ca265b81b015cd94ce2d69527688ba826242a129
[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 <signal.h>
21 #include <stddef.h>
22 #include <errno.h>
23 #include <sys/types.h>
24 #ifndef LISP_FEATURE_WIN32
25 #include <sys/wait.h>
26 #endif
27
28 #ifdef LISP_FEATURE_MACH_EXCEPTION_HANDLER
29 #include <mach/mach.h>
30 #include <mach/mach_error.h>
31 #include <mach/mach_types.h>
32 #endif
33
34 #include "runtime.h"
35 #include "validate.h"           /* for BINDING_STACK_SIZE etc */
36 #include "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
48 #ifdef LISP_FEATURE_WIN32
49 /*
50  * Win32 doesn't have SIGSTKSZ, and we're not switching stacks anyway,
51  * so define it arbitrarily
52  */
53 #define SIGSTKSZ 1024
54 #endif
55
56 #if defined(LISP_FEATURE_DARWIN) && defined(LISP_FEATURE_SB_THREAD)
57 #define DELAY_THREAD_POST_MORTEM 5
58 #define LOCK_CREATE_THREAD
59 #endif
60
61 #ifdef LISP_FEATURE_FREEBSD
62 #define CREATE_CLEANUP_THREAD
63 #define LOCK_CREATE_THREAD
64 #endif
65
66 #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 #endif
123
124 static int
125 initial_thread_trampoline(struct thread *th)
126 {
127     lispobj function;
128 #if defined(LISP_FEATURE_X86) || defined(LISP_FEATURE_X86_64)
129     lispobj *args = NULL;
130 #endif
131 #ifdef LISP_FEATURE_SB_THREAD
132     pthread_setspecific(lisp_thread, (void *)1);
133 #endif
134 #if defined(LISP_FEATURE_SB_THREAD) && defined(LISP_FEATURE_PPC)
135     /* SIG_STOP_FOR_GC defaults to blocked on PPC? */
136     unblock_gc_signals(0,0);
137 #endif
138     function = th->no_tls_value_marker;
139     th->no_tls_value_marker = NO_TLS_VALUE_MARKER_WIDETAG;
140     if(arch_os_thread_init(th)==0) return 1;
141     link_thread(th);
142     th->os_thread=thread_self();
143 #ifndef LISP_FEATURE_WIN32
144     protect_control_stack_hard_guard_page(1, NULL);
145     protect_binding_stack_hard_guard_page(1, NULL);
146     protect_alien_stack_hard_guard_page(1, NULL);
147     protect_control_stack_guard_page(1, NULL);
148     protect_binding_stack_guard_page(1, NULL);
149     protect_alien_stack_guard_page(1, NULL);
150 #endif
151
152 #if defined(LISP_FEATURE_X86) || defined(LISP_FEATURE_X86_64)
153     return call_into_lisp_first_time(function,args,0);
154 #else
155     return funcall0(function);
156 #endif
157 }
158
159 #ifdef LISP_FEATURE_SB_THREAD
160 #define THREAD_STATE_LOCK_SIZE \
161     (sizeof(pthread_mutex_t))+(sizeof(pthread_cond_t))
162 #else
163 #define THREAD_STATE_LOCK_SIZE 0
164 #endif
165
166 #define THREAD_STRUCT_SIZE (thread_control_stack_size + BINDING_STACK_SIZE + \
167                             ALIEN_STACK_SIZE +                               \
168                             THREAD_STATE_LOCK_SIZE +                         \
169                             dynamic_values_bytes +                           \
170                             32 * SIGSTKSZ +                                  \
171                             THREAD_ALIGNMENT_BYTES)
172
173 #ifdef LISP_FEATURE_SB_THREAD
174 /* THREAD POST MORTEM CLEANUP
175  *
176  * Memory allocated for the thread stacks cannot be reclaimed while
177  * the thread is still alive, so we need a mechanism for post mortem
178  * cleanups. FIXME: We actually have three, for historical reasons as
179  * the saying goes. Do we really need three? Nikodemus guesses that
180  * not anymore, now that we properly call pthread_attr_destroy before
181  * freeing the stack. */
182
183 static struct thread_post_mortem *
