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     FSHOW((stderr, "Deallocating mach port %x\n", THREAD_STRUCT_TO_EXCEPTION_PORT(th)));
 396     mach_port_move_member(current_mach_task,
 397                           THREAD_STRUCT_TO_EXCEPTION_PORT(th),
 398                           MACH_PORT_NULL);
 399     mach_port_deallocate(current_mach_task,
 400                          THREAD_STRUCT_TO_EXCEPTION_PORT(th));
 401     mach_port_destroy(current_mach_task,
 402                       THREAD_STRUCT_TO_EXCEPTION_PORT(th));
 403 #endif
 404
 405     schedule_thread_post_mortem(th);
 406     FSHOW((stderr,"/exiting thread %lu\n", thread_self()));
 407     return result;
 408 }
 409
 410 #endif /* LISP_FEATURE_SB_THREAD */
 411
 412 static void
 413 free_thread_struct(struct thread *th)
 414 {
 415     if (th->interrupt_data)
 416         os_invalidate((os_vm_address_t) th->interrupt_data,
 417                       (sizeof (struct interrupt_data)));
 418     os_invalidate((os_vm_address_t) th->os_address,
 419                   THREAD_STRUCT_SIZE);
 420 }
 421
 422 #ifdef LISP_FEATURE_SB_THREAD
 423 /* FIXME: should be MAX_INTERRUPTS -1 ? */
 424 const unsigned int tls_index_start =
 425   MAX_INTERRUPTS + sizeof(struct thread)/sizeof(lispobj);
 426 #endif
 427
 428 /* this is called from any other thread to create the new one, and
 429  * initialize all parts of it that can be initialized from another
 430  * thread
 431  */
 432
 433 static struct thread *
 434 create_thread_struct(lispobj initial_function) {
 435     union per_thread_data *per_thread;
 436     struct thread *th=0;        /*  subdue gcc */
 437     void *spaces=0;
 438     void *aligned_spaces=0;
 439 #ifdef LISP_FEATURE_SB_THREAD
 440     unsigned int i;
 441 #endif
 442
 443     /* May as well allocate all the spaces at once: it saves us from
 444      * having to decide what to do if only some of the allocations
 445      * succeed. SPACES must be appropriately aligned, since the GC
 446      * expects the control stack to start at a page boundary -- and
 447      * the OS may have even more rigorous requirements. We can't rely
 448      * on the alignment passed from os_validate, since that might
 449      * assume the current (e.g. 4k) pagesize, while we calculate with
 450      * the biggest (e.g. 64k) pagesize allowed by the ABI. */
 451     spaces=os_validate(0, THREAD_STRUCT_SIZE);
 452     if(!spaces)
 453         return NULL;
 454     /* Aligning up is safe as THREAD_STRUCT_SIZE has
 455      * THREAD_ALIGNMENT_BYTES padding. */
 456     aligned_spaces = (void *)((((unsigned long)(char *)spaces)
 457                                + THREAD_ALIGNMENT_BYTES-1)
 458                               &~(unsigned long)(THREAD_ALIGNMENT_BYTES-1));
 459     per_thread=(union per_thread_data *)
 460         (aligned_spaces+
 461          thread_control_stack_size+
 462          BINDING_STACK_SIZE+
 463          ALIEN_STACK_SIZE +
 464          THREAD_STATE_LOCK_SIZE);
 465
 466 #ifdef LISP_FEATURE_SB_THREAD
 467     for(i = 0; i < (dynamic_values_bytes / sizeof(lispobj)); i++)
 468         per_thread->dynamic_values[i] = NO_TLS_VALUE_MARKER_WIDETAG;
 469     if (all_threads == 0) {
 470         if(SymbolValue(FREE_TLS_INDEX,0)==UNBOUND_MARKER_WIDETAG) {
 471             SetSymbolValue(FREE_TLS_INDEX,tls_index_start << WORD_SHIFT,0);
 472             SetSymbolValue(TLS_INDEX_LOCK,make_fixnum(0),0);
 473         }
 474 #define STATIC_TLS_INIT(sym,field) \
 475   ((struct symbol *)(sym-OTHER_POINTER_LOWTAG))->tls_index= \
 476   (THREAD_SLOT_OFFSET_WORDS(field) << WORD_SHIFT)
 477
 478         STATIC_TLS_INIT(BINDING_STACK_START,binding_stack_start);
 479 #ifdef BINDING_STACK_POINTER
 480         STATIC_TLS_INIT(BINDING_STACK_POINTER,binding_stack_pointer);
 481 #endif
 482         STATIC_TLS_INIT(CONTROL_STACK_START,control_stack_start);
 483         STATIC_TLS_INIT(CONTROL_STACK_END,control_stack_end);
