2 #include "pthread_impl.h"
3 #include "stdio_impl.h"
13 weak_alias(dummy_0, __acquire_ptc);
14 weak_alias(dummy_0, __release_ptc);
15 weak_alias(dummy_0, __pthread_tsd_run_dtors);
16 weak_alias(dummy_0, __do_orphaned_stdio_locks);
17 weak_alias(dummy_0, __dl_thread_cleanup);
18 weak_alias(dummy_0, __membarrier_init);
20 static int tl_lock_count;
21 static int tl_lock_waiters;
25 int tid = __pthread_self()->tid;
26 int val = __thread_list_lock;
31 while ((val = a_cas(&__thread_list_lock, 0, tid)))
32 __wait(&__thread_list_lock, &tl_lock_waiters, val, 0);
35 void __tl_unlock(void)
41 a_store(&__thread_list_lock, 0);
42 if (tl_lock_waiters) __wake(&__thread_list_lock, 1, 0);
45 void __tl_sync(pthread_t td)
48 int val = __thread_list_lock;
50 __wait(&__thread_list_lock, &tl_lock_waiters, val, 0);
51 if (tl_lock_waiters) __wake(&__thread_list_lock, 1, 0);
54 _Noreturn void __pthread_exit(void *result)
56 pthread_t self = __pthread_self();
59 self->canceldisable = 1;
60 self->cancelasync = 0;
61 self->result = result;
63 while (self->cancelbuf) {
64 void (*f)(void *) = self->cancelbuf->__f;
65 void *x = self->cancelbuf->__x;
66 self->cancelbuf = self->cancelbuf->__next;
70 __pthread_tsd_run_dtors();
72 /* Access to target the exiting thread with syscalls that use
73 * its kernel tid is controlled by killlock. For detached threads,
74 * any use past this point would have undefined behavior, but for
75 * joinable threads it's a valid usage that must be handled.
76 * Signals must be blocked since pthread_kill must be AS-safe. */
77 __block_app_sigs(&set);
80 /* The thread list lock must be AS-safe, and thus depends on
81 * application signals being blocked above. */
84 /* If this is the only thread in the list, don't proceed with
85 * termination of the thread, but restore the previous lock and
86 * signal state to prepare for exit to call atexit handlers. */
87 if (self->next == self) {
89 UNLOCK(self->killlock);
94 /* At this point we are committed to thread termination. */
96 /* Process robust list in userspace to handle non-pshared mutexes
97 * and the detached thread case where the robust list head will
98 * be invalid when the kernel would process it. */
100 volatile void *volatile *rp;
101 while ((rp=self->robust_list.head) && rp != &self->robust_list.head) {
102 pthread_mutex_t *m = (void *)((char *)rp
103 - offsetof(pthread_mutex_t, _m_next));
104 int waiters = m->_m_waiters;
105 int priv = (m->_m_type & 128) ^ 128;
106 self->robust_list.pending = rp;
107 self->robust_list.head = *rp;
108 int cont = a_swap(&m->_m_lock, 0x40000000);
109 self->robust_list.pending = 0;
110 if (cont < 0 || waiters)
111 __wake(&m->_m_lock, 1, priv);
115 __do_orphaned_stdio_locks();
116 __dl_thread_cleanup();
118 /* Last, unlink thread from the list. This change will not be visible
119 * until the lock is released, which only happens after SYS_exit
120 * has been called, via the exit futex address pointing at the lock.
