2001-06-15 Roland McGrath <roland@frob.com>

	* rt/Makefile (aio-routines, clock-routines, timer-routines): New
	variables, broken out of librt-routines.
	(librt-routines): Set using those.
	(extra-libs, extra-libc-others, tests): Set these unconditionally.
	(distribute): No need for $(librt-map) or $(librt-routines:=.c) here.
	* sysdeps/generic/aio_fsync.c: New file.
	* sysdeps/generic/aio_misc.c: New file.
	* sysdeps/generic/aio_write.c: New file.
	* sysdeps/generic/aio_write64.c: New file.
	* sysdeps/generic/aio_read64.c: New file.
	* sysdeps/generic/aio_read.c: New file.
	* sysdeps/generic/aio_notify.c: New file.
	* sysdeps/generic/lio_listio64.c: New file.
	* sysdeps/generic/lio_listio.c: New file.
	* sysdeps/generic/aio_cancel.c: New file.
	* sysdeps/generic/aio_misc.h: New file.
	* sysdeps/generic/aio_suspend.c: New file.
	* rt/aio_cancel.c: Moved to ...
	* sysdeps/pthread/aio_cancel.c: ... here.
	* rt/aio_fsync.c: Moved to ...
	* sysdeps/pthread/aio_fsync.c: ... here.
	* rt/aio_misc.c: Moved to ...
	* sysdeps/pthread/aio_misc.c: ... here.
	* rt/aio_misc.h: Moved to ...
	* sysdeps/pthread/aio_misc.h: ... here.
	* rt/aio_notify.c: Moved to ...
	* sysdeps/pthread/aio_notify.c: ... here.
	* rt/aio_read.c: Moved to ...
	* sysdeps/pthread/aio_read.c: ... here.
	* rt/aio_read64.c: Moved to ...
	* sysdeps/pthread/aio_read64.c: ... here.
	* rt/aio_suspend.c: Moved to ...
	* sysdeps/pthread/aio_suspend.c: ... here.
	* rt/aio_write.c: Moved to ...
	* sysdeps/pthread/aio_write.c: ... here.
	* rt/aio_write64.c: Moved to ...
	* sysdeps/pthread/aio_write64.c: ... here.
	* rt/lio_listio.c: Moved to ...
	* sysdeps/pthread/lio_listio.c: ... here.
	* rt/lio_listio64.c: Moved to ...
	* sysdeps/pthread/lio_listio64.c: ... here.

12 files changed, 1781 insertions, 0 deletions

diff --git a/sysdeps/pthread/aio_cancel.c b/sysdeps/pthread/aio_cancel.c
new file mode 100644
index 0000000000..c1cfea50c8
--- /dev/null
+++ b/sysdeps/pthread/aio_cancel.c
@@ -0,0 +1,142 @@
+/* Cancel requests associated with given file descriptor.
+   Copyright (C) 1997, 1998, 2000 Free Software Foundation, Inc.
+   This file is part of the GNU C Library.
+   Contributed by Ulrich Drepper <drepper@cygnus.com>, 1997.
+
+   The GNU C Library 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.
+
+   The GNU C Library 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.
+
+   You should have received a copy of the GNU Library General Public
+   License along with the GNU C Library; see the file COPYING.LIB.  If not,
+   write to the Free Software Foundation, Inc., 59 Temple Place - Suite 330,
+   Boston, MA 02111-1307, USA.  */
+
+
+/* We use an UGLY hack to prevent gcc from finding us cheating.  The
+   implementation of aio_cancel and aio_cancel64 are identical and so
+   we want to avoid code duplication by using aliases.  But gcc sees
+   the different parameter lists and prints a warning.  We define here
+   a function so that aio_cancel64 has no prototype.  */
+#define aio_cancel64 XXX
+#include <aio.h>
+/* And undo the hack.  */
+#undef aio_cancel64
+
+#include <assert.h>
+#include <errno.h>
+
+#include "aio_misc.h"
+
+
+int
+aio_cancel (fildes, aiocbp)
+     int fildes;
+     struct aiocb *aiocbp;
+{
+  struct requestlist *req = NULL;
+  int result = AIO_ALLDONE;
+
+  /* Request the mutex.  */
+  pthread_mutex_lock (&__aio_requests_mutex);
+
+  /* We are asked to cancel a specific AIO request.  */
+  if (aiocbp != NULL)
+    {
+      /* If the AIO request is not for this descriptor it has no value
+	 to look for the request block.  */
+      if (aiocbp->aio_fildes == fildes)
+	{
+	  struct requestlist *last = NULL;
+
+	  req = __aio_find_req_fd (fildes);
+
+	  if (req == NULL)
+	    {
+	    not_found:
+	      pthread_mutex_unlock (&__aio_requests_mutex);
+	      __set_errno (EINVAL);
+	      return -1;
+	    }
+
+	  while (req->aiocbp != (aiocb_union *) aiocbp)
+	    {
+	      last = req;
+	      req = req->next_prio;
+	      if (req == NULL)
+		goto not_found;
+	    }
+
+	  /* Don't remove the entry if a thread is already working on it.  */
+	  if (req->running == allocated)
+	    {
+	      result = AIO_NOTCANCELED;
+	      req = NULL;
+	    }
+	  else
+	    {
+	      /* We can remove the entry.  */
+	      __aio_remove_request (last, req, 0);
+
+	      result = AIO_CANCELED;
+
+	      req->next_prio = NULL;
+	    }
+	}
+    }
+  else
+    {
+      /* Find the beginning of the list of all requests for this
+	 desriptor.  */
+      req = __aio_find_req_fd (fildes);
+
+      /* If any request is worked on by a thread it must be the first.
+	 So either we can delete all requests or all but the first.  */
+      if (req != NULL)
+	{
+	  if (req->running == allocated)
+	    {
+	      struct requestlist *old = req;
+	      req = req->next_prio;
+	      old->next_prio = NULL;
+
+	      result = AIO_NOTCANCELED;
+
+	      if (req != NULL)
+		__aio_remove_request (old, req, 1);
+	    }
+	  else
+	    {
+	      result = AIO_CANCELED;
+
+	      /* We can remove the entry.  */
+	      __aio_remove_request (NULL, req, 1);
+	    }
+	}
+    }
+
+  /* Mark requests as canceled and send signal.  */
+  while (req != NULL)
+    {
+      struct requestlist *old = req;
+      assert (req->running == yes || req->running == queued);
+      req->aiocbp->aiocb.__error_code = ECANCELED;
+      req->aiocbp->aiocb.__return_value = -1;
+      __aio_notify (req);
+      req = req->next_prio;
+      __aio_free_request (old);
+    }
+
+  /* Release the mutex.  */
+  pthread_mutex_unlock (&__aio_requests_mutex);
+
+  return result;
+}
+
+weak_alias (aio_cancel, aio_cancel64)
diff --git a/sysdeps/pthread/aio_fsync.c b/sysdeps/pthread/aio_fsync.c
new file mode 100644
index 0000000000..298b4ec87d
--- /dev/null
+++ b/sysdeps/pthread/aio_fsync.c
@@ -0,0 +1,50 @@
+/* Synchronize I/O in given file descriptor.
+   Copyright (C) 1997, 1999 Free Software Foundation, Inc.
+   This file is part of the GNU C Library.
+   Contributed by Ulrich Drepper <drepper@cygnus.com>, 1997.
+
+   The GNU C Library 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.
+
+   The GNU C Library 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.
