fix stale locks left behind when pthread_create fails
[musl] / src / thread / pthread_create.c
index 87bf816..d11dcfa 100644 (file)
@@ -1,14 +1,15 @@
 #include "pthread_impl.h"
 #include "stdio_impl.h"
+#include <sys/mman.h>
 
 static void dummy_0()
 {
 }
-weak_alias(dummy_0, __synccall_lock);
-weak_alias(dummy_0, __synccall_unlock);
+weak_alias(dummy_0, __acquire_ptc);
+weak_alias(dummy_0, __release_ptc);
 weak_alias(dummy_0, __pthread_tsd_run_dtors);
 
-void pthread_exit(void *result)
+_Noreturn void pthread_exit(void *result)
 {
        pthread_t self = pthread_self();
        int n;
@@ -24,45 +25,55 @@ void pthread_exit(void *result)
 
        __pthread_tsd_run_dtors();
 
-       __lock(&self->exitlock);
+       __lock(self->exitlock);
 
        /* Mark this thread dead before decrementing count */
-       __lock(&self->killlock);
+       __lock(self->killlock);
        self->dead = 1;
-       a_store(&self->killlock, 0);
+       __unlock(self->killlock);
 
        do n = libc.threads_minus_1;
        while (n && a_cas(&libc.threads_minus_1, n, n-1)!=n);
        if (!n) exit(0);
 
        if (self->detached && self->map_base) {
-               __syscall(SYS_rt_sigprocmask, SIG_BLOCK, (uint64_t[]){-1},0,8);
+               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);
        }
 
-       __syscall(SYS_exit, 0);
+       for (;;) __syscall(SYS_exit, 0);
 }
 
-void __do_cleanup_push(struct __ptcb *cb, void (*f)(void *), void *x)
+void __do_cleanup_push(struct __ptcb *cb)
 {
        struct pthread *self = pthread_self();
-       cb->__f = f;
-       cb->__x = x;
        cb->__next = self->cancelbuf;
        self->cancelbuf = cb;
 }
 
-void __do_cleanup_pop(struct __ptcb *cb, int run)
+void __do_cleanup_pop(struct __ptcb *cb)
 {
        __pthread_self()->cancelbuf = cb->__next;
-       if (run) cb->__f(cb->__x);
 }
 
 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);
+               }
+               __syscall(SYS_rt_sigprocmask, SIG_SETMASK,
+                       self->sigmask, 0, __SYSCALL_SSLEN);
+       }
        if (self->unblock_cancel)
-               __syscall(SYS_rt_sigprocmask, SIG_UNBLOCK, SIGPT_SET, 0, 8);
+               __syscall(SYS_rt_sigprocmask, SIG_UNBLOCK,
+                       SIGPT_SET, 0, __SYSCALL_SSLEN);
        pthread_exit(self->start(self->start_arg));
        return 0;
 }
@@ -83,13 +94,17 @@ 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)
+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;
-       size_t size = DEFAULT_STACK_SIZE + DEFAULT_GUARD_SIZE;
-       size_t guard = DEFAULT_GUARD_SIZE;
+       size_t size, guard;
        struct pthread *self = pthread_self(), *new;
-       unsigned char *map, *stack, *tsd;
+       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) {
@@ -98,21 +113,49 @@ int pthread_create(pthread_t *res, const pthread_attr_t *attr, void *(*entry)(vo
                init_file_lock(__stdin_used);
                init_file_lock(__stdout_used);
                init_file_lock(__stderr_used);
-               __syscall(SYS_rt_sigprocmask, SIG_UNBLOCK, SIGPT_SET, 0, 8);
                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 (attr) {
-               guard = ROUND(attr->_a_guardsize + DEFAULT_GUARD_SIZE);
-               size = guard + ROUND(attr->_a_stacksize + DEFAULT_STACK_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;
        }
-       size += __pthread_tsd_size;
-       map = mmap(0, size, PROT_READ|PROT_WRITE|PROT_EXEC, MAP_PRIVATE|MAP_ANON, -1, 0);
-       if (map == MAP_FAILED) return EAGAIN;
-       if (guard) mprotect(map, guard, PROT_NONE);
 
-       tsd = map + size - __pthread_tsd_size;
-       new = (void *)(tsd - sizeof *new - PAGE_SIZE%sizeof *new);
+       new = __copy_tls(tsd - libc.tls_size);
        new->map_base = map;
        new->map_size = size;
        new->pid = self->pid;
@@ -121,22 +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;
-       if (attr) new->detached = attr->_a_detach;
+       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;
-       stack = (void *)new;
-
-       __synccall_lock();
+       new->canary = self->canary;
 
        a_inc(&libc.threads_minus_1);
-       ret = __clone(start, stack, 0x7d8f00, new, &new->tid, new, &new->tid);
+       ret = __clone(start, stack, flags, new, &new->tid, TP_ADJ(new), &new->tid);
 
-       __synccall_unlock();
+       __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;
+fail:
+       __release_ptc();
+       return EAGAIN;
 }