cc9eebd0ed3cec0c8a8385be10b5f5a843d574d6
[sbcl.git] / 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 "runtime.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 "thread.h"
37 #include "arch.h"
38 #include "target-arch-os.h"
39 #include "os.h"
40 #include "globals.h"
41 #include "dynbind.h"
42 #include "genesis/cons.h"
43 #include "genesis/fdefn.h"
44 #include "interr.h"             /* for lose() */
45 #include "alloc.h"
46 #include "gc-internal.h"
47 #include "cpputil.h"
48 #include "pseudo-atomic.h"
49 #include "interrupt.h"
50 #include "lispregs.h"
51
52 #if defined(LISP_FEATURE_WIN32) && defined(LISP_FEATURE_SB_THREAD)
53 # define IMMEDIATE_POST_MORTEM
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 #ifdef LISP_FEATURE_SB_THREAD
67 struct thread_post_mortem {
68 #ifdef DELAY_THREAD_POST_MORTEM
69     struct thread_post_mortem *next;
70 #endif
71     os_thread_t os_thread;
72     pthread_attr_t *os_attr;
73     os_vm_address_t os_address;
74 };
75
76 #ifdef DELAY_THREAD_POST_MORTEM
77 static int pending_thread_post_mortem_count = 0;
78 pthread_mutex_t thread_post_mortem_lock = PTHREAD_MUTEX_INITIALIZER;
79 #endif
80 static struct thread_post_mortem * volatile pending_thread_post_mortem = 0;
81 #endif
82
83 int dynamic_values_bytes=TLS_SIZE*sizeof(lispobj);  /* same for all threads */
84 struct thread *all_threads;
85 extern struct interrupt_data * global_interrupt_data;
86
87 #ifdef LISP_FEATURE_SB_THREAD
88 pthread_mutex_t all_threads_lock = PTHREAD_MUTEX_INITIALIZER;
89 #ifdef LOCK_CREATE_THREAD
90 static pthread_mutex_t create_thread_lock = PTHREAD_MUTEX_INITIALIZER;
91 #endif
92 #ifdef LISP_FEATURE_GCC_TLS
93 __thread struct thread *current_thread;
94 #endif
95 pthread_key_t lisp_thread = 0;
96 #endif
97
98 #if defined(LISP_FEATURE_X86) || defined(LISP_FEATURE_X86_64)
99 extern lispobj call_into_lisp_first_time(lispobj fun, lispobj *args, int nargs);
100 #endif
101
102 static void
103 link_thread(struct thread *th)
104 {
105     if (all_threads) all_threads->prev=th;
106     th->next=all_threads;
107     th->prev=0;
108     all_threads=th;
109 }
110
111 #ifdef LISP_FEATURE_SB_THREAD
112 static void
113 unlink_thread(struct thread *th)
114 {
115     if (th->prev)
116         th->prev->next = th->next;
117     else
118         all_threads = th->next;
119     if (th->next)
120         th->next->prev = th->prev;
121 }
122
123 /* Only access thread state with blockables blocked. */
124 lispobj
125 thread_state(struct thread *thread)
126 {
127     lispobj state;
128     sigset_t old;
129     block_blockable_signals(NULL, &old);
130     os_sem_wait(thread->state_sem, "thread_state");
131     state = thread->state;
132     os_sem_post(thread->state_sem, "thread_state");
133     thread_sigmask(SIG_SETMASK, &old, NULL);
134     return state;
135 }
136
137 void
138 set_thread_state(struct thread *thread, lispobj state)
139 {
140     int i, waitcount = 0;
141     sigset_t old;
142     block_blockable_signals(NULL, &old);
143     os_sem_wait(thread->state_sem, "set_thread_state");
144     if (thread->state != state) {
145         if ((STATE_STOPPED==state) ||
146             (STATE_DEAD==state)) {
147             waitcount = thread->state_not_running_waitcount;
148             thread->state_not_running_waitcount = 0;
149             for (i=0; i<waitcount; i++)
150                 os_sem_post(thread->state_not_running_sem, "set_thread_state (not running)");
151         }
152         if ((STATE_RUNNING==state) ||
153             (STATE_DEAD==state)) {
154             waitcount = thread->state_not_stopped_waitcount;
155             thread->state_not_stopped_waitcount = 0;
156             for (i=0; i<waitcount; i++)
157                 os_sem_post(thread->state_not_stopped_sem, "set_thread_state (not stopped)");
158         }
159         thread->state = state;
160     }
161     os_sem_post(thread->state_sem, "set_thread_state");
162     thread_sigmask(SIG_SETMASK, &old, NULL);
163 }
164
165 void
166 wait_for_thread_state_change(struct thread *thread, lispobj state)
167 {
168     sigset_t old;
169     os_sem_t *wait_sem;
170     block_blockable_signals(NULL, &old);