+
+   You should have received a copy of the GNU Library General Public
+   License along with the GNU C Library; see the file COPYING.LIB.  If not,
+   write to the Free Software Foundation, Inc., 59 Temple Place - Suite 330,
+   Boston, MA 02111-1307, USA.  */
+
+
+/* We use an UGLY hack to prevent gcc from finding us cheating.  The
+   implementation of aio_fsync and aio_fsync64 are identical and so
+   we want to avoid code duplication by using aliases.  But gcc sees
+   the different parameter lists and prints a warning.  We define here
+   a function so that aio_fsync64 has no prototype.  */
+#define aio_fsync64 XXX
+#include <aio.h>
+/* And undo the hack.  */
+#undef aio_fsync64
+#include <errno.h>
+
+#include "aio_misc.h"
+
+
+int
+aio_fsync (int op, struct aiocb *aiocbp)
+{
+  if (op != O_DSYNC && op != O_SYNC)
+    {
+      __set_errno (EINVAL);
+      return -1;
+    }
+
+  return (__aio_enqueue_request ((aiocb_union *) aiocbp,
+				 op == O_SYNC ? LIO_SYNC : LIO_DSYNC) == NULL
+	  ? -1 : 0);
+}
+
+weak_alias (aio_fsync, aio_fsync64)
diff --git a/sysdeps/pthread/aio_misc.c b/sysdeps/pthread/aio_misc.c
new file mode 100644
index 0000000000..1026e061c8
--- /dev/null
+++ b/sysdeps/pthread/aio_misc.c
@@ -0,0 +1,708 @@
+/* Handle general operations.
+   Copyright (C) 1997, 1998, 1999, 2000, 2001 Free Software Foundation, Inc.
+   This file is part of the GNU C Library.
+   Contributed by Ulrich Drepper <drepper@cygnus.com>, 1997.
+
+   The GNU C Library 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.
+
+   The GNU C Library 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.
+
+   You should have received a copy of the GNU Library General Public
+   License along with the GNU C Library; see the file COPYING.LIB.  If not,
+   write to the Free Software Foundation, Inc., 59 Temple Place - Suite 330,
+   Boston, MA 02111-1307, USA.  */
+
+#include <aio.h>
+#include <assert.h>
+#include <errno.h>
+#include <limits.h>
+#include <pthread.h>
+#include <stdlib.h>
+#include <unistd.h>
+#include <sys/stat.h>
+#include <sys/time.h>
+
+#include "aio_misc.h"
+
+static void add_request_to_runlist (struct requestlist *newrequest);
+
+/* Pool of request list entries.  */
+static struct requestlist **pool;
+
+/* Number of total and allocated pool entries.  */
+static size_t pool_max_size;
+static size_t pool_size;
+
+/* We implement a two dimensional array but allocate each row separately.
+   The macro below determines how many entries should be used per row.
+   It should better be a power of two.  */
+#define ENTRIES_PER_ROW	32
+
+/* How many rows we allocate at once.  */
+#define ROWS_STEP	8
+
+/* List of available entries.  */
+static struct requestlist *freelist;
+
+/* List of request waiting to be processed.  */
+static struct requestlist *runlist;
+
+/* Structure list of all currently processed requests.  */
+static struct requestlist *requests;
+
+/* Number of threads currently running.  */
+static int nthreads;
+
+/* Number of threads waiting for work to arrive. */
+static int idle_thread_count;
+
+
+/* These are the values used to optimize the use of AIO.  The user can
+   overwrite them by using the `aio_init' function.  */
+static struct aioinit optim =
+{
+  20,	/* int aio_threads;	Maximal number of threads.  */
+  64,	/* int aio_num;		Number of expected simultanious requests. */
+  0,
+  0,
+  0,
+  0,
+  1,
+  0
+};
+
+
+/* Since the list is global we need a mutex protecting it.  */
+pthread_mutex_t __aio_requests_mutex = PTHREAD_RECURSIVE_MUTEX_INITIALIZER_NP;
+
+/* When you add a request to the list and there are idle threads present,
+   you signal this condition variable. When a thread finishes work, it waits
+   on this condition variable for a time before it actually exits. */
+pthread_cond_t __aio_new_request_notification = PTHREAD_COND_INITIALIZER;
+
+
+/* Functions to handle request list pool.  */
+static struct requestlist *
+get_elem (void)
+{
+  struct requestlist *result;
+
+  if (freelist == NULL)
+    {
+      struct requestlist *new_row;
+      int cnt;
+
+      assert (sizeof (struct aiocb) == sizeof (struct aiocb64));
+
+      if (pool_size + 1 >= pool_max_size)
+	{
+	  size_t new_max_size = pool_max_size + ROWS_STEP;
+	  struct requestlist **new_tab;
+
+	  new_tab = (struct requestlist **)
+	    realloc (pool, new_max_size * sizeof (struct requestlist *));
+
+	  if (new_tab == NULL)
+	    return NULL;
+
+	  pool_max_size = new_max_size;
+	  pool = new_tab;
+	}
+
+      /* Allocate the new row.  */
+      cnt = pool_size == 0 ? optim.aio_num : ENTRIES_PER_ROW;
+      new_row = (struct requestlist *) calloc (cnt,
+					       sizeof (struct requestlist));
+      if (new_row == NULL)
+	return NULL;
+
+      pool[pool_size++] = new_row;
+
+      /* Put all the new entries in the freelist.  */
+      do
+	{
+	  new_row->next_prio = freelist;
+	  freelist = new_row++;
+	}
+      while (--cnt > 0);
+    }
+
+  result = freelist;
+  freelist = freelist->next_prio;
+
+  return result;
+}
+
+
+void
+internal_function
+__aio_free_request (struct requestlist *elem)
+{
+  elem->running = no;
+  elem->next_prio = freelist;
+  freelist = elem;
+}
+
+
+struct requestlist *
+internal_function
+__aio_find_req (aiocb_union *elem)
+{
+  struct requestlist *runp = requests;
+  int fildes = elem->aiocb.aio_fildes;
+
+  while (runp != NULL && runp->aiocbp->aiocb.aio_fildes < fildes)
+    runp = runp->next_fd;
+
+  if (runp != NULL)
+    {
+      if (runp->aiocbp->aiocb.aio_fildes != fildes)
+	runp = NULL;
+      else
+	while (runp != NULL && runp->aiocbp != elem)
+	  runp = runp->next_prio;
+    }
+
+  return runp;
+}
+
+
+struct requestlist *
+internal_function
+__aio_find_req_fd (int fildes)
+{
+  struct requestlist *runp = requests;
+
+  while (runp != NULL && runp->aiocbp->aiocb.aio_fildes < fildes)
+    runp = runp->next_fd;
+
+  return (runp != NULL && runp->aiocbp->aiocb.aio_fildes == fildes
+	  ? runp : NULL);
+}
+
+
+void
+internal_function
+__aio_remove_request (struct requestlist *last, struct requestlist *req,
+		      int all)
+{
+  assert (req->running == yes || req->running == queued
+	  || req->running == done);
+
+  if (last != NULL)
+    last->next_prio = all ? NULL : req->next_prio;
+  else
+    {
+      if (all || req->next_prio == NULL)
+	{
+	  if (req->last_fd != NULL)
+	    req->last_fd->next_fd = req->next_fd;
+	  else
+	    requests = req->next_fd;
+	  if (req->next_fd != NULL)
+	    req->next_fd->last_fd = req->last_fd;
+	}
+      else
+	{
+	  if (req->last_fd != NULL)
+	    req->last_fd->next_fd = req->next_prio;
+	  else
+	    requests = req->next_prio;
+
+	  if (req->next_fd != NULL)
+	    req->next_fd->last_fd = req->next_prio;
+
+	  req->next_prio->last_fd = req->last_fd;
+	  req->next_prio->next_fd = req->next_fd;
+
+	  /* Mark this entry as runnable.  */
+	  req->next_prio->running = yes;
+	}
+
+      if (req->running == yes)
+	{
+	  struct requestlist *runp = runlist;
+
+	  last = NULL;
+	  while (runp != NULL)
+	    {
+	      if (runp == req)
+		{
+		  if (last == NULL)
+		    runlist = runp->next_run;
+		  else
