about summary refs log tree commit diff
path: root/rt
diff options
context:
space:
mode:
Diffstat (limited to 'rt')
-rw-r--r--rt/Makefile3
-rw-r--r--rt/aio.h39
-rw-r--r--rt/aio_cancel.c195
-rw-r--r--rt/aio_fsync.c2
-rw-r--r--rt/aio_misc.c588
-rw-r--r--rt/aio_misc.h81
-rw-r--r--rt/aio_notify.c97
-rw-r--r--rt/aio_read.c2
-rw-r--r--rt/aio_read64.c2
-rw-r--r--rt/aio_suspend.c79
-rw-r--r--rt/aio_write.c2
-rw-r--r--rt/aio_write64.c2
-rw-r--r--rt/lio_listio.c112
-rw-r--r--rt/lio_listio64.c113
14 files changed, 978 insertions, 339 deletions
diff --git a/rt/Makefile b/rt/Makefile
index 4eb9dfcb7a..e0f03d78a0 100644
--- a/rt/Makefile
+++ b/rt/Makefile
@@ -28,7 +28,8 @@ extra-libs-others := $(extra-libs)
 
 librt-routines := aio_cancel aio_error aio_fsync aio_misc aio_read	\
 		  aio_read64 aio_return aio_suspend aio_write		\
-		  aio_write64 lio_listio lio_listio64 aio_sigqueue
+		  aio_write64 lio_listio lio_listio64 aio_sigqueue	\
+		  aio_notify
 
 librt-map := librt.map
 
diff --git a/rt/aio.h b/rt/aio.h
index 561776c2d3..a49b8c7a67 100644
--- a/rt/aio.h
+++ b/rt/aio.h
@@ -43,8 +43,6 @@ struct aiocb
   struct sigevent aio_sigevent;	/* Signal number and value.  */
 
   /* Internal members.  */
-  struct aiocb *__last_fd;
-  struct aiocb *__next_fd;
   struct aiocb *__next_prio;
   int __abs_prio;
   int __policy;
@@ -53,10 +51,11 @@ struct aiocb
 
 #ifndef __USE_FILE_OFFSET64
   __off_t aio_offset;		/* File offset.  */
+  char __pad[sizeof (__off64_t) - sizeof (__off_t)];
 #else
   __off64_t aio_offset;		/* File offset.  */
 #endif
-  int __unused[32];
+  char __unused[32];
 };
 
 /* The same for the 64bit offsets.  */
@@ -71,20 +70,33 @@ struct aiocb64
   struct sigevent aio_sigevent;	/* Signal number and value.  */
 
   /* Internal members.  */
-  struct aiocb64 *__last_fd;
-  struct aiocb64 *__next_fd;
-  struct aiocb64 *__next_prio;
   int __abs_prio;
   int __policy;
   int __error_code;
   __ssize_t __return_value;
 
   __off64_t aio_offset;		/* File offset.  */
-  int __unused[32];
+  char __unused[32];
 };
 #endif
 
 
+#ifdef __USE_GNU
+/* To customize the implementation one can use the following struct.
+   This implementation follows the one in Irix.  */
+struct aioinit
+  {
+    int aio_threads;		/* Maximal number of threads.  */
+    int aio_num;		/* Number of expected simultanious requests. */
+    int aio_locks;		/* Not used.  */
+    int aio_usedba;		/* Not used.  */
+    int aio_debug;		/* Not used.  */
+    int aio_numusers;		/* Not used.  */
+    int aio_reserved[2];
+  };
+#endif
+
+
 /* Return values of cancelation function.  */
 enum
 {
@@ -104,12 +116,8 @@ enum
 #define LIO_READ LIO_READ
   LIO_WRITE,
 #define LIO_WRITE LIO_WRITE
-  LIO_NOP,
+  LIO_NOP
 #define LIO_NOP LIO_NOP
-  __LIO_DSYNC,
-  __LIO_SYNC,
-  __LIO_READ64 = LIO_READ | 128,
-  __LIO_WRITE64 = LIO_WRITE | 128
 };
 
 
@@ -123,6 +131,13 @@ enum
 };
 
 
+/* Allow user to specify optimization.  */
+#ifdef __USE_GNU
+extern void __aio_init __P ((__const struct aioinit *__init));
+extern void aio_init __P ((__const struct aioinit *__init));
+#endif
+
+
 /* Enqueue read request for given number of bytes and the given priority.  */
 #ifndef __USE_FILE_OFFSET64
 extern int aio_read __P ((struct aiocb *__aiocbp));
diff --git a/rt/aio_cancel.c b/rt/aio_cancel.c
index f2d9389650..e802a8533f 100644
--- a/rt/aio_cancel.c
+++ b/rt/aio_cancel.c
@@ -30,7 +30,6 @@
 #undef aio_cancel64
 
 #include <errno.h>
-#include <pthread.h>
 
 #include "aio_misc.h"
 
@@ -43,128 +42,106 @@ aio_cancel (fildes, aiocbp)
      int fildes;
      struct aiocb *aiocbp;
 {
-  struct aiocb *firstp;
+  struct requestlist *req = NULL;
   int result = AIO_ALLDONE;
 
