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 "runtime.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 #if defined(LISP_FEATURE_WIN32) && defined(LISP_FEATURE_SB_THREAD)
  53 # define IMMEDIATE_POST_MORTEM
  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 # ifdef LISP_FEATURE_X86_64
 101     __attribute__((sysv_abi))
 102 # endif
 103     ;
 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 #ifndef LISP_FEATURE_SB_SAFEPOINT
 128 /* Only access thread state with blockables blocked. */
 129 lispobj
 130 thread_state(struct thread *thread)
 131 {
 132     lispobj state;
 133     sigset_t old;
 134     block_blockable_signals(NULL, &old);
 135     os_sem_wait(thread->state_sem, "thread_state");
 136     state = thread->state;
 137     os_sem_post(thread->state_sem, "thread_state");
 138     thread_sigmask(SIG_SETMASK, &old, NULL);
 139     return state;
 140 }
 141
 142 void
 143 set_thread_state(struct thread *thread, lispobj state)
 144 {
 145     int i, waitcount = 0;
 146     sigset_t old;
 147     block_blockable_signals(NULL, &old);
 148     os_sem_wait(thread->state_sem, "set_thread_state");
 149     if (thread->state != state) {
 150         if ((STATE_STOPPED==state) ||
 151             (STATE_DEAD==state)) {
 152             waitcount = thread->state_not_running_waitcount;
 153             thread->state_not_running_waitcount = 0;
 154             for (i=0; i<waitcount; i++)
 155                 os_sem_post(thread->state_not_running_sem, "set_thread_state (not running)");
 156         }
 157         if ((STATE_RUNNING==state) ||
 158             (STATE_DEAD==state)) {
 159             waitcount = thread->state_not_stopped_waitcount;
 160             thread->state_not_stopped_waitcount = 0;
 161             for (i=0; i<waitcount; i++)
 162                 os_sem_post(thread->state_not_stopped_sem, "set_thread_state (not stopped)");
 163         }
 164         thread->state = state;
 165     }
 166     os_sem_post(thread->state_sem, "set_thread_state");
 167     thread_sigmask(SIG_SETMASK, &old, NULL);
 168 }
 169
 170 void
 171 wait_for_thread_state_change(struct thread *thread, lispobj state)
 172 {
 173     sigset_t old;
 174     os_sem_t *wait_sem;
 175     block_blockable_signals(NULL, &old);
 176   start:
 177     os_sem_wait(thread->state_sem, "wait_for_thread_state_change");
 178     if (thread->state == state) {
 179         switch (state) {
 180         case STATE_RUNNING:
 181             wait_sem = thread->state_not_running_sem;
 182             thread->state_not_running_waitcount++;
 183             break;
 184         case STATE_STOPPED:
 185             wait_sem = thread->state_not_stopped_sem;
 186             thread->state_not_stopped_waitcount++;
 187             break;
 188         default:
 189             lose("Invalid state in wait_for_thread_state_change: "OBJ_FMTX"\n", state);
 190         }
 191     } else {
 192         wait_sem = NULL;
 193     }
 194     os_sem_post(thread->state_sem, "wait_for_thread_state_change");
 195     if (wait_sem) {
 196         os_sem_wait(wait_sem, "wait_for_thread_state_change");
 197         goto start;
 198     }
 199     thread_sigmask(SIG_SETMASK, &old, NULL);
 200 }
 201 #endif /* sb-safepoint */
 202 #endif /* sb-thread */
 203
 204 static int
 205 initial_thread_trampoline(struct thread *th)
 206 {
 207     lispobj function;
 208 #if defined(LISP_FEATURE_X86) || defined(LISP_FEATURE_X86_64)
 209     lispobj *args = NULL;
 210 #endif
 211 #ifdef LISP_FEATURE_SB_THREAD
 212     pthread_setspecific(lisp_thread, (void *)1);
 213 #endif
 214 #if defined(THREADS_USING_GCSIGNAL) && defined(LISP_FEATURE_PPC)
 215     /* SIG_STOP_FOR_GC defaults to blocked on PPC? */
 216     unblock_gc_signals(0,0);
 217 #endif
 218     function = th->no_tls_value_marker;
 219     th->no_tls_value_marker = NO_TLS_VALUE_MARKER_WIDETAG;
 220     if(arch_os_thread_init(th)==0) return 1;
 221     link_thread(th);
 222     th->os_thread=thread_self();
 223 #ifndef LISP_FEATURE_WIN32
 224     protect_control_stack_hard_guard_page(1, NULL);
 225 #endif
 226     protect_binding_stack_hard_guard_page(1, NULL);
 227     protect_alien_stack_hard_guard_page(1, NULL);
 228 #ifndef LISP_FEATURE_WIN32
 229     protect_control_stack_guard_page(1, NULL);
 230 #endif
 231     protect_binding_stack_guard_page(1, NULL);
 232     protect_alien_stack_guard_page(1, NULL);
 233
 234 #if defined(LISP_FEATURE_X86) || defined(LISP_FEATURE_X86_64)
 235     return call_into_lisp_first_time(function,args,0);
 236 #else
 237     return funcall0(function);
 238 #endif
 239 }
 240
 241 #ifdef LISP_FEATURE_SB_THREAD
 242
 243 # if defined(IMMEDIATE_POST_MORTEM)
 244
 245 /*
 246  * If this feature is set, we are running on a stack managed by the OS,
 247  * and no fancy delays are required for anything.  Just do it.
