#define ALIEN_STACK_SIZE (1*1024*1024) /* 1Mb size chosen at random */
+struct freeable_stack {
+ os_thread_t os_thread;
+ os_vm_address_t stack;
+};
+
+static struct freeable_stack * volatile freeable_stack = 0;
+
int dynamic_values_bytes=4096*sizeof(lispobj); /* same for all threads */
-struct thread *all_threads;
-volatile lispobj all_threads_lock;
+struct thread * volatile all_threads;
extern struct interrupt_data * global_interrupt_data;
extern int linux_no_threads_p;
#ifdef LISP_FEATURE_SB_THREAD
-/* When trying to get all_threads_lock one should make sure that
- * sig_stop_for_gc is not blocked. Else there would be a possible
- * deadlock: gc locks it, other thread blocks signals, gc sends stop
- * request to other thread and waits, other thread blocks on lock. */
-void check_sig_stop_for_gc_can_arrive_or_lose()
-{
- /* Get the current sigmask, by blocking the empty set. */
- sigset_t empty,current;
- sigemptyset(&empty);
- thread_sigmask(SIG_BLOCK, &empty, ¤t);
- if (sigismember(¤t,SIG_STOP_FOR_GC))
- lose("SIG_STOP_FOR_GC is blocked\n");
- if (SymbolValue(INTERRUPTS_ENABLED,arch_os_get_current_thread()) == NIL)
- lose("interrupts disabled\n");
- if (arch_pseudo_atomic_atomic(NULL))
- lose("n pseudo atomic\n");
-}
-
-#ifdef QSHOW_SIGNALS
-#define FSHOW_SIGNAL FSHOW
-#else
-#define FSHOW_SIGNAL(args)
+pthread_mutex_t all_threads_lock = PTHREAD_MUTEX_INITIALIZER;
#endif
-#define GET_ALL_THREADS_LOCK(name) \
- { \
- sigset_t _newset,_oldset; \
- sigemptyset(&_newset); \
- sigaddset_blockable(&_newset); \
- sigdelset(&_newset,SIG_STOP_FOR_GC); \
- thread_sigmask(SIG_BLOCK, &_newset, &_oldset); \
- check_sig_stop_for_gc_can_arrive_or_lose(); \
- FSHOW_SIGNAL((stderr,"/%s:waiting on lock=%ld, thread=%ld\n",name, \
- all_threads_lock,arch_os_get_current_thread()->os_thread)); \
- get_spinlock(&all_threads_lock,(long)arch_os_get_current_thread()); \
- FSHOW_SIGNAL((stderr,"/%s:got lock, thread=%ld\n", \
- name,arch_os_get_current_thread()->os_thread));
-
-#define RELEASE_ALL_THREADS_LOCK(name) \
- FSHOW_SIGNAL((stderr,"/%s:released lock\n",name)); \
- release_spinlock(&all_threads_lock); \
- thread_sigmask(SIG_SETMASK,&_oldset,0); \
- }
+#if defined(LISP_FEATURE_X86) || defined(LISP_FEATURE_X86_64)
+extern lispobj call_into_lisp_first_time(lispobj fun, lispobj *args, int nargs);
#endif
+static void
+link_thread(struct thread *th)
+{
+ if (all_threads) all_threads->prev=th;
+ th->next=all_threads;
+ th->prev=0;
+ all_threads=th;
+}
-#if defined(LISP_FEATURE_X86) || defined(LISP_FEATURE_X86_64)
-extern lispobj call_into_lisp_first_time(lispobj fun, lispobj *args, int nargs);
+#ifdef LISP_FEATURE_SB_THREAD
+static void
+unlink_thread(struct thread *th)
+{
+ if (th->prev)
+ th->prev->next = th->next;
+ else
+ all_threads = th->next;
+ if (th->next)
+ th->next->prev = th->prev;
+}
#endif
-int
+static int
initial_thread_trampoline(struct thread *th)
{
lispobj function;
#if defined(LISP_FEATURE_X86) || defined(LISP_FEATURE_X86_64)
lispobj *args = NULL;
#endif
- function = th->unbound_marker;
- th->unbound_marker = UNBOUND_MARKER_WIDETAG;
+ function = th->no_tls_value_marker;
+ th->no_tls_value_marker = NO_TLS_VALUE_MARKER_WIDETAG;
if(arch_os_thread_init(th)==0) return 1;
+ link_thread(th);
+ th->os_thread=thread_self();
+ protect_control_stack_guard_page(1);
- if(th->os_thread < 1) lose("th->os_thread not set up right");
- th->state=STATE_RUNNING;
#if defined(LISP_FEATURE_X86) || defined(LISP_FEATURE_X86_64)
return call_into_lisp_first_time(function,args,0);
#else
#endif
}