184 plan_thread_post_mortem(struct thread *corpse)
185 {
186     if (corpse) {
187         struct thread_post_mortem *post_mortem = malloc(sizeof(struct thread_post_mortem));
188         gc_assert(post_mortem);
189         post_mortem->os_thread = corpse->os_thread;
190         post_mortem->os_attr = corpse->os_attr;
191         post_mortem->os_address = corpse->os_address;
192 #ifdef DELAY_THREAD_POST_MORTEM
193         post_mortem->next = NULL;
194 #endif
195         return post_mortem;
196     } else {
197         /* FIXME: When does this happen? */
198         return NULL;
199     }
200 }
201
202 static void
203 perform_thread_post_mortem(struct thread_post_mortem *post_mortem)
204 {
205 #ifdef CREATE_POST_MORTEM_THREAD
206     pthread_detach(pthread_self());
207 #endif
208     if (post_mortem) {
209         gc_assert(!pthread_join(post_mortem->os_thread, NULL));
210         gc_assert(!pthread_attr_destroy(post_mortem->os_attr));
211         free(post_mortem->os_attr);
212         os_invalidate(post_mortem->os_address, THREAD_STRUCT_SIZE);
213         free(post_mortem);
214     }
215 }
216
217 static void
218 schedule_thread_post_mortem(struct thread *corpse)
219 {
220     struct thread_post_mortem *post_mortem = NULL;
221     if (corpse) {
222         post_mortem = plan_thread_post_mortem(corpse);
223
224 #ifdef DELAY_THREAD_POST_MORTEM
225         pthread_mutex_lock(&thread_post_mortem_lock);
226         /* First stick the new post mortem to the end of the queue. */
227         if (pending_thread_post_mortem) {
228             struct thread_post_mortem *next = pending_thread_post_mortem;
229             while (next->next) {
230                 next = next->next;
231             }
232             next->next = post_mortem;
233         } else {
234             pending_thread_post_mortem = post_mortem;
235         }
236         /* Then, if there are enough things in the queue, clean up one
237          * from the head -- or increment the count, and null out the
238          * post_mortem we have. */
239         if (pending_thread_post_mortem_count > DELAY_THREAD_POST_MORTEM) {
240             post_mortem = pending_thread_post_mortem;
241             pending_thread_post_mortem = post_mortem->next;
242         } else {
243             pending_thread_post_mortem_count++;
244             post_mortem = NULL;
245         }
246         pthread_mutex_unlock(&thread_post_mortem_lock);
247         /* Finally run, the cleanup, if any. */
248         perform_thread_post_mortem(post_mortem);
249 #elif defined(CREATE_POST_MORTEM_THREAD)
250         gc_assert(!pthread_create(&thread, NULL, perform_thread_post_mortem, post_mortem));
251 #else
252         post_mortem = (struct thread_post_mortem *)
253             swap_lispobjs((lispobj *)(void *)&pending_thread_post_mortem,
254                           (lispobj)post_mortem);
255         perform_thread_post_mortem(post_mortem);
256 #endif
257     }
258 }
259
260 /* this is the first thing that runs in the child (which is why the
261  * silly calling convention).  Basically it calls the user's requested
262  * lisp function after doing arch_os_thread_init and whatever other
263  * bookkeeping needs to be done
264  */
265 int
266 new_thread_trampoline(struct thread *th)
267 {
268     lispobj function;
269     int result, lock_ret;
270
271     FSHOW((stderr,"/creating thread %lu\n", thread_self()));
272     check_deferrables_blocked_or_lose(0);
273     check_gc_signals_unblocked_or_lose(0);
