hurd: Global signal disposition

This adds _hurd_sigstate_set_global_rcv used by libpthread to enable
POSIX-confirming behavior of signals on a per-thread basis.

This also provides a sigstate destructor _hurd_sigstate_delete, and a
global process signal state, which needs to be locked and check when
global disposition is enabled, thus the addition of _hurd_sigstate_lock
_hurd_sigstate_actions _hurd_sigstate_pending _hurd_sigstate_unlock helpers.

This also updates all the glibc code accordingly.

This also drops support for get_int(INIT_SIGMASK), which did not make sense
any more since we do not have a single signal thread any more.

During fork/spawn, this also reinitializes the child global sigstate's
lock. That cures an issue that would very rarely cause a deadlock in the
child in fork, tries to unlock ss' critical section lock at the end of
fork.  This will typically (always?) be observed in /bin/sh, which is not
surprising as that is the foremost caller of fork.

To reproduce an intermediate state, add an endless loop if
_hurd_global_sigstate is locked after __proc_dostop (cast through
volatile); that is, while still being in the fork's parent process.

When that triggers (use the libtool testsuite), the signal thread has
already locked ss (which is _hurd_global_sigstate), and is stuck at
hurdsig.c:685 in post_signal, trying to lock _hurd_siglock (which the
main thread already has locked and keeps locked until after
__task_create).  This is the case that ss->thread == MACH_PORT_NULL, that
is, a global signal.  In the main thread, between __proc_dostop and
__task_create is the __thread_abort call on the signal thread which would
abort any current kernel operation (but leave ss locked).  Later in fork,
in the parent, when _hurd_siglock is unlocked in fork, the parent's
signal thread can proceed and will unlock eventually the global sigstate.
In the client, _hurd_siglock will likewise be unlocked, but the global
sigstate never will be, as the client's signal thread has been configured
to restart execution from _hurd_msgport_receive.  Thus, when the child
tries to unlock ss' critical section lock at the end of fork, it will
first lock the global sigstate, will spin trying to lock it, which can
never be successful, and we get our deadlock.

Options seem to be:

  * Move the locking of _hurd_siglock earlier in post_signal -- but that
    may generally impact performance, if this locking isn't generally
    needed anyway?

    On the other hand, would it actually make sense to wait here until we
    are not any longer in a critical section (which is meant to disable
    signal delivery anyway (but not for preempted signals?))?

  * Clear the global sigstate in the fork's child with the rationale that
    we're anyway restarting the signal thread from a clean state.  This
    has now been implemented.

Why has this problem not been observed before Jérémie's patches?  (Or has
it?  Perhaps even more rarely?)  In _S_msg_sig_post, the signal is now
posted to a *global receiver thread*, whereas previously it was posted to
the *designated signal-receiving thread*.  The latter one was in a
critical section in fork, so didn't try to handle the signal until after
leaving the critical section?  (Not completely analyzed and verified.)

Another question is what the signal is that is being received
during/around the time __proc_dostop executes.

1 files changed, 217 insertions, 63 deletions

diff --git a/hurd/hurdsig.c b/hurd/hurdsig.c
index 5ac8c2a..13a81ed 100644
--- a/hurd/hurdsig.c
+++ b/hurd/hurdsig.c
@@ -46,9 +46,6 @@ mach_port_t _hurd_msgport;
 /* Thread listening on it.  */
 thread_t _hurd_msgport_thread;
 
-/* Thread which receives task-global signals.  */
-thread_t _hurd_sigthread;
-
 /* These are set up by _hurdsig_init.  */
 unsigned long int __hurd_sigthread_stack_base;
 unsigned long int __hurd_sigthread_stack_end;
@@ -56,6 +53,9 @@ unsigned long int __hurd_sigthread_stack_end;
 /* Linked-list of per-thread signal state.  */
 struct hurd_sigstate *_hurd_sigstates;
 
+/* Sigstate for the task-global signals.  */
+struct hurd_sigstate *_hurd_global_sigstate;
+
 /* Timeout for RPC's after interrupt_operation. */
 mach_msg_timeout_t _hurd_interrupted_rpc_timeout = 60000;
 
@@ -84,7 +84,7 @@ _hurd_thread_sigstate (thread_t thread)
     {
       ss = malloc (sizeof (*ss));
       if (ss == NULL)
-	__libc_fatal ("hurd: Can't allocate thread sigstate\n");
+	__libc_fatal ("hurd: Can't allocate sigstate\n");
       ss->thread = thread;
       __spin_lock_init (&ss->lock);
 
@@ -98,16 +98,19 @@ _hurd_thread_sigstate (thread_t thread)
       ss->intr_port = MACH_PORT_NULL;
       ss->context = NULL;
 
