Redesigned how cancellation unblocks a thread from internal cancellation points (sem_wait, pthread_join, pthread_cond_{wait,timedwait}). Cancellation won't eat a signal in any of these functions (*required* by POSIX and Single Unix Spec!).

2000-01-03  Kaz Kylheku  <kaz@ashi.footprints.net>

	Redesigned how cancellation unblocks a thread from internal
	cancellation points (sem_wait, pthread_join,
	pthread_cond_{wait,timedwait}).
	Cancellation won't eat a signal in any of these functions
	(*required* by POSIX and Single Unix Spec!).
	* condvar.c: spontaneous wakeup on pthread_cond_timedwait won't eat a
	simultaneous condition variable signal (not required by POSIX
	or Single Unix Spec, but nice).
	* spinlock.c: __pthread_lock queues back any received restarts
	that don't belong to it instead of assuming ownership of lock
	upon any restart; fastlock can no longer be acquired by two threads
	simultaneously.
	* restart.h: restarts queue even on kernels that don't have
	queued real time signals (2.0, early 2.1), thanks to atomic counter,
	avoiding a rare race condition in pthread_cond_timedwait.

13 files changed, 610 insertions, 102 deletions

diff --git a/linuxthreads/ChangeLog b/linuxthreads/ChangeLog
index 37be0e829f..78e8552a6f 100644
--- a/linuxthreads/ChangeLog
+++ b/linuxthreads/ChangeLog
@@ -1,3 +1,21 @@
+2000-01-03  Kaz Kylheku  <kaz@ashi.footprints.net>
+
+	Redesigned how cancellation unblocks a thread from internal
+	cancellation points (sem_wait, pthread_join,
+	pthread_cond_{wait,timedwait}).
+	Cancellation won't eat a signal in any of these functions
+	(*required* by POSIX and Single Unix Spec!).
+	* condvar.c: spontaneous wakeup on pthread_cond_timedwait won't eat a
+	simultaneous condition variable signal (not required by POSIX
+	or Single Unix Spec, but nice).
+	* spinlock.c: __pthread_lock queues back any received restarts
+	that don't belong to it instead of assuming ownership of lock
+	upon any restart; fastlock can no longer be acquired by two threads
+	simultaneously.
+	* restart.h: restarts queue even on kernels that don't have
+	queued real time signals (2.0, early 2.1), thanks to atomic counter,
+	avoiding a rare race condition in pthread_cond_timedwait.
+
 1999-12-31  Andreas Jaeger  <aj@suse.de>
 
 	* internals.h: Remove duplicate prototype declarations.
diff --git a/linuxthreads/cancel.c b/linuxthreads/cancel.c
index c45cac97a3..8fd8c1e60f 100644
--- a/linuxthreads/cancel.c
+++ b/linuxthreads/cancel.c
@@ -52,16 +52,54 @@ int pthread_cancel(pthread_t thread)
 {
   pthread_handle handle = thread_handle(thread);
   int pid;
+  int dorestart = 0;
+  pthread_descr th;
+  pthread_extricate_if *pextricate;
 
   __pthread_lock(&handle->h_lock, NULL);
   if (invalid_handle(handle, thread)) {
     __pthread_unlock(&handle->h_lock);
     return ESRCH;
   }
-  handle->h_descr->p_canceled = 1;
-  pid = handle->h_descr->p_pid;
+
+  th = handle->h_descr;
+
+  if (th->p_canceled) {
+    __pthread_unlock(&handle->h_lock);
+    return 0;
+  }
+
+  pextricate = th->p_extricate;
+  th->p_canceled = 1;
+  pid = th->p_pid;
+
+  /* If the thread has registered an extrication interface, then
+     invoke the interface. If it returns 1, then we succeeded in
+     dequeuing the thread from whatever waiting object it was enqueued
+     with. In that case, it is our responsibility to wake it up. 
+     And also to set the p_woken_by_cancel flag so the woken thread
+     can tell that it was woken by cancellation. */
+
+  if (pextricate != NULL) {
+    dorestart = pextricate->pu_extricate_func(pextricate->pu_object, th);
+    th->p_woken_by_cancel = dorestart;
+  }
+
   __pthread_unlock(&handle->h_lock);
-   kill(pid, __pthread_sig_cancel);
+
+  /* If the thread has suspended or is about to, then we unblock it by
+     issuing a restart, instead of a cancel signal. Otherwise we send
+     the cancel signal to unblock the thread from a cancellation point,
+     or to initiate asynchronous cancellation. The restart is needed so
+     we have proper accounting of restarts; suspend decrements the thread's
+     resume count, and restart() increments it.  This also means that suspend's
+     handling of the cancel signal is obsolete. */
+
+  if (dorestart)
+    restart(th);
+  else 
+    kill(pid, __pthread_sig_cancel);
+
   return 0;
 }
 
diff --git a/linuxthreads/condvar.c b/linuxthreads/condvar.c
index 2ea7513c68..87a93a9115 100644
--- a/linuxthreads/condvar.c
+++ b/linuxthreads/condvar.c
@@ -25,6 +25,22 @@
 #include "queue.h"
 #include "restart.h"
 
+static int pthread_cond_timedwait_relative_old(pthread_cond_t *,
+    pthread_mutex_t *, const struct timespec *);
+
+static int pthread_cond_timedwait_relative_new(pthread_cond_t *,
+    pthread_mutex_t *, const struct timespec *);
+
+static int (*pthread_cond_tw_rel)(pthread_cond_t *, pthread_mutex_t *,
+    const struct timespec *) = pthread_cond_timedwait_relative_old;
+
+/* initialize this module */
+void __pthread_init_condvar(int rt_sig_available)
+{
+  if (rt_sig_available)
+    pthread_cond_tw_rel = pthread_cond_timedwait_relative_new;
+}
+
 int pthread_cond_init(pthread_cond_t *cond,
                       const pthread_condattr_t *cond_attr)
 {
@@ -39,54 +55,125 @@ int pthread_cond_destroy(pthread_cond_t *cond)
   return 0;
 }
 
