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 CONTROL_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 QUEUE_FREEABLE_THREAD_STACKS
  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 freeable_stack {
  69 #ifdef QUEUE_FREEABLE_THREAD_STACKS
  70     struct freeable_stack *next;
  71 #endif
  72     os_thread_t os_thread;
  73     os_vm_address_t os_address;
  74 };
  75
  76
  77 #ifdef QUEUE_FREEABLE_THREAD_STACKS
  78 static struct freeable_stack * volatile freeable_stack_queue = 0;
  79 static int freeable_stack_count = 0;
  80 pthread_mutex_t freeable_stack_lock = PTHREAD_MUTEX_INITIALIZER;
  81 #else
  82 static struct freeable_stack * volatile freeable_stack = 0;
  83 #endif
  84
  85 int dynamic_values_bytes=4096*sizeof(lispobj);  /* same for all threads */
  86 struct thread * volatile all_threads;
  87 extern struct interrupt_data * global_interrupt_data;
  88
  89 #ifdef LISP_FEATURE_SB_THREAD
  90 pthread_mutex_t all_threads_lock = PTHREAD_MUTEX_INITIALIZER;
  91 #ifdef LOCK_CREATE_THREAD
  92 static pthread_mutex_t create_thread_lock = PTHREAD_MUTEX_INITIALIZER;
  93 #endif
  94 #endif
  95
  96 #if defined(LISP_FEATURE_X86) || defined(LISP_FEATURE_X86_64)
  97 extern lispobj call_into_lisp_first_time(lispobj fun, lispobj *args, int nargs);
  98 #endif
  99
 100 static void
 101 link_thread(struct thread *th)
 102 {
 103     if (all_threads) all_threads->prev=th;
 104     th->next=all_threads;
 105     th->prev=0;
 106     all_threads=th;
 107 }
 108
 109 #ifdef LISP_FEATURE_SB_THREAD
 110 static void
 111 unlink_thread(struct thread *th)
 112 {
 113     if (th->prev)
 114         th->prev->next = th->next;
 115     else
 116         all_threads = th->next;
 117     if (th->next)
 118         th->next->prev = th->prev;
 119 }
 120 #endif
 121
 122 static int
 123 initial_thread_trampoline(struct thread *th)
 124 {
 125     lispobj function;
 126 #if defined(LISP_FEATURE_X86) || defined(LISP_FEATURE_X86_64)
 127     lispobj *args = NULL;
 128 #endif
 129     function = th->no_tls_value_marker;
 130     th->no_tls_value_marker = NO_TLS_VALUE_MARKER_WIDETAG;
 131     if(arch_os_thread_init(th)==0) return 1;
 132     link_thread(th);
 133     th->os_thread=thread_self();
 134 #ifndef LISP_FEATURE_WIN32
 135     protect_control_stack_guard_page(1);
 136 #endif
 137
 138 #if defined(LISP_FEATURE_X86) || defined(LISP_FEATURE_X86_64)
 139     return call_into_lisp_first_time(function,args,0);
 140 #else
 141     return funcall0(function);
 142 #endif
 143 }
 144
 145 #define THREAD_STRUCT_SIZE (THREAD_CONTROL_STACK_SIZE + BINDING_STACK_SIZE + \
 146                             ALIEN_STACK_SIZE + dynamic_values_bytes + \
 147                             32 * SIGSTKSZ + \
 148                             BACKEND_PAGE_SIZE)
 149
 150 #ifdef LISP_FEATURE_SB_THREAD
 151
 152 #ifdef QUEUE_FREEABLE_THREAD_STACKS
 153
 154 static void
 155 queue_freeable_thread_stack(struct thread *thread_to_be_cleaned_up)
 156 {
 157      struct freeable_stack *new_freeable_stack = 0;
 158      if (thread_to_be_cleaned_up) {
 159         /* FIXME: os_validate is mmap -- for small things like these
 160          * malloc would probably perform better. */
 161         new_freeable_stack = (struct freeable_stack *)
 162             os_validate(0, sizeof(struct freeable_stack));
 163         new_freeable_stack->next = NULL;
