about summary refs log tree commit diff
path: root/linuxthreads/pthread.c
diff options
context:
space:
mode:
Diffstat (limited to 'linuxthreads/pthread.c')
-rw-r--r--linuxthreads/pthread.c1407
1 files changed, 1407 insertions, 0 deletions
diff --git a/linuxthreads/pthread.c b/linuxthreads/pthread.c
new file mode 100644
index 0000000000..24f0eb02b0
--- /dev/null
+++ b/linuxthreads/pthread.c
@@ -0,0 +1,1407 @@
+
+/* Linuxthreads - a simple clone()-based implementation of Posix        */
+/* threads for Linux.                                                   */
+/* Copyright (C) 1996 Xavier Leroy (Xavier.Leroy@inria.fr)              */
+/*                                                                      */
+/* This program is free software; you can redistribute it and/or        */
+/* modify it under the terms of the GNU Library General Public License  */
+/* as published by the Free Software Foundation; either version 2       */
+/* of the License, or (at your option) any later version.               */
+/*                                                                      */
+/* This program is distributed in the hope that it will be useful,      */
+/* but WITHOUT ANY WARRANTY; without even the implied warranty of       */
+/* MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the        */
+/* GNU Library General Public License for more details.                 */
+
+/* Thread creation, initialization, and basic low-level routines */
+
+#include <errno.h>
+#include <stddef.h>
+#include <stdio.h>
+#include <stdlib.h>
+#include <string.h>
+#include <unistd.h>
+#include <fcntl.h>
+#include <sys/wait.h>
+#include <sys/resource.h>
+#include <sys/time.h>
+#include <shlib-compat.h>
+#include "pthread.h"
+#include "internals.h"
+#include "spinlock.h"
+#include "restart.h"
+#include "smp.h"
+#include <ldsodefs.h>
+#include <tls.h>
+#include <version.h>
+#include <not-cancel.h>
+
+/* Sanity check.  */
+#if !defined __SIGRTMIN || (__SIGRTMAX - __SIGRTMIN) < 3
+# error "This must not happen"
+#endif
+
+#if !(USE_TLS && HAVE___THREAD)
+/* These variables are used by the setup code.  */
+extern int _errno;
+extern int _h_errno;
+
+/* We need the global/static resolver state here.  */
+# include <resolv.h>
+# undef _res
+
+extern struct __res_state _res;
+#endif
+
+#ifdef USE_TLS
+
+/* We need only a few variables.  */
+#define manager_thread __pthread_manager_threadp
+pthread_descr __pthread_manager_threadp attribute_hidden;
+
+#else
+
+/* Descriptor of the initial thread */
+
+struct _pthread_descr_struct __pthread_initial_thread = {
+  .p_header.data.self = &__pthread_initial_thread,
+  .p_nextlive = &__pthread_initial_thread,
+  .p_prevlive = &__pthread_initial_thread,
+  .p_tid = PTHREAD_THREADS_MAX,
+  .p_lock = &__pthread_handles[0].h_lock,
+  .p_start_args = PTHREAD_START_ARGS_INITIALIZER(NULL),
+#if !(USE_TLS && HAVE___THREAD)
+  .p_errnop = &_errno,
+  .p_h_errnop = &_h_errno,
+  .p_resp = &_res,
+#endif
+  .p_userstack = 1,
+  .p_resume_count = __ATOMIC_INITIALIZER,
+  .p_alloca_cutoff = __MAX_ALLOCA_CUTOFF
+};
+
+/* Descriptor of the manager thread; none of this is used but the error
+   variables, the p_pid and p_priority fields,
+   and the address for identification.  */
+
+#define manager_thread (&__pthread_manager_thread)
+struct _pthread_descr_struct __pthread_manager_thread = {
+  .p_header.data.self = &__pthread_manager_thread,
+  .p_header.data.multiple_threads = 1,
+  .p_lock = &__pthread_handles[1].h_lock,
+  .p_start_args = PTHREAD_START_ARGS_INITIALIZER(__pthread_manager),
+#if !(USE_TLS && HAVE___THREAD)
+  .p_errnop = &__pthread_manager_thread.p_errno,
+#endif
+  .p_nr = 1,
+  .p_resume_count = __ATOMIC_INITIALIZER,
+  .p_alloca_cutoff = PTHREAD_STACK_MIN / 4
+};
+#endif
+
+/* Pointer to the main thread (the father of the thread manager thread) */
+/* Originally, this is the initial thread, but this changes after fork() */
+
+#ifdef USE_TLS
+pthread_descr __pthread_main_thread;
+#else
+pthread_descr __pthread_main_thread = &__pthread_initial_thread;
+#endif
+
+/* Limit between the stack of the initial thread (above) and the
+   stacks of other threads (below). Aligned on a STACK_SIZE boundary. */
+
+char *__pthread_initial_thread_bos;
+
+/* File descriptor for sending requests to the thread manager. */
+/* Initially -1, meaning that the thread manager is not running. */
+
+int __pthread_manager_request = -1;
+
+int __pthread_multiple_threads attribute_hidden;
+
+/* Other end of the pipe for sending requests to the thread manager. */
+
+int __pthread_manager_reader;
+
+/* Limits of the thread manager stack */
+
+char *__pthread_manager_thread_bos;
+char *__pthread_manager_thread_tos;
+
+/* For process-wide exit() */
+
+int __pthread_exit_requested;
+int __pthread_exit_code;
+
+/* Maximum stack size.  */
+size_t __pthread_max_stacksize;
+
+/* Nozero if the machine has more than one processor.  */
+int __pthread_smp_kernel;
+
+
+#if !__ASSUME_REALTIME_SIGNALS
+/* Pointers that select new or old suspend/resume functions
+   based on availability of rt signals. */
+
+void (*__pthread_restart)(pthread_descr) = __pthread_restart_old;
+void (*__pthread_suspend)(pthread_descr) = __pthread_suspend_old;
+int (*__pthread_timedsuspend)(pthread_descr, const struct timespec *) = __pthread_timedsuspend_old;
+#endif	/* __ASSUME_REALTIME_SIGNALS */
+
+/* Communicate relevant LinuxThreads constants to gdb */
+
+const int __pthread_threads_max = PTHREAD_THREADS_MAX;
+const int __pthread_sizeof_handle = sizeof(struct pthread_handle_struct);
+const int __pthread_offsetof_descr = offsetof(struct pthread_handle_struct,
+                                              h_descr);
+const int __pthread_offsetof_pid = offsetof(struct _pthread_descr_struct,
+                                            p_pid);
+const int __linuxthreads_pthread_sizeof_descr
+  = sizeof(struct _pthread_descr_struct);
+
+const int __linuxthreads_initial_report_events;
+
+const char __linuxthreads_version[] = VERSION;
+
+/* Forward declarations */
+
+static void pthread_onexit_process(int retcode, void *arg);
+#ifndef HAVE_Z_NODELETE
+static void pthread_atexit_process(void *arg, int retcode);
+static void pthread_atexit_retcode(void *arg, int retcode);
+#endif
+static void pthread_handle_sigcancel(int sig);
+static void pthread_handle_sigrestart(int sig);
+static void pthread_handle_sigdebug(int sig);
+
+/* Signal numbers used for the communication.