+/* Function called by pthread_cancel to remove the thread from
+   waiting on a condition variable queue. */
+
+static int cond_extricate_func(void *obj, pthread_descr th)
+{
+  volatile pthread_descr self = thread_self();
+  pthread_cond_t *cond = obj;
+  int did_remove = 0;
+
+  __pthread_lock(&cond->__c_lock, self);
+  did_remove = remove_from_queue(&cond->__c_waiting, th);
+  __pthread_unlock(&cond->__c_lock);
+
+  return did_remove;
+}
+
 int pthread_cond_wait(pthread_cond_t *cond, pthread_mutex_t *mutex)
 {
   volatile pthread_descr self = thread_self();
+  pthread_extricate_if extr;
+  int already_canceled = 0;
+
+  /* Set up extrication interface */
+  extr.pu_object = cond;
+  extr.pu_extricate_func = cond_extricate_func;
+
+  /* Register extrication interface */
+  __pthread_set_own_extricate_if(self, &extr);
+
+  /* Atomically enqueue thread for waiting, but only if it is not
+     canceled. If the thread is canceled, then it will fall through the
+     suspend call below, and then call pthread_exit without
+     having to worry about whether it is still on the condition variable queue.
+     This depends on pthread_cancel setting p_canceled before calling the
+     extricate function. */
 
   __pthread_lock(&cond->__c_lock, self);
-  enqueue(&cond->__c_waiting, self);
+  if (!(THREAD_GETMEM(self, p_canceled)
+      && THREAD_GETMEM(self, p_cancelstate) == PTHREAD_CANCEL_ENABLE))
+    enqueue(&cond->__c_waiting, self);
+  else
+    already_canceled = 1;
   __pthread_unlock(&cond->__c_lock);
+
+  if (already_canceled) {
+    __pthread_set_own_extricate_if(self, 0);
+    pthread_exit(PTHREAD_CANCELED);
+  }
+
   pthread_mutex_unlock(mutex);
-  suspend_with_cancellation(self);
-  pthread_mutex_lock(mutex);
-  /* This is a cancellation point */
-  if (THREAD_GETMEM(self, p_canceled)
+
+  suspend(self);
+  __pthread_set_own_extricate_if(self, 0);
+
+  /* Check for cancellation again, to provide correct cancellation
+     point behavior */
+
+  if (THREAD_GETMEM(self, p_woken_by_cancel)
       && THREAD_GETMEM(self, p_cancelstate) == PTHREAD_CANCEL_ENABLE) {
-    /* Remove ourselves from the waiting queue if we're still on it */
-    __pthread_lock(&cond->__c_lock, self);
-    remove_from_queue(&cond->__c_waiting, self);
-    __pthread_unlock(&cond->__c_lock);
+    THREAD_SETMEM(self, p_woken_by_cancel, 0);
+    pthread_mutex_lock(mutex);
     pthread_exit(PTHREAD_CANCELED);
   }
+
+  pthread_mutex_lock(mutex);
   return 0;
 }
 
+/* The following function is used on kernels that don't have rt signals.
+   SIGUSR1 is used as the restart signal. The different code is needed
+   because that ordinary signal does not queue. */
+
 static int
-pthread_cond_timedwait_relative(pthread_cond_t *cond,
+pthread_cond_timedwait_relative_old(pthread_cond_t *cond,
 				pthread_mutex_t *mutex,
 				const struct timespec * reltime)
 {
   volatile pthread_descr self = thread_self();
   sigset_t unblock, initial_mask;
-  int retsleep;
+  int retsleep, already_canceled, was_signalled;
   sigjmp_buf jmpbuf;
+  pthread_extricate_if extr;
+
+requeue_and_wait_again:
 
