#define LOCK_CREATE_THREAD
#endif
-#define ALIEN_STACK_SIZE (1*1024*1024) /* 1Mb size chosen at random */
-
#ifdef LISP_FEATURE_SB_THREAD
struct thread_post_mortem {
#ifdef DELAY_THREAD_POST_MORTEM
#endif
int dynamic_values_bytes=TLS_SIZE*sizeof(lispobj); /* same for all threads */
-struct thread * volatile all_threads;
+struct thread *all_threads;
extern struct interrupt_data * global_interrupt_data;
#ifdef LISP_FEATURE_SB_THREAD
link_thread(th);
th->os_thread=thread_self();
#ifndef LISP_FEATURE_WIN32
- protect_control_stack_guard_page(1);
+ protect_control_stack_guard_page(1, NULL);
+ protect_binding_stack_guard_page(1, NULL);
+ protect_alien_stack_guard_page(1, NULL);
#endif
#if defined(LISP_FEATURE_X86) || defined(LISP_FEATURE_X86_64)
#endif
}
+#ifdef LISP_FEATURE_SB_THREAD
+#define THREAD_STATE_LOCK_SIZE \
+ (sizeof(pthread_mutex_t))+(sizeof(pthread_cond_t))
+#else
+#define THREAD_STATE_LOCK_SIZE 0
+#endif
+
#define THREAD_STRUCT_SIZE (thread_control_stack_size + BINDING_STACK_SIZE + \
- ALIEN_STACK_SIZE + dynamic_values_bytes + \
+ ALIEN_STACK_SIZE + \
+ THREAD_STATE_LOCK_SIZE + \
+ dynamic_values_bytes + \
32 * SIGSTKSZ + \
THREAD_ALIGNMENT_BYTES)
int result, lock_ret;
FSHOW((stderr,"/creating thread %lu\n", thread_self()));
+ check_deferrables_blocked_or_lose(0);
+ check_gc_signals_unblocked_or_lose(0);
function = th->no_tls_value_marker;
th->no_tls_value_marker = NO_TLS_VALUE_MARKER_WIDETAG;
if(arch_os_thread_init(th)==0) {
}
th->os_thread=thread_self();
- protect_control_stack_guard_page(1);
+ protect_control_stack_guard_page(1, NULL);
+ protect_binding_stack_guard_page(1, NULL);
+ protect_alien_stack_guard_page(1, NULL);
/* Since GC can only know about this thread from the all_threads
- * list and we're just adding this thread to it there is no danger
- * of deadlocking even with SIG_STOP_FOR_GC blocked (which it is
- * not). */
+ * list and we're just adding this thread to it, there is no
+ * danger of deadlocking even with SIG_STOP_FOR_GC blocked (which
+ * it is not). */
lock_ret = pthread_mutex_lock(&all_threads_lock);
gc_assert(lock_ret == 0);
link_thread(th);
result = funcall0(function);
/* Block GC */
- block_blockable_signals();
- th->state=STATE_DEAD;
+ block_blockable_signals(0, 0);
+ set_thread_state(th, STATE_DEAD);
/* SIG_STOP_FOR_GC is blocked and GC might be waiting for this
* thread, but since we are already dead it won't wait long. */
gc_assert(lock_ret == 0);
if(th->tls_cookie>=0) arch_os_thread_cleanup(th);
+ pthread_mutex_destroy(th->state_lock);
+ pthread_cond_destroy(th->state_cond);
+
os_invalidate((os_vm_address_t)th->interrupt_data,
(sizeof (struct interrupt_data)));
(aligned_spaces+
thread_control_stack_size+
BINDING_STACK_SIZE+
- ALIEN_STACK_SIZE);
+ ALIEN_STACK_SIZE +
+ THREAD_STATE_LOCK_SIZE);
#ifdef LISP_FEATURE_SB_THREAD
for(i = 0; i < (dynamic_values_bytes / sizeof(lispobj)); i++)
th->os_thread=0;
#ifdef LISP_FEATURE_SB_THREAD
th->os_attr=malloc(sizeof(pthread_attr_t));
+ th->state_lock=(pthread_mutex_t *)((void *)th->alien_stack_start +
+ ALIEN_STACK_SIZE);
+ pthread_mutex_init(th->state_lock, NULL);
+ th->state_cond=(pthread_cond_t *)((void *)th->state_lock +
+ (sizeof(pthread_mutex_t)));
+ pthread_cond_init(th->state_cond, NULL);
#endif
th->state=STATE_RUNNING;
#ifdef LISP_FEATURE_STACK_GROWS_DOWNWARD_NOT_UPWARD
return 0;
}
th->interrupt_data->pending_handler = 0;
+ th->interrupt_data->gc_blocked_deferrables = 0;
th->no_tls_value_marker=initial_function;
th->stepping = NIL;
{
/* The new thread inherits the restrictive signal mask set here,