+   In these variables we keep track of the used variables.  If the
+   platform does not support any real-time signals we will define the
+   values to some unreasonable value which will signal failing of all
+   the functions below.  */
+int __pthread_sig_restart = __SIGRTMIN;
+int __pthread_sig_cancel = __SIGRTMIN + 1;
+int __pthread_sig_debug = __SIGRTMIN + 2;
+
+extern int __libc_current_sigrtmin_private (void);
+
+#if !__ASSUME_REALTIME_SIGNALS
+static int rtsigs_initialized;
+
+static void
+init_rtsigs (void)
+{
+  if (rtsigs_initialized)
+    return;
+
+  if (__libc_current_sigrtmin_private () == -1)
+    {
+      __pthread_sig_restart = SIGUSR1;
+      __pthread_sig_cancel = SIGUSR2;
+      __pthread_sig_debug = 0;
+    }
+  else
+    {
+      __pthread_restart = __pthread_restart_new;
+      __pthread_suspend = __pthread_wait_for_restart_signal;
+      __pthread_timedsuspend = __pthread_timedsuspend_new;
+    }
+
+  rtsigs_initialized = 1;
+}
+#endif
+
+
+/* Initialize the pthread library.
+   Initialization is split in two functions:
+   - a constructor function that blocks the __pthread_sig_restart signal
+     (must do this very early, since the program could capture the signal
+      mask with e.g. sigsetjmp before creating the first thread);
+   - a regular function called from pthread_create when needed. */
+
+static void pthread_initialize(void) __attribute__((constructor));
+
+#ifndef HAVE_Z_NODELETE
+extern void *__dso_handle __attribute__ ((weak));
+#endif
+
+
+#if defined USE_TLS && !defined SHARED
+extern void __libc_setup_tls (size_t tcbsize, size_t tcbalign);
+#endif
+
+struct pthread_functions __pthread_functions =
+  {
+#if !(USE_TLS && HAVE___THREAD)
+    .ptr_pthread_internal_tsd_set = __pthread_internal_tsd_set,
+    .ptr_pthread_internal_tsd_get = __pthread_internal_tsd_get,
+    .ptr_pthread_internal_tsd_address = __pthread_internal_tsd_address,
+#endif
+    .ptr_pthread_fork = __pthread_fork,
+    .ptr_pthread_attr_destroy = __pthread_attr_destroy,
+#if SHLIB_COMPAT(libpthread, GLIBC_2_0, GLIBC_2_1)
+    .ptr___pthread_attr_init_2_0 = __pthread_attr_init_2_0,
+#endif
+    .ptr___pthread_attr_init_2_1 = __pthread_attr_init_2_1,
+    .ptr_pthread_attr_getdetachstate = __pthread_attr_getdetachstate,
+    .ptr_pthread_attr_setdetachstate = __pthread_attr_setdetachstate,
+    .ptr_pthread_attr_getinheritsched = __pthread_attr_getinheritsched,
+    .ptr_pthread_attr_setinheritsched = __pthread_attr_setinheritsched,
+    .ptr_pthread_attr_getschedparam = __pthread_attr_getschedparam,
+    .ptr_pthread_attr_setschedparam = __pthread_attr_setschedparam,
+    .ptr_pthread_attr_getschedpolicy = __pthread_attr_getschedpolicy,
+    .ptr_pthread_attr_setschedpolicy = __pthread_attr_setschedpolicy,
+    .ptr_pthread_attr_getscope = __pthread_attr_getscope,
+    .ptr_pthread_attr_setscope = __pthread_attr_setscope,
+    .ptr_pthread_condattr_destroy = __pthread_condattr_destroy,
+    .ptr_pthread_condattr_init = __pthread_condattr_init,
+    .ptr___pthread_cond_broadcast = __pthread_cond_broadcast,
+    .ptr___pthread_cond_destroy = __pthread_cond_destroy,
+    .ptr___pthread_cond_init = __pthread_cond_init,
+    .ptr___pthread_cond_signal = __pthread_cond_signal,
+    .ptr___pthread_cond_wait = __pthread_cond_wait,
+    .ptr___pthread_cond_timedwait = __pthread_cond_timedwait,
+    .ptr_pthread_equal = __pthread_equal,
+    .ptr___pthread_exit = __pthread_exit,
+    .ptr_pthread_getschedparam = __pthread_getschedparam,
+    .ptr_pthread_setschedparam = __pthread_setschedparam,
+    .ptr_pthread_mutex_destroy = __pthread_mutex_destroy,
+    .ptr_pthread_mutex_init = __pthread_mutex_init,
+    .ptr_pthread_mutex_lock = __pthread_mutex_lock,
+    .ptr_pthread_mutex_trylock = __pthread_mutex_trylock,
+    .ptr_pthread_mutex_unlock = __pthread_mutex_unlock,
+    .ptr_pthread_self = __pthread_self,
+    .ptr_pthread_setcancelstate = __pthread_setcancelstate,
+    .ptr_pthread_setcanceltype = __pthread_setcanceltype,
+    .ptr_pthread_do_exit = __pthread_do_exit,
+    .ptr_pthread_thread_self = __pthread_thread_self,
+    .ptr_pthread_cleanup_upto = __pthread_cleanup_upto,
+    .ptr_pthread_sigaction = __pthread_sigaction,
+    .ptr_pthread_sigwait = __pthread_sigwait,
+    .ptr_pthread_raise = __pthread_raise,
+    .ptr__pthread_cleanup_push = _pthread_cleanup_push,
+    .ptr__pthread_cleanup_pop = _pthread_cleanup_pop
+  };
+#ifdef SHARED
+# define ptr_pthread_functions &__pthread_functions
+#else
+# define ptr_pthread_functions NULL
+#endif
+
+static int *__libc_multiple_threads_ptr;
+
+/* Do some minimal initialization which has to be done during the
+   startup of the C library.  */
+void
+__pthread_initialize_minimal(void)
+{
+#ifdef USE_TLS
+  pthread_descr self;
+
+  /* First of all init __pthread_handles[0] and [1] if needed.  */
+# if __LT_SPINLOCK_INIT != 0
+  __pthread_handles[0].h_lock = __LOCK_INITIALIZER;
+  __pthread_handles[1].h_lock = __LOCK_INITIALIZER;
+# endif
+# ifndef SHARED
+  /* Unlike in the dynamically linked case the dynamic linker has not
+     taken care of initializing the TLS data structures.  */
+  __libc_setup_tls (TLS_TCB_SIZE, TLS_TCB_ALIGN);
+# elif !USE___THREAD
+  if (__builtin_expect (GL(dl_tls_dtv_slotinfo_list) == NULL, 0))
+    {
+      tcbhead_t *tcbp;
+
+      /* There is no actual TLS being used, so the thread register
+	 was not initialized in the dynamic linker.  */
+
+      /* We need to install special hooks so that the malloc and memalign
+	 calls in _dl_tls_setup and _dl_allocate_tls won't cause full
+	 malloc initialization that will try to set up its thread state.  */
+
+      extern void __libc_malloc_pthread_startup (bool first_time);
+      __libc_malloc_pthread_startup (true);
+
+      if (__builtin_expect (_dl_tls_setup (), 0)
+	  || __builtin_expect ((tcbp = _dl_allocate_tls (NULL)) == NULL, 0))
+	{
+	  static const char msg[] = "\
+cannot allocate TLS data structures for initial thread\n";
+	  TEMP_FAILURE_RETRY (write_not_cancel (STDERR_FILENO,
+						msg, sizeof msg - 1));
+	  abort ();
+	}
+      const char *lossage = TLS_INIT_TP (tcbp, 0);
+      if (__builtin_expect (lossage != NULL, 0))
+	{
+	  static const char msg[] = "cannot set up thread-local storage: ";
+	  const char nl = '\n';
+	  TEMP_FAILURE_RETRY (write_not_cancel (STDERR_FILENO,
+						msg, sizeof msg - 1));
+	  TEMP_FAILURE_RETRY (write_not_cancel (STDERR_FILENO,
+						lossage, strlen (lossage)));
+	  TEMP_FAILURE_RETRY (write_not_cancel (STDERR_FILENO, &nl, 1));
+	}
+
+      /* Though it was allocated with libc's malloc, that was done without
+	 the user's __malloc_hook installed.  A later realloc that uses
+	 the hooks might not work with that block from the plain malloc.
