src/runtime/thread.c

   1 #include <stdlib.h>
   2 #include <stdio.h>
   3 #include <string.h>
   4 #include <sched.h>
   5 #include <signal.h>
   6 #include <stddef.h>
   7 #include <errno.h>
   8 #include <sys/types.h>
   9 #include <sys/wait.h>
  10
  11 #include "sbcl.h"
  12 #include "runtime.h"
  13 #include "validate.h"           /* for CONTROL_STACK_SIZE etc */
  14 #include "alloc.h"
  15 #include "thread.h"
  16 #include "arch.h"
  17 #include "target-arch-os.h"
  18 #include "os.h"
  19 #include "globals.h"
  20 #include "dynbind.h"
  21 #include "genesis/cons.h"
  22 #include "genesis/fdefn.h"
  23 #include "interr.h"             /* for lose() */
  24 #include "gc-internal.h"
  25
  26 #define ALIEN_STACK_SIZE (1*1024*1024) /* 1Mb size chosen at random */
  27
  28 struct freeable_stack {
  29     os_thread_t os_thread;
  30     os_vm_address_t stack;
  31 };
  32
  33 static struct freeable_stack * volatile freeable_stack = 0;
  34
  35 int dynamic_values_bytes=4096*sizeof(lispobj);  /* same for all threads */
  36 struct thread * volatile all_threads;
  37 extern struct interrupt_data * global_interrupt_data;
  38 extern int linux_no_threads_p;
  39
  40 #ifdef LISP_FEATURE_SB_THREAD
  41 pthread_mutex_t all_threads_lock = PTHREAD_MUTEX_INITIALIZER;
  42 #endif
  43
  44 #if defined(LISP_FEATURE_X86) || defined(LISP_FEATURE_X86_64)
  45 extern lispobj call_into_lisp_first_time(lispobj fun, lispobj *args, int nargs);
  46 #endif
  47
  48 static void
  49 link_thread(struct thread *th)
  50 {
  51     if (all_threads) all_threads->prev=th;
  52     th->next=all_threads;
  53     th->prev=0;
  54     all_threads=th;
  55 }
  56
  57 #ifdef LISP_FEATURE_SB_THREAD
  58 static void
  59 unlink_thread(struct thread *th)
  60 {
  61     if (th->prev)
  62         th->prev->next = th->next;
  63     else
  64         all_threads = th->next;
  65     if (th->next)
  66         th->next->prev = th->prev;
  67 }
  68 #endif
  69
  70 static int
  71 initial_thread_trampoline(struct thread *th)
  72 {
  73     lispobj function;
  74 #if defined(LISP_FEATURE_X86) || defined(LISP_FEATURE_X86_64)
  75     lispobj *args = NULL;
  76 #endif
  77     function = th->no_tls_value_marker;
  78     th->no_tls_value_marker = NO_TLS_VALUE_MARKER_WIDETAG;
  79     if(arch_os_thread_init(th)==0) return 1;
  80     link_thread(th);
  81     th->os_thread=thread_self();
  82     protect_control_stack_guard_page(1);
  83
  84 #if defined(LISP_FEATURE_X86) || defined(LISP_FEATURE_X86_64)
  85     return call_into_lisp_first_time(function,args,0);
  86 #else
  87     return funcall0(function);
  88 #endif
  89 }
  90
  91 #define THREAD_STRUCT_SIZE (THREAD_CONTROL_STACK_SIZE + BINDING_STACK_SIZE + \
  92                             ALIEN_STACK_SIZE + dynamic_values_bytes + \
  93                             32 * SIGSTKSZ)
  94
  95 #ifdef LISP_FEATURE_SB_THREAD
  96
  97 static void
  98 free_thread_stack_later(struct thread *thread_to_be_cleaned_up)
  99 {
 100     struct freeable_stack *new_freeable_stack = 0;
 101     if (thread_to_be_cleaned_up) {
 102         new_freeable_stack = (struct freeable_stack *)
 103             os_validate(0, sizeof(struct freeable_stack));
 104         new_freeable_stack->os_thread = thread_to_be_cleaned_up->os_thread;
 105         new_freeable_stack->stack = (os_vm_address_t)
 106             thread_to_be_cleaned_up->control_stack_start;
 107     }
 108     new_freeable_stack = (struct freeable_stack *)
 109         swap_lispobjs((lispobj *)(void *)&freeable_stack,
 110                       (lispobj)new_freeable_stack);
 111     if (new_freeable_stack) {
 112         FSHOW((stderr,"/reaping %lu\n", new_freeable_stack->os_thread));
 113         /* Under NPTL pthread_join really waits until the thread
 114          * exists and the stack can be safely freed. This is sadly not
 115          * mandated by the pthread spec. */
 116         gc_assert(pthread_join(new_freeable_stack->os_thread, NULL) == 0);
