_page_protected = T;
 514     th->alien_stack_start=
 515         (lispobj*)((void*)th->binding_stack_start+BINDING_STACK_SIZE);
 516     set_binding_stack_pointer(th,th->binding_stack_start);
 517     th->this=th;
 518     th->os_thread=0;
 519
 520 #ifdef LISP_FEATURE_SB_SAFEPOINT
 521     th->pc_around_foreign_call = 0;
 522     th->csp_around_foreign_call = csp_page;
 523 #endif
 524
 525 #ifdef LISP_FEATURE_SB_THREAD
 526     /* Contrary to the "allocate all the spaces at once" comment above,
 527      * the os_attr is allocated separately.  We cannot put it into the
 528      * nonpointer data, because it's used for post_mortem and freed
 529      * separately */
 530     th->os_attr=malloc(sizeof(pthread_attr_t));
 531     th->nonpointer_data = nonpointer_data;
 532     th->state_sem=&nonpointer_data->state_sem;
 533     th->state_not_running_sem=&nonpointer_data->state_not_running_sem;
 534     th->state_not_stopped_sem=&nonpointer_data->state_not_stopped_sem;
 535     th->state_not_running_waitcount = 0;
 536     th->state_not_stopped_waitcount = 0;
 537     os_sem_init(th->state_sem, 1);
 538     os_sem_init(th->state_not_running_sem, 0);
 539     os_sem_init(th->state_not_stopped_sem, 0);
 540 # ifdef LISP_FEATURE_SB_SAFEPOINT
 541     pthread_mutex_init(thread_qrl(th), NULL);
 542 # endif
 543 #endif
 544     th->state=STATE_RUNNING;
 545 #ifdef LISP_FEATURE_STACK_GROWS_DOWNWARD_NOT_UPWARD
 546     th->alien_stack_pointer=((void *)th->alien_stack_start
 547                              + ALIEN_STACK_SIZE-N_WORD_BYTES);
 548 #else
 549     th->alien_stack_pointer=((void *)th->alien_stack_start);
 550 #endif
 551 #if defined(LISP_FEATURE_X86) || defined (LISP_FEATURE_X86_64) || defined(LISP_FEATURE_SB_THREAD)
 552     th->pseudo_atomic_bits=0;
 553 #endif
 554 #ifdef LISP_FEATURE_GENCGC
 555     gc_set_region_empty(&th->alloc_region);
 556 #endif
 557 #ifdef LISP_FEATURE_SB_THREAD
 558     /* This parallels the same logic in globals.c for the
 559      * single-threaded foreign_function_call_active, KLUDGE and
 560      * all. */
 561 #if defined(LISP_FEATURE_X86) || defined(LISP_FEATURE_X86_64)
 562     th->foreign_function_call_active = 0;
 563 #else
 564     th->foreign_function_call_active = 1;
 565 #endif
 566 #endif
 567
 568 #ifndef LISP_FEATURE_SB_THREAD
 569     /* the tls-points-into-struct-thread trick is only good for threaded
 570      * sbcl, because unithread sbcl doesn't have tls.  So, we copy the
 571      * appropriate values from struct thread here, and make sure that
 572      * we use the appropriate SymbolValue macros to access any of the
 573      * variable quantities from the C runtime.  It's not quite OAOOM,
 574      * it just feels like it */
 575     SetSymbolValue(BINDING_STACK_START,(lispobj)th->binding_stack_start,th);
 576     SetSymbolValue(CONTROL_STACK_START,(lispobj)th->control_stack_start,th);
 577     SetSymbolValue(CONTROL_STACK_END,(lispobj)th->control_stack_end,th);
 578 #if defined(LISP_FEATURE_X86) || defined (LISP_FEATURE_X86_64)
 579     SetSymbolValue(ALIEN_STACK,(lispobj)th->alien_stack_pointer,th);
 580     SetSymbolValue(PSEUDO_ATOMIC_BITS,(lispobj)th->pseudo_atomic_bits,th);
 581 #endif
 582 #endif
 583     bind_variable(CURRENT_CATCH_BLOCK,make_fixnum(0),th);
 584     bind_variable(CURRENT_UNWIND_PROTECT_BLOCK,make_fixnum(0),th);
 585     bind_variable(FREE_INTERRUPT_CONTEXT_INDEX,make_fixnum(0),th);
