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 "alloc.h"
  37 #include "thread.h"
  38 #include "arch.h"
  39 #include "target-arch-os.h"
  40 #include "os.h"
  41 #include "globals.h"
  42 #include "dynbind.h"
  43 #include "genesis/cons.h"
  44 #include "genesis/fdefn.h"
  45 #include "interr.h"             /* for lose() */
  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 #define ALIEN_STACK_SIZE (1*1024*1024) /* 1Mb size chosen at random */
  67
  68 struct thread_post_mortem {
  69 #ifdef DELAY_THREAD_POST_MORTEM
  70     struct thread_post_mortem *next;
  71 #endif
  72     os_thread_t os_thread;
  73     pthread_attr_t *os_attr;
  74     os_vm_address_t os_address;
  75 };
  76
  77
  78 #ifdef DELAY_THREAD_POST_MORTEM
  79 static int pending_thread_post_mortem_count = 0;
  80 pthread_mutex_t thread_post_mortem_lock = PTHREAD_MUTEX_INITIALIZER;
  81 #endif
  82 static struct thread_post_mortem * volatile pending_thread_post_mortem = 0;
  83
  84 int dynamic_values_bytes=TLS_SIZE*sizeof(lispobj);  /* same for all threads */
  85 struct thread * volatile all_threads;
  86 extern struct interrupt_data * global_interrupt_data;
  87
  88 #ifdef LISP_FEATURE_SB_THREAD
  89 pthread_mutex_t all_threads_lock = PTHREAD_MUTEX_INITIALIZER;
  90 #ifdef LOCK_CREATE_THREAD
  91 static pthread_mutex_t create_thread_lock = PTHREAD_MUTEX_INITIALIZER;
  92 #endif
  93 #ifdef LISP_FEATURE_GCC_TLS
  94 __thread struct thread *current_thread;
  95 #endif
  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     function = th->no_tls_value_marker;
 132     th->no_tls_value_marker = NO_TLS_VALUE_MARKER_WIDETAG;
 133     if(arch_os_thread_init(th)==0) return 1;
 134     link_thread(th);
 135     th->os_thread=thread_self();
 136 #ifndef LISP_FEATURE_WIN32
 137     protect_control_stack_guard_page(1);
 138 #endif
 139
 140 #if defined(LISP_FEATURE_X86) || defined(LISP_FEATURE_X86_64)
 141     return call_into_lisp_first_time(function,args,0);
 142 #else
 143     return funcall0(function);
 144 #endif
 145 }
 146
 147 #define THREAD_STRUCT_SIZE (thread_control_stack_size + BINDING_STACK_SIZE + \
 148                             ALIEN_STACK_SIZE + dynamic_values_bytes +        \
 149                             32 * SIGSTKSZ +                                  \
 150                             THREAD_ALIGNMENT_BYTES)
 151
 152 #ifdef LISP_FEATURE_SB_THREAD
 153 /* THREAD POST MORTEM CLEANUP
 154  *
 155  * Memory allocated for the thread stacks cannot be reclaimed while
 156  * the thread is still alive, so we need a mechanism for post mortem
 157  * cleanups. FIXME: We actually have three, for historical reasons as
 158  * the saying goes. Do we really need three? Nikodemus guesses that
 159  * not anymore, now that we properly call pthread_attr_destroy before
 160  * freeing the stack. */
 161
 162 static struct thread_post_mortem *
 163 plan_thread_post_mortem(struct thread *corpse)
 164 {
 165     if (corpse) {
 166         struct thread_post_mortem *post_mortem = malloc(sizeof(struct thread_post_mortem));
 167         gc_assert(post_mortem);
 168         post_mortem->os_thread = corpse->os_thread;
 169         post_mortem->os_attr = corpse->os_attr;
 170         post_mortem->os_address = corpse->os_address;
 171 #ifdef DELAY_THREAD_POST_MORTEM
 172         post_mortem->next = NULL;
 173 #endif
 174         return post_mortem;
 175     } else {
 176         /* FIXME: When does this happen? */
 177         return NULL;
 178     }
 179 }
 180
 181 static void
 182 perform_thread_post_mortem(struct thread_post_mortem *post_mortem)