274     pthread_setspecific(lisp_thread, (void *)1);
275     function = th->no_tls_value_marker;
276     th->no_tls_value_marker = NO_TLS_VALUE_MARKER_WIDETAG;
277     if(arch_os_thread_init(th)==0) {
278         /* FIXME: handle error */
279         lose("arch_os_thread_init failed\n");
280     }
281
282     th->os_thread=thread_self();
283     protect_control_stack_guard_page(1, NULL);
284     protect_binding_stack_guard_page(1, NULL);
285     protect_alien_stack_guard_page(1, NULL);
286     /* Since GC can only know about this thread from the all_threads
287      * list and we're just adding this thread to it, there is no
288      * danger of deadlocking even with SIG_STOP_FOR_GC blocked (which
289      * it is not). */
290     lock_ret = pthread_mutex_lock(&all_threads_lock);
291     gc_assert(lock_ret == 0);
292     link_thread(th);
293     lock_ret = pthread_mutex_unlock(&all_threads_lock);
294     gc_assert(lock_ret == 0);
295
296     result = funcall0(function);
297
298     /* Block GC */
299     block_blockable_signals(0, 0);
300     set_thread_state(th, STATE_DEAD);
301
302     /* SIG_STOP_FOR_GC is blocked and GC might be waiting for this
303      * thread, but since we are already dead it won't wait long. */
304     lock_ret = pthread_mutex_lock(&all_threads_lock);
305     gc_assert(lock_ret == 0);
306
307     gc_alloc_update_page_tables(BOXED_PAGE_FLAG, &th->alloc_region);
308     unlink_thread(th);
309     pthread_mutex_unlock(&all_threads_lock);
310     gc_assert(lock_ret == 0);
311
312     if(th->tls_cookie>=0) arch_os_thread_cleanup(th);
313     pthread_mutex_destroy(th->state_lock);
314     pthread_cond_destroy(th->state_cond);
315
316     os_invalidate((os_vm_address_t)th->interrupt_data,
317                   (sizeof (struct interrupt_data)));
318
319 #ifdef LISP_FEATURE_MACH_EXCEPTION_HANDLER
320     FSHOW((stderr, "Deallocating mach port %x\n", THREAD_STRUCT_TO_EXCEPTION_PORT(th)));
321     mach_port_move_member(current_mach_task,
322                           THREAD_STRUCT_TO_EXCEPTION_PORT(th),
323                           MACH_PORT_NULL);
324     mach_port_deallocate(current_mach_task,
325                          THREAD_STRUCT_TO_EXCEPTION_PORT(th));
326     mach_port_destroy(current_mach_task,
327                       THREAD_STRUCT_TO_EXCEPTION_PORT(th));
328 #endif
329
330     schedule_thread_post_mortem(th);
331     FSHOW((stderr,"/exiting thread %lu\n", thread_self()));
332     return result;
333 }
334
335 #endif /* LISP_FEATURE_SB_THREAD */
336
337 static void
338 free_thread_struct(struct thread *th)
339 {
340     if (th->interrupt_data)
341         os_invalidate((os_vm_address_t) th->interrupt_data,
342                       (sizeof (struct interrupt_data)));
343     os_invalidate((os_vm_address_t) th->os_address,
344                   THREAD_STRUCT_SIZE);
345 }
346
347 #ifdef LISP_FEATURE_SB_THREAD
348 /* FIXME: should be MAX_INTERRUPTS -1 ? */
349 const unsigned int tls_index_start =
350   MAX_INTERRUPTS + sizeof(struct thread)/sizeof(lispobj);
351 #endif
352
353 /* this is called from any other thread to create the new one, and
354  * initialize all parts of it that can be initialized from another
355  * thread
356  */
357
358 static struct thread *
359 create_thread_struct(lispobj initial_function) {
360     union per_thread_data *per_thread;
361     struct thread *th=0;        /*  subdue gcc */
362     void *spaces=0;
363     void *aligned_spaces=0;
364 #ifdef LISP_FEATURE_SB_THREAD