 586     bind_variable(INTERRUPT_PENDING, NIL,th);
 587     bind_variable(INTERRUPTS_ENABLED,T,th);
 588     bind_variable(ALLOW_WITH_INTERRUPTS,T,th);
 589     bind_variable(GC_PENDING,NIL,th);
 590     bind_variable(ALLOC_SIGNAL,NIL,th);
 591 #ifdef PINNED_OBJECTS
 592     bind_variable(PINNED_OBJECTS,NIL,th);
 593 #endif
 594 #ifdef LISP_FEATURE_SB_THREAD
 595     bind_variable(STOP_FOR_GC_PENDING,NIL,th);
 596 #endif
 597 #if defined(LISP_FEATURE_SB_SAFEPOINT)
 598     bind_variable(GC_SAFE,NIL,th);
 599     bind_variable(IN_SAFEPOINT,NIL,th);
 600 #endif
 601 #ifdef LISP_FEATURE_SB_THRUPTION
 602     bind_variable(THRUPTION_PENDING,NIL,th);
 603     bind_variable(RESTART_CLUSTERS,NIL,th);
 604 #endif
 605 #ifndef LISP_FEATURE_C_STACK_IS_CONTROL_STACK
 606     access_control_stack_pointer(th)=th->control_stack_start;
 607 #endif
 608
 609     th->interrupt_data = (struct interrupt_data *)
 610         os_validate(0,(sizeof (struct interrupt_data)));
 611     if (!th->interrupt_data) {
 612         free_thread_struct(th);
 613         return 0;
 614     }
 615     th->interrupt_data->pending_handler = 0;
 616     th->interrupt_data->gc_blocked_deferrables = 0;
 617 #ifdef LISP_FEATURE_PPC
 618     th->interrupt_data->allocation_trap_context = 0;
 619 #endif
 620     th->no_tls_value_marker=initial_function;
 621
 622     th->stepping = NIL;
 623     return th;
 624 }
 625
 626 void create_initial_thread(lispobj initial_function) {
 627     struct thread *th=create_thread_struct(initial_function);
 628 #ifdef LISP_FEATURE_SB_THREAD
 629     pthread_key_create(&lisp_thread, 0);
 630 #endif
 631     if(th) {
 632         initial_thread_trampoline(th); /* no return */
 633     } else lose("can't create initial thread\n");
 634 }
 635
 636 #ifdef LISP_FEATURE_SB_THREAD
 637
 638 #ifndef __USE_XOPEN2K
 639 extern int pthread_attr_setstack (pthread_attr_t *__attr, void *__stackaddr,
 640                                   size_t __stacksize);
 641 #endif
 642
 643 boolean create_os_thread(struct thread *th,os_thread_t *kid_tid)
 644 {
 645     /* The new thread inherits the restrictive signal mask set here,
 646      * and enables signals again when it is set up properly. */
 647     sigset_t oldset;
 648     boolean r=1;
 649     int retcode = 0, initcode;
 650
 651     FSHOW_SIGNAL((stderr,"/create_os_thread: creating new thread\n"));
 652
 653     /* Blocking deferrable signals is enough, no need to block
 654      * SIG_STOP_FOR_GC because the child process is not linked onto
 655      * all_threads until it's ready. */
 656     block_deferrable_signals(0, &oldset);
 657
 658 #ifdef LOCK_CREATE_THREAD
 659     retcode = pthread_mutex_lock(&create_thread_lock);
 660     gc_assert(retcode == 0);
 661     FSHOW_SIGNAL((stderr,"/create_os_thread: got lock\n"));
 662 #endif
 663
 664     if((initcode = pthread_attr_init(th->os_attr)) ||
 665        /* call_into_lisp_first_time switches the stack for the initial
 666         * thread. For the others, we use this. */
 667        (pthread_attr_setstack(th->os_attr,th->control_stack_start,
 668                               thread_control_stack_size)) ||
 669        (retcode = pthread_create
 670         (kid_tid,th->os_attr,(void *(*)(void *))new_thread_trampoline,th))) {
 671         FSHOW_SIGNAL((stderr, "init = %d\n", initcode));
