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