365     unsigned int i;
366 #endif
367
368     /* May as well allocate all the spaces at once: it saves us from
369      * having to decide what to do if only some of the allocations
370      * succeed. SPACES must be appropriately aligned, since the GC
371      * expects the control stack to start at a page boundary -- and
372      * the OS may have even more rigorous requirements. We can't rely
373      * on the alignment passed from os_validate, since that might
374      * assume the current (e.g. 4k) pagesize, while we calculate with
375      * the biggest (e.g. 64k) pagesize allowed by the ABI. */
376     spaces=os_validate(0, THREAD_STRUCT_SIZE);
377     if(!spaces)
378         return NULL;
379     /* Aligning up is safe as THREAD_STRUCT_SIZE has
380      * THREAD_ALIGNMENT_BYTES padding. */
381     aligned_spaces = (void *)((((unsigned long)(char *)spaces)
382                                + THREAD_ALIGNMENT_BYTES-1)
383                               &~(unsigned long)(THREAD_ALIGNMENT_BYTES-1));
384     per_thread=(union per_thread_data *)
385         (aligned_spaces+
386          thread_control_stack_size+
387          BINDING_STACK_SIZE+
388          ALIEN_STACK_SIZE +
389          THREAD_STATE_LOCK_SIZE);
390
391 #ifdef LISP_FEATURE_SB_THREAD
392     for(i = 0; i < (dynamic_values_bytes / sizeof(lispobj)); i++)
393         per_thread->dynamic_values[i] = NO_TLS_VALUE_MARKER_WIDETAG;
394     if (all_threads == 0) {
395         if(SymbolValue(FREE_TLS_INDEX,0)==UNBOUND_MARKER_WIDETAG) {
396             SetSymbolValue(FREE_TLS_INDEX,make_fixnum(tls_index_start),0);
397             SetSymbolValue(TLS_INDEX_LOCK,make_fixnum(0),0);
398         }
399 #define STATIC_TLS_INIT(sym,field) \
400   ((struct symbol *)(sym-OTHER_POINTER_LOWTAG))->tls_index= \
401   make_fixnum(THREAD_SLOT_OFFSET_WORDS(field))
402
403         STATIC_TLS_INIT(BINDING_STACK_START,binding_stack_start);
404 #ifdef BINDING_STACK_POINTER
405         STATIC_TLS_INIT(BINDING_STACK_POINTER,binding_stack_pointer);
406 #endif
407         STATIC_TLS_INIT(CONTROL_STACK_START,control_stack_start);
408         STATIC_TLS_INIT(CONTROL_STACK_END,control_stack_end);
409 #ifdef ALIEN_STACK
410         STATIC_TLS_INIT(ALIEN_STACK,alien_stack_pointer);
411 #endif
412 #if defined(LISP_FEATURE_X86) || defined (LISP_FEATURE_X86_64)
413         STATIC_TLS_INIT(PSEUDO_ATOMIC_BITS,pseudo_atomic_bits);
414 #endif
415 #undef STATIC_TLS_INIT
416     }
417 #endif
418
419     th=&per_thread->thread;
420     th->os_address = spaces;
421     th->control_stack_start = aligned_spaces;
422     th->binding_stack_start=
423         (lispobj*)((void*)th->control_stack_start+thread_control_stack_size);
424     th->control_stack_end = th->binding_stack_start;
425     th->control_stack_guard_page_protected = T;
426     th->alien_stack_start=
427         (lispobj*)((void*)th->binding_stack_start+BINDING_STACK_SIZE);
428     set_binding_stack_pointer(th,th->binding_stack_start);
429     th->this=th;
430     th->os_thread=0;
431 #ifdef LISP_FEATURE_SB_THREAD
432     th->os_attr=malloc(sizeof(pthread_attr_t));
433     th->state_lock=(pthread_mutex_t *)((void *)th->alien_stack_start +
434                                        ALIEN_STACK_SIZE);
435     pthread_mutex_init(th->state_lock, NULL);
436     th->state_cond=(pthread_cond_t *)((void *)th->state_lock +
437                                       (sizeof(pthread_mutex_t)));
438     pthread_cond_init(th->state_cond, NULL);
439 #endif
440     th->state=STATE_RUNNING;