 117         os_invalidate(new_freeable_stack->stack, THREAD_STRUCT_SIZE);
 118         os_invalidate((os_vm_address_t) new_freeable_stack,
 119                       sizeof(struct freeable_stack));
 120     }
 121 }
 122
 123 /* this is the first thing that runs in the child (which is why the
 124  * silly calling convention).  Basically it calls the user's requested
 125  * lisp function after doing arch_os_thread_init and whatever other
 126  * bookkeeping needs to be done
 127  */
 128 int
 129 new_thread_trampoline(struct thread *th)
 130 {
 131     lispobj function;
 132     int result;
 133     FSHOW((stderr,"/creating thread %lu\n", thread_self()));
 134     function = th->no_tls_value_marker;
 135     th->no_tls_value_marker = NO_TLS_VALUE_MARKER_WIDETAG;
 136     if(arch_os_thread_init(th)==0) {
 137         /* FIXME: handle error */
 138         lose("arch_os_thread_init failed\n");
 139     }
 140
 141     th->os_thread=thread_self();
 142     protect_control_stack_guard_page(1);
 143     /* Since GC can only know about this thread from the all_threads
 144      * list and we're just adding this thread to it there is no danger
 145      * of deadlocking even with SIG_STOP_FOR_GC blocked (which it is
 146      * not). */
 147     pthread_mutex_lock(&all_threads_lock);
 148     link_thread(th);
 149     pthread_mutex_unlock(&all_threads_lock);
 150
 151     result = funcall0(function);
 152     th->state=STATE_DEAD;
 153
 154     /* SIG_STOP_FOR_GC is blocked and GC might be waiting for this
 155      * thread, but since we are already dead it won't wait long. */
 156     pthread_mutex_lock(&all_threads_lock);
 157     gc_alloc_update_page_tables(0, &th->alloc_region);
 158     unlink_thread(th);
 159     pthread_mutex_unlock(&all_threads_lock);
 160
 161     if(th->tls_cookie>=0) arch_os_thread_cleanup(th);
 162     os_invalidate((os_vm_address_t)th->interrupt_data,
 163                   (sizeof (struct interrupt_data)));
 164     free_thread_stack_later(th);
 165     FSHOW((stderr,"/exiting thread %lu\n", thread_self()));
 166     return result;
 167 }
 168
 169 #endif /* LISP_FEATURE_SB_THREAD */
 170
 171 static void
 172 free_thread_struct(struct thread *th)
 173 {
 174     if (th->interrupt_data)
 175         os_invalidate((os_vm_address_t) th->interrupt_data,
 176                       (sizeof (struct interrupt_data)));
 177     os_invalidate((os_vm_address_t) th->control_stack_start,
 178                   THREAD_STRUCT_SIZE);
 179 }
 180
 181 /* this is called from any other thread to create the new one, and
 182  * initialize all parts of it that can be initialized from another
 183  * thread
 184  */
 185
 186 static struct thread *
 187 create_thread_struct(lispobj initial_function) {
 188     union per_thread_data *per_thread;
 189     struct thread *th=0;        /*  subdue gcc */
 190     void *spaces=0;
 191 #ifdef LISP_FEATURE_SB_THREAD
 192     int i;
 193 #endif
 194
 195     /* may as well allocate all the spaces at once: it saves us from
 196      * having to decide what to do if only some of the allocations
 197      * succeed */
 198     spaces=os_validate(0, THREAD_STRUCT_SIZE);
 199     if(!spaces)
 200          return NULL;
 201     per_thread=(union per_thread_data *)
 202         (spaces+
 203          THREAD_CONTROL_STACK_SIZE+
 204          BINDING_STACK_SIZE+
 205          ALIEN_STACK_SIZE);
 206
 207 #ifdef LISP_FEATURE_SB_THREAD
 208     for(i = 0; i < (dynamic_values_bytes / sizeof(lispobj)); i++)
 209         per_thread->dynamic_values[i] = NO_TLS_VALUE_MARKER_WIDETAG;
 210     if (all_threads == 0) {
 211         if(SymbolValue(FREE_TLS_INDEX,0)==UNBOUND_MARKER_WIDETAG) {
 212             SetSymbolValue
 213                 (FREE_TLS_INDEX,
 214                  /* FIXME: should be MAX_INTERRUPTS -1 ? */
 215                  make_fixnum(MAX_INTERRUPTS+
 216                              sizeof(struct thread)/sizeof(lispobj)),