121 * This needs to happen after any possible calls to LOCK() that might
122 * skip locking if process appears single-threaded. */
123 if (!--libc.threads_minus_1) libc.need_locks = -1;
124 self->next->prev = self->prev;
125 self->prev->next = self->next;
126 self->prev = self->next = self;
128 /* This atomic potentially competes with a concurrent pthread_detach
129 * call; the loser is responsible for freeing thread resources. */
130 int state = a_cas(&self->detach_state, DT_JOINABLE, DT_EXITING);
132 if (state==DT_DETACHED && self->map_base) {
133 /* Detached threads must block even implementation-internal
134 * signals, since they will not have a stack in their last
135 * moments of existence. */
136 __block_all_sigs(&set);
138 /* Robust list will no longer be valid, and was already
139 * processed above, so unregister it with the kernel. */
140 if (self->robust_list.off)
141 __syscall(SYS_set_robust_list, 0, 3*sizeof(long));
143 /* Since __unmapself bypasses the normal munmap code path,
144 * explicitly wait for vmlock holders first. */
147 /* The following call unmaps the thread's stack mapping
148 * and then exits without touching the stack. */
149 __unmapself(self->map_base, self->map_size);
152 /* Wake any joiner. */
153 __wake(&self->detach_state, 1, 1);
155 /* After the kernel thread exits, its tid may be reused. Clear it
156 * to prevent inadvertent use and inform functions that would use
157 * it that it's no longer available. */
159 UNLOCK(self->killlock);
161 for (;;) __syscall(SYS_exit, 0);
164 void __do_cleanup_push(struct __ptcb *cb)
166 struct pthread *self = __pthread_self();
167 cb->__next = self->cancelbuf;
168 self->cancelbuf = cb;
171 void __do_cleanup_pop(struct __ptcb *cb)
173 __pthread_self()->cancelbuf = cb->__next;
177 void *(*start_func)(void *);
179 volatile int control;
180 unsigned long sig_mask[_NSIG/8/sizeof(long)];
183 static int start(void *p)
185 struct start_args *args = p;
186 int state = args->control;
188 if (a_cas(&args->control, 1, 2)==1)
189 __wait(&args->control, 0, 2, 1);
191 __syscall(SYS_set_tid_address, &args->control);
192 for (;;) __syscall(SYS_exit, 0);
195 __syscall(SYS_rt_sigprocmask, SIG_SETMASK, &args->sig_mask, 0, _NSIG/8);
196 __pthread_exit(args->start_func(args->start_arg));
200 static int start_c11(void *p)
202 struct start_args *args = p;
203 int (*start)(void*) = (int(*)(void*)) args->start_func;
204 __pthread_exit((void *)(uintptr_t)start(args->start_arg));
208 #define ROUND(x) (((x)+PAGE_SIZE-1)&-PAGE_SIZE)
210 /* pthread_key_create.c overrides this */
211 static volatile size_t dummy = 0;
212 weak_alias(dummy, __pthread_tsd_size);
213 static void *dummy_tsd[1] = { 0 };
214 weak_alias(dummy_tsd, __pthread_tsd_main);
216 static FILE *volatile dummy_file = 0;
217 weak_alias(dummy_file, __stdin_used);
218 weak_alias(dummy_file, __stdout_used);
219 weak_alias(dummy_file, __stderr_used);
221 static void init_file_lock(FILE *f)
223 if (f && f->lock<0) f->lock = 0;
226 int __pthread_create(pthread_t *restrict res, const pthread_attr_t *restrict attrp, void *(*entry)(void *), void *restrict arg)
228 int ret, c11 = (attrp == __ATTRP_C11_THREAD);
230 struct pthread *self, *new;
231 unsigned char *map = 0, *stack = 0, *tsd = 0, *stack_limit;
232 unsigned flags = CLONE_VM | CLONE_FS | CLONE_FILES | CLONE_SIGHAND
233 | CLONE_THREAD | CLONE_SYSVSEM | CLONE_SETTLS
234 | CLONE_PARENT_SETTID | CLONE_CHILD_CLEARTID | CLONE_DETACHED;
235 pthread_attr_t attr = { 0 };
238 if (!libc.can_do_threads) return ENOSYS;
239 self = __pthread_self();
240 if (!libc.threaded) {
241 for (FILE *f=*__ofl_lock(); f; f=f->next)
244 init_file_lock(__stdin_used);