+	 So we record this block as unfreeable just as the dynamic linker
+	 does when it allocates the DTV before the libc malloc exists.  */
+      GL(dl_initial_dtv) = GET_DTV (tcbp);
+
+      __libc_malloc_pthread_startup (false);
+    }
+# endif
+
+  self = THREAD_SELF;
+
+  /* The memory for the thread descriptor was allocated elsewhere as
+     part of the TLS allocation.  We have to initialize the data
+     structure by hand.  This initialization must mirror the struct
+     definition above.  */
+  self->p_nextlive = self->p_prevlive = self;
+  self->p_tid = PTHREAD_THREADS_MAX;
+  self->p_lock = &__pthread_handles[0].h_lock;
+# ifndef HAVE___THREAD
+  self->p_errnop = &_errno;
+  self->p_h_errnop = &_h_errno;
+# endif
+  /* self->p_start_args need not be initialized, it's all zero.  */
+  self->p_userstack = 1;
+# if __LT_SPINLOCK_INIT != 0
+  self->p_resume_count = (struct pthread_atomic) __ATOMIC_INITIALIZER;
+# endif
+  self->p_alloca_cutoff = __MAX_ALLOCA_CUTOFF;
+
+  /* Another variable which points to the thread descriptor.  */
+  __pthread_main_thread = self;
+
+  /* And fill in the pointer the the thread __pthread_handles array.  */
+  __pthread_handles[0].h_descr = self;
+
+#else  /* USE_TLS */
+
+  /* First of all init __pthread_handles[0] and [1].  */
+# if __LT_SPINLOCK_INIT != 0
+  __pthread_handles[0].h_lock = __LOCK_INITIALIZER;
+  __pthread_handles[1].h_lock = __LOCK_INITIALIZER;
+# endif
+  __pthread_handles[0].h_descr = &__pthread_initial_thread;
+  __pthread_handles[1].h_descr = &__pthread_manager_thread;
+
+  /* If we have special thread_self processing, initialize that for the
+     main thread now.  */
+# ifdef INIT_THREAD_SELF
+  INIT_THREAD_SELF(&__pthread_initial_thread, 0);
+# endif
+#endif
+
+#if HP_TIMING_AVAIL
+# ifdef USE_TLS
+  self->p_cpuclock_offset = GL(dl_cpuclock_offset);
+# else
+  __pthread_initial_thread.p_cpuclock_offset = GL(dl_cpuclock_offset);
+# endif
+#endif
+
+  __libc_multiple_threads_ptr = __libc_pthread_init (ptr_pthread_functions);
+}
+
+
+void
+__pthread_init_max_stacksize(void)
+{
+  struct rlimit limit;
+  size_t max_stack;
+
+  getrlimit(RLIMIT_STACK, &limit);
+#ifdef FLOATING_STACKS
+  if (limit.rlim_cur == RLIM_INFINITY)
+    limit.rlim_cur = ARCH_STACK_MAX_SIZE;
+# ifdef NEED_SEPARATE_REGISTER_STACK
+  max_stack = limit.rlim_cur / 2;
+# else
+  max_stack = limit.rlim_cur;
+# endif
+#else
+  /* Play with the stack size limit to make sure that no stack ever grows
+     beyond STACK_SIZE minus one page (to act as a guard page). */
+# ifdef NEED_SEPARATE_REGISTER_STACK
+  /* STACK_SIZE bytes hold both the main stack and register backing
+     store. The rlimit value applies to each individually.  */
+  max_stack = STACK_SIZE/2 - __getpagesize ();
+# else
+  max_stack = STACK_SIZE - __getpagesize();
+# endif
+  if (limit.rlim_cur > max_stack) {
+    limit.rlim_cur = max_stack;
+    setrlimit(RLIMIT_STACK, &limit);
+  }
+#endif
+  __pthread_max_stacksize = max_stack;
+  if (max_stack / 4 < __MAX_ALLOCA_CUTOFF)
+    {
+#ifdef USE_TLS
+      pthread_descr self = THREAD_SELF;
+      self->p_alloca_cutoff = max_stack / 4;
+#else
+      __pthread_initial_thread.p_alloca_cutoff = max_stack / 4;
+#endif
+    }
+}
+
+#ifdef SHARED
+# if USE___THREAD
+/* When using __thread for this, we do it in libc so as not
+   to give libpthread its own TLS segment just for this.  */
+extern void **__libc_dl_error_tsd (void) __attribute__ ((const));
+# else
+static void ** __attribute__ ((const))
+__libc_dl_error_tsd (void)
+{
+  return &thread_self ()->p_libc_specific[_LIBC_TSD_KEY_DL_ERROR];
+}
+# endif
+#endif
+
+#ifdef USE_TLS
+static inline void __attribute__((always_inline))
+init_one_static_tls (pthread_descr descr, struct link_map *map)
+{
+# if TLS_TCB_AT_TP
+  dtv_t *dtv = GET_DTV (descr);
+  void *dest = (char *) descr - map->l_tls_offset;
+# elif TLS_DTV_AT_TP
+  dtv_t *dtv = GET_DTV ((pthread_descr) ((char *) descr + TLS_PRE_TCB_SIZE));
+  void *dest = (char *) descr + map->l_tls_offset + TLS_PRE_TCB_SIZE;
+# else
+#  error "Either TLS_TCB_AT_TP or TLS_DTV_AT_TP must be defined"
+# endif
+
+  /* Fill in the DTV slot so that a later LD/GD access will find it.  */
+  dtv[map->l_tls_modid].pointer = dest;
+
+  /* Initialize the memory.  */
+  memset (__mempcpy (dest, map->l_tls_initimage, map->l_tls_initimage_size),
+	  '\0', map->l_tls_blocksize - map->l_tls_initimage_size);
+}
+
+static void
+__pthread_init_static_tls (struct link_map *map)
+{
+  size_t i;
+
+  for (i = 0; i < PTHREAD_THREADS_MAX; ++i)
+    if (__pthread_handles[i].h_descr != NULL && i != 1)
+      {
+        __pthread_lock (&__pthread_handles[i].h_lock, NULL);
+	if (__pthread_handles[i].h_descr != NULL)
+	  init_one_static_tls (__pthread_handles[i].h_descr, map);
+        __pthread_unlock (&__pthread_handles[i].h_lock);
+      }
+}
+#endif
+
+static void pthread_initialize(void)
+{
+  struct sigaction sa;
+  sigset_t mask;
+
+  /* If already done (e.g. by a constructor called earlier!), bail out */
+  if (__pthread_initial_thread_bos != NULL) return;
+#ifdef TEST_FOR_COMPARE_AND_SWAP
+  /* Test if compare-and-swap is available */
+  __pthread_has_cas = compare_and_swap_is_available();
+#endif
+#ifdef FLOATING_STACKS
+  /* We don't need to know the bottom of the stack.  Give the pointer some
+     value to signal that initialization happened.  */
+  __pthread_initial_thread_bos = (void *) -1l;
+#else
+  /* Determine stack size limits .  */
+  __pthread_init_max_stacksize ();
+# ifdef _STACK_GROWS_UP
+  /* The initial thread already has all the stack it needs */
+  __pthread_initial_thread_bos = (char *)
+    ((long)CURRENT_STACK_FRAME &~ (STACK_SIZE - 1));
+# else
+  /* For the initial stack, reserve at least STACK_SIZE bytes of stack
+     below the current stack address, and align that on a
+     STACK_SIZE boundary. */
+  __pthread_initial_thread_bos =
+    (char *)(((long)CURRENT_STACK_FRAME - 2 * STACK_SIZE) & ~(STACK_SIZE - 1));
+# endif
+#endif
+#ifdef USE_TLS
+  /* Update the descriptor for the initial thread. */
+  THREAD_SETMEM (((pthread_descr) NULL), p_pid, __getpid());
+# ifndef HAVE___THREAD
+  /* Likewise for the resolver state _res.  */
+  THREAD_SETMEM (((pthread_descr) NULL), p_resp, &_res);
+# endif
+#else
+  /* Update the descriptor for the initial thread. */
+  __pthread_initial_thread.p_pid = __getpid();
+  /* Likewise for the resolver state _res.  */
+  __pthread_initial_thread.p_resp = &_res;
+#endif
+#if !__ASSUME_REALTIME_SIGNALS
+  /* Initialize real-time signals. */
+  init_rtsigs ();
+#endif
+  /* Setup signal handlers for the initial thread.