 248  */
 249 static void
 250 schedule_thread_post_mortem(struct thread *corpse)
 251 {
 252     pthread_detach(pthread_self());
 253     gc_assert(!pthread_attr_destroy(corpse->os_attr));
 254     free(corpse->os_attr);
 255 #if defined(LISP_FEATURE_WIN32)
 256     os_invalidate_free(corpse->os_address, THREAD_STRUCT_SIZE);
 257 #else
 258     os_invalidate(corpse->os_address, THREAD_STRUCT_SIZE);
 259 #endif
 260 }
 261
 262 # else
 263
 264 /* THREAD POST MORTEM CLEANUP
 265  *
 266  * Memory allocated for the thread stacks cannot be reclaimed while
 267  * the thread is still alive, so we need a mechanism for post mortem
 268  * cleanups. FIXME: We actually have three, for historical reasons as
 269  * the saying goes. Do we really need three? Nikodemus guesses that
 270  * not anymore, now that we properly call pthread_attr_destroy before
 271  * freeing the stack. */
 272
 273 static struct thread_post_mortem *
 274 plan_thread_post_mortem(struct thread *corpse)
 275 {
 276     if (corpse) {
 277         struct thread_post_mortem *post_mortem = malloc(sizeof(struct thread_post_mortem));
 278         gc_assert(post_mortem);
 279         post_mortem->os_thread = corpse->os_thread;
 280         post_mortem->os_attr = corpse->os_attr;
 281         post_mortem->os_address = corpse->os_address;
 282 #ifdef DELAY_THREAD_POST_MORTEM
 283         post_mortem->next = NULL;
 284 #endif
 285         return post_mortem;
 286     } else {
 287         /* FIXME: When does this happen? */
 288         return NULL;
 289     }
 290 }
 291
 292 static void
 293 perform_thread_post_mortem(struct thread_post_mortem *post_mortem)
 294 {
 295 #ifdef CREATE_POST_MORTEM_THREAD
 296     pthread_detach(pthread_self());
 297 #endif
 298     if (post_mortem) {
 299         gc_assert(!pthread_join(post_mortem->os_thread, NULL));
 300         gc_assert(!pthread_attr_destroy(post_mortem->os_attr));
 301         free(post_mortem->os_attr);
 302         os_invalidate(post_mortem->os_address, THREAD_STRUCT_SIZE);
 303         free(post_mortem);
 304     }
 305 }
 306
 307 static void
 308 schedule_thread_post_mortem(struct thread *corpse)
 309 {
 310     struct thread_post_mortem *post_mortem = NULL;
 311     if (corpse) {
 312         post_mortem = plan_thread_post_mortem(corpse);
 313
 314 #ifdef DELAY_THREAD_POST_MORTEM
 315         pthread_mutex_lock(&thread_post_mortem_lock);
 316         /* First stick the new post mortem to the end of the queue. */
 317         if (pending_thread_post_mortem) {
 318             struct thread_post_mortem *next = pending_thread_post_mortem;
 319             while (next->next) {
 320                 next = next->next;
 321             }
 322             next->next = post_mortem;
 323         } else {
 324             pending_thread_post_mortem = post_mortem;
 325         }
 326         /* Then, if there are enough things in the queue, clean up one
 327          * from the head -- or increment the count, and null out the
 328          * post_mortem we have. */
 329         if (pending_thread_post_mortem_count > DELAY_THREAD_POST_MORTEM) {
 330             post_mortem = pending_thread_post_mortem;
 331             pending_thread_post_mortem = post_mortem->next;
 332         } else {
 333             pending_thread_post_mortem_count++;
 334             post_mortem = NULL;
 335         }
 336         pthread_mutex_unlock(&thread_post_mortem_lock);
 337         /* Finally run, the cleanup, if any. */
 338         perform_thread_post_mortem(post_mortem);
 339 #elif defined(CREATE_POST_MORTEM_THREAD)
 340         gc_assert(!pthread_create(&thread, NULL, perform_thread_post_mortem, post_mortem));
 341 #else
 342         post_mortem = (struct thread_post_mortem *)
 343             swap_lispobjs((lispobj *)(void *)&pending_thread_post_mortem,
 344                           (lispobj)post_mortem);
 345         perform_thread_post_mortem(post_mortem);
 346 #endif
 347     }
 348 }
 349
 350 # endif /* !IMMEDIATE_POST_MORTEM */
 351
 352 /* this is the first thing that runs in the child (which is why the
 353  * silly calling convention).  Basically it calls the user's requested
 354  * lisp function after doing arch_os_thread_init and whatever other
 355  * bookkeeping needs to be done
 356  */
 357 int
 358 new_thread_trampoline(struct thread *th)
