X-Git-Url: http://nsz.repo.hu/git/?a=blobdiff_plain;f=src%2Fthread%2Fpthread_create.c;h=10f1b7d8cd39605d42dfea19370c88b9183e6036;hb=4554f155dd23a65fcdfd39f1d5af8af55ba37694;hp=b6c475d0ad7349c45acba92d8886f15b5ac22f45;hpb=e882756311c7b06e59fcc8e582f03852b7dcfd30;p=musl

diff --git a/src/thread/pthread_create.c b/src/thread/pthread_create.c
index b6c475d0..10f1b7d8 100644
--- a/src/thread/pthread_create.c
+++ b/src/thread/pthread_create.c
@@ -1,224 +1,386 @@
+#define _GNU_SOURCE
 #include "pthread_impl.h"
-
-void __pthread_unwind_next(struct __ptcb *cb)
+#include "stdio_impl.h"
+#include "libc.h"
+#include "lock.h"
+#include <sys/mman.h>
+#include <string.h>
+#include <stddef.h>
+
+static void dummy_0()
 {
-	int i, j, not_finished;
-	pthread_t self;
-
-	if (cb->__next) longjmp((void *)cb->__next->__jb, 1);
-
-	self = pthread_self();
-	if (self->cancel) self->result = PTHREAD_CANCELLED;
-
-	if (!a_fetch_add(&libc.threads_minus_1, -1))
-		exit(0);
-
-	LOCK(&self->exitlock);
-
-	not_finished = self->tsd_used;
-	for (j=0; not_finished && j<PTHREAD_DESTRUCTOR_ITERATIONS; j++) {
-		not_finished = 0;
-		for (i=0; i<PTHREAD_KEYS_MAX; i++) {
-			if (self->tsd[i] && libc.tsd_keys[i]) {
-				void *tmp = self->tsd[i];
-				self->tsd[i] = 0;
-				libc.tsd_keys[i](tmp);
-				not_finished = 1;
-			}
-		}
-	}
-
-	if (self->detached && self->map_base)
-		__unmapself(self->map_base, self->map_size);
-
-	__syscall_exit(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_orphaned_stdio_locks);
+weak_alias(dummy_0, __dl_thread_cleanup);
+weak_alias(dummy_0, __membarrier_init);
 
-static void docancel(struct pthread *self)
+static int tl_lock_count;
+static int tl_lock_waiters;
+
+void __tl_lock(void)
 {
-	struct __ptcb cb = { .__next = self->cancelbuf };
-	__pthread_unwind_next(&cb);
+	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);
 }
 
-static void cancel_handler(int sig, siginfo_t *si, void *ctx)
+void __tl_unlock(void)
 {
-	struct pthread *self = __pthread_self();
-	self->cancel = 1;
-	if (self->canceldisable || (!self->cancelasync && !self->cancelpoint))
+	if (tl_lock_count) {
+		tl_lock_count--;
 		return;
-	docancel(self);
+	}
+	a_store(&__thread_list_lock, 0);
+	if (tl_lock_waiters) __wake(&__thread_list_lock, 1, 0);
 }
 
-static void cancelpt(int x)
+void __tl_sync(pthread_t td)
 {
-	struct pthread *self = __pthread_self();
-	if (self->canceldisable) return;
-	self->cancelpoint = x;
-	if (self->cancel) docancel(self);
+	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);
 }
 
-/* "rsyscall" is a mechanism by which a thread can synchronously force all
- * other threads to perform an arbitrary syscall. It is necessary to work
- * around the non-conformant implementation of setuid() et al on Linux,
- * which affect only the calling thread and not the whole process. This
- * implementation performs some tricks with signal delivery to work around
- * the fact that it does not keep any list of threads in userspace. */
-
-static struct {
-	volatile int lock, hold, blocks, cnt;
-	unsigned long arg[6];
-	int nr;
-	int err;
-} rs;
-
-static void rsyscall_handler(int sig, siginfo_t *si, void *ctx)
+_Noreturn void __pthread_exit(void *result)
 {
-	if (rs.cnt == libc.threads_minus_1) return;
+	pthread_t self = __pthread_self();
+	sigset_t set;
 
-	if (syscall6(rs.nr, rs.arg[0], rs.arg[1], rs.arg[2],
-		rs.arg[3], rs.arg[4], rs.arg[5]) < 0 && !rs.err) rs.err=errno;
+	self->canceldisable = 1;
+	self->cancelasync = 0;
+	self->result = result;
 