+     Since signal handlers are shared between threads, these settings
+     will be inherited by all other threads. */
+  sa.sa_handler = pthread_handle_sigrestart;
+  sigemptyset(&sa.sa_mask);
+  sa.sa_flags = 0;
+  __libc_sigaction(__pthread_sig_restart, &sa, NULL);
+  sa.sa_handler = pthread_handle_sigcancel;
+  sigaddset(&sa.sa_mask, __pthread_sig_restart);
+  // sa.sa_flags = 0;
+  __libc_sigaction(__pthread_sig_cancel, &sa, NULL);
+  if (__pthread_sig_debug > 0) {
+    sa.sa_handler = pthread_handle_sigdebug;
+    sigemptyset(&sa.sa_mask);
+    // sa.sa_flags = 0;
+    __libc_sigaction(__pthread_sig_debug, &sa, NULL);
+  }
+  /* Initially, block __pthread_sig_restart. Will be unblocked on demand. */
+  sigemptyset(&mask);
+  sigaddset(&mask, __pthread_sig_restart);
+  sigprocmask(SIG_BLOCK, &mask, NULL);
+  /* And unblock __pthread_sig_cancel if it has been blocked. */
+  sigdelset(&mask, __pthread_sig_restart);
+  sigaddset(&mask, __pthread_sig_cancel);
+  sigprocmask(SIG_UNBLOCK, &mask, NULL);
+  /* Register an exit function to kill all other threads. */
+  /* Do it early so that user-registered atexit functions are called
+     before pthread_*exit_process. */
+#ifndef HAVE_Z_NODELETE
+  if (__builtin_expect (&__dso_handle != NULL, 1))
+    __cxa_atexit ((void (*) (void *)) pthread_atexit_process, NULL,
+		  __dso_handle);
+  else
+#endif
+    __on_exit (pthread_onexit_process, NULL);
+  /* How many processors.  */
+  __pthread_smp_kernel = is_smp_system ();
+
+#ifdef SHARED
+  /* Transfer the old value from the dynamic linker's internal location.  */
+  *__libc_dl_error_tsd () = *(*GL(dl_error_catch_tsd)) ();
+  GL(dl_error_catch_tsd) = &__libc_dl_error_tsd;
+
+  /* Make __rtld_lock_{,un}lock_recursive use pthread_mutex_{,un}lock,
+     keep the lock count from the ld.so implementation.  */
+  GL(dl_rtld_lock_recursive) = (void *) __pthread_mutex_lock;
+  GL(dl_rtld_unlock_recursive) = (void *) __pthread_mutex_unlock;
+  unsigned int rtld_lock_count = GL(dl_load_lock).mutex.__m_count;
+  GL(dl_load_lock).mutex.__m_count = 0;
+  while (rtld_lock_count-- > 0)
+    __pthread_mutex_lock (&GL(dl_load_lock).mutex);
+#endif
+
+#ifdef USE_TLS
+  GL(dl_init_static_tls) = &__pthread_init_static_tls;
+#endif
+}
+
+void __pthread_initialize(void)
+{
+  pthread_initialize();
+}
+
+int __pthread_initialize_manager(void)
+{
+  int manager_pipe[2];
+  int pid;
+  struct pthread_request request;
+  int report_events;
+  pthread_descr mgr;
+#ifdef USE_TLS
+  tcbhead_t *tcbp;
+#endif
+
+  __pthread_multiple_threads = 1;
+#if TLS_MULTIPLE_THREADS_IN_TCB || !defined USE_TLS || !TLS_DTV_AT_TP
+  __pthread_main_thread->p_multiple_threads = 1;
+#endif
+  *__libc_multiple_threads_ptr = 1;
+
+#ifndef HAVE_Z_NODELETE
+  if (__builtin_expect (&__dso_handle != NULL, 1))
+    __cxa_atexit ((void (*) (void *)) pthread_atexit_retcode, NULL,
+		  __dso_handle);
+#endif
+
+  if (__pthread_max_stacksize == 0)
+    __pthread_init_max_stacksize ();
+  /* If basic initialization not done yet (e.g. we're called from a
+     constructor run before our constructor), do it now */
+  if (__pthread_initial_thread_bos == NULL) pthread_initialize();
+  /* Setup stack for thread manager */
+  __pthread_manager_thread_bos = malloc(THREAD_MANAGER_STACK_SIZE);
+  if (__pthread_manager_thread_bos == NULL) return -1;
+  __pthread_manager_thread_tos =
+    __pthread_manager_thread_bos + THREAD_MANAGER_STACK_SIZE;
+  /* Setup pipe to communicate with thread manager */
+  if (pipe(manager_pipe) == -1) {
+    free(__pthread_manager_thread_bos);
+    return -1;
+  }
+
+#ifdef USE_TLS
+  /* Allocate memory for the thread descriptor and the dtv.  */
+  tcbp = _dl_allocate_tls (NULL);
+  if (tcbp == NULL) {
+    free(__pthread_manager_thread_bos);
+    close_not_cancel(manager_pipe[0]);
+    close_not_cancel(manager_pipe[1]);
+    return -1;
+  }
+
+# if TLS_TCB_AT_TP
+  mgr = (pthread_descr) tcbp;
+# elif TLS_DTV_AT_TP
+  /* pthread_descr is located right below tcbhead_t which _dl_allocate_tls
+     returns.  */
+  mgr = (pthread_descr) ((char *) tcbp - TLS_PRE_TCB_SIZE);
+# endif
+  __pthread_handles[1].h_descr = manager_thread = mgr;
+
+  /* Initialize the descriptor.  */
+#if !defined USE_TLS || !TLS_DTV_AT_TP
+  mgr->p_header.data.tcb = tcbp;
+  mgr->p_header.data.self = mgr;
+  mgr->p_header.data.multiple_threads = 1;
+#elif TLS_MULTIPLE_THREADS_IN_TCB
+  mgr->p_multiple_threads = 1;
+#endif
+  mgr->p_lock = &__pthread_handles[1].h_lock;
+# ifndef HAVE___THREAD
+  mgr->p_errnop = &mgr->p_errno;
+# endif
+  mgr->p_start_args = (struct pthread_start_args) PTHREAD_START_ARGS_INITIALIZER(__pthread_manager);
+  mgr->p_nr = 1;
+# if __LT_SPINLOCK_INIT != 0
+  self->p_resume_count = (struct pthread_atomic) __ATOMIC_INITIALIZER;
+# endif
+  mgr->p_alloca_cutoff = PTHREAD_STACK_MIN / 4;
+#else
+  mgr = &__pthread_manager_thread;
+#endif
+
+  __pthread_manager_request = manager_pipe[1]; /* writing end */
+  __pthread_manager_reader = manager_pipe[0]; /* reading end */
+
+  /* Start the thread manager */
+  pid = 0;
+#ifdef USE_TLS
+  if (__linuxthreads_initial_report_events != 0)