 672         FSHOW_SIGNAL((stderr, "pthread_create returned %d, errno %d\n",
 673                       retcode, errno));
 674         if(retcode < 0) {
 675             perror("create_os_thread");
 676         }
 677         r=0;
 678     }
 679
 680 #ifdef LOCK_CREATE_THREAD
 681     retcode = pthread_mutex_unlock(&create_thread_lock);
 682     gc_assert(retcode == 0);
 683     FSHOW_SIGNAL((stderr,"/create_os_thread: released lock\n"));
 684 #endif
 685     thread_sigmask(SIG_SETMASK,&oldset,0);
 686     return r;
 687 }
 688
 689 os_thread_t create_thread(lispobj initial_function) {
 690     struct thread *th, *thread = arch_os_get_current_thread();
 691     os_thread_t kid_tid = 0;
 692
 693     /* Must defend against async unwinds. */
 694     if (SymbolValue(INTERRUPTS_ENABLED, thread) != NIL)
 695         lose("create_thread is not safe when interrupts are enabled.\n");
 696
 697     /* Assuming that a fresh thread struct has no lisp objects in it,
 698      * linking it to all_threads can be left to the thread itself
 699      * without fear of gc lossage. initial_function violates this
 700      * assumption and must stay pinned until the child starts up. */
 701     th = create_thread_struct(initial_function);
 702     if (th && !create_os_thread(th,&kid_tid)) {
 703         free_thread_struct(th);
 704         kid_tid = 0;
 705     }
 706     return kid_tid;
 707 }
 708
 709 /* stopping the world is a two-stage process.  From this thread we signal
 710  * all the others with SIG_STOP_FOR_GC.  The handler for this signal does
 711  * the usual pseudo-atomic checks (we don't want to stop a thread while
 712  * it's in the middle of allocation) then waits for another SIG_STOP_FOR_GC.
 713  */
 714 /*
 715  * (With SB-SAFEPOINT, see the definitions in safepoint.c instead.)
 716  */
 717 #ifndef LISP_FEATURE_SB_SAFEPOINT
 718
 719 /* To avoid deadlocks when gc stops the world all clients of each
 720  * mutex must enable or disable SIG_STOP_FOR_GC for the duration of
 721  * holding the lock, but they must agree on which. */
 722 void gc_stop_the_world()
 723 {
 724     struct thread *p,*th=arch_os_get_current_thread();
 725     int status, lock_ret;
 726 #ifdef LOCK_CREATE_THREAD
 727     /* KLUDGE: Stopping the thread during pthread_create() causes deadlock
 728      * on FreeBSD. */
 729     FSHOW_SIGNAL((stderr,"/gc_stop_the_world:waiting on create_thread_lock\n"));
 730     lock_ret = pthread_mutex_lock(&create_thread_lock);
 731     gc_assert(lock_ret == 0);
 732     FSHOW_SIGNAL((stderr,"/gc_stop_the_world:got create_thread_lock\n"));
 733 #endif
 734     FSHOW_SIGNAL((stderr,"/gc_stop_the_world:waiting on lock\n"));
 735     /* keep threads from starting while the world is stopped. */
 736     lock_ret = pthread_mutex_lock(&all_threads_lock);      \
 737     gc_assert(lock_ret == 0);
 738
 739     FSHOW_SIGNAL((stderr,"/gc_stop_the_world:got lock\n"));
 740     /* stop all other threads by sending them SIG_STOP_FOR_GC */
 741     for(p=all_threads; p; p=p->next) {
 742         gc_assert(p->os_thread != 0);
 743         FSHOW_SIGNAL((stderr,"/gc_stop_the_world: thread=%lu, state=%x\n",
 744                       p->os_thread, thread_state(p)));
 745         if((p!=th) && ((thread_state(p)==STATE_RUNNING))) {
 746             FSHOW_SIGNAL((stderr,"/gc_stop_the_world: suspending thread %lu\n",
 747                           p->os_thread));