 183 {
 184 #ifdef CREATE_POST_MORTEM_THREAD
 185     pthread_detach(pthread_self());
 186 #endif
 187     if (post_mortem) {
 188         gc_assert(!pthread_join(post_mortem->os_thread, NULL));
 189         gc_assert(!pthread_attr_destroy(post_mortem->os_attr));
 190         free(post_mortem->os_attr);
 191         os_invalidate(post_mortem->os_address, THREAD_STRUCT_SIZE);
 192         free(post_mortem);
 193     }
 194 }
 195
 196 static void
 197 schedule_thread_post_mortem(struct thread *corpse)
 198 {
 199     struct thread_post_mortem *post_mortem = NULL;
 200     if (corpse) {
 201         post_mortem = plan_thread_post_mortem(corpse);
 202
 203 #ifdef DELAY_THREAD_POST_MORTEM
 204         pthread_mutex_lock(&thread_post_mortem_lock);
 205         /* First stick the new post mortem to the end of the queue. */
 206         if (pending_thread_post_mortem) {
 207             struct thread_post_mortem *next = pending_thread_post_mortem;
 208             while (next->next) {
 209                 next = next->next;
 210             }
 211             next->next = post_mortem;
 212         } else {
 213             pending_thread_post_mortem = post_mortem;
 214         }
 215         /* Then, if there are enough things in the queue, clean up one
 216          * from the head -- or increment the count, and null out the
 217          * post_mortem we have. */
 218         if (pending_thread_post_mortem_count > DELAY_THREAD_POST_MORTEM) {
 219             post_mortem = pending_thread_post_mortem;
 220             pending_thread_post_mortem = post_mortem->next;
 221         } else {
 222             pending_thread_post_mortem_count++;
 223             post_mortem = NULL;
 224         }
 225         pthread_mutex_unlock(&thread_post_mortem_lock);
 226         /* Finally run, the cleanup, if any. */
 227         perform_thread_post_mortem(post_mortem);
 228 #elif defined(CREATE_POST_MORTEM_THREAD)
 229         gc_assert(!pthread_create(&thread, NULL, perform_thread_post_mortem, post_mortem));
 230 #else
 231         post_mortem = (struct thread_post_mortem *)
 232             swap_lispobjs((lispobj *)(void *)&pending_thread_post_mortem,
 233                           (lispobj)post_mortem);
 234         perform_thread_post_mortem(post_mortem);
 235 #endif
 236     }
 237 }
 238
 239 /* this is the first thing that runs in the child (which is why the
 240  * silly calling convention).  Basically it calls the user's requested
 241  * lisp function after doing arch_os_thread_init and whatever other
 242  * bookkeeping needs to be done
 243  */
 244 int
 245 new_thread_trampoline(struct thread *th)
 246 {
 247     lispobj function;
 248     int result, lock_ret;
 249
 250     FSHOW((stderr,"/creating thread %lu\n", thread_self()));
 251     function = th->no_tls_value_marker;
 252     th->no_tls_value_marker = NO_TLS_VALUE_MARKER_WIDETAG;
 253     if(arch_os_thread_init(th)==0) {
 254         /* FIXME: handle error */
 255         lose("arch_os_thread_init failed\n");
 256     }
 257
 258     th->os_thread=thread_self();
 259     protect_control_stack_guard_page(1);
 260     /* Since GC can only know about this thread from the all_threads
 261      * list and we're just adding this thread to it there is no danger
 262      * of deadlocking even with SIG_STOP_FOR_GC blocked (which it is
 263      * not). */
 264     lock_ret = pthread_mutex_lock(&all_threads_lock);
 265     gc_assert(lock_ret == 0);
 266     link_thread(th);
 267     lock_ret = pthread_mutex_unlock(&all_threads_lock);
 268     gc_assert(lock_ret == 0);
 269