441 #ifdef LISP_FEATURE_STACK_GROWS_DOWNWARD_NOT_UPWARD
442     th->alien_stack_pointer=((void *)th->alien_stack_start
443                              + ALIEN_STACK_SIZE-N_WORD_BYTES);
444 #else
445     th->alien_stack_pointer=((void *)th->alien_stack_start);
446 #endif
447 #if defined(LISP_FEATURE_X86) || defined (LISP_FEATURE_X86_64) || defined(LISP_FEATURE_SB_THREAD)
448     th->pseudo_atomic_bits=0;
449 #endif
450 #ifdef LISP_FEATURE_GENCGC
451     gc_set_region_empty(&th->alloc_region);
452 #endif
453 #ifdef LISP_FEATURE_SB_THREAD
454     /* This parallels the same logic in globals.c for the
455      * single-threaded foreign_function_call_active, KLUDGE and
456      * all. */
457 #if defined(LISP_FEATURE_X86) || defined(LISP_FEATURE_X86_64)
458     th->foreign_function_call_active = 0;
459 #else
460     th->foreign_function_call_active = 1;
461 #endif
462 #endif
463
464 #ifndef LISP_FEATURE_SB_THREAD
465     /* the tls-points-into-struct-thread trick is only good for threaded
466      * sbcl, because unithread sbcl doesn't have tls.  So, we copy the
467      * appropriate values from struct thread here, and make sure that
468      * we use the appropriate SymbolValue macros to access any of the
469      * variable quantities from the C runtime.  It's not quite OAOOM,
470      * it just feels like it */
471     SetSymbolValue(BINDING_STACK_START,(lispobj)th->binding_stack_start,th);
472     SetSymbolValue(CONTROL_STACK_START,(lispobj)th->control_stack_start,th);
473     SetSymbolValue(CONTROL_STACK_END,(lispobj)th->control_stack_end,th);
474 #if defined(LISP_FEATURE_X86) || defined (LISP_FEATURE_X86_64)
475     SetSymbolValue(ALIEN_STACK,(lispobj)th->alien_stack_pointer,th);
476     SetSymbolValue(PSEUDO_ATOMIC_BITS,(lispobj)th->pseudo_atomic_bits,th);
477 #endif
478 #endif
479     bind_variable(CURRENT_CATCH_BLOCK,make_fixnum(0),th);
480     bind_variable(CURRENT_UNWIND_PROTECT_BLOCK,make_fixnum(0),th);
481     bind_variable(FREE_INTERRUPT_CONTEXT_INDEX,make_fixnum(0),th);
482     bind_variable(INTERRUPT_PENDING, NIL,th);
483     bind_variable(INTERRUPTS_ENABLED,T,th);
484     bind_variable(ALLOW_WITH_INTERRUPTS,T,th);
485     bind_variable(GC_PENDING,NIL,th);
486     bind_variable(ALLOC_SIGNAL,NIL,th);
487 #ifdef PINNED_OBJECTS
488     bind_variable(PINNED_OBJECTS,NIL,th);
489 #endif
490 #ifdef LISP_FEATURE_SB_THREAD
491     bind_variable(STOP_FOR_GC_PENDING,NIL,th);
492 #endif
493 #ifndef LISP_FEATURE_C_STACK_IS_CONTROL_STACK
494     access_control_stack_pointer(th)=th->control_stack_start;
495 #endif
496
497     th->interrupt_data = (struct interrupt_data *)
498         os_validate(0,(sizeof (struct interrupt_data)));
499     if (!th->interrupt_data) {
500         free_thread_struct(th);
501         return 0;
502     }
503     th->interrupt_data->pending_handler = 0;
504     th->interrupt_data->gc_blocked_deferrables = 0;
505 #ifdef LISP_FEATURE_PPC
506     th->interrupt_data->allocation_trap_context = 0;
507 #endif
508     th->no_tls_value_marker=initial_function;
509
510     th->stepping = NIL;
511     return th;
512 }
513
514 #ifdef LISP_FEATURE_MACH_EXCEPTION_HANDLER
515 mach_port_t setup_mach_exception_handling_thread();
516 kern_return_t mach_thread_init(mach_port_t thread_exception_port);
517
518 #endif
519
520 void create_initial_thread(lispobj initial_function) {
521     struct thread *th=create_thread_struct(initial_function);
522 #ifdef LISP_FEATURE_SB_THREAD
523     pthread_key_create(&lisp_thread, 0);
