source: trunk/gc6.8/pthread_stop_world.c

Last change on this file was 132, checked in by cinc, 19 years ago

Boehm-Demers-Weiser garbage collector. Single-threaded for OS/2.
File size: 16.0 KB
Line 
1#include "private/pthread_support.h"
2
3#if defined(GC_PTHREADS) && !defined(GC_SOLARIS_THREADS) \
4 && !defined(GC_WIN32_THREADS) && !defined(GC_DARWIN_THREADS)
5
6#include <signal.h>
7#include <semaphore.h>
8#include <errno.h>
9#include <unistd.h>
10#include <sys/time.h>
11#ifndef HPUX
12# include <sys/select.h>
13 /* Doesn't exist on HP/UX 11.11. */
14#endif
15
16#if DEBUG_THREADS
17
18#ifndef NSIG
19# if defined(MAXSIG)
20# define NSIG (MAXSIG+1)
21# elif defined(_NSIG)
22# define NSIG _NSIG
23# elif defined(__SIGRTMAX)
24# define NSIG (__SIGRTMAX+1)
25# else
26 --> please fix it
27# endif
28#endif
29
30void GC_print_sig_mask()
31{
32 sigset_t blocked;
33 int i;
34
35 if (pthread_sigmask(SIG_BLOCK, NULL, &blocked) != 0)
36 ABORT("pthread_sigmask");
37 GC_printf0("Blocked: ");
38 for (i = 1; i < NSIG; i++) {
39 if (sigismember(&blocked, i)) { GC_printf1("%ld ",(long) i); }
40 }
41 GC_printf0("\n");
42}
43
44#endif
45
46/* Remove the signals that we want to allow in thread stopping */
47/* handler from a set. */
48void GC_remove_allowed_signals(sigset_t *set)
49{
50# ifdef NO_SIGNALS
51 if (sigdelset(set, SIGINT) != 0
52 || sigdelset(set, SIGQUIT) != 0
53 || sigdelset(set, SIGABRT) != 0
54 || sigdelset(set, SIGTERM) != 0) {
55 ABORT("sigdelset() failed");
56 }
57# endif
58
59# ifdef MPROTECT_VDB
60 /* Handlers write to the thread structure, which is in the heap, */
61 /* and hence can trigger a protection fault. */
62 if (sigdelset(set, SIGSEGV) != 0
63# ifdef SIGBUS
64 || sigdelset(set, SIGBUS) != 0
65# endif
66 ) {
67 ABORT("sigdelset() failed");
68 }
69# endif
70}
71
72static sigset_t suspend_handler_mask;
73
74volatile sig_atomic_t GC_stop_count;
75 /* Incremented at the beginning of GC_stop_world. */
76
77volatile sig_atomic_t GC_world_is_stopped = FALSE;
78 /* FALSE ==> it is safe for threads to restart, i.e. */
79 /* they will see another suspend signal before they */
80 /* are expected to stop (unless they have voluntarily */
81 /* stopped). */
82
83void GC_brief_async_signal_safe_sleep()
84{
85 struct timeval tv;
86 tv.tv_sec = 0;
87 tv.tv_usec = 1000 * TIME_LIMIT / 2;
88 select(0, 0, 0, 0, &tv);
89}
90
91#ifdef GC_OSF1_THREADS
92 GC_bool GC_retry_signals = TRUE;
93#else
94 GC_bool GC_retry_signals = FALSE;
95#endif
96
97/*
98 * We use signals to stop threads during GC.
99 *
100 * Suspended threads wait in signal handler for SIG_THR_RESTART.
101 * That's more portable than semaphores or condition variables.
102 * (We do use sem_post from a signal handler, but that should be portable.)
103 *
104 * The thread suspension signal SIG_SUSPEND is now defined in gc_priv.h.
105 * Note that we can't just stop a thread; we need it to save its stack
106 * pointer(s) and acknowledge.
