0.9.2.18: various error &co reporting improvements and build tweaks
[sbcl.git] / 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 int dynamic_values_bytes=4096*sizeof(lispobj);  /* same for all threads */
29 struct thread *all_threads;
30 volatile lispobj all_threads_lock;
31 extern struct interrupt_data * global_interrupt_data;
32 extern int linux_no_threads_p;
33
34 #ifdef LISP_FEATURE_SB_THREAD
35 /* When trying to get all_threads_lock one should make sure that
36  * sig_stop_for_gc is not blocked. Else there would be a possible
37  * deadlock: gc locks it, other thread blocks signals, gc sends stop
38  * request to other thread and waits, other thread blocks on lock. */
39 void check_sig_stop_for_gc_can_arrive_or_lose()
40 {
41     /* Get the current sigmask, by blocking the empty set. */
42     sigset_t empty,current;
43     sigemptyset(&empty);
44     thread_sigmask(SIG_BLOCK, &empty, &current);
45     if (sigismember(&current,SIG_STOP_FOR_GC))
46         lose("SIG_STOP_FOR_GC is blocked\n");
47     if (SymbolValue(INTERRUPTS_ENABLED,arch_os_get_current_thread()) == NIL)
48         lose("interrupts disabled\n");
49     if (arch_pseudo_atomic_atomic(NULL))
50         lose("n pseudo atomic\n");
51 }
52
53 #ifdef QSHOW_SIGNALS
54 #define FSHOW_SIGNAL FSHOW
55 #else
56 #define FSHOW_SIGNAL(args)
57 #endif
58
59 #define GET_ALL_THREADS_LOCK(name) \
60     { \
61         sigset_t _newset,_oldset; \
62         sigemptyset(&_newset); \
63         sigaddset_blockable(&_newset); \
64         sigdelset(&_newset,SIG_STOP_FOR_GC); \
65         thread_sigmask(SIG_BLOCK, &_newset, &_oldset); \
66         check_sig_stop_for_gc_can_arrive_or_lose(); \
67         FSHOW_SIGNAL((stderr,"/%s:waiting on lock=%ld, thread=%ld\n",name, \
68                all_threads_lock,arch_os_get_current_thread()->os_thread)); \
69         get_spinlock(&all_threads_lock,(long)arch_os_get_current_thread()); \
70         FSHOW_SIGNAL((stderr,"/%s:got lock, thread=%ld\n", \
71                name,arch_os_get_current_thread()->os_thread));
72
73 #define RELEASE_ALL_THREADS_LOCK(name) \
74         FSHOW_SIGNAL((stderr,"/%s:released lock\n",name)); \
75         release_spinlock(&all_threads_lock); \
76         thread_sigmask(SIG_SETMASK,&_oldset,0); \
77     }
78 #endif
79
80 int
81 initial_thread_trampoline(struct thread *th)
82 {
83     lispobj function;
84 #if defined(LISP_FEATURE_X86) || defined(LISP_FEATURE_X86_64)
85     lispobj *args = NULL;
86 #endif
87
88     function = th->unbound_marker;
89     th->unbound_marker = UNBOUND_MARKER_WIDETAG;
90     if(arch_os_thread_init(th)==0) return 1;
91
92     if(th->os_thread < 1) lose("th->os_thread not set up right");
93     th->state=STATE_RUNNING;
94 #if defined(LISP_FEATURE_X86) || defined(LISP_FEATURE_X86_64)
95     return call_into_lisp_first_time(function,args,0);
96 #else
97     return funcall0(function);
98 #endif
99 }
100
101 #ifdef LISP_FEATURE_SB_THREAD
102
103 /* this is the first thing that runs in the child (which is why the
104  * silly calling convention).  Basically it calls the user's requested
105  * lisp function after doing arch_os_thread_init and whatever other
106  * bookkeeping needs to be done
107  */
108 int
109 new_thread_trampoline(struct thread *th)
110 {
111     lispobj function;
112     function = th->unbound_marker;
113     th->unbound_marker = UNBOUND_MARKER_WIDETAG;
114     if(arch_os_thread_init(th)==0) return 1;
115
116     /* wait here until our thread is linked into all_threads: see below */
117     while(th->os_thread<1) sched_yield();
118
119     th->state=STATE_RUNNING;
120     return funcall0(function);
