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 #include "runtime.h"
  29 #include "validate.h"           /* for CONTROL_STACK_SIZE etc */
  30 #include "alloc.h"
  31 #include "thread.h"
  32 #include "arch.h"
  33 #include "target-arch-os.h"
  34 #include "os.h"
  35 #include "globals.h"
  36 #include "dynbind.h"
  37 #include "genesis/cons.h"
  38 #include "genesis/fdefn.h"
  39 #include "interr.h"             /* for lose() */
  40 #include "gc-internal.h"
  41
  42 #ifdef LISP_FEATURE_WIN32
  43 /*
  44  * Win32 doesn't have SIGSTKSZ, and we're not switching stacks anyway,
  45  * so define it arbitrarily
  46  */
  47 #define SIGSTKSZ 1024
  48 #endif
  49
  50 #define ALIEN_STACK_SIZE (1*1024*1024) /* 1Mb size chosen at random */
  51
  52 struct freeable_stack {
  53     os_thread_t os_thread;
  54     os_vm_address_t stack;
  55 };
  56
  57 static struct freeable_stack * volatile freeable_stack = 0;
  58
  59 int dynamic_values_bytes=4096*sizeof(lispobj);  /* same for all threads */
  60 struct thread * volatile all_threads;
  61 extern struct interrupt_data * global_interrupt_data;
  62 extern int linux_no_threads_p;
  63
  64 #ifdef LISP_FEATURE_SB_THREAD
  65 pthread_mutex_t all_threads_lock = PTHREAD_MUTEX_INITIALIZER;
  66 #endif
  67
  68 #if defined(LISP_FEATURE_X86) || defined(LISP_FEATURE_X86_64)
  69 extern lispobj call_into_lisp_first_time(lispobj fun, lispobj *args, int nargs);
  70 #endif
  71
  72 static void
  73 link_thread(struct thread *th)
  74 {
  75     if (all_threads) all_threads->prev=th;
  76     th->next=all_threads;
  77     th->prev=0;
  78     all_threads=th;
  79 }
  80
  81 #ifdef LISP_FEATURE_SB_THREAD
  82 static void
  83 unlink_thread(struct thread *th)
  84 {
  85     if (th->prev)
  86         th->prev->next = th->next;
  87     else
  88         all_threads = th->next;
  89     if (th->next)
  90         th->next->prev = th->prev;
  91 }
  92 #endif
  93
  94 static int
  95 initial_thread_trampoline(struct thread *th)
  96 {
  97     lispobj function;
  98 #if defined(LISP_FEATURE_X86) || defined(LISP_FEATURE_X86_64)
  99     lispobj *args = NULL;
 100 #endif
 101     function = th->no_tls_value_marker;
 102     th->no_tls_value_marker = NO_TLS_VALUE_MARKER_WIDETAG;
 103     if(arch_os_thread_init(th)==0) return 1;
 104     link_thread(th);
 105     th->os_thread=thread_self();
 106 #ifndef LISP_FEATURE_WIN32
 107     protect_control_stack_guard_page(1);
 108 #endif
 109
 110 #if defined(LISP_FEATURE_X86) || defined(LISP_FEATURE_X86_64)
 111     return call_into_lisp_first_time(function,args,0);
 112 #else
 113     return funcall0(function);
 114 #endif
 115 }
 116
 117 #define THREAD_STRUCT_SIZE (THREAD_CONTROL_STACK_SIZE + BINDING_STACK_SIZE + \
 118                             ALIEN_STACK_SIZE + dynamic_values_bytes + \
 119                             32 * SIGSTKSZ)
 120
 121 #ifdef LISP_FEATURE_SB_THREAD
 122
 123 static void
 124 free_thread_stack_later(struct thread *thread_to_be_cleaned_up)
 125 {
 126     struct freeable_stack *new_freeable_stack = 0;
 127     if (thread_to_be_cleaned_up) {
 128         new_freeable_stack = (struct freeable_stack *)
 129             os_validate(0, sizeof(struct freeable_stack));
 130         new_freeable_stack->os_thread = thread_to_be_cleaned_up->os_thread;
 131         new_freeable_stack->stack = (os_vm_address_t)
 132             thread_to_be_cleaned_up->control_stack_start;
 133     }
 134     new_freeable_stack = (struct freeable_stack *)
 135         swap_lispobjs((lispobj *)(void *)&freeable_stack,
 136                       (lispobj)new_freeable_stack);
 137     if (new_freeable_stack) {
 138         FSHOW((stderr,"/reaping %lu\n", new_freeable_stack->os_thread));
 139         /* Under NPTL pthread_join really waits until the thread
 140          * exists and the stack can be safely freed. This is sadly not
 141          * mandated by the pthread spec. */