 359 {
 360     lispobj function;
 361     int result, lock_ret;
 362
 363     FSHOW((stderr,"/creating thread %lu\n", thread_self()));
 364     check_deferrables_blocked_or_lose(0);
 365 #ifndef LISP_FEATURE_SB_SAFEPOINT
 366     check_gc_signals_unblocked_or_lose(0);
 367 #endif
 368     pthread_setspecific(lisp_thread, (void *)1);
 369     function = th->no_tls_value_marker;
 370     th->no_tls_value_marker = NO_TLS_VALUE_MARKER_WIDETAG;
 371     if(arch_os_thread_init(th)==0) {
 372         /* FIXME: handle error */
 373         lose("arch_os_thread_init failed\n");
 374     }
 375
 376     th->os_thread=thread_self();
 377     protect_control_stack_guard_page(1, NULL);
 378     protect_binding_stack_guard_page(1, NULL);
 379     protect_alien_stack_guard_page(1, NULL);
 380     /* Since GC can only know about this thread from the all_threads
 381      * list and we're just adding this thread to it, there is no
 382      * danger of deadlocking even with SIG_STOP_FOR_GC blocked (which
 383      * it is not). */
 384 #ifdef LISP_FEATURE_SB_SAFEPOINT
 385     *th->csp_around_foreign_call = (lispobj)&function;
 386 #endif
 387     lock_ret = pthread_mutex_lock(&all_threads_lock);
 388     gc_assert(lock_ret == 0);
 389     link_thread(th);
 390     lock_ret = pthread_mutex_unlock(&all_threads_lock);
 391     gc_assert(lock_ret == 0);
 392
 393     /* Kludge: Changed the order of some steps between the safepoint/
 394      * non-safepoint versions of this code.  Can we unify this more?
 395      */
 396 #ifdef LISP_FEATURE_SB_SAFEPOINT
 397     gc_state_lock();
 398     gc_state_wait(GC_NONE);
 399     gc_state_unlock();
 400     WITH_GC_AT_SAFEPOINTS_ONLY() {
 401         result = funcall0(function);
 402         block_blockable_signals(0, 0);
 403         gc_alloc_update_page_tables(BOXED_PAGE_FLAG, &th->alloc_region);
 404     }
 405     lock_ret = pthread_mutex_lock(&all_threads_lock);
 406     gc_assert(lock_ret == 0);
 407     unlink_thread(th);
 408     lock_ret = pthread_mutex_unlock(&all_threads_lock);
 409     gc_assert(lock_ret == 0);
 410 #else
 411     result = funcall0(function);
 412
 413     /* Block GC */
 414     block_blockable_signals(0, 0);
 415     set_thread_state(th, STATE_DEAD);
 416
 417     /* SIG_STOP_FOR_GC is blocked and GC might be waiting for this
 418      * thread, but since we are already dead it won't wait long. */
 419     lock_ret = pthread_mutex_lock(&all_threads_lock);
 420     gc_assert(lock_ret == 0);
 421
 422     gc_alloc_update_page_tables(BOXED_PAGE_FLAG, &th->alloc_region);
 423     unlink_thread(th);
 424     pthread_mutex_unlock(&all_threads_lock);
 425     gc_assert(lock_ret == 0);
 426 #endif
 427
 428     if(th->tls_cookie>=0) arch_os_thread_cleanup(th);
 429 #ifndef LISP_FEATURE_SB_SAFEPOINT
 430     os_sem_destroy(th->state_sem);
 431     os_sem_destroy(th->state_not_running_sem);
 432     os_sem_destroy(th->state_not_stopped_sem);
 433 #endif
 434
 435 #if defined(LISP_FEATURE_WIN32)
 436     free((os_vm_address_t)th->interrupt_data);
 437 #else
 438     os_invalidate((os_vm_address_t)th->interrupt_data,
 439                   (sizeof (struct interrupt_data)));
 440 #endif
 441
 442 #ifdef LISP_FEATURE_MACH_EXCEPTION_HANDLER
 443     mach_lisp_thread_destroy(th);
 444 #endif
 445
 446 #if defined(LISP_FEATURE_WIN32)
 447     int i;
 448     for (i = 0; i<
 449              (int) (sizeof(th->private_events.events)/
 450                     sizeof(th->private_events.events[0])); ++i) {
 451       CloseHandle(th->private_events.events[i]);
 452     }
 453     TlsSetValue(OUR_TLS_INDEX,NULL);
 454 #endif
 455
 456     schedule_thread_post_mortem(th);
 457     FSHOW((stderr,"/exiting thread %lu\n", thread_self()));
 458     return result;
 459 }
 460
 461 #endif /* LISP_FEATURE_SB_THREAD */
 462
 463 static void
 464 free_thread_struct(struct thread *th)
 465 {
 466 #if defined(LISP_FEATURE_WIN32)
 467     if (th->interrupt_data) {
 468         os_invalidate_free((os_vm_address_t) th->interrupt_data,
 469                       (sizeof (struct interrupt_data)));
 470     }
 471     os_invalidate_free((os_vm_address_t) th->os_address,
 472                   THREAD_STRUCT_SIZE);
 473 #else
 474     if (th->interrupt_data)
 475         os_invalidate((os_vm_address_t) th->interrupt_data,