171   start:
172     os_sem_wait(thread->state_sem, "wait_for_thread_state_change");
173     if (thread->state == state) {
174         switch (state) {
175         case STATE_RUNNING:
176             wait_sem = thread->state_not_running_sem;
177             thread->state_not_running_waitcount++;
178             break;
179         case STATE_STOPPED:
180             wait_sem = thread->state_not_stopped_sem;
181             thread->state_not_stopped_waitcount++;
182             break;
183         default:
184             lose("Invalid state in wait_for_thread_state_change: "OBJ_FMTX"\n", state);
185         }
186     } else {
187         wait_sem = NULL;
188     }
189     os_sem_post(thread->state_sem, "wait_for_thread_state_change");
190     if (wait_sem) {
191         os_sem_wait(wait_sem, "wait_for_thread_state_change");
192         goto start;
193     }
194     thread_sigmask(SIG_SETMASK, &old, NULL);
195 }
196 #endif
197
198 static int
199 initial_thread_trampoline(struct thread *th)
200 {
201     lispobj function;
202 #if defined(LISP_FEATURE_X86) || defined(LISP_FEATURE_X86_64)
203     lispobj *args = NULL;
204 #endif
205 #ifdef LISP_FEATURE_SB_THREAD
206     pthread_setspecific(lisp_thread, (void *)1);
207 #endif
208 #if defined(THREADS_USING_GCSIGNAL) && defined(LISP_FEATURE_PPC)
209     /* SIG_STOP_FOR_GC defaults to blocked on PPC? */
210     unblock_gc_signals(0,0);
211 #endif
212     function = th->no_tls_value_marker;
213     th->no_tls_value_marker = NO_TLS_VALUE_MARKER_WIDETAG;
214     if(arch_os_thread_init(th)==0) return 1;
215 #ifdef LISP_FEATURE_SB_SAFEPOINT
216     pthread_mutex_lock(thread_qrl(th));
217 #endif
218     link_thread(th);
219     th->os_thread=thread_self();
220 #ifndef LISP_FEATURE_WIN32
221     protect_control_stack_hard_guard_page(1, NULL);
222 #endif
223     protect_binding_stack_hard_guard_page(1, NULL);
224     protect_alien_stack_hard_guard_page(1, NULL);
225 #ifndef LISP_FEATURE_WIN32
226     protect_control_stack_guard_page(1, NULL);
227 #endif
228     protect_binding_stack_guard_page(1, NULL);
229     protect_alien_stack_guard_page(1, NULL);
230
231 #if defined(LISP_FEATURE_X86) || defined(LISP_FEATURE_X86_64)
232     return call_into_lisp_first_time(function,args,0);
233 #else
234     return funcall0(function);
235 #endif
236 }
237
238 #ifdef LISP_FEATURE_SB_THREAD
239
240 # if defined(IMMEDIATE_POST_MORTEM)
241
242 /*
243  * If this feature is set, we are running on a stack managed by the OS,
244  * and no fancy delays are required for anything.  Just do it.
245  */
246 static void
247 schedule_thread_post_mortem(struct thread *corpse)
248 {
249     pthread_detach(pthread_self());
250     gc_assert(!pthread_attr_destroy(corpse->os_attr));
251     free(corpse->os_attr);
252 #if defined(LISP_FEATURE_WIN32)
253     os_invalidate_free(corpse->os_address, THREAD_STRUCT_SIZE);
254 #else
255     os_invalidate(corpse->os_address, THREAD_STRUCT_SIZE);
256 #endif
257 }
258
259 # else
260
261 /* THREAD POST MORTEM CLEANUP
262  *
263  * Memory allocated for the thread stacks cannot be reclaimed while
264  * the thread is still alive, so we need a mechanism for post mortem
265  * cleanups. FIXME: We actually have three, for historical reasons as
266  * the saying goes. Do we really need three? Nikodemus guesses that
267  * not anymore, now that we properly call pthread_attr_destroy before
268  * freeing the stack. */
269
270 static struct thread_post_mortem *
271 plan_thread_post_mortem(struct thread *corpse)
272 {
273     if (corpse) {
274         struct thread_post_mortem *post_mortem = malloc(sizeof(struct thread_post_mortem));
275         gc_assert(post_mortem);
276         post_mortem->os_thread = corpse->os_thread;
277         post_mortem->os_attr = corpse->os_attr;
278         post_mortem->os_address = corpse->os_address;
279 #ifdef DELAY_THREAD_POST_MORTEM
280         post_mortem->next = NULL;
281 #endif
282         return post_mortem;
283     } else {
284         /* FIXME: When does this happen? */
285         return NULL;
286     }
287 }
288
289 static void
290 perform_thread_post_mortem(struct thread_post_mortem *post_mortem)
