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