-  /* Request the semaphore.  */
-  sem_wait (&__aio_requests_sema);
+  /* Request the mutex.  */
+  pthread_mutex_lock (&__aio_requests_mutex);
 
-  /* Search for the list of requests associated with the given file
-     descriptor.  */
-  for (firstp = (struct aiocb *) __aio_requests; firstp != NULL;
-       firstp = firstp->__next_fd)
-    if (firstp->aio_fildes == fildes)
-      break;
-
-  /* If the file descriptor is not found all work seems to done
-     already.  Otherwise try to cancel the request(s).  */
-  if (firstp != NULL)
+  /* We are asked to cancel a specific AIO request.  */
+  if (aiocbp != NULL)
     {
-      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)
 	{
-	  /* Locate the entry corresponding to the AIOCBP parameter.  */
-	  if (aiocbp == firstp)
-	    /* The requests is currently handled, therefore don't
-	       cancel it and signal this to the user.  */
-	    result = AIO_NOTCANCELED;
-	  else
+	  struct requestlist *last = NULL;
+
+	  req = __aio_find_req_fd (fildes);
+
+	  while (req->aiocbp != (aiocb_union *) aiocbp)
 	    {
-	      while (firstp->__next_prio != NULL
-		     && aiocbp != firstp->__next_prio)
-		firstp = firstp->__next_prio;
-
-	      if (firstp->__next_prio != NULL)
-		{
-		  /* The request the user wants to cancel is in the
-		     queue.  Simply remove it.  */
-		  firstp->__next_prio = aiocbp->__next_prio;
-
-		  /* Mark as canceled.  */
-		  aiocbp->__error_code = ECANCELED;
-		  aiocbp->__return_value = -1;
-
-		  /* Send the signal to notify about canceled
-		     processing of the request.  */
-		  if (aiocbp->aio_sigevent.sigev_notify == SIGEV_THREAD)
-		    {
-		      /* We have to start a thread.  */
-		      pthread_t tid;
-		      pthread_attr_t attr, *pattr;
-
-		      pattr = (pthread_attr_t *)
-			aiocbp->aio_sigevent.sigev_notify_attributes;
-		      if (pattr == NULL)
-			{
-			  pthread_attr_init (&attr);
-			  pthread_attr_setdetachstate (&attr,
-						       PTHREAD_CREATE_DETACHED);
-			  pattr = &attr;
-			}
-
-		      pthread_create (&tid, pattr,
-				      (void *(*) (void *))
-				      aiocbp->aio_sigevent.sigev_notify_function,
-				      aiocbp->aio_sigevent.sigev_value.sival_ptr);
-		    }
-		  else if (aiocbp->aio_sigevent.sigev_notify == SIGEV_SIGNAL)
-		    /* We have to send a signal.  */
-		    __aio_sigqueue (aiocbp->aio_sigevent.sigev_signo,
-				    aiocbp->aio_sigevent.sigev_value);
-
-		  result = AIO_CANCELED;
-		}
+	      last = req;
+	      req = req->next_prio;
 	    }
-	}
-      else
-	{
-	  /* First dequeue all waiting requests.  */
-	  aiocbp = firstp;
 
-	  while ((firstp = firstp->__next_prio) != NULL)
+	  /* Don't remove the entry if a thread is already working on it.  */
+	  if (req->running == allocated)
+	    result = AIO_NOTCANCELED;
+	  else
 	    {
-	      firstp->__error_code = ECANCELED;
-	      firstp->__return_value = -1;
-
-
-	      /* Send the signal to notify about canceled processing
-		 of the request.  */
-	      if (firstp->aio_sigevent.sigev_notify == SIGEV_THREAD)
-		{
-		  /* We have to start a thread.  */
-		  pthread_t tid;
-		  pthread_attr_t attr, *pattr;
-
-		  pattr = (pthread_attr_t *)
-		    aiocbp->aio_sigevent.sigev_notify_attributes;
-		  if (pattr == NULL)
-		    {
-		      pthread_attr_init (&attr);
-		      pthread_attr_setdetachstate (&attr,
-						   PTHREAD_CREATE_DETACHED);
-		      pattr = &attr;
-		    }
-
-		  pthread_create (&tid, pattr,
-				  (void *(*) (void *))
-				  firstp->aio_sigevent.sigev_notify_function,
-				  firstp->aio_sigevent.sigev_value.sival_ptr);
-		}
-	      else if (firstp->aio_sigevent.sigev_notify == SIGEV_SIGNAL)
-		/* We have to send a signal.  */
-		__aio_sigqueue (firstp->aio_sigevent.sigev_signo,
-				firstp->aio_sigevent.sigev_value);
+	      /* We can remove the entry.  */
+	      if (last != NULL)
+		last->next_prio = req->next_prio;
+	      else
+		if (req->next_prio == NULL)
+		  {
+		    if (req->last_fd != NULL)
+		      req->last_fd->next_fd = 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;
+		    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;
+		  }
+
+	      result = AIO_CANCELED;
 	    }
 