+		    last->next_run = runp->next_run;
+		  break;
+		}
+	      last = runp;
+	      runp = runp->next_run;
+	    }
+	}
+    }
+}
+
+
+/* The thread handler.  */
+static void *handle_fildes_io (void *arg);
+
+
+/* User optimization.  */
+void
+__aio_init (const struct aioinit *init)
+{
+  /* Get the mutex.  */
+  pthread_mutex_lock (&__aio_requests_mutex);
+
+  /* Only allow writing new values if the table is not yet allocated.  */
+  if (pool == NULL)
+    {
+      optim.aio_threads = init->aio_threads < 1 ? 1 : init->aio_threads;
+      optim.aio_num = (init->aio_num < ENTRIES_PER_ROW
+		       ? ENTRIES_PER_ROW
+		       : init->aio_num & ~ENTRIES_PER_ROW);
+    }
+
+  if (init->aio_idle_time != 0)
+    optim.aio_idle_time = init->aio_idle_time;
+
+  /* Release the mutex.  */
+  pthread_mutex_unlock (&__aio_requests_mutex);
+}
+weak_alias (__aio_init, aio_init)
+
+
+/* The main function of the async I/O handling.  It enqueues requests
+   and if necessary starts and handles threads.  */
+struct requestlist *
+internal_function
+__aio_enqueue_request (aiocb_union *aiocbp, int operation)
+{
+  int result = 0;
+  int policy, prio;
+  struct sched_param param;
+  struct requestlist *last, *runp, *newp;
+  int running = no;
+
+  if (operation == LIO_SYNC || operation == LIO_DSYNC)
+    aiocbp->aiocb.aio_reqprio = 0;
+  else if (aiocbp->aiocb.aio_reqprio < 0
+	   || aiocbp->aiocb.aio_reqprio > AIO_PRIO_DELTA_MAX)
+    {
+      /* Invalid priority value.  */
+      __set_errno (EINVAL);
+      aiocbp->aiocb.__error_code = EINVAL;
+      aiocbp->aiocb.__return_value = -1;
+      return NULL;
+    }
+
+  /* Compute priority for this request.  */
+  pthread_getschedparam (pthread_self (), &policy, &param);
+  prio = param.sched_priority - aiocbp->aiocb.aio_reqprio;
+
+  /* Get the mutex.  */
+  pthread_mutex_lock (&__aio_requests_mutex);
+
+  last = NULL;
+  runp = requests;
+  /* First look whether the current file descriptor is currently
+     worked with.  */
+  while (runp != NULL
+	 && runp->aiocbp->aiocb.aio_fildes < aiocbp->aiocb.aio_fildes)
+    {
+      last = runp;
+      runp = runp->next_fd;
+    }
+
+  /* Get a new element for the waiting list.  */
+  newp = get_elem ();
+  if (newp == NULL)
+    {
+      pthread_mutex_unlock (&__aio_requests_mutex);
+      __set_errno (EAGAIN);
+      return NULL;
+    }
+  newp->aiocbp = aiocbp;
+  newp->caller_pid = (aiocbp->aiocb.aio_sigevent.sigev_notify == SIGEV_SIGNAL
+		      ? getpid () : 0);
+  newp->waiting = NULL;
+
+  aiocbp->aiocb.__abs_prio = prio;
+  aiocbp->aiocb.__policy = policy;
+  aiocbp->aiocb.aio_lio_opcode = operation;
+  aiocbp->aiocb.__error_code = EINPROGRESS;
+  aiocbp->aiocb.__return_value = 0;
+
+  if (runp != NULL
+      && runp->aiocbp->aiocb.aio_fildes == aiocbp->aiocb.aio_fildes)
+    {
+      /* The current file descriptor is worked on.  It makes no sense
+	 to start another thread since this new thread would fight
+	 with the running thread for the resources.  But we also cannot
+	 say that the thread processing this desriptor shall immediately
+	 after finishing the current job process this request if there
+	 are other threads in the running queue which have a higher
+	 priority.  */
+
+      /* Simply enqueue it after the running one according to the
+	 priority.  */
+      while (runp->next_prio != NULL
+	     && runp->next_prio->aiocbp->aiocb.__abs_prio >= prio)
+	runp = runp->next_prio;
+
+      newp->next_prio = runp->next_prio;
+      runp->next_prio = newp;
+
+      running = queued;
+    }
+  else
+    {
+      running = yes;
+      /* Enqueue this request for a new descriptor.  */
+      if (last == NULL)
+	{
+	  newp->last_fd = NULL;
+	  newp->next_fd = requests;
+	  if (requests != NULL)
+	    requests->last_fd = newp;
+	  requests = newp;
+	}
+      else
+	{
+	  newp->next_fd = last->next_fd;
+	  newp->last_fd = last;
+	  last->next_fd = newp;
+	  if (newp->next_fd != NULL)
+	    newp->next_fd->last_fd = newp;
+	}
+
+      newp->next_prio = NULL;
+    }
+
+  if (running == yes)
+    {
+      /* We try to create a new thread for this file descriptor.  The
+	 function which gets called will handle all available requests
+	 for this descriptor and when all are processed it will
+	 terminate.
+
+	 If no new thread can be created or if the specified limit of
+	 threads for AIO is reached we queue the request.  */
+
+      /* See if we need to and are able to create a thread.  */
+      if (nthreads < optim.aio_threads && idle_thread_count == 0)
+	{
+	  pthread_t thid;
+	  pthread_attr_t attr;
+
+	  /* Make sure the thread is created detached.  */
+	  pthread_attr_init (&attr);
+	  pthread_attr_setdetachstate (&attr, PTHREAD_CREATE_DETACHED);
+
+	  running = newp->running = allocated;
+
+	  /* Now try to start a thread.  */
+	  if (pthread_create (&thid, &attr, handle_fildes_io, newp) == 0)
+	    /* We managed to enqueue the request.  All errors which can
+	       happen now can be recognized by calls to `aio_return' and
+	       `aio_error'.  */
+	    ++nthreads;
+	  else
+	    {
+	      /* Reset the running flag.  The new request is not running.  */
+	      running = newp->running = yes;
+
+	      if (nthreads == 0)
+		/* We cannot create a thread in the moment and there is
+		   also no thread running.  This is a problem.  `errno' is
+		   set to EAGAIN if this is only a temporary problem.  */
+		result = -1;
+	    }
+	}
+    }
+
+  /* Enqueue the request in the run queue if it is not yet running.  */
+  if (running == yes && result == 0)
+    {
+      add_request_to_runlist (newp);
+
+      /* If there is a thread waiting for work, then let it know that we
+	 have just given it something to do. */
+      if (idle_thread_count > 0)
+	pthread_cond_signal (&__aio_new_request_notification);
+    }
+
+  if (result == 0)
+    newp->running = running;
+  else
+    {
+      /* Something went wrong.  */
+      __aio_free_request (newp);
+      newp = NULL;
+    }
+
+  /* Release the mutex.  */
+  pthread_mutex_unlock (&__aio_requests_mutex);
+
+  return newp;
+}
+
+
+static void *
+__attribute__ ((noreturn))
+handle_fildes_io (void *arg)
+{
+  pthread_t self = pthread_self ();
+  struct sched_param param;
+  struct requestlist *runp = (struct requestlist *) arg;
+  aiocb_union *aiocbp;
+  int policy;
+  int fildes;
+
+  pthread_getschedparam (self, &policy, &param);
+
+  do
+    {
+      /* If runp is NULL, then we were created to service the work queue
+	 in general, not to handle any particular request. In that case we
+	 skip the "do work" stuff on the first pass, and go directly to the
+	 "get work off the work queue" part of this loop, which is near the
+	 end. */
+      if (runp == NULL)
+	pthread_mutex_lock (&__aio_requests_mutex);
+      else
+	{
+	  /* Hopefully this request is marked as running.  */
+	  assert (runp->running == allocated);
+
+	  /* Update our variables.  */
+	  aiocbp = runp->aiocbp;
+	  fildes = aiocbp->aiocb.aio_fildes;
+
+	  /* Change the priority to the requested value (if necessary).  */
+	  if (aiocbp->aiocb.__abs_prio != param.sched_priority
+	      || aiocbp->aiocb.__policy != policy)