 476                       (sizeof (struct interrupt_data)));
 477     os_invalidate((os_vm_address_t) th->os_address,
 478                   THREAD_STRUCT_SIZE);
 479 #endif
 480 }
 481
 482 #ifdef LISP_FEATURE_SB_THREAD
 483 /* FIXME: should be MAX_INTERRUPTS -1 ? */
 484 const unsigned int tls_index_start =
 485   MAX_INTERRUPTS + sizeof(struct thread)/sizeof(lispobj);
 486 #endif
 487
 488 /* this is called from any other thread to create the new one, and
 489  * initialize all parts of it that can be initialized from another
 490  * thread
 491  */
 492
 493 static struct thread *
 494 create_thread_struct(lispobj initial_function) {
 495     union per_thread_data *per_thread;
 496     struct thread *th=0;        /*  subdue gcc */
 497     void *spaces=0;
 498     void *aligned_spaces=0;
 499 #if defined(LISP_FEATURE_SB_THREAD) || defined(LISP_FEATURE_WIN32)
 500     unsigned int i;
 501 #endif
 502
 503     /* May as well allocate all the spaces at once: it saves us from
 504      * having to decide what to do if only some of the allocations
 505      * succeed. SPACES must be appropriately aligned, since the GC
 506      * expects the control stack to start at a page boundary -- and
 507      * the OS may have even more rigorous requirements. We can't rely
 508      * on the alignment passed from os_validate, since that might
 509      * assume the current (e.g. 4k) pagesize, while we calculate with
 510      * the biggest (e.g. 64k) pagesize allowed by the ABI. */
 511     spaces=os_validate(0, THREAD_STRUCT_SIZE);
 512     if(!spaces)
 513         return NULL;
 514     /* Aligning up is safe as THREAD_STRUCT_SIZE has
 515      * THREAD_ALIGNMENT_BYTES padding. */
 516     aligned_spaces = (void *)((((uword_t)(char *)spaces)
 517                                + THREAD_ALIGNMENT_BYTES-1)
 518                               &~(uword_t)(THREAD_ALIGNMENT_BYTES-1));
 519     void* csp_page=
 520         (aligned_spaces+
 521          thread_control_stack_size+
 522          BINDING_STACK_SIZE+
 523          ALIEN_STACK_SIZE);
 524     per_thread=(union per_thread_data *)
 525         (csp_page + THREAD_CSP_PAGE_SIZE);
 526     struct nonpointer_thread_data *nonpointer_data
 527         = (void *) &per_thread->dynamic_values[TLS_SIZE];
 528
 529 #ifdef LISP_FEATURE_SB_THREAD
 530     for(i = 0; i < (dynamic_values_bytes / sizeof(lispobj)); i++)
 531         per_thread->dynamic_values[i] = NO_TLS_VALUE_MARKER_WIDETAG;
 532     if (all_threads == 0) {
 533         if(SymbolValue(FREE_TLS_INDEX,0)==UNBOUND_MARKER_WIDETAG) {
 534             SetSymbolValue(FREE_TLS_INDEX,tls_index_start << WORD_SHIFT,0);
 535             SetSymbolValue(TLS_INDEX_LOCK,make_fixnum(0),0);
 536         }
 537 #define STATIC_TLS_INIT(sym,field) \
 538   ((struct symbol *)(sym-OTHER_POINTER_LOWTAG))->tls_index= \
 539   (THREAD_SLOT_OFFSET_WORDS(field) << WORD_SHIFT)
 540
 541         STATIC_TLS_INIT(BINDING_STACK_START,binding_stack_start);
 542 #ifdef BINDING_STACK_POINTER
 543         STATIC_TLS_INIT(BINDING_STACK_POINTER,binding_stack_pointer);
 544 #endif
 545         STATIC_TLS_INIT(CONTROL_STACK_START,control_stack_start);
 546         STATIC_TLS_INIT(CONTROL_STACK_END,control_stack_end);
 547 #ifdef ALIEN_STACK
 548         STATIC_TLS_INIT(ALIEN_STACK,alien_stack_pointer);
 549 #endif
 550 #if defined(LISP_FEATURE_X86) || defined (LISP_FEATURE_X86_64)
 551         STATIC_TLS_INIT(PSEUDO_ATOMIC_BITS,pseudo_atomic_bits);
 552 #endif
 553 #undef STATIC_TLS_INIT
 554     }
 555 #endif
 556
 557     th=&per_thread->thread;
 558     th->os_address = spaces;
 559     th->control_stack_start = aligned_spaces;
 560     th->binding_stack_start=
 561         (lispobj*)((void*)th->control_stack_start+thread_control_stack_size);
 562     th->control_stack_end = th->binding_stack_start;
 563     th->control_stack_guard_page_protected = T;
 564     th->alien_stack_start=
 565         (lispobj*)((void*)th->binding_stack_start+BINDING_STACK_SIZE);
 566     set_binding_stack_pointer(th,th->binding_stack_start);
 567     th->this=th;
 568     th->os_thread=0;
 569
 570 #ifdef LISP_FEATURE_SB_SAFEPOINT
 571     th->pc_around_foreign_call = 0;
 572     th->csp_around_foreign_call = csp_page;
 573 #endif
 574
 575 #ifdef LISP_FEATURE_SB_THREAD