+    THREAD_SETMEM (((pthread_descr) NULL), p_report_events,
+		   __linuxthreads_initial_report_events);
+  report_events = THREAD_GETMEM (((pthread_descr) NULL), p_report_events);
+#else
+  if (__linuxthreads_initial_report_events != 0)
+    __pthread_initial_thread.p_report_events
+      = __linuxthreads_initial_report_events;
+  report_events = __pthread_initial_thread.p_report_events;
+#endif
+  if (__builtin_expect (report_events, 0))
+    {
+      /* It's a bit more complicated.  We have to report the creation of
+	 the manager thread.  */
+      int idx = __td_eventword (TD_CREATE);
+      uint32_t mask = __td_eventmask (TD_CREATE);
+      uint32_t event_bits;
+
+#ifdef USE_TLS
+      event_bits = THREAD_GETMEM_NC (((pthread_descr) NULL),
+				     p_eventbuf.eventmask.event_bits[idx]);
+#else
+      event_bits = __pthread_initial_thread.p_eventbuf.eventmask.event_bits[idx];
+#endif
+
+      if ((mask & (__pthread_threads_events.event_bits[idx] | event_bits))
+	  != 0)
+	{
+	  __pthread_lock(mgr->p_lock, NULL);
+
+#ifdef NEED_SEPARATE_REGISTER_STACK
+	  pid = __clone2(__pthread_manager_event,
+			 (void **) __pthread_manager_thread_bos,
+			 THREAD_MANAGER_STACK_SIZE,
+			 CLONE_VM | CLONE_FS | CLONE_FILES | CLONE_SIGHAND,
+			 mgr);
+#elif _STACK_GROWS_UP
+	  pid = __clone(__pthread_manager_event,
+			(void **) __pthread_manager_thread_bos,
+			CLONE_VM | CLONE_FS | CLONE_FILES | CLONE_SIGHAND,
+			mgr);
+#else
+	  pid = __clone(__pthread_manager_event,
+			(void **) __pthread_manager_thread_tos,
+			CLONE_VM | CLONE_FS | CLONE_FILES | CLONE_SIGHAND,
+			mgr);
+#endif
+
+	  if (pid != -1)
+	    {
+	      /* Now fill in the information about the new thread in
+	         the newly created thread's data structure.  We cannot let
+	         the new thread do this since we don't know whether it was
+	         already scheduled when we send the event.  */
+	      mgr->p_eventbuf.eventdata = mgr;
+	      mgr->p_eventbuf.eventnum = TD_CREATE;
+	      __pthread_last_event = mgr;
+	      mgr->p_tid = 2* PTHREAD_THREADS_MAX + 1;
+	      mgr->p_pid = pid;
+
+	      /* Now call the function which signals the event.  */
+	      __linuxthreads_create_event ();
+	    }
+
+	  /* Now restart the thread.  */
+	  __pthread_unlock(mgr->p_lock);
+	}
+    }
+
+  if (__builtin_expect (pid, 0) == 0)
+    {
+#ifdef NEED_SEPARATE_REGISTER_STACK
+      pid = __clone2(__pthread_manager, (void **) __pthread_manager_thread_bos,
+		     THREAD_MANAGER_STACK_SIZE,
+		     CLONE_VM | CLONE_FS | CLONE_FILES | CLONE_SIGHAND, mgr);
+#elif _STACK_GROWS_UP
+      pid = __clone(__pthread_manager, (void **) __pthread_manager_thread_bos,
+		    CLONE_VM | CLONE_FS | CLONE_FILES | CLONE_SIGHAND, mgr);
+#else
+      pid = __clone(__pthread_manager, (void **) __pthread_manager_thread_tos,
+		    CLONE_VM | CLONE_FS | CLONE_FILES | CLONE_SIGHAND, mgr);
+#endif
+    }
+  if (__builtin_expect (pid, 0) == -1) {
+#ifdef USE_TLS
+    _dl_deallocate_tls (tcbp, true);
+#endif
+    free(__pthread_manager_thread_bos);
+    close_not_cancel(manager_pipe[0]);
+    close_not_cancel(manager_pipe[1]);
+    return -1;
+  }
+  mgr->p_tid = 2* PTHREAD_THREADS_MAX + 1;
+  mgr->p_pid = pid;
+  /* Make gdb aware of new thread manager */
+  if (__builtin_expect (__pthread_threads_debug, 0) && __pthread_sig_debug > 0)
+    {
+      raise(__pthread_sig_debug);
+      /* We suspend ourself and gdb will wake us up when it is
+	 ready to handle us. */
+      __pthread_wait_for_restart_signal(thread_self());
+    }
+  /* Synchronize debugging of the thread manager */
+  request.req_kind = REQ_DEBUG;
+  TEMP_FAILURE_RETRY(write_not_cancel(__pthread_manager_request,
+				      (char *) &request, sizeof(request)));
+  return 0;
+}
+
+/* Thread creation */
+
+int __pthread_create_2_1(pthread_t *thread, const pthread_attr_t *attr,
+			 void * (*start_routine)(void *), void *arg)
+{
+  pthread_descr self = thread_self();
+  struct pthread_request request;
+  int retval;
+  if (__builtin_expect (__pthread_manager_request, 0) < 0) {
+    if (__pthread_initialize_manager() < 0) return EAGAIN;
+  }
+  request.req_thread = self;
+  request.req_kind = REQ_CREATE;
+  request.req_args.create.attr = attr;
+  request.req_args.create.fn = start_routine;
+  request.req_args.create.arg = arg;
+  sigprocmask(SIG_SETMASK, (const sigset_t *) NULL,
+              &request.req_args.create.mask);
+  TEMP_FAILURE_RETRY(write_not_cancel(__pthread_manager_request,
+				      (char *) &request, sizeof(request)));
+  suspend(self);
+  retval = THREAD_GETMEM(self, p_retcode);
+  if (__builtin_expect (retval, 0) == 0)
+    *thread = (pthread_t) THREAD_GETMEM(self, p_retval);
+  return retval;
+}
+
+versioned_symbol (libpthread, __pthread_create_2_1, pthread_create, GLIBC_2_1);
+
+#if SHLIB_COMPAT (libpthread, GLIBC_2_0, GLIBC_2_1)
+
+int __pthread_create_2_0(pthread_t *thread, const pthread_attr_t *attr,
+			 void * (*start_routine)(void *), void *arg)
+{
+  /* The ATTR attribute is not really of type `pthread_attr_t *'.  It has
+     the old size and access to the new members might crash the program.