 142         gc_assert(pthread_join(new_freeable_stack->os_thread, NULL) == 0);
 143         os_invalidate(new_freeable_stack->stack, THREAD_STRUCT_SIZE);
 144         os_invalidate((os_vm_address_t) new_freeable_stack,
 145                       sizeof(struct freeable_stack));
 146     }
 147 }
 148
 149 /* this is the first thing that runs in the child (which is why the
 150  * silly calling convention).  Basically it calls the user's requested
 151  * lisp function after doing arch_os_thread_init and whatever other
 152  * bookkeeping needs to be done
 153  */
 154 int
 155 new_thread_trampoline(struct thread *th)
 156 {
 157     lispobj function;
 158     int result;
 159     FSHOW((stderr,"/creating thread %lu\n", thread_self()));
 160     function = th->no_tls_value_marker;
 161     th->no_tls_value_marker = NO_TLS_VALUE_MARKER_WIDETAG;
 162     if(arch_os_thread_init(th)==0) {
 163         /* FIXME: handle error */
 164         lose("arch_os_thread_init failed\n");
 165     }
 166
 167     th->os_thread=thread_self();
 168     protect_control_stack_guard_page(1);
 169     /* Since GC can only know about this thread from the all_threads
 170      * list and we're just adding this thread to it there is no danger
 171      * of deadlocking even with SIG_STOP_FOR_GC blocked (which it is
 172      * not). */
 173     pthread_mutex_lock(&all_threads_lock);
 174     link_thread(th);
 175     pthread_mutex_unlock(&all_threads_lock);
 176
 177     result = funcall0(function);
 178     th->state=STATE_DEAD;
 179
 180     /* SIG_STOP_FOR_GC is blocked and GC might be waiting for this
 181      * thread, but since we are already dead it won't wait long. */
 182     pthread_mutex_lock(&all_threads_lock);
 183     gc_alloc_update_page_tables(0, &th->alloc_region);
 184     unlink_thread(th);
 185     pthread_mutex_unlock(&all_threads_lock);
 186
 187     if(th->tls_cookie>=0) arch_os_thread_cleanup(th);
 188     os_invalidate((os_vm_address_t)th->interrupt_data,
 189                   (sizeof (struct interrupt_data)));
 190     free_thread_stack_later(th);
 191     FSHOW((stderr,"/exiting thread %lu\n", thread_self()));
 192     return result;
 193 }
 194
 195 #endif /* LISP_FEATURE_SB_THREAD */
 196
 197 static void
 198 free_thread_struct(struct thread *th)
 199 {
 200     if (th->interrupt_data)
 201         os_invalidate((os_vm_address_t) th->interrupt_data,
 202                       (sizeof (struct interrupt_data)));
 203     os_invalidate((os_vm_address_t) th->control_stack_start,
 204                   THREAD_STRUCT_SIZE);
 205 }
 206
 207 /* this is called from any other thread to create the new one, and
 208  * initialize all parts of it that can be initialized from another
 209  * thread
 210  */
 211
 212 static struct thread *
 213 create_thread_struct(lispobj initial_function) {
 214     union per_thread_data *per_thread;
 215     struct thread *th=0;        /*  subdue gcc */
 216     void *spaces=0;
 217 #ifdef LISP_FEATURE_SB_THREAD
 218     int i;
 219 #endif
 220
 221     /* may as well allocate all the spaces at once: it saves us from
 222      * having to decide what to do if only some of the allocations
 223      * succeed */
 224     spaces=os_validate(0, THREAD_STRUCT_SIZE);
 225     if(!spaces)
 226          return NULL;
 227     per_thread=(union per_thread_data *)
 228         (spaces+
 229          THREAD_CONTROL_STACK_SIZE+
 230          BINDING_STACK_SIZE+
 231          ALIEN_STACK_SIZE);
 232
 233 #ifdef LISP_FEATURE_SB_THREAD
 234     for(i = 0; i < (dynamic_values_bytes / sizeof(lispobj)); i++)
 235         per_thread->dynamic_values[i] = NO_TLS_VALUE_MARKER_WIDETAG;
 236     if (all_threads == 0) {
 237         if(SymbolValue(FREE_TLS_INDEX,0)==UNBOUND_MARKER_WIDETAG) {
 238             SetSymbolValue
 239                 (FREE_TLS_INDEX,
 240                  /* FIXME: should be MAX_INTERRUPTS -1 ? */
 241                  make_fixnum(MAX_INTERRUPTS+
 242                              sizeof(struct thread)/sizeof(lispobj)),