107 */
108
109#ifndef SIG_THR_RESTART
110# if defined(GC_HPUX_THREADS) || defined(GC_OSF1_THREADS) || defined(GC_NETBSD_THREADS)
111# ifdef _SIGRTMIN
112# define SIG_THR_RESTART _SIGRTMIN + 5
113# else
114# define SIG_THR_RESTART SIGRTMIN + 5
115# endif
116# else
117# define SIG_THR_RESTART SIGXCPU
118# endif
119#endif
120
121sem_t GC_suspend_ack_sem;
122
123#ifdef GC_NETBSD_THREADS
124# define GC_NETBSD_THREADS_WORKAROUND
125 /* It seems to be necessary to wait until threads have restarted. */
126 /* But it is unclear why that is the case. */
127 sem_t GC_restart_ack_sem;
128#endif
129
130void GC_suspend_handler_inner(ptr_t sig_arg);
131
132#if defined(IA64) || defined(HP_PA)
133extern void GC_with_callee_saves_pushed();
134
135void GC_suspend_handler(int sig)
136{
137 int old_errno = errno;
138 GC_with_callee_saves_pushed(GC_suspend_handler_inner, (ptr_t)(word)sig);
139 errno = old_errno;
140}
141
142#else
143/* We believe that in all other cases the full context is already */
144/* in the signal handler frame. */
145void GC_suspend_handler(int sig)
146{
147 int old_errno = errno;
148 GC_suspend_handler_inner((ptr_t)(word)sig);
149 errno = old_errno;
150}
151#endif
152
153void GC_suspend_handler_inner(ptr_t sig_arg)
154{
155 int sig = (int)(word)sig_arg;
156 int dummy;
157 pthread_t my_thread = pthread_self();
158 GC_thread me;
159# ifdef PARALLEL_MARK
160 word my_mark_no = GC_mark_no;
161 /* Marker can't proceed until we acknowledge. Thus this is */
162 /* guaranteed to be the mark_no correspending to our */
163 /* suspension, i.e. the marker can't have incremented it yet. */
164# endif
165 word my_stop_count = GC_stop_count;
166
167 if (sig != SIG_SUSPEND) ABORT("Bad signal in suspend_handler");
168
169#if DEBUG_THREADS
170 GC_printf1("Suspending 0x%lx\n", my_thread);
171#endif
172
173 me = GC_lookup_thread(my_thread);
174 /* The lookup here is safe, since I'm doing this on behalf */
175 /* of a thread which holds the allocation lock in order */
176 /* to stop the world. Thus concurrent modification of the */
177 /* data structure is impossible. */
178 if (me -> stop_info.last_stop_count == my_stop_count) {
179 /* Duplicate signal. OK if we are retrying. */
180 if (!GC_retry_signals) {
181 WARN("Duplicate suspend signal in thread %lx\n",
182 pthread_self());
183 }
184 return;
185 }
186# ifdef SPARC
187 me -> stop_info.stack_ptr = (ptr_t)GC_save_regs_in_stack();
188# else
189 me -> stop_info.stack_ptr = (ptr_t)(&dummy);
190# endif
191# ifdef IA64
192 me -> backing_store_ptr = (ptr_t)GC_save_regs_in_stack();
193# endif
194
195 /* Tell the thread that wants to stop the world that this */
196 /* thread has been stopped. Note that sem_post() is */
197 /* the only async-signal-safe primitive in LinuxThreads. */
198 sem_post(&GC_suspend_ack_sem);
199 me -> stop_info.last_stop_count = my_stop_count;
200
201 /* Wait until that thread tells us to restart by sending */
202 /* this thread a SIG_THR_RESTART signal. */
203 /* SIG_THR_RESTART should be masked at this point. Thus there */
204 /* is no race. */
205 /* We do not continue until we receive a SIG_THR_RESTART, */
206 /* but we do not take that as authoritative. (We may be */
207 /* accidentally restarted by one of the user signals we */
208 /* don't block.) After we receive the signal, we use a */
209 /* primitive and expensive mechanism to wait until it's */
210 /* really safe to proceed. Under normal circumstances, */
211 /* this code should not be executed. */
212 sigsuspend(&suspend_handler_mask); /* Wait for signal */
213 while (GC_world_is_stopped && GC_stop_count == my_stop_count) {
214 GC_brief_async_signal_safe_sleep();
215# if DEBUG_THREADS
216 GC_err_printf0("Sleeping in signal handler");
217# endif
218 }
219 /* If the RESTART signal gets lost, we can still lose. That should be */
220 /* less likely than losing the SUSPEND signal, since we don't do much */
221 /* between the sem_post and sigsuspend. */
222 /* We'd need more handshaking to work around that. */
223 /* Simply dropping the sigsuspend call should be safe, but is unlikely */
224 /* to be efficient. */
225
226#if DEBUG_THREADS
227 GC_printf1("Continuing 0x%lx\n", my_thread);
228#endif
229}
230
231void GC_restart_handler(int sig)
232{
233 pthread_t my_thread = pthread_self();
234
235 if (sig != SIG_THR_RESTART) ABORT("Bad signal in suspend_handler");
236
237#ifdef GC_NETBSD_THREADS_WORKAROUND
238 sem_post(&GC_restart_ack_sem);
239#endif
240
241 /*
242 ** Note: even if we don't do anything useful here,
243 ** it would still be necessary to have a signal handler,
244 ** rather than ignoring the signals, otherwise
245 ** the signals will not be delivered at all, and
246 ** will thus not interrupt the sigsuspend() above.