291 {
292 #ifdef CREATE_POST_MORTEM_THREAD
293     pthread_detach(pthread_self());
294 #endif
295     if (post_mortem) {
296         gc_assert(!pthread_join(post_mortem->os_thread, NULL));
297         gc_assert(!pthread_attr_destroy(post_mortem->os_attr));
298         free(post_mortem->os_attr);
299         os_invalidate(post_mortem->os_address, THREAD_STRUCT_SIZE);
300         free(post_mortem);
301     }
302 }
303
304 static void
305 schedule_thread_post_mortem(struct thread *corpse)
306 {
307     struct thread_post_mortem *post_mortem = NULL;
308     if (corpse) {
309         post_mortem = plan_thread_post_mortem(corpse);
310
311 #ifdef DELAY_THREAD_POST_MORTEM
312         pthread_mutex_lock(&thread_post_mortem_lock);
313         /* First stick the new post mortem to the end of the queue. */
314         if (pending_thread_post_mortem) {
315             struct thread_post_mortem *next = pending_thread_post_mortem;
316             while (next->next) {
317                 next = next->next;
318             }
319             next->next = post_mortem;
320         } else {
321             pending_thread_post_mortem = post_mortem;
322         }
323         /* Then, if there are enough things in the queue, clean up one
324          * from the head -- or increment the count, and null out the
325          * post_mortem we have. */
326         if (pending_thread_post_mortem_count > DELAY_THREAD_POST_MORTEM) {
327             post_mortem = pending_thread_post_mortem;
328             pending_thread_post_mortem = post_mortem->next;
329         } else {
330             pending_thread_post_mortem_count++;
331             post_mortem = NULL;
332         }
333         pthread_mutex_unlock(&thread_post_mortem_lock);
334         /* Finally run, the cleanup, if any. */
335         perform_thread_post_mortem(post_mortem);
336 #elif defined(CREATE_POST_MORTEM_THREAD)
337         gc_assert(!pthread_create(&thread, NULL, perform_thread_post_mortem, post_mortem));
338 #else
339         post_mortem = (struct thread_post_mortem *)
340             swap_lispobjs((lispobj *)(void *)&pending_thread_post_mortem,
341                           (lispobj)post_mortem);
342         perform_thread_post_mortem(post_mortem);
343 #endif
344     }
345 }
346
347 # endif /* !IMMEDIATE_POST_MORTEM */
348
349 /* this is the first thing that runs in the child (which is why the
350  * silly calling convention).  Basically it calls the user's requested
351  * lisp function after doing arch_os_thread_init and whatever other
352  * bookkeeping needs to be done
353  */
354 int
355 new_thread_trampoline(struct thread *th)
356 {
357     lispobj function;
358     int result, lock_ret;
359
360     FSHOW((stderr,"/creating thread %lu\n", thread_self()));
361     check_deferrables_blocked_or_lose(0);
362 #ifndef LISP_FEATURE_SB_SAFEPOINT
363     check_gc_signals_unblocked_or_lose(0);
364 #endif
365     pthread_setspecific(lisp_thread, (void *)1);
366     function = th->no_tls_value_marker;
367     th->no_tls_value_marker = NO_TLS_VALUE_MARKER_WIDETAG;
368     if(arch_os_thread_init(th)==0) {
369         /* FIXME: handle error */
370         lose("arch_os_thread_init failed\n");
371     }
372
373     th->os_thread=thread_self();
374     protect_control_stack_guard_page(1, NULL);
375     protect_binding_stack_guard_page(1, NULL);
376     protect_alien_stack_guard_page(1, NULL);
377     /* Since GC can only know about this thread from the all_threads
378      * list and we're just adding this thread to it, there is no
379      * danger of deadlocking even with SIG_STOP_FOR_GC blocked (which
380      * it is not). */
381 #ifdef LISP_FEATURE_SB_SAFEPOINT
382     *th->csp_around_foreign_call = (lispobj)&function;
383     pthread_mutex_lock(thread_qrl(th));
384 #endif
385     lock_ret = pthread_mutex_lock(&all_threads_lock);
386     gc_assert(lock_ret == 0);
387     link_thread(th);
388     lock_ret = pthread_mutex_unlock(&all_threads_lock);
389     gc_assert(lock_ret == 0);
390
391     /* Kludge: Changed the order of some steps between the safepoint/
392      * non-safepoint versions of this code.  Can we unify this more?