121 }
122 #endif /* LISP_FEATURE_SB_THREAD */
123
124 /* this is called from any other thread to create the new one, and
125  * initialize all parts of it that can be initialized from another
126  * thread
127  */
128
129 struct thread * create_thread_struct(lispobj initial_function) {
130     union per_thread_data *per_thread;
131     struct thread *th=0;        /*  subdue gcc */
132     void *spaces=0;
133
134     /* may as well allocate all the spaces at once: it saves us from
135      * having to decide what to do if only some of the allocations
136      * succeed */
137     spaces=os_validate(0,
138                        THREAD_CONTROL_STACK_SIZE+
139                        BINDING_STACK_SIZE+
140                        ALIEN_STACK_SIZE+
141                        dynamic_values_bytes+
142                        32*SIGSTKSZ);
143     if(!spaces)
144          return NULL;
145     per_thread=(union per_thread_data *)
146         (spaces+
147          THREAD_CONTROL_STACK_SIZE+
148          BINDING_STACK_SIZE+
149          ALIEN_STACK_SIZE);
150
151     if(all_threads) {
152         memcpy(per_thread,arch_os_get_current_thread(),
153                dynamic_values_bytes);
154     } else {
155 #ifdef LISP_FEATURE_SB_THREAD
156         int i;
157         for(i=0;i<(dynamic_values_bytes/sizeof(lispobj));i++)
158             per_thread->dynamic_values[i]=UNBOUND_MARKER_WIDETAG;
159         if(SymbolValue(FREE_TLS_INDEX,0)==UNBOUND_MARKER_WIDETAG)
160             SetSymbolValue
161                 (FREE_TLS_INDEX,
162                  make_fixnum(MAX_INTERRUPTS+
163                              sizeof(struct thread)/sizeof(lispobj)),
164                  0);
165 #define STATIC_TLS_INIT(sym,field) \
166   ((struct symbol *)(sym-OTHER_POINTER_LOWTAG))->tls_index= \
167   make_fixnum(THREAD_SLOT_OFFSET_WORDS(field))
168
169         STATIC_TLS_INIT(BINDING_STACK_START,binding_stack_start);
170         STATIC_TLS_INIT(BINDING_STACK_POINTER,binding_stack_pointer);
171         STATIC_TLS_INIT(CONTROL_STACK_START,control_stack_start);
172         STATIC_TLS_INIT(CONTROL_STACK_END,control_stack_end);
173         STATIC_TLS_INIT(ALIEN_STACK,alien_stack_pointer);
174 #if defined(LISP_FEATURE_X86) || defined (LISP_FEATURE_X86_64)
175         STATIC_TLS_INIT(PSEUDO_ATOMIC_ATOMIC,pseudo_atomic_atomic);
176         STATIC_TLS_INIT(PSEUDO_ATOMIC_INTERRUPTED,pseudo_atomic_interrupted);
177 #endif
178 #undef STATIC_TLS_INIT
179 #endif
180     }
181
182     th=&per_thread->thread;
183     th->control_stack_start = spaces;
184     th->binding_stack_start=
185         (lispobj*)((void*)th->control_stack_start+THREAD_CONTROL_STACK_SIZE);
186     th->control_stack_end = th->binding_stack_start;
187     th->alien_stack_start=
188         (lispobj*)((void*)th->binding_stack_start+BINDING_STACK_SIZE);
189     th->binding_stack_pointer=th->binding_stack_start;
190     th->this=th;
191     th->os_thread=0;
192     th->interrupt_fun=NIL;
193     th->interrupt_fun_lock=0;
194     th->state=STATE_STARTING;
195 #ifdef LISP_FEATURE_STACK_GROWS_DOWNWARD_NOT_UPWARD
196     th->alien_stack_pointer=((void *)th->alien_stack_start
197                              + ALIEN_STACK_SIZE-N_WORD_BYTES);
198 #else
199     th->alien_stack_pointer=((void *)th->alien_stack_start);
200 #endif
201 #if defined(LISP_FEATURE_X86) || defined (LISP_FEATURE_X86_64)
202     th->pseudo_atomic_interrupted=0;
203     th->pseudo_atomic_atomic=0;
204 #endif
205 #ifdef LISP_FEATURE_GENCGC
206     gc_set_region_empty(&th->alloc_region);
207 #endif
208
209 #ifndef LISP_FEATURE_SB_THREAD
210     /* the tls-points-into-struct-thread trick is only good for threaded