-	  /* We have to signal that not all requests could be canceled
-	     since the first requests is currently processed.  */
-	  result = AIO_NOTCANCELED;
-
-	  aiocbp->__next_prio = NULL;
+	  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;
+	  }
+	else
+	  {
+	    /* Remove entry from the file descriptor list.  */
+	    if (req->last_fd != NULL)
+	      req->last_fd->next_fd = req->next_fd;
+	    if (req->next_fd != NULL)
+	      req->next_fd->last_fd = req->last_fd;
+
+	    result = AIO_CANCELED;
+	  }
+    }
+
+  /* Mark requests as canceled and send signal.  */
+  while (req != NULL)
+    {
+      struct requestlist *old = req;
+      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 semaphore.  */
-  sem_post (&__aio_requests_sema);
+  /* Release the mutex.  */
+  pthread_mutex_unlock (&__aio_requests_mutex);
 
   return result;
 }
diff --git a/rt/aio_fsync.c b/rt/aio_fsync.c
index 6daaca1e97..c813654b61 100644
--- a/rt/aio_fsync.c
+++ b/rt/aio_fsync.c
@@ -36,7 +36,7 @@ int
 aio_fsync (int op, struct aiocb *aiocbp)
 {
   return __aio_enqueue_request ((aiocb_union *) aiocbp,
-				op == O_SYNC ? __LIO_SYNC : __LIO_DSYNC, 1);
+				op == O_SYNC ? LIO_SYNC : LIO_DSYNC) != NULL;
 }
 
 weak_alias (aio_fsync, aio_fsync64)
diff --git a/rt/aio_misc.c b/rt/aio_misc.c
index e4bb12c500..6ea30c2158 100644
--- a/rt/aio_misc.c
+++ b/rt/aio_misc.c
@@ -21,7 +21,6 @@
 #include <aio.h>
 #include <errno.h>
 #include <pthread.h>
-#include <semaphore.h>
 #include <stdlib.h>
 #include <unistd.h>
 #include <sys/stat.h>
@@ -29,40 +28,199 @@
 #include "aio_misc.h"
 
 
-/* We need a list of pending operations.  This is sorted according to
-   the priority given in the aio_reqprio member.  */
-aiocb_union *__aio_requests;
+/* Pool of request list entries.  */
+static struct requestlist **pool;
 
-/* Since the list is global we need a semaphore protecting it.  */
-sem_t __aio_requests_sema;
+/* Number of total and allocated pool entries.  */
+static size_t pool_tab_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	16
 
-/* The initialization function.  It gets automatically called if any
-   aio_* function is used in the program.  */
-static void
-__attribute__ ((unused))
-aio_initialize (void)
+/* The row table is incremented in units of this.  */
+#define ROW_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;
+
+
+/* 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.  */
+  256,	/* int aio_num;		Number of expected simultanious requests. */
+  0,
+  0,
+  0,
+  0,
+  { 0, }
+};
+
+
+/* Since the list is global we need a mutex protecting it.  */
+pthread_mutex_t __aio_requests_mutex = PTHREAD_RECURSIVE_MUTEX_INITIALIZER_NP;
+
+
+/* Functions to handle request list pool.  */
+static struct requestlist *
+get_elem (void)
 {
-  /* Initialize the semaphore.  We allow exactly one user at a time.  */
-  sem_init (&__aio_requests_sema, 0, 1);
+  struct requestlist *result;
+
+  if (freelist == NULL)
+    {
+      struct requestlist *new_row;
+      size_t new_size;
+
+      /* Compute new size.  */
+      new_size = pool_size ? pool_size + ENTRIES_PER_ROW : optim.aio_num;
+
+      if ((new_size / ENTRIES_PER_ROW) >= pool_tab_size)
+	{
+	  size_t new_tab_size = new_size / ENTRIES_PER_ROW;
+	  struct requestlist **new_tab;
+
+	  new_tab = (struct requestlist **)
+	    realloc (pool, (new_tab_size * sizeof (struct requestlist *)));
+
+	  if (new_tab == NULL)
+	    return NULL;
+
+	  pool_tab_size = new_tab_size;
+	  pool = new_tab;
+	}
+
+      if (pool_size == 0)
+	{
+	  size_t cnt;
+
+	  new_row = (struct requestlist *)
+	    calloc (new_size, sizeof (struct requestlist));
+
+	  if (new_row == NULL)
+	    return NULL;
+
+	  for (cnt = 0; cnt < new_size / ENTRIES_PER_ROW; ++cnt)
+	    pool[cnt] = &new_row[cnt * ENTRIES_PER_ROW];
+	}
+      else
+	{
+	  /* Allocat one new row.  */
+	  new_row = (struct requestlist *)
+	    calloc (ENTRIES_PER_ROW, sizeof (struct requestlist));
+	  if (new_row == NULL)
+	    return NULL;
+
+	  pool[new_size / ENTRIES_PER_ROW] = new_row;
+	}
+
+      /* Put all the new entries in the freelist.  */
+      do
+	{
+	  new_row->next_prio = freelist;
+	  freelist = new_row++;
+	}
+      while (++pool_size < new_size);
+    }
+
+  result = freelist;
+  freelist = freelist->next_prio;
+
+  return result;
 }
 