393      */
394 #ifdef LISP_FEATURE_SB_SAFEPOINT
395     WITH_GC_AT_SAFEPOINTS_ONLY() {
396         result = funcall0(function);
397         block_blockable_signals(0, 0);
398         gc_alloc_update_page_tables(BOXED_PAGE_FLAG, &th->alloc_region);
399     }
400     lock_ret = pthread_mutex_lock(&all_threads_lock);
401     gc_assert(lock_ret == 0);
402     unlink_thread(th);
403     lock_ret = pthread_mutex_unlock(&all_threads_lock);
404     gc_assert(lock_ret == 0);
405     pthread_mutex_unlock(thread_qrl(th));
406     set_thread_state(th,STATE_DEAD);
407 #else
408     result = funcall0(function);
409
410     /* Block GC */
411     block_blockable_signals(0, 0);
412     set_thread_state(th, STATE_DEAD);
413
414     /* SIG_STOP_FOR_GC is blocked and GC might be waiting for this
415      * thread, but since we are already dead it won't wait long. */
416     lock_ret = pthread_mutex_lock(&all_threads_lock);
417     gc_assert(lock_ret == 0);
418
419     gc_alloc_update_page_tables(BOXED_PAGE_FLAG, &th->alloc_region);
420     unlink_thread(th);
421     pthread_mutex_unlock(&all_threads_lock);
422     gc_assert(lock_ret == 0);
423 #endif
424
425     if(th->tls_cookie>=0) arch_os_thread_cleanup(th);
426     os_sem_destroy(th->state_sem);
427     os_sem_destroy(th->state_not_running_sem);
428     os_sem_destroy(th->state_not_stopped_sem);
429
430 #if defined(LISP_FEATURE_WIN32)
431     free((os_vm_address_t)th->interrupt_data);
432 #else
433     os_invalidate((os_vm_address_t)th->interrupt_data,
434                   (sizeof (struct interrupt_data)));
435 #endif
436
437 #ifdef LISP_FEATURE_MACH_EXCEPTION_HANDLER
438     mach_lisp_thread_destroy(th);
439 #endif
440
441 #if defined(LISP_FEATURE_WIN32)
442     int i;
443     for (i = 0; i<
444              (int) (sizeof(th->private_events.events)/
445                     sizeof(th->private_events.events[0])); ++i) {
446       CloseHandle(th->private_events.events[i]);
447     }
448     TlsSetValue(OUR_TLS_INDEX,NULL);
449 #endif
450
451     schedule_thread_post_mortem(th);
452     FSHOW((stderr,"/exiting thread %lu\n", thread_self()));
453     return result;
454 }
455
456 #endif /* LISP_FEATURE_SB_THREAD */
457
458 static void
459 free_thread_struct(struct thread *th)
460 {
461 #if defined(LISP_FEATURE_WIN32)
462     if (th->interrupt_data) {
463         os_invalidate_free((os_vm_address_t) th->interrupt_data,
464                       (sizeof (struct interrupt_data)));
465     }
466     os_invalidate_free((os_vm_address_t) th->os_address,
467                   THREAD_STRUCT_SIZE);
468 #else
469     if (th->interrupt_data)
470         os_invalidate((os_vm_address_t) th->interrupt_data,
471                       (sizeof (struct interrupt_data)));
472     os_invalidate((os_vm_address_t) th->os_address,
473                   THREAD_STRUCT_SIZE);
474 #endif
475 }
476
477 #ifdef LISP_FEATURE_SB_THREAD
478 /* FIXME: should be MAX_INTERRUPTS -1 ? */
479 const unsigned int tls_index_start =
480   MAX_INTERRUPTS + sizeof(struct thread)/sizeof(lispobj);
481 #endif
482
483 /* this is called from any other thread to create the new one, and
484  * initialize all parts of it that can be initialized from another
485  * thread
486  */
487
488 static struct thread *
489 create_thread_struct(lispobj initial_function) {
490     union per_thread_data *per_thread;
491     struct thread *th=0;        /*  subdue gcc */
492     void *spaces=0;
493     void *aligned_spaces=0;
494 #if defined(LISP_FEATURE_SB_THREAD) || defined(LISP_FEATURE_WIN32)
495     unsigned int i;
496 #endif
497
498     /* May as well allocate all the spaces at once: it saves us from
499      * having to decide what to do if only some of the allocations
500      * succeed. SPACES must be appropriately aligned, since the GC
501      * expects the control stack to start at a page boundary -- and
502      * the OS may have even more rigorous requirements. We can't rely
503      * on the alignment passed from os_validate, since that might
504      * assume the current (e.g. 4k) pagesize, while we calculate with
505      * the biggest (e.g. 64k) pagesize allowed by the ABI. */
506     spaces=os_validate(0, THREAD_STRUCT_SIZE);
507     if(!spaces)
508         return NULL;
509     /* Aligning up is safe as THREAD_STRUCT_SIZE has
510      * THREAD_ALIGNMENT_BYTES padding. */
511     aligned_spaces = (void *)((((unsigned long)(char *)spaces)
512                                + THREAD_ALIGNMENT_BYTES-1)
513                               &~(unsigned long)(THREAD_ALIGNMENT_BYTES-1));
514     void* csp_page=
515         (aligned_spaces+
516          thread_control_stack_size+
517          BINDING_STACK_SIZE+
518          ALIEN_STACK_SIZE);
519     per_thread=(union per_thread_data *)
520         (csp_page + THREAD_CSP_PAGE_SIZE);
521     struct nonpointer_thread_data *nonpointer_data
522         = (void *) &per_thread->dynamic_values[TLS_SIZE];
523
524 #ifdef LISP_FEATURE_SB_THREAD
525     for(i = 0; i < (dynamic_values_bytes / sizeof(lispobj)); i++)
526         per_thread->dynamic_values[i] = NO_TLS_VALUE_MARKER_WIDETAG;