 748             /* We already hold all_thread_lock, P can become DEAD but
 749              * cannot exit, ergo it's safe to use pthread_kill. */
 750             status=pthread_kill(p->os_thread,SIG_STOP_FOR_GC);
 751             if (status==ESRCH) {
 752                 /* This thread has exited. */
 753                 gc_assert(thread_state(p)==STATE_DEAD);
 754             } else if (status) {
 755                 lose("cannot send suspend thread=%lu: %d, %s\n",
 756                      p->os_thread,status,strerror(status));
 757             }
 758         }
 759     }
 760     FSHOW_SIGNAL((stderr,"/gc_stop_the_world:signals sent\n"));
 761     for(p=all_threads;p;p=p->next) {
 762         if (p!=th) {
 763             FSHOW_SIGNAL
 764                 ((stderr,
 765                   "/gc_stop_the_world: waiting for thread=%lu: state=%x\n",
 766                   p->os_thread, thread_state(p)));
 767             wait_for_thread_state_change(p, STATE_RUNNING);
 768             if (p->state == STATE_RUNNING)
 769                 lose("/gc_stop_the_world: unexpected state");
 770         }
 771     }
 772     FSHOW_SIGNAL((stderr,"/gc_stop_the_world:end\n"));
 773 }
 774
 775 void gc_start_the_world()
 776 {
 777     struct thread *p,*th=arch_os_get_current_thread();
 778     int lock_ret;
 779     /* if a resumed thread creates a new thread before we're done with
 780      * this loop, the new thread will get consed on the front of
 781      * all_threads, but it won't have been stopped so won't need
 782      * restarting */
 783     FSHOW_SIGNAL((stderr,"/gc_start_the_world:begin\n"));
 784     for(p=all_threads;p;p=p->next) {
 785         gc_assert(p->os_thread!=0);
 786         if (p!=th) {
 787             lispobj state = thread_state(p);
 788             if (state != STATE_DEAD) {
 789                 if(state != STATE_STOPPED) {
 790                     lose("gc_start_the_world: wrong thread state is %d\n",
 791                          fixnum_value(state));
 792                 }
 793                 FSHOW_SIGNAL((stderr, "/gc_start_the_world: resuming %lu\n",
 794                               p->os_thread));
 795                 set_thread_state(p, STATE_RUNNING);
 796             }
 797         }
 798     }
 799
 800     lock_ret = pthread_mutex_unlock(&all_threads_lock);
 801     gc_assert(lock_ret == 0);
 802 #ifdef LOCK_CREATE_THREAD
 803     lock_ret = pthread_mutex_unlock(&create_thread_lock);
 804     gc_assert(lock_ret == 0);
 805 #endif
 806
 807     FSHOW_SIGNAL((stderr,"/gc_start_the_world:end\n"));
 808 }
 809
 810 #endif /* !LISP_FEATURE_SB_SAFEPOINT */
 811 #endif /* !LISP_FEATURE_SB_THREAD */
 812
 813 int
 814 thread_yield()
 815 {
 816 #ifdef LISP_FEATURE_SB_THREAD
 817     return sched_yield();
 818 #else
 819     return 0;
 820 #endif
 821 }
 822
 823 int
 824 wake_thread(os_thread_t os_thread)
 825 {
 826 #ifdef LISP_FEATURE_WIN32
 827 # define SIGPIPE 1
 828 #endif
 829 #if !defined(LISP_FEATURE_SB_THRUPTION) || defined(LISP_FEATURE_WIN32)
 830     return kill_safely(os_thread, SIGPIPE);
 831 #else
 832     return wake_thread_posix(os_thread);
 833 #endif
 834 }
 835
 836 /* If the thread id given does not belong to a running thread (it has
 837  * exited or never even existed) pthread_kill _may_ fail with ESRCH,
 838  * but it is also allowed to just segfault, see
 839  * <http://udrepper.livejournal.com/16844.html>.