-text_set_element (__libc_subinit, aio_initialize);
+
+void
+__aio_free_req (struct requestlist *elem)
+{
+  elem->running = no;
+  elem->next_prio = freelist;
+  freelist = elem;
+}
+
+
+struct requestlist *
+__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 *
+__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);
+}
 
 
 /* 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);
+    }
+
+  /* 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.  */
-int
-__aio_enqueue_request (aiocb_union *aiocbp, int operation, int require_lock)
+struct requestlist *
+__aio_enqueue_request (aiocb_union *aiocbp, int operation)
 {
-  int result;
+  int result = 0;
   int policy, prio;
   struct sched_param param;
-  aiocb_union *runp;
+  struct requestlist *last, *runp, *newp;
+  int running = no;
 
   if (aiocbp->aiocb.aio_reqprio < 0
       || aiocbp->aiocb.aio_reqprio > AIO_PRIO_DELTA_MAX)
@@ -71,94 +229,160 @@ __aio_enqueue_request (aiocb_union *aiocbp, int operation, int require_lock)
       __set_errno (EINVAL);
       aiocbp->aiocb.__error_code = EINVAL;
       aiocbp->aiocb.__return_value = -1;
-      return -1;
-    }
-
-  if (pthread_getschedparam (pthread_self (), &policy, &param) < 0)
-    {
-      /* Something went wrong.  */
-      aiocbp->aiocb.__error_code = errno;
-      aiocbp->aiocb.__return_value = -1;
-      return -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);
 
-  /* Get the semaphore.  */
-  if (require_lock)
-    sem_wait (&__aio_requests_sema);
-
-  runp = __aio_requests;
+  last = NULL;
+  runp = requests;
   /* First look whether the current file descriptor is currently
      worked with.  */
-  while (runp != NULL && runp->aiocb.aio_fildes < aiocbp->aiocb.aio_fildes)
-    runp = (aiocb_union *) runp->aiocb.__next_fd;
+  while (runp != NULL
+	 && runp->aiocbp->aiocb.aio_fildes < aiocbp->aiocb.aio_fildes)
+    {
+      last = runp;
+      runp = runp->next_fd;
+    }
 
-  if (runp != NULL)
+  /* Get a new element for the waiting list.  */
+  newp = get_elem ();
+  if (newp == NULL)
+    {
+      __set_errno (EAGAIN);
+      pthread_mutex_unlock (&__aio_requests_mutex);
+      return NULL;
+    }
+  newp->aiocbp = aiocbp;
+  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 have to
-	 wait for the previous one to terminate.  Simply enqueue it
-	 after the running one according to the priority.  */
-      while (runp->aiocb.__next_prio != NULL
-	     && runp->aiocb.__next_prio->__abs_prio >= prio)
-	runp = (aiocb_union *) runp->aiocb.__next_prio;
-
-      aiocbp->aiocb.__next_prio = runp->aiocb.__next_prio;
-      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;
-      runp->aiocb.__next_prio = (struct aiocb *) aiocbp;
-
-      result = 0;
+	 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 imeediately
+	 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
     {
-      /* We create a new thread for this file descriptor.  The
+      /* 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 == no)
+    {
+      /* 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.  */
-      pthread_t thid;
-      pthread_attr_t attr;
-
-      /* First enqueue the request (the list is empty).  */
-      aiocbp->aiocb.__next_fd = NULL;
-      aiocbp->aiocb.__last_fd = NULL;
-
-      aiocbp->aiocb.__next_prio = 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;
-
-      /* 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 (pthread_create (&thid, &attr, handle_fildes_io, aiocbp) < 0)
+	 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 can create a thread.  */
+      if (nthreads < optim.aio_threads)
 	{
-	  result = -1;
-	  aiocbp->aiocb.__error_code = errno;
-	  aiocbp->aiocb.__return_value = -1;
+	  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 (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'.  */
+	      running = allocated;
+	      ++nthreads;
+	    }
+	  else 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)
+    {
+      if (runlist == NULL || runlist->aiocbp->aiocb.__abs_prio < prio)
+	{
+	  newp->next_run = runlist;
+	  runlist = newp;
 	}
       else
-	/* We managed to enqueue the request.  All errors which can
-	   happen now can be recognized by calls to `aio_return' and
-	   `aio_error'.  */
-	  result = 0;
+	{
+	  runp = runlist;
+
+	  while (runp->next_run != NULL
+		 && runp->next_run->aiocbp->aiocb.__abs_prio >= prio)
+	    runp = runp->next_run;
+
+	  newp->next_run = runp->next_run;
+	  runp->next_run = newp;
+	}
     }
 
-  /* Release the semaphore.  */
-  if (require_lock)
-    sem_post (&__aio_requests_sema);
+  if (result == 0)
+    newp->running = running;
+  else
+    {
+      /* Something went wrong.  */
+      __aio_free_req (newp);
+      newp = NULL;
+    }
 
-  return result;
+  /* Release the mutex.  */
+  pthread_mutex_unlock (&__aio_requests_mutex);
+
+  return newp;
 }
 