 217                  0);
 218             SetSymbolValue(TLS_INDEX_LOCK,make_fixnum(0),0);
 219         }
 220 #define STATIC_TLS_INIT(sym,field) \
 221   ((struct symbol *)(sym-OTHER_POINTER_LOWTAG))->tls_index= \
 222   make_fixnum(THREAD_SLOT_OFFSET_WORDS(field))
 223
 224         STATIC_TLS_INIT(BINDING_STACK_START,binding_stack_start);
 225         STATIC_TLS_INIT(BINDING_STACK_POINTER,binding_stack_pointer);
 226         STATIC_TLS_INIT(CONTROL_STACK_START,control_stack_start);
 227         STATIC_TLS_INIT(CONTROL_STACK_END,control_stack_end);
 228         STATIC_TLS_INIT(ALIEN_STACK,alien_stack_pointer);
 229 #if defined(LISP_FEATURE_X86) || defined (LISP_FEATURE_X86_64)
 230         STATIC_TLS_INIT(PSEUDO_ATOMIC_ATOMIC,pseudo_atomic_atomic);
 231         STATIC_TLS_INIT(PSEUDO_ATOMIC_INTERRUPTED,pseudo_atomic_interrupted);
 232 #endif
 233 #undef STATIC_TLS_INIT
 234     }
 235 #endif
 236
 237     th=&per_thread->thread;
 238     th->control_stack_start = spaces;
 239     th->binding_stack_start=
 240         (lispobj*)((void*)th->control_stack_start+THREAD_CONTROL_STACK_SIZE);
 241     th->control_stack_end = th->binding_stack_start;
 242     th->alien_stack_start=
 243         (lispobj*)((void*)th->binding_stack_start+BINDING_STACK_SIZE);
 244     th->binding_stack_pointer=th->binding_stack_start;
 245     th->this=th;
 246     th->os_thread=0;
 247     th->state=STATE_RUNNING;
 248 #ifdef LISP_FEATURE_STACK_GROWS_DOWNWARD_NOT_UPWARD
 249     th->alien_stack_pointer=((void *)th->alien_stack_start
 250                              + ALIEN_STACK_SIZE-N_WORD_BYTES);
 251 #else
 252     th->alien_stack_pointer=((void *)th->alien_stack_start);
 253 #endif
 254 #if defined(LISP_FEATURE_X86) || defined (LISP_FEATURE_X86_64)
 255     th->pseudo_atomic_interrupted=0;
 256     th->pseudo_atomic_atomic=0;
 257 #endif
 258 #ifdef LISP_FEATURE_GENCGC
 259     gc_set_region_empty(&th->alloc_region);
 260 #endif
 261
 262 #ifndef LISP_FEATURE_SB_THREAD
 263     /* the tls-points-into-struct-thread trick is only good for threaded
 264      * sbcl, because unithread sbcl doesn't have tls.  So, we copy the
 265      * appropriate values from struct thread here, and make sure that
 266      * we use the appropriate SymbolValue macros to access any of the
 267      * variable quantities from the C runtime.  It's not quite OAOOM,
 268      * it just feels like it */
 269     SetSymbolValue(BINDING_STACK_START,(lispobj)th->binding_stack_start,th);
 270     SetSymbolValue(CONTROL_STACK_START,(lispobj)th->control_stack_start,th);
 271     SetSymbolValue(CONTROL_STACK_END,(lispobj)th->control_stack_end,th);
 272 #if defined(LISP_FEATURE_X86) || defined (LISP_FEATURE_X86_64)
 273     SetSymbolValue(BINDING_STACK_POINTER,(lispobj)th->binding_stack_pointer,th);
 274     SetSymbolValue(ALIEN_STACK,(lispobj)th->alien_stack_pointer,th);
 275     SetSymbolValue(PSEUDO_ATOMIC_ATOMIC,(lispobj)th->pseudo_atomic_atomic,th);
 276     SetSymbolValue(PSEUDO_ATOMIC_INTERRUPTED,th->pseudo_atomic_interrupted,th);
 277 #else
 278     current_binding_stack_pointer=th->binding_stack_pointer;
 279     current_control_stack_pointer=th->control_stack_start;
 280 #endif
 281 #endif
 282     bind_variable(CURRENT_CATCH_BLOCK,make_fixnum(0),th);
 283     bind_variable(CURRENT_UNWIND_PROTECT_BLOCK,make_fixnum(0),th);
 284     bind_variable(FREE_INTERRUPT_CONTEXT_INDEX,make_fixnum(0),th);
 285     bind_variable(INTERRUPT_PENDING, NIL,th);
 286     bind_variable(INTERRUPTS_ENABLED,T,th);
 287     bind_variable(GC_PENDING,NIL,th);
 288 #ifdef LISP_FEATURE_SB_THREAD
 289     bind_variable(STOP_FOR_GC_PENDING,NIL,th);
 290 #endif
 291
 292     th->interrupt_data = (struct interrupt_data *)
 293         os_validate(0,(sizeof (struct interrupt_data)));
 294     if (!th->interrupt_data) {
 295         free_thread_struct(th);
 296         return 0;
 297     }
 298     th->interrupt_data->pending_handler = 0;