 243                  0);
 244             SetSymbolValue(TLS_INDEX_LOCK,make_fixnum(0),0);
 245         }
 246 #define STATIC_TLS_INIT(sym,field) \
 247   ((struct symbol *)(sym-OTHER_POINTER_LOWTAG))->tls_index= \
 248   make_fixnum(THREAD_SLOT_OFFSET_WORDS(field))
 249
 250         STATIC_TLS_INIT(BINDING_STACK_START,binding_stack_start);
 251         STATIC_TLS_INIT(BINDING_STACK_POINTER,binding_stack_pointer);
 252         STATIC_TLS_INIT(CONTROL_STACK_START,control_stack_start);
 253         STATIC_TLS_INIT(CONTROL_STACK_END,control_stack_end);
 254         STATIC_TLS_INIT(ALIEN_STACK,alien_stack_pointer);
 255 #if defined(LISP_FEATURE_X86) || defined (LISP_FEATURE_X86_64)
 256         STATIC_TLS_INIT(PSEUDO_ATOMIC_ATOMIC,pseudo_atomic_atomic);
 257         STATIC_TLS_INIT(PSEUDO_ATOMIC_INTERRUPTED,pseudo_atomic_interrupted);
 258 #endif
 259 #undef STATIC_TLS_INIT
 260     }
 261 #endif
 262
 263     th=&per_thread->thread;
 264     th->control_stack_start = spaces;
 265     th->binding_stack_start=
 266         (lispobj*)((void*)th->control_stack_start+THREAD_CONTROL_STACK_SIZE);
 267     th->control_stack_end = th->binding_stack_start;
 268     th->alien_stack_start=
 269         (lispobj*)((void*)th->binding_stack_start+BINDING_STACK_SIZE);
 270     th->binding_stack_pointer=th->binding_stack_start;
 271     th->this=th;
 272     th->os_thread=0;
 273     th->state=STATE_RUNNING;
 274 #ifdef LISP_FEATURE_STACK_GROWS_DOWNWARD_NOT_UPWARD
 275     th->alien_stack_pointer=((void *)th->alien_stack_start
 276                              + ALIEN_STACK_SIZE-N_WORD_BYTES);
 277 #else
 278     th->alien_stack_pointer=((void *)th->alien_stack_start);
 279 #endif
 280 #if defined(LISP_FEATURE_X86) || defined (LISP_FEATURE_X86_64)
 281     th->pseudo_atomic_interrupted=0;
 282     th->pseudo_atomic_atomic=0;
 283 #endif
 284 #ifdef LISP_FEATURE_GENCGC
 285     gc_set_region_empty(&th->alloc_region);
 286 #endif
 287
 288 #ifndef LISP_FEATURE_SB_THREAD
 289     /* the tls-points-into-struct-thread trick is only good for threaded
 290      * sbcl, because unithread sbcl doesn't have tls.  So, we copy the
 291      * appropriate values from struct thread here, and make sure that
 292      * we use the appropriate SymbolValue macros to access any of the
 293      * variable quantities from the C runtime.  It's not quite OAOOM,
 294      * it just feels like it */
 295     SetSymbolValue(BINDING_STACK_START,(lispobj)th->binding_stack_start,th);
 296     SetSymbolValue(CONTROL_STACK_START,(lispobj)th->control_stack_start,th);
 297     SetSymbolValue(CONTROL_STACK_END,(lispobj)th->control_stack_end,th);
 298 #if defined(LISP_FEATURE_X86) || defined (LISP_FEATURE_X86_64)
 299     SetSymbolValue(BINDING_STACK_POINTER,(lispobj)th->binding_stack_pointer,th);
 300     SetSymbolValue(ALIEN_STACK,(lispobj)th->alien_stack_pointer,th);
 301     SetSymbolValue(PSEUDO_ATOMIC_ATOMIC,(lispobj)th->pseudo_atomic_atomic,th);
 302     SetSymbolValue(PSEUDO_ATOMIC_INTERRUPTED,th->pseudo_atomic_interrupted,th);
 303 #else
 304     current_binding_stack_pointer=th->binding_stack_pointer;
 305     current_control_stack_pointer=th->control_stack_start;
 306 #endif
 307 #endif
 308     bind_variable(CURRENT_CATCH_BLOCK,make_fixnum(0),th);
 309     bind_variable(CURRENT_UNWIND_PROTECT_BLOCK,make_fixnum(0),th);
 310     bind_variable(FREE_INTERRUPT_CONTEXT_INDEX,make_fixnum(0),th);
 311     bind_variable(INTERRUPT_PENDING, NIL,th);
 312     bind_variable(INTERRUPTS_ENABLED,T,th);
 313     bind_variable(GC_PENDING,NIL,th);
 314 #ifdef LISP_FEATURE_SB_THREAD
 315     bind_variable(STOP_FOR_GC_PENDING,NIL,th);
 316 #endif
 317
 318     th->interrupt_data = (struct interrupt_data *)
 319         os_validate(0,(sizeof (struct interrupt_data)));
 320     if (!th->interrupt_data) {
 321         free_thread_struct(th);
 322         return 0;
 323     }
 324     th->interrupt_data->pending_handler = 0;