+#define THREAD_STRUCT_SIZE (THREAD_CONTROL_STACK_SIZE + BINDING_STACK_SIZE + \
+ ALIEN_STACK_SIZE + dynamic_values_bytes + \
+ 32 * SIGSTKSZ)
+
#ifdef LISP_FEATURE_SB_THREAD
+static void
+free_thread_stack_later(struct thread *thread_to_be_cleaned_up)
+{
+ struct freeable_stack *new_freeable_stack = 0;
+ if (thread_to_be_cleaned_up) {
+ new_freeable_stack = (struct freeable_stack *)
+ os_validate(0, sizeof(struct freeable_stack));
+ new_freeable_stack->os_thread = thread_to_be_cleaned_up->os_thread;
+ new_freeable_stack->stack = (os_vm_address_t)
+ thread_to_be_cleaned_up->control_stack_start;
+ }
+ new_freeable_stack = (struct freeable_stack *)
+ swap_lispobjs((lispobj *)(void *)&freeable_stack,
+ (lispobj)new_freeable_stack);
+ if (new_freeable_stack) {
+ FSHOW((stderr,"/reaping %lu\n", new_freeable_stack->os_thread));
+ /* Under NPTL pthread_join really waits until the thread
+ * exists and the stack can be safely freed. This is sadly not
+ * mandated by the pthread spec. */
+ gc_assert(pthread_join(new_freeable_stack->os_thread, NULL) == 0);
+ os_invalidate(new_freeable_stack->stack, THREAD_STRUCT_SIZE);
+ os_invalidate((os_vm_address_t) new_freeable_stack,
+ sizeof(struct freeable_stack));
+ }
+}
+
/* this is the first thing that runs in the child (which is why the
* silly calling convention). Basically it calls the user's requested
* lisp function after doing arch_os_thread_init and whatever other
new_thread_trampoline(struct thread *th)
{
lispobj function;
- function = th->unbound_marker;
- th->unbound_marker = UNBOUND_MARKER_WIDETAG;
- if(arch_os_thread_init(th)==0) return 1;
+ int result;
+ FSHOW((stderr,"/creating thread %lu\n", thread_self()));
+ function = th->no_tls_value_marker;
+ th->no_tls_value_marker = NO_TLS_VALUE_MARKER_WIDETAG;
+ if(arch_os_thread_init(th)==0) {
+ /* FIXME: handle error */
+ lose("arch_os_thread_init failed\n");
+ }
- /* wait here until our thread is linked into all_threads: see below */
- while(th->os_thread<1) sched_yield();
+ th->os_thread=thread_self();
+ protect_control_stack_guard_page(1);
+ /* 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). */
+ pthread_mutex_lock(&all_threads_lock);
+ link_thread(th);
+ pthread_mutex_unlock(&all_threads_lock);
+
+ result = funcall0(function);
+ th->state=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. */
+ pthread_mutex_lock(&all_threads_lock);
+ gc_alloc_update_page_tables(0, &th->alloc_region);
+ unlink_thread(th);
+ pthread_mutex_unlock(&all_threads_lock);
- th->state=STATE_RUNNING;
- return funcall0(function);
+ if(th->tls_cookie>=0) arch_os_thread_cleanup(th);
+ os_invalidate((os_vm_address_t)th->interrupt_data,
+ (sizeof (struct interrupt_data)));
+ free_thread_stack_later(th);
+ FSHOW((stderr,"/exiting thread %lu\n", thread_self()));
+ return result;
}
+
#endif /* LISP_FEATURE_SB_THREAD */
+static void
+free_thread_struct(struct thread *th)
+{
+ if (th->interrupt_data)
+ os_invalidate((os_vm_address_t) th->interrupt_data,
+ (sizeof (struct interrupt_data)));
+ os_invalidate((os_vm_address_t) th->control_stack_start,
+ THREAD_STRUCT_SIZE);
+}
+
/* this is called from any other thread to create the new one, and
* initialize all parts of it that can be initialized from another
* thread
*/
-struct thread * create_thread_struct(lispobj initial_function) {
+static struct thread *
+create_thread_struct(lispobj initial_function) {
union per_thread_data *per_thread;
struct thread *th=0; /* subdue gcc */
void *spaces=0;
+#ifdef LISP_FEATURE_SB_THREAD
+ int i;
+#endif
