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 #ifdef LISP_FEATURE_GCC_TLS
  95 __thread struct thread *current_thread;
  96 #endif
  97 #endif
  98
  99 #if defined(LISP_FEATURE_X86) || defined(LISP_FEATURE_X86_64)
 100 extern lispobj call_into_lisp_first_time(lispobj fun, lispobj *args, int nargs);
 101 #endif
 102
 103 static void
 104 link_thread(struct thread *th)
 105 {
 106     if (all_threads) all_threads->prev=th;
 107     th->next=all_threads;
 108     th->prev=0;
 109     all_threads=th;
 110 }
 111
 112 #ifdef LISP_FEATURE_SB_THREAD
 113 static void
 114 unlink_thread(struct thread *th)
 115 {
 116     if (th->prev)
 117         th->prev->next = th->next;
 118     else
 119         all_threads = th->next;
 120     if (th->next)
 121         th->next->prev = th->prev;
 122 }
 123 #endif
 124
 125 static int
 126 initial_thread_trampoline(struct thread *th)
 127 {
 128     lispobj function;
 129 #if defined(LISP_FEATURE_X86) || defined(LISP_FEATURE_X86_64)
 130     lispobj *args = NULL;
 131 #endif
 132     function = th->no_tls_value_marker;
 133     th->no_tls_value_marker = NO_TLS_VALUE_MARKER_WIDETAG;
 134     if(arch_os_thread_init(th)==0) return 1;
 135     link_thread(th);
 136     th->os_thread=thread_self();
 137 #ifndef LISP_FEATURE_WIN32
 138     protect_control_stack_guard_page(1);
 139 #endif
 140
 141 #if defined(LISP_FEATURE_X86) || defined(LISP_FEATURE_X86_64)
 142     return call_into_lisp_first_time(function,args,0);
 143 #else
 144     return funcall0(function);
 145 #endif
 146 }
 147
 148 #define THREAD_STRUCT_SIZE (THREAD_CONTROL_STACK_SIZE + BINDING_STACK_SIZE + \
 149                             ALIEN_STACK_SIZE + dynamic_values_bytes + \
 150                             32 * SIGSTKSZ + \
 151                             BACKEND_PAGE_SIZE)
 152
 153 #ifdef LISP_FEATURE_SB_THREAD
 154
 155 #ifdef QUEUE_FREEABLE_THREAD_STACKS
 156
 157 static void
 158 queue_freeable_thread_stack(struct thread *thread_to_be_cleaned_up)
 159 {
 160      struct freeable_stack *new_freeable_stack = 0;
 161      if (thread_to_be_cleaned_up) {
 162         /* FIXME: os_validate is mmap -- for small things like these
 163          * malloc would probably perform better. */
 164         new_freeable_stack = (struct freeable_stack *)
 165             os_validate(0, sizeof(struct freeable_stack));
 166         new_freeable_stack->next = NULL;
 167         new_freeable_stack->os_thread = thread_to_be_cleaned_up->os_thread;
 168         new_freeable_stack->os_address = thread_to_be_cleaned_up->os_address;
 169         pthread_mutex_lock(&freeable_stack_lock);
 170         if (freeable_stack_queue) {
 171             struct freeable_stack *next;
 172             next = freeable_stack_queue;
 173             while (next->next) {
 174                 next = next->next;
 175             }
 176             next->next = new_freeable_stack;
 177         } else {
 178             freeable_stack_queue = new_freeable_stack;
 179         }
 180         freeable_stack_count++;
 181         pthread_mutex_unlock(&freeable_stack_lock);
 182     }
 183 }
 184
 185 #define FREEABLE_STACK_QUEUE_SIZE 4
 186
 187 static void
 188 free_freeable_stacks() {
 189     if (freeable_stack_queue && (freeable_stack_count > FREEABLE_STACK_QUEUE_SIZE)) {
 190         struct freeable_stack* old;
 191         pthread_mutex_lock(&freeable_stack_lock);
 192         old = freeable_stack_queue;
 193         freeable_stack_queue = old->next;
 194         freeable_stack_count--;