 325     th->no_tls_value_marker=initial_function;
 326     return th;
 327 }
 328
 329 void create_initial_thread(lispobj initial_function) {
 330     struct thread *th=create_thread_struct(initial_function);
 331     if(th) {
 332         initial_thread_trampoline(th); /* no return */
 333     } else lose("can't create initial thread\n");
 334 }
 335
 336 #ifdef LISP_FEATURE_SB_THREAD
 337
 338 #ifndef __USE_XOPEN2K
 339 extern int pthread_attr_setstack (pthread_attr_t *__attr, void *__stackaddr,
 340                                   size_t __stacksize);
 341 #endif
 342
 343 boolean create_os_thread(struct thread *th,os_thread_t *kid_tid)
 344 {
 345     /* The new thread inherits the restrictive signal mask set here,
 346      * and enables signals again when it is set up properly. */
 347     pthread_attr_t attr;
 348     sigset_t newset,oldset;
 349     boolean r=1;
 350     sigemptyset(&newset);
 351     /* Blocking deferrable signals is enough, no need to block
 352      * SIG_STOP_FOR_GC because the child process is not linked onto
 353      * all_threads until it's ready. */
 354     sigaddset_deferrable(&newset);
 355     thread_sigmask(SIG_BLOCK, &newset, &oldset);
 356
 357     if((pthread_attr_init(&attr)) ||
 358        (pthread_attr_setstack(&attr,th->control_stack_start,
 359                               THREAD_CONTROL_STACK_SIZE-16)) ||
 360        (pthread_create
 361         (kid_tid,&attr,(void *(*)(void *))new_thread_trampoline,th)))
 362         r=0;
 363     thread_sigmask(SIG_SETMASK,&oldset,0);
 364     return r;
 365 }
 366
 367 os_thread_t create_thread(lispobj initial_function) {
 368     struct thread *th;
 369     os_thread_t kid_tid;
 370
 371     if(linux_no_threads_p) return 0;
 372
 373     /* Assuming that a fresh thread struct has no lisp objects in it,
 374      * linking it to all_threads can be left to the thread itself
 375      * without fear of gc lossage. initial_function violates this
 376      * assumption and must stay pinned until the child starts up. */
 377     th = create_thread_struct(initial_function);
 378     if(th==0) return 0;
 379
 380     if (create_os_thread(th,&kid_tid)) {
 381         return kid_tid;
 382     } else {
 383         free_thread_struct(th);
 384         return 0;
 385     }
 386 }
 387
 388 /* Send the signo to os_thread, retry if the rt signal queue is
 389  * full. */
 390 static int kill_thread_safely(os_thread_t os_thread, int signo)
 391 {
 392     int r;
 393     /* The man page does not mention EAGAIN as a valid return value
 394      * for either pthread_kill or kill. But that's theory, this is
 395      * practice. By waiting here we assume that the delivery of this
 396      * signal is not necessary for the delivery of the signals in the
 397      * queue. In other words, we _assume_ there are no deadlocks. */
 398     while ((r=pthread_kill(os_thread,signo))==EAGAIN) {
 399         /* wait a bit then try again in the hope of the rt signal
 400          * queue not being full */
 401         FSHOW_SIGNAL((stderr,"/rt signal queue full\n"));
 402         /* FIXME: some kind of backoff (random, exponential) would be
 403          * nice. */
 404         sleep(1);
 405     }
 406     return r;
 407 }
 408
 409 int signal_interrupt_thread(os_thread_t os_thread)
 410 {
 411     int status = kill_thread_safely(os_thread, SIG_INTERRUPT_THREAD);
 412     if (status == 0) {
 413         return 0;
 414     } else if (status == ESRCH) {
 415         return -1;
 416     } else {
 417         lose("cannot send SIG_INTERRUPT_THREAD to thread=%lu: %d, %s\n",
 418              os_thread, status, strerror(status));
 419     }
 420 }
 421
 422 /* stopping the world is a two-stage process.  From this thread we signal
 423  * all the others with SIG_STOP_FOR_GC.  The handler for this signal does
 424  * the usual pseudo-atomic checks (we don't want to stop a thread while
 425  * it's in the middle of allocation) then waits for another SIG_STOP_FOR_GC.