-  /* Wait on the condition */
+  retsleep = 0;
+  already_canceled = 0;
+  was_signalled = 0;
+
+  /* Set up extrication interface */
+  extr.pu_object = cond;
+  extr.pu_extricate_func = cond_extricate_func;
+
+  /* Register extrication interface */
+  __pthread_set_own_extricate_if(self, &extr);
+
+  /* Enqueue to wait on the condition and check for cancellation. */
   __pthread_lock(&cond->__c_lock, self);
-  enqueue(&cond->__c_waiting, self);
+  if (!(THREAD_GETMEM(self, p_canceled)
+      && THREAD_GETMEM(self, p_cancelstate) == PTHREAD_CANCEL_ENABLE))
+    enqueue(&cond->__c_waiting, self);
+  else
+    already_canceled = 1;
   __pthread_unlock(&cond->__c_lock);
+
+  if (already_canceled) {
+    __pthread_set_own_extricate_if(self, 0);
+    pthread_exit(PTHREAD_CANCELED);
+  }
+
   pthread_mutex_unlock(mutex);
- continue_waiting:
-  /* Set up a longjmp handler for the restart and cancel signals */
-  if (sigsetjmp(jmpbuf, 1) == 0) {
-    THREAD_SETMEM(self, p_signal_jmp, &jmpbuf);
-    THREAD_SETMEM(self, p_cancel_jmp, &jmpbuf);
-    THREAD_SETMEM(self, p_signal, 0);
-    /* Check for cancellation */
-    if (THREAD_GETMEM(self, p_canceled)
-	&& THREAD_GETMEM(self, p_cancelstate) == PTHREAD_CANCEL_ENABLE) {
-      retsleep = -1;
-    } else {
+
+  if (atomic_decrement(&self->p_resume_count) == 0) {
+    /* Set up a longjmp handler for the restart signal, unblock
+       the signal and sleep. */
+
+    if (sigsetjmp(jmpbuf, 1) == 0) {
+      THREAD_SETMEM(self, p_signal_jmp, &jmpbuf);
+      THREAD_SETMEM(self, p_signal, 0);
       /* Unblock the restart signal */
       sigemptyset(&unblock);
       sigaddset(&unblock, __pthread_sig_restart);
@@ -95,37 +182,180 @@ pthread_cond_timedwait_relative(pthread_cond_t *cond,
       retsleep = __libc_nanosleep(reltime, NULL);
       /* Block the restart signal again */
       sigprocmask(SIG_SETMASK, &initial_mask, NULL);
+      was_signalled = 0;
+    } else {
+      retsleep = -1;
+      was_signalled = 1;
     }
-  } else {
-    retsleep = -1;
+    THREAD_SETMEM(self, p_signal_jmp, NULL);
   }
-  THREAD_SETMEM(self, p_signal_jmp, NULL);
-  THREAD_SETMEM(self, p_cancel_jmp, NULL);
-  /* Here, either the condition was signaled (self->p_signal != 0)
-                   or we got canceled (self->p_canceled != 0)
-                   or the timeout occurred (retsleep == 0)
-                   or another interrupt occurred (retsleep == -1) */
-  /* This is a cancellation point */
-  if (THREAD_GETMEM(self, p_canceled)
+
+  /* Now was_signalled is true if we exited the above code
+     due to the delivery of a restart signal.  In that case,
+     we know we have been dequeued and resumed and that the
+     resume count is balanced.  Otherwise, there are some
+     cases to consider. First, try to bump up the resume count
+     back to zero. If it goes to 1, it means restart() was
+     invoked on this thread. The signal must be consumed
+     and the count bumped down and everything is cool.
+     Otherwise, no restart was delivered yet, so we remove
+     the thread from the queue. If this succeeds, it's a clear
+     case of timeout. If we fail to remove from the queue, then we
+     must wait for a restart. */
+
+  if (!was_signalled) {
+    if (atomic_increment(&self->p_resume_count) != -1) {
+      __pthread_wait_for_restart_signal(self);
+      atomic_decrement(&self->p_resume_count); /* should be zero now! */
+    } else {
+      int was_on_queue;
+      __pthread_lock(&cond->__c_lock, self);
+      was_on_queue = remove_from_queue(&cond->__c_waiting, self);
+      __pthread_unlock(&cond->__c_lock);
+
+      if (was_on_queue) {
+	__pthread_set_own_extricate_if(self, 0);
+	pthread_mutex_lock(mutex);
+
+	if (retsleep == 0)
+	  return ETIMEDOUT;
+	/* Woken by a signal: resume waiting as
+	   required by Single Unix Specification. */
+	goto requeue_and_wait_again;
+      }
+
+      suspend(self);
+    }
+  }
+
+  __pthread_set_own_extricate_if(self, 0);
+
+  /* The remaining logic is the same as in other cancellable waits,
+     such as pthread_join sem_wait or pthread_cond wait. */
+
+  if (THREAD_GETMEM(self, p_woken_by_cancel)
       && THREAD_GETMEM(self, p_cancelstate) == PTHREAD_CANCEL_ENABLE) {
-    __pthread_lock(&cond->__c_lock, self);
-    remove_from_queue(&cond->__c_waiting, self);
-    __pthread_unlock(&cond->__c_lock);
+    THREAD_SETMEM(self, p_woken_by_cancel, 0);
     pthread_mutex_lock(mutex);
     pthread_exit(PTHREAD_CANCELED);
   }
-  /* If not signaled: also remove ourselves and return an error code, but
-     only if the timeout has elapsed.  If not, just continue waiting. */
-  if (THREAD_GETMEM(self, p_signal) == 0) {
-    if (retsleep != 0)
-      goto continue_waiting;
+
+  pthread_mutex_lock(mutex);
+  return 0;
+}
+
+/* The following function is used on new (late 2.1 and 2.2 and higher) kernels
+   that have rt signals which queue. */
+
+static int
+pthread_cond_timedwait_relative_new(pthread_cond_t *cond,
+				pthread_mutex_t *mutex,
+				const struct timespec * reltime)
+{
+  volatile pthread_descr self = thread_self();
+  sigset_t unblock, initial_mask;
+  int retsleep, already_canceled, was_signalled;
+  sigjmp_buf jmpbuf;
+  pthread_extricate_if extr;
+
+ requeue_and_wait_again:
+
+  retsleep = 0;
+  already_canceled = 0;
+  was_signalled = 0;
+
+  /* Set up extrication interface */
+  extr.pu_object = cond;
+  extr.pu_extricate_func = cond_extricate_func;
+
+  /* Register extrication interface */
+  __pthread_set_own_extricate_if(self, &extr);
+
+  /* Enqueue to wait on the condition and check for cancellation. */
+  __pthread_lock(&cond->__c_lock, self);
+  if (!(THREAD_GETMEM(self, p_canceled)
+      && THREAD_GETMEM(self, p_cancelstate) == PTHREAD_CANCEL_ENABLE))
+    enqueue(&cond->__c_waiting, self);
+  else
+    already_canceled = 1;
+  __pthread_unlock(&cond->__c_lock);
+
+  if (already_canceled) {
+    __pthread_set_own_extricate_if(self, 0);
+    pthread_exit(PTHREAD_CANCELED);
+  }
+
+  pthread_mutex_unlock(mutex);
+
+  /* Set up a longjmp handler for the restart signal, unblock
+     the signal and sleep. */
+
+  if (sigsetjmp(jmpbuf, 1) == 0) {
+    THREAD_SETMEM(self, p_signal_jmp, &jmpbuf);
+    THREAD_SETMEM(self, p_signal, 0);
+    /* Unblock the restart signal */
+    sigemptyset(&unblock);
+    sigaddset(&unblock, __pthread_sig_restart);
+    sigprocmask(SIG_UNBLOCK, &unblock, &initial_mask);
+    /* Sleep for the required duration */
+    retsleep = __libc_nanosleep(reltime, NULL);
+    /* Block the restart signal again */
+    sigprocmask(SIG_SETMASK, &initial_mask, NULL);
+    was_signalled = 0;
+  } else {
+    retsleep = -1;
+    was_signalled = 1;
+  }
+  THREAD_SETMEM(self, p_signal_jmp, NULL);
+
+  /* Now was_signalled is true if we exited the above code
+     due to the delivery of a restart signal.  In that case,
+     everything is cool. We have been removed from the queue
+     by the other thread, and consumed its signal.
+
+     Otherwise we this thread woke up spontaneously, or due to a signal other
+     than restart. The next thing to do is to try to remove the thread
+     from the queue. This may fail due to a race against another thread
+     trying to do the same. In the failed case, we know we were signalled,
+     and we may also have to consume a restart signal. */
+
+  if (!was_signalled) {
+    int was_on_queue;
+
+    /* __pthread_lock will queue back any spurious restarts that
+       may happen to it. */
+
     __pthread_lock(&cond->__c_lock, self);
-    remove_from_queue(&cond->__c_waiting, self);
+    was_on_queue = remove_from_queue(&cond->__c_waiting, self);
     __pthread_unlock(&cond->__c_lock);
+
+    if (was_on_queue) {
+      __pthread_set_own_extricate_if(self, 0);
+      pthread_mutex_lock(mutex);
+
+      if (retsleep == 0)
+	return ETIMEDOUT;
+      /* Woken by a signal: resume waiting as
+	 required by Single Unix Specification. */
+      goto requeue_and_wait_again;
+    }
+
+    /* Eat the outstanding restart() from the signaller */
+    suspend(self);
+  }
+
+  __pthread_set_own_extricate_if(self, 0);
+
+  /* The remaining logic is the same as in other cancellable waits,
+     such as pthread_join sem_wait or pthread_cond wait. */
+
+  if (THREAD_GETMEM(self, p_woken_by_cancel)
+      && THREAD_GETMEM(self, p_cancelstate) == PTHREAD_CANCEL_ENABLE) {
+    THREAD_SETMEM(self, p_woken_by_cancel, 0);
     pthread_mutex_lock(mutex);
-    return ETIMEDOUT;
+    pthread_exit(PTHREAD_CANCELED);
   }
-  /* Otherwise, return normally */
+
   pthread_mutex_lock(mutex);
   return 0;
 }
@@ -144,7 +374,9 @@ int pthread_cond_timedwait(pthread_cond_t *cond, pthread_mutex_t *mutex,
     reltime.tv_sec -= 1;
   }
   if (reltime.tv_sec < 0) return ETIMEDOUT;
-  return pthread_cond_timedwait_relative(cond, mutex, &reltime);
+
+  /* Indirect call through pointer! */
+  return pthread_cond_tw_rel(cond, mutex, &reltime);
 }
 