 195         gc_assert(pthread_join(old->os_thread, NULL) == 0);
 196         FSHOW((stderr, "freeing thread %x stack\n", old->os_thread));
 197         os_invalidate(old->os_address, THREAD_STRUCT_SIZE);
 198         os_invalidate((os_vm_address_t)old, sizeof(struct freeable_stack));
 199         pthread_mutex_unlock(&freeable_stack_lock);
 200     }
 201 }
 202
 203 #elif defined(CREATE_CLEANUP_THREAD)
 204 static void *
 205 cleanup_thread(void *arg)
 206 {
 207     struct freeable_stack *freeable = arg;
 208     pthread_t self = pthread_self();
 209
 210     FSHOW((stderr, "/cleaner thread(%p): joining %p\n",
 211            self, freeable->os_thread));
 212     gc_assert(pthread_join(freeable->os_thread, NULL) == 0);
 213     FSHOW((stderr, "/cleaner thread(%p): free stack %p\n",
 214            self, freeable->stack));
 215     os_invalidate(freeable->os_address, THREAD_STRUCT_SIZE);
 216     free(freeable);
 217
 218     pthread_detach(self);
 219
 220     return NULL;
 221 }
 222
 223 static void
 224 create_cleanup_thread(struct thread *thread_to_be_cleaned_up)
 225 {
 226     pthread_t thread;
 227     int result;
 228
 229     if (thread_to_be_cleaned_up) {
 230         struct freeable_stack *freeable =
 231             malloc(sizeof(struct freeable_stack));
 232         gc_assert(freeable != NULL);
 233         freeable->os_thread = thread_to_be_cleaned_up->os_thread;
 234         freeable->os_address =
 235             (os_vm_address_t) thread_to_be_cleaned_up->os_address;
 236         result = pthread_create(&thread, NULL, cleanup_thread, freeable);
 237         gc_assert(result == 0);
 238         sched_yield();
 239     }
 240 }
 241
 242 #else
 243 static void
 244 free_thread_stack_later(struct thread *thread_to_be_cleaned_up)
 245 {
 246     struct freeable_stack *new_freeable_stack = 0;
 247     if (thread_to_be_cleaned_up) {
 248         new_freeable_stack = (struct freeable_stack *)
 249             os_validate(0, sizeof(struct freeable_stack));
 250         new_freeable_stack->os_thread = thread_to_be_cleaned_up->os_thread;
 251         new_freeable_stack->os_address = (os_vm_address_t)
 252             thread_to_be_cleaned_up->os_address;
 253     }
 254     new_freeable_stack = (struct freeable_stack *)
 255         swap_lispobjs((lispobj *)(void *)&freeable_stack,
 256                       (lispobj)new_freeable_stack);
 257     if (new_freeable_stack) {
 258         FSHOW((stderr,"/reaping %p\n", (void*) new_freeable_stack->os_thread));
 259         /* Under NPTL pthread_join really waits until the thread
 260          * exists and the stack can be safely freed. This is sadly not
 261          * mandated by the pthread spec. */
 262         gc_assert(pthread_join(new_freeable_stack->os_thread, NULL) == 0);
 263         os_invalidate(new_freeable_stack->os_address, THREAD_STRUCT_SIZE);
 264         os_invalidate((os_vm_address_t) new_freeable_stack,
 265                       sizeof(struct freeable_stack));
 266     }
 267 }
 268 #endif
 269
 270 /* this is the first thing that runs in the child (which is why the
 271  * silly calling convention).  Basically it calls the user's requested
 272  * lisp function after doing arch_os_thread_init and whatever other
 273  * bookkeeping needs to be done
 274  */
 275 int
 276 new_thread_trampoline(struct thread *th)
 277 {
 278     lispobj function;
 279     int result, lock_ret;
 280
 281     FSHOW((stderr,"/creating thread %lu\n", thread_self()));
 282     function = th->no_tls_value_marker;
 283     th->no_tls_value_marker = NO_TLS_VALUE_MARKER_WIDETAG;