 270     result = funcall0(function);
 271
 272     /* Block GC */
 273     block_blockable_signals();
 274     th->state=STATE_DEAD;
 275
 276     /* SIG_STOP_FOR_GC is blocked and GC might be waiting for this
 277      * thread, but since we are already dead it won't wait long. */
 278     lock_ret = pthread_mutex_lock(&all_threads_lock);
 279     gc_assert(lock_ret == 0);
 280
 281     gc_alloc_update_page_tables(0, &th->alloc_region);
 282     unlink_thread(th);
 283     pthread_mutex_unlock(&all_threads_lock);
 284     gc_assert(lock_ret == 0);
 285
 286     if(th->tls_cookie>=0) arch_os_thread_cleanup(th);
 287     os_invalidate((os_vm_address_t)th->interrupt_data,
 288                   (sizeof (struct interrupt_data)));
 289
 290 #ifdef LISP_FEATURE_MACH_EXCEPTION_HANDLER
 291     FSHOW((stderr, "Deallocating mach port %x\n", THREAD_STRUCT_TO_EXCEPTION_PORT(th)));
 292     mach_port_move_member(mach_task_self(),
 293                           THREAD_STRUCT_TO_EXCEPTION_PORT(th),
 294                           MACH_PORT_NULL);
 295     mach_port_deallocate(mach_task_self(),
 296                          THREAD_STRUCT_TO_EXCEPTION_PORT(th));
 297     mach_port_destroy(mach_task_self(),
 298                       THREAD_STRUCT_TO_EXCEPTION_PORT(th));
 299 #endif
 300
 301     schedule_thread_post_mortem(th);
 302     FSHOW((stderr,"/exiting thread %p\n", thread_self()));
 303     return result;
 304 }
 305
 306 #endif /* LISP_FEATURE_SB_THREAD */
 307
 308 static void
 309 free_thread_struct(struct thread *th)
 310 {
 311     if (th->interrupt_data)
 312         os_invalidate((os_vm_address_t) th->interrupt_data,
 313                       (sizeof (struct interrupt_data)));
 314     os_invalidate((os_vm_address_t) th->os_address,
 315                   THREAD_STRUCT_SIZE);
 316 }
 317
 318 /* this is called from any other thread to create the new one, and
 319  * initialize all parts of it that can be initialized from another
 320  * thread
 321  */
 322
 323 static struct thread *
 324 create_thread_struct(lispobj initial_function) {
 325     union per_thread_data *per_thread;
 326     struct thread *th=0;        /*  subdue gcc */
 327     void *spaces=0;
 328     void *aligned_spaces=0;
 329 #ifdef LISP_FEATURE_SB_THREAD
 330     unsigned int i;
 331 #endif
 332
 333     /* May as well allocate all the spaces at once: it saves us from
 334      * having to decide what to do if only some of the allocations
 335      * succeed. SPACES must be appropriately aligned, since the GC
 336      * expects the control stack to start at a page boundary -- and
 337      * the OS may have even more rigorous requirements. We can't rely
 338      * on the alignment passed from os_validate, since that might
 339      * assume the current (e.g. 4k) pagesize, while we calculate with
 340      * the biggest (e.g. 64k) pagesize allowed by the ABI. */
 341     spaces=os_validate(0, THREAD_STRUCT_SIZE);
 342     if(!spaces)
 343         return NULL;
 344     /* Aligning up is safe as THREAD_STRUCT_SIZE has
 345      * THREAD_ALIGNMENT_BYTES padding. */
 346     aligned_spaces = (void *)((((unsigned long)(char *)spaces)
 347                                + THREAD_ALIGNMENT_BYTES-1)
 348                               &~(unsigned long)(THREAD_ALIGNMENT_BYTES-1));
 349     per_thread=(union per_thread_data *)
 350         (aligned_spaces+
 351          thread_control_stack_size+
 352          BINDING_STACK_SIZE+
 353          ALIEN_STACK_SIZE);
 354
 355 #ifdef LISP_FEATURE_SB_THREAD
 356     for(i = 0; i < (dynamic_values_bytes / sizeof(lispobj)); i++)
