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