 284     if(arch_os_thread_init(th)==0) {
 285         /* FIXME: handle error */
 286         lose("arch_os_thread_init failed\n");
 287     }
 288
 289     th->os_thread=thread_self();
 290     protect_control_stack_guard_page(1);
 291     /* Since GC can only know about this thread from the all_threads
 292      * list and we're just adding this thread to it there is no danger
 293      * of deadlocking even with SIG_STOP_FOR_GC blocked (which it is
 294      * not). */
 295     lock_ret = pthread_mutex_lock(&all_threads_lock);
 296     gc_assert(lock_ret == 0);
 297     link_thread(th);
 298     lock_ret = pthread_mutex_unlock(&all_threads_lock);
 299     gc_assert(lock_ret == 0);
 300
 301     result = funcall0(function);
 302
 303     /* Block GC */
 304     block_blockable_signals();
 305     th->state=STATE_DEAD;
 306
 307     /* SIG_STOP_FOR_GC is blocked and GC might be waiting for this
 308      * thread, but since we are already dead it won't wait long. */
 309     lock_ret = pthread_mutex_lock(&all_threads_lock);
 310     gc_assert(lock_ret == 0);
 311
 312     gc_alloc_update_page_tables(0, &th->alloc_region);
 313     unlink_thread(th);
 314     pthread_mutex_unlock(&all_threads_lock);
 315     gc_assert(lock_ret == 0);
 316
 317     if(th->tls_cookie>=0) arch_os_thread_cleanup(th);
 318     os_invalidate((os_vm_address_t)th->interrupt_data,
 319                   (sizeof (struct interrupt_data)));
 320
 321 #ifdef LISP_FEATURE_MACH_EXCEPTION_HANDLER
 322     FSHOW((stderr, "Deallocating mach port %x\n", THREAD_STRUCT_TO_EXCEPTION_PORT(th)));
 323     mach_port_move_member(mach_task_self(),
 324                           THREAD_STRUCT_TO_EXCEPTION_PORT(th),
 325                           MACH_PORT_NULL);
 326     mach_port_deallocate(mach_task_self(),
 327                          THREAD_STRUCT_TO_EXCEPTION_PORT(th));
 328     mach_port_destroy(mach_task_self(),
 329                       THREAD_STRUCT_TO_EXCEPTION_PORT(th));
 330 #endif
 331
 332 #ifdef QUEUE_FREEABLE_THREAD_STACKS
 333     queue_freeable_thread_stack(th);
 334 #elif defined(CREATE_CLEANUP_THREAD)
 335     create_cleanup_thread(th);
 336 #else
 337     free_thread_stack_later(th);
 338 #endif
 339
 340     FSHOW((stderr,"/exiting thread %p\n", thread_self()));
 341     return result;
 342 }
 343
 344 #endif /* LISP_FEATURE_SB_THREAD */
 345
 346 static void
 347 free_thread_struct(struct thread *th)
 348 {
 349     if (th->interrupt_data)
 350         os_invalidate((os_vm_address_t) th->interrupt_data,
 351                       (sizeof (struct interrupt_data)));
 352     os_invalidate((os_vm_address_t) th->os_address,
 353                   THREAD_STRUCT_SIZE);
 354 }
 355
 356 /* this is called from any other thread to create the new one, and
 357  * initialize all parts of it that can be initialized from another
 358  * thread
 359  */
 360
 361 static struct thread *
 362 create_thread_struct(lispobj initial_function) {
 363     union per_thread_data *per_thread;
 364     struct thread *th=0;        /*  subdue gcc */
 365     void *spaces=0;
 366     void *aligned_spaces=0;
 367 #ifdef LISP_FEATURE_SB_THREAD
 368     int i;
 369 #endif
 370
 371 #ifdef CREATE_CLEANUP_THREAD
 372     /* Give a chance for cleanup threads to run. */
 373     sched_yield();
 374 #endif
 375     /* May as well allocate all the spaces at once: it saves us from
 376      * having to decide what to do if only some of the allocations
 377      * succeed.  SPACES must be page-aligned, since the GC expects the
