X-Git-Url: http://nsz.repo.hu/git/?p=musl;a=blobdiff_plain;f=src%2Fthread%2Fpthread_create.c;h=d11dcfafa15153251b6f79b72dbc9772d512a7b0;hp=c73c52114074c66d3c04bf44b3cd18a1d2d78341;hb=72768ea99e67162b7b42d9cd8917cf9a2c00f1f1;hpb=19eb13b9a4cf2f787f60b6e2a6d26a6cd7d3ffd2

diff --git a/src/thread/pthread_create.c b/src/thread/pthread_create.c
index c73c5211..d11dcfaf 100644
--- a/src/thread/pthread_create.c
+++ b/src/thread/pthread_create.c
@@ -1,189 +1,161 @@
 #include "pthread_impl.h"
+#include "stdio_impl.h"
+#include <sys/mman.h>
 
-void __pthread_unwind_next(struct __ptcb *cb)
+static void dummy_0()
 {
-	int i, j, not_finished;
-	pthread_t self;
-
-	if (cb->__next) longjmp((void *)cb->__next->__jb, 1);
+}
+weak_alias(dummy_0, __acquire_ptc);
+weak_alias(dummy_0, __release_ptc);
+weak_alias(dummy_0, __pthread_tsd_run_dtors);
 
-	self = pthread_self();
-	if (self->cancel) self->result = PTHREAD_CANCELLED;
+_Noreturn void pthread_exit(void *result)
+{
+	pthread_t self = pthread_self();
+	int n;
 
-	LOCK(&self->exitlock);
+	self->result = result;
 
-	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;
-			}
-		}
+	while (self->cancelbuf) {
+		void (*f)(void *) = self->cancelbuf->__f;
+		void *x = self->cancelbuf->__x;
+		self->cancelbuf = self->cancelbuf->__next;
+		f(x);
 	}
 
-	syscall4(__NR_sigprocmask, SIG_BLOCK, (long)(uint64_t[1]){-1}, 0, 8);
-
-	if (!a_fetch_add(&libc.threads_minus_1, -1))
-		exit(0);
-
-	if (self->detached && self->map_base)
-		__unmapself(self->map_base, self->map_size);
-
-	__syscall_exit(0);
-}
-
-static void docancel(struct pthread *self)
-{
-	struct __ptcb cb = { .__next = self->cancelbuf };
-	__pthread_unwind_next(&cb);
-}
+	__pthread_tsd_run_dtors();
 
-static void cancel_handler(int sig, siginfo_t *si, void *ctx)
-{
-	struct pthread *self = __pthread_self();
-	self->cancel = 1;
-	if (self->canceldisable || (!self->cancelasync && !self->cancelpoint))
-		return;
-	docancel(self);
-}
+	__lock(self->exitlock);
 
-static void cancelpt(int x)
-{
-	struct pthread *self = __pthread_self();
-	if (self->canceldisable) return;
-	self->cancelpoint = x;
-	if (self->cancel) docancel(self);
-}
+	/* Mark this thread dead before decrementing count */
+	__lock(self->killlock);
+	self->dead = 1;
+	__unlock(self->killlock);
 
-/* "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)
-{
-	if (rs.cnt == libc.threads_minus_1) return;
+	do n = libc.threads_minus_1;
+	while (n && a_cas(&libc.threads_minus_1, n, n-1)!=n);
+	if (!n) exit(0);
 
-	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;
+	if (self->detached && self->map_base) {
+		if (self->detached == 2)
+			__syscall(SYS_set_tid_address, 0);
+		__syscall(SYS_rt_sigprocmask, SIG_BLOCK,
+			SIGALL_SET, 0, __SYSCALL_SSLEN);
+		__unmapself(self->map_base, self->map_size);
+	}
 
-	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);
+	for (;;) __syscall(SYS_exit, 0);
 }
 