 576     /* Contrary to the "allocate all the spaces at once" comment above,
 577      * the os_attr is allocated separately.  We cannot put it into the
 578      * nonpointer data, because it's used for post_mortem and freed
 579      * separately */
 580     th->os_attr=malloc(sizeof(pthread_attr_t));
 581     th->nonpointer_data = nonpointer_data;
 582 # ifndef LISP_FEATURE_SB_SAFEPOINT
 583     th->state_sem=&nonpointer_data->state_sem;
 584     th->state_not_running_sem=&nonpointer_data->state_not_running_sem;
 585     th->state_not_stopped_sem=&nonpointer_data->state_not_stopped_sem;
 586     os_sem_init(th->state_sem, 1);
 587     os_sem_init(th->state_not_running_sem, 0);
 588     os_sem_init(th->state_not_stopped_sem, 0);
 589 # endif
 590     th->state_not_running_waitcount = 0;
 591     th->state_not_stopped_waitcount = 0;
 592 #endif
 593     th->state=STATE_RUNNING;
 594 #ifdef LISP_FEATURE_STACK_GROWS_DOWNWARD_NOT_UPWARD
 595     th->alien_stack_pointer=((void *)th->alien_stack_start
 596                              + ALIEN_STACK_SIZE-N_WORD_BYTES);
 597 #else
 598     th->alien_stack_pointer=((void *)th->alien_stack_start);
 599 #endif
 600 #if defined(LISP_FEATURE_X86) || defined (LISP_FEATURE_X86_64) || defined(LISP_FEATURE_SB_THREAD)
 601     th->pseudo_atomic_bits=0;
 602 #endif
 603 #ifdef LISP_FEATURE_GENCGC
 604     gc_set_region_empty(&th->alloc_region);
 605 #endif
 606 #ifdef LISP_FEATURE_SB_THREAD
 607     /* This parallels the same logic in globals.c for the
 608      * single-threaded foreign_function_call_active, KLUDGE and
 609      * all. */
 610 #if defined(LISP_FEATURE_X86) || defined(LISP_FEATURE_X86_64)
 611     th->foreign_function_call_active = 0;
 612 #else
 613     th->foreign_function_call_active = 1;
 614 #endif
 615 #endif
 616
 617 #ifndef LISP_FEATURE_SB_THREAD
 618     /* the tls-points-into-struct-thread trick is only good for threaded
 619      * sbcl, because unithread sbcl doesn't have tls.  So, we copy the
 620      * appropriate values from struct thread here, and make sure that
 621      * we use the appropriate SymbolValue macros to access any of the
 622      * variable quantities from the C runtime.  It's not quite OAOOM,
 623      * it just feels like it */
 624     SetSymbolValue(BINDING_STACK_START,(lispobj)th->binding_stack_start,th);
 625     SetSymbolValue(CONTROL_STACK_START,(lispobj)th->control_stack_start,th);
 626     SetSymbolValue(CONTROL_STACK_END,(lispobj)th->control_stack_end,th);
 627 #if defined(LISP_FEATURE_X86) || defined (LISP_FEATURE_X86_64)
 628     SetSymbolValue(ALIEN_STACK,(lispobj)th->alien_stack_pointer,th);
 629     SetSymbolValue(PSEUDO_ATOMIC_BITS,(lispobj)th->pseudo_atomic_bits,th);
 630 #endif
 631 #endif
 632     bind_variable(CURRENT_CATCH_BLOCK,make_fixnum(0),th);
 633     bind_variable(CURRENT_UNWIND_PROTECT_BLOCK,make_fixnum(0),th);
 634     bind_variable(FREE_INTERRUPT_CONTEXT_INDEX,make_fixnum(0),th);
 635     bind_variable(INTERRUPT_PENDING, NIL,th);
 636     bind_variable(INTERRUPTS_ENABLED,T,th);
 637     bind_variable(ALLOW_WITH_INTERRUPTS,T,th);
 638     bind_variable(GC_PENDING,NIL,th);
 639     bind_variable(ALLOC_SIGNAL,NIL,th);
 640 #ifdef PINNED_OBJECTS
 641     bind_variable(PINNED_OBJECTS,NIL,th);
 642 #endif
 643 #ifdef LISP_FEATURE_SB_THREAD
 644     bind_variable(STOP_FOR_GC_PENDING,NIL,th);
 645 #endif
 646 #if defined(LISP_FEATURE_SB_SAFEPOINT)
 647     bind_variable(GC_SAFE,NIL,th);
 648     bind_variable(IN_SAFEPOINT,NIL,th);
 649 #endif
 650 #ifdef LISP_FEATURE_SB_THRUPTION
 651     bind_variable(THRUPTION_PENDING,NIL,th);
 652     bind_variable(RESTART_CLUSTERS,NIL,th);
 653 #endif
 654 #ifndef LISP_FEATURE_C_STACK_IS_CONTROL_STACK
 655     access_control_stack_pointer(th)=th->control_stack_start;
 656 #endif
 657
 658 #if defined(LISP_FEATURE_WIN32)
 659     th->interrupt_data = (struct interrupt_data *)
 660         calloc((sizeof (struct interrupt_data)),1);
 661 #else
 662     th->interrupt_data = (struct interrupt_data *)
 663         os_validate(0,(sizeof (struct interrupt_data)));
 664 #endif
 665     if (!th->interrupt_data) {
 666         free_thread_struct(th);
 667         return 0;
 668     }
 669     th->interrupt_data->pending_handler = 0;
 670     th->interrupt_data->gc_blocked_deferrables = 0;