 int pthread_cond_signal(pthread_cond_t *cond)
diff --git a/linuxthreads/internals.h b/linuxthreads/internals.h
index c5fde1f883..8af2fca335 100644
--- a/linuxthreads/internals.h
+++ b/linuxthreads/internals.h
@@ -86,6 +86,27 @@ struct pthread_key_struct {
 
 typedef struct _pthread_descr_struct * pthread_descr;
 
+/* Callback interface for removing the thread from waiting on an
+   object if it is cancelled while waiting or about to wait.
+   This hold a pointer to the object, and a pointer to a function
+   which ``extricates'' the thread from its enqueued state.
+   The function takes two arguments: pointer to the wait object,
+   and a pointer to the thread. It returns 1 if an extrication
+   actually occured, and hence the thread must also be signalled.
+   It returns 0 if the thread had already been extricated. */
+
+typedef struct _pthread_extricate_struct {
+    void *pu_object;
+    int (*pu_extricate_func)(void *, pthread_descr);
+} pthread_extricate_if;
+
+/* Atomic counter made possible by compare_and_swap */
+
+struct pthread_atomic {
+  long p_count;
+  int p_spinlock;
+};
+
 struct _pthread_descr_struct {
   pthread_descr p_nextlive, p_prevlive;
                                 /* Double chaining of active threads */
@@ -96,6 +117,7 @@ struct _pthread_descr_struct {
   int p_priority;               /* Thread priority (== 0 if not realtime) */
   struct _pthread_fastlock * p_lock; /* Spinlock for synchronized accesses */
   int p_signal;                 /* last signal received */
+  struct pthread_atomic p_resume_count; /* number of times restart() was called on thread */
   sigjmp_buf * p_signal_jmp;    /* where to siglongjmp on a signal or NULL */
   sigjmp_buf * p_cancel_jmp;    /* where to siglongjmp on a cancel or NULL */
   char p_terminated;            /* true if terminated e.g. by pthread_exit */
@@ -108,6 +130,8 @@ struct _pthread_descr_struct {
   char p_cancelstate;           /* cancellation state */
   char p_canceltype;            /* cancellation type (deferred/async) */
   char p_canceled;              /* cancellation request pending */
+  char p_woken_by_cancel;       /* cancellation performed wakeup */
+  pthread_extricate_if *p_extricate; /* See above */
   int * p_errnop;               /* pointer to used errno variable */
   int p_errno;                  /* error returned by last system call */
   int * p_h_errnop;             /* pointer to used h_errno variable */
@@ -353,6 +377,7 @@ void __pthread_manager_sighandler(int sig);
 void __pthread_reset_main_thread(void);
 void __fresetlockfiles(void);
 void __pthread_manager_adjust_prio(int thread_prio);
+void __pthread_set_own_extricate_if(pthread_descr self, pthread_extricate_if *peif);
 