+     We convert the struct now.  */
+  pthread_attr_t new_attr;
+
+  if (attr != NULL)
+    {
+      size_t ps = __getpagesize ();
+
+      memcpy (&new_attr, attr,
+	      (size_t) &(((pthread_attr_t*)NULL)->__guardsize));
+      new_attr.__guardsize = ps;
+      new_attr.__stackaddr_set = 0;
+      new_attr.__stackaddr = NULL;
+      new_attr.__stacksize = STACK_SIZE - ps;
+      attr = &new_attr;
+    }
+  return __pthread_create_2_1 (thread, attr, start_routine, arg);
+}
+compat_symbol (libpthread, __pthread_create_2_0, pthread_create, GLIBC_2_0);
+#endif
+
+/* Simple operations on thread identifiers */
+
+pthread_descr __pthread_thread_self(void)
+{
+  return thread_self();
+}
+
+pthread_t __pthread_self(void)
+{
+  pthread_descr self = thread_self();
+  return THREAD_GETMEM(self, p_tid);
+}
+strong_alias (__pthread_self, pthread_self);
+
+int __pthread_equal(pthread_t thread1, pthread_t thread2)
+{
+  return thread1 == thread2;
+}
+strong_alias (__pthread_equal, pthread_equal);
+
+/* Helper function for thread_self in the case of user-provided stacks */
+
+#ifndef THREAD_SELF
+
+pthread_descr __pthread_find_self(void)
+{
+  char * sp = CURRENT_STACK_FRAME;
+  pthread_handle h;
+
+  /* __pthread_handles[0] is the initial thread, __pthread_handles[1] is
+     the manager threads handled specially in thread_self(), so start at 2 */
+  h = __pthread_handles + 2;
+# ifdef _STACK_GROWS_UP
+  while (! (sp >= (char *) h->h_descr && sp < h->h_descr->p_guardaddr)) h++;
+# else
+  while (! (sp <= (char *) h->h_descr && sp >= h->h_bottom)) h++;
+# endif
+  return h->h_descr;
+}
+
+#else
+
+pthread_descr __pthread_self_stack(void)
+{
+  char *sp = CURRENT_STACK_FRAME;
+  pthread_handle h;
+
+  if (sp >= __pthread_manager_thread_bos && sp < __pthread_manager_thread_tos)
+    return manager_thread;
+  h = __pthread_handles + 2;
+# ifdef USE_TLS
+#  ifdef _STACK_GROWS_UP
+  while (h->h_descr == NULL
+	 || ! (sp >= h->h_descr->p_stackaddr && sp < h->h_descr->p_guardaddr))
+    h++;
+#  else
+  while (h->h_descr == NULL
+	 || ! (sp <= (char *) h->h_descr->p_stackaddr && sp >= h->h_bottom))
+    h++;
+#  endif
+# else
+#  ifdef _STACK_GROWS_UP
+  while (! (sp >= (char *) h->h_descr && sp < h->h_descr->p_guardaddr))
+    h++;
+#  else
+  while (! (sp <= (char *) h->h_descr && sp >= h->h_bottom))
+    h++;
+#  endif
+# endif
+  return h->h_descr;
+}
+
+#endif
+
+/* Thread scheduling */
+
+int __pthread_setschedparam(pthread_t thread, int policy,
+                            const struct sched_param *param)
+{
+  pthread_handle handle = thread_handle(thread);
+  pthread_descr th;
+
+  __pthread_lock(&handle->h_lock, NULL);
+  if (__builtin_expect (invalid_handle(handle, thread), 0)) {
+    __pthread_unlock(&handle->h_lock);
+    return ESRCH;
+  }
+  th = handle->h_descr;
+  if (__builtin_expect (__sched_setscheduler(th->p_pid, policy, param) == -1,
+			0)) {
+    __pthread_unlock(&handle->h_lock);
+    return errno;
+  }
+  th->p_priority = policy == SCHED_OTHER ? 0 : param->sched_priority;
+  __pthread_unlock(&handle->h_lock);
+  if (__pthread_manager_request >= 0)
+    __pthread_manager_adjust_prio(th->p_priority);
+  return 0;
+}
+strong_alias (__pthread_setschedparam, pthread_setschedparam);
+
+int __pthread_getschedparam(pthread_t thread, int *policy,
+                            struct sched_param *param)
+{
+  pthread_handle handle = thread_handle(thread);
+  int pid, pol;
+
+  __pthread_lock(&handle->h_lock, NULL);
+  if (__builtin_expect (invalid_handle(handle, thread), 0)) {
+    __pthread_unlock(&handle->h_lock);
+    return ESRCH;
+  }
+  pid = handle->h_descr->p_pid;
+  __pthread_unlock(&handle->h_lock);
+  pol = __sched_getscheduler(pid);
+  if (__builtin_expect (pol, 0) == -1) return errno;
+  if (__sched_getparam(pid, param) == -1) return errno;
+  *policy = pol;
+  return 0;
+}
+strong_alias (__pthread_getschedparam, pthread_getschedparam);
+
+int __pthread_yield (void)
+{
+  /* For now this is equivalent with the POSIX call.  */
+  return sched_yield ();
+}
+weak_alias (__pthread_yield, pthread_yield)
+
+/* Process-wide exit() request */
+
+static void pthread_onexit_process(int retcode, void *arg)
+{
+  if (__builtin_expect (__pthread_manager_request, 0) >= 0) {
+    struct pthread_request request;
+    pthread_descr self = thread_self();
+
+    request.req_thread = self;
+    request.req_kind = REQ_PROCESS_EXIT;
+    request.req_args.exit.code = retcode;
+    TEMP_FAILURE_RETRY(write_not_cancel(__pthread_manager_request,
+					(char *) &request, sizeof(request)));
+    suspend(self);
+    /* Main thread should accumulate times for thread manager and its
+       children, so that timings for main thread account for all threads. */
+    if (self == __pthread_main_thread)
+      {
+#ifdef USE_TLS
+	waitpid(manager_thread->p_pid, NULL, __WCLONE);
+#else
+	waitpid(__pthread_manager_thread.p_pid, NULL, __WCLONE);
+#endif
+	/* Since all threads have been asynchronously terminated
+           (possibly holding locks), free cannot be used any more.