 484 #ifdef ALIEN_STACK
 485         STATIC_TLS_INIT(ALIEN_STACK,alien_stack_pointer);
 486 #endif
 487 #if defined(LISP_FEATURE_X86) || defined (LISP_FEATURE_X86_64)
 488         STATIC_TLS_INIT(PSEUDO_ATOMIC_BITS,pseudo_atomic_bits);
 489 #endif
 490 #undef STATIC_TLS_INIT
 491     }
 492 #endif
 493
 494     th=&per_thread->thread;
 495     th->os_address = spaces;
 496     th->control_stack_start = aligned_spaces;
 497     th->binding_stack_start=
 498         (lispobj*)((void*)th->control_stack_start+thread_control_stack_size);
 499     th->control_stack_end = th->binding_stack_start;
 500     th->control_stack_guard_page_protected = T;
 501     th->alien_stack_start=
 502         (lispobj*)((void*)th->binding_stack_start+BINDING_STACK_SIZE);
 503     set_binding_stack_pointer(th,th->binding_stack_start);
 504     th->this=th;
 505     th->os_thread=0;
 506 #ifdef LISP_FEATURE_SB_THREAD
 507     th->os_attr=malloc(sizeof(pthread_attr_t));
 508     th->state_sem=(os_sem_t *)((void *)th->alien_stack_start + ALIEN_STACK_SIZE);
 509     th->state_not_running_sem=(os_sem_t *)
 510         ((void *)th->state_sem + (sizeof(os_sem_t)));
 511     th->state_not_stopped_sem=(os_sem_t *)
 512         ((void *)th->state_not_running_sem + (sizeof(os_sem_t)));
 513     th->state_not_running_waitcount = 0;
 514     th->state_not_stopped_waitcount = 0;
 515     os_sem_init(th->state_sem, 1);
 516     os_sem_init(th->state_not_running_sem, 0);
 517     os_sem_init(th->state_not_stopped_sem, 0);
 518 #endif
 519     th->state=STATE_RUNNING;
 520 #ifdef LISP_FEATURE_STACK_GROWS_DOWNWARD_NOT_UPWARD
 521     th->alien_stack_pointer=((void *)th->alien_stack_start
 522                              + ALIEN_STACK_SIZE-N_WORD_BYTES);
 523 #else
 524     th->alien_stack_pointer=((void *)th->alien_stack_start);
 525 #endif
 526 #if defined(LISP_FEATURE_X86) || defined (LISP_FEATURE_X86_64) || defined(LISP_FEATURE_SB_THREAD)
 527     th->pseudo_atomic_bits=0;
 528 #endif
 529 #ifdef LISP_FEATURE_GENCGC
 530     gc_set_region_empty(&th->alloc_region);
 531 #endif
 532 #ifdef LISP_FEATURE_SB_THREAD
 533     /* This parallels the same logic in globals.c for the
 534      * single-threaded foreign_function_call_active, KLUDGE and
 535      * all. */
 536 #if defined(LISP_FEATURE_X86) || defined(LISP_FEATURE_X86_64)
 537     th->foreign_function_call_active = 0;
 538 #else
 539     th->foreign_function_call_active = 1;
 540 #endif
 541 #endif
 542
 543 #ifndef LISP_FEATURE_SB_THREAD
 544     /* the tls-points-into-struct-thread trick is only good for threaded
 545      * sbcl, because unithread sbcl doesn't have tls.  So, we copy the
 546      * appropriate values from struct thread here, and make sure that
 547      * we use the appropriate SymbolValue macros to access any of the
 548      * variable quantities from the C runtime.  It's not quite OAOOM,
 549      * it just feels like it */
 550     SetSymbolValue(BINDING_STACK_START,(lispobj)th->binding_stack_start,th);
 551     SetSymbolValue(CONTROL_STACK_START,(lispobj)th->control_stack_start,th);
 552     SetSymbolValue(CONTROL_STACK_END,(lispobj)th->control_stack_end,th);
 553 #if defined(LISP_FEATURE_X86) || defined (LISP_FEATURE_X86_64)
 554     SetSymbolValue(ALIEN_STACK,(lispobj)th->alien_stack_pointer,th);
 555     SetSymbolValue(PSEUDO_ATOMIC_BITS,(lispobj)th->pseudo_atomic_bits,th);
 556 #endif
 557 #endif
 558     bind_variable(CURRENT_CATCH_BLOCK,make_fixnum(0),th);
 559     bind_variable(CURRENT_UNWIND_PROTECT_BLOCK,make_fixnum(0),th);
 560     bind_variable(FREE_INTERRUPT_CONTEXT_INDEX,make_fixnum(0),th);
 561     bind_variable(INTERRUPT_PENDING, NIL,th);
 562     bind_variable(INTERRUPTS_ENABLED,T,th);
 563     bind_variable(ALLOW_WITH_INTERRUPTS,T,th);
 564     bind_variable(GC_PENDING,NIL,th);
 565     bind_variable(ALLOC_SIGNAL,NIL,th);
 566 #ifdef PINNED_OBJECTS
