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