weak_alias(dummy_0, __pthread_tsd_run_dtors);
weak_alias(dummy_0, __do_orphaned_stdio_locks);
weak_alias(dummy_0, __dl_thread_cleanup);
+weak_alias(dummy_0, __membarrier_init);
+
+static int tl_lock_count;
+static int tl_lock_waiters;
+
+void __tl_lock(void)
+{
+ int tid = __pthread_self()->tid;
+ int val = __thread_list_lock;
+ if (val == tid) {
+ tl_lock_count++;
+ return;
+ }
+ while ((val = a_cas(&__thread_list_lock, 0, tid)))
+ __wait(&__thread_list_lock, &tl_lock_waiters, val, 0);
+}
+
+void __tl_unlock(void)
+{
+ if (tl_lock_count) {
+ tl_lock_count--;
+ return;
+ }
+ a_store(&__thread_list_lock, 0);
+ if (tl_lock_waiters) __wake(&__thread_list_lock, 1, 0);
+}
+
+void __tl_sync(pthread_t td)
+{
+ a_barrier();
+ int val = __thread_list_lock;
+ if (!val) return;
+ __wait(&__thread_list_lock, &tl_lock_waiters, val, 0);
+ if (tl_lock_waiters) __wake(&__thread_list_lock, 1, 0);
+}
_Noreturn void __pthread_exit(void *result)
{
/* Access to target the exiting thread with syscalls that use
* its kernel tid is controlled by killlock. For detached threads,
* any use past this point would have undefined behavior, but for
- * joinable threads it's a valid usage that must be handled. */
+ * joinable threads it's a valid usage that must be handled.
+ * Signals must be blocked since pthread_kill must be AS-safe. */
+ __block_app_sigs(&set);
LOCK(self->killlock);
- /* Block all signals before decrementing the live thread count.
- * This is important to ensure that dynamically allocated TLS
- * is not under-allocated/over-committed, and possibly for other
- * reasons as well. */
- __block_all_sigs(&set);
-
- /* It's impossible to determine whether this is "the last thread"
- * until performing the atomic decrement, since multiple threads
- * could exit at the same time. For the last thread, revert the
- * decrement, restore the tid, and unblock signals to give the
- * atexit handlers and stdio cleanup code a consistent state. */
- if (a_fetch_add(&libc.threads_minus_1, -1)==0) {
- libc.threads_minus_1 = 0;
+ /* The thread list lock must be AS-safe, and thus depends on
+ * application signals being blocked above. */
+ __tl_lock();
+
+ /* If this is the only thread in the list, don't proceed with
+ * termination of the thread, but restore the previous lock and
+ * signal state to prepare for exit to call atexit handlers. */
+ if (self->next == self) {
+ __tl_unlock();
UNLOCK(self->killlock);
__restore_sigs(&set);
exit(0);
}
+ /* At this point we are committed to thread termination. */
+
/* Process robust list in userspace to handle non-pshared mutexes
* and the detached thread case where the robust list head will
* be invalid when the kernel would process it. */
__do_orphaned_stdio_locks();
__dl_thread_cleanup();
+ /* Last, unlink thread from the list. This change will not be visible
+ * until the lock is released, which only happens after SYS_exit
+ * has been called, via the exit futex address pointing at the lock.
+ * This needs to happen after any possible calls to LOCK() that might
+ * skip locking if process appears single-threaded. */
+ if (!--libc.threads_minus_1) libc.need_locks = -1;
+ self->next->prev = self->prev;
+ self->prev->next = self->next;
+ self->prev = self->next = self;
+
/* This atomic potentially competes with a concurrent pthread_detach
* call; the loser is responsible for freeing thread resources. */
int state = a_cas(&self->detach_state, DT_JOINABLE, DT_EXITING);
- if (state>=DT_DETACHED && self->map_base) {
- /* Detached threads must avoid the kernel clear_child_tid
- * feature, since the virtual address will have been
- * unmapped and possibly already reused by a new mapping
- * at the time the kernel would perform the write. In
- * the case of threads that started out detached, the
- * initial clone flags are correct, but if the thread was
- * detached later, we need to clear it here. */
- if (state == DT_DYNAMIC) __syscall(SYS_set_tid_address, 0);
+ if (state==DT_DETACHED && self->map_base) {
+ /* Detached threads must block even implementation-internal
+ * signals, since they will not have a stack in their last
+ * moments of existence. */
+ __block_all_sigs(&set);
/* Robust list will no longer be valid, and was already
* processed above, so unregister it with the kernel. */
__unmapself(self->map_base, self->map_size);
}
+ /* Wake any joiner. */
+ __wake(&self->detach_state, 1, 1);