 
@@ -167,140 +391,156 @@ handle_fildes_io (void *arg)
 {
   pthread_t self = pthread_self ();
   struct sched_param param;
-  aiocb_union *runp = (aiocb_union *) arg;
+  struct requestlist *runp = (struct requestlist *) arg;
+  aiocb_union *aiocbp = runp->aiocbp;
   int policy;
-  int fildes = runp->aiocb.aio_fildes;	/* This is always the same.  */
+  int fildes = runp->aiocbp->aiocb.aio_fildes;
 
   pthread_getschedparam (self, &policy, &param);
 
   do
     {
       /* Change the priority to the requested value (if necessary).  */
-      if (runp->aiocb.__abs_prio != param.sched_priority
-	  || runp->aiocb.__policy != policy)
+      if (aiocbp->aiocb.__abs_prio != param.sched_priority
+	  || aiocbp->aiocb.__policy != policy)
 	{
-	  param.sched_priority = runp->aiocb.__abs_prio;
-	  policy = runp->aiocb.__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 ((runp->aiocb.aio_lio_opcode & 127) == LIO_READ)
+      if ((aiocbp->aiocb.aio_lio_opcode & 127) == LIO_READ)
 	{
-	  if (runp->aiocb.aio_lio_opcode & 128)
-	    runp->aiocb.__return_value =
+	  if (aiocbp->aiocb.aio_lio_opcode & 128)
+	    aiocbp->aiocb.__return_value =
 	      TEMP_FAILURE_RETRY (__pread64 (fildes,
-					     (void *) runp->aiocb64.aio_buf,
-					     runp->aiocb64.aio_nbytes,
-					     runp->aiocb64.aio_offset));
+					     (void *) aiocbp->aiocb64.aio_buf,
+					     aiocbp->aiocb64.aio_nbytes,
+					     aiocbp->aiocb64.aio_offset));
 	  else
-	    runp->aiocb.__return_value =
+	    aiocbp->aiocb.__return_value =
 	      TEMP_FAILURE_RETRY (__pread (fildes,
-					   (void *) runp->aiocb.aio_buf,
-					   runp->aiocb.aio_nbytes,
-					   runp->aiocb.aio_offset));
+					   (void *) aiocbp->aiocb.aio_buf,
+					   aiocbp->aiocb.aio_nbytes,
+					   aiocbp->aiocb.aio_offset));
 	}
-      else if ((runp->aiocb.aio_lio_opcode & 127) == LIO_WRITE)
+      else if ((aiocbp->aiocb.aio_lio_opcode & 127) == LIO_WRITE)
 	{
-	  if (runp->aiocb.aio_lio_opcode & 128)
-	    runp->aiocb.__return_value =
+	  if (aiocbp->aiocb.aio_lio_opcode & 128)
+	    aiocbp->aiocb.__return_value =
 	      TEMP_FAILURE_RETRY (__pwrite64 (fildes,
-					      (const void *) runp->aiocb64.aio_buf,
-					      runp->aiocb64.aio_nbytes,
-					      runp->aiocb64.aio_offset));
+					      (const void *) aiocbp->aiocb64.aio_buf,
+					      aiocbp->aiocb64.aio_nbytes,
+					      aiocbp->aiocb64.aio_offset));
 	  else
-	    runp->aiocb.__return_value =
+	    aiocbp->aiocb.__return_value =
 	      TEMP_FAILURE_RETRY (__pwrite (fildes,
-					    (const void *) runp->aiocb.aio_buf,
-					    runp->aiocb.aio_nbytes,
-					    runp->aiocb.aio_offset));
+					    (const void *) aiocbp->aiocb.aio_buf,
+					    aiocbp->aiocb.aio_nbytes,
+					    aiocbp->aiocb.aio_offset));
 	}
-      else if (runp->aiocb.aio_lio_opcode == __LIO_DSYNC)
-	runp->aiocb.__return_value = TEMP_FAILURE_RETRY (fdatasync (fildes));
-      else if (runp->aiocb.aio_lio_opcode == __LIO_SYNC)
-	runp->aiocb.__return_value = TEMP_FAILURE_RETRY (fsync (fildes));
+      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.  */
-	  runp->aiocb.__return_value = -1;
+	  aiocbp->aiocb.__return_value = -1;
 	  __set_errno (EINVAL);
 	}
 
-      if (runp->aiocb.__return_value == -1)
-	runp->aiocb.__error_code = errno;
+      /* Get the mutex.  */
+      pthread_mutex_lock (&__aio_requests_mutex);
+
+      if (aiocbp->aiocb.__return_value == -1)
+	aiocbp->aiocb.__error_code = errno;
       else
-	runp->aiocb.__error_code = 0;
+	aiocbp->aiocb.__error_code = 0;
 
       /* Send the signal to notify about finished processing of the
 	 request.  */
-      if (runp->aiocb.aio_sigevent.sigev_notify == SIGEV_THREAD)
+      __aio_notify (runp);
+
+      /* Now dequeue the current request.  */
+      if (runp->next_prio == NULL)
 	{
-	  /* We have to start a thread.  */
-	  pthread_t tid;
-	  pthread_attr_t attr, *pattr;
+	  /* No outstanding request for this descriptor.  Process the
+	     runlist if necessary.  */
+	  if (runp->next_fd != NULL)
+	    runp->next_fd->last_fd = runp->last_fd;
+	  if (runp->last_fd != NULL)
+	    runp->last_fd->next_fd = runp->next_fd;
+	}
+      else
+	{
+	  runp->next_prio->last_fd = runp->last_fd;
+	  runp->next_prio->next_fd = runp->next_fd;
+	  runp->next_prio->running = yes;
+	  if (runp->next_fd != NULL)
+	    runp->next_fd->last_fd = runp->next_prio;
+	  if (runp->last_fd != NULL)
+	    runp->last_fd->next_fd = runp->next_prio;
+	}
+
+      /* Free the old element.  */
+      __aio_free_req (runp);
 