 840  *
 841  * Relying on thread ids can easily backfire since ids are recycled
 842  * (NPTL recycles them extremely fast) so a signal can be sent to
 843  * another process if the one it was sent to exited.
 844  *
 845  * For these reasons, we must make sure that the thread is still alive
 846  * when the pthread_kill is called and return if the thread is
 847  * exiting.
 848  *
 849  * Note (DFL, 2011-06-22): At the time of writing, this function is only
 850  * used for INTERRUPT-THREAD, hence the wake_thread special-case for
 851  * Windows is OK. */
 852 int
 853 kill_safely(os_thread_t os_thread, int signal)
 854 {
 855     FSHOW_SIGNAL((stderr,"/kill_safely: %lu, %d\n", os_thread, signal));
 856     {
 857 #ifdef LISP_FEATURE_SB_THREAD
 858         sigset_t oldset;
 859         struct thread *thread;
 860         /* Frequent special case: resignalling to self.  The idea is
 861          * that leave_region safepoint will acknowledge the signal, so
 862          * there is no need to take locks, roll thread to safepoint
 863          * etc. */
 864         /* Kludge (on safepoint builds): At the moment, this isn't just
 865          * an optimization; rather it masks the fact that
 866          * gc_stop_the_world() grabs the all_threads mutex without
 867          * releasing it, and since we're not using recursive pthread
 868          * mutexes, the pthread_mutex_lock() around the all_threads loop
 869          * would go wrong.  Why are we running interruptions while
 870          * stopping the world though?  Test case is (:ASYNC-UNWIND
 871          * :SPECIALS), especially with s/10/100/ in both loops. */
 872         if (os_thread == pthread_self()) {
 873             pthread_kill(os_thread, signal);
 874             return 0;
 875         }
 876
 877         /* pthread_kill is not async signal safe and we don't want to be
 878          * interrupted while holding the lock. */
 879         block_deferrable_signals(0, &oldset);
 880         pthread_mutex_lock(&all_threads_lock);
 881         for (thread = all_threads; thread; thread = thread->next) {
 882             if (thread->os_thread == os_thread) {
 883                 int status = pthread_kill(os_thread, signal);
 884                 if (status)
 885                     lose("kill_safely: pthread_kill failed with %d\n", status);
 886                 break;
 887             }
 888         }
 889         pthread_mutex_unlock(&all_threads_lock);
 890         thread_sigmask(SIG_SETMASK,&oldset,0);
 891         if (thread)
 892             return 0;
 893         else
 894             return -1;
 895 #else
 896         int status;
 897         if (os_thread != 0)
 898             lose("kill_safely: who do you want to kill? %d?\n", os_thread);
 899         /* Dubious (as in don't know why it works) workaround for the
 900          * signal sometimes not being generated on darwin. */
 901 #ifdef LISP_FEATURE_DARWIN
 902         {
 903             sigset_t oldset;
 904             sigprocmask(SIG_BLOCK, &deferrable_sigset, &oldset);
 905             status = raise(signal);
 906             sigprocmask(SIG_SETMASK,&oldset,0);
 907         }
 908 #else
 909         status = raise(signal);
 910 #endif
 911         if (status == 0) {
 912             return 0;
 913         } else {
 914             lose("cannot raise signal %d, %d %s\n",
 915                  signal, status, strerror(errno));
 916         }
 917 #endif
 918     }
 919 }