 378      * control stack to start at a page boundary.  We can't rely on the
 379      * alignment passed from os_validate, since that might assume the
 380      * current (e.g. 4k) pagesize, while we calculate with the biggest
 381      * (e.g. 64k) pagesize allowed by the ABI.  */
 382     spaces=os_validate(0, THREAD_STRUCT_SIZE);
 383     if(!spaces)
 384         return NULL;
 385     /* Aligning up is safe as THREAD_STRUCT_SIZE has BACKEND_PAGE_SIZE
 386      * padding. */
 387     aligned_spaces = (void *)((((unsigned long)(char *)spaces)
 388                                + BACKEND_PAGE_SIZE - 1)
 389                               & ~(unsigned long)(BACKEND_PAGE_SIZE - 1));
 390     per_thread=(union per_thread_data *)
 391         (aligned_spaces+
 392          THREAD_CONTROL_STACK_SIZE+
 393          BINDING_STACK_SIZE+
 394          ALIEN_STACK_SIZE);
 395
 396 #ifdef LISP_FEATURE_SB_THREAD
 397     for(i = 0; i < (dynamic_values_bytes / sizeof(lispobj)); i++)
 398         per_thread->dynamic_values[i] = NO_TLS_VALUE_MARKER_WIDETAG;
 399     if (all_threads == 0) {
 400         if(SymbolValue(FREE_TLS_INDEX,0)==UNBOUND_MARKER_WIDETAG) {
 401             SetSymbolValue
 402                 (FREE_TLS_INDEX,
 403                  /* FIXME: should be MAX_INTERRUPTS -1 ? */
 404                  make_fixnum(MAX_INTERRUPTS+
 405                              sizeof(struct thread)/sizeof(lispobj)),
 406                  0);
 407             SetSymbolValue(TLS_INDEX_LOCK,make_fixnum(0),0);
 408         }
 409 #define STATIC_TLS_INIT(sym,field) \
 410   ((struct symbol *)(sym-OTHER_POINTER_LOWTAG))->tls_index= \
 411   make_fixnum(THREAD_SLOT_OFFSET_WORDS(field))
 412
 413         STATIC_TLS_INIT(BINDING_STACK_START,binding_stack_start);
 414         STATIC_TLS_INIT(BINDING_STACK_POINTER,binding_stack_pointer);
 415         STATIC_TLS_INIT(CONTROL_STACK_START,control_stack_start);
 416         STATIC_TLS_INIT(CONTROL_STACK_END,control_stack_end);
 417         STATIC_TLS_INIT(ALIEN_STACK,alien_stack_pointer);
 418 #if defined(LISP_FEATURE_X86) || defined (LISP_FEATURE_X86_64)
 419         STATIC_TLS_INIT(PSEUDO_ATOMIC_BITS,pseudo_atomic_bits);
 420 #endif
 421 #undef STATIC_TLS_INIT
 422     }
 423 #endif
 424
 425     th=&per_thread->thread;
 426     th->os_address = spaces;
 427     th->control_stack_start = aligned_spaces;
 428     th->binding_stack_start=
 429         (lispobj*)((void*)th->control_stack_start+THREAD_CONTROL_STACK_SIZE);
 430     th->control_stack_end = th->binding_stack_start;
 431     th->alien_stack_start=
 432         (lispobj*)((void*)th->binding_stack_start+BINDING_STACK_SIZE);
 433     th->binding_stack_pointer=th->binding_stack_start;
 434     th->this=th;
 435     th->os_thread=0;
 436     th->state=STATE_RUNNING;
 437 #ifdef LISP_FEATURE_STACK_GROWS_DOWNWARD_NOT_UPWARD
 438     th->alien_stack_pointer=((void *)th->alien_stack_start
 439                              + ALIEN_STACK_SIZE-N_WORD_BYTES);
 440 #else
 441     th->alien_stack_pointer=((void *)th->alien_stack_start);
 442 #endif
 443 #if defined(LISP_FEATURE_X86) || defined (LISP_FEATURE_X86_64)
 444     th->pseudo_atomic_bits=0;
 445 #endif
 446 #ifdef LISP_FEATURE_GENCGC
 447     gc_set_region_empty(&th->alloc_region);
 448 #endif
 449
 450 #ifndef LISP_FEATURE_SB_THREAD
 451     /* the tls-points-into-struct-thread trick is only good for threaded
 452      * sbcl, because unithread sbcl doesn't have tls.  So, we copy the
 453      * appropriate values from struct thread here, and make sure that