 164         new_freeable_stack->os_thread = thread_to_be_cleaned_up->os_thread;
 165         new_freeable_stack->os_address = thread_to_be_cleaned_up->os_address;
 166         pthread_mutex_lock(&freeable_stack_lock);
 167         if (freeable_stack_queue) {
 168             struct freeable_stack *next;
 169             next = freeable_stack_queue;
 170             while (next->next) {
 171                 next = next->next;
 172             }
 173             next->next = new_freeable_stack;
 174         } else {
 175             freeable_stack_queue = new_freeable_stack;
 176         }
 177         freeable_stack_count++;
 178         pthread_mutex_unlock(&freeable_stack_lock);
 179     }
 180 }
 181
 182 #define FREEABLE_STACK_QUEUE_SIZE 4
 183
 184 static void
 185 free_freeable_stacks() {
 186     if (freeable_stack_queue && (freeable_stack_count > FREEABLE_STACK_QUEUE_SIZE)) {
 187         struct freeable_stack* old;
 188         pthread_mutex_lock(&freeable_stack_lock);
 189         old = freeable_stack_queue;
 190         freeable_stack_queue = old->next;
 191         freeable_stack_count--;
 192         gc_assert(pthread_join(old->os_thread, NULL) == 0);
 193         FSHOW((stderr, "freeing thread %x stack\n", old->os_thread));
 194         os_invalidate(old->os_address, THREAD_STRUCT_SIZE);
 195         os_invalidate((os_vm_address_t)old, sizeof(struct freeable_stack));
 196         pthread_mutex_unlock(&freeable_stack_lock);
 197     }
 198 }
 199
 200 #elif defined(CREATE_CLEANUP_THREAD)
 201 static void *
 202 cleanup_thread(void *arg)
 203 {
 204     struct freeable_stack *freeable = arg;
 205     pthread_t self = pthread_self();
 206
 207     FSHOW((stderr, "/cleaner thread(%p): joining %p\n",
 208            self, freeable->os_thread));
 209     gc_assert(pthread_join(freeable->os_thread, NULL) == 0);
 210     FSHOW((stderr, "/cleaner thread(%p): free stack %p\n",
 211            self, freeable->stack));
 212     os_invalidate(freeable->os_address, THREAD_STRUCT_SIZE);
 213     free(freeable);
 214
 215     pthread_detach(self);
 216
 217     return NULL;
 218 }
 219
 220 static void
 221 create_cleanup_thread(struct thread *thread_to_be_cleaned_up)
 222 {
 223     pthread_t thread;
 224     int result;
 225
 226     if (thread_to_be_cleaned_up) {
 227         struct freeable_stack *freeable =
 228             malloc(sizeof(struct freeable_stack));
 229         gc_assert(freeable != NULL);
 230         freeable->os_thread = thread_to_be_cleaned_up->os_thread;
 231         freeable->os_address =
 232             (os_vm_address_t) thread_to_be_cleaned_up->os_address;
 233         result = pthread_create(&thread, NULL, cleanup_thread, freeable);
 234         gc_assert(result == 0);
 235         sched_yield();
 236     }
 237 }
 238
 239 #else
 240 static void
 241 free_thread_stack_later(struct thread *thread_to_be_cleaned_up)
 242 {
 243     struct freeable_stack *new_freeable_stack = 0;
 244     if (thread_to_be_cleaned_up) {
 245         new_freeable_stack = (struct freeable_stack *)
 246             os_validate(0, sizeof(struct freeable_stack));
 247         new_freeable_stack->os_thread = thread_to_be_cleaned_up->os_thread;
 248         new_freeable_stack->os_address = (os_vm_address_t)
 249             thread_to_be_cleaned_up->os_address;
 250     }
 251     new_freeable_stack = (struct freeable_stack *)