+           For mtrace, we'd like to print something though.  */
+	/* #ifdef USE_TLS
+	   tcbhead_t *tcbp = (tcbhead_t *) manager_thread;
+	   # if TLS_DTV_AT_TP
+	   tcbp = (tcbhead_t) ((char *) tcbp + TLS_PRE_TCB_SIZE);
+	   # endif
+	   _dl_deallocate_tls (tcbp, true);
+	   #endif
+	   free (__pthread_manager_thread_bos); */
+	__pthread_manager_thread_bos = __pthread_manager_thread_tos = NULL;
+      }
+  }
+}
+
+#ifndef HAVE_Z_NODELETE
+static int __pthread_atexit_retcode;
+
+static void pthread_atexit_process(void *arg, int retcode)
+{
+  pthread_onexit_process (retcode ?: __pthread_atexit_retcode, arg);
+}
+
+static void pthread_atexit_retcode(void *arg, int retcode)
+{
+  __pthread_atexit_retcode = retcode;
+}
+#endif
+
+/* The handler for the RESTART signal just records the signal received
+   in the thread descriptor, and optionally performs a siglongjmp
+   (for pthread_cond_timedwait). */
+
+static void pthread_handle_sigrestart(int sig)
+{
+  pthread_descr self = check_thread_self();
+  THREAD_SETMEM(self, p_signal, sig);
+  if (THREAD_GETMEM(self, p_signal_jmp) != NULL)
+    siglongjmp(*THREAD_GETMEM(self, p_signal_jmp), 1);
+}
+
+/* The handler for the CANCEL signal checks for cancellation
+   (in asynchronous mode), for process-wide exit and exec requests.
+   For the thread manager thread, redirect the signal to
+   __pthread_manager_sighandler. */
+
+static void pthread_handle_sigcancel(int sig)
+{
+  pthread_descr self = check_thread_self();
+  sigjmp_buf * jmpbuf;
+
+  if (self == manager_thread)
+    {
+      __pthread_manager_sighandler(sig);
+      return;
+    }
+  if (__builtin_expect (__pthread_exit_requested, 0)) {
+    /* Main thread should accumulate times for thread manager and its
+       children, so that timings for main thread account for all threads. */
+    if (self == __pthread_main_thread) {
+#ifdef USE_TLS
+      waitpid(manager_thread->p_pid, NULL, __WCLONE);
+#else
+      waitpid(__pthread_manager_thread.p_pid, NULL, __WCLONE);
+#endif
+    }
+    _exit(__pthread_exit_code);
+  }
+  if (__builtin_expect (THREAD_GETMEM(self, p_canceled), 0)
+      && THREAD_GETMEM(self, p_cancelstate) == PTHREAD_CANCEL_ENABLE) {
+    if (THREAD_GETMEM(self, p_canceltype) == PTHREAD_CANCEL_ASYNCHRONOUS)
+      __pthread_do_exit(PTHREAD_CANCELED, CURRENT_STACK_FRAME);
+    jmpbuf = THREAD_GETMEM(self, p_cancel_jmp);
+    if (jmpbuf != NULL) {
+      THREAD_SETMEM(self, p_cancel_jmp, NULL);
+      siglongjmp(*jmpbuf, 1);
+    }
+  }
+}
+
+/* Handler for the DEBUG signal.
+   The debugging strategy is as follows:
+   On reception of a REQ_DEBUG request (sent by new threads created to
+   the thread manager under debugging mode), the thread manager throws
+   __pthread_sig_debug to itself. The debugger (if active) intercepts
+   this signal, takes into account new threads and continue execution
+   of the thread manager by propagating the signal because it doesn't
+   know what it is specifically done for. In the current implementation,
+   the thread manager simply discards it. */
+
+static void pthread_handle_sigdebug(int sig)
+{
+  /* Nothing */
+}
+
+/* Reset the state of the thread machinery after a fork().
+   Close the pipe used for requests and set the main thread to the forked
+   thread.
+   Notice that we can't free the stack segments, as the forked thread
+   may hold pointers into them. */
+
+void __pthread_reset_main_thread(void)
+{
+  pthread_descr self = thread_self();
+
+  if (__pthread_manager_request != -1) {
+    /* Free the thread manager stack */
+    free(__pthread_manager_thread_bos);
+    __pthread_manager_thread_bos = __pthread_manager_thread_tos = NULL;
+    /* Close the two ends of the pipe */
+    close_not_cancel(__pthread_manager_request);
+    close_not_cancel(__pthread_manager_reader);
+    __pthread_manager_request = __pthread_manager_reader = -1;
+  }
+
+  /* Update the pid of the main thread */
+  THREAD_SETMEM(self, p_pid, __getpid());
+  /* Make the forked thread the main thread */
+  __pthread_main_thread = self;
+  THREAD_SETMEM(self, p_nextlive, self);
+  THREAD_SETMEM(self, p_prevlive, self);
+#if !(USE_TLS && HAVE___THREAD)
+  /* Now this thread modifies the global variables.  */
+  THREAD_SETMEM(self, p_errnop, &_errno);
+  THREAD_SETMEM(self, p_h_errnop, &_h_errno);
+  THREAD_SETMEM(self, p_resp, &_res);
+#endif
+
+#ifndef FLOATING_STACKS
+  /* This is to undo the setrlimit call in __pthread_init_max_stacksize.
+     XXX This can be wrong if the user set the limit during the run.  */
+ {
+   struct rlimit limit;
+   if (getrlimit (RLIMIT_STACK, &limit) == 0
+       && limit.rlim_cur != limit.rlim_max)
+     {
+       limit.rlim_cur = limit.rlim_max;
+       setrlimit(RLIMIT_STACK, &limit);
+     }
+ }
+#endif
+}
+
+/* Process-wide exec() request */
+
+void __pthread_kill_other_threads_np(void)
+{
+  struct sigaction sa;
+  /* Terminate all other threads and thread manager */
+  pthread_onexit_process(0, NULL);
+  /* Make current thread the main thread in case the calling thread
+     changes its mind, does not exec(), and creates new threads instead. */
+  __pthread_reset_main_thread();
+
+  /* Reset the signal handlers behaviour for the signals the
+     implementation uses since this would be passed to the new
+     process.  */
+  sigemptyset(&sa.sa_mask);
+  sa.sa_flags = 0;
+  sa.sa_handler = SIG_DFL;
+  __libc_sigaction(__pthread_sig_restart, &sa, NULL);
+  __libc_sigaction(__pthread_sig_cancel, &sa, NULL);
+  if (__pthread_sig_debug > 0)
+    __libc_sigaction(__pthread_sig_debug, &sa, NULL);
+}
+weak_alias (__pthread_kill_other_threads_np, pthread_kill_other_threads_np)
+
+/* Concurrency symbol level.  */
+static int current_level;
+
+int __pthread_setconcurrency(int level)
+{
+  /* We don't do anything unless we have found a useful interpretation.  */
+  current_level = level;
+  return 0;
+}
+weak_alias (__pthread_setconcurrency, pthread_setconcurrency)
+
+int __pthread_getconcurrency(void)
+{
+  return current_level;
+}
+weak_alias (__pthread_getconcurrency, pthread_getconcurrency)
+
+/* Primitives for controlling thread execution */
+
+void __pthread_wait_for_restart_signal(pthread_descr self)
+{
+  sigset_t mask;
+
+  sigprocmask(SIG_SETMASK, NULL, &mask); /* Get current signal mask */
+  sigdelset(&mask, __pthread_sig_restart); /* Unblock the restart signal */
+  THREAD_SETMEM(self, p_signal, 0);
+  do {
+    __pthread_sigsuspend(&mask);	/* Wait for signal.  Must not be a
+					   cancellation point. */
+  } while (THREAD_GETMEM(self, p_signal) !=__pthread_sig_restart);
+
+  READ_MEMORY_BARRIER(); /* See comment in __pthread_restart_new */
+}
+
+#if !__ASSUME_REALTIME_SIGNALS
+/* The _old variants are for 2.0 and early 2.1 kernels which don't have RT
+   signals.