 299     th->no_tls_value_marker=initial_function;
 300     return th;
 301 }
 302
 303 void create_initial_thread(lispobj initial_function) {
 304     struct thread *th=create_thread_struct(initial_function);
 305     if(th) {
 306         initial_thread_trampoline(th); /* no return */
 307     } else lose("can't create initial thread");
 308 }
 309
 310 #ifdef LISP_FEATURE_SB_THREAD
 311
 312 #ifndef __USE_XOPEN2K
 313 extern int pthread_attr_setstack (pthread_attr_t *__attr, void *__stackaddr,
 314                                   size_t __stacksize);
 315 #endif
 316
 317 boolean create_os_thread(struct thread *th,os_thread_t *kid_tid)
 318 {
 319     /* The new thread inherits the restrictive signal mask set here,
 320      * and enables signals again when it is set up properly. */
 321     pthread_attr_t attr;
 322     sigset_t newset,oldset;
 323     boolean r=1;
 324     sigemptyset(&newset);
 325     /* Blocking deferrable signals is enough, no need to block
 326      * SIG_STOP_FOR_GC because the child process is not linked onto
 327      * all_threads until it's ready. */
 328     sigaddset_deferrable(&newset);
 329     thread_sigmask(SIG_BLOCK, &newset, &oldset);
 330
 331     if((pthread_attr_init(&attr)) ||
 332        (pthread_attr_setstack(&attr,th->control_stack_start,
 333                               THREAD_CONTROL_STACK_SIZE-16)) ||
 334        (pthread_create
 335         (kid_tid,&attr,(void *(*)(void *))new_thread_trampoline,th)))
 336         r=0;
 337     thread_sigmask(SIG_SETMASK,&oldset,0);
 338     return r;
 339 }
 340
 341 os_thread_t create_thread(lispobj initial_function) {
 342     struct thread *th;
 343     os_thread_t kid_tid;
 344
 345     if(linux_no_threads_p) return 0;
 346
 347     /* Assuming that a fresh thread struct has no lisp objects in it,
 348      * linking it to all_threads can be left to the thread itself
 349      * without fear of gc lossage. initial_function violates this
 350      * assumption and must stay pinned until the child starts up. */
 351     th = create_thread_struct(initial_function);
 352     if(th==0) return 0;
 353
 354     if (create_os_thread(th,&kid_tid)) {
 355         return kid_tid;
 356     } else {
 357         free_thread_struct(th);
 358         return 0;
 359     }
 360 }
 361
 362 /* Send the signo to os_thread, retry if the rt signal queue is
 363  * full. */
 364 static int kill_thread_safely(os_thread_t os_thread, int signo)
 365 {
 366     int r;
 367     /* The man page does not mention EAGAIN as a valid return value
 368      * for either pthread_kill or kill. But that's theory, this is
 369      * practice. By waiting here we assume that the delivery of this
 370      * signal is not necessary for the delivery of the signals in the
 371      * queue. In other words, we _assume_ there are no deadlocks. */
 372     while ((r=pthread_kill(os_thread,signo))==EAGAIN) {
 373         /* wait a bit then try again in the hope of the rt signal
 374          * queue not being full */
 375         FSHOW_SIGNAL((stderr,"/rt signal queue full\n"));
 376         /* FIXME: some kind of backoff (random, exponential) would be
 377          * nice. */
 378         sleep(1);
 379     }
 380     return r;
 381 }
 382
 383 int signal_interrupt_thread(os_thread_t os_thread)
 384 {
 385     int status = kill_thread_safely(os_thread, SIG_INTERRUPT_THREAD);
 386     if (status == 0) {
 387         return 0;
 388     } else if (status == ESRCH) {
 389         return -1;
 390     } else {
 391         lose("cannot send SIG_INTERRUPT_THREAD to thread=%lu: %d, %s",
 392              os_thread, status, strerror(status));
 393     }
 394 }
 395
 396 /* stopping the world is a two-stage process.  From this thread we signal
 397  * all the others with SIG_STOP_FOR_GC.  The handler for this signal does
 398  * the usual pseudo-atomic checks (we don't want to stop a thread while
 399  * it's in the middle of allocation) then waits for another SIG_STOP_FOR_GC.