+
/* After the kernel thread exits, its tid may be reused. Clear it
* to prevent inadvertent use and inform functions that would use
* it that it's no longer available. */
struct start_args {
void *(*start_func)(void *);
void *start_arg;
- pthread_attr_t *attr;
- volatile int *perr;
+ volatile int control;
unsigned long sig_mask[_NSIG/8/sizeof(long)];
};
static int start(void *p)
{
struct start_args *args = p;
- if (args->attr) {
- pthread_t self = __pthread_self();
- int ret = -__syscall(SYS_sched_setscheduler, self->tid,
- args->attr->_a_policy, &args->attr->_a_prio);
- if (a_swap(args->perr, ret)==-2)
- __wake(args->perr, 1, 1);
- if (ret) {
- self->detach_state = DT_DYNAMIC;
- __pthread_exit(0);
+ int state = args->control;
+ if (state) {
+ if (a_cas(&args->control, 1, 2)==1)
+ __wait(&args->control, 0, 2, 1);
+ if (args->control) {
+ __syscall(SYS_set_tid_address, &args->control);
+ for (;;) __syscall(SYS_exit, 0);
}
}
__syscall(SYS_rt_sigprocmask, SIG_SETMASK, &args->sig_mask, 0, _NSIG/8);
static void *dummy_tsd[1] = { 0 };
weak_alias(dummy_tsd, __pthread_tsd_main);
-volatile int __block_new_threads = 0;
-
static FILE *volatile dummy_file = 0;
weak_alias(dummy_file, __stdin_used);
weak_alias(dummy_file, __stdout_used);
| CLONE_PARENT_SETTID | CLONE_CHILD_CLEARTID | CLONE_DETACHED;
pthread_attr_t attr = { 0 };
sigset_t set;
- volatile int err = -1;
if (!libc.can_do_threads) return ENOSYS;
self = __pthread_self();
init_file_lock(__stderr_used);
__syscall(SYS_rt_sigprocmask, SIG_UNBLOCK, SIGPT_SET, 0, _NSIG/8);
self->tsd = (void **)__pthread_tsd_main;
+ __membarrier_init();
libc.threaded = 1;
}
if (attrp && !c11) attr = *attrp;
attr._a_guardsize = __default_guardsize;
}
- if (__block_new_threads) __wait(&__block_new_threads, 0, 1, 1);
-
if (attr._a_stackaddr) {
size_t need = libc.tls_size + __pthread_tsd_size;
size = attr._a_stacksize;
new->locale = &libc.global_locale;
if (attr._a_detach) {
new->detach_state = DT_DETACHED;
- flags -= CLONE_CHILD_CLEARTID;
} else {
new->detach_state = DT_JOINABLE;
}
new->robust_list.head = &new->robust_list.head;
new->CANARY = self->CANARY;
+ new->sysinfo = self->sysinfo;
- /* Setup argument structure for the new thread on its stack. */
+ /* Setup argument structure for the new thread on its stack.
+ * It's safe to access from the caller only until the thread
+ * list is unlocked. */
stack -= (uintptr_t)stack % sizeof(uintptr_t);
stack -= sizeof(struct start_args);
struct start_args *args = (void *)stack;
args->start_func = entry;
args->start_arg = arg;
- if (attr._a_sched) {
- args->attr = &attr;
- args->perr = &err;
- } else {
- args->attr = 0;
- args->perr = 0;
- }
+ args->control = attr._a_sched ? 1 : 0;
+ /* Application signals (but not the synccall signal) must be
+ * blocked before the thread list lock can be taken, to ensure
+ * that the lock is AS-safe. */
__block_app_sigs(&set);
/* Ensure SIGCANCEL is unblocked in new thread. This requires
args->sig_mask[(SIGCANCEL-1)/8/sizeof(long)] &=
~(1UL<<((SIGCANCEL-1)%(8*sizeof(long))));
- a_inc(&libc.threads_minus_1);
- ret = __clone((c11 ? start_c11 : start), stack, flags, args, &new->tid, TP_ADJ(new), &new->detach_state);
+ __tl_lock();
+ if (!libc.threads_minus_1++) libc.need_locks = 1;
+ ret = __clone((c11 ? start_c11 : start), stack, flags, args, &new->tid, TP_ADJ(new), &__thread_list_lock);
+ /* All clone failures translate to EAGAIN. If explicit scheduling
+ * was requested, attempt it before unlocking the thread list so
+ * that the failed thread is never exposed and so that we can
+ * clean up all transient resource usage before returning. */
+ if (ret < 0) {
+ ret = -EAGAIN;
+ } else if (attr._a_sched) {
+ ret = __syscall(SYS_sched_setscheduler,
+ new->tid, attr._a_policy, &attr._a_prio);
+ if (a_swap(&args->control, ret ? 3 : 0)==2)
+ __wake(&args->control, 1, 1);
+ if (ret)
+ __wait(&args->control, 0, 3, 0);
+ }
+
+ if (ret >= 0) {
+ new->next = self->next;
+ new->prev = self;
+ new->next->prev = new;
+ new->prev->next = new;
+ } else {
+ if (!--libc.threads_minus_1) libc.need_locks = 0;
+ }
+ __tl_unlock();
__restore_sigs(&set);
__release_ptc();
if (ret < 0) {
- a_dec(&libc.threads_minus_1);
if (map) __munmap(map, size);
- return EAGAIN;
- }
-
- if (attr._a_sched) {
- if (a_cas(&err, -1, -2)==-1)
- __wait(&err, 0, -2, 1);
- ret = err;
- if (ret) return ret;
+ return -ret;
}
*res = new;