524 #endif
525     if(th) {
526 #ifdef LISP_FEATURE_MACH_EXCEPTION_HANDLER
527         setup_mach_exception_handling_thread();
528 #endif
529         initial_thread_trampoline(th); /* no return */
530     } else lose("can't create initial thread\n");
531 }
532
533 #ifdef LISP_FEATURE_SB_THREAD
534
535 #ifndef __USE_XOPEN2K
536 extern int pthread_attr_setstack (pthread_attr_t *__attr, void *__stackaddr,
537                                   size_t __stacksize);
538 #endif
539
540 boolean create_os_thread(struct thread *th,os_thread_t *kid_tid)
541 {
542     /* The new thread inherits the restrictive signal mask set here,
543      * and enables signals again when it is set up properly. */
544     sigset_t oldset;
545     boolean r=1;
546     int retcode = 0, initcode;
547
548     FSHOW_SIGNAL((stderr,"/create_os_thread: creating new thread\n"));
549
550     /* Blocking deferrable signals is enough, no need to block
551      * SIG_STOP_FOR_GC because the child process is not linked onto
552      * all_threads until it's ready. */
553     block_deferrable_signals(0, &oldset);
554
555 #ifdef LOCK_CREATE_THREAD
556     retcode = pthread_mutex_lock(&create_thread_lock);
557     gc_assert(retcode == 0);
558     FSHOW_SIGNAL((stderr,"/create_os_thread: got lock\n"));
559 #endif
560
561     if((initcode = pthread_attr_init(th->os_attr)) ||
562        /* call_into_lisp_first_time switches the stack for the initial
563         * thread. For the others, we use this. */
564        (pthread_attr_setstack(th->os_attr,th->control_stack_start,
565                               thread_control_stack_size)) ||
566        (retcode = pthread_create
567         (kid_tid,th->os_attr,(void *(*)(void *))new_thread_trampoline,th))) {
568         FSHOW_SIGNAL((stderr, "init = %d\n", initcode));
569         FSHOW_SIGNAL((stderr, "pthread_create returned %d, errno %d\n",
570                       retcode, errno));
571         if(retcode < 0) {
572             perror("create_os_thread");
573         }
574         r=0;
575     }
576
577 #ifdef LOCK_CREATE_THREAD
578     retcode = pthread_mutex_unlock(&create_thread_lock);
579     gc_assert(retcode == 0);
580     FSHOW_SIGNAL((stderr,"/create_os_thread: released lock\n"));
581 #endif
582     thread_sigmask(SIG_SETMASK,&oldset,0);
583     return r;
584 }
585
586 os_thread_t create_thread(lispobj initial_function) {
587     struct thread *th, *thread = arch_os_get_current_thread();
588     os_thread_t kid_tid = 0;
589
590     /* Must defend against async unwinds. */
591     if (SymbolValue(INTERRUPTS_ENABLED, thread) != NIL)
592         lose("create_thread is not safe when interrupts are enabled.\n");
593
594     /* Assuming that a fresh thread struct has no lisp objects in it,
595      * linking it to all_threads can be left to the thread itself
596      * without fear of gc lossage. initial_function violates this
597      * assumption and must stay pinned until the child starts up. */
598     th = create_thread_struct(initial_function);
599     if (th && !create_os_thread(th,&kid_tid)) {
600         free_thread_struct(th);
601         kid_tid = 0;
602     }
603     return kid_tid;
604 }
605
606 /* stopping the world is a two-stage process.  From this thread we signal
607  * all the others with SIG_STOP_FOR_GC.  The handler for this signal does
608  * the usual pseudo-atomic checks (we don't want to stop a thread while
609  * it's in the middle of allocation) then waits for another SIG_STOP_FOR_GC.