 400  */
 401
 402 /* To avoid deadlocks when gc stops the world all clients of each
 403  * mutex must enable or disable SIG_STOP_FOR_GC for the duration of
 404  * holding the lock, but they must agree on which. */
 405 void gc_stop_the_world()
 406 {
 407     struct thread *p,*th=arch_os_get_current_thread();
 408     int status;
 409     FSHOW_SIGNAL((stderr,"/gc_stop_the_world:waiting on lock, thread=%lu\n",
 410                   th->os_thread));
 411     /* keep threads from starting while the world is stopped. */
 412     pthread_mutex_lock(&all_threads_lock); \
 413     FSHOW_SIGNAL((stderr,"/gc_stop_the_world:got lock, thread=%lu\n",
 414                   th->os_thread));
 415     /* stop all other threads by sending them SIG_STOP_FOR_GC */
 416     for(p=all_threads; p; p=p->next) {
 417         gc_assert(p->os_thread != 0);
 418         if((p!=th) && ((p->state==STATE_RUNNING))) {
 419             FSHOW_SIGNAL((stderr,"/gc_stop_the_world: suspending %lu\n",
 420                           p->os_thread));
 421             status=kill_thread_safely(p->os_thread,SIG_STOP_FOR_GC);
 422             if (status==ESRCH) {
 423                 /* This thread has exited. */
 424                 gc_assert(p->state==STATE_DEAD);
 425             } else if (status) {
 426                 lose("cannot send suspend thread=%lu: %d, %s",
 427                      p->os_thread,status,strerror(status));
 428             }
 429         }
 430     }
 431     FSHOW_SIGNAL((stderr,"/gc_stop_the_world:signals sent\n"));
 432     /* wait for the running threads to stop or finish */
 433     for(p=all_threads;p;) {
 434         if((p!=th) && (p->state==STATE_RUNNING)) {
 435             sched_yield();
 436         } else {
 437             p=p->next;
 438         }
 439     }
 440     FSHOW_SIGNAL((stderr,"/gc_stop_the_world:end\n"));
 441 }
 442
 443 void gc_start_the_world()
 444 {
 445     struct thread *p,*th=arch_os_get_current_thread();
 446     int status;
 447     /* if a resumed thread creates a new thread before we're done with
 448      * this loop, the new thread will get consed on the front of
 449      * all_threads, but it won't have been stopped so won't need
 450      * restarting */
 451     FSHOW_SIGNAL((stderr,"/gc_start_the_world:begin\n"));
 452     for(p=all_threads;p;p=p->next) {
 453         gc_assert(p->os_thread!=0);
 454         if((p!=th) && (p->state!=STATE_DEAD)) {
 455             if(p->state!=STATE_SUSPENDED) {
 456                 lose("gc_start_the_world: wrong thread state is %d\n",
 457                      fixnum_value(p->state));
 458             }
 459             FSHOW_SIGNAL((stderr, "/gc_start_the_world: resuming %lu\n",
 460                           p->os_thread));
 461             p->state=STATE_RUNNING;
 462             status=kill_thread_safely(p->os_thread,SIG_STOP_FOR_GC);
 463             if (status) {
 464                 lose("cannot resume thread=%lu: %d, %s",
 465                      p->os_thread,status,strerror(status));
 466             }
 467         }
 468     }
 469     /* If we waited here until all threads leave STATE_SUSPENDED, then
 470      * SIG_STOP_FOR_GC wouldn't need to be a rt signal. That has some
 471      * performance implications, but does away with the 'rt signal
 472      * queue full' problem. */
 473     pthread_mutex_unlock(&all_threads_lock); \
 474     FSHOW_SIGNAL((stderr,"/gc_start_the_world:end\n"));
 475 }
 476 #endif