+	    {
+	      param.sched_priority = aiocbp->aiocb.__abs_prio;
+	      policy = aiocbp->aiocb.__policy;
+	      pthread_setschedparam (self, policy, &param);
+	    }
+
+	  /* Process request pointed to by RUNP.  We must not be disturbed
+	     by signals.  */
+	  if ((aiocbp->aiocb.aio_lio_opcode & 127) == LIO_READ)
+	    {
+	      if (aiocbp->aiocb.aio_lio_opcode & 128)
+		aiocbp->aiocb.__return_value =
+		  TEMP_FAILURE_RETRY (__pread64 (fildes, (void *)
+						 aiocbp->aiocb64.aio_buf,
+						 aiocbp->aiocb64.aio_nbytes,
+						 aiocbp->aiocb64.aio_offset));
+	      else
+		aiocbp->aiocb.__return_value =
+		  TEMP_FAILURE_RETRY (pread (fildes,
+					     (void *) aiocbp->aiocb.aio_buf,
+					     aiocbp->aiocb.aio_nbytes,
+					     aiocbp->aiocb.aio_offset));
+
+	      if (aiocbp->aiocb.__return_value == -1 && errno == ESPIPE)
+		/* The Linux kernel is different from others.  It returns
+		   ESPIPE if using pread on a socket.  Other platforms
+		   simply ignore the offset parameter and behave like
+		   read.  */
+		aiocbp->aiocb.__return_value =
+		  TEMP_FAILURE_RETRY (read (fildes,
+					    (void *) aiocbp->aiocb64.aio_buf,
+					    aiocbp->aiocb64.aio_nbytes));
+	    }
+	  else if ((aiocbp->aiocb.aio_lio_opcode & 127) == LIO_WRITE)
+	    {
+	      if (aiocbp->aiocb.aio_lio_opcode & 128)
+		aiocbp->aiocb.__return_value =
+		  TEMP_FAILURE_RETRY (__pwrite64 (fildes, (const void *)
+						  aiocbp->aiocb64.aio_buf,
+						  aiocbp->aiocb64.aio_nbytes,
+						  aiocbp->aiocb64.aio_offset));
+	      else
+		aiocbp->aiocb.__return_value =
+		  TEMP_FAILURE_RETRY (pwrite (fildes, (const void *)
+					      aiocbp->aiocb.aio_buf,
+					      aiocbp->aiocb.aio_nbytes,
+					      aiocbp->aiocb.aio_offset));
+
+	      if (aiocbp->aiocb.__return_value == -1 && errno == ESPIPE)
+		/* The Linux kernel is different from others.  It returns
+		   ESPIPE if using pwrite on a socket.  Other platforms
+		   simply ignore the offset parameter and behave like
+		   write.  */
+		aiocbp->aiocb.__return_value =
+		  TEMP_FAILURE_RETRY (write (fildes,
+					     (void *) aiocbp->aiocb64.aio_buf,
+					     aiocbp->aiocb64.aio_nbytes));
+	    }
+	  else if (aiocbp->aiocb.aio_lio_opcode == LIO_DSYNC)
+	    aiocbp->aiocb.__return_value =
+	      TEMP_FAILURE_RETRY (fdatasync (fildes));
+	  else if (aiocbp->aiocb.aio_lio_opcode == LIO_SYNC)
+	    aiocbp->aiocb.__return_value =
+	      TEMP_FAILURE_RETRY (fsync (fildes));
+	  else
+	    {
+	      /* This is an invalid opcode.  */
+	      aiocbp->aiocb.__return_value = -1;
+	      __set_errno (EINVAL);
+	    }
+
+	  /* Get the mutex.  */
+	  pthread_mutex_lock (&__aio_requests_mutex);
+
+	  /* In theory we would need here a write memory barrier since the
+	     callers test using aio_error() whether the request finished
+	     and once this value != EINPROGRESS the field __return_value
+	     must be committed to memory.
+
+	     But since the pthread_mutex_lock call involves write memory
+	     barriers as well it is not necessary.  */
+
+	  if (aiocbp->aiocb.__return_value == -1)
+	    aiocbp->aiocb.__error_code = errno;
+	  else
+	    aiocbp->aiocb.__error_code = 0;
+
+	  /* Send the signal to notify about finished processing of the
+	     request.  */
+	  __aio_notify (runp);
+
+	  /* For debugging purposes we reset the running flag of the
+	     finished request.  */
+	  assert (runp->running == allocated);
+	  runp->running = done;
+
+	  /* Now dequeue the current request.  */
+	  __aio_remove_request (NULL, runp, 0);
+	  if (runp->next_prio != NULL)
+	    add_request_to_runlist (runp->next_prio);
+
+	  /* Free the old element.  */
+	  __aio_free_request (runp);
+	}
+
+      runp = runlist;
+
+      /* If the runlist is empty, then we sleep for a while, waiting for
+	 something to arrive in it. */
+      if (runp == NULL && optim.aio_idle_time >= 0)
+	{
+	  struct timeval now;
+	  struct timespec wakeup_time;
+
+	  ++idle_thread_count;
+	  gettimeofday (&now, NULL);
+	  wakeup_time.tv_sec = now.tv_sec + optim.aio_idle_time;
+	  wakeup_time.tv_nsec = now.tv_usec * 1000;
+	  if (wakeup_time.tv_nsec > 1000000000)
+	    {
+	      wakeup_time.tv_nsec -= 1000000000;
+	      ++wakeup_time.tv_sec;
+	    }
+	  pthread_cond_timedwait (&__aio_new_request_notification,
+				  &__aio_requests_mutex,
+				  &wakeup_time);
+	  --idle_thread_count;
+	  runp = runlist;
+	}
+
+      if (runp == NULL)
+	--nthreads;
+      else
+	{
+	  assert (runp->running == yes);
+	  runp->running = allocated;
+	  runlist = runp->next_run;
+
+	  /* If we have a request to process, and there's still another in
+	     the run list, then we need to either wake up or create a new
+	     thread to service the request that is still in the run list. */
+	  if (runlist != NULL)
+	    {
+	      /* There are at least two items in the work queue to work on.
+		 If there are other idle threads, then we should wake them
+		 up for these other work elements; otherwise, we should try
+		 to create a new thread. */
+	      if (idle_thread_count > 0)
+		pthread_cond_signal (&__aio_new_request_notification);
+	      else if (nthreads < optim.aio_threads)
+		{
+		  pthread_t thid;
+		  pthread_attr_t attr;
+
+		  /* Make sure the thread is created detached.  */
+		  pthread_attr_init (&attr);
+		  pthread_attr_setdetachstate (&attr, PTHREAD_CREATE_DETACHED);
+
+		  /* Now try to start a thread. If we fail, no big deal,
+		     because we know that there is at least one thread (us)
+		     that is working on AIO operations. */
+		  if (pthread_create (&thid, &attr, handle_fildes_io, NULL)
+		      == 0)
+		    ++nthreads;
+		}
+	    }
+	}
+
+      /* Release the mutex.  */
+      pthread_mutex_unlock (&__aio_requests_mutex);
+    }
+  while (runp != NULL);
+
+  pthread_exit (NULL);
+}
+
+
+/* Free allocated resources.  */
+static void
+__attribute__ ((unused))
+free_res (void)
+{
+  size_t row;
+
+  for (row = 0; row < pool_max_size; ++row)
+    free (pool[row]);
+
+  free (pool);
+}
+text_set_element (__libc_subfreeres, free_res);
+
+
+/* Add newrequest to the runlist. The __abs_prio flag of newrequest must
+   be correctly set to do this. Also, you had better set newrequest's
+   "running" flag to "yes" before you release your lock or you'll throw an
+   assertion. */
+static void
+add_request_to_runlist (struct requestlist *newrequest)
+{
+  int prio = newrequest->aiocbp->aiocb.__abs_prio;
+  struct requestlist *runp;
+
+  if (runlist == NULL || runlist->aiocbp->aiocb.__abs_prio < prio)
+    {
+      newrequest->next_run = runlist;
+      runlist = newrequest;
+    }
+  else
+    {
+      runp = runlist;
+
+      while (runp->next_run != NULL
+	     && runp->next_run->aiocbp->aiocb.__abs_prio >= prio)
+	runp = runp->next_run;
+
+      newrequest->next_run = runp->next_run;
+      runp->next_run = newrequest;
+    }
+}
diff --git a/sysdeps/pthread/aio_misc.h b/sysdeps/pthread/aio_misc.h
new file mode 100644
index 0000000000..372b5556ca
--- /dev/null
+++ b/sysdeps/pthread/aio_misc.h
@@ -0,0 +1,127 @@
+/* Copyright (C) 1997, 1999, 2000 Free Software Foundation, Inc.