247 */
248
249#if DEBUG_THREADS
250 GC_printf1("In GC_restart_handler for 0x%lx\n", pthread_self());
251#endif
252}
253
254# ifdef IA64
255# define IF_IA64(x) x
256# else
257# define IF_IA64(x)
258# endif
259/* We hold allocation lock. Should do exactly the right thing if the */
260/* world is stopped. Should not fail if it isn't. */
261void GC_push_all_stacks()
262{
263 GC_bool found_me = FALSE;
264 int i;
265 GC_thread p;
266 ptr_t lo, hi;
267 /* On IA64, we also need to scan the register backing store. */
268 IF_IA64(ptr_t bs_lo; ptr_t bs_hi;)
269 pthread_t me = pthread_self();
270
271 if (!GC_thr_initialized) GC_thr_init();
272 #if DEBUG_THREADS
273 GC_printf1("Pushing stacks from thread 0x%lx\n", (unsigned long) me);
274 #endif
275 for (i = 0; i < THREAD_TABLE_SZ; i++) {
276 for (p = GC_threads[i]; p != 0; p = p -> next) {
277 if (p -> flags & FINISHED) continue;
278 if (pthread_equal(p -> id, me)) {
279# ifdef SPARC
280 lo = (ptr_t)GC_save_regs_in_stack();
281# else
282 lo = GC_approx_sp();
283# endif
284 found_me = TRUE;
285 IF_IA64(bs_hi = (ptr_t)GC_save_regs_in_stack();)
286 } else {
287 lo = p -> stop_info.stack_ptr;
288 IF_IA64(bs_hi = p -> backing_store_ptr;)
289 }
290 if ((p -> flags & MAIN_THREAD) == 0) {
291 hi = p -> stack_end;
292 IF_IA64(bs_lo = p -> backing_store_end);
293 } else {
294 /* The original stack. */
295 hi = GC_stackbottom;
296 IF_IA64(bs_lo = BACKING_STORE_BASE;)
297 }
298 #if DEBUG_THREADS
299 GC_printf3("Stack for thread 0x%lx = [%lx,%lx)\n",
300 (unsigned long) p -> id,
301 (unsigned long) lo, (unsigned long) hi);
302 #endif
303 if (0 == lo) ABORT("GC_push_all_stacks: sp not set!\n");
304# ifdef STACK_GROWS_UP
305 /* We got them backwards! */
306 GC_push_all_stack(hi, lo);
307# else
308 GC_push_all_stack(lo, hi);
309# endif
310# ifdef IA64
311# if DEBUG_THREADS
312 GC_printf3("Reg stack for thread 0x%lx = [%lx,%lx)\n",
313 (unsigned long) p -> id,
314 (unsigned long) bs_lo, (unsigned long) bs_hi);
315# endif
316 if (pthread_equal(p -> id, me)) {
317 /* FIXME: This may add an unbounded number of entries, */
318 /* and hence overflow the mark stack, which is bad. */
319 GC_push_all_eager(bs_lo, bs_hi);
320 } else {
321 GC_push_all_stack(bs_lo, bs_hi);
322 }
323# endif
324 }
325 }
326 if (!found_me && !GC_in_thread_creation)
327 ABORT("Collecting from unknown thread.");
328}
329
330/* There seems to be a very rare thread stopping problem. To help us */
331/* debug that, we save the ids of the stopping thread. */
332pthread_t GC_stopping_thread;
333int GC_stopping_pid;
334
335/* We hold the allocation lock. Suspend all threads that might */
336/* still be running. Return the number of suspend signals that */
337/* were sent. */
338int GC_suspend_all()
339{
340 int n_live_threads = 0;
341 int i;
342 GC_thread p;
343 int result;
344 pthread_t my_thread = pthread_self();
345
346 GC_stopping_thread = my_thread; /* debugging only. */