-	a_inc(&rs.cnt);
-	__wake(&rs.cnt, 1, 1);
-	while(rs.hold)
-		__wait(&rs.hold, 0, 1, 1);
-	a_dec(&rs.cnt);
-	if (!rs.cnt) __wake(&rs.cnt, 1, 1);
-}
+	while (self->cancelbuf) {
+		void (*f)(void *) = self->cancelbuf->__f;
+		void *x = self->cancelbuf->__x;
+		self->cancelbuf = self->cancelbuf->__next;
+		f(x);
+	}
 
-static int rsyscall(int nr, long a, long b, long c, long d, long e, long f)
-{
-	int i, ret;
-	sigset_t set = { 0 };
-	struct pthread *self = __pthread_self();
-	sigaddset(&set, SIGSYSCALL);
-
-	LOCK(&rs.lock);
-	while ((i=rs.blocks))
-		__wait(&rs.blocks, 0, i, 1);
-
-	__libc_sigprocmask(SIG_BLOCK, &set, 0);
-
-	rs.nr = nr;
-	rs.arg[0] = a; rs.arg[1] = b;
-	rs.arg[2] = c; rs.arg[3] = d;
-	rs.arg[4] = d; rs.arg[5] = f;
-	rs.hold = 1;
-	rs.err = 0;
-	rs.cnt = 0;
-
-	/* Dispatch signals until all threads respond */
-	for (i=libc.threads_minus_1; i; i--)
-		sigqueue(self->pid, SIGSYSCALL, (union sigval){0});
-	while ((i=rs.cnt) < libc.threads_minus_1) {
-		sigqueue(self->pid, SIGSYSCALL, (union sigval){0});
-		__wait(&rs.cnt, 0, i, 1);
+	__pthread_tsd_run_dtors();
+
+	/* 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.
+	 * Signals must be blocked since pthread_kill must be AS-safe. */
+	__block_app_sigs(&set);
+	LOCK(self->killlock);
+
+	/* 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);
 	}
 
-	/* Handle any lingering signals with no-op */
-	__libc_sigprocmask(SIG_UNBLOCK, &set, 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. */
+	__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_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 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);
 
-	/* Resume other threads' signal handlers and wait for them */
-	rs.hold = 0;
-	__wake(&rs.hold, -1, 0);
-	while((i=rs.cnt)) __wait(&rs.cnt, 0, i, 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);
 
-	if (rs.err) errno = rs.err, ret = -1;
-	else ret = syscall6(nr, a, b, c, d, e, f);
+	for (;;) __syscall(SYS_exit, 0);
+}
 
-	UNLOCK(&rs.lock);
-	return ret;
+void __do_cleanup_push(struct __ptcb *cb)
+{
+	struct pthread *self = __pthread_self();
+	cb->__next = self->cancelbuf;
+	self->cancelbuf = cb;
 }
 
-static void init_threads()
+void __do_cleanup_pop(struct __ptcb *cb)
 {
-	struct sigaction sa = { .sa_flags = SA_SIGINFO | SA_RESTART };
-	libc.lock = __lock;
-	libc.cancelpt = cancelpt;
-	libc.rsyscall = rsyscall;
-	sa.sa_sigaction = cancel_handler;
-	__libc_sigaction(SIGCANCEL, &sa, 0);
-	sigaddset(&sa.sa_mask, SIGSYSCALL);
-	sigaddset(&sa.sa_mask, SIGCANCEL);
-	sa.sa_sigaction = rsyscall_handler;
-	__libc_sigaction(SIGSYSCALL, &sa, 0);
-	sigprocmask(SIG_UNBLOCK, &sa.sa_mask, 0);
+	__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)
 {
-	struct pthread *self = p;
-	pthread_exit(self->start(self->start_arg));
+	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);
+		}
+	}
+	__syscall(SYS_rt_sigprocmask, SIG_SETMASK, &args->sig_mask, 0, _NSIG/8);
+	__pthread_exit(args->start_func(args->start_arg));
 	return 0;
 }
 
-int __uniclone(void *, int (*)(), void *);
+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;
+}
 
 #define ROUND(x) (((x)+PAGE_SIZE-1)&-PAGE_SIZE)
 
 /* pthread_key_create.c overrides this */
-static const size_t dummy = 0;
+static volatile size_t dummy = 0;
 weak_alias(dummy, __pthread_tsd_size);
+static void *dummy_tsd[1] = { 0 };
+weak_alias(dummy_tsd, __pthread_tsd_main);
+
+static FILE *volatile dummy_file = 0;
+weak_alias(dummy_file, __stdin_used);
+weak_alias(dummy_file, __stdout_used);
+weak_alias(dummy_file, __stderr_used);
+
+static void init_file_lock(FILE *f)
+{
+	if (f && f->lock<0) f->lock = 0;
+}
 