+   On these kernels, we use SIGUSR1 and SIGUSR2 for restart and cancellation.
+   Since the restart signal does not queue, we use an atomic counter to create
+   queuing semantics. This is needed to resolve a rare race condition in
+   pthread_cond_timedwait_relative. */
+
+void __pthread_restart_old(pthread_descr th)
+{
+  if (atomic_increment(&th->p_resume_count) == -1)
+    kill(th->p_pid, __pthread_sig_restart);
+}
+
+void __pthread_suspend_old(pthread_descr self)
+{
+  if (atomic_decrement(&self->p_resume_count) <= 0)
+    __pthread_wait_for_restart_signal(self);
+}
+
+int
+__pthread_timedsuspend_old(pthread_descr self, const struct timespec *abstime)
+{
+  sigset_t unblock, initial_mask;
+  int was_signalled = 0;
+  sigjmp_buf jmpbuf;
+
+  if (atomic_decrement(&self->p_resume_count) == 0) {
+    /* Set up a longjmp handler for the restart signal, unblock
+       the signal and sleep. */
+
+    if (sigsetjmp(jmpbuf, 1) == 0) {
+      THREAD_SETMEM(self, p_signal_jmp, &jmpbuf);
+      THREAD_SETMEM(self, p_signal, 0);
+      /* Unblock the restart signal */
+      sigemptyset(&unblock);
+      sigaddset(&unblock, __pthread_sig_restart);
+      sigprocmask(SIG_UNBLOCK, &unblock, &initial_mask);
+
+      while (1) {
+	struct timeval now;
+	struct timespec reltime;
+
+	/* Compute a time offset relative to now.  */
+	__gettimeofday (&now, NULL);
+	reltime.tv_nsec = abstime->tv_nsec - now.tv_usec * 1000;
+	reltime.tv_sec = abstime->tv_sec - now.tv_sec;
+	if (reltime.tv_nsec < 0) {
+	  reltime.tv_nsec += 1000000000;
+	  reltime.tv_sec -= 1;
+	}
+
+	/* Sleep for the required duration. If woken by a signal,
+	   resume waiting as required by Single Unix Specification.  */
+	if (reltime.tv_sec < 0 || __libc_nanosleep(&reltime, NULL) == 0)
+	  break;
+      }
+
+      /* Block the restart signal again */
+      sigprocmask(SIG_SETMASK, &initial_mask, NULL);
+      was_signalled = 0;
+    } else {
+      was_signalled = 1;
+    }
+    THREAD_SETMEM(self, p_signal_jmp, NULL);
+  }
+
+  /* Now was_signalled is true if we exited the above code
+     due to the delivery of a restart signal.  In that case,
+     we know we have been dequeued and resumed and that the
+     resume count is balanced.  Otherwise, there are some
+     cases to consider. First, try to bump up the resume count
+     back to zero. If it goes to 1, it means restart() was
+     invoked on this thread. The signal must be consumed
+     and the count bumped down and everything is cool. We
+     can return a 1 to the caller.
+     Otherwise, no restart was delivered yet, so a potential
+     race exists; we return a 0 to the caller which must deal
+     with this race in an appropriate way; for example by
+     atomically removing the thread from consideration for a
+     wakeup---if such a thing fails, it means a restart is
+     being delivered. */
+
+  if (!was_signalled) {
+    if (atomic_increment(&self->p_resume_count) != -1) {
+      __pthread_wait_for_restart_signal(self);
+      atomic_decrement(&self->p_resume_count); /* should be zero now! */
+      /* woke spontaneously and consumed restart signal */
+      return 1;
+    }
+    /* woke spontaneously but did not consume restart---caller must resolve */
+    return 0;
+  }
+  /* woken due to restart signal */
+  return 1;
+}
+#endif /* __ASSUME_REALTIME_SIGNALS */
+
+void __pthread_restart_new(pthread_descr th)
+{
+  /* The barrier is proabably not needed, in which case it still documents
+     our assumptions. The intent is to commit previous writes to shared
+     memory so the woken thread will have a consistent view.  Complementary
+     read barriers are present to the suspend functions. */
+  WRITE_MEMORY_BARRIER();
+  kill(th->p_pid, __pthread_sig_restart);
+}
+
+/* There is no __pthread_suspend_new because it would just
+   be a wasteful wrapper for __pthread_wait_for_restart_signal */
+
+int
+__pthread_timedsuspend_new(pthread_descr self, const struct timespec *abstime)
+{
+  sigset_t unblock, initial_mask;
+  int was_signalled = 0;
+  sigjmp_buf jmpbuf;
+
+  if (sigsetjmp(jmpbuf, 1) == 0) {
+    THREAD_SETMEM(self, p_signal_jmp, &jmpbuf);
+    THREAD_SETMEM(self, p_signal, 0);
+    /* Unblock the restart signal */
+    sigemptyset(&unblock);
+    sigaddset(&unblock, __pthread_sig_restart);
+    sigprocmask(SIG_UNBLOCK, &unblock, &initial_mask);
+
+    while (1) {
+      struct timeval now;
+      struct timespec reltime;
+
+      /* Compute a time offset relative to now.  */
+      __gettimeofday (&now, NULL);
+      reltime.tv_nsec = abstime->tv_nsec - now.tv_usec * 1000;
+      reltime.tv_sec = abstime->tv_sec - now.tv_sec;
+      if (reltime.tv_nsec < 0) {
+	reltime.tv_nsec += 1000000000;
+	reltime.tv_sec -= 1;
+      }
+
+      /* Sleep for the required duration. If woken by a signal,
+	 resume waiting as required by Single Unix Specification.  */
+      if (reltime.tv_sec < 0 || __libc_nanosleep(&reltime, NULL) == 0)
+	break;
+    }
+
+    /* Block the restart signal again */
+    sigprocmask(SIG_SETMASK, &initial_mask, NULL);
+    was_signalled = 0;
+  } else {
+    was_signalled = 1;
+  }
+  THREAD_SETMEM(self, p_signal_jmp, NULL);
+
+  /* Now was_signalled is true if we exited the above code
+     due to the delivery of a restart signal.  In that case,
+     everything is cool. We have been removed from whatever
+     we were waiting on by the other thread, and consumed its signal.
+
+     Otherwise we this thread woke up spontaneously, or due to a signal other
+     than restart. This is an ambiguous case  that must be resolved by
+     the caller; the thread is still eligible for a restart wakeup
+     so there is a race. */
+
+  READ_MEMORY_BARRIER(); /* See comment in __pthread_restart_new */
+  return was_signalled;
+}
+
+
+/* Debugging aid */
+
+#ifdef DEBUG
+#include <stdarg.h>
+
+void __pthread_message(const char * fmt, ...)
+{
+  char buffer[1024];
+  va_list args;
+  sprintf(buffer, "%05d : ", __getpid());
+  va_start(args, fmt);
+  vsnprintf(buffer + 8, sizeof(buffer) - 8, fmt, args);
+  va_end(args);
+  TEMP_FAILURE_RETRY(write_not_cancel(2, buffer, strlen(buffer)));
+}
+
+#endif