 252         swap_lispobjs((lispobj *)(void *)&freeable_stack,
 253                       (lispobj)new_freeable_stack);
 254     if (new_freeable_stack) {
 255         FSHOW((stderr,"/reaping %p\n", (void*) new_freeable_stack->os_thread));
 256         /* Under NPTL pthread_join really waits until the thread
 257          * exists and the stack can be safely freed. This is sadly not
 258          * mandated by the pthread spec. */
 259         gc_assert(pthread_join(new_freeable_stack->os_thread, NULL) == 0);
 260         os_invalidate(new_freeable_stack->os_address, THREAD_STRUCT_SIZE);
 261         os_invalidate((os_vm_address_t) new_freeable_stack,
 262                       sizeof(struct freeable_stack));
 263     }
 264 }
 265 #endif
 266
 267 /* this is the first thing that runs in the child (which is why the
 268  * silly calling convention).  Basically it calls the user's requested
 269  * lisp function after doing arch_os_thread_init and whatever other
 270  * bookkeeping needs to be done
 271  */
 272 int
 273 new_thread_trampoline(struct thread *th)
 274 {
 275     lispobj function;
 276     int result, lock_ret;
 277
 278     FSHOW((stderr,"/creating thread %lu\n", thread_self()));
 279     function = th->no_tls_value_marker;
 280     th->no_tls_value_marker = NO_TLS_VALUE_MARKER_WIDETAG;
 281     if(arch_os_thread_init(th)==0) {
 282         /* FIXME: handle error */
 283         lose("arch_os_thread_init failed\n");
 284     }
 285
 286     th->os_thread=thread_self();
 287     protect_control_stack_guard_page(1);
 288     /* Since GC can only know about this thread from the all_threads
 289      * list and we're just adding this thread to it there is no danger
 290      * of deadlocking even with SIG_STOP_FOR_GC blocked (which it is
 291      * not). */
 292     lock_ret = pthread_mutex_lock(&all_threads_lock);
 293     gc_assert(lock_ret == 0);
 294     link_thread(th);
 295     lock_ret = pthread_mutex_unlock(&all_threads_lock);
 296     gc_assert(lock_ret == 0);
 297
 298     result = funcall0(function);
 299
 300     /* Block GC */
 301     block_blockable_signals();
 302     th->state=STATE_DEAD;
 303
 304     /* SIG_STOP_FOR_GC is blocked and GC might be waiting for this
 305      * thread, but since we are already dead it won't wait long. */
 306     lock_ret = pthread_mutex_lock(&all_threads_lock);
 307     gc_assert(lock_ret == 0);
 308
 309     gc_alloc_update_page_tables(0, &th->alloc_region);
 310     unlink_thread(th);
 311     pthread_mutex_unlock(&all_threads_lock);
 312     gc_assert(lock_ret == 0);
 313
 314     if(th->tls_cookie>=0) arch_os_thread_cleanup(th);
 315     os_invalidate((os_vm_address_t)th->interrupt_data,
 316                   (sizeof (struct interrupt_data)));
 317
 318 #ifdef LISP_FEATURE_MACH_EXCEPTION_HANDLER
 319     FSHOW((stderr, "Deallocating mach port %x\n", THREAD_STRUCT_TO_EXCEPTION_PORT(th)));
 320     mach_port_move_member(mach_task_self(),
 321                           THREAD_STRUCT_TO_EXCEPTION_PORT(th),
 322                           MACH_PORT_NULL);
 323     mach_port_deallocate(mach_task_self(),
 324                          THREAD_STRUCT_TO_EXCEPTION_PORT(th));
 325     mach_port_destroy(mach_task_self(),
 326                       THREAD_STRUCT_TO_EXCEPTION_PORT(th));
 327 #endif
 328
 329 #ifdef QUEUE_FREEABLE_THREAD_STACKS
 330     queue_freeable_thread_stack(th);
 331 #elif defined(CREATE_CLEANUP_THREAD)