211      * sbcl, because unithread sbcl doesn't have tls.  So, we copy the
212      * appropriate values from struct thread here, and make sure that
213      * we use the appropriate SymbolValue macros to access any of the
214      * variable quantities from the C runtime.  It's not quite OAOOM,
215      * it just feels like it */
216     SetSymbolValue(BINDING_STACK_START,(lispobj)th->binding_stack_start,th);
217     SetSymbolValue(CONTROL_STACK_START,(lispobj)th->control_stack_start,th);
218     SetSymbolValue(CONTROL_STACK_END,(lispobj)th->control_stack_end,th);
219 #if defined(LISP_FEATURE_X86) || defined (LISP_FEATURE_X86_64)
220     SetSymbolValue(BINDING_STACK_POINTER,(lispobj)th->binding_stack_pointer,th);
221     SetSymbolValue(ALIEN_STACK,(lispobj)th->alien_stack_pointer,th);
222     SetSymbolValue(PSEUDO_ATOMIC_ATOMIC,(lispobj)th->pseudo_atomic_atomic,th);
223     SetSymbolValue(PSEUDO_ATOMIC_INTERRUPTED,th->pseudo_atomic_interrupted,th);
224 #else
225     current_binding_stack_pointer=th->binding_stack_pointer;
226     current_control_stack_pointer=th->control_stack_start;
227 #endif
228 #endif
229     bind_variable(CURRENT_CATCH_BLOCK,make_fixnum(0),th);
230     bind_variable(CURRENT_UNWIND_PROTECT_BLOCK,make_fixnum(0),th);
231     bind_variable(FREE_INTERRUPT_CONTEXT_INDEX,make_fixnum(0),th);
232     bind_variable(INTERRUPT_PENDING, NIL,th);
233     bind_variable(INTERRUPTS_ENABLED,T,th);
234
235     th->interrupt_data = (struct interrupt_data *)
236         os_validate(0,(sizeof (struct interrupt_data)));
237     if(all_threads)
238         memcpy(th->interrupt_data,
239                arch_os_get_current_thread()->interrupt_data,
240                sizeof (struct interrupt_data));
241     else
242         memcpy(th->interrupt_data,global_interrupt_data,
243                sizeof (struct interrupt_data));
244
245     th->unbound_marker=initial_function;
246     return th;
247 }
248
249 void link_thread(struct thread *th,os_thread_t kid_tid)
250 {
251     if (all_threads) all_threads->prev=th;
252     th->next=all_threads;
253     th->prev=0;
254     all_threads=th;
255     /* note that th->os_thread is 0 at this time.  We rely on
256      * all_threads_lock to ensure that we don't have >1 thread with
257      * os_thread=0 on the list at once
258      */
259     protect_control_stack_guard_page(th,1);
260     /* child will not start until this is set */
261     th->os_thread=kid_tid;
262 }
263
264 void create_initial_thread(lispobj initial_function) {
265     struct thread *th=create_thread_struct(initial_function);
266     os_thread_t kid_tid=thread_self();
267     if(th && kid_tid>0) {
268         link_thread(th,kid_tid);
269         initial_thread_trampoline(all_threads); /* no return */
270     } else lose("can't create initial thread");
271 }
272
273 #ifdef LISP_FEATURE_SB_THREAD
274
275 boolean create_os_thread(struct thread *th,os_thread_t *kid_tid)
276 {
277     /* The new thread inherits the restrictive signal mask set here,
278      * and enables signals again when it is set up properly. */
279     pthread_attr_t attr;
280     sigset_t newset,oldset;
281     boolean r=1;
282     sigemptyset(&newset);
283     sigaddset_blockable(&newset);
284     thread_sigmask(SIG_BLOCK, &newset, &oldset);
285     
286     if((pthread_attr_init(&attr)) ||
287        (pthread_attr_setstack(&attr,th->control_stack_start,
288                               THREAD_CONTROL_STACK_SIZE-16)) ||
289        (pthread_create
290         (kid_tid,&attr,(void *(*)(void *))new_thread_trampoline,th)))
291         r=0;
292     thread_sigmask(SIG_SETMASK,&oldset,0);