/* may as well allocate all the spaces at once: it saves us from
* having to decide what to do if only some of the allocations
* succeed */
- spaces=os_validate(0,
- THREAD_CONTROL_STACK_SIZE+
- BINDING_STACK_SIZE+
- ALIEN_STACK_SIZE+
- dynamic_values_bytes+
- 32*SIGSTKSZ);
+ spaces=os_validate(0, THREAD_STRUCT_SIZE);
if(!spaces)
return NULL;
per_thread=(union per_thread_data *)
BINDING_STACK_SIZE+
ALIEN_STACK_SIZE);
- if(all_threads) {
- memcpy(per_thread,arch_os_get_current_thread(),
- dynamic_values_bytes);
- } else {
#ifdef LISP_FEATURE_SB_THREAD
- int i;
- for(i=0;i<(dynamic_values_bytes/sizeof(lispobj));i++)
- per_thread->dynamic_values[i]=UNBOUND_MARKER_WIDETAG;
- if(SymbolValue(FREE_TLS_INDEX,0)==UNBOUND_MARKER_WIDETAG)
+ for(i = 0; i < (dynamic_values_bytes / sizeof(lispobj)); i++)
+ per_thread->dynamic_values[i] = NO_TLS_VALUE_MARKER_WIDETAG;
+ if (all_threads == 0) {
+ if(SymbolValue(FREE_TLS_INDEX,0)==UNBOUND_MARKER_WIDETAG) {
SetSymbolValue
(FREE_TLS_INDEX,
+ /* FIXME: should be MAX_INTERRUPTS -1 ? */
make_fixnum(MAX_INTERRUPTS+
sizeof(struct thread)/sizeof(lispobj)),
0);
+ SetSymbolValue(TLS_INDEX_LOCK,make_fixnum(0),0);
+ }
#define STATIC_TLS_INIT(sym,field) \
((struct symbol *)(sym-OTHER_POINTER_LOWTAG))->tls_index= \
make_fixnum(THREAD_SLOT_OFFSET_WORDS(field))
STATIC_TLS_INIT(PSEUDO_ATOMIC_INTERRUPTED,pseudo_atomic_interrupted);
#endif
#undef STATIC_TLS_INIT
-#endif
}
+#endif
th=&per_thread->thread;
th->control_stack_start = spaces;
th->binding_stack_pointer=th->binding_stack_start;
th->this=th;
th->os_thread=0;
- th->interrupt_fun=NIL;
- th->interrupt_fun_lock=0;
- th->state=STATE_STARTING;
+ th->state=STATE_RUNNING;
#ifdef LISP_FEATURE_STACK_GROWS_DOWNWARD_NOT_UPWARD
th->alien_stack_pointer=((void *)th->alien_stack_start
+ ALIEN_STACK_SIZE-N_WORD_BYTES);
bind_variable(FREE_INTERRUPT_CONTEXT_INDEX,make_fixnum(0),th);
bind_variable(INTERRUPT_PENDING, NIL,th);
bind_variable(INTERRUPTS_ENABLED,T,th);
+ bind_variable(GC_PENDING,NIL,th);
+#ifdef LISP_FEATURE_SB_THREAD
+ bind_variable(STOP_FOR_GC_PENDING,NIL,th);
+#endif
th->interrupt_data = (struct interrupt_data *)
os_validate(0,(sizeof (struct interrupt_data)));
- if(all_threads)
- memcpy(th->interrupt_data,
- arch_os_get_current_thread()->interrupt_data,
- sizeof (struct interrupt_data));
- else
- memcpy(th->interrupt_data,global_interrupt_data,
- sizeof (struct interrupt_data));
-
- th->unbound_marker=initial_function;
+ if (!th->interrupt_data) {
+ free_thread_struct(th);
+ return 0;
+ }
+ th->interrupt_data->pending_handler = 0;
+ th->no_tls_value_marker=initial_function;
return th;
}
-void link_thread(struct thread *th,os_thread_t kid_tid)
-{
- if (all_threads) all_threads->prev=th;
- th->next=all_threads;
- th->prev=0;
- all_threads=th;
- /* note that th->os_thread is 0 at this time. We rely on
- * all_threads_lock to ensure that we don't have >1 thread with
- * os_thread=0 on the list at once
- */
- protect_control_stack_guard_page(th,1);
- /* child will not start until this is set */
- th->os_thread=kid_tid;
-}
-
void create_initial_thread(lispobj initial_function) {
struct thread *th=create_thread_struct(initial_function);
- os_thread_t kid_tid=thread_self();
- if(th && kid_tid>0) {
- link_thread(th,kid_tid);
- initial_thread_trampoline(all_threads); /* no return */
- } else lose("can't create initial thread");
+ if(th) {
+ initial_thread_trampoline(th); /* no return */
+ } else lose("can't create initial thread\n");
}
#ifdef LISP_FEATURE_SB_THREAD
#ifndef __USE_XOPEN2K
