add comments on some of the pthread_exit logic
authorRich Felker <dalias@aerifal.cx>
Fri, 26 Apr 2013 20:04:30 +0000 (16:04 -0400)
committerRich Felker <dalias@aerifal.cx>
Fri, 26 Apr 2013 20:04:30 +0000 (16:04 -0400)
src/thread/pthread_create.c

index 0cd2d6c29260d5f92f0b13f5ca8e82f596be770c..6a37ee9be5c760a89923a394276406a027a45e22 100644 (file)
@@ -32,6 +32,10 @@ _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;
@@ -39,8 +43,17 @@ _Noreturn void pthread_exit(void *result)
        if (!n) exit(0);
 
        if (self->detached && self->map_base) {
-               if (self->detached == 2)
-                       __syscall(SYS_set_tid_address, 0);
+               /* 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);
        }