-static int rsyscall(int nr, long a, long b, long c, long d, long e, long f)
+void __do_cleanup_push(struct __ptcb *cb)
 {
-	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);
-	}
-
-	/* Handle any lingering signals with no-op */
-	__libc_sigprocmask(SIG_UNBLOCK, &set, 0);
-
-	/* 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);
-
-	if (rs.err) errno = rs.err, ret = -1;
-	else ret = syscall6(nr, a, b, c, d, e, f);
-
-	UNLOCK(&rs.lock);
-	return ret;
+	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;
 }
 
 static int start(void *p)
 {
-	struct pthread *self = 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);
+		}
+		__syscall(SYS_rt_sigprocmask, SIG_SETMASK,
+			self->sigmask, 0, __SYSCALL_SSLEN);
+	}
+	if (self->unblock_cancel)
+		__syscall(SYS_rt_sigprocmask, SIG_UNBLOCK,
+			SIGPT_SET, 0, __SYSCALL_SSLEN);
 	pthread_exit(self->start(self->start_arg));
 	return 0;
 }
 
-int __uniclone(void *, int (*)(), void *);
-
 #define ROUND(x) (((x)+PAGE_SIZE-1)&-PAGE_SIZE)
 
 /* pthread_key_create.c overrides this */
 static const size_t dummy = 0;
 weak_alias(dummy, __pthread_tsd_size);
 
-int pthread_create(pthread_t *res, const pthread_attr_t *attr, void *(*entry)(void *), void *arg)
+static FILE *const 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;
+}
+
+void *__copy_tls(unsigned char *);
+
+int pthread_create(pthread_t *restrict res, const pthread_attr_t *restrict attrp, void *(*entry)(void *), void *restrict arg)
 {
-	static int init;
 	int ret;
 	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);
+	unsigned char *map = 0, *stack = 0, *tsd = 0;
+	unsigned flags = 0x7d8f00;
+	int do_sched = 0;
+	pthread_attr_t attr = {0};
+
+	if (!self) return ENOSYS;
+	if (!libc.threaded) {
+		for (FILE *f=libc.ofl_head; f; f=f->next)
+			init_file_lock(f);
+		init_file_lock(__stdin_used);
+		init_file_lock(__stdout_used);
+		init_file_lock(__stderr_used);
+		libc.threaded = 1;
+	}
+	if (attrp) attr = *attrp;
+
+	__acquire_ptc();
+
+	if (attr._a_stackaddr) {
+		size_t need = libc.tls_size + __pthread_tsd_size;
+		size = attr._a_stacksize + DEFAULT_STACK_SIZE;
+		stack = (void *)(attr._a_stackaddr & -16);
+		/* 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;
+		} else {
+			size = ROUND(need);
+			guard = 0;
+		}
+	} else {
+		guard = ROUND(DEFAULT_GUARD_SIZE + attr._a_guardsize);
+		size = guard + ROUND(DEFAULT_STACK_SIZE + 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)) {
+				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;
+	}
+
+	new = __copy_tls(tsd - libc.tls_size);
 	new->map_base = map;
 	new->map_size = size;
 	new->pid = self->pid;
@@ -192,35 +164,45 @@ int pthread_create(pthread_t *res, const pthread_attr_t *attr, void *(*entry)(vo
 	new->start_arg = arg;
 	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);
+	if (attr._a_detach) {
+		new->detached = 1;
+		flags -= 0x200000;
+	}
+	if (attr._a_sched) {
+		do_sched = new->startlock[0] = 1;
+		__syscall(SYS_rt_sigprocmask, SIG_BLOCK,
+			SIGALL_SET, self->sigmask, __SYSCALL_SSLEN);
+	}
+	new->unblock_cancel = self->cancel;
+	new->canary = self->canary;
 
 	a_inc(&libc.threads_minus_1);
-	ret = __uniclone(stack, start, new);
+	ret = __clone(start, stack, flags, new, &new->tid, TP_ADJ(new), &new->tid);
 
-	a_dec(&rs.blocks);
-	if (rs.lock) __wake(&rs.blocks, 1, 1);
+	__release_ptc();
+
+	if (do_sched) {
+		__syscall(SYS_rt_sigprocmask, SIG_SETMASK,
+			new->sigmask, 0, __SYSCALL_SSLEN);
+	}
 
 	if (ret < 0) {
 		a_dec(&libc.threads_minus_1);
-		munmap(map, size);
+		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;
+	}
+
 	*res = new;
 	return 0;
-}
-
-void pthread_exit(void *result)
-{
-	struct pthread *self = pthread_self();
-	self->result = result;
-	docancel(self);
+fail:
+	__release_ptc();
+	return EAGAIN;
 }