527     if (all_threads == 0) {
528         if(SymbolValue(FREE_TLS_INDEX,0)==UNBOUND_MARKER_WIDETAG) {
529             SetSymbolValue(FREE_TLS_INDEX,tls_index_start << WORD_SHIFT,0);
530             SetSymbolValue(TLS_INDEX_LOCK,make_fixnum(0),0);
531         }
532 #define STATIC_TLS_INIT(sym,field) \
533   ((struct symbol *)(sym-OTHER_POINTER_LOWTAG))->tls_index= \
534   (THREAD_SLOT_OFFSET_WORDS(field) << WORD_SHIFT)
535
536         STATIC_TLS_INIT(BINDING_STACK_START,binding_stack_start);
537 #ifdef BINDING_STACK_POINTER
538         STATIC_TLS_INIT(BINDING_STACK_POINTER,binding_stack_pointer);
539 #endif
540         STATIC_TLS_INIT(CONTROL_STACK_START,control_stack_start);
541         STATIC_TLS_INIT(CONTROL_STACK_END,control_stack_end);
542 #ifdef ALIEN_STACK
543         STATIC_TLS_INIT(ALIEN_STACK,alien_stack_pointer);
544 #endif
545 #if defined(LISP_FEATURE_X86) || defined (LISP_FEATURE_X86_64)
546         STATIC_TLS_INIT(PSEUDO_ATOMIC_BITS,pseudo_atomic_bits);
547 #endif
548 #undef STATIC_TLS_INIT
549     }
550 #endif
551
552     th=&per_thread->thread;
553     th->os_address = spaces;
554     th->control_stack_start = aligned_spaces;
555     th->binding_stack_start=
556         (lispobj*)((void*)th->control_stack_start+thread_control_stack_size);
557     th->control_stack_end = th->binding_stack_start;
558     th->control_stack_guard_page_protected = T;
559     th->alien_stack_start=
560         (lispobj*)((void*)th->binding_stack_start+BINDING_STACK_SIZE);
561     set_binding_stack_pointer(th,th->binding_stack_start);
562     th->this=th;
563     th->os_thread=0;
564
565 #ifdef LISP_FEATURE_SB_SAFEPOINT
566     th->pc_around_foreign_call = 0;
567     th->csp_around_foreign_call = csp_page;
568 #endif
569
570 #ifdef LISP_FEATURE_SB_THREAD
571     /* Contrary to the "allocate all the spaces at once" comment above,
572      * the os_attr is allocated separately.  We cannot put it into the
573      * nonpointer data, because it's used for post_mortem and freed
574      * separately */
575     th->os_attr=malloc(sizeof(pthread_attr_t));
576     th->nonpointer_data = nonpointer_data;
577     th->state_sem=&nonpointer_data->state_sem;
578     th->state_not_running_sem=&nonpointer_data->state_not_running_sem;
579     th->state_not_stopped_sem=&nonpointer_data->state_not_stopped_sem;
580     th->state_not_running_waitcount = 0;
581     th->state_not_stopped_waitcount = 0;
582     os_sem_init(th->state_sem, 1);
583     os_sem_init(th->state_not_running_sem, 0);
584     os_sem_init(th->state_not_stopped_sem, 0);
585 # ifdef LISP_FEATURE_SB_SAFEPOINT
586     pthread_mutex_init(thread_qrl(th), NULL);
587 # endif
588 #endif
589     th->state=STATE_RUNNING;
590 #ifdef LISP_FEATURE_STACK_GROWS_DOWNWARD_NOT_UPWARD
591     th->alien_stack_pointer=((void *)th->alien_stack_start
592                              + ALIEN_STACK_SIZE-N_WORD_BYTES);
593 #else
594     th->alien_stack_pointer=((void *)th->alien_stack_start);
595 #endif
596 #if defined(LISP_FEATURE_X86) || defined (LISP_FEATURE_X86_64) || defined(LISP_FEATURE_SB_THREAD)
597     th->pseudo_atomic_bits=0;
598 #endif
599 #ifdef LISP_FEATURE_GENCGC
600     gc_set_region_empty(&th->alloc_region);
601 #endif
602 #ifdef LISP_FEATURE_SB_THREAD
603     /* This parallels the same logic in globals.c for the
604      * single-threaded foreign_function_call_active, KLUDGE and
605      * all. */
606 #if defined(LISP_FEATURE_X86) || defined(LISP_FEATURE_X86_64)
607     th->foreign_function_call_active = 0;
608 #else
609     th->foreign_function_call_active = 1;
610 #endif
611 #endif
612
613 #ifndef LISP_FEATURE_SB_THREAD
614     /* the tls-points-into-struct-thread trick is only good for threaded
615      * sbcl, because unithread sbcl doesn't have tls.  So, we copy the
616      * appropriate values from struct thread here, and make sure that
617      * we use the appropriate SymbolValue macros to access any of the
618      * variable quantities from the C runtime.  It's not quite OAOOM,
619      * it just feels like it */
620     SetSymbolValue(BINDING_STACK_START,(lispobj)th->binding_stack_start,th);
621     SetSymbolValue(CONTROL_STACK_START,(lispobj)th->control_stack_start,th);
622     SetSymbolValue(CONTROL_STACK_END,(lispobj)th->control_stack_end,th);
623 #if defined(LISP_FEATURE_X86) || defined (LISP_FEATURE_X86_64)
624     SetSymbolValue(ALIEN_STACK,(lispobj)th->alien_stack_pointer,th);
625     SetSymbolValue(PSEUDO_ATOMIC_BITS,(lispobj)th->pseudo_atomic_bits,th);
626 #endif
627 #endif
628     bind_variable(CURRENT_CATCH_BLOCK,make_fixnum(0),th);
629     bind_variable(CURRENT_UNWIND_PROTECT_BLOCK,make_fixnum(0),th);
630     bind_variable(FREE_INTERRUPT_CONTEXT_INDEX,make_fixnum(0),th);
631     bind_variable(INTERRUPT_PENDING, NIL,th);
632     bind_variable(INTERRUPTS_ENABLED,T,th);
633     bind_variable(ALLOW_WITH_INTERRUPTS,T,th);
634     bind_variable(GC_PENDING,NIL,th);
635     bind_variable(ALLOC_SIGNAL,NIL,th);
636 #ifdef PINNED_OBJECTS
637     bind_variable(PINNED_OBJECTS,NIL,th);
638 #endif
639 #ifdef LISP_FEATURE_SB_THREAD
640     bind_variable(STOP_FOR_GC_PENDING,NIL,th);