extern int pthread_attr_setstack (pthread_attr_t *__attr, void *__stackaddr,
- size_t __stacksize);
+ size_t __stacksize);
#endif
boolean create_os_thread(struct thread *th,os_thread_t *kid_tid)
sigset_t newset,oldset;
boolean r=1;
sigemptyset(&newset);
- sigaddset_blockable(&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((pthread_attr_init(&attr)) ||
return r;
}
-struct thread *create_thread(lispobj initial_function) {
+os_thread_t create_thread(lispobj initial_function) {
struct thread *th;
- os_thread_t kid_tid=0;
- boolean success;
+ os_thread_t kid_tid;
if(linux_no_threads_p) return 0;
- th=create_thread_struct(initial_function);
+ /* 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;
- /* we must not be interrupted here after a successful
- * create_os_thread, because the kid will be waiting for its
- * thread struct to be linked */
- GET_ALL_THREADS_LOCK("create_thread")
-
- success=create_os_thread(th,&kid_tid);
- if (success)
- link_thread(th,kid_tid);
- else
- os_invalidate((os_vm_address_t) th->control_stack_start,
- ((sizeof (lispobj))
- * (th->control_stack_end-th->control_stack_start)) +
- BINDING_STACK_SIZE+ALIEN_STACK_SIZE+dynamic_values_bytes+
- 32*SIGSTKSZ);
-
- RELEASE_ALL_THREADS_LOCK("create_thread")
-
- if (success)
- return th;
- else
+ if (create_os_thread(th,&kid_tid)) {
+ return kid_tid;
+ } else {
+ free_thread_struct(th);
return 0;
+ }
}
-/* called from lisp from the thread object finalizer */
-void reap_dead_thread(struct thread *th)
+/* Send the signo to os_thread, retry if the rt signal queue is
+ * full. */
+static int kill_thread_safely(os_thread_t os_thread, int signo)
{
- if(th->state!=STATE_DEAD)
- lose("thread %p is not joinable, state=%d\n",th,th->state);
-#ifdef LISP_FEATURE_GENCGC
- {
- sigset_t newset,oldset;
- sigemptyset(&newset);
- sigaddset_blockable(&newset);
- thread_sigmask(SIG_BLOCK, &newset, &oldset);
- gc_alloc_update_page_tables(0, &th->alloc_region);
- release_spinlock(&all_threads_lock);
- thread_sigmask(SIG_SETMASK,&oldset,0);
+ 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);
}
-#endif
- GET_ALL_THREADS_LOCK("reap_dead_thread")
- FSHOW((stderr,"/reap_dead_thread: reaping %ld\n",th->os_thread));
- if(th->prev)
- th->prev->next=th->next;
- else all_threads=th->next;
- if(th->next)
- th->next->prev=th->prev;
- RELEASE_ALL_THREADS_LOCK("reap_dead_thread")
- if(th->tls_cookie>=0) arch_os_thread_cleanup(th);
- gc_assert(pthread_join(th->os_thread,NULL)==0);
- os_invalidate((os_vm_address_t) th->control_stack_start,
- ((sizeof (lispobj))
- * (th->control_stack_end-th->control_stack_start)) +
- BINDING_STACK_SIZE+ALIEN_STACK_SIZE+dynamic_values_bytes+
- 32*SIGSTKSZ);
+ return r;
}
-int interrupt_thread(struct thread *th, lispobj function)
+int signal_interrupt_thread(os_thread_t os_thread)
{
- /* A thread may also become dead after this test. */
- if ((th->state != STATE_DEAD)) {
- /* In clone_threads, if A and B both interrupt C at
- * approximately the same time, it does not matter: the
- * second signal will be masked until the handler has
- * returned from the first one. In pthreads though, we
- * can't put the knowledge of what function to call into
- * the siginfo, so we have to store it in the destination
- * thread, and do it in such a way that A won't clobber
- * B's interrupt. Hence this stupid linked list.
- *
- * This does depend on SIG_INTERRUPT_THREAD being queued
- * (as POSIX RT signals are): we need to keep
- * interrupt_fun data for exactly as many signals as are
- * going to be received by the destination thread.