 454      * we use the appropriate SymbolValue macros to access any of the
 455      * variable quantities from the C runtime.  It's not quite OAOOM,
 456      * it just feels like it */
 457     SetSymbolValue(BINDING_STACK_START,(lispobj)th->binding_stack_start,th);
 458     SetSymbolValue(CONTROL_STACK_START,(lispobj)th->control_stack_start,th);
 459     SetSymbolValue(CONTROL_STACK_END,(lispobj)th->control_stack_end,th);
 460 #if defined(LISP_FEATURE_X86) || defined (LISP_FEATURE_X86_64)
 461     SetSymbolValue(BINDING_STACK_POINTER,(lispobj)th->binding_stack_pointer,th);
 462     SetSymbolValue(ALIEN_STACK,(lispobj)th->alien_stack_pointer,th);
 463     SetSymbolValue(PSEUDO_ATOMIC_BITS,(lispobj)th->pseudo_atomic_bits,th);
 464 #else
 465     current_binding_stack_pointer=th->binding_stack_pointer;
 466     current_control_stack_pointer=th->control_stack_start;
 467 #endif
 468 #endif
 469     bind_variable(CURRENT_CATCH_BLOCK,make_fixnum(0),th);
 470     bind_variable(CURRENT_UNWIND_PROTECT_BLOCK,make_fixnum(0),th);
 471     bind_variable(FREE_INTERRUPT_CONTEXT_INDEX,make_fixnum(0),th);
 472     bind_variable(INTERRUPT_PENDING, NIL,th);
 473     bind_variable(INTERRUPTS_ENABLED,T,th);
 474     bind_variable(ALLOW_WITH_INTERRUPTS,T,th);
 475     bind_variable(GC_PENDING,NIL,th);
 476 #ifdef LISP_FEATURE_SB_THREAD
 477     bind_variable(STOP_FOR_GC_PENDING,NIL,th);
 478 #endif
 479
 480     th->interrupt_data = (struct interrupt_data *)
 481         os_validate(0,(sizeof (struct interrupt_data)));
 482     if (!th->interrupt_data) {
 483         free_thread_struct(th);
 484         return 0;
 485     }
 486     th->interrupt_data->pending_handler = 0;
 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     if(th) {
 502 #ifdef LISP_FEATURE_MACH_EXCEPTION_HANDLER
 503         kern_return_t ret;
 504
 505         setup_mach_exception_handling_thread();
 506 #endif
 507         initial_thread_trampoline(th); /* no return */
 508     } else lose("can't create initial thread\n");
 509 }
 510
 511 #ifdef LISP_FEATURE_SB_THREAD
 512
 513 #ifndef __USE_XOPEN2K
 514 extern int pthread_attr_setstack (pthread_attr_t *__attr, void *__stackaddr,
 515                                   size_t __stacksize);
 516 #endif
 517
 518 boolean create_os_thread(struct thread *th,os_thread_t *kid_tid)
 519 {
 520     /* The new thread inherits the restrictive signal mask set here,
 521      * and enables signals again when it is set up properly. */
 522     pthread_attr_t attr;
 523     sigset_t newset,oldset;
 524     boolean r=1;
 525     int retcode, initcode, sizecode, addrcode;
 526
 527     FSHOW_SIGNAL((stderr,"/create_os_thread: creating new thread\n"));
 528
 529 #ifdef LOCK_CREATE_THREAD
 530     retcode = pthread_mutex_lock(&create_thread_lock);
 531     gc_assert(retcode == 0);
 532     FSHOW_SIGNAL((stderr,"/create_os_thread: got lock\n"));
 533 #endif
 534     sigemptyset(&newset);
 535     /* Blocking deferrable signals is enough, no need to block
 536      * SIG_STOP_FOR_GC because the child process is not linked onto
 537      * all_threads until it's ready. */
 538     sigaddset_deferrable(&newset);
 539     thread_sigmask(SIG_BLOCK, &newset, &oldset);
 540
 541 #if defined(LISP_FEATURE_DARWIN)
 542 #define CONTROL_STACK_ADJUST 8192 /* darwin wants page-aligned stacks */
 543 #else
 544 #define CONTROL_STACK_ADJUST 16
 545 #endif
 546
 547     if((initcode = pthread_attr_init(&attr)) ||
 548        /* FIXME: why do we even have this in the first place? */