245 init_file_lock(__stdout_used);
246 init_file_lock(__stderr_used);
247 __syscall(SYS_rt_sigprocmask, SIG_UNBLOCK, SIGPT_SET, 0, _NSIG/8);
248 self->tsd = (void **)__pthread_tsd_main;
252 if (attrp && !c11) attr = *attrp;
256 attr._a_stacksize = __default_stacksize;
257 attr._a_guardsize = __default_guardsize;
260 if (attr._a_stackaddr) {
261 size_t need = libc.tls_size + __pthread_tsd_size;
262 size = attr._a_stacksize;
263 stack = (void *)(attr._a_stackaddr & -16);
264 stack_limit = (void *)(attr._a_stackaddr - size);
265 /* Use application-provided stack for TLS only when
266 * it does not take more than ~12% or 2k of the
267 * application's stack space. */
268 if (need < size/8 && need < 2048) {
269 tsd = stack - __pthread_tsd_size;
270 stack = tsd - libc.tls_size;
271 memset(stack, 0, need);
277 guard = ROUND(attr._a_guardsize);
278 size = guard + ROUND(attr._a_stacksize
279 + libc.tls_size + __pthread_tsd_size);
284 map = __mmap(0, size, PROT_NONE, MAP_PRIVATE|MAP_ANON, -1, 0);
285 if (map == MAP_FAILED) goto fail;
286 if (__mprotect(map+guard, size-guard, PROT_READ|PROT_WRITE)
287 && errno != ENOSYS) {
292 map = __mmap(0, size, PROT_READ|PROT_WRITE, MAP_PRIVATE|MAP_ANON, -1, 0);
293 if (map == MAP_FAILED) goto fail;
295 tsd = map + size - __pthread_tsd_size;
297 stack = tsd - libc.tls_size;
298 stack_limit = map + guard;
302 new = __copy_tls(tsd - libc.tls_size);
304 new->map_size = size;
306 new->stack_size = stack - stack_limit;
307 new->guard_size = guard;
309 new->tsd = (void *)tsd;
310 new->locale = &libc.global_locale;
311 if (attr._a_detach) {
312 new->detach_state = DT_DETACHED;
314 new->detach_state = DT_JOINABLE;
316 new->robust_list.head = &new->robust_list.head;
317 new->canary = self->canary;
318 new->sysinfo = self->sysinfo;
320 /* Setup argument structure for the new thread on its stack.
321 * It's safe to access from the caller only until the thread
322 * list is unlocked. */
323 stack -= (uintptr_t)stack % sizeof(uintptr_t);
324 stack -= sizeof(struct start_args);
325 struct start_args *args = (void *)stack;
326 args->start_func = entry;
327 args->start_arg = arg;
328 args->control = attr._a_sched ? 1 : 0;
330 /* Application signals (but not the synccall signal) must be
331 * blocked before the thread list lock can be taken, to ensure
332 * that the lock is AS-safe. */
333 __block_app_sigs(&set);
335 /* Ensure SIGCANCEL is unblocked in new thread. This requires
336 * working with a copy of the set so we can restore the
337 * original mask in the calling thread. */
338 memcpy(&args->sig_mask, &set, sizeof args->sig_mask);
339 args->sig_mask[(SIGCANCEL-1)/8/sizeof(long)] &=
340 ~(1UL<<((SIGCANCEL-1)%(8*sizeof(long))));
343 if (!libc.threads_minus_1++) libc.need_locks = 1;
344 ret = __clone((c11 ? start_c11 : start), stack, flags, args, &new->tid, TP_ADJ(new), &__thread_list_lock);
346 /* All clone failures translate to EAGAIN. If explicit scheduling
347 * was requested, attempt it before unlocking the thread list so
348 * that the failed thread is never exposed and so that we can
349 * clean up all transient resource usage before returning. */
352 } else if (attr._a_sched) {
353 ret = __syscall(SYS_sched_setscheduler,
354 new->tid, attr._a_policy, &attr._a_prio);
355 if (a_swap(&args->control, ret ? 3 : 0)==2)
356 __wake(&args->control, 1, 1);
358 __wait(&args->control, 0, 3, 0);
362 new->next = self->next;
364 new->next->prev = new;
365 new->prev->next = new;
367 if (!--libc.threads_minus_1) libc.need_locks = 0;
370 __restore_sigs(&set);
374 if (map) __munmap(map, size);
385 weak_alias(__pthread_exit, pthread_exit);
386 weak_alias(__pthread_create, pthread_create);