+   This file is part of the GNU C Library.
+
+   The GNU C Library 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.
+
+   The GNU C Library 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.
+
+   You should have received a copy of the GNU Library General Public
+   License along with the GNU C Library; see the file COPYING.LIB.  If not,
+   write to the Free Software Foundation, Inc., 59 Temple Place - Suite 330,
+   Boston, MA 02111-1307, USA.  */
+
+#ifndef _AIO_MISC_H
+#define _AIO_MISC_H	1
+
+#include <aio.h>
+#include <pthread.h>
+
+
+/* Extend the operation enum.  */
+enum
+{
+  LIO_DSYNC = LIO_READ + 1,
+  LIO_SYNC,
+  LIO_READ64 = LIO_READ | 128,
+  LIO_WRITE64 = LIO_WRITE | 128
+};
+
+
+/* Union of the two request types.  */
+typedef union
+  {
+    struct aiocb aiocb;
+    struct aiocb64 aiocb64;
+  } aiocb_union;
+
+
+/* Used to synchronize.  */
+struct waitlist
+  {
+    struct waitlist *next;
+
+    pthread_cond_t *cond;
+    volatile int *counterp;
+    /* The next field is used in asynchronous `lio_listio' operations.  */
+    struct sigevent *sigevp;
+    /* XXX See requestlist, it's used to work around the broken signal
+       handling in Linux.  */
+    pid_t caller_pid;
+  };
+
+
+/* Status of a request.  */
+enum
+{
+  no,
+  queued,
+  yes,
+  allocated,
+  done
+};
+
+
+/* Used to queue requests..  */
+struct requestlist
+  {
+    int running;
+
+    struct requestlist *last_fd;
+    struct requestlist *next_fd;
+    struct requestlist *next_prio;
+    struct requestlist *next_run;
+
+    /* Pointer to the actual data.  */
+    aiocb_union *aiocbp;
+
+    /* PID of the initiator thread.
+       XXX This is only necessary for the broken signal handling on Linux.  */
+    pid_t caller_pid;
+
+    /* List of waiting processes.  */
+    struct waitlist *waiting;
+  };
+
+
+/* Lock for global I/O list of requests.  */
+extern pthread_mutex_t __aio_requests_mutex;
+
+
+/* Enqueue request.  */
+extern struct requestlist *__aio_enqueue_request (aiocb_union *aiocbp,
+						  int operation)
+     internal_function;
+
+/* Find request entry for given AIO control block.  */
+extern struct requestlist *__aio_find_req (aiocb_union *elem)
+     internal_function;
+
+/* Find request entry for given file descriptor.  */
+extern struct requestlist *__aio_find_req_fd (int fildes) internal_function;
+
+/* Remove request from the list.  */
+extern void __aio_remove_request (struct requestlist *last,
+				  struct requestlist *req, int all)
+     internal_function;
+
+/* Release the entry for the request.  */
+extern void __aio_free_request (struct requestlist *req) internal_function;
+
+/* Notify initiator of request and tell this everybody listening.  */
+extern void __aio_notify (struct requestlist *req) internal_function;
+
+/* Notify initiator of request.  */
+extern int __aio_notify_only (struct sigevent *sigev, pid_t caller_pid)
+     internal_function;
+
+/* Send the signal.  */
+extern int __aio_sigqueue (int sig, const union sigval val, pid_t caller_pid)
+     internal_function;
+
+#endif /* aio_misc.h */
diff --git a/sysdeps/pthread/aio_notify.c b/sysdeps/pthread/aio_notify.c
new file mode 100644
index 0000000000..9b17750d03
--- /dev/null
+++ b/sysdeps/pthread/aio_notify.c
@@ -0,0 +1,119 @@
+/* Notify initiator of AIO request.
+   Copyright (C) 1997,98,99,2000,01 Free Software Foundation, Inc.
+   This file is part of the GNU C Library.
+   Contributed by Ulrich Drepper <drepper@cygnus.com>, 1997.
+
+   The GNU C Library 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.
+
+   The GNU C Library 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.
+
+   You should have received a copy of the GNU Library General Public
+   License along with the GNU C Library; see the file COPYING.LIB.  If not,
+   write to the Free Software Foundation, Inc., 59 Temple Place - Suite 330,
+   Boston, MA 02111-1307, USA.  */
+
+#include <errno.h>
+#include <pthread.h>
+#include <stdlib.h>
+#include <unistd.h>
+#include "aio_misc.h"
+
+
+static void *
+notify_func_wrapper (void *arg)
+{
+  struct sigevent *sigev = arg;
+  sigev->sigev_notify_function (sigev->sigev_value);
+  return NULL;
+}
+
+
+int
+internal_function
+__aio_notify_only (struct sigevent *sigev, pid_t caller_pid)
+{
+  int result = 0;
+
+  /* Send the signal to notify about finished processing of the request.  */
+  if (sigev->sigev_notify == SIGEV_THREAD)
+    {
+      /* We have to start a thread.  */
+      pthread_t tid;
+      pthread_attr_t attr, *pattr;
+
+      pattr = (pthread_attr_t *) sigev->sigev_notify_attributes;
+      if (pattr == NULL)
+	{
+	  pthread_attr_init (&attr);
+	  pthread_attr_setdetachstate (&attr, PTHREAD_CREATE_DETACHED);
+	  pattr = &attr;
+	}
+
+      if (pthread_create (&tid, pattr, notify_func_wrapper, sigev) < 0)
+	result = -1;
+    }
+  else if (sigev->sigev_notify == SIGEV_SIGNAL)
+    {
+      /* We have to send a signal.  */
+#if _POSIX_REALTIME_SIGNALS
+      /* Note that the standard gives us the option of using a plain
+	 non-queuing signal here when SA_SIGINFO is not set for the signal.  */
+      if (__aio_sigqueue (sigev->sigev_signo, sigev->sigev_value, caller_pid)
+	  < 0)
+	result = -1;
+#else
+      /* There are no queued signals on this system at all.  */
+      result = raise (sigev->sigev_signo);
+#endif
+    }
+
+  return result;
+}
+
+
+void
+internal_function
+__aio_notify (struct requestlist *req)
+{
+  struct waitlist *waitlist;
+  struct aiocb *aiocbp = &req->aiocbp->aiocb;
+
+  if (__aio_notify_only (&aiocbp->aio_sigevent, req->caller_pid) != 0)
+    {
+      /* XXX What shall we do if already an error is set by
+	 read/write/fsync?  */
+      aiocbp->__error_code = errno;
+      aiocbp->__return_value = -1;
+    }
+
+  /* Now also notify possibly waiting threads.  */
+  waitlist = req->waiting;
+  while (waitlist != NULL)
+    {
+      struct waitlist *next = waitlist->next;
+
+      /* Decrement the counter.  This is used in both cases.  */
+      --*waitlist->counterp;
+
+      if (waitlist->sigevp == NULL)
+	pthread_cond_signal (waitlist->cond);
+      else
+	/* This is part of a asynchronous `lio_listio' operation.  If
+	   this request is the last one, send the signal.  */
+	if (*waitlist->counterp == 0)
+	  {
+	    __aio_notify_only (waitlist->sigevp, waitlist->caller_pid);
+	    /* This is tricky.  See lio_listio.c for the reason why
+	       this works.  */
+	    free ((void *) waitlist->counterp);
+	  }
+
+      waitlist = next;
+    }
+}
diff --git a/sysdeps/pthread/aio_read.c b/sysdeps/pthread/aio_read.c
new file mode 100644
index 0000000000..4b5d404190
--- /dev/null
+++ b/sysdeps/pthread/aio_read.c
@@ -0,0 +1,32 @@
+/* Asynchronous read.