-	  pattr = (pthread_attr_t *)
-	    runp->aiocb.aio_sigevent.sigev_notify_attributes;
-	  if (pattr == NULL)
+      runp = freelist;
+      if (runp != NULL)
+	{
+	  /* We must not run requests which are not marked `running'.  */
+	  if (runp->running == yes)
 	    {
-	      pthread_attr_init (&attr);
-	      pthread_attr_setdetachstate (&attr, PTHREAD_CREATE_DETACHED);
-	      pattr = &attr;
+	      freelist = runp->next_run;
+	      runp->running = allocated;
 	    }
-
-	  if (pthread_create (&tid,
-			      (pthread_attr_t *)
-			      runp->aiocb.aio_sigevent.sigev_notify_attributes,
-			      (void *(*) (void *))
-			      runp->aiocb.aio_sigevent.sigev_notify_function,
-			      runp->aiocb.aio_sigevent.sigev_value.sival_ptr)
-	      < 0)
+	  else
 	    {
-	      /* XXX What shall we do if already an error is set by
-		 read/write/fsync?  */
-	      runp->aiocb.__error_code = errno;
-	      runp->aiocb.__return_value = -1;
+	      struct requestlist *old;
+
+	      do
+		{
+		  old = runp;
+		  runp = runp->next_run;
+		}
+	      while (runp != NULL && runp->running != yes);
+
+	      if (runp != NULL)
+		old->next_run = runp->next_run;
 	    }
 	}
-      else if (runp->aiocb.aio_sigevent.sigev_notify == SIGEV_SIGNAL)
-	/* We have to send a signal.  */
-	if (__aio_sigqueue (runp->aiocb.aio_sigevent.sigev_signo,
-			    runp->aiocb.aio_sigevent.sigev_value) < 0)
-	  {
-	    /* XXX What shall we do if already an error is set by
-	       read/write/fsync?  */
-	    runp->aiocb.__error_code = errno;
-	    runp->aiocb.__return_value = -1;
-	  }
-
-      /* Get the semaphore.  */
-      sem_wait (&__aio_requests_sema);
 
-      /* Now dequeue the current request.  */
-      if (runp->aiocb.__next_prio == NULL)
-	{
-	  if (runp->aiocb.__next_fd != NULL)
-	    runp->aiocb.__next_fd->__last_fd = runp->aiocb.__last_fd;
-	  if (runp->aiocb.__last_fd != NULL)
-	    runp->aiocb.__last_fd->__next_fd = runp->aiocb.__next_fd;
-	  runp = NULL;
-	}
-      else
-	{
-	  runp->aiocb.__next_prio->__last_fd = runp->aiocb.__last_fd;
-	  runp->aiocb.__next_prio->__next_fd = runp->aiocb.__next_fd;
-	  if (runp->aiocb.__next_fd != NULL)
-	    runp->aiocb.__next_fd->__last_fd = runp->aiocb.__next_prio;
-	  if (runp->aiocb.__last_fd != NULL)
-	    runp->aiocb.__last_fd->__next_fd = runp->aiocb.__next_prio;
-	  runp = (aiocb_union *) runp->aiocb.__next_prio;
-	}
+      /* If no request to work on we will stop the thread.  */
+      if (runp == NULL)
+	--nthreads;
 
-      /* Release the semaphore.  */
-      sem_post (&__aio_requests_sema);
+      /* 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;
+
+  /* The first block of rows as specified in OPTIM is allocated in
+     one chunk.  */
+  free (pool[0]);
+
+  for (row = optim.aio_num / ENTRIES_PER_ROW; row < pool_tab_size; ++row)
+    free (pool[row]);
+
+  free (pool);
+}
+
+text_set_element (__libc_subfreeres, free_res);
diff --git a/rt/aio_misc.h b/rt/aio_misc.h
index c2eb9fcea1..af8e2479ba 100644
--- a/rt/aio_misc.h
+++ b/rt/aio_misc.h
@@ -19,7 +19,19 @@
 #ifndef _AIO_MISC_H
 #define _AIO_MISC_H	1
 
-#include <semaphore.h>
+#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
@@ -28,18 +40,73 @@ typedef union
     struct aiocb64 aiocb64;
   } aiocb_union;
 
-/* List of enqueued requests.  */
-extern aiocb_union *__aio_requests;
+
+/* 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;
+  };
+
+
+/* Status of a request.  */
+enum
+{
+  no,
+  queued,
+  yes,
+  allocated
+};
+
+
+/* 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;
+
+    /* List of waiting processes.  */
+    struct waitlist *waiting;
+  };
+
 