 357         per_thread->dynamic_values[i] = NO_TLS_VALUE_MARKER_WIDETAG;
 358     if (all_threads == 0) {
 359         if(SymbolValue(FREE_TLS_INDEX,0)==UNBOUND_MARKER_WIDETAG) {
 360             SetSymbolValue
 361                 (FREE_TLS_INDEX,
 362                  /* FIXME: should be MAX_INTERRUPTS -1 ? */
 363                  make_fixnum(MAX_INTERRUPTS+
 364                              sizeof(struct thread)/sizeof(lispobj)),
 365                  0);
 366             SetSymbolValue(TLS_INDEX_LOCK,make_fixnum(0),0);
 367         }
 368 #define STATIC_TLS_INIT(sym,field) \
 369   ((struct symbol *)(sym-OTHER_POINTER_LOWTAG))->tls_index= \
 370   make_fixnum(THREAD_SLOT_OFFSET_WORDS(field))
 371
 372         STATIC_TLS_INIT(BINDING_STACK_START,binding_stack_start);
 373         STATIC_TLS_INIT(BINDING_STACK_POINTER,binding_stack_pointer);
 374         STATIC_TLS_INIT(CONTROL_STACK_START,control_stack_start);
 375         STATIC_TLS_INIT(CONTROL_STACK_END,control_stack_end);
 376         STATIC_TLS_INIT(ALIEN_STACK,alien_stack_pointer);
 377 #if defined(LISP_FEATURE_X86) || defined (LISP_FEATURE_X86_64)
 378         STATIC_TLS_INIT(PSEUDO_ATOMIC_BITS,pseudo_atomic_bits);
 379 #endif
 380 #undef STATIC_TLS_INIT
 381     }
 382 #endif
 383
 384     th=&per_thread->thread;
 385     th->os_address = spaces;
 386     th->control_stack_start = aligned_spaces;
 387     th->binding_stack_start=
 388         (lispobj*)((void*)th->control_stack_start+thread_control_stack_size);
 389     th->control_stack_end = th->binding_stack_start;
 390     th->alien_stack_start=
 391         (lispobj*)((void*)th->binding_stack_start+BINDING_STACK_SIZE);
 392     th->binding_stack_pointer=th->binding_stack_start;
 393     th->this=th;
 394     th->os_thread=0;
 395     th->os_attr=malloc(sizeof(pthread_attr_t));
 396     th->state=STATE_RUNNING;
 397 #ifdef LISP_FEATURE_STACK_GROWS_DOWNWARD_NOT_UPWARD
 398     th->alien_stack_pointer=((void *)th->alien_stack_start
 399                              + ALIEN_STACK_SIZE-N_WORD_BYTES);
 400 #else
 401     th->alien_stack_pointer=((void *)th->alien_stack_start);
 402 #endif
 403 #if defined(LISP_FEATURE_X86) || defined (LISP_FEATURE_X86_64)
 404     th->pseudo_atomic_bits=0;
 405 #endif
 406 #ifdef LISP_FEATURE_GENCGC
 407     gc_set_region_empty(&th->alloc_region);
 408 #endif
 409
 410 #ifndef LISP_FEATURE_SB_THREAD
 411     /* the tls-points-into-struct-thread trick is only good for threaded
 412      * sbcl, because unithread sbcl doesn't have tls.  So, we copy the
 413      * appropriate values from struct thread here, and make sure that
 414      * we use the appropriate SymbolValue macros to access any of the
 415      * variable quantities from the C runtime.  It's not quite OAOOM,
 416      * it just feels like it */
 417     SetSymbolValue(BINDING_STACK_START,(lispobj)th->binding_stack_start,th);
 418     SetSymbolValue(CONTROL_STACK_START,(lispobj)th->control_stack_start,th);
 419     SetSymbolValue(CONTROL_STACK_END,(lispobj)th->control_stack_end,th);
 420 #if defined(LISP_FEATURE_X86) || defined (LISP_FEATURE_X86_64)
 421     SetSymbolValue(BINDING_STACK_POINTER,(lispobj)th->binding_stack_pointer,th);
 422     SetSymbolValue(ALIEN_STACK,(lispobj)th->alien_stack_pointer,th);
 423     SetSymbolValue(PSEUDO_ATOMIC_BITS,(lispobj)th->pseudo_atomic_bits,th);
 424 #else
 425     current_binding_stack_pointer=th->binding_stack_pointer;
 426     current_control_stack_pointer=th->control_stack_start;
 427 #endif
 428 #endif
 429     bind_variable(CURRENT_CATCH_BLOCK,make_fixnum(0),th);