* and enables signals again when it is set up properly. */
- sigset_t newset,oldset;
+ sigset_t oldset;
boolean r=1;
int retcode = 0, initcode;
FSHOW_SIGNAL((stderr,"/create_os_thread: creating new thread\n"));
+ /* Blocking deferrable signals is enough, no need to block
+ * SIG_STOP_FOR_GC because the child process is not linked onto
+ * all_threads until it's ready. */
+ block_deferrable_signals(0, &oldset);
+
#ifdef LOCK_CREATE_THREAD
retcode = pthread_mutex_lock(&create_thread_lock);
gc_assert(retcode == 0);
FSHOW_SIGNAL((stderr,"/create_os_thread: got lock\n"));
#endif
- sigemptyset(&newset);
- /* Blocking deferrable signals is enough, no need to block
- * SIG_STOP_FOR_GC because the child process is not linked onto
- * all_threads until it's ready. */
- sigaddset_deferrable(&newset);
- thread_sigmask(SIG_BLOCK, &newset, &oldset);
if((initcode = pthread_attr_init(th->os_attr)) ||
- /* call_into_lisp_first_time switches the stack for the initial thread. For the
- * others, we use this. */
- (pthread_attr_setstack(th->os_attr,th->control_stack_start,thread_control_stack_size)) ||
+ /* call_into_lisp_first_time switches the stack for the initial
+ * thread. For the others, we use this. */
+ (pthread_attr_setstack(th->os_attr,th->control_stack_start,
+ thread_control_stack_size)) ||
(retcode = pthread_create
(kid_tid,th->os_attr,(void *(*)(void *))new_thread_trampoline,th))) {
FSHOW_SIGNAL((stderr, "init = %d\n", initcode));
- FSHOW_SIGNAL((stderr, printf("pthread_create returned %d, errno %d\n", retcode, errno)));
- FSHOW_SIGNAL((stderr, "wanted stack size %d, min stack size %d\n",
- cstack_size, PTHREAD_STACK_MIN));
+ FSHOW_SIGNAL((stderr, "pthread_create returned %d, errno %d\n",
+ retcode, errno));
if(retcode < 0) {
perror("create_os_thread");
}
r=0;
}
- thread_sigmask(SIG_SETMASK,&oldset,0);
#ifdef LOCK_CREATE_THREAD
retcode = pthread_mutex_unlock(&create_thread_lock);
gc_assert(retcode == 0);
FSHOW_SIGNAL((stderr,"/create_os_thread: released lock\n"));
#endif
+ thread_sigmask(SIG_SETMASK,&oldset,0);
return r;
}
os_thread_t create_thread(lispobj initial_function) {
- struct thread *th;
- os_thread_t kid_tid;
+ struct thread *th, *thread = arch_os_get_current_thread();
+ os_thread_t kid_tid = 0;
+
+ /* Must defend against async unwinds. */
+ if (SymbolValue(INTERRUPTS_ENABLED, thread) != NIL)
+ lose("create_thread is not safe when interrupts are enabled.\n");
/* Assuming that a fresh thread struct has no lisp objects in it,
* linking it to all_threads can be left to the thread itself
* without fear of gc lossage. initial_function violates this
* assumption and must stay pinned until the child starts up. */
th = create_thread_struct(initial_function);
- if(th==0) return 0;
-
- if (create_os_thread(th,&kid_tid)) {
- return kid_tid;
- } else {
+ if (th && !create_os_thread(th,&kid_tid)) {
free_thread_struct(th);
- return 0;
- }
-}
-
-/* Send the signo to os_thread, retry if the rt signal queue is
- * full. */
-int
-kill_thread_safely(os_thread_t os_thread, int signo)
-{
- int r;
- /* The man page does not mention EAGAIN as a valid return value
- * for either pthread_kill or kill. But that's theory, this is
- * practice. By waiting here we assume that the delivery of this
- * signal is not necessary for the delivery of the signals in the
- * queue. In other words, we _assume_ there are no deadlocks. */
- while ((r=pthread_kill(os_thread,signo))==EAGAIN) {
- /* wait a bit then try again in the hope of the rt signal