 /* Lock for global I/O list of requests.  */
-extern sem_t __aio_requests_sema;
+extern pthread_mutex_t __aio_requests_mutex;
 
 
 /* Enqueue request.  */
-extern int __aio_enqueue_request (aiocb_union *aiocbp, int operation,
-				  int require_lock);
+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;
+
+/* 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) internal_function;
 
 /* Send the signal.  */
-extern int __aio_sigqueue (int sig, const union sigval val);
+extern int __aio_sigqueue (int sig, const union sigval val) internal_function;
 
 #endif /* aio_misc.h */
diff --git a/rt/aio_notify.c b/rt/aio_notify.c
new file mode 100644
index 0000000000..be2e1e246b
--- /dev/null
+++ b/rt/aio_notify.c
@@ -0,0 +1,97 @@
+/* Notify initiator of AIO request.
+   Copyright (C) 1997 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 <pthread.h>
+#include <stdlib.h>
+#include "aio_misc.h"
+
+int
+__aio_notify_only (struct sigevent *sigev)
+{
+  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,
+			  (void *(*) (void *)) sigev->sigev_notify_function,
+			  sigev->sigev_value.sival_ptr) < 0)
+	result = -1;
+    }
+  else if (sigev->sigev_notify == SIGEV_SIGNAL)
+    /* We have to send a signal.  */
+    if (__aio_sigqueue (sigev->sigev_signo, sigev->sigev_value) < 0)
+      result = -1;
+
+  return result;
+}
+
+
+void
+__aio_notify (struct requestlist *req)
+{
+  struct waitlist *waitlist;
+  struct aiocb *aiocbp = &req->aiocbp->aiocb;
+
+  if (__aio_notify_only (&aiocbp->aio_sigevent) != 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);
+	    /* This is tricky.  See lio_listio.c for the reason why
+	       this works.  */
+	    free ((void *) waitlist->counterp);
+	  }
+
+      waitlist = next;
+    }
+}
diff --git a/rt/aio_read.c b/rt/aio_read.c
index 8286ba9334..871ce99767 100644
--- a/rt/aio_read.c
+++ b/rt/aio_read.c
@@ -27,5 +27,5 @@ int
 aio_read (aiocbp)
      struct aiocb *aiocbp;
 {
-  return __aio_enqueue_request ((aiocb_union *) aiocbp, LIO_READ, 1);
+  return __aio_enqueue_request ((aiocb_union *) aiocbp, LIO_READ) != NULL;
 }
diff --git a/rt/aio_read64.c b/rt/aio_read64.c
index bf808fb3da..b194dda6db 100644
--- a/rt/aio_read64.c
+++ b/rt/aio_read64.c
@@ -27,5 +27,5 @@ int
 aio_read64 (aiocbp)
      struct aiocb64 *aiocbp;
 {
-  return __aio_enqueue_request ((aiocb_union *) aiocbp, __LIO_READ64, 1);
+  return __aio_enqueue_request ((aiocb_union *) aiocbp, LIO_READ64) != NULL;
 }
diff --git a/rt/aio_suspend.c b/rt/aio_suspend.c
index 75bf9ba94b..6123b7baa0 100644
--- a/rt/aio_suspend.c
+++ b/rt/aio_suspend.c
@@ -40,16 +40,89 @@ aio_suspend (list, nent, timeout)
      int nent;
      const struct timespec *timeout;
 {
+  pthread_cond_t cond;
+  struct waitlist waitlist[nent];
+  struct requestlist *requestlist[nent];
   int cnt;
+  int result = 0;
+
+  /* Request the mutex.  */
+  pthread_mutex_lock (&__aio_requests_mutex);
 
   /* First look whether there is already a terminated request.  */
   for (cnt = 0; cnt < nent; ++cnt)
     if (list[cnt] != NULL && list[cnt]->__error_code != EINPROGRESS)
-      return 0;
+      break;
+
+  if (cnt == nent)
+    {
+      int oldstate;
+
+      /* 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 = NULL;
+		waitlist[cnt].sigevp = NULL;
+		requestlist[cnt]->waiting = &waitlist[cnt];
+	      }
+	  }
+
+      /* 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
+	result = pthread_cond_timedwait (&cond, &__aio_requests_mutex,
+					 timeout);
+
+      /* 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);
+
+      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;
+	}
+    }
 
-  /* XXX We have to write code which waits.  */
+  /* Release the mutex.  */
+  pthread_mutex_unlock (&__aio_requests_mutex);
 