+   Copyright (C) 1997, 1999 Free Software Foundation, Inc.
+   This file is part of the GNU C Library.
+   Contributed by Ulrich Drepper <drepper@cygnus.com>, 1997.
+
+   The GNU C Library 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.
+
+   The GNU C Library 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.
+
+   You should have received a copy of the GNU Library General Public
+   License along with the GNU C Library; see the file COPYING.LIB.  If not,
+   write to the Free Software Foundation, Inc., 59 Temple Place - Suite 330,
+   Boston, MA 02111-1307, USA.  */
+
+#include <aio.h>
+
+#include "aio_misc.h"
+
+
+int
+aio_read (aiocbp)
+     struct aiocb *aiocbp;
+{
+  return (__aio_enqueue_request ((aiocb_union *) aiocbp, LIO_READ) == NULL
+	  ? -1 : 0);
+}
diff --git a/sysdeps/pthread/aio_read64.c b/sysdeps/pthread/aio_read64.c
new file mode 100644
index 0000000000..0006ad6303
--- /dev/null
+++ b/sysdeps/pthread/aio_read64.c
@@ -0,0 +1,32 @@
+/* Asynchronous read, 64bit offset version.
+   Copyright (C) 1997, 1999 Free Software Foundation, Inc.
+   This file is part of the GNU C Library.
+   Contributed by Ulrich Drepper <drepper@cygnus.com>, 1997.
+
+   The GNU C Library 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.
+
+   The GNU C Library 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.
+
+   You should have received a copy of the GNU Library General Public
+   License along with the GNU C Library; see the file COPYING.LIB.  If not,
+   write to the Free Software Foundation, Inc., 59 Temple Place - Suite 330,
+   Boston, MA 02111-1307, USA.  */
+
+#include <aio.h>
+
+#include "aio_misc.h"
+
+
+int
+aio_read64 (aiocbp)
+     struct aiocb64 *aiocbp;
+{
+  return (__aio_enqueue_request ((aiocb_union *) aiocbp, LIO_READ64) == NULL
+	  ? -1 : 0);
+}
diff --git a/sysdeps/pthread/aio_suspend.c b/sysdeps/pthread/aio_suspend.c
new file mode 100644
index 0000000000..5ac2d23ffb
--- /dev/null
+++ b/sysdeps/pthread/aio_suspend.c
@@ -0,0 +1,151 @@
+/* Suspend until termination of a requests.
+   Copyright (C) 1997, 1998, 1999, 2000 Free Software Foundation, Inc.
+   This file is part of the GNU C Library.
+   Contributed by Ulrich Drepper <drepper@cygnus.com>, 1997.
+
+   The GNU C Library 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.
+
+   The GNU C Library 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.
+
+   You should have received a copy of the GNU Library General Public
+   License along with the GNU C Library; see the file COPYING.LIB.  If not,
+   write to the Free Software Foundation, Inc., 59 Temple Place - Suite 330,
+   Boston, MA 02111-1307, USA.  */
+
+
+/* We use an UGLY hack to prevent gcc from finding us cheating.  The
+   implementations of aio_suspend and aio_suspend64 are identical and so
+   we want to avoid code duplication by using aliases.  But gcc sees
+   the different parameter lists and prints a warning.  We define here
+   a function so that aio_suspend64 has no prototype.  */
+#define aio_suspend64 XXX
+#include <aio.h>
+/* And undo the hack.  */
+#undef aio_suspend64
+
+#include <errno.h>
+#include <stdlib.h>
+#include <sys/time.h>
+
+#include "aio_misc.h"
+
+
+int
+aio_suspend (list, nent, timeout)
+     const struct aiocb *const list[];
+     int nent;
+     const struct timespec *timeout;
+{
+  struct waitlist waitlist[nent];
+  struct requestlist *requestlist[nent];
+  pthread_cond_t cond = PTHREAD_COND_INITIALIZER;
+  int cnt;
+  int result = 0;
+  int dummy;
+  int none = 1;
+
+  /* Request the mutex.  */
+  pthread_mutex_lock (&__aio_requests_mutex);
+
+  /* There is not yet a finished request.  Signal the request that
+     we are working for it.  */
+  for (cnt = 0; cnt < nent; ++cnt)
+    if (list[cnt] != NULL && list[cnt]->__error_code == EINPROGRESS)
+      {
+	requestlist[cnt] = __aio_find_req ((aiocb_union *) list[cnt]);
+
+	if (requestlist[cnt] != NULL)
+	  {
+	    waitlist[cnt].cond = &cond;
+	    waitlist[cnt].next = requestlist[cnt]->waiting;
+	    waitlist[cnt].counterp = &dummy;
+	    waitlist[cnt].sigevp = NULL;
+	    waitlist[cnt].caller_pid = 0;	/* Not needed.  */
+	    requestlist[cnt]->waiting = &waitlist[cnt];
+	    none = 0;
+	  }
+      }
+
+  /* If there is a not finished request wait for it.  */
+  if (!none)
+    {
+      int oldstate;
+
+      /* Since `pthread_cond_wait'/`pthread_cond_timedwait' are cancelation
+	 points we must be careful.  We added entries to the waiting lists
+	 which we must remove.  So defer cancelation for now.  */
+      pthread_setcancelstate (PTHREAD_CANCEL_DISABLE, &oldstate);
+
+      if (timeout == NULL)
+	result = pthread_cond_wait (&cond, &__aio_requests_mutex);
+      else
+	{
+	  /* We have to convert the relative timeout value into an
+	     absolute time value with pthread_cond_timedwait expects.  */
+	  struct timeval now;
+	  struct timespec abstime;
+
+	  __gettimeofday (&now, NULL);
+	  abstime.tv_nsec = timeout->tv_nsec + now.tv_usec * 1000;
+	  abstime.tv_sec = timeout->tv_sec + now.tv_sec;
+	  if (abstime.tv_nsec >= 1000000000)
+	    {
+	      abstime.tv_nsec -= 1000000000;
+	      abstime.tv_sec += 1;
+	    }
+
+	  result = pthread_cond_timedwait (&cond, &__aio_requests_mutex,
+					   &abstime);
+	}
+
+      /* Now remove the entry in the waiting list for all requests
+	 which didn't terminate.  */
+      for (cnt = 0; cnt < nent; ++cnt)
+	if (list[cnt] != NULL && list[cnt]->__error_code == EINPROGRESS
+	    && requestlist[cnt] != NULL)
+	  {
+	    struct waitlist **listp = &requestlist[cnt]->waiting;
+
+	    /* There is the chance that we cannot find our entry anymore.