 430     bind_variable(CURRENT_UNWIND_PROTECT_BLOCK,make_fixnum(0),th);
 431     bind_variable(FREE_INTERRUPT_CONTEXT_INDEX,make_fixnum(0),th);
 432     bind_variable(INTERRUPT_PENDING, NIL,th);
 433     bind_variable(INTERRUPTS_ENABLED,T,th);
 434     bind_variable(ALLOW_WITH_INTERRUPTS,T,th);
 435     bind_variable(GC_PENDING,NIL,th);
 436 #ifdef LISP_FEATURE_SB_THREAD
 437     bind_variable(STOP_FOR_GC_PENDING,NIL,th);
 438 #endif
 439
 440     th->interrupt_data = (struct interrupt_data *)
 441         os_validate(0,(sizeof (struct interrupt_data)));
 442     if (!th->interrupt_data) {
 443         free_thread_struct(th);
 444         return 0;
 445     }
 446     th->interrupt_data->pending_handler = 0;
 447     th->no_tls_value_marker=initial_function;
 448
 449     th->stepping = NIL;
 450     return th;
 451 }
 452
 453 #ifdef LISP_FEATURE_MACH_EXCEPTION_HANDLER
 454 mach_port_t setup_mach_exception_handling_thread();
 455 kern_return_t mach_thread_init(mach_port_t thread_exception_port);
 456
 457 #endif
 458
 459 void create_initial_thread(lispobj initial_function) {
 460     struct thread *th=create_thread_struct(initial_function);
 461     if(th) {
 462 #ifdef LISP_FEATURE_MACH_EXCEPTION_HANDLER
 463         setup_mach_exception_handling_thread();
 464 #endif
 465         initial_thread_trampoline(th); /* no return */
 466     } else lose("can't create initial thread\n");
 467 }
 468
 469 #ifdef LISP_FEATURE_SB_THREAD
 470
 471 #ifndef __USE_XOPEN2K
 472 extern int pthread_attr_setstack (pthread_attr_t *__attr, void *__stackaddr,
 473                                   size_t __stacksize);
 474 #endif
 475
 476 boolean create_os_thread(struct thread *th,os_thread_t *kid_tid)
 477 {
 478     /* The new thread inherits the restrictive signal mask set here,
 479      * and enables signals again when it is set up properly. */
 480     sigset_t newset,oldset;
 481     boolean r=1;
 482     int retcode = 0, initcode;
 483
 484     FSHOW_SIGNAL((stderr,"/create_os_thread: creating new thread\n"));
 485
 486 #ifdef LOCK_CREATE_THREAD
 487     retcode = pthread_mutex_lock(&create_thread_lock);
 488     gc_assert(retcode == 0);
 489     FSHOW_SIGNAL((stderr,"/create_os_thread: got lock\n"));
 490 #endif
 491     sigemptyset(&newset);
 492     /* Blocking deferrable signals is enough, no need to block
 493      * SIG_STOP_FOR_GC because the child process is not linked onto
 494      * all_threads until it's ready. */
 495     sigaddset_deferrable(&newset);
 496     thread_sigmask(SIG_BLOCK, &newset, &oldset);
 497
 498     if((initcode = pthread_attr_init(th->os_attr)) ||
 499        /* call_into_lisp_first_time switches the stack for the initial thread. For the
 500         * others, we use this. */
 501        (pthread_attr_setstack(th->os_attr,th->control_stack_start,thread_control_stack_size)) ||
 502        (retcode = pthread_create
 503         (kid_tid,th->os_attr,(void *(*)(void *))new_thread_trampoline,th))) {
 504         FSHOW_SIGNAL((stderr, "init = %d\n", initcode));
 505         FSHOW_SIGNAL((stderr, printf("pthread_create returned %d, errno %d\n", retcode, errno)));
 506         FSHOW_SIGNAL((stderr, "wanted stack size %d, min stack size %d\n",
 507                       cstack_size, PTHREAD_STACK_MIN));
 508         if(retcode < 0) {
 509             perror("create_os_thread");
 510         }
 511         r=0;
 512     }
 513
 514     thread_sigmask(SIG_SETMASK,&oldset,0);
 515 #ifdef LOCK_CREATE_THREAD
 516     retcode = pthread_mutex_unlock(&create_thread_lock);
 517     gc_assert(retcode == 0);