347 GC_stopping_pid = getpid(); /* debugging only. */
348 for (i = 0; i < THREAD_TABLE_SZ; i++) {
349 for (p = GC_threads[i]; p != 0; p = p -> next) {
350 if (p -> id != my_thread) {
351 if (p -> flags & FINISHED) continue;
352 if (p -> stop_info.last_stop_count == GC_stop_count) continue;
353 if (p -> thread_blocked) /* Will wait */ continue;
354 n_live_threads++;
355 #if DEBUG_THREADS
356 GC_printf1("Sending suspend signal to 0x%lx\n", p -> id);
357 #endif
358
359 result = pthread_kill(p -> id, SIG_SUSPEND);
360 switch(result) {
361 case ESRCH:
362 /* Not really there anymore. Possible? */
363 n_live_threads--;
364 break;
365 case 0:
366 break;
367 default:
368 ABORT("pthread_kill failed");
369 }
370 }
371 }
372 }
373 return n_live_threads;
374}
375
376/* Caller holds allocation lock. */
377void GC_stop_world()
378{
379 int i;
380 int n_live_threads;
381 int code;
382
383 #if DEBUG_THREADS
384 GC_printf1("Stopping the world from 0x%lx\n", pthread_self());
385 #endif
386
387 /* Make sure all free list construction has stopped before we start. */
388 /* No new construction can start, since free list construction is */
389 /* required to acquire and release the GC lock before it starts, */
390 /* and we have the lock. */
391# ifdef PARALLEL_MARK
392 GC_acquire_mark_lock();
393 GC_ASSERT(GC_fl_builder_count == 0);
394 /* We should have previously waited for it to become zero. */
395# endif /* PARALLEL_MARK */
396 ++GC_stop_count;
397 GC_world_is_stopped = TRUE;
398 n_live_threads = GC_suspend_all();
399
400 if (GC_retry_signals) {
401 unsigned long wait_usecs = 0; /* Total wait since retry. */
402# define WAIT_UNIT 3000
403# define RETRY_INTERVAL 100000
404 for (;;) {
405 int ack_count;
406
407 sem_getvalue(&GC_suspend_ack_sem, &ack_count);
408 if (ack_count == n_live_threads) break;
409 if (wait_usecs > RETRY_INTERVAL) {
410 int newly_sent = GC_suspend_all();
411
412# ifdef CONDPRINT
413 if (GC_print_stats) {
414 GC_printf1("Resent %ld signals after timeout\n",
415 newly_sent);
416 }
417# endif
418 sem_getvalue(&GC_suspend_ack_sem, &ack_count);
419 if (newly_sent < n_live_threads - ack_count) {
420 WARN("Lost some threads during GC_stop_world?!\n",0);
421 n_live_threads = ack_count + newly_sent;
422 }
423 wait_usecs = 0;
424 }
425 usleep(WAIT_UNIT);
426 wait_usecs += WAIT_UNIT;
427 }
428 }
429 for (i = 0; i < n_live_threads; i++) {
430 while (0 != (code = sem_wait(&GC_suspend_ack_sem))) {
431 if (errno != EINTR) {
432 GC_err_printf1("Sem_wait returned %ld\n", (unsigned long)code);
433 ABORT("sem_wait for handler failed");
434 }
435 }
436 }
437# ifdef PARALLEL_MARK
438 GC_release_mark_lock();
439# endif
440 #if DEBUG_THREADS
441 GC_printf1("World stopped from 0x%lx\n", pthread_self());
442 #endif
443 GC_stopping_thread = 0; /* debugging only */
444}
445
446/* Caller holds allocation lock, and has held it continuously since */
447/* the world stopped. */