 549        (pthread_attr_setstack(&attr,th->control_stack_start,
 550                               THREAD_CONTROL_STACK_SIZE-CONTROL_STACK_ADJUST)) ||
 551 #undef CONTROL_STACK_ADJUST
 552        (retcode = pthread_create
 553         (kid_tid,&attr,(void *(*)(void *))new_thread_trampoline,th))) {
 554         FSHOW_SIGNAL((stderr, "init, size, addr = %d, %d, %d\n", initcode, sizecode, addrcode));
 555         FSHOW_SIGNAL((stderr, printf("pthread_create returned %d, errno %d\n", retcode, errno)));
 556         FSHOW_SIGNAL((stderr, "wanted stack size %d, min stack size %d\n",
 557                       THREAD_CONTROL_STACK_SIZE-16, PTHREAD_STACK_MIN));
 558         if(retcode < 0) {
 559             perror("create_os_thread");
 560         }
 561         r=0;
 562     }
 563
 564 #ifdef QUEUE_FREEABLE_THREAD_STACKS
 565     free_freeable_stacks();
 566 #endif
 567     thread_sigmask(SIG_SETMASK,&oldset,0);
 568 #ifdef LOCK_CREATE_THREAD
 569     retcode = pthread_mutex_unlock(&create_thread_lock);
 570     gc_assert(retcode == 0);
 571     FSHOW_SIGNAL((stderr,"/create_os_thread: released lock\n"));
 572 #endif
 573     return r;
 574 }
 575
 576 os_thread_t create_thread(lispobj initial_function) {
 577     struct thread *th;
 578     os_thread_t kid_tid;
 579
 580     /* Assuming that a fresh thread struct has no lisp objects in it,
 581      * linking it to all_threads can be left to the thread itself
 582      * without fear of gc lossage. initial_function violates this
 583      * assumption and must stay pinned until the child starts up. */
 584     th = create_thread_struct(initial_function);
 585     if(th==0) return 0;
 586
 587     if (create_os_thread(th,&kid_tid)) {
 588         return kid_tid;
 589     } else {
 590         free_thread_struct(th);
 591         return 0;
 592     }
 593 }
 594
 595 /* Send the signo to os_thread, retry if the rt signal queue is
 596  * full. */
 597 int
 598 kill_thread_safely(os_thread_t os_thread, int signo)
 599 {
 600     int r;
 601     /* The man page does not mention EAGAIN as a valid return value
 602      * for either pthread_kill or kill. But that's theory, this is
 603      * practice. By waiting here we assume that the delivery of this
 604      * signal is not necessary for the delivery of the signals in the
 605      * queue. In other words, we _assume_ there are no deadlocks. */
 606     while ((r=pthread_kill(os_thread,signo))==EAGAIN) {
 607         /* wait a bit then try again in the hope of the rt signal
 608          * queue not being full */
 609         FSHOW_SIGNAL((stderr,"/rt signal queue full\n"));
 610         /* FIXME: some kind of backoff (random, exponential) would be
 611          * nice. */
 612         sleep(1);
 613     }
 614     return r;
 615 }
 616
 617 int signal_interrupt_thread(os_thread_t os_thread)
 618 {
 619     int status = kill_thread_safely(os_thread, SIG_INTERRUPT_THREAD);
 620     if (status == 0) {
 621         return 0;
 622     } else if (status == ESRCH) {
 623         return -1;
 624     } else {
 625         lose("cannot send SIG_INTERRUPT_THREAD to thread=%lu: %d, %s\n",
 626              os_thread, status, strerror(status));
 627     }
 628 }
 629
 630 /* stopping the world is a two-stage process.  From this thread we signal
 631  * all the others with SIG_STOP_FOR_GC.  The handler for this signal does
 632  * the usual pseudo-atomic checks (we don't want to stop a thread while
 633  * it's in the middle of allocation) then waits for another SIG_STOP_FOR_GC.