 332     create_cleanup_thread(th);
 333 #else
 334     free_thread_stack_later(th);
 335 #endif
 336
 337     FSHOW((stderr,"/exiting thread %p\n", thread_self()));
 338     return result;
 339 }
 340
 341 #endif /* LISP_FEATURE_SB_THREAD */
 342
 343 static void
 344 free_thread_struct(struct thread *th)
 345 {
 346     if (th->interrupt_data)
 347         os_invalidate((os_vm_address_t) th->interrupt_data,
 348                       (sizeof (struct interrupt_data)));
 349     os_invalidate((os_vm_address_t) th->os_address,
 350                   THREAD_STRUCT_SIZE);
 351 }
 352
 353 /* this is called from any other thread to create the new one, and
 354  * initialize all parts of it that can be initialized from another
 355  * thread
 356  */
 357
 358 static struct thread *
 359 create_thread_struct(lispobj initial_function) {
 360     union per_thread_data *per_thread;
 361     struct thread *th=0;        /*  subdue gcc */
 362     void *spaces=0;
 363     void *aligned_spaces=0;
 364 #ifdef LISP_FEATURE_SB_THREAD
 365     int i;
 366 #endif
 367
 368 #ifdef CREATE_CLEANUP_THREAD
 369     /* Give a chance for cleanup threads to run. */
 370     sched_yield();
 371 #endif
 372     /* May as well allocate all the spaces at once: it saves us from
 373      * having to decide what to do if only some of the allocations
 374      * succeed.  SPACES must be page-aligned, since the GC expects the
 375      * control stack to start at a page boundary.  We can't rely on the
 376      * alignment passed from os_validate, since that might assume the
 377      * current (e.g. 4k) pagesize, while we calculate with the biggest
 378      * (e.g. 64k) pagesize allowed by the ABI.  */
 379     spaces=os_validate(0, THREAD_STRUCT_SIZE);
 380     if(!spaces)
 381         return NULL;
 382     /* Aligning up is safe as THREAD_STRUCT_SIZE has BACKEND_PAGE_SIZE
 383      * padding. */
 384     aligned_spaces = (void *)((((unsigned long)(char *)spaces)
 385                                + BACKEND_PAGE_SIZE - 1)
 386                               & ~(unsigned long)(BACKEND_PAGE_SIZE - 1));
 387     per_thread=(union per_thread_data *)
 388         (aligned_spaces+
 389          THREAD_CONTROL_STACK_SIZE+
 390          BINDING_STACK_SIZE+
 391          ALIEN_STACK_SIZE);
 392
 393 #ifdef LISP_FEATURE_SB_THREAD
 394     for(i = 0; i < (dynamic_values_bytes / sizeof(lispobj)); i++)
 395         per_thread->dynamic_values[i] = NO_TLS_VALUE_MARKER_WIDETAG;
 396     if (all_threads == 0) {
 397         if(SymbolValue(FREE_TLS_INDEX,0)==UNBOUND_MARKER_WIDETAG) {
 398             SetSymbolValue
 399                 (FREE_TLS_INDEX,
 400                  /* FIXME: should be MAX_INTERRUPTS -1 ? */
 401                  make_fixnum(MAX_INTERRUPTS+
 402                              sizeof(struct thread)/sizeof(lispobj)),
 403                  0);
 404             SetSymbolValue(TLS_INDEX_LOCK,make_fixnum(0),0);
 405         }
 406 #define STATIC_TLS_INIT(sym,field) \
 407   ((struct symbol *)(sym-OTHER_POINTER_LOWTAG))->tls_index= \
 408   make_fixnum(THREAD_SLOT_OFFSET_WORDS(field))
 409
 410         STATIC_TLS_INIT(BINDING_STACK_START,binding_stack_start);
 411         STATIC_TLS_INIT(BINDING_STACK_POINTER,binding_stack_pointer);
 412         STATIC_TLS_INIT(CONTROL_STACK_START,control_stack_start);
 413         STATIC_TLS_INIT(CONTROL_STACK_END,control_stack_end);