641 #endif
642 #if defined(LISP_FEATURE_SB_SAFEPOINT)
643     bind_variable(GC_SAFE,NIL,th);
644     bind_variable(IN_SAFEPOINT,NIL,th);
645 #endif
646 #ifdef LISP_FEATURE_SB_THRUPTION
647     bind_variable(THRUPTION_PENDING,NIL,th);
648     bind_variable(RESTART_CLUSTERS,NIL,th);
649 #endif
650 #ifndef LISP_FEATURE_C_STACK_IS_CONTROL_STACK
651     access_control_stack_pointer(th)=th->control_stack_start;
652 #endif
653
654 #if defined(LISP_FEATURE_WIN32)
655     th->interrupt_data = (struct interrupt_data *)
656         calloc((sizeof (struct interrupt_data)),1);
657 #else
658     th->interrupt_data = (struct interrupt_data *)
659         os_validate(0,(sizeof (struct interrupt_data)));
660 #endif
661     if (!th->interrupt_data) {
662         free_thread_struct(th);
663         return 0;
664     }
665     th->interrupt_data->pending_handler = 0;
666     th->interrupt_data->gc_blocked_deferrables = 0;
667 #ifdef GENCGC_IS_PRECISE
668     th->interrupt_data->allocation_trap_context = 0;
669 #endif
670     th->no_tls_value_marker=initial_function;
671
672 #if defined(LISP_FEATURE_WIN32)
673     for (i = 0; i<sizeof(th->private_events.events)/
674            sizeof(th->private_events.events[0]); ++i) {
675       th->private_events.events[i] = CreateEvent(NULL,FALSE,FALSE,NULL);
676     }
677 #endif
678     th->stepping = NIL;
679     return th;
680 }
681
682 void create_initial_thread(lispobj initial_function) {
683     struct thread *th=create_thread_struct(initial_function);
684 #ifdef LISP_FEATURE_SB_THREAD
685     pthread_key_create(&lisp_thread, 0);
686 #endif
687     if(th) {
688         initial_thread_trampoline(th); /* no return */
689     } else lose("can't create initial thread\n");
690 }
691
692 #ifdef LISP_FEATURE_SB_THREAD
693
694 #ifndef __USE_XOPEN2K
695 extern int pthread_attr_setstack (pthread_attr_t *__attr, void *__stackaddr,
696                                   size_t __stacksize);
697 #endif
698
699 boolean create_os_thread(struct thread *th,os_thread_t *kid_tid)
700 {
701     /* The new thread inherits the restrictive signal mask set here,
702      * and enables signals again when it is set up properly. */
703     sigset_t oldset;
704     boolean r=1;
705     int retcode = 0, initcode;
706
707     FSHOW_SIGNAL((stderr,"/create_os_thread: creating new thread\n"));
708
709     /* Blocking deferrable signals is enough, no need to block
710      * SIG_STOP_FOR_GC because the child process is not linked onto
711      * all_threads until it's ready. */
712     block_deferrable_signals(0, &oldset);
713
714 #ifdef LOCK_CREATE_THREAD
715     retcode = pthread_mutex_lock(&create_thread_lock);
716     gc_assert(retcode == 0);
717     FSHOW_SIGNAL((stderr,"/create_os_thread: got lock\n"));
718 #endif
719
720     if((initcode = pthread_attr_init(th->os_attr)) ||
721        /* call_into_lisp_first_time switches the stack for the initial
722         * thread. For the others, we use this. */
723 #if defined(LISP_FEATURE_WIN32)
724        (pthread_attr_setstacksize(th->os_attr, thread_control_stack_size)) ||
725 #else
726        (pthread_attr_setstack(th->os_attr,th->control_stack_start,
727                               thread_control_stack_size)) ||
728 #endif
729        (retcode = pthread_create
730         (kid_tid,th->os_attr,(void *(*)(void *))new_thread_trampoline,th))) {
731         FSHOW_SIGNAL((stderr, "init = %d\n", initcode));
732         FSHOW_SIGNAL((stderr, "pthread_create returned %d, errno %d\n",
733                       retcode, errno));
734         if(retcode < 0) {
735             perror("create_os_thread");
736         }
737         r=0;
738     }
739
740 #ifdef LOCK_CREATE_THREAD
741     retcode = pthread_mutex_unlock(&create_thread_lock);
742     gc_assert(retcode == 0);
743     FSHOW_SIGNAL((stderr,"/create_os_thread: released lock\n"));
744 #endif
745     thread_sigmask(SIG_SETMASK,&oldset,0);
746     return r;
747 }
748
749 os_thread_t create_thread(lispobj initial_function) {
750     struct thread *th, *thread = arch_os_get_current_thread();
751     os_thread_t kid_tid = 0;
752
753     /* Must defend against async unwinds. */
754     if (SymbolValue(INTERRUPTS_ENABLED, thread) != NIL)
755         lose("create_thread is not safe when interrupts are enabled.\n");
756
757     /* Assuming that a fresh thread struct has no lisp objects in it,
758      * linking it to all_threads can be left to the thread itself
759      * without fear of gc lossage. initial_function violates this
760      * assumption and must stay pinned until the child starts up. */
761     th = create_thread_struct(initial_function);
762     if (th && !create_os_thread(th,&kid_tid)) {
763         free_thread_struct(th);
764         kid_tid = 0;
765     }
766     return kid_tid;
767 }
768
769 /* stopping the world is a two-stage process.  From this thread we signal
770  * all the others with SIG_STOP_FOR_GC.  The handler for this signal does
771  * the usual pseudo-atomic checks (we don't want to stop a thread while
772  * it's in the middle of allocation) then waits for another SIG_STOP_FOR_GC.