 567     bind_variable(PINNED_OBJECTS,NIL,th);
 568 #endif
 569 #ifdef LISP_FEATURE_SB_THREAD
 570     bind_variable(STOP_FOR_GC_PENDING,NIL,th);
 571 #endif
 572 #ifndef LISP_FEATURE_C_STACK_IS_CONTROL_STACK
 573     access_control_stack_pointer(th)=th->control_stack_start;
 574 #endif
 575
 576     th->interrupt_data = (struct interrupt_data *)
 577         os_validate(0,(sizeof (struct interrupt_data)));
 578     if (!th->interrupt_data) {
 579         free_thread_struct(th);
 580         return 0;
 581     }
 582     th->interrupt_data->pending_handler = 0;
 583     th->interrupt_data->gc_blocked_deferrables = 0;
 584 #ifdef LISP_FEATURE_PPC
 585     th->interrupt_data->allocation_trap_context = 0;
 586 #endif
 587     th->no_tls_value_marker=initial_function;
 588
 589     th->stepping = NIL;
 590     return th;
 591 }
 592
 593 #ifdef LISP_FEATURE_MACH_EXCEPTION_HANDLER
 594 mach_port_t setup_mach_exception_handling_thread();
 595 kern_return_t mach_thread_init(mach_port_t thread_exception_port);
 596
 597 #endif
 598
 599 void create_initial_thread(lispobj initial_function) {
 600     struct thread *th=create_thread_struct(initial_function);
 601 #ifdef LISP_FEATURE_SB_THREAD
 602     pthread_key_create(&lisp_thread, 0);
 603 #endif
 604     if(th) {
 605         initial_thread_trampoline(th); /* no return */
 606     } else lose("can't create initial thread\n");
 607 }
 608
 609 #ifdef LISP_FEATURE_SB_THREAD
 610
 611 #ifndef __USE_XOPEN2K
 612 extern int pthread_attr_setstack (pthread_attr_t *__attr, void *__stackaddr,
 613                                   size_t __stacksize);
 614 #endif
 615
 616 boolean create_os_thread(struct thread *th,os_thread_t *kid_tid)
 617 {
 618     /* The new thread inherits the restrictive signal mask set here,
 619      * and enables signals again when it is set up properly. */
 620     sigset_t oldset;
 621     boolean r=1;
 622     int retcode = 0, initcode;
 623
 624     FSHOW_SIGNAL((stderr,"/create_os_thread: creating new thread\n"));
 625
 626     /* Blocking deferrable signals is enough, no need to block
 627      * SIG_STOP_FOR_GC because the child process is not linked onto
 628      * all_threads until it's ready. */
 629     block_deferrable_signals(0, &oldset);
 630
 631 #ifdef LOCK_CREATE_THREAD
 632     retcode = pthread_mutex_lock(&create_thread_lock);
 633     gc_assert(retcode == 0);
 634     FSHOW_SIGNAL((stderr,"/create_os_thread: got lock\n"));
 635 #endif
 636
 637     if((initcode = pthread_attr_init(th->os_attr)) ||
 638        /* call_into_lisp_first_time switches the stack for the initial
 639         * thread. For the others, we use this. */
 640        (pthread_attr_setstack(th->os_attr,th->control_stack_start,
 641                               thread_control_stack_size)) ||
 642        (retcode = pthread_create
 643         (kid_tid,th->os_attr,(void *(*)(void *))new_thread_trampoline,th))) {
 644         FSHOW_SIGNAL((stderr, "init = %d\n", initcode));
 645         FSHOW_SIGNAL((stderr, "pthread_create returned %d, errno %d\n",
 646                       retcode, errno));
 647         if(retcode < 0) {
 648             perror("create_os_thread");
 649         }
 650         r=0;
 651     }
 652
 653 #ifdef LOCK_CREATE_THREAD
 654     retcode = pthread_mutex_unlock(&create_thread_lock);
 655     gc_assert(retcode == 0);
 656     FSHOW_SIGNAL((stderr,"/create_os_thread: released lock\n"));
 657 #endif
 658     thread_sigmask(SIG_SETMASK,&oldset,0);
 659     return r;
 660 }
 661
 662 os_thread_t create_thread(lispobj initial_function) {
 663     struct thread *th, *thread = arch_os_get_current_thread();
 664     os_thread_t kid_tid = 0;
 665
 666     /* Must defend against async unwinds. */
 667     if (SymbolValue(INTERRUPTS_ENABLED, thread) != NIL)
 668         lose("create_thread is not safe when interrupts are enabled.\n");
 669
 670     /* Assuming that a fresh thread struct has no lisp objects in it,