 414         STATIC_TLS_INIT(ALIEN_STACK,alien_stack_pointer);
 415 #if defined(LISP_FEATURE_X86) || defined (LISP_FEATURE_X86_64)
 416         STATIC_TLS_INIT(PSEUDO_ATOMIC_BITS,pseudo_atomic_bits);
 417 #endif
 418 #undef STATIC_TLS_INIT
 419     }
 420 #endif
 421
 422     th=&per_thread->thread;
 423     th->os_address = spaces;
 424     th->control_stack_start = aligned_spaces;
 425     th->binding_stack_start=
 426         (lispobj*)((void*)th->control_stack_start+THREAD_CONTROL_STACK_SIZE);
 427     th->control_stack_end = th->binding_stack_start;
 428     th->alien_stack_start=
 429         (lispobj*)((void*)th->binding_stack_start+BINDING_STACK_SIZE);
 430     th->binding_stack_pointer=th->binding_stack_start;
 431     th->this=th;
 432     th->os_thread=0;
 433     th->state=STATE_RUNNING;
 434 #ifdef LISP_FEATURE_STACK_GROWS_DOWNWARD_NOT_UPWARD
 435     th->alien_stack_pointer=((void *)th->alien_stack_start
 436                              + ALIEN_STACK_SIZE-N_WORD_BYTES);
 437 #else
 438     th->alien_stack_pointer=((void *)th->alien_stack_start);
 439 #endif
 440 #if defined(LISP_FEATURE_X86) || defined (LISP_FEATURE_X86_64)
 441     th->pseudo_atomic_bits=0;
 442 #endif
 443 #ifdef LISP_FEATURE_GENCGC
 444     gc_set_region_empty(&th->alloc_region);
 445 #endif
 446
 447 #ifndef LISP_FEATURE_SB_THREAD
 448     /* the tls-points-into-struct-thread trick is only good for threaded
 449      * sbcl, because unithread sbcl doesn't have tls.  So, we copy the
 450      * appropriate values from struct thread here, and make sure that
 451      * we use the appropriate SymbolValue macros to access any of the
 452      * variable quantities from the C runtime.  It's not quite OAOOM,
 453      * it just feels like it */
 454     SetSymbolValue(BINDING_STACK_START,(lispobj)th->binding_stack_start,th);
 455     SetSymbolValue(CONTROL_STACK_START,(lispobj)th->control_stack_start,th);
 456     SetSymbolValue(CONTROL_STACK_END,(lispobj)th->control_stack_end,th);
 457 #if defined(LISP_FEATURE_X86) || defined (LISP_FEATURE_X86_64)
 458     SetSymbolValue(BINDING_STACK_POINTER,(lispobj)th->binding_stack_pointer,th);
 459     SetSymbolValue(ALIEN_STACK,(lispobj)th->alien_stack_pointer,th);
 460     SetSymbolValue(PSEUDO_ATOMIC_BITS,(lispobj)th->pseudo_atomic_bits,th);
 461 #else
 462     current_binding_stack_pointer=th->binding_stack_pointer;
 463     current_control_stack_pointer=th->control_stack_start;
 464 #endif
 465 #endif
 466     bind_variable(CURRENT_CATCH_BLOCK,make_fixnum(0),th);
 467     bind_variable(CURRENT_UNWIND_PROTECT_BLOCK,make_fixnum(0),th);
 468     bind_variable(FREE_INTERRUPT_CONTEXT_INDEX,make_fixnum(0),th);
 469     bind_variable(INTERRUPT_PENDING, NIL,th);
 470     bind_variable(INTERRUPTS_ENABLED,T,th);
 471     bind_variable(ALLOW_WITH_INTERRUPTS,T,th);
 472     bind_variable(GC_PENDING,NIL,th);
 473 #ifdef LISP_FEATURE_SB_THREAD
 474     bind_variable(STOP_FOR_GC_PENDING,NIL,th);
 475 #endif
 476
 477     th->interrupt_data = (struct interrupt_data *)
 478         os_validate(0,(sizeof (struct interrupt_data)));
 479     if (!th->interrupt_data) {
 480         free_thread_struct(th);
 481         return 0;
 482     }
 483     th->interrupt_data->pending_handler = 0;
 484     th->no_tls_value_marker=initial_function;
 485
 486     th->stepping = NIL;
 487     return th;
 488 }
 489