610  */
611
612 /* To avoid deadlocks when gc stops the world all clients of each
613  * mutex must enable or disable SIG_STOP_FOR_GC for the duration of
614  * holding the lock, but they must agree on which. */
615 void gc_stop_the_world()
616 {
617     struct thread *p,*th=arch_os_get_current_thread();
618     int status, lock_ret;
619 #ifdef LOCK_CREATE_THREAD
620     /* KLUDGE: Stopping the thread during pthread_create() causes deadlock
621      * on FreeBSD. */
622     FSHOW_SIGNAL((stderr,"/gc_stop_the_world:waiting on create_thread_lock\n"));
623     lock_ret = pthread_mutex_lock(&create_thread_lock);
624     gc_assert(lock_ret == 0);
625     FSHOW_SIGNAL((stderr,"/gc_stop_the_world:got create_thread_lock\n"));
626 #endif
627     FSHOW_SIGNAL((stderr,"/gc_stop_the_world:waiting on lock\n"));
628     /* keep threads from starting while the world is stopped. */
629     lock_ret = pthread_mutex_lock(&all_threads_lock);      \
630     gc_assert(lock_ret == 0);
631
632     FSHOW_SIGNAL((stderr,"/gc_stop_the_world:got lock\n"));
633     /* stop all other threads by sending them SIG_STOP_FOR_GC */
634     for(p=all_threads; p; p=p->next) {
635         gc_assert(p->os_thread != 0);
636         FSHOW_SIGNAL((stderr,"/gc_stop_the_world: thread=%lu, state=%x\n",
637                       p->os_thread, thread_state(p)));
638         if((p!=th) && ((thread_state(p)==STATE_RUNNING))) {
639             FSHOW_SIGNAL((stderr,"/gc_stop_the_world: suspending thread %lu\n",
640                           p->os_thread));
641             /* We already hold all_thread_lock, P can become DEAD but
642              * cannot exit, ergo it's safe to use pthread_kill. */
643             status=pthread_kill(p->os_thread,SIG_STOP_FOR_GC);
644             if (status==ESRCH) {
645                 /* This thread has exited. */
646                 gc_assert(thread_state(p)==STATE_DEAD);
647             } else if (status) {
648                 lose("cannot send suspend thread=%lu: %d, %s\n",
649                      p->os_thread,status,strerror(status));
650             }
651         }
652     }
653     FSHOW_SIGNAL((stderr,"/gc_stop_the_world:signals sent\n"));
654     for(p=all_threads;p;p=p->next) {
655         if (p!=th) {
656             FSHOW_SIGNAL
657                 ((stderr,
658                   "/gc_stop_the_world: waiting for thread=%lu: state=%x\n",
659                   p->os_thread, thread_state(p)));
660             wait_for_thread_state_change(p, STATE_RUNNING);
661             if (p->state == STATE_RUNNING)
662                 lose("/gc_stop_the_world: unexpected state");
663         }
664     }
665     FSHOW_SIGNAL((stderr,"/gc_stop_the_world:end\n"));
666 }
667
668 void gc_start_the_world()
669 {
670     struct thread *p,*th=arch_os_get_current_thread();
671     int lock_ret;
672     /* if a resumed thread creates a new thread before we're done with
673      * this loop, the new thread will get consed on the front of
674      * all_threads, but it won't have been stopped so won't need
675      * restarting */
676     FSHOW_SIGNAL((stderr,"/gc_start_the_world:begin\n"));
677     for(p=all_threads;p;p=p->next) {
678         gc_assert(p->os_thread!=0);
679         if (p!=th) {
680             lispobj state = thread_state(p);
681             if (state != STATE_DEAD) {
682                 if(state != STATE_SUSPENDED) {
683                     lose("gc_start_the_world: wrong thread state is %d\n",