448void GC_start_world()
449{
450 pthread_t my_thread = pthread_self();
451 register int i;
452 register GC_thread p;
453 register int n_live_threads = 0;
454 register int result;
455#ifdef GC_NETBSD_THREADS_WORKAROUND
456 int code;
457#endif
458
459# if DEBUG_THREADS
460 GC_printf0("World starting\n");
461# endif
462
463 GC_world_is_stopped = FALSE;
464 for (i = 0; i < THREAD_TABLE_SZ; i++) {
465 for (p = GC_threads[i]; p != 0; p = p -> next) {
466 if (p -> id != my_thread) {
467 if (p -> flags & FINISHED) continue;
468 if (p -> thread_blocked) continue;
469 n_live_threads++;
470 #if DEBUG_THREADS
471 GC_printf1("Sending restart signal to 0x%lx\n", p -> id);
472 #endif
473 result = pthread_kill(p -> id, SIG_THR_RESTART);
474 switch(result) {
475 case ESRCH:
476 /* Not really there anymore. Possible? */
477 n_live_threads--;
478 break;
479 case 0:
480 break;
481 default:
482 ABORT("pthread_kill failed");
483 }
484 }
485 }
486 }
487#ifdef GC_NETBSD_THREADS_WORKAROUND
488 for (i = 0; i < n_live_threads; i++)
489 while (0 != (code = sem_wait(&GC_restart_ack_sem)))
490 if (errno != EINTR) {
491 GC_err_printf1("sem_wait() returned %ld\n", (unsigned long)code);
492 ABORT("sem_wait() for restart handler failed");
493 }
494#endif
495 #if DEBUG_THREADS
496 GC_printf0("World started\n");
497 #endif
498}
499
500void GC_stop_init() {
501 struct sigaction act;
502
503 if (sem_init(&GC_suspend_ack_sem, 0, 0) != 0)
504 ABORT("sem_init failed");
505#ifdef GC_NETBSD_THREADS_WORKAROUND
506 if (sem_init(&GC_restart_ack_sem, 0, 0) != 0)
507 ABORT("sem_init failed");
508#endif
509
510 act.sa_flags = SA_RESTART;
511 if (sigfillset(&act.sa_mask) != 0) {
512 ABORT("sigfillset() failed");
513 }
514 GC_remove_allowed_signals(&act.sa_mask);
515 /* SIG_THR_RESTART is set in the resulting mask. */
516 /* It is unmasked by the handler when necessary. */
517 act.sa_handler = GC_suspend_handler;
518 if (sigaction(SIG_SUSPEND, &act, NULL) != 0) {
519 ABORT("Cannot set SIG_SUSPEND handler");
520 }
521
522 act.sa_handler = GC_restart_handler;
523 if (sigaction(SIG_THR_RESTART, &act, NULL) != 0) {
524 ABORT("Cannot set SIG_THR_RESTART handler");
525 }
526
527 /* Inititialize suspend_handler_mask. It excludes SIG_THR_RESTART. */
528 if (sigfillset(&suspend_handler_mask) != 0) ABORT("sigfillset() failed");
529 GC_remove_allowed_signals(&suspend_handler_mask);
530 if (sigdelset(&suspend_handler_mask, SIG_THR_RESTART) != 0)
531 ABORT("sigdelset() failed");
532
533 /* Check for GC_RETRY_SIGNALS. */
534 if (0 != GETENV("GC_RETRY_SIGNALS")) {
535 GC_retry_signals = TRUE;
536 }
537 if (0 != GETENV("GC_NO_RETRY_SIGNALS")) {
538 GC_retry_signals = FALSE;
539 }
540# ifdef CONDPRINT
541 if (GC_print_stats && GC_retry_signals) {
542 GC_printf0("Will retry suspend signal if necessary.\n");
543 }
544# endif
545}
546
547#endif
Note: See TracBrowser for help on using the repository browser.