 671      * linking it to all_threads can be left to the thread itself
 672      * without fear of gc lossage. initial_function violates this
 673      * assumption and must stay pinned until the child starts up. */
 674     th = create_thread_struct(initial_function);
 675     if (th && !create_os_thread(th,&kid_tid)) {
 676         free_thread_struct(th);
 677         kid_tid = 0;
 678     }
 679     return kid_tid;
 680 }
 681
 682 /* stopping the world is a two-stage process.  From this thread we signal
 683  * all the others with SIG_STOP_FOR_GC.  The handler for this signal does
 684  * the usual pseudo-atomic checks (we don't want to stop a thread while
 685  * it's in the middle of allocation) then waits for another SIG_STOP_FOR_GC.
 686  */
 687
 688 /* To avoid deadlocks when gc stops the world all clients of each
 689  * mutex must enable or disable SIG_STOP_FOR_GC for the duration of
 690  * holding the lock, but they must agree on which. */
 691 void gc_stop_the_world()
 692 {
 693     struct thread *p,*th=arch_os_get_current_thread();
 694     int status, lock_ret;
 695 #ifdef LOCK_CREATE_THREAD
 696     /* KLUDGE: Stopping the thread during pthread_create() causes deadlock
 697      * on FreeBSD. */
 698     FSHOW_SIGNAL((stderr,"/gc_stop_the_world:waiting on create_thread_lock\n"));
 699     lock_ret = pthread_mutex_lock(&create_thread_lock);
 700     gc_assert(lock_ret == 0);
 701     FSHOW_SIGNAL((stderr,"/gc_stop_the_world:got create_thread_lock\n"));
 702 #endif
 703     FSHOW_SIGNAL((stderr,"/gc_stop_the_world:waiting on lock\n"));
 704     /* keep threads from starting while the world is stopped. */
 705     lock_ret = pthread_mutex_lock(&all_threads_lock);      \
 706     gc_assert(lock_ret == 0);
 707
 708     FSHOW_SIGNAL((stderr,"/gc_stop_the_world:got lock\n"));
 709     /* stop all other threads by sending them SIG_STOP_FOR_GC */
 710     for(p=all_threads; p; p=p->next) {
 711         gc_assert(p->os_thread != 0);
 712         FSHOW_SIGNAL((stderr,"/gc_stop_the_world: thread=%lu, state=%x\n",
 713                       p->os_thread, thread_state(p)));
 714         if((p!=th) && ((thread_state(p)==STATE_RUNNING))) {
 715             FSHOW_SIGNAL((stderr,"/gc_stop_the_world: suspending thread %lu\n",
 716                           p->os_thread));
 717             /* We already hold all_thread_lock, P can become DEAD but
 718              * cannot exit, ergo it's safe to use pthread_kill. */
 719             status=pthread_kill(p->os_thread,SIG_STOP_FOR_GC);
 720             if (status==ESRCH) {
 721                 /* This thread has exited. */
 722                 gc_assert(thread_state(p)==STATE_DEAD);
 723             } else if (status) {
 724                 lose("cannot send suspend thread=%lu: %d, %s\n",
 725                      p->os_thread,status,strerror(status));
 726             }
 727         }
 728     }
 729     FSHOW_SIGNAL((stderr,"/gc_stop_the_world:signals sent\n"));
 730     for(p=all_threads;p;p=p->next) {
 731         if (p!=th) {
 732             FSHOW_SIGNAL
 733                 ((stderr,
 734                   "/gc_stop_the_world: waiting for thread=%lu: state=%x\n",
 735                   p->os_thread, thread_state(p)));
 736             wait_for_thread_state_change(p, STATE_RUNNING);
 737             if (p->state == STATE_RUNNING)
 738                 lose("/gc_stop_the_world: unexpected state");
 739         }
 740     }
 741     FSHOW_SIGNAL((stderr,"/gc_stop_the_world:end\n"));
 742 }
 743
 744 void gc_start_the_world()
 745 {
 746     struct thread *p,*th=arch_os_get_current_thread();
 747     int lock_ret;
 748     /* if a resumed thread creates a new thread before we're done with
 749      * this loop, the new thread will get consed on the front of
 750      * all_threads, but it won't have been stopped so won't need
 751      * restarting */