 634  */
 635
 636 /* To avoid deadlocks when gc stops the world all clients of each
 637  * mutex must enable or disable SIG_STOP_FOR_GC for the duration of
 638  * holding the lock, but they must agree on which. */
 639 void gc_stop_the_world()
 640 {
 641     struct thread *p,*th=arch_os_get_current_thread();
 642     int status, lock_ret;
 643 #ifdef LOCK_CREATE_THREAD
 644     /* KLUDGE: Stopping the thread during pthread_create() causes deadlock
 645      * on FreeBSD. */
 646     FSHOW_SIGNAL((stderr,"/gc_stop_the_world:waiting on create_thread_lock, thread=%lu\n",
 647                   th->os_thread));
 648     lock_ret = pthread_mutex_lock(&create_thread_lock);
 649     gc_assert(lock_ret == 0);
 650     FSHOW_SIGNAL((stderr,"/gc_stop_the_world:got create_thread_lock, thread=%lu\n",
 651                   th->os_thread));
 652 #endif
 653     FSHOW_SIGNAL((stderr,"/gc_stop_the_world:waiting on lock, thread=%lu\n",
 654                   th->os_thread));
 655     /* keep threads from starting while the world is stopped. */
 656     lock_ret = pthread_mutex_lock(&all_threads_lock);      \
 657     gc_assert(lock_ret == 0);
 658
 659     FSHOW_SIGNAL((stderr,"/gc_stop_the_world:got lock, thread=%lu\n",
 660                   th->os_thread));
 661     /* stop all other threads by sending them SIG_STOP_FOR_GC */
 662     for(p=all_threads; p; p=p->next) {
 663         gc_assert(p->os_thread != 0);
 664         FSHOW_SIGNAL((stderr,"/gc_stop_the_world: p->state: %x\n", p->state));
 665         if((p!=th) && ((p->state==STATE_RUNNING))) {
 666             FSHOW_SIGNAL((stderr,"/gc_stop_the_world: suspending %x, os_thread %x\n",
 667                           p, p->os_thread));
 668             status=kill_thread_safely(p->os_thread,SIG_STOP_FOR_GC);
 669             if (status==ESRCH) {
 670                 /* This thread has exited. */
 671                 gc_assert(p->state==STATE_DEAD);
 672             } else if (status) {
 673                 lose("cannot send suspend thread=%lu: %d, %s\n",
 674                      p->os_thread,status,strerror(status));
 675             }
 676         }
 677     }
 678     FSHOW_SIGNAL((stderr,"/gc_stop_the_world:signals sent\n"));
 679     /* wait for the running threads to stop or finish */
 680     for(p=all_threads;p;) {
 681         FSHOW_SIGNAL((stderr,"/gc_stop_the_world: th: %p, p: %p\n", th, p));
 682         if((p!=th) && (p->state==STATE_RUNNING)) {
 683             sched_yield();
 684         } else {
 685             p=p->next;
 686         }
 687     }
 688     FSHOW_SIGNAL((stderr,"/gc_stop_the_world:end\n"));
 689 }
 690
 691 void gc_start_the_world()
 692 {
 693     struct thread *p,*th=arch_os_get_current_thread();
 694     int status, lock_ret;
 695     /* if a resumed thread creates a new thread before we're done with
 696      * this loop, the new thread will get consed on the front of
 697      * all_threads, but it won't have been stopped so won't need
 698      * restarting */
 699     FSHOW_SIGNAL((stderr,"/gc_start_the_world:begin\n"));
 700     for(p=all_threads;p;p=p->next) {
 701         gc_assert(p->os_thread!=0);
 702         if((p!=th) && (p->state!=STATE_DEAD)) {
 703             if(p->state!=STATE_SUSPENDED) {
 704                 lose("gc_start_the_world: wrong thread state is %d\n",
 705                      fixnum_value(p->state));
 706             }
 707             FSHOW_SIGNAL((stderr, "/gc_start_the_world: resuming %lu\n",
 708                           p->os_thread));
 709             p->state=STATE_RUNNING;
 710
 711 #if defined(SIG_RESUME_FROM_GC)
 712             status=kill_thread_safely(p->os_thread,SIG_RESUME_FROM_GC);
 713 #else
 714             status=kill_thread_safely(p->os_thread,SIG_STOP_FOR_GC);
 715 #endif
 716             if (status) {
 717                 lose("cannot resume thread=%lu: %d, %s\n",
 718                      p->os_thread,status,strerror(status));
 719             }
 720         }
 721     }
 722     /* If we waited here until all threads leave STATE_SUSPENDED, then
 723      * SIG_STOP_FOR_GC wouldn't need to be a rt signal. That has some
 724      * performance implications, but does away with the 'rt signal
 725      * queue full' problem. */
 726
 727     lock_ret = pthread_mutex_unlock(&all_threads_lock);
 728     gc_assert(lock_ret == 0);
 729 #ifdef LOCK_CREATE_THREAD
 730     lock_ret = pthread_mutex_unlock(&create_thread_lock);
 731     gc_assert(lock_ret == 0);
 732 #endif
 733
 734     FSHOW_SIGNAL((stderr,"/gc_start_the_world:end\n"));
 735 }
 736 #endif