 671 #ifdef GENCGC_IS_PRECISE
 672     th->interrupt_data->allocation_trap_context = 0;
 673 #endif
 674     th->no_tls_value_marker=initial_function;
 675
 676 #if defined(LISP_FEATURE_WIN32)
 677     for (i = 0; i<sizeof(th->private_events.events)/
 678            sizeof(th->private_events.events[0]); ++i) {
 679       th->private_events.events[i] = CreateEvent(NULL,FALSE,FALSE,NULL);
 680     }
 681     th->synchronous_io_handle_and_flag = 0;
 682 #endif
 683     th->stepping = NIL;
 684     return th;
 685 }
 686
 687 void create_initial_thread(lispobj initial_function) {
 688     struct thread *th=create_thread_struct(initial_function);
 689 #ifdef LISP_FEATURE_SB_THREAD
 690     pthread_key_create(&lisp_thread, 0);
 691 #endif
 692     if(th) {
 693         initial_thread_trampoline(th); /* no return */
 694     } else lose("can't create initial thread\n");
 695 }
 696
 697 #ifdef LISP_FEATURE_SB_THREAD
 698
 699 #ifndef __USE_XOPEN2K
 700 extern int pthread_attr_setstack (pthread_attr_t *__attr, void *__stackaddr,
 701                                   size_t __stacksize);
 702 #endif
 703
 704 boolean create_os_thread(struct thread *th,os_thread_t *kid_tid)
 705 {
 706     /* The new thread inherits the restrictive signal mask set here,
 707      * and enables signals again when it is set up properly. */
 708     sigset_t oldset;
 709     boolean r=1;
 710     int retcode = 0, initcode;
 711
 712     FSHOW_SIGNAL((stderr,"/create_os_thread: creating new thread\n"));
 713
 714     /* Blocking deferrable signals is enough, no need to block
 715      * SIG_STOP_FOR_GC because the child process is not linked onto
 716      * all_threads until it's ready. */
 717     block_deferrable_signals(0, &oldset);
 718
 719 #ifdef LOCK_CREATE_THREAD
 720     retcode = pthread_mutex_lock(&create_thread_lock);
 721     gc_assert(retcode == 0);
 722     FSHOW_SIGNAL((stderr,"/create_os_thread: got lock\n"));
 723 #endif
 724
 725     if((initcode = pthread_attr_init(th->os_attr)) ||
 726        /* call_into_lisp_first_time switches the stack for the initial
 727         * thread. For the others, we use this. */
 728 #if defined(LISP_FEATURE_WIN32)
 729        (pthread_attr_setstacksize(th->os_attr, thread_control_stack_size)) ||
 730 #else
 731        (pthread_attr_setstack(th->os_attr,th->control_stack_start,
 732                               thread_control_stack_size)) ||
 733 #endif
 734        (retcode = pthread_create
 735         (kid_tid,th->os_attr,(void *(*)(void *))new_thread_trampoline,th))) {
 736         FSHOW_SIGNAL((stderr, "init = %d\n", initcode));
 737         FSHOW_SIGNAL((stderr, "pthread_create returned %d, errno %d\n",
 738                       retcode, errno));
 739         if(retcode < 0) {
 740             perror("create_os_thread");
 741         }
 742         r=0;
 743     }
 744
 745 #ifdef LOCK_CREATE_THREAD
 746     retcode = pthread_mutex_unlock(&create_thread_lock);
 747     gc_assert(retcode == 0);
 748     FSHOW_SIGNAL((stderr,"/create_os_thread: released lock\n"));
 749 #endif
 750     thread_sigmask(SIG_SETMASK,&oldset,0);
 751     return r;
 752 }
 753
 754 os_thread_t create_thread(lispobj initial_function) {
 755     struct thread *th, *thread = arch_os_get_current_thread();
 756     os_thread_t kid_tid = 0;
 757
 758     /* Must defend against async unwinds. */
 759     if (SymbolValue(INTERRUPTS_ENABLED, thread) != NIL)
 760         lose("create_thread is not safe when interrupts are enabled.\n");
 761
 762     /* Assuming that a fresh thread struct has no lisp objects in it,
 763      * linking it to all_threads can be left to the thread itself
 764      * without fear of gc lossage. initial_function violates this
 765      * assumption and must stay pinned until the child starts up. */
 766     th = create_thread_struct(initial_function);
 767     if (th && !create_os_thread(th,&kid_tid)) {
 768         free_thread_struct(th);
 769         kid_tid = 0;
 770     }
 771     return kid_tid;
 772 }
 773
 774 /* stopping the world is a two-stage process.  From this thread we signal
 775  * all the others with SIG_STOP_FOR_GC.  The handler for this signal does
 776  * the usual pseudo-atomic checks (we don't want to stop a thread while
 777  * it's in the middle of allocation) then waits for another SIG_STOP_FOR_GC.