 extern int __pthread_attr_setguardsize (pthread_attr_t *__attr,
 					size_t __guardsize);
@@ -372,6 +397,21 @@ extern int __pthread_mutexattr_gettype (const pthread_mutexattr_t *__attr,
 					int *__kind);
 extern void __pthread_kill_other_threads_np (void);
 
+void __pthread_restart_old(pthread_descr th);
+void __pthread_suspend_old(pthread_descr self);
+
+void __pthread_restart_new(pthread_descr th);
+void __pthread_suspend_new(pthread_descr self);
+
+void __pthread_wait_for_restart_signal(pthread_descr self);
+
+void __pthread_init_condvar(int rt_sig_available);
+
+/* Global pointers to old or new suspend functions */
+
+extern void (*__pthread_restart)(pthread_descr);
+extern void (*__pthread_suspend)(pthread_descr);
+
 /* Prototypes for the function without cancelation support when the
    normal version has it.  */
 extern int __libc_close (int fd);
diff --git a/linuxthreads/join.c b/linuxthreads/join.c
index 71db541391..5e6b78ab3b 100644
--- a/linuxthreads/join.c
+++ b/linuxthreads/join.c
@@ -79,12 +79,37 @@ void pthread_exit(void * retval)
   _exit(0);
 }
 
+/* Function called by pthread_cancel to remove the thread from
+   waiting on a condition variable queue. */
+
+static int join_extricate_func(void *obj, pthread_descr th)
+{
+  volatile pthread_descr self = thread_self();
+  pthread_handle handle = obj;
+  pthread_descr jo;
+  int did_remove = 0;
+
+  __pthread_lock(&handle->h_lock, self);
+  jo = handle->h_descr;
+  did_remove = jo->p_joining != NULL;
+  jo->p_joining = NULL;
+  __pthread_unlock(&handle->h_lock);
+
+  return did_remove;
+}
+
 int pthread_join(pthread_t thread_id, void ** thread_return)
 {
   volatile pthread_descr self = thread_self();
   struct pthread_request request;
   pthread_handle handle = thread_handle(thread_id);
   pthread_descr th;
+  pthread_extricate_if extr;
+  int already_canceled = 0;
+
+  /* Set up extrication interface */
+  extr.pu_object = handle;
+  extr.pu_extricate_func = join_extricate_func;
 
   __pthread_lock(&handle->h_lock, self);
   if (invalid_handle(handle, thread_id)) {
@@ -103,13 +128,28 @@ int pthread_join(pthread_t thread_id, void ** thread_return)
   }
   /* If not terminated yet, suspend ourselves. */
   if (! th->p_terminated) {
-    th->p_joining = self;
+    /* Register extrication interface */
+    __pthread_set_own_extricate_if(self, &extr); 
+    if (!(THREAD_GETMEM(self, p_canceled)
+	&& THREAD_GETMEM(self, p_cancelstate) == PTHREAD_CANCEL_ENABLE))
+      th->p_joining = self;
+    else
+      already_canceled = 1;
     __pthread_unlock(&handle->h_lock);
-    suspend_with_cancellation(self);
+
+    if (already_canceled) {
+      __pthread_set_own_extricate_if(self, 0); 
+      pthread_exit(PTHREAD_CANCELED);
+    }
+
+    suspend(self);
+    /* Deregister extrication interface */
+    __pthread_set_own_extricate_if(self, 0); 
+
     /* This is a cancellation point */
-    if (THREAD_GETMEM(self, p_canceled)
+    if (THREAD_GETMEM(self, p_woken_by_cancel)
 	&& THREAD_GETMEM(self, p_cancelstate) == PTHREAD_CANCEL_ENABLE) {
-      th->p_joining = NULL;
+      THREAD_SETMEM(self, p_woken_by_cancel, 0);
       pthread_exit(PTHREAD_CANCELED);
     }
     __pthread_lock(&handle->h_lock, self);
diff --git a/linuxthreads/oldsemaphore.c b/linuxthreads/oldsemaphore.c
index 72d12d20c3..62d2812693 100644
--- a/linuxthreads/oldsemaphore.c
+++ b/linuxthreads/oldsemaphore.c
@@ -68,13 +68,29 @@ int __old_sem_init(old_sem_t *sem, int pshared, unsigned int value)
   return 0;
 }
 