 426  */
 427
 428 /* To avoid deadlocks when gc stops the world all clients of each
 429  * mutex must enable or disable SIG_STOP_FOR_GC for the duration of
 430  * holding the lock, but they must agree on which. */
 431 void gc_stop_the_world()
 432 {
 433     struct thread *p,*th=arch_os_get_current_thread();
 434     int status;
 435     FSHOW_SIGNAL((stderr,"/gc_stop_the_world:waiting on lock, thread=%lu\n",
 436                   th->os_thread));
 437     /* keep threads from starting while the world is stopped. */
 438     pthread_mutex_lock(&all_threads_lock); \
 439     FSHOW_SIGNAL((stderr,"/gc_stop_the_world:got lock, thread=%lu\n",
 440                   th->os_thread));
 441     /* stop all other threads by sending them SIG_STOP_FOR_GC */
 442     for(p=all_threads; p; p=p->next) {
 443         gc_assert(p->os_thread != 0);
 444         if((p!=th) && ((p->state==STATE_RUNNING))) {
 445             FSHOW_SIGNAL((stderr,"/gc_stop_the_world: suspending %lu\n",
 446                           p->os_thread));
 447             status=kill_thread_safely(p->os_thread,SIG_STOP_FOR_GC);
 448             if (status==ESRCH) {
 449                 /* This thread has exited. */
 450                 gc_assert(p->state==STATE_DEAD);
 451             } else if (status) {
 452                 lose("cannot send suspend thread=%lu: %d, %s\n",
 453                      p->os_thread,status,strerror(status));
 454             }
 455         }
 456     }
 457     FSHOW_SIGNAL((stderr,"/gc_stop_the_world:signals sent\n"));
 458     /* wait for the running threads to stop or finish */
 459     for(p=all_threads;p;) {
 460         if((p!=th) && (p->state==STATE_RUNNING)) {
 461             sched_yield();
 462         } else {
 463             p=p->next;
 464         }
 465     }
 466     FSHOW_SIGNAL((stderr,"/gc_stop_the_world:end\n"));
 467 }
 468
 469 void gc_start_the_world()
 470 {
 471     struct thread *p,*th=arch_os_get_current_thread();
 472     int status;
 473     /* if a resumed thread creates a new thread before we're done with
 474      * this loop, the new thread will get consed on the front of
 475      * all_threads, but it won't have been stopped so won't need
 476      * restarting */
 477     FSHOW_SIGNAL((stderr,"/gc_start_the_world:begin\n"));
 478     for(p=all_threads;p;p=p->next) {
 479         gc_assert(p->os_thread!=0);
 480         if((p!=th) && (p->state!=STATE_DEAD)) {
 481             if(p->state!=STATE_SUSPENDED) {
 482                 lose("gc_start_the_world: wrong thread state is %d\n",
 483                      fixnum_value(p->state));
 484             }
 485             FSHOW_SIGNAL((stderr, "/gc_start_the_world: resuming %lu\n",
 486                           p->os_thread));
 487             p->state=STATE_RUNNING;
 488             status=kill_thread_safely(p->os_thread,SIG_STOP_FOR_GC);
 489             if (status) {
 490                 lose("cannot resume thread=%lu: %d, %s\n",
 491                      p->os_thread,status,strerror(status));
 492             }
 493         }
 494     }
 495     /* If we waited here until all threads leave STATE_SUSPENDED, then
 496      * SIG_STOP_FOR_GC wouldn't need to be a rt signal. That has some
 497      * performance implications, but does away with the 'rt signal
 498      * queue full' problem. */
 499     pthread_mutex_unlock(&all_threads_lock); \
 500     FSHOW_SIGNAL((stderr,"/gc_start_the_world:end\n"));
 501 }
 502 #endif