 490 #ifdef LISP_FEATURE_MACH_EXCEPTION_HANDLER
 491 mach_port_t setup_mach_exception_handling_thread();
 492 kern_return_t mach_thread_init(mach_port_t thread_exception_port);
 493
 494 #endif
 495
 496 void create_initial_thread(lispobj initial_function) {
 497     struct thread *th=create_thread_struct(initial_function);
 498     if(th) {
 499 #ifdef LISP_FEATURE_MACH_EXCEPTION_HANDLER
 500         kern_return_t ret;
 501
 502         setup_mach_exception_handling_thread();
 503 #endif
 504         initial_thread_trampoline(th); /* no return */
 505     } else lose("can't create initial thread\n");
 506 }
 507
 508 #ifdef LISP_FEATURE_SB_THREAD
 509
 510 #ifndef __USE_XOPEN2K
 511 extern int pthread_attr_setstack (pthread_attr_t *__attr, void *__stackaddr,
 512                                   size_t __stacksize);
 513 #endif
 514
 515 boolean create_os_thread(struct thread *th,os_thread_t *kid_tid)
 516 {
 517     /* The new thread inherits the restrictive signal mask set here,
 518      * and enables signals again when it is set up properly. */
 519     pthread_attr_t attr;
 520     sigset_t newset,oldset;
 521     boolean r=1;
 522     int retcode, initcode, sizecode, addrcode;
 523
 524     FSHOW_SIGNAL((stderr,"/create_os_thread: creating new thread\n"));
 525
 526 #ifdef LOCK_CREATE_THREAD
 527     retcode = pthread_mutex_lock(&create_thread_lock);
 528     gc_assert(retcode == 0);
 529     FSHOW_SIGNAL((stderr,"/create_os_thread: got lock\n"));
 530 #endif
 531     sigemptyset(&newset);
 532     /* Blocking deferrable signals is enough, no need to block
 533      * SIG_STOP_FOR_GC because the child process is not linked onto
 534      * all_threads until it's ready. */
 535     sigaddset_deferrable(&newset);
 536     thread_sigmask(SIG_BLOCK, &newset, &oldset);
 537
 538 #if defined(LISP_FEATURE_DARWIN)
 539 #define CONTROL_STACK_ADJUST 8192 /* darwin wants page-aligned stacks */
 540 #else
 541 #define CONTROL_STACK_ADJUST 16
 542 #endif
 543
 544     if((initcode = pthread_attr_init(&attr)) ||
 545        /* FIXME: why do we even have this in the first place? */
 546        (pthread_attr_setstack(&attr,th->control_stack_start,
 547                               THREAD_CONTROL_STACK_SIZE-CONTROL_STACK_ADJUST)) ||
 548 #undef CONTROL_STACK_ADJUST
 549        (retcode = pthread_create
 550         (kid_tid,&attr,(void *(*)(void *))new_thread_trampoline,th))) {
 551         FSHOW_SIGNAL((stderr, "init, size, addr = %d, %d, %d\n", initcode, sizecode, addrcode));
 552         FSHOW_SIGNAL((stderr, printf("pthread_create returned %d, errno %d\n", retcode, errno)));
 553         FSHOW_SIGNAL((stderr, "wanted stack size %d, min stack size %d\n",
 554                       THREAD_CONTROL_STACK_SIZE-16, PTHREAD_STACK_MIN));
 555         if(retcode < 0) {
 556             perror("create_os_thread");
 557         }
 558         r=0;
 559     }
 560
 561 #ifdef QUEUE_FREEABLE_THREAD_STACKS
 562     free_freeable_stacks();
 563 #endif
 564     thread_sigmask(SIG_SETMASK,&oldset,0);
 565 #ifdef LOCK_CREATE_THREAD
 566     retcode = pthread_mutex_unlock(&create_thread_lock);
 567     gc_assert(retcode == 0);
 568     FSHOW_SIGNAL((stderr,"/create_os_thread: released lock\n"));
 569 #endif
 570     return r;
 571 }
 572
 573 os_thread_t create_thread(lispobj initial_function) {
 574     struct thread *th;
 575     os_thread_t kid_tid;
 576
