add comments on some of the pthread_exit logic

[musl] / src / thread / pthread_create.c

diff --git a/src/thread/pthread_create.c b/src/thread/pthread_create.c
index 6163f50..6a37ee9 100644 (file)
--- a/src/thread/pthread_create.c
+++ b/src/thread/pthread_create.c
@@ -32,15 +32,28 @@ _Noreturn void pthread_exit(void *result)
        self->dead = 1;
        __unlock(self->killlock);
 
+       /* 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. */
+       __syscall(SYS_rt_sigprocmask, SIG_BLOCK, SIGALL_SET, 0, _NSIG/8);
+
        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) {
-               if (self->detached == 2)
-                       __syscall(SYS_set_tid_address, 0);
-               __syscall(SYS_rt_sigprocmask, SIG_BLOCK,
-                       SIGALL_SET, 0, _NSIG/8);
+               /* 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);
+
+               /* The following call unmaps the thread's stack mapping
+                * and then exits without touching the stack. */
                __unmapself(self->map_base, self->map_size);
        }
 
@@ -123,7 +136,7 @@ int pthread_create(pthread_t *restrict res, const pthread_attr_t *restrict attrp
                size_t need = libc.tls_size + __pthread_tsd_size;
                size = attr._a_stacksize + DEFAULT_STACK_SIZE;
                stack = (void *)(attr._a_stackaddr & -16);
-               stack_limit = attr._a_stackaddr - size;
+               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. */