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)
{
- if (!a_cas(&__thread_list_lock, 0, 1)) return;
- do {
- a_cas(&__thread_list_lock, 1, 2);
- __futexwait(&__thread_list_lock, 2, 0);
- } while (a_cas(&__thread_list_lock, 0, 2));
+ 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 (a_swap(&__thread_list_lock, 0)==2)
- __wake(&__thread_list_lock, 1, 0);
+ 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();
- if (!__thread_list_lock) return;
- a_cas(&__thread_list_lock, 1, 2);
- __wait(&__thread_list_lock, 0, 2, 0);
- __wake(&__thread_list_lock, 1, 0);
+ 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)
__pthread_tsd_run_dtors();
+ __block_app_sigs(&set);
+
+ /* 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) {
+ /* Since __unmapself bypasses the normal munmap code path,
+ * explicitly wait for vmlock holders first. This must be
+ * done before any locks are taken, to avoid lock ordering
+ * issues that could lead to deadlock. */
+ __vm_wait();
+ }
+
/* 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. */
LOCK(self->killlock);
- /* The thread list lock must be AS-safe, and thus requires
- * application signals to be blocked before it can be taken. */
- __block_app_sigs(&set);
+ /* 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
* signal state to prepare for exit to call atexit handlers. */
if (self->next == self) {
__tl_unlock();
- __restore_sigs(&set);
UNLOCK(self->killlock);
+ self->detach_state = state;
+ __restore_sigs(&set);
exit(0);
}
- /* At this point we are committed to thread termination. Unlink
- * the 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. */
- libc.threads_minus_1--;
- self->next->prev = self->prev;
- self->prev->next = self->next;
- self->prev = self->next = self;
+ /* At this point we are committed to thread termination. */
+
+ /* 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. At this point the killlock
+ * may be released, since functions that use it will consistently
+ * see the thread as having exited. Release it now so that no
+ * remaining locks (except thread list) are held if we end up
+ * resetting need_locks below. */
+ self->tid = 0;
+ UNLOCK(self->killlock);
/* Process robust list in userspace to handle non-pshared mutexes
* and the detached thread case where the robust list head will
__do_orphaned_stdio_locks();
__dl_thread_cleanup();
- /* 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);
+ /* 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;
if (state==DT_DETACHED && self->map_base) {
/* Detached threads must block even implementation-internal
if (self->robust_list.off)
__syscall(SYS_set_robust_list, 0, 3*sizeof(long));
- /* Since __unmapself bypasses the normal munmap code path,
- * explicitly wait for vmlock holders first. */
- __vm_wait();
-
/* The following call unmaps the thread's stack mapping
* and then exits without touching the stack. */
__unmapself(self->map_base, self->map_size);
}
/* Wake any joiner. */
+ a_store(&self->detach_state, DT_EXITED);
__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. */
- self->tid = 0;
- UNLOCK(self->killlock);
-
for (;;) __syscall(SYS_exit, 0);
}
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_DETACHED;
- __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);
| 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;
new->detach_state = DT_JOINABLE;
}
new->robust_list.head = &new->robust_list.head;
- new->CANARY = self->CANARY;
+ new->canary = self->canary;
+ new->sysinfo = self->sysinfo;
/* Setup argument structure for the new thread on its stack.
* It's safe to access from the caller only until the thread
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
~(1UL<<((SIGCANCEL-1)%(8*sizeof(long))));
__tl_lock();
- libc.threads_minus_1++;
+ 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);
- /* If clone succeeded, new thread must be linked on the thread
- * list before unlocking it, even if scheduling may still fail. */
+ /* 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) {
- 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;
projects / musl / blobdiff