 518     FSHOW_SIGNAL((stderr,"/create_os_thread: released lock\n"));
 519 #endif
 520     return r;
 521 }
 522
 523 os_thread_t create_thread(lispobj initial_function) {
 524     struct thread *th;
 525     os_thread_t kid_tid;
 526
 527     /* Assuming that a fresh thread struct has no lisp objects in it,
 528      * linking it to all_threads can be left to the thread itself
 529      * without fear of gc lossage. initial_function violates this
 530      * assumption and must stay pinned until the child starts up. */
 531     th = create_thread_struct(initial_function);
 532     if(th==0) return 0;
 533
 534     if (create_os_thread(th,&kid_tid)) {
 535         return kid_tid;
 536     } else {
 537         free_thread_struct(th);
 538         return 0;
 539     }
 540 }
 541
 542 /* Send the signo to os_thread, retry if the rt signal queue is
 543  * full. */
 544 int
 545 kill_thread_safely(os_thread_t os_thread, int signo)
 546 {
 547     int r;
 548     /* The man page does not mention EAGAIN as a valid return value
 549      * for either pthread_kill or kill. But that's theory, this is
 550      * practice. By waiting here we assume that the delivery of this
 551      * signal is not necessary for the delivery of the signals in the
 552      * queue. In other words, we _assume_ there are no deadlocks. */
 553     while ((r=pthread_kill(os_thread,signo))==EAGAIN) {
 554         /* wait a bit then try again in the hope of the rt signal
 555          * queue not being full */
 556         FSHOW_SIGNAL((stderr,"/rt signal queue full\n"));
 557         /* FIXME: some kind of backoff (random, exponential) would be
 558          * nice. */
 559         sleep(1);
 560     }
 561     return r;
 562 }
 563
 564 int signal_interrupt_thread(os_thread_t os_thread)
 565 {
 566     int status = kill_thread_safely(os_thread, SIG_INTERRUPT_THREAD);
 567     if (status == 0) {
 568         return 0;
 569     } else if (status == ESRCH) {
 570         return -1;
 571     } else {
 572         lose("cannot send SIG_INTERRUPT_THREAD to thread=%lu: %d, %s\n",
 573              os_thread, status, strerror(status));
 574     }
 575 }
 576
 577 /* stopping the world is a two-stage process.  From this thread we signal
 578  * all the others with SIG_STOP_FOR_GC.  The handler for this signal does
 579  * the usual pseudo-atomic checks (we don't want to stop a thread while
 580  * it's in the middle of allocation) then waits for another SIG_STOP_FOR_GC.
 581  */
 582
 583 /* To avoid deadlocks when gc stops the world all clients of each
 584  * mutex must enable or disable SIG_STOP_FOR_GC for the duration of
 585  * holding the lock, but they must agree on which. */
 586 void gc_stop_the_world()
 587 {
 588     struct thread *p,*th=arch_os_get_current_thread();
 589     int status, lock_ret;
 590 #ifdef LOCK_CREATE_THREAD
 591     /* KLUDGE: Stopping the thread during pthread_create() causes deadlock
 592      * on FreeBSD. */
 593     FSHOW_SIGNAL((stderr,"/gc_stop_the_world:waiting on create_thread_lock, thread=%lu\n",
 594                   th->os_thread));
 595     lock_ret = pthread_mutex_lock(&create_thread_lock);
 596     gc_assert(lock_ret == 0);
 597     FSHOW_SIGNAL((stderr,"/gc_stop_the_world:got create_thread_lock, thread=%lu\n",
 598                   th->os_thread));
 599 #endif
 600     FSHOW_SIGNAL((stderr,"/gc_stop_the_world:waiting on lock, thread=%lu\n",
 601                   th->os_thread));
 602     /* keep threads from starting while the world is stopped. */
 603     lock_ret = pthread_mutex_lock(&all_threads_lock);      \