684                          fixnum_value(state));
685                 }
686                 FSHOW_SIGNAL((stderr, "/gc_start_the_world: resuming %lu\n",
687                               p->os_thread));
688                 set_thread_state(p, STATE_RUNNING);
689             }
690         }
691     }
692
693     lock_ret = pthread_mutex_unlock(&all_threads_lock);
694     gc_assert(lock_ret == 0);
695 #ifdef LOCK_CREATE_THREAD
696     lock_ret = pthread_mutex_unlock(&create_thread_lock);
697     gc_assert(lock_ret == 0);
698 #endif
699
700     FSHOW_SIGNAL((stderr,"/gc_start_the_world:end\n"));
701 }
702 #endif
703
704 int
705 thread_yield()
706 {
707 #ifdef LISP_FEATURE_SB_THREAD
708     return sched_yield();
709 #else
710     return 0;
711 #endif
712 }
713
714 /* If the thread id given does not belong to a running thread (it has
715  * exited or never even existed) pthread_kill _may_ fail with ESRCH,
716  * but it is also allowed to just segfault, see
717  * <http://udrepper.livejournal.com/16844.html>.
718  *
719  * Relying on thread ids can easily backfire since ids are recycled
720  * (NPTL recycles them extremely fast) so a signal can be sent to
721  * another process if the one it was sent to exited.
722  *
723  * We send signals in two places: signal_interrupt_thread sends a
724  * signal that's harmless if delivered to another thread, but
725  * SIG_STOP_FOR_GC is fatal.
726  *
727  * For these reasons, we must make sure that the thread is still alive
728  * when the pthread_kill is called and return if the thread is
729  * exiting. */
730 int
731 kill_safely(os_thread_t os_thread, int signal)
732 {
733     FSHOW_SIGNAL((stderr,"/kill_safely: %lu, %d\n", os_thread, signal));
734     {
735 #ifdef LISP_FEATURE_SB_THREAD
736         sigset_t oldset;
737         struct thread *thread;
738         /* pthread_kill is not async signal safe and we don't want to be
739          * interrupted while holding the lock. */
740         block_deferrable_signals(0, &oldset);
741         pthread_mutex_lock(&all_threads_lock);
742         for (thread = all_threads; thread; thread = thread->next) {
743             if (thread->os_thread == os_thread) {
744                 int status = pthread_kill(os_thread, signal);
745                 if (status)
746                     lose("kill_safely: pthread_kill failed with %d\n", status);
747                 break;
748             }
749         }
750         pthread_mutex_unlock(&all_threads_lock);
751         thread_sigmask(SIG_SETMASK,&oldset,0);
752         if (thread)
753             return 0;
754         else
755             return -1;
756 #else
757         int status;
758         if (os_thread != 0)
759             lose("kill_safely: who do you want to kill? %d?\n", os_thread);
760         /* Dubious (as in don't know why it works) workaround for the
761          * signal sometimes not being generated on darwin. */
762 #ifdef LISP_FEATURE_DARWIN
763         {
764             sigset_t oldset;
765             sigprocmask(SIG_BLOCK, &deferrable_sigset, &oldset);
766             status = raise(signal);
767             sigprocmask(SIG_SETMASK,&oldset,0);
768         }
769 #else
770         status = raise(signal);
771 #endif
772         if (status == 0) {
773             return 0;
774         } else {
775             lose("cannot raise signal %d, %d %s\n",
776                  signal, status, strerror(errno));
777         }
778 #endif
779     }
780 }