- * queue not being full */
- FSHOW_SIGNAL((stderr,"/rt signal queue full\n"));
- /* FIXME: some kind of backoff (random, exponential) would be
- * nice. */
- sleep(1);
- }
- return r;
-}
-
-int signal_interrupt_thread(os_thread_t os_thread)
-{
- int status = kill_thread_safely(os_thread, SIG_INTERRUPT_THREAD);
- if (status == 0) {
- return 0;
- } else if (status == ESRCH) {
- return -1;
- } else {
- lose("cannot send SIG_INTERRUPT_THREAD to thread=%lu: %d, %s\n",
- os_thread, status, strerror(status));
+ kid_tid = 0;
}
+ return kid_tid;
}
/* stopping the world is a two-stage process. From this thread we signal
#ifdef LOCK_CREATE_THREAD
/* KLUDGE: Stopping the thread during pthread_create() causes deadlock
* on FreeBSD. */
- FSHOW_SIGNAL((stderr,"/gc_stop_the_world:waiting on create_thread_lock, thread=%lu\n",
- th->os_thread));
+ FSHOW_SIGNAL((stderr,"/gc_stop_the_world:waiting on create_thread_lock\n"));
lock_ret = pthread_mutex_lock(&create_thread_lock);
gc_assert(lock_ret == 0);
- FSHOW_SIGNAL((stderr,"/gc_stop_the_world:got create_thread_lock, thread=%lu\n",
- th->os_thread));
+ FSHOW_SIGNAL((stderr,"/gc_stop_the_world:got create_thread_lock\n"));
#endif
- FSHOW_SIGNAL((stderr,"/gc_stop_the_world:waiting on lock, thread=%lu\n",
- th->os_thread));
+ FSHOW_SIGNAL((stderr,"/gc_stop_the_world:waiting on lock\n"));
/* keep threads from starting while the world is stopped. */
lock_ret = pthread_mutex_lock(&all_threads_lock); \
gc_assert(lock_ret == 0);
- FSHOW_SIGNAL((stderr,"/gc_stop_the_world:got lock, thread=%lu\n",
- th->os_thread));
+ FSHOW_SIGNAL((stderr,"/gc_stop_the_world:got lock\n"));
/* stop all other threads by sending them SIG_STOP_FOR_GC */
for(p=all_threads; p; p=p->next) {
gc_assert(p->os_thread != 0);
- FSHOW_SIGNAL((stderr,"/gc_stop_the_world: p->state: %x\n", p->state));
- if((p!=th) && ((p->state==STATE_RUNNING))) {
- FSHOW_SIGNAL((stderr,"/gc_stop_the_world: suspending %x, os_thread %x\n",
- p, p->os_thread));
- status=kill_thread_safely(p->os_thread,SIG_STOP_FOR_GC);
+ FSHOW_SIGNAL((stderr,"/gc_stop_the_world: thread=%lu, state=%x\n",
+ p->os_thread, thread_state(p)));
+ if((p!=th) && ((thread_state(p)==STATE_RUNNING))) {
+ FSHOW_SIGNAL((stderr,"/gc_stop_the_world: suspending thread %lu\n",
+ p->os_thread));
+ /* We already hold all_thread_lock, P can become DEAD but
+ * cannot exit, ergo it's safe to use pthread_kill. */
+ status=pthread_kill(p->os_thread,SIG_STOP_FOR_GC);
if (status==ESRCH) {
/* This thread has exited. */
- gc_assert(p->state==STATE_DEAD);
+ gc_assert(thread_state(p)==STATE_DEAD);
} else if (status) {
lose("cannot send suspend thread=%lu: %d, %s\n",
p->os_thread,status,strerror(status));
}
}
FSHOW_SIGNAL((stderr,"/gc_stop_the_world:signals sent\n"));
- /* wait for the running threads to stop or finish */
- for(p=all_threads;p;) {
- FSHOW_SIGNAL((stderr,"/gc_stop_the_world: th: %p, p: %p\n", th, p));
- if((p!=th) && (p->state==STATE_RUNNING)) {
- sched_yield();
- } else {
- p=p->next;
+ for(p=all_threads;p;p=p->next) {
+ if (p!=th) {
+ FSHOW_SIGNAL
+ ((stderr,
+ "/gc_stop_the_world: waiting for thread=%lu: state=%x\n",
+ p->os_thread, thread_state(p)));
+ wait_for_thread_state_change(p, STATE_RUNNING);
+ if (p->state == STATE_RUNNING)
+ lose("/gc_stop_the_world: unexpected state");
}
}
FSHOW_SIGNAL((stderr,"/gc_stop_the_world:end\n"));
void gc_start_the_world()
{
struct thread *p,*th=arch_os_get_current_thread();
- int status, lock_ret;
+ int lock_ret;
/* if a resumed thread creates a new thread before we're done with
