#include "pthread_impl.h"
#include "stdio_impl.h"
#include "libc.h"
+#include "lock.h"
#include <sys/mman.h>
#include <string.h>
-
-void *__mmap(void *, size_t, int, int, int, off_t);
-int __munmap(void *, size_t);
-int __mprotect(void *, size_t, int);
+#include <stddef.h>
static void dummy_0()
{
weak_alias(dummy_0, __acquire_ptc);
weak_alias(dummy_0, __release_ptc);
weak_alias(dummy_0, __pthread_tsd_run_dtors);
-weak_alias(dummy_0, __do_private_robust_list);
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)
{
pthread_t self = __pthread_self();
sigset_t set;
+ self->canceldisable = 1;
+ self->cancelasync = 0;
self->result = result;
while (self->cancelbuf) {
__pthread_tsd_run_dtors();
- __lock(self->exitlock);
-
- /* Mark this thread dead before decrementing count */
- __lock(self->killlock);
- self->dead = 1;
-
- /* 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);
-
- /* Wait to unlock the kill lock, which governs functions like
- * pthread_kill which target a thread id, until signals have
- * been blocked. This precludes observation of the thread id
- * as a live thread (with application code running in it) after
- * the thread was reported dead by ESRCH being returned. */
- __unlock(self->killlock);
-
- /* 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 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;
+ /* 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. */
+ 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);
+ __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();
__restore_sigs(&set);
+ UNLOCK(self->killlock);
exit(0);
}
- if (self->locale != &libc.global_locale) {
- a_dec(&libc.uselocale_cnt);
- if (self->locale->ctype_utf8)
- a_dec(&libc.bytelocale_cnt_minus_1);
+ /* 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. */
+ __vm_lock();
+ volatile void *volatile *rp;
+ while ((rp=self->robust_list.head) && rp != &self->robust_list.head) {
+ pthread_mutex_t *m = (void *)((char *)rp
+ - offsetof(pthread_mutex_t, _m_next));
+ int waiters = m->_m_waiters;
+ int priv = (m->_m_type & 128) ^ 128;
+ self->robust_list.pending = rp;
+ self->robust_list.head = *rp;
+ int cont = a_swap(&m->_m_lock, 0x40000000);
+ self->robust_list.pending = 0;
+ if (cont < 0 || waiters)
+ __wake(&m->_m_lock, 1, priv);
}
+ __vm_unlock();
- __do_private_robust_list();
__do_orphaned_stdio_locks();
-
- if (self->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 (== 2), we need to clear it here. */
- if (self->detached == 2) __syscall(SYS_set_tid_address, 0);
+ __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 libc.threads_minus_1 is zero. */
+ libc.threads_minus_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 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. */
+ 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. */
+ __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);
}
void __do_cleanup_push(struct __ptcb *cb)
{
- if (!libc.has_thread_pointer) return;
struct pthread *self = __pthread_self();
cb->__next = self->cancelbuf;
self->cancelbuf = cb;
void __do_cleanup_pop(struct __ptcb *cb)
{
- if (!libc.has_thread_pointer) return;
__pthread_self()->cancelbuf = cb->__next;
}
+struct start_args {
+ void *(*start_func)(void *);
+ void *start_arg;
+ volatile int control;
+ unsigned long sig_mask[_NSIG/8/sizeof(long)];
+};
+
static int start(void *p)
{
- pthread_t self = p;
- if (self->startlock[0]) {
- __wait(self->startlock, 0, 1, 1);
- if (self->startlock[0]) {
- self->detached = 2;
- pthread_exit(0);
+ struct start_args *args = p;
+ 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);
}
- __restore_sigs(self->sigmask);
}
- if (self->unblock_cancel)
- __syscall(SYS_rt_sigprocmask, SIG_UNBLOCK,
- SIGPT_SET, 0, _NSIG/8);
- pthread_exit(self->start(self->start_arg));
+ __syscall(SYS_rt_sigprocmask, SIG_SETMASK, &args->sig_mask, 0, _NSIG/8);
+ __pthread_exit(args->start_func(args->start_arg));
+ return 0;
+}
+
+static int start_c11(void *p)
+{
+ struct start_args *args = p;