 577     /* Assuming that a fresh thread struct has no lisp objects in it,
 578      * linking it to all_threads can be left to the thread itself
 579      * without fear of gc lossage. initial_function violates this
 580      * assumption and must stay pinned until the child starts up. */
 581     th = create_thread_struct(initial_function);
 582     if(th==0) return 0;
 583
 584     if (create_os_thread(th,&kid_tid)) {
 585         return kid_tid;
 586     } else {
 587         free_thread_struct(th);
 588         return 0;
 589     }
 590 }
 591
 592 /* Send the signo to os_thread, retry if the rt signal queue is
 593  * full. */
 594 int
 595 kill_thread_safely(os_thread_t os_thread, int signo)
 596 {
 597     int r;
 598     /* The man page does not mention EAGAIN as a valid return value
 599      * for either pthread_kill or kill. But that's theory, this is
 600      * practice. By waiting here we assume that the delivery of this
 601      * signal is not necessary for the delivery of the signals in the
 602      * queue. In other words, we _assume_ there are no deadlocks. */
 603     while ((r=pthread_kill(os_thread,signo))==EAGAIN) {
 604         /* wait a bit then try again in the hope of the rt signal
 605          * queue not being full */
 606         FSHOW_SIGNAL((stderr,"/rt signal queue full\n"));
 607         /* FIXME: some kind of backoff (random, exponential) would be
 608          * nice. */
 609         sleep(1);
 610     }
 611     return r;
 612 }
 613
 614 int signal_interrupt_thread(os_thread_t os_thread)
 615 {
 616     int status = kill_thread_safely(os_thread, SIG_INTERRUPT_THREAD);
 617     if (status == 0) {
 618         return 0;
 619     } else if (status == ESRCH) {
 620         return -1;
 621     } else {
 622         lose("cannot send SIG_INTERRUPT_THREAD to thread=%lu: %d, %s\n",
 623              os_thread, status, strerror(status));
 624     }
 625 }
 626
 627 /* stopping the world is a two-stage process.  From this thread we signal
 628  * all the others with SIG_STOP_FOR_GC.  The handler for this signal does
 629  * the usual pseudo-atomic checks (we don't want to stop a thread while
 630  * it's in the middle of allocation) then waits for another SIG_STOP_FOR_GC.
 631  */
 632
 633 /* To avoid deadlocks when gc stops the world all clients of each
 634  * mutex must enable or disable SIG_STOP_FOR_GC for the duration of
 635  * holding the lock, but they must agree on which. */
 636 void gc_stop_the_world()
 637 {
 638     struct thread *p,*th=arch_os_get_current_thread();
 639     int status, lock_ret;
 640 #ifdef LOCK_CREATE_THREAD
 641     /* KLUDGE: Stopping the thread during pthread_create() causes deadlock
 642      * on FreeBSD. */
 643     FSHOW_SIGNAL((stderr,"/gc_stop_the_world:waiting on create_thread_lock, thread=%lu\n",
 644                   th->os_thread));
 645     lock_ret = pthread_mutex_lock(&create_thread_lock);
 646     gc_assert(lock_ret == 0);
 647     FSHOW_SIGNAL((stderr,"/gc_stop_the_world:got create_thread_lock, thread=%lu\n",
 648                   th->os_thread));
 649 #endif
 650     FSHOW_SIGNAL((stderr,"/gc_stop_the_world:waiting on lock, thread=%lu\n",
 651                   th->os_thread));
 652     /* keep threads from starting while the world is stopped. */
 653     lock_ret = pthread_mutex_lock(&all_threads_lock);      \
 654     gc_assert(lock_ret == 0);
 655