* this loop, the new thread will get consed on the front of
* all_threads, but it won't have been stopped so won't need
FSHOW_SIGNAL((stderr,"/gc_start_the_world:begin\n"));
for(p=all_threads;p;p=p->next) {
gc_assert(p->os_thread!=0);
- if((p!=th) && (p->state!=STATE_DEAD)) {
- if(p->state!=STATE_SUSPENDED) {
- lose("gc_start_the_world: wrong thread state is %d\n",
- fixnum_value(p->state));
- }
- FSHOW_SIGNAL((stderr, "/gc_start_the_world: resuming %lu\n",
- p->os_thread));
- p->state=STATE_RUNNING;
-
-#if defined(SIG_RESUME_FROM_GC)
- status=kill_thread_safely(p->os_thread,SIG_RESUME_FROM_GC);
-#else
- status=kill_thread_safely(p->os_thread,SIG_STOP_FOR_GC);
-#endif
- if (status) {
- lose("cannot resume thread=%lu: %d, %s\n",
- p->os_thread,status,strerror(status));
+ if (p!=th) {
+ lispobj state = thread_state(p);
+ if (state != STATE_DEAD) {
+ if(state != STATE_SUSPENDED) {
+ lose("gc_start_the_world: wrong thread state is %d\n",
+ fixnum_value(state));
+ }
+ FSHOW_SIGNAL((stderr, "/gc_start_the_world: resuming %lu\n",
+ p->os_thread));
+ set_thread_state(p, STATE_RUNNING);
}
}
}
- /* If we waited here until all threads leave STATE_SUSPENDED, then
- * SIG_STOP_FOR_GC wouldn't need to be a rt signal. That has some
- * performance implications, but does away with the 'rt signal
- * queue full' problem. */
lock_ret = pthread_mutex_unlock(&all_threads_lock);
gc_assert(lock_ret == 0);
return 0;
#endif
}
+
+/* If the thread id given does not belong to a running thread (it has
+ * exited or never even existed) pthread_kill _may_ fail with ESRCH,
+ * but it is also allowed to just segfault, see
+ * <http://udrepper.livejournal.com/16844.html>.
+ *
+ * Relying on thread ids can easily backfire since ids are recycled
+ * (NPTL recycles them extremely fast) so a signal can be sent to
+ * another process if the one it was sent to exited.
+ *
+ * We send signals in two places: signal_interrupt_thread sends a
+ * signal that's harmless if delivered to another thread, but
+ * SIG_STOP_FOR_GC is fatal.
+ *
+ * For these reasons, we must make sure that the thread is still alive
+ * when the pthread_kill is called and return if the thread is
+ * exiting. */
+int
+kill_safely(os_thread_t os_thread, int signal)
+{
+ FSHOW_SIGNAL((stderr,"/kill_safely: %lu, %d\n", os_thread, signal));
+ {
+#ifdef LISP_FEATURE_SB_THREAD
+ sigset_t oldset;
+ struct thread *thread;
+ /* pthread_kill is not async signal safe and we don't want to be
+ * interrupted while holding the lock. */
+ block_deferrable_signals(0, &oldset);
+ pthread_mutex_lock(&all_threads_lock);
+ for (thread = all_threads; thread; thread = thread->next) {
+ if (thread->os_thread == os_thread) {
+ int status = pthread_kill(os_thread, signal);
+ if (status)
+ lose("kill_safely: pthread_kill failed with %d\n", status);
+ break;
+ }
+ }
+ pthread_mutex_unlock(&all_threads_lock);
+ thread_sigmask(SIG_SETMASK,&oldset,0);
+ if (thread)
+ return 0;
+ else
+ return -1;
+#else
+ int status;
+ if (os_thread != 0)
+ lose("kill_safely: who do you want to kill? %d?\n", os_thread);
+ /* Dubious (as in don't know why it works) workaround for the
+ * signal sometimes not being generated on darwin. */
+#ifdef LISP_FEATURE_DARWIN
+ {
+ sigset_t oldset;
+ sigprocmask(SIG_BLOCK, &deferrable_sigset, &oldset);
+ status = raise(signal);
+ sigprocmask(SIG_SETMASK,&oldset,0);
+ }
+#else
+ status = raise(signal);
+#endif
+ if (status == 0) {
+ return 0;
+ } else {
+ lose("cannot raise signal %d, %d %s\n",
+ signal, status, strerror(errno));
+ }
+#endif
+ }
+}