293     return r;
294 }
295
296 struct thread *create_thread(lispobj initial_function) {
297     struct thread *th;
298     os_thread_t kid_tid=0;
299     boolean success;
300
301     if(linux_no_threads_p) return 0;
302
303     th=create_thread_struct(initial_function);
304     if(th==0) return 0;
305
306     /* we must not be interrupted here after a successful
307      * create_os_thread, because the kid will be waiting for its
308      * thread struct to be linked */
309     GET_ALL_THREADS_LOCK("create_thread")
310
311     success=create_os_thread(th,&kid_tid);
312     if (success)
313         link_thread(th,kid_tid);
314     else
315         os_invalidate((os_vm_address_t) th->control_stack_start,
316                       ((sizeof (lispobj))
317                        * (th->control_stack_end-th->control_stack_start)) +
318                       BINDING_STACK_SIZE+ALIEN_STACK_SIZE+dynamic_values_bytes+
319                       32*SIGSTKSZ);
320
321     RELEASE_ALL_THREADS_LOCK("create_thread")
322
323     if (success)
324         return th;
325     else
326         return 0;
327 }
328 #endif
329
330 #if defined LISP_FEATURE_SB_THREAD
331 /* This is not needed unless #+SB-THREAD, as there's a trivial null
332  * unithread definition. */
333
334 /* called from lisp from the thread object finalizer */
335 void reap_dead_thread(struct thread *th)
336 {
337     if(th->state!=STATE_DEAD)
338         lose("thread %lx is not joinable, state=%d\n",th,th->state);
339 #ifdef LISP_FEATURE_GENCGC
340     {
341         sigset_t newset,oldset;
342         sigemptyset(&newset);
343         sigaddset_blockable(&newset);
344         thread_sigmask(SIG_BLOCK, &newset, &oldset);
345         gc_alloc_update_page_tables(0, &th->alloc_region);
346         release_spinlock(&all_threads_lock);
347         thread_sigmask(SIG_SETMASK,&oldset,0);
348     }
349 #endif
350     GET_ALL_THREADS_LOCK("reap_dead_thread")
351     FSHOW((stderr,"/reap_dead_thread: reaping %ld\n",th->os_thread));
352     if(th->prev)
353         th->prev->next=th->next;
354     else all_threads=th->next;
355     if(th->next)
356         th->next->prev=th->prev;
357     RELEASE_ALL_THREADS_LOCK("reap_dead_thread")
358     if(th->tls_cookie>=0) arch_os_thread_cleanup(th);
359     gc_assert(pthread_join(th->os_thread,NULL)==0);
360     os_invalidate((os_vm_address_t) th->control_stack_start,
361                   ((sizeof (lispobj))
362                    * (th->control_stack_end-th->control_stack_start)) +
363                   BINDING_STACK_SIZE+ALIEN_STACK_SIZE+dynamic_values_bytes+
364                   32*SIGSTKSZ);
365 }
366
367 int interrupt_thread(struct thread *th, lispobj function)
368 {
369     /* A thread may also become dead after this test. */
370     if ((th->state != STATE_DEAD)) {
371         /* In clone_threads, if A and B both interrupt C at
372          * approximately the same time, it does not matter: the
373          * second signal will be masked until the handler has
374          * returned from the first one.  In pthreads though, we
375          * can't put the knowledge of what function to call into
376          * the siginfo, so we have to store it in the destination
377          * thread, and do it in such a way that A won't clobber
378          * B's interrupt.  Hence this stupid linked list.
379          *
380          * This does depend on SIG_INTERRUPT_THREAD being queued
381          * (as POSIX RT signals are): we need to keep
382          * interrupt_fun data for exactly as many signals as are
383          * going to be received by the destination thread.