+	       This could happen if the request terminated and restarted
+	       again.  */
+	    while (*listp != NULL && *listp != &waitlist[cnt])
+	      listp = &(*listp)->next;
+
+	    if (*listp != NULL)
+	      *listp = (*listp)->next;
+	  }
+
+      /* Now it's time to restore the cancelation state.  */
+      pthread_setcancelstate (oldstate, NULL);
+
+      /* Release the conditional variable.  */
+      if (pthread_cond_destroy (&cond) != 0)
+	/* This must never happen.  */
+	abort ();
+
+      if (result != 0)
+	{
+	  /* An error occurred.  Possibly it's EINTR.  We have to translate
+	     the timeout error report of `pthread_cond_timedwait' to the
+	     form expected from `aio_suspend'.  */
+	  if (result == ETIMEDOUT)
+	    __set_errno (EAGAIN);
+
+	  result = -1;
+	}
+    }
+
+  /* Release the mutex.  */
+  pthread_mutex_unlock (&__aio_requests_mutex);
+
+  return result;
+}
+
+weak_alias (aio_suspend, aio_suspend64)
diff --git a/sysdeps/pthread/aio_write.c b/sysdeps/pthread/aio_write.c
new file mode 100644
index 0000000000..77ad0a685b
--- /dev/null
+++ b/sysdeps/pthread/aio_write.c
@@ -0,0 +1,32 @@
+/* Asynchronous write.
+   Copyright (C) 1997, 1999 Free Software Foundation, Inc.
+   This file is part of the GNU C Library.
+   Contributed by Ulrich Drepper <drepper@cygnus.com>, 1997.
+
+   The GNU C Library 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.
+
+   The GNU C Library 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.
+
+   You should have received a copy of the GNU Library General Public
+   License along with the GNU C Library; see the file COPYING.LIB.  If not,
+   write to the Free Software Foundation, Inc., 59 Temple Place - Suite 330,
+   Boston, MA 02111-1307, USA.  */
+
+#include <aio.h>
+
+#include "aio_misc.h"
+
+
+int
+aio_write (aiocbp)
+     struct aiocb *aiocbp;
+{
+  return (__aio_enqueue_request ((aiocb_union *) aiocbp, LIO_WRITE) == NULL
+	  ? -1 : 0);
+}
diff --git a/sysdeps/pthread/aio_write64.c b/sysdeps/pthread/aio_write64.c
new file mode 100644
index 0000000000..bc95e50b49
--- /dev/null
+++ b/sysdeps/pthread/aio_write64.c
@@ -0,0 +1,32 @@
+/* Asynchronous write, 64bit offset version.
+   Copyright (C) 1997, 1999 Free Software Foundation, Inc.
+   This file is part of the GNU C Library.
+   Contributed by Ulrich Drepper <drepper@cygnus.com>, 1997.
+
+   The GNU C Library 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.
+
+   The GNU C Library 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.
+
+   You should have received a copy of the GNU Library General Public
+   License along with the GNU C Library; see the file COPYING.LIB.  If not,
+   write to the Free Software Foundation, Inc., 59 Temple Place - Suite 330,
+   Boston, MA 02111-1307, USA.  */
+
+#include <aio.h>
+
+#include "aio_misc.h"
+
+
+int
+aio_write64 (aiocbp)
+     struct aiocb64 *aiocbp;
+{
+  return (__aio_enqueue_request ((aiocb_union *) aiocbp, LIO_WRITE64) == NULL
+	  ? -1 : 0);
+}
diff --git a/sysdeps/pthread/lio_listio.c b/sysdeps/pthread/lio_listio.c
new file mode 100644
index 0000000000..e1df603fd8
--- /dev/null
+++ b/sysdeps/pthread/lio_listio.c
@@ -0,0 +1,187 @@
+/* Enqueue and list of read or write requests.
+   Copyright (C) 1997, 1998, 1999, 2000, 2001 Free Software Foundation, Inc.
+   This file is part of the GNU C Library.
+   Contributed by Ulrich Drepper <drepper@cygnus.com>, 1997.
+
+   The GNU C Library 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.
+
+   The GNU C Library 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.
+
+   You should have received a copy of the GNU Library General Public
+   License along with the GNU C Library; see the file COPYING.LIB.  If not,
+   write to the Free Software Foundation, Inc., 59 Temple Place - Suite 330,
+   Boston, MA 02111-1307, USA.  */
+
+#include <aio.h>
+#include <assert.h>
+#include <errno.h>
+#include <stdlib.h>
+#include <unistd.h>
+
+#include "aio_misc.h"
+
+
+/* We need this special structure to handle asynchronous I/O.  */
+struct async_waitlist
+  {
+    int counter;
+    struct sigevent sigev;
+    struct waitlist list[0];
+  };
+
+
+int
+lio_listio (mode, list, nent, sig)
+     int mode;
+     struct aiocb *const list[];
+     int nent;
+     struct sigevent *sig;
+{
+  struct sigevent defsigev;
+  struct requestlist *requests[nent];
+  int cnt;
+  volatile int total = 0;
+  int result = 0;
+
+  /* Check arguments.  */
+  if (mode != LIO_WAIT && mode != LIO_NOWAIT)
+    {
+      __set_errno (EINVAL);
+      return -1;
+    }
+
+  if (sig == NULL)
+    {
+      defsigev.sigev_notify = SIGEV_NONE;
+      sig = &defsigev;
+    }
+
+  /* Request the mutex.  */
+  pthread_mutex_lock (&__aio_requests_mutex);
+
+  /* Now we can enqueue all requests.  Since we already acquired the
+     mutex the enqueue function need not do this.  */
+  for (cnt = 0; cnt < nent; ++cnt)
+    if (list[cnt] != NULL && list[cnt]->aio_lio_opcode != LIO_NOP)
+      {
+	list[cnt]->aio_sigevent.sigev_notify = SIGEV_NONE;
+	requests[cnt] = __aio_enqueue_request ((aiocb_union *) list[cnt],
+					       list[cnt]->aio_lio_opcode);
+
+	if (requests[cnt] != NULL)
+	  /* Successfully enqueued.  */
+	  ++total;
+	else
+	  /* Signal that we've seen an error.  `errno' and the error code
+	     of the aiocb will tell more.  */
+	  result = -1;
+      }
+    else
+      requests[cnt] = NULL;
+
+  if (total == 0)
+    {
+      /* We don't have anything to do except signalling if we work
+	 asynchronously.  */
+
+      /* Release the mutex.  We do this before raising a signal since the
+	 signal handler might do a `siglongjmp' and then the mutex is
+	 locked forever.  */
+      pthread_mutex_unlock (&__aio_requests_mutex);
+
+      if (mode == LIO_NOWAIT)
+	__aio_notify_only (sig,
+			   sig->sigev_notify == SIGEV_SIGNAL ? getpid () : 0);
+
+      return result;
+    }
+  else if (mode == LIO_WAIT)
+    {
+      pthread_cond_t cond = PTHREAD_COND_INITIALIZER;
+      struct waitlist waitlist[nent];
+      int oldstate;
+
+      total = 0;
+      for (cnt = 0; cnt < nent; ++cnt)
+	{
+	  assert (requests[cnt] == NULL || list[cnt] != NULL);
+
+	  if (requests[cnt] != NULL && list[cnt]->aio_lio_opcode != LIO_NOP)
+	    {
+	      waitlist[cnt].cond = &cond;
+	      waitlist[cnt].next = requests[cnt]->waiting;
+	      waitlist[cnt].counterp = &total;
+	      waitlist[cnt].sigevp = NULL;
+	      waitlist[cnt].caller_pid = 0;	/* Not needed.  */
+	      requests[cnt]->waiting = &waitlist[cnt];
+	      ++total;
+	    }
+	}
+
+      /* Since `pthread_cond_wait'/`pthread_cond_timedwait' are cancelation
+	 points we must be careful.  We added entries to the waiting lists
+	 which we must remove.  So defer cancelation for now.  */
+      pthread_setcancelstate (PTHREAD_CANCEL_DISABLE, &oldstate);
+
+      while (total > 0)
+	pthread_cond_wait (&cond, &__aio_requests_mutex);
+
+      /* Now it's time to restore the cancelation state.  */
+      pthread_setcancelstate (oldstate, NULL);
+
+      /* Release the conditional variable.  */
+      if (pthread_cond_destroy (&cond) != 0)
+	/* This must never happen.  */
+	abort ();
+    }
+  else
+    {
+      struct async_waitlist *waitlist;
+
+      waitlist = (struct async_waitlist *)
+	malloc (sizeof (struct async_waitlist)
+		+ (nent * sizeof (struct waitlist)));
+
+      if (waitlist == NULL)
+	{
+	  __set_errno (EAGAIN);
+	  result = -1;
+	}
+      else
+	{
+	  pid_t caller_pid = sig->sigev_notify == SIGEV_SIGNAL ? getpid () : 0;
+	  total = 0;
+
+	  for (cnt = 0; cnt < nent; ++cnt)
+	    {
+	      assert (requests[cnt] == NULL || list[cnt] != NULL);
+
+	      if (requests[cnt] != NULL
+		  && list[cnt]->aio_lio_opcode != LIO_NOP)
+		{
+		  waitlist->list[cnt].cond = NULL;
+		  waitlist->list[cnt].next = requests[cnt]->waiting;
+		  waitlist->list[cnt].counterp = &waitlist->counter;
+		  waitlist->list[cnt].sigevp = &waitlist->sigev;
+		  waitlist->list[cnt].caller_pid = caller_pid;
+		  requests[cnt]->waiting = &waitlist->list[cnt];
+		  ++total;
+		}
+	    }
+
+	  waitlist->counter = total;
+	  waitlist->sigev = *sig;
+	}
+    }
+
+  /* Release the mutex.  */
+  pthread_mutex_unlock (&__aio_requests_mutex);
+
+  return result;
+}
diff --git a/sysdeps/pthread/lio_listio64.c b/sysdeps/pthread/lio_listio64.c
new file mode 100644
index 0000000000..5df2596c0f
--- /dev/null
+++ b/sysdeps/pthread/lio_listio64.c
@@ -0,0 +1,169 @@
+/* Enqueue and list of read or write requests, 64bit offset version.