773  */
774 /*
775  * (With SB-SAFEPOINT, see the definitions in safepoint.c instead.)
776  */
777 #ifndef LISP_FEATURE_SB_SAFEPOINT
778
779 /* To avoid deadlocks when gc stops the world all clients of each
780  * mutex must enable or disable SIG_STOP_FOR_GC for the duration of
781  * holding the lock, but they must agree on which. */
782 void gc_stop_the_world()
783 {
784     struct thread *p,*th=arch_os_get_current_thread();
785     int status, lock_ret;
786 #ifdef LOCK_CREATE_THREAD
787     /* KLUDGE: Stopping the thread during pthread_create() causes deadlock
788      * on FreeBSD. */
789     FSHOW_SIGNAL((stderr,"/gc_stop_the_world:waiting on create_thread_lock\n"));
790     lock_ret = pthread_mutex_lock(&create_thread_lock);
791     gc_assert(lock_ret == 0);
792     FSHOW_SIGNAL((stderr,"/gc_stop_the_world:got create_thread_lock\n"));
793 #endif
794     FSHOW_SIGNAL((stderr,"/gc_stop_the_world:waiting on lock\n"));
795     /* keep threads from starting while the world is stopped. */
796     lock_ret = pthread_mutex_lock(&all_threads_lock);      \
797     gc_assert(lock_ret == 0);
798
799     FSHOW_SIGNAL((stderr,"/gc_stop_the_world:got lock\n"));
800     /* stop all other threads by sending them SIG_STOP_FOR_GC */
801     for(p=all_threads; p; p=p->next) {
802         gc_assert(p->os_thread != 0);
803         FSHOW_SIGNAL((stderr,"/gc_stop_the_world: thread=%lu, state=%x\n",
804                       p->os_thread, thread_state(p)));
805         if((p!=th) && ((thread_state(p)==STATE_RUNNING))) {
806             FSHOW_SIGNAL((stderr,"/gc_stop_the_world: suspending thread %lu\n",
807                           p->os_thread));
808             /* We already hold all_thread_lock, P can become DEAD but
809              * cannot exit, ergo it's safe to use pthread_kill. */
810             status=pthread_kill(p->os_thread,SIG_STOP_FOR_GC);
811             if (status==ESRCH) {
812                 /* This thread has exited. */
813                 gc_assert(thread_state(p)==STATE_DEAD);
814             } else if (status) {
815                 lose("cannot send suspend thread=%lu: %d, %s\n",
816                      p->os_thread,status,strerror(status));
817             }
818         }
819     }
820     FSHOW_SIGNAL((stderr,"/gc_stop_the_world:signals sent\n"));
821     for(p=all_threads;p;p=p->next) {
822         if (p!=th) {
823             FSHOW_SIGNAL
824                 ((stderr,
825                   "/gc_stop_the_world: waiting for thread=%lu: state=%x\n",
826                   p->os_thread, thread_state(p)));
827             wait_for_thread_state_change(p, STATE_RUNNING);
828             if (p->state == STATE_RUNNING)
829                 lose("/gc_stop_the_world: unexpected state");
830         }
831     }
832     FSHOW_SIGNAL((stderr,"/gc_stop_the_world:end\n"));
833 }
834
835 void gc_start_the_world()
836 {
837     struct thread *p,*th=arch_os_get_current_thread();
838     int lock_ret;
839     /* if a resumed thread creates a new thread before we're done with
840      * this loop, the new thread will get consed on the front of
841      * all_threads, but it won't have been stopped so won't need
842      * restarting */
843     FSHOW_SIGNAL((stderr,"/gc_start_the_world:begin\n"));
844     for(p=all_threads;p;p=p->next) {
845         gc_assert(p->os_thread!=0);
846         if (p!=th) {
847             lispobj state = thread_state(p);
848             if (state != STATE_DEAD) {
849                 if(state != STATE_STOPPED) {
850                     lose("gc_start_the_world: wrong thread state is %d\n",
851                          fixnum_value(state));
852                 }
853                 FSHOW_SIGNAL((stderr, "/gc_start_the_world: resuming %lu\n",
854                               p->os_thread));
855                 set_thread_state(p, STATE_RUNNING);
856             }
857         }
858     }
859
860     lock_ret = pthread_mutex_unlock(&all_threads_lock);
861     gc_assert(lock_ret == 0);
862 #ifdef LOCK_CREATE_THREAD
863     lock_ret = pthread_mutex_unlock(&create_thread_lock);
864     gc_assert(lock_ret == 0);
865 #endif
866
867     FSHOW_SIGNAL((stderr,"/gc_start_the_world:end\n"));
868 }
869
870 #endif /* !LISP_FEATURE_SB_SAFEPOINT */
871 #endif /* !LISP_FEATURE_SB_THREAD */
872
873 int
874 thread_yield()
875 {