384          */
385         struct cons *c=alloc_cons(function,NIL);
386         int kill_status;
387         /* interrupt_thread_handler locks this spinlock with
388          * interrupts blocked and it does so for the sake of
389          * arrange_return_to_lisp_function, so we must also block
390          * them. */
391         sigset_t newset,oldset;
392         sigemptyset(&newset);
393         sigaddset_blockable(&newset);
394         thread_sigmask(SIG_BLOCK, &newset, &oldset);
395         get_spinlock(&th->interrupt_fun_lock,
396                      (long)arch_os_get_current_thread());
397         kill_status=thread_kill(th->os_thread,SIG_INTERRUPT_THREAD);
398         if(kill_status==0) {
399             ((struct cons *)native_pointer(c))->cdr=th->interrupt_fun;
400             th->interrupt_fun=c;
401         }
402         release_spinlock(&th->interrupt_fun_lock);
403         thread_sigmask(SIG_SETMASK,&oldset,0);
404         return (kill_status ? -1 : 0);
405     }
406     errno=EPERM; return -1;
407 }
408
409 /* stopping the world is a two-stage process.  From this thread we signal
410  * all the others with SIG_STOP_FOR_GC.  The handler for this signal does
411  * the usual pseudo-atomic checks (we don't want to stop a thread while
412  * it's in the middle of allocation) then waits for another SIG_STOP_FOR_GC.
413  */
414
415 /* To avoid deadlocks when gc stops the world all clients of each
416  * mutex must enable or disable SIG_STOP_FOR_GC for the duration of
417  * holding the lock, but they must agree on which. */
418 void gc_stop_the_world()
419 {
420     struct thread *p,*th=arch_os_get_current_thread();
421     FSHOW_SIGNAL((stderr,"/gc_stop_the_world:waiting on lock, thread=%ld\n",
422                   th->os_thread));
423     /* keep threads from starting while the world is stopped. */
424     get_spinlock(&all_threads_lock,(long)th);
425     FSHOW_SIGNAL((stderr,"/gc_stop_the_world:got lock, thread=%ld\n",
426                   th->os_thread));
427     /* stop all other threads by sending them SIG_STOP_FOR_GC */
428     for(p=all_threads; p; p=p->next) {
429         while(p->state==STATE_STARTING) sched_yield();
430         if((p!=th) && (p->state==STATE_RUNNING)) {
431             FSHOW_SIGNAL((stderr,"/gc_stop_the_world:sending sig_stop to %ld\n",
432                           p->os_thread));
433             if(thread_kill(p->os_thread,SIG_STOP_FOR_GC)==-1) {
434                 /* we can't kill the thread; assume because it died
435                  * since we last checked */
436                 p->state=STATE_DEAD;
437                 FSHOW_SIGNAL((stderr,"/gc_stop_the_world:assuming %ld dead\n",
438                    p->os_thread));
439             }
440         }
441     }
442     FSHOW_SIGNAL((stderr,"/gc_stop_the_world:signals sent\n"));
443     /* wait for the running threads to stop or finish */
444     for(p=all_threads;p;) {
445         gc_assert(p->os_thread!=0);
446         gc_assert(p->state!=STATE_STARTING);
447         if((p==th) || (p->state==STATE_SUSPENDED) ||
448            (p->state==STATE_DEAD)) {
449             p=p->next;
450         }
451     }
452     FSHOW_SIGNAL((stderr,"/gc_stop_the_world:end\n"));
453 }
454
455 void gc_start_the_world()
456 {
457     struct thread *p,*th=arch_os_get_current_thread();
458     /* if a resumed thread creates a new thread before we're done with
459      * this loop, the new thread will get consed on the front of
460      * all_threads, but it won't have been stopped so won't need
461      * restarting */
462     FSHOW_SIGNAL((stderr,"/gc_start_the_world:begin\n"));
463     for(p=all_threads;p;p=p->next) {
464         gc_assert(p->os_thread!=0);
465         if((p!=th) && (p->state!=STATE_DEAD)) {
466             if(p->state!=STATE_SUSPENDED) {
467                 lose("gc_start_the_world: wrong thread state is %ld\n",
468                      fixnum_value(p->state));
469             }
470             thread_kill(p->os_thread,SIG_STOP_FOR_GC);
471         }
472     }
473     /* we must wait for all threads to leave stopped state else we
474      * risk signal accumulation and lose any meaning of
475      * thread->state */
476     for(p=all_threads;p;) {
477         if((p==th) || (p->state!=STATE_SUSPENDED)) {
478             p=p->next;
479         }
480     }
481     release_spinlock(&all_threads_lock);
482     FSHOW_SIGNAL((stderr,"/gc_start_the_world:end\n"));
483 }
484 #endif