-  return -1;
+  return result;
 }
 
 weak_alias (aio_suspend, aio_suspend64)
diff --git a/rt/aio_write.c b/rt/aio_write.c
index 80b827c1aa..8801d89110 100644
--- a/rt/aio_write.c
+++ b/rt/aio_write.c
@@ -27,5 +27,5 @@ int
 aio_write (aiocbp)
      struct aiocb *aiocbp;
 {
-  return __aio_enqueue_request ((aiocb_union *) aiocbp, LIO_WRITE, 1);
+  return __aio_enqueue_request ((aiocb_union *) aiocbp, LIO_WRITE) != NULL;
 }
diff --git a/rt/aio_write64.c b/rt/aio_write64.c
index 477ce2cbe4..4d599b1062 100644
--- a/rt/aio_write64.c
+++ b/rt/aio_write64.c
@@ -27,5 +27,5 @@ int
 aio_write64 (aiocbp)
      struct aiocb64 *aiocbp;
 {
-  return __aio_enqueue_request ((aiocb_union *) aiocbp, __LIO_WRITE64, 1);
+  return __aio_enqueue_request ((aiocb_union *) aiocbp, LIO_WRITE64) != NULL;
 }
diff --git a/rt/lio_listio.c b/rt/lio_listio.c
index 73df5c2f5c..b389c6a5f6 100644
--- a/rt/lio_listio.c
+++ b/rt/lio_listio.c
@@ -20,11 +20,20 @@
 
 #include <aio.h>
 #include <errno.h>
-#include <semaphore.h>
+#include <stdlib.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;
@@ -32,8 +41,9 @@ lio_listio (mode, list, nent, sig)
      int nent;
      struct sigevent *sig;
 {
+  struct requestlist *requests[nent];
   int cnt;
-  int total = 0;
+  volatile int total = 0;
   int result = 0;
 
   /* Check arguments.  */
@@ -43,26 +53,100 @@ lio_listio (mode, list, nent, sig)
       return -1;
     }
 
-  /* Request the semaphore.  */
-  sem_wait (&__aio_requests_sema);
+  /* Request the mutex.  */
+  pthread_mutex_lock (&__aio_requests_mutex);
 
   /* Now we can enqueue all requests.  Since we already acquired the
-     semaphore the enqueue function need not do this.  */
+     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)
-      if (__aio_enqueue_request ((aiocb_union *) list[cnt],
-				 list[cnt]->aio_lio_opcode, 0) >= 0)
-	/* Successfully enqueued.  */
-	++total;
+      {
+	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;
+      }
+
+  if (total == 0)
+    {
+      /* We don't have anything to do except signalling if we work
+	 asynchronously.  */
+      if (mode == LIO_NOWAIT)
+	__aio_notify_only (sig);
+    }
+  else if (mode == LIO_WAIT)
+    {
+      pthread_cond_t cond;
+      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 = NULL;
+	    waitlist[cnt].sigevp = NULL;
+	    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)
+	if (pthread_cond_wait (&cond, &__aio_requests_mutex) == 0)
+	  --total;
+
+      /* Now it's time to restore the cancelation state.  */
+      pthread_setcancelstate (oldstate, NULL);
+    }
+  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
-	/* Signal that we've seen an error.  `errno' and the error code
-	   of the aiocb will tell more.  */
-	result = -1;
+	{
+	  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;
+		requests[cnt]->waiting = &waitlist->list[cnt];
+		++total;
+	      }
 
+	  waitlist->counter = total;
+	  waitlist->sigev = *sig;
+	}
+    }
 
-  /* Release the semaphore.  */
-  sem_post (&__aio_requests_sema);
+  /* Release the mutex.  */
+  pthread_mutex_unlock (&__aio_requests_mutex);
 
   return result;
 }
diff --git a/rt/lio_listio64.c b/rt/lio_listio64.c
index b17e9edae2..19f56adab1 100644
--- a/rt/lio_listio64.c
+++ b/rt/lio_listio64.c
@@ -20,11 +20,20 @@
 
 #include <aio.h>
 #include <errno.h>
-#include <semaphore.h>
+#include <stdlib.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;
@@ -32,8 +41,9 @@ lio_listio64 (mode, list, nent, sig)
      int nent;
      struct sigevent *sig;
 {
+  struct requestlist *requests[nent];
   int cnt;
-  int total = 0;
+  volatile int total = 0;
   int result = 0;
 
   /* Check arguments.  */
@@ -43,24 +53,99 @@ lio_listio64 (mode, list, nent, sig)
       return -1;
     }
 
-  /* Request the semaphore.  */
-  sem_wait (&__aio_requests_sema);
+  /* Request the mutex.  */
+  pthread_mutex_lock (&__aio_requests_mutex);
 
   /* Now we can enqueue all requests.  Since we already acquired the
-     semaphore the enqueue function need not do this.  */
+     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)
-      if (__aio_enqueue_request ((aiocb_union *) list[cnt],
-				 list[cnt]->aio_lio_opcode | 128, 0) >= 0)
-	/* Successfully enqueued.  */
-	++total;
+      {
+	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.  */
+      if (mode == LIO_NOWAIT)
+	__aio_notify_only (sig);
+    }
+  else if (mode == LIO_WAIT)
+    {
+      pthread_cond_t cond;
+      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;
+	    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);
+    }
+  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
-	/* Signal that we've seen an error.  `errno' and the error code
-	   of the aiocb will tell more.  */
-	result = -1;
+	{
+	  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;
+		requests[cnt]->waiting = &waitlist->list[cnt];
+		++total;
+	      }
+
+	  waitlist->counter = total;
+	  waitlist->sigev = *sig;
+	}
+    }
 
-  /* Release the semaphore.  */
-  sem_post (&__aio_requests_sema);
+  /* Release the mutex.  */
+  pthread_mutex_unlock (&__aio_requests_mutex);
 
   return result;
 }