+/* Function called by pthread_cancel to remove the thread from
+   waiting inside __old_sem_wait. Here we simply unconditionally
+   indicate that the thread is to be woken, by returning 1. */
+
+static int old_sem_extricate_func(void *obj, pthread_descr th)
+{
+    return 1;
+}
+
 int __old_sem_wait(old_sem_t * sem)
 {
     long oldstatus, newstatus;
     volatile pthread_descr self = thread_self();
     pthread_descr * th;
+    pthread_extricate_if extr;
+
+    /* Set up extrication interface */
+    extr.pu_object = 0;
+    extr.pu_extricate_func = old_sem_extricate_func;
 
     while (1) {
+	/* Register extrication interface */
+	__pthread_set_own_extricate_if(self, &extr); 
 	do {
             oldstatus = sem->sem_status;
             if ((oldstatus & 1) && (oldstatus != 1))
@@ -85,11 +101,15 @@ int __old_sem_wait(old_sem_t * sem)
 	    }
 	}
 	while (! sem_compare_and_swap(sem, oldstatus, newstatus));
-	if (newstatus & 1)
+	if (newstatus & 1) {
 	    /* We got the semaphore. */
+	  __pthread_set_own_extricate_if(self, 0); 
 	    return 0;
+	}
 	/* Wait for sem_post or cancellation */
-	suspend_with_cancellation(self);
+	suspend(self);
+	__pthread_set_own_extricate_if(self, 0); 
+
 	/* This is a cancellation point */
 	if (self->p_canceled && self->p_cancelstate == PTHREAD_CANCEL_ENABLE) {
 	    /* Remove ourselves from the waiting list if we're still on it */
diff --git a/linuxthreads/pthread.c b/linuxthreads/pthread.c
index dfde08f432..f082e95478 100644
--- a/linuxthreads/pthread.c
+++ b/linuxthreads/pthread.c
@@ -43,6 +43,7 @@ struct _pthread_descr_struct __pthread_initial_thread = {
   0,                          /* int p_priority */
   &__pthread_handles[0].h_lock, /* struct _pthread_fastlock * p_lock */
   0,                          /* int p_signal */
+  ATOMIC_INITIALIZER,         /* struct pthread_atomic p_resume_count */
   NULL,                       /* sigjmp_buf * p_signal_buf */
   NULL,                       /* sigjmp_buf * p_cancel_buf */
   0,                          /* char p_terminated */
@@ -55,6 +56,8 @@ struct _pthread_descr_struct __pthread_initial_thread = {
   0,                          /* char p_cancelstate */
   0,                          /* char p_canceltype */
   0,                          /* char p_canceled */
+  0,                          /* char p_woken_by_cancel */
+  NULL,                       /* struct pthread_extricate_if *p_extricate */
   NULL,                       /* int *p_errnop */
   0,                          /* int p_errno */
   NULL,                       /* int *p_h_errnop */
@@ -86,6 +89,7 @@ struct _pthread_descr_struct __pthread_manager_thread = {
   0,                          /* int p_priority */
   &__pthread_handles[1].h_lock, /* struct _pthread_fastlock * p_lock */
   0,                          /* int p_signal */
+  ATOMIC_INITIALIZER,         /* struct pthread_atomic p_resume_count */
   NULL,                       /* sigjmp_buf * p_signal_buf */
   NULL,                       /* sigjmp_buf * p_cancel_buf */
   0,                          /* char p_terminated */
@@ -98,6 +102,8 @@ struct _pthread_descr_struct __pthread_manager_thread = {
   0,                          /* char p_cancelstate */
   0,                          /* char p_canceltype */
   0,                          /* char p_canceled */
+  0,                          /* char p_woken_by_cancel */
+  NULL,                       /* struct pthread_extricate_if *p_extricate */
   &__pthread_manager_thread.p_errno, /* int *p_errnop */
   0,                          /* int p_errno */
   NULL,                       /* int *p_h_errnop */
@@ -144,6 +150,12 @@ char *__pthread_manager_thread_tos = NULL;
 int __pthread_exit_requested = 0;
 int __pthread_exit_code = 0;
 
+/* Pointers that select new or old suspend/resume functions 
+   based on availability of rt signals. */
+
+void (*__pthread_restart)(pthread_descr) = __pthread_restart_old;
+void (*__pthread_suspend)(pthread_descr) = __pthread_suspend_old;
+
 /* Communicate relevant LinuxThreads constants to gdb */
 
 const int __pthread_threads_max = PTHREAD_THREADS_MAX;
@@ -215,13 +227,18 @@ init_rtsigs (void)
       __pthread_sig_cancel = SIGUSR2;
       __pthread_sig_debug = 0;
 #endif
+      __pthread_init_condvar(0);
     }
   else
     {
 #if __SIGRTMAX - __SIGRTMIN >= 3
       current_rtmin = __SIGRTMIN + 3;
+      __pthread_restart = __pthread_restart_new;
+      __pthread_suspend = __pthread_wait_for_restart_signal;
+      __pthread_init_condvar(1);
 #else
       current_rtmin = __SIGRTMIN;
+      __pthread_init_condvar(0);
 #endif
 
       current_rtmax = __SIGRTMAX;
@@ -447,7 +464,7 @@ int __pthread_initialize_manager(void)
       raise(__pthread_sig_debug);
       /* We suspend ourself and gdb will wake us up when it is
 	 ready to handle us. */
-      suspend(thread_self());
+      __pthread_wait_for_restart_signal(thread_self());
     }
   /* Synchronize debugging of the thread manager */
   request.req_kind = REQ_DEBUG;
@@ -770,6 +787,53 @@ int __pthread_getconcurrency(void)
 }
 weak_alias (__pthread_getconcurrency, pthread_getconcurrency)
 