 752     FSHOW_SIGNAL((stderr,"/gc_start_the_world:begin\n"));
 753     for(p=all_threads;p;p=p->next) {
 754         gc_assert(p->os_thread!=0);
 755         if (p!=th) {
 756             lispobj state = thread_state(p);
 757             if (state != STATE_DEAD) {
 758                 if(state != STATE_STOPPED) {
 759                     lose("gc_start_the_world: wrong thread state is %d\n",
 760                          fixnum_value(state));
 761                 }
 762                 FSHOW_SIGNAL((stderr, "/gc_start_the_world: resuming %lu\n",
 763                               p->os_thread));
 764                 set_thread_state(p, STATE_RUNNING);
 765             }
 766         }
 767     }
 768
 769     lock_ret = pthread_mutex_unlock(&all_threads_lock);
 770     gc_assert(lock_ret == 0);
 771 #ifdef LOCK_CREATE_THREAD
 772     lock_ret = pthread_mutex_unlock(&create_thread_lock);
 773     gc_assert(lock_ret == 0);
 774 #endif
 775
 776     FSHOW_SIGNAL((stderr,"/gc_start_the_world:end\n"));
 777 }
 778 #endif
 779
 780 int
 781 thread_yield()
 782 {
 783 #ifdef LISP_FEATURE_SB_THREAD
 784     return sched_yield();
 785 #else
 786     return 0;
 787 #endif
 788 }
 789
 790 /* If the thread id given does not belong to a running thread (it has
 791  * exited or never even existed) pthread_kill _may_ fail with ESRCH,
 792  * but it is also allowed to just segfault, see
 793  * <http://udrepper.livejournal.com/16844.html>.
 794  *
 795  * Relying on thread ids can easily backfire since ids are recycled
 796  * (NPTL recycles them extremely fast) so a signal can be sent to
 797  * another process if the one it was sent to exited.
 798  *
 799  * We send signals in two places: signal_interrupt_thread sends a
 800  * signal that's harmless if delivered to another thread, but
 801  * SIG_STOP_FOR_GC is fatal.
 802  *
 803  * For these reasons, we must make sure that the thread is still alive
 804  * when the pthread_kill is called and return if the thread is
 805  * exiting. */
 806 int
 807 kill_safely(os_thread_t os_thread, int signal)
 808 {
 809     FSHOW_SIGNAL((stderr,"/kill_safely: %lu, %d\n", os_thread, signal));
 810     {
 811 #ifdef LISP_FEATURE_SB_THREAD
 812         sigset_t oldset;
 813         struct thread *thread;
 814         /* pthread_kill is not async signal safe and we don't want to be
 815          * interrupted while holding the lock. */
 816         block_deferrable_signals(0, &oldset);
 817         pthread_mutex_lock(&all_threads_lock);
 818         for (thread = all_threads; thread; thread = thread->next) {
 819             if (thread->os_thread == os_thread) {
 820                 int status = pthread_kill(os_thread, signal);
 821                 if (status)
 822                     lose("kill_safely: pthread_kill failed with %d\n", status);
 823                 break;
 824             }
 825         }
 826         pthread_mutex_unlock(&all_threads_lock);
 827         thread_sigmask(SIG_SETMASK,&oldset,0);
 828         if (thread)
 829             return 0;
 830         else
 831             return -1;
 832 #else
 833         int status;
 834         if (os_thread != 0)
 835             lose("kill_safely: who do you want to kill? %d?\n", os_thread);
 836         /* Dubious (as in don't know why it works) workaround for the
 837          * signal sometimes not being generated on darwin. */
 838 #ifdef LISP_FEATURE_DARWIN
 839         {
 840             sigset_t oldset;
 841             sigprocmask(SIG_BLOCK, &deferrable_sigset, &oldset);
 842             status = raise(signal);
 843             sigprocmask(SIG_SETMASK,&oldset,0);
 844         }
 845 #else
 846         status = raise(signal);
 847 #endif
 848         if (status == 0) {
 849             return 0;
 850         } else {
 851             lose("cannot raise signal %d, %d %s\n",
 852                  signal, status, strerror(errno));
 853         }
 854 #endif
 855     }
 856 }