876 #ifdef LISP_FEATURE_SB_THREAD
877     return sched_yield();
878 #else
879     return 0;
880 #endif
881 }
882
883 int
884 wake_thread(os_thread_t os_thread)
885 {
886 #if defined(LISP_FEATURE_WIN32)
887     return kill_safely(os_thread, 1);
888 #elif !defined(LISP_FEATURE_SB_THRUPTION)
889     return kill_safely(os_thread, SIGPIPE);
890 #else
891     return wake_thread_posix(os_thread);
892 #endif
893 }
894
895 /* If the thread id given does not belong to a running thread (it has
896  * exited or never even existed) pthread_kill _may_ fail with ESRCH,
897  * but it is also allowed to just segfault, see
898  * <http://udrepper.livejournal.com/16844.html>.
899  *
900  * Relying on thread ids can easily backfire since ids are recycled
901  * (NPTL recycles them extremely fast) so a signal can be sent to
902  * another process if the one it was sent to exited.
903  *
904  * For these reasons, we must make sure that the thread is still alive
905  * when the pthread_kill is called and return if the thread is
906  * exiting.
907  *
908  * Note (DFL, 2011-06-22): At the time of writing, this function is only
909  * used for INTERRUPT-THREAD, hence the wake_thread special-case for
910  * Windows is OK. */
911 int
912 kill_safely(os_thread_t os_thread, int signal)
913 {
914     FSHOW_SIGNAL((stderr,"/kill_safely: %lu, %d\n", os_thread, signal));
915     {
916 #ifdef LISP_FEATURE_SB_THREAD
917         sigset_t oldset;
918         struct thread *thread;
919         /* Frequent special case: resignalling to self.  The idea is
920          * that leave_region safepoint will acknowledge the signal, so
921          * there is no need to take locks, roll thread to safepoint
922          * etc. */
923         /* Kludge (on safepoint builds): At the moment, this isn't just
924          * an optimization; rather it masks the fact that
925          * gc_stop_the_world() grabs the all_threads mutex without
926          * releasing it, and since we're not using recursive pthread
927          * mutexes, the pthread_mutex_lock() around the all_threads loop
928          * would go wrong.  Why are we running interruptions while
929          * stopping the world though?  Test case is (:ASYNC-UNWIND
930          * :SPECIALS), especially with s/10/100/ in both loops. */
931         if (os_thread == pthread_self()) {
932             pthread_kill(os_thread, signal);
933 #ifdef LISP_FEATURE_WIN32
934             check_pending_thruptions(NULL);
935 #endif
936             return 0;
937         }
938
939         /* pthread_kill is not async signal safe and we don't want to be
940          * interrupted while holding the lock. */
941         block_deferrable_signals(0, &oldset);
942         pthread_mutex_lock(&all_threads_lock);
943         for (thread = all_threads; thread; thread = thread->next) {
944             if (thread->os_thread == os_thread) {
945                 int status = pthread_kill(os_thread, signal);
946                 if (status)
947                     lose("kill_safely: pthread_kill failed with %d\n", status);
948 #if defined(LISP_FEATURE_WIN32) && defined(LISP_FEATURE_SB_THRUPTION)
949                 wake_thread_win32(thread);
950 #endif
951                 break;
952             }
953         }
954         pthread_mutex_unlock(&all_threads_lock);
955         thread_sigmask(SIG_SETMASK,&oldset,0);
956         if (thread)
957             return 0;
958         else
959             return -1;
960 #elif defined(LISP_FEATURE_WIN32)
961         return 0;
962 #else
963         int status;
964         if (os_thread != 0)
965             lose("kill_safely: who do you want to kill? %d?\n", os_thread);
966         /* Dubious (as in don't know why it works) workaround for the
967          * signal sometimes not being generated on darwin. */
968 #ifdef LISP_FEATURE_DARWIN
969         {
970             sigset_t oldset;
971             sigprocmask(SIG_BLOCK, &deferrable_sigset, &oldset);
972             status = raise(signal);
973             sigprocmask(SIG_SETMASK,&oldset,0);
974         }
975 #else
976         status = raise(signal);
977 #endif
978         if (status == 0) {
979             return 0;
980         } else {
981             lose("cannot raise signal %d, %d %s\n",
982                  signal, status, strerror(errno));
983         }
984 #endif
985     }
986 }