+void __pthread_set_own_extricate_if(pthread_descr self, pthread_extricate_if *peif)
+{
+  __pthread_lock(self->p_lock, self);
+  THREAD_SETMEM(self, p_extricate, peif);
+  __pthread_unlock(self->p_lock);
+}
+
+/* Primitives for controlling thread execution */
+
+void __pthread_wait_for_restart_signal(pthread_descr self)
+{
+  sigset_t mask;
+
+  sigprocmask(SIG_SETMASK, NULL, &mask); /* Get current signal mask */
+  sigdelset(&mask, __pthread_sig_restart); /* Unblock the restart signal */
+  do {
+    self->p_signal = 0;
+    sigsuspend(&mask);                   /* Wait for signal */
+  } while (self->p_signal !=__pthread_sig_restart );
+}
+
+/* The _old variants are for 2.0 and early 2.1 kernels which don't have RT signals.
+   On these kernels, we use SIGUSR1 and SIGUSR2 for restart and cancellation.
+   Since the restart signal does not queue, we use an atomic counter to create
+   queuing semantics. This is needed to resolve a rare race condition in
+   pthread_cond_timedwait_relative. */
+    
+void __pthread_restart_old(pthread_descr th)
+{
+  if (atomic_increment(&th->p_resume_count) == -1)
+    kill(th->p_pid, __pthread_sig_restart);
+}
+
+void __pthread_suspend_old(pthread_descr self)
+{
+  if (atomic_decrement(&self->p_resume_count) <= 0)
+    __pthread_wait_for_restart_signal(self);
+}
+
+void __pthread_restart_new(pthread_descr th)
+{
+    kill(th->p_pid, __pthread_sig_restart);
+}
+
+/* There is no __pthread_suspend_new because it would just
+   be a wasteful wrapper for __pthread_wait_for_restart_signal */
+
 /* Debugging aid */
 
 #ifdef DEBUG
diff --git a/linuxthreads/queue.h b/linuxthreads/queue.h
index fa8c5d861d..f87322f84a 100644
--- a/linuxthreads/queue.h
+++ b/linuxthreads/queue.h
@@ -43,13 +43,14 @@ static inline pthread_descr dequeue(pthread_descr * q)
   return th;
 }
 
-static inline void remove_from_queue(pthread_descr * q, pthread_descr th)
+static inline int remove_from_queue(pthread_descr * q, pthread_descr th)
 {
   for (; *q != NULL; q = &((*q)->p_nextwaiting)) {
     if (*q == th) {
       *q = th->p_nextwaiting;
       th->p_nextwaiting = NULL;
-      return;
+      return 1;
     }
   }
+  return 0;
 }
diff --git a/linuxthreads/restart.h b/linuxthreads/restart.h
index 749201391d..702d7d15c6 100644
--- a/linuxthreads/restart.h
+++ b/linuxthreads/restart.h
@@ -18,41 +18,10 @@
 
 static inline void restart(pthread_descr th)
 {
-  kill(th->p_pid, __pthread_sig_restart);
+  __pthread_restart(th); /* see pthread.c */ 
 }
 
 static inline void suspend(pthread_descr self)
 {
-  sigset_t mask;
-
-  sigprocmask(SIG_SETMASK, NULL, &mask); /* Get current signal mask */
-  sigdelset(&mask, __pthread_sig_restart); /* Unblock the restart signal */
-  do {
-    self->p_signal = 0;
-    sigsuspend(&mask);                   /* Wait for signal */
-  } while (self->p_signal !=__pthread_sig_restart );
-}
-
-#define suspend_with_cancellation(self) \
-{									      \
-  sigset_t mask;							      \
-  sigjmp_buf jmpbuf;							      \
-									      \
-  sigprocmask(SIG_SETMASK, NULL, &mask); /* Get current signal mask */	      \
-  sigdelset(&mask, __pthread_sig_restart); /* Unblock the restart signal */   \
-  /* No need to save the signal mask, we'll restore it ourselves */	      \
-  if (sigsetjmp(jmpbuf, 0) == 0) {					      \
-    self->p_cancel_jmp = &jmpbuf;					      \
-    if (! (self->p_canceled						      \
-	   && self->p_cancelstate == PTHREAD_CANCEL_ENABLE)) {		      \
-      do {								      \
-	self->p_signal = 0;						      \
-        sigsuspend(&mask);               /* Wait for a signal */	      \
-      } while (self->p_signal != __pthread_sig_restart);		      \
-    }									      \
-    self->p_cancel_jmp = NULL;						      \
-  } else {								      \
-    sigaddset(&mask, __pthread_sig_restart); /* Reblock the restart signal */ \
-    sigprocmask(SIG_SETMASK, &mask, NULL);				      \
-  }									      \
+  __pthread_suspend(self); /* see pthread.c */
 }
diff --git a/linuxthreads/rwlock.c b/linuxthreads/rwlock.c
index 1fb18a3f9e..1d78b78cdf 100644
--- a/linuxthreads/rwlock.c
+++ b/linuxthreads/rwlock.c
@@ -23,8 +23,8 @@
 #include <pthread.h>
 #include "internals.h"
 #include "queue.h"
-#include "restart.h"
 #include "spinlock.h"
+#include "restart.h"
 
 int
 pthread_rwlock_init (pthread_rwlock_t *rwlock,
diff --git a/linuxthreads/semaphore.c b/linuxthreads/semaphore.c
index cfad97aced..0297b3a1e3 100644
--- a/linuxthreads/semaphore.c
+++ b/linuxthreads/semaphore.c
@@ -38,9 +38,31 @@ int __new_sem_init(sem_t *sem, int pshared, unsigned int value)
   return 0;
 }
 