 604     gc_assert(lock_ret == 0);
 605
 606     FSHOW_SIGNAL((stderr,"/gc_stop_the_world:got lock, thread=%lu\n",
 607                   th->os_thread));
 608     /* stop all other threads by sending them SIG_STOP_FOR_GC */
 609     for(p=all_threads; p; p=p->next) {
 610         gc_assert(p->os_thread != 0);
 611         FSHOW_SIGNAL((stderr,"/gc_stop_the_world: p->state: %x\n", p->state));
 612         if((p!=th) && ((p->state==STATE_RUNNING))) {
 613             FSHOW_SIGNAL((stderr,"/gc_stop_the_world: suspending %x, os_thread %x\n",
 614                           p, p->os_thread));
 615             status=kill_thread_safely(p->os_thread,SIG_STOP_FOR_GC);
 616             if (status==ESRCH) {
 617                 /* This thread has exited. */
 618                 gc_assert(p->state==STATE_DEAD);
 619             } else if (status) {
 620                 lose("cannot send suspend thread=%lu: %d, %s\n",
 621                      p->os_thread,status,strerror(status));
 622             }
 623         }
 624     }
 625     FSHOW_SIGNAL((stderr,"/gc_stop_the_world:signals sent\n"));
 626     /* wait for the running threads to stop or finish */
 627     for(p=all_threads;p;) {
 628         FSHOW_SIGNAL((stderr,"/gc_stop_the_world: th: %p, p: %p\n", th, p));
 629         if((p!=th) && (p->state==STATE_RUNNING)) {
 630             sched_yield();
 631         } else {
 632             p=p->next;
 633         }
 634     }
 635     FSHOW_SIGNAL((stderr,"/gc_stop_the_world:end\n"));
 636 }
 637
 638 void gc_start_the_world()
 639 {
 640     struct thread *p,*th=arch_os_get_current_thread();
 641     int status, lock_ret;
 642     /* if a resumed thread creates a new thread before we're done with
 643      * this loop, the new thread will get consed on the front of
 644      * all_threads, but it won't have been stopped so won't need
 645      * restarting */
 646     FSHOW_SIGNAL((stderr,"/gc_start_the_world:begin\n"));
 647     for(p=all_threads;p;p=p->next) {
 648         gc_assert(p->os_thread!=0);
 649         if((p!=th) && (p->state!=STATE_DEAD)) {
 650             if(p->state!=STATE_SUSPENDED) {
 651                 lose("gc_start_the_world: wrong thread state is %d\n",
 652                      fixnum_value(p->state));
 653             }
 654             FSHOW_SIGNAL((stderr, "/gc_start_the_world: resuming %lu\n",
 655                           p->os_thread));
 656             p->state=STATE_RUNNING;
 657
 658 #if defined(SIG_RESUME_FROM_GC)
 659             status=kill_thread_safely(p->os_thread,SIG_RESUME_FROM_GC);
 660 #else
 661             status=kill_thread_safely(p->os_thread,SIG_STOP_FOR_GC);
 662 #endif
 663             if (status) {
 664                 lose("cannot resume thread=%lu: %d, %s\n",
 665                      p->os_thread,status,strerror(status));
 666             }
 667         }
 668     }
 669     /* If we waited here until all threads leave STATE_SUSPENDED, then
 670      * SIG_STOP_FOR_GC wouldn't need to be a rt signal. That has some
 671      * performance implications, but does away with the 'rt signal
 672      * queue full' problem. */
 673
 674     lock_ret = pthread_mutex_unlock(&all_threads_lock);
 675     gc_assert(lock_ret == 0);
 676 #ifdef LOCK_CREATE_THREAD
 677     lock_ret = pthread_mutex_unlock(&create_thread_lock);
 678     gc_assert(lock_ret == 0);
 679 #endif
 680
 681     FSHOW_SIGNAL((stderr,"/gc_start_the_world:end\n"));
 682 }
 683 #endif
 684
 685 int
 686 thread_yield()
 687 {
 688 #ifdef LISP_FEATURE_SB_THREAD
 689     return sched_yield();
 690 #else
 691     return 0;
 692 #endif
 693 }