+   Copyright (C) 1997, 1998, 1999 Free Software Foundation, Inc.
+   This file is part of the GNU C Library.
+   Contributed by Ulrich Drepper <drepper@cygnus.com>, 1997.
+
+   The GNU C Library 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.
+
+   The GNU C Library 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.
+
+   You should have received a copy of the GNU Library General Public
+   License along with the GNU C Library; see the file COPYING.LIB.  If not,
+   write to the Free Software Foundation, Inc., 59 Temple Place - Suite 330,
+   Boston, MA 02111-1307, USA.  */
+
+#include <aio.h>
+#include <errno.h>
+#include <stdlib.h>
+#include <unistd.h>
+
+#include "aio_misc.h"
+
+
+/* We need this special structure to handle asynchronous I/O.  */
+struct async_waitlist
+  {
+    int counter;
+    struct sigevent sigev;
+    struct waitlist list[0];
+  };
+
+
+int
+lio_listio64 (mode, list, nent, sig)
+     int mode;
+     struct aiocb64 *const list[];
+     int nent;
+     struct sigevent *sig;
+{
+  struct requestlist *requests[nent];
+  int cnt;
+  volatile int total = 0;
+  int result = 0;
+
+  /* Check arguments.  */
+  if (mode != LIO_WAIT && mode != LIO_NOWAIT)
+    {
+      __set_errno (EINVAL);
+      return -1;
+    }
+
+  /* Request the mutex.  */
+  pthread_mutex_lock (&__aio_requests_mutex);
+
+  /* Now we can enqueue all requests.  Since we already acquired the
+     mutex the enqueue function need not do this.  */
+  for (cnt = 0; cnt < nent; ++cnt)
+    if (list[cnt] != NULL && list[cnt]->aio_lio_opcode != LIO_NOP)
+      {
+	requests[cnt] = __aio_enqueue_request ((aiocb_union *) list[cnt],
+					       (list[cnt]->aio_lio_opcode
+						| 128));
+	if (requests[cnt] != NULL)
+	  /* Successfully enqueued.  */
+	  ++total;
+	else
+	  /* Signal that we've seen an error.  `errno' and the error code
+	     of the aiocb will tell more.  */
+	  result = -1;
+      }
+
+  if (total == 0)
+    {
+      /* We don't have anything to do except signalling if we work
+	 asynchronously.  */
+
+      /* Release the mutex.  We do this before raising a signal since the
+	 signal handler might do a `siglongjmp' and then the mutex is
+	 locked forever.  */
+      pthread_mutex_unlock (&__aio_requests_mutex);
+
+      if (mode == LIO_NOWAIT)
+	__aio_notify_only (sig,
+			   sig->sigev_notify == SIGEV_SIGNAL ? getpid () : 0);
+
+      return result;
+    }
+  else if (mode == LIO_WAIT)
+    {
+      pthread_cond_t cond = PTHREAD_COND_INITIALIZER;
+      struct waitlist waitlist[nent];
+      int oldstate;
+
+      total = 0;
+      for (cnt = 0; cnt < nent; ++cnt)
+	if (list[cnt] != NULL && list[cnt]->aio_lio_opcode != LIO_NOP
+	    && requests[cnt] != NULL)
+	  {
+	    waitlist[cnt].cond = &cond;
+	    waitlist[cnt].next = requests[cnt]->waiting;
+	    waitlist[cnt].counterp = &total;
+	    waitlist[cnt].sigevp = NULL;
+	    waitlist[cnt].caller_pid = 0;	/* Not needed.  */
+	    requests[cnt]->waiting = &waitlist[cnt];
+	    ++total;
+	  }
+
+      /* Since `pthread_cond_wait'/`pthread_cond_timedwait' are cancelation
+	 points we must be careful.  We added entries to the waiting lists
+	 which we must remove.  So defer cancelation for now.  */
+      pthread_setcancelstate (PTHREAD_CANCEL_DISABLE, &oldstate);
+
+      while (total > 0)
+	pthread_cond_wait (&cond, &__aio_requests_mutex);
+
+      /* Now it's time to restore the cancelation state.  */
+      pthread_setcancelstate (oldstate, NULL);
+
+      /* Release the conditional variable.  */
+      if (pthread_cond_destroy (&cond) != 0)
+	/* This must never happen.  */
+	abort ();
+    }
+  else
+    {
+      struct async_waitlist *waitlist;
+
+      waitlist = (struct async_waitlist *)
+	malloc (sizeof (struct async_waitlist)
+		+ (nent * sizeof (struct waitlist)));
+
+      if (waitlist == NULL)
+	{
+	  __set_errno (EAGAIN);
+	  result = -1;
+	}
+      else
+	{
+	  pid_t caller_pid = sig->sigev_notify == SIGEV_SIGNAL ? getpid () : 0;
+	  total = 0;
+
+	  for (cnt = 0; cnt < nent; ++cnt)
+	    if (list[cnt] != NULL && list[cnt]->aio_lio_opcode != LIO_NOP
+		&& requests[cnt] != NULL)
+	      {
+		waitlist->list[cnt].cond = NULL;
+		waitlist->list[cnt].next = requests[cnt]->waiting;
+		waitlist->list[cnt].counterp = &waitlist->counter;
+		waitlist->list[cnt].sigevp = &waitlist->sigev;
+		waitlist->list[cnt].caller_pid = caller_pid;
+		requests[cnt]->waiting = &waitlist->list[cnt];
+		++total;
+	      }
+
+	  waitlist->counter = total;
+	  waitlist->sigev = *sig;
+	}
+    }
+
+  /* Release the mutex.  */
+  pthread_mutex_unlock (&__aio_requests_mutex);
+
+  return result;
+}