+/* Function called by pthread_cancel to remove the thread from
+   waiting inside __new_sem_wait. */
+
+static int new_sem_extricate_func(void *obj, pthread_descr th)
+{
+  volatile pthread_descr self = thread_self();
+  sem_t *sem = obj;
+  int did_remove = 0;
+
+  __pthread_lock((struct _pthread_fastlock *) &sem->__sem_lock, self);
+  did_remove = remove_from_queue(&sem->__sem_waiting, th);
+  __pthread_unlock((struct _pthread_fastlock *) &sem->__sem_lock);
+
+  return did_remove;
+}
+
 int __new_sem_wait(sem_t * sem)
 {
   volatile pthread_descr self = thread_self();
+  pthread_extricate_if extr;
+  int already_canceled = 0;
+
+  /* Set up extrication interface */
+  extr.pu_object = sem;
+  extr.pu_extricate_func = new_sem_extricate_func;
 
   __pthread_lock((struct _pthread_fastlock *) &sem->__sem_lock, self);
   if (sem->__sem_value > 0) {
@@ -48,17 +70,31 @@ int __new_sem_wait(sem_t * sem)
     __pthread_unlock((struct _pthread_fastlock *) &sem->__sem_lock);
     return 0;
   }
-  enqueue(&sem->__sem_waiting, self);
-  /* Wait for sem_post or cancellation */
+  /* Register extrication interface */
+  __pthread_set_own_extricate_if(self, &extr); 
+  /* Enqueue only if not already cancelled. */
+  if (!(THREAD_GETMEM(self, p_canceled)
+      && THREAD_GETMEM(self, p_cancelstate) == PTHREAD_CANCEL_ENABLE))
+    enqueue(&sem->__sem_waiting, self);
+  else
+    already_canceled = 1;
   __pthread_unlock((struct _pthread_fastlock *) &sem->__sem_lock);
-  suspend_with_cancellation(self);
-  /* This is a cancellation point */
-  if (THREAD_GETMEM(self, p_canceled)
+
+  if (already_canceled) {
+    __pthread_set_own_extricate_if(self, 0); 
+    pthread_exit(PTHREAD_CANCELED);
+  } 
+
+  /* Wait for sem_post or cancellation, or fall through if already canceled */
+  suspend(self);
+  __pthread_set_own_extricate_if(self, 0); 
+
+  /* Terminate only if the wakeup came from cancellation. */
+  /* Otherwise ignore cancellation because we got the semaphore. */
+
+  if (THREAD_GETMEM(self, p_woken_by_cancel)
       && THREAD_GETMEM(self, p_cancelstate) == PTHREAD_CANCEL_ENABLE) {
-    /* Remove ourselves from the waiting list if we're still on it */
-    __pthread_lock((struct _pthread_fastlock *) &sem->__sem_lock, self);
-    remove_from_queue(&sem->__sem_waiting, self);
-    __pthread_unlock((struct _pthread_fastlock *) &sem->__sem_lock);
+    THREAD_SETMEM(self, p_woken_by_cancel, 0);
     pthread_exit(PTHREAD_CANCELED);
   }
   /* We got the semaphore */
diff --git a/linuxthreads/spinlock.c b/linuxthreads/spinlock.c
index ce6ff9e310..b1a99d9753 100644
--- a/linuxthreads/spinlock.c
+++ b/linuxthreads/spinlock.c
@@ -40,6 +40,7 @@ void internal_function __pthread_lock(struct _pthread_fastlock * lock,
 				      pthread_descr self)
 {
   long oldstatus, newstatus;
+  int spurious_wakeup_count = 0;
 
   do {
     oldstatus = lock->__status;
@@ -56,7 +57,28 @@ void internal_function __pthread_lock(struct _pthread_fastlock * lock,
     }
   } while(! compare_and_swap(&lock->__status, oldstatus, newstatus,
                              &lock->__spinlock));
-  if (oldstatus != 0) suspend(self);
+
+  /* Suspend with guard against spurious wakeup. 
+     This can happen in pthread_cond_timedwait_relative, when the thread
+     wakes up due to timeout and is still on the condvar queue, and then
+     locks the queue to remove itself. At that point it may still be on the
+     queue, and may be resumed by a condition signal. */
+
+  if (oldstatus != 0) {
+    for (;;) {
+      suspend(self);
+      if (self->p_nextlock != NULL) {
+	/* Count resumes that don't belong to us. */
+	spurious_wakeup_count++;
+	continue;
+      }
+      break;
+    }
+  }
+
+  /* Put back any resumes we caught that don't belong to us. */
+  while (spurious_wakeup_count--)
+    restart(self);
 }
 
 void internal_function __pthread_unlock(struct _pthread_fastlock * lock)
diff --git a/linuxthreads/spinlock.h b/linuxthreads/spinlock.h
index 29f030406c..aae18a27b4 100644
--- a/linuxthreads/spinlock.h
+++ b/linuxthreads/spinlock.h
@@ -72,3 +72,31 @@ static inline int __pthread_trylock (struct _pthread_fastlock * lock)
 }
 
 #define LOCK_INITIALIZER {0, 0}
+
+/* Operations on pthread_atomic, which is defined in internals.h */
+
+static inline long atomic_increment(struct pthread_atomic *pa)
+{
+    long oldval;
+
+    do {
+	oldval = pa->p_count;
+    } while (!compare_and_swap(&pa->p_count, oldval, oldval + 1, &pa->p_spinlock));
+
+    return oldval;
+}
+
+
+static inline long atomic_decrement(struct pthread_atomic *pa)
+{
+    long oldval;
+
+    do {
+	oldval = pa->p_count;
+    } while (!compare_and_swap(&pa->p_count, oldval, oldval - 1, &pa->p_spinlock));
+
+    return oldval;
+}
+
+#define ATOMIC_INITIALIZER { 0, 0 }
+