- */
- lispobj c=alloc_cons(function,NIL);
- int kill_status;
- /* interrupt_thread_handler locks this spinlock with
- * interrupts blocked and it does so for the sake of
- * arrange_return_to_lisp_function, so we must also block
- * them. */
- sigset_t newset,oldset;
- sigemptyset(&newset);
- sigaddset_blockable(&newset);
- thread_sigmask(SIG_BLOCK, &newset, &oldset);
- get_spinlock(&th->interrupt_fun_lock,
- (long)arch_os_get_current_thread());
- kill_status=thread_kill(th->os_thread,SIG_INTERRUPT_THREAD);
- if(kill_status==0) {
- ((struct cons *)native_pointer(c))->cdr=th->interrupt_fun;
- th->interrupt_fun=c;
- }
- release_spinlock(&th->interrupt_fun_lock);
- thread_sigmask(SIG_SETMASK,&oldset,0);
- return (kill_status ? -1 : 0);
+ 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));
}
- errno=EPERM; return -1;
}
/* stopping the world is a two-stage process. From this thread we signal
void gc_stop_the_world()
{
struct thread *p,*th=arch_os_get_current_thread();
- FSHOW_SIGNAL((stderr,"/gc_stop_the_world:waiting on lock, thread=%ld\n",
+ int status;
+ FSHOW_SIGNAL((stderr,"/gc_stop_the_world:waiting on lock, thread=%lu\n",
th->os_thread));
/* keep threads from starting while the world is stopped. */
- get_spinlock(&all_threads_lock,(long)th);
- FSHOW_SIGNAL((stderr,"/gc_stop_the_world:got lock, thread=%ld\n",
+ pthread_mutex_lock(&all_threads_lock); \
+ FSHOW_SIGNAL((stderr,"/gc_stop_the_world:got lock, thread=%lu\n",
th->os_thread));
/* stop all other threads by sending them SIG_STOP_FOR_GC */
for(p=all_threads; p; p=p->next) {
- while(p->state==STATE_STARTING) sched_yield();
- if((p!=th) && (p->state==STATE_RUNNING)) {
- FSHOW_SIGNAL((stderr,"/gc_stop_the_world:sending sig_stop to %ld\n",
+ gc_assert(p->os_thread != 0);
+ if((p!=th) && ((p->state==STATE_RUNNING))) {
+ FSHOW_SIGNAL((stderr,"/gc_stop_the_world: suspending %lu\n",
p->os_thread));
- if(thread_kill(p->os_thread,SIG_STOP_FOR_GC)==-1) {
- /* we can't kill the thread; assume because it died
- * since we last checked */
- p->state=STATE_DEAD;
- FSHOW_SIGNAL((stderr,"/gc_stop_the_world:assuming %ld dead\n",
- p->os_thread));
+ status=kill_thread_safely(p->os_thread,SIG_STOP_FOR_GC);
+ if (status==ESRCH) {
+ /* This thread has exited. */
+ gc_assert(p->state==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;) {
- gc_assert(p->os_thread!=0);
- gc_assert(p->state!=STATE_STARTING);
- if((p==th) || (p->state==STATE_SUSPENDED) ||
- (p->state==STATE_DEAD)) {
+ if((p!=th) && (p->state==STATE_RUNNING)) {
+ sched_yield();
+ } else {
p=p->next;
}
}
void gc_start_the_world()
{
struct thread *p,*th=arch_os_get_current_thread();
+ int status;
/* 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
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 %ld\n",
+ lose("gc_start_the_world: wrong thread state is %d\n",
fixnum_value(p->state));
}
- thread_kill(p->os_thread,SIG_STOP_FOR_GC);
- }
- }
- /* we must wait for all threads to leave stopped state else we
- * risk signal accumulation and lose any meaning of
- * thread->state */
- for(p=all_threads;p;) {
- if((p==th) || (p->state!=STATE_SUSPENDED)) {
- p=p->next;
+ FSHOW_SIGNAL((stderr, "/gc_start_the_world: resuming %lu\n",
+ p->os_thread));
+ p->state=STATE_RUNNING;
+ status=kill_thread_safely(p->os_thread,SIG_STOP_FOR_GC);
+ if (status) {
+ lose("cannot resume thread=%lu: %d, %s\n",
+ p->os_thread,status,strerror(status));
+ }
}
}
- release_spinlock(&all_threads_lock);
+ /* 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. */
+ pthread_mutex_unlock(&all_threads_lock); \
FSHOW_SIGNAL((stderr,"/gc_start_the_world:end\n"));
}
#endif
projects / sbcl.git / blobdiff