-int pthread_create(pthread_t *res, const pthread_attr_t *attr, void *(*entry)(void *), void *arg)
+int __pthread_create(pthread_t *restrict res, const pthread_attr_t *restrict attrp, void *(*entry)(void *), void *restrict arg)
 {
-	static int init;
-	int ret;
+	int ret, c11 = (attrp == __ATTRP_C11_THREAD);
 	size_t size, guard;
-	struct pthread *self = pthread_self(), *new;
-	unsigned char *map, *stack, *tsd;
-	static const pthread_attr_t default_attr;
-
-	if (!self) return errno = ENOSYS;
-	if (!init && ++init) init_threads();
-
-	if (!attr) attr = &default_attr;
-	guard = ROUND(attr->_a_guardsize + DEFAULT_GUARD_SIZE);
-	size = guard + ROUND(attr->_a_stacksize + DEFAULT_STACK_SIZE);
-	size += __pthread_tsd_size;
-	map = mmap(0, size, PROT_READ|PROT_WRITE|PROT_EXEC, MAP_PRIVATE|MAP_ANON, -1, 0);
-	if (!map) return EAGAIN;
-	mprotect(map, guard, PROT_NONE);
-
-	tsd = map + size - __pthread_tsd_size;
-	new = (void *)(tsd - sizeof *new - PAGE_SIZE%sizeof *new);
+	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;
+	pthread_attr_t attr = { 0 };
+	sigset_t set;
+
+	if (!libc.can_do_threads) return ENOSYS;
+	self = __pthread_self();
+	if (!libc.threaded) {
+		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 && !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;
+		stack = (void *)(attr._a_stackaddr & -16);
+		stack_limit = (void *)(attr._a_stackaddr - size);
+		/* Use application-provided stack for TLS only when
+		 * it does not take more than ~12% or 2k of the
+		 * application's stack space. */
+		if (need < size/8 && need < 2048) {
+			tsd = stack - __pthread_tsd_size;
+			stack = tsd - libc.tls_size;
+			memset(stack, 0, need);
+		} else {
+			size = ROUND(need);
+		}
+		guard = 0;
+	} else {
+		guard = ROUND(attr._a_guardsize);
+		size = guard + ROUND(attr._a_stacksize
+			+ libc.tls_size +  __pthread_tsd_size);
+	}
+
+	if (!tsd) {
+		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)
+			    && errno != ENOSYS) {
+				__munmap(map, size);
+				goto fail;
+			}
+		} else {
+			map = __mmap(0, size, PROT_READ|PROT_WRITE, MAP_PRIVATE|MAP_ANON, -1, 0);
+			if (map == MAP_FAILED) goto fail;
+		}
+		tsd = map + size - __pthread_tsd_size;
+		if (!stack) {
+			stack = tsd - libc.tls_size;
+			stack_limit = map + guard;
+		}
+	}
+
+	new = __copy_tls(tsd - libc.tls_size);
 	new->map_base = map;
 	new->map_size = size;
-	new->pid = self->pid;
-	new->errno_ptr = &new->errno_val;
-	new->start = entry;
-	new->start_arg = arg;
+	new->stack = stack;
+	new->stack_size = stack - stack_limit;
+	new->guard_size = guard;
 	new->self = new;
 	new->tsd = (void *)tsd;
-	new->detached = attr->_a_detach;
-	new->attr = *attr;
-	memcpy(new->tlsdesc, self->tlsdesc, sizeof new->tlsdesc);
-	new->tlsdesc[1] = (uintptr_t)new;
-	stack = (void *)((uintptr_t)new-1 & ~(uintptr_t)15);
-
-	/* We must synchronize new thread creation with rsyscall
-	 * delivery. This looks to be the least expensive way: */
-	a_inc(&rs.blocks);
-	while (rs.lock) __wait(&rs.lock, 0, 1, 1);
-
-	a_inc(&libc.threads_minus_1);
-	ret = __uniclone(stack, start, new);
+	new->locale = &libc.global_locale;
+	if (attr._a_detach) {
+		new->detach_state = DT_DETACHED;
+	} 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.
+	 * 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();
+	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);
+	}
 
-	a_dec(&rs.blocks);
-	if (rs.lock) __wake(&rs.blocks, 1, 1);
+	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);
-		munmap(map, size);
-		return EAGAIN;
+		if (map) __munmap(map, size);
+		return -ret;
 	}
+
 	*res = new;
 	return 0;
+fail:
+	__release_ptc();
+	return EAGAIN;
 }
 
-void pthread_exit(void *result)
-{
-	struct pthread *self = pthread_self();
-	self->result = result;
-	docancel(self);
-}
+weak_alias(__pthread_exit, pthread_exit);
+weak_alias(__pthread_create, pthread_create);