+ int (*start)(void*) = (int(*)(void*)) args->start_func;
+ __pthread_exit((void *)(uintptr_t)start(args->start_arg));
return 0;
}
if (f && f->lock<0) f->lock = 0;
}
-void *__copy_tls(unsigned char *);
-
int __pthread_create(pthread_t *restrict res, const pthread_attr_t *restrict attrp, void *(*entry)(void *), void *restrict arg)
{
- int ret;
+ int ret, c11 = (attrp == __ATTRP_C11_THREAD);
size_t size, guard;
struct pthread *self, *new;
unsigned char *map = 0, *stack = 0, *tsd = 0, *stack_limit;
unsigned flags = CLONE_VM | CLONE_FS | CLONE_FILES | CLONE_SIGHAND
| CLONE_THREAD | CLONE_SYSVSEM | CLONE_SETTLS
| CLONE_PARENT_SETTID | CLONE_CHILD_CLEARTID | CLONE_DETACHED;
- int do_sched = 0;
- pthread_attr_t attr = {0};
+ pthread_attr_t attr = { 0 };
+ sigset_t set;
if (!libc.can_do_threads) return ENOSYS;
self = __pthread_self();
if (!libc.threaded) {
- for (FILE *f=libc.ofl_head; f; f=f->next)
+ for (FILE *f=*__ofl_lock(); f; f=f->next)
init_file_lock(f);
+ __ofl_unlock();
init_file_lock(__stdin_used);
init_file_lock(__stdout_used);
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) attr = *attrp;
+ if (attrp && !c11) attr = *attrp;
__acquire_ptc();
+ if (!attrp || c11) {
+ attr._a_stacksize = __default_stacksize;
+ attr._a_guardsize = __default_guardsize;
+ }
if (attr._a_stackaddr) {
size_t need = libc.tls_size + __pthread_tsd_size;
- size = attr._a_stacksize + DEFAULT_STACK_SIZE;
+ size = attr._a_stacksize;
stack = (void *)(attr._a_stackaddr & -16);
stack_limit = (void *)(attr._a_stackaddr - size);
/* Use application-provided stack for TLS only when
memset(stack, 0, need);
} else {
size = ROUND(need);
- guard = 0;
}
+ guard = 0;
} else {
- guard = ROUND(DEFAULT_GUARD_SIZE + attr._a_guardsize);
- size = guard + ROUND(DEFAULT_STACK_SIZE + attr._a_stacksize
+ guard = ROUND(attr._a_guardsize);
+ size = guard + ROUND(attr._a_stacksize
+ libc.tls_size + __pthread_tsd_size);
}
if (guard) {
map = __mmap(0, size, PROT_NONE, MAP_PRIVATE|MAP_ANON, -1, 0);
if (map == MAP_FAILED) goto fail;
- if (__mprotect(map+guard, size-guard, PROT_READ|PROT_WRITE)) {
+ if (__mprotect(map+guard, size-guard, PROT_READ|PROT_WRITE)
+ && errno != ENOSYS) {
__munmap(map, size);
goto fail;
}
new->map_size = size;
new->stack = stack;
new->stack_size = stack - stack_limit;
- new->start = entry;
- new->start_arg = arg;
+ new->guard_size = guard;
new->self = new;
new->tsd = (void *)tsd;
new->locale = &libc.global_locale;
if (attr._a_detach) {
- new->detached = 1;
- flags -= CLONE_CHILD_CLEARTID;
+ new->detach_state = DT_DETACHED;
+ } else {
+ new->detach_state = DT_JOINABLE;
}
- if (attr._a_sched) {
- do_sched = new->startlock[0] = 1;
- __block_app_sigs(new->sigmask);
+ 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.
+ * 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;
+ 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
+ * working with a copy of the set so we can restore the
+ * original mask in the calling thread. */
+ memcpy(&args->sig_mask, &set, sizeof args->sig_mask);
+ args->sig_mask[(SIGCANCEL-1)/8/sizeof(long)] &=
+ ~(1UL<<((SIGCANCEL-1)%(8*sizeof(long))));
+
+ __tl_lock();
+ libc.threads_minus_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);
}
- new->unblock_cancel = self->cancel;
- new->canary = self->canary;
-
- a_inc(&libc.threads_minus_1);
- ret = __clone(start, stack, flags, new, &new->tid, TP_ADJ(new), &new->tid);
-
- __release_ptc();
- if (do_sched) {
- __restore_sigs(new->sigmask);
+ if (ret >= 0) {
+ new->next = self->next;
+ new->prev = self;
+ new->next->prev = new;
+ new->prev->next = new;
+ } else {
+ libc.threads_minus_1--;
}
+ __tl_unlock();
+ __restore_sigs(&set);
+ __release_ptc();
if (ret < 0) {
- a_dec(&libc.threads_minus_1);
if (map) __munmap(map, size);
- return EAGAIN;
- }
-
- if (do_sched) {
- ret = __syscall(SYS_sched_setscheduler, new->tid,
- attr._a_policy, &attr._a_prio);
- a_store(new->startlock, ret<0 ? 2 : 0);
- __wake(new->startlock, 1, 1);
- if (ret < 0) return -ret;
+ return -ret;
}
*res = new;
projects / musl / blobdiff