 778  */
 779 /*
 780  * (With SB-SAFEPOINT, see the definitions in safepoint.c instead.)
 781  */
 782 #ifndef LISP_FEATURE_SB_SAFEPOINT
 783
 784 /* To avoid deadlocks when gc stops the world all clients of each
 785  * mutex must enable or disable SIG_STOP_FOR_GC for the duration of
 786  * holding the lock, but they must agree on which. */
 787 void gc_stop_the_world()
 788 {
 789     struct thread *p,*th=arch_os_get_current_thread();
 790     int status, lock_ret;
 791 #ifdef LOCK_CREATE_THREAD
 792     /* KLUDGE: Stopping the thread during pthread_create() causes deadlock
 793      * on FreeBSD. */
 794     FSHOW_SIGNAL((stderr,"/gc_stop_the_world:waiting on create_thread_lock\n"));
 795     lock_ret = pthread_mutex_lock(&create_thread_lock);
 796     gc_assert(lock_ret == 0);
 797     FSHOW_SIGNAL((stderr,"/gc_stop_the_world:got create_thread_lock\n"));
 798 #endif
 799     FSHOW_SIGNAL((stderr,"/gc_stop_the_world:waiting on lock\n"));
 800     /* keep threads from starting while the world is stopped. */
 801     lock_ret = pthread_mutex_lock(&all_threads_lock);      \
 802     gc_assert(lock_ret == 0);
 803
 804     FSHOW_SIGNAL((stderr,"/gc_stop_the_world:got lock\n"));
 805     /* stop all other threads by sending them SIG_STOP_FOR_GC */
 806     for(p=all_threads; p; p=p->next) {
 807         gc_assert(p->os_thread != 0);
 808         FSHOW_SIGNAL((stderr,"/gc_stop_the_world: thread=%lu, state=%x\n",
 809                       p->os_thread, thread_state(p)));
 810         if((p!=th) && ((thread_state(p)==STATE_RUNNING))) {
 811             FSHOW_SIGNAL((stderr,"/gc_stop_the_world: suspending thread %lu\n",
 812                           p->os_thread));
 813             /* We already hold all_thread_lock, P can become DEAD but
 814              * cannot exit, ergo it's safe to use pthread_kill. */
 815             status=pthread_kill(p->os_thread,SIG_STOP_FOR_GC);
 816             if (status==ESRCH) {
 817                 /* This thread has exited. */
 818                 gc_assert(thread_state(p)==STATE_DEAD);
 819             } else if (status) {
 820                 lose("cannot send suspend thread=%lu: %d, %s\n",
 821                      p->os_thread,status,strerror(status));
 822             }
 823         }
 824     }
 825     FSHOW_SIGNAL((stderr,"/gc_stop_the_world:signals sent\n"));
 826     for(p=all_threads;p;p=p->next) {
 827         if (p!=th) {
 828             FSHOW_SIGNAL
 829                 ((stderr,
 830                   "/gc_stop_the_world: waiting for thread=%lu: state=%x\n",
 831                   p->os_thread, thread_state(p)));
 832             wait_for_thread_state_change(p, STATE_RUNNING);
 833             if (p->state == STATE_RUNNING)
 834                 lose("/gc_stop_the_world: unexpected state");
 835         }
 836     }
 837     FSHOW_SIGNAL((stderr,"/gc_stop_the_world:end\n"));
 838 }
 839
 840 void gc_start_the_world()
 841 {
 842     struct thread *p,*th=arch_os_get_current_thread();
 843     int lock_ret;
 844     /* if a resumed thread creates a new thread before we're done with
 845      * this loop, the new thread will get consed on the front of
 846      * all_threads, but it won't have been stopped so won't need
 847      * restarting */
 848     FSHOW_SIGNAL((stderr,"/gc_start_the_world:begin\n"));
 849     for(p=all_threads;p;p=p->next) {
 850         gc_assert(p->os_thread!=0);
 851         if (p!=th) {
 852             lispobj state = thread_state(p);
 853             if (state != STATE_DEAD) {
 854                 if(state != STATE_STOPPED) {
 855                     lose("gc_start_the_world: wrong thread state is %d\n",
 856                          fixnum_value(state));
 857                 }
 858                 FSHOW_SIGNAL((stderr, "/gc_start_the_world: resuming %lu\n",
 859                               p->os_thread));
 860                 set_thread_state(p, STATE_RUNNING);
 861             }
 862         }
 863     }
 864
 865     lock_ret = pthread_mutex_unlock(&all_threads_lock);
 866     gc_assert(lock_ret == 0);
 867 #ifdef LOCK_CREATE_THREAD
 868     lock_ret = pthread_mutex_unlock(&create_thread_lock);
 869     gc_assert(lock_ret == 0);
 870 #endif
 871
 872     FSHOW_SIGNAL((stderr,"/gc_start_the_world:end\n"));
 873 }
 874
 875 #endif /* !LISP_FEATURE_SB_SAFEPOINT */
 876 #endif /* !LISP_FEATURE_SB_THREAD */
 877
 878 int
 879 thread_yield()
 880 {
 881 #ifdef LISP_FEATURE_SB_THREAD
 882     return sched_yield();
 883 #else
 884     return 0;
 885 #endif
 886 }
 887
 888 int
 889 wake_thread(os_thread_t os_thread)
 890 {
 891 #if defined(LISP_FEATURE_WIN32)
 892     return kill_safely(os_thread, 1);
 893 #elif !defined(LISP_FEATURE_SB_THRUPTION)
 894     return kill_safely(os_thread, SIGPIPE);
 895 #else
 896     return wake_thread_posix(os_thread);
 897 #endif
 898 }
 899
 900 /* If the thread id given does not belong to a running thread (it has
 901  * exited or never even existed) pthread_kill _may_ fail with ESRCH,
 902  * but it is also allowed to just segfault, see
 903  * <http://udrepper.livejournal.com/16844.html>.