 656     FSHOW_SIGNAL((stderr,"/gc_stop_the_world:got lock, thread=%lu\n",
 657                   th->os_thread));
 658     /* stop all other threads by sending them SIG_STOP_FOR_GC */
 659     for(p=all_threads; p; p=p->next) {
 660         gc_assert(p->os_thread != 0);
 661         FSHOW_SIGNAL((stderr,"/gc_stop_the_world: p->state: %x\n", p->state));
 662         if((p!=th) && ((p->state==STATE_RUNNING))) {
 663             FSHOW_SIGNAL((stderr,"/gc_stop_the_world: suspending %x, os_thread %x\n",
 664                           p, p->os_thread));
 665             status=kill_thread_safely(p->os_thread,SIG_STOP_FOR_GC);
 666             if (status==ESRCH) {
 667                 /* This thread has exited. */
 668                 gc_assert(p->state==STATE_DEAD);
 669             } else if (status) {
 670                 lose("cannot send suspend thread=%lu: %d, %s\n",
 671                      p->os_thread,status,strerror(status));
 672             }
 673         }
 674     }
 675     FSHOW_SIGNAL((stderr,"/gc_stop_the_world:signals sent\n"));
 676     /* wait for the running threads to stop or finish */
 677     for(p=all_threads;p;) {
 678         FSHOW_SIGNAL((stderr,"/gc_stop_the_world: th: %p, p: %p\n", th, p));
 679         if((p!=th) && (p->state==STATE_RUNNING)) {
 680             sched_yield();
 681         } else {
 682             p=p->next;
 683         }
 684     }
 685     FSHOW_SIGNAL((stderr,"/gc_stop_the_world:end\n"));
 686 }
 687
 688 void gc_start_the_world()
 689 {
 690     struct thread *p,*th=arch_os_get_current_thread();
 691     int status, lock_ret;
 692     /* if a resumed thread creates a new thread before we're done with
 693      * this loop, the new thread will get consed on the front of
 694      * all_threads, but it won't have been stopped so won't need
 695      * restarting */
 696     FSHOW_SIGNAL((stderr,"/gc_start_the_world:begin\n"));
 697     for(p=all_threads;p;p=p->next) {
 698         gc_assert(p->os_thread!=0);
 699         if((p!=th) && (p->state!=STATE_DEAD)) {
 700             if(p->state!=STATE_SUSPENDED) {
 701                 lose("gc_start_the_world: wrong thread state is %d\n",
 702                      fixnum_value(p->state));
 703             }
 704             FSHOW_SIGNAL((stderr, "/gc_start_the_world: resuming %lu\n",
 705                           p->os_thread));
 706             p->state=STATE_RUNNING;
 707
 708 #if defined(SIG_RESUME_FROM_GC)
 709             status=kill_thread_safely(p->os_thread,SIG_RESUME_FROM_GC);
 710 #else
 711             status=kill_thread_safely(p->os_thread,SIG_STOP_FOR_GC);
 712 #endif
 713             if (status) {
 714                 lose("cannot resume thread=%lu: %d, %s\n",
 715                      p->os_thread,status,strerror(status));
 716             }
 717         }
 718     }
 719     /* If we waited here until all threads leave STATE_SUSPENDED, then
 720      * SIG_STOP_FOR_GC wouldn't need to be a rt signal. That has some
 721      * performance implications, but does away with the 'rt signal
 722      * queue full' problem. */
 723
 724     lock_ret = pthread_mutex_unlock(&all_threads_lock);
 725     gc_assert(lock_ret == 0);
 726 #ifdef LOCK_CREATE_THREAD
 727     lock_ret = pthread_mutex_unlock(&create_thread_lock);
 728     gc_assert(lock_ret == 0);
 729 #endif
 730
 731     FSHOW_SIGNAL((stderr,"/gc_start_the_world:end\n"));
 732 }
 733 #endif