 904  *
 905  * Relying on thread ids can easily backfire since ids are recycled
 906  * (NPTL recycles them extremely fast) so a signal can be sent to
 907  * another process if the one it was sent to exited.
 908  *
 909  * For these reasons, we must make sure that the thread is still alive
 910  * when the pthread_kill is called and return if the thread is
 911  * exiting.
 912  *
 913  * Note (DFL, 2011-06-22): At the time of writing, this function is only
 914  * used for INTERRUPT-THREAD, hence the wake_thread special-case for
 915  * Windows is OK. */
 916 int
 917 kill_safely(os_thread_t os_thread, int signal)
 918 {
 919     FSHOW_SIGNAL((stderr,"/kill_safely: %lu, %d\n", os_thread, signal));
 920     {
 921 #ifdef LISP_FEATURE_SB_THREAD
 922         sigset_t oldset;
 923         struct thread *thread;
 924         /* Frequent special case: resignalling to self.  The idea is
 925          * that leave_region safepoint will acknowledge the signal, so
 926          * there is no need to take locks, roll thread to safepoint
 927          * etc. */
 928         /* Kludge (on safepoint builds): At the moment, this isn't just
 929          * an optimization; rather it masks the fact that
 930          * gc_stop_the_world() grabs the all_threads mutex without
 931          * releasing it, and since we're not using recursive pthread
 932          * mutexes, the pthread_mutex_lock() around the all_threads loop
 933          * would go wrong.  Why are we running interruptions while
 934          * stopping the world though?  Test case is (:ASYNC-UNWIND
 935          * :SPECIALS), especially with s/10/100/ in both loops. */
 936         if (os_thread == pthread_self()) {
 937             pthread_kill(os_thread, signal);
 938 #ifdef LISP_FEATURE_WIN32
 939             check_pending_thruptions(NULL);
 940 #endif
 941             return 0;
 942         }
 943
 944         /* pthread_kill is not async signal safe and we don't want to be
 945          * interrupted while holding the lock. */
 946         block_deferrable_signals(0, &oldset);
 947         pthread_mutex_lock(&all_threads_lock);
 948         for (thread = all_threads; thread; thread = thread->next) {
 949             if (thread->os_thread == os_thread) {
 950                 int status = pthread_kill(os_thread, signal);
 951                 if (status)
 952                     lose("kill_safely: pthread_kill failed with %d\n", status);
 953 #if defined(LISP_FEATURE_WIN32) && defined(LISP_FEATURE_SB_THRUPTION)
 954                 wake_thread_win32(thread);
 955 #endif
 956                 break;
 957             }
 958         }
 959         pthread_mutex_unlock(&all_threads_lock);
 960         thread_sigmask(SIG_SETMASK,&oldset,0);
 961         if (thread)
 962             return 0;
 963         else
 964             return -1;
 965 #elif defined(LISP_FEATURE_WIN32)
 966         return 0;
 967 #else
 968         int status;
 969         if (os_thread != 0)
 970             lose("kill_safely: who do you want to kill? %d?\n", os_thread);
 971         /* Dubious (as in don't know why it works) workaround for the
 972          * signal sometimes not being generated on darwin. */
 973 #ifdef LISP_FEATURE_DARWIN
 974         {
 975             sigset_t oldset;
 976             sigprocmask(SIG_BLOCK, &deferrable_sigset, &oldset);
 977             status = raise(signal);
 978             sigprocmask(SIG_SETMASK,&oldset,0);
 979         }
 980 #else
 981         status = raise(signal);
 982 #endif
 983         if (status == 0) {
 984             return 0;
 985         } else {
 986             lose("cannot raise signal %d, %d %s\n",
 987                  signal, status, strerror(errno));
 988         }
 989 #endif
 990     }
 991 }