-      /* Initialize the sigaction vector from the default signal receiving
-	 thread's state, and its from the system defaults.  */
-      if (thread == _hurd_sigthread)
-	default_sigaction (ss->actions);
+      if (thread == MACH_PORT_NULL)
+	{
+	  /* Process-wide sigstate, use the system defaults.  */
+	  default_sigaction (ss->actions);
+
+	  /* The global sigstate is not added to the _hurd_sigstates list.
+	     It is created with _hurd_thread_sigstate (MACH_PORT_NULL)
+	     but should be accessed through _hurd_global_sigstate.  */
+	}
       else
 	{
-	  struct hurd_sigstate *s;
-	  for (s = _hurd_sigstates; s != NULL; s = s->next)
-	    if (s->thread == _hurd_sigthread)
-	      break;
+	  /* Use the global actions as a default for new threads.  */
+	  struct hurd_sigstate *s = _hurd_global_sigstate;
 	  if (s)
 	    {
 	      __spin_lock (&s->lock);
@@ -116,15 +119,115 @@ _hurd_thread_sigstate (thread_t thread)
 	    }
 	  else
 	    default_sigaction (ss->actions);
-	}
 
-      ss->next = _hurd_sigstates;
-      _hurd_sigstates = ss;
+	  ss->next = _hurd_sigstates;
+	  _hurd_sigstates = ss;
+	}
     }
   __mutex_unlock (&_hurd_siglock);
   return ss;
 }
 libc_hidden_def (_hurd_thread_sigstate)
+
+/* Destroy a sigstate structure.  Called by libpthread just before the
+ * corresponding thread is terminated (the kernel thread port must remain valid
+ * until this function is called.) */
+void
+_hurd_sigstate_delete (thread_t thread)
+{
+  struct hurd_sigstate **ssp, *ss;
+
+  __mutex_lock (&_hurd_siglock);
+  for (ssp = &_hurd_sigstates; *ssp; ssp = &(*ssp)->next)
+    if ((*ssp)->thread == thread)
+      break;
+
+  ss = *ssp;
+  if (ss)
+    *ssp = ss->next;
+
+  __mutex_unlock (&_hurd_siglock);
+  if (ss)
+    free (ss);
+}
+
+/* Make SS a global receiver, with pthread signal semantics.  */
+void
+_hurd_sigstate_set_global_rcv (struct hurd_sigstate *ss)
+{
+  assert (ss->thread != MACH_PORT_NULL);
+  ss->actions[0].sa_handler = SIG_IGN;
+}
+
+/* Check whether SS is a global receiver.  */
+static int
+sigstate_is_global_rcv (const struct hurd_sigstate *ss)
+{
+  return (_hurd_global_sigstate != NULL)
+	 && (ss->actions[0].sa_handler == SIG_IGN);
+}
+libc_hidden_def (_hurd_sigstate_delete)
+
+/* Lock/unlock a hurd_sigstate structure.  If the accessors below require
+   it, the global sigstate will be locked as well.  */
+void
+_hurd_sigstate_lock (struct hurd_sigstate *ss)
+{
+  if (sigstate_is_global_rcv (ss))
+    __spin_lock (&_hurd_global_sigstate->lock);
+  __spin_lock (&ss->lock);
+}
+void
+_hurd_sigstate_unlock (struct hurd_sigstate *ss)
+{
+  __spin_unlock (&ss->lock);
+  if (sigstate_is_global_rcv (ss))
+    __spin_unlock (&_hurd_global_sigstate->lock);
+}
+libc_hidden_def (_hurd_sigstate_set_global_rcv)
+
+/* Retreive a thread's full set of pending signals, including the global
+   ones if appropriate.  SS must be locked.  */
+sigset_t
+_hurd_sigstate_pending (const struct hurd_sigstate *ss)
+{
+  sigset_t pending = ss->pending;
+  if (sigstate_is_global_rcv (ss))
+    __sigorset (&pending, &pending, &_hurd_global_sigstate->pending);
+  return pending;
+}
+
+/* Clear a pending signal and return the associated detailed
+   signal information. SS must be locked, and must have signal SIGNO
+   pending, either directly or through the global sigstate.  */
+static struct hurd_signal_detail
+sigstate_clear_pending (struct hurd_sigstate *ss, int signo)
+{
+  if (sigstate_is_global_rcv (ss)
+      && __sigismember (&_hurd_global_sigstate->pending, signo))
+    {
+      __sigdelset (&_hurd_global_sigstate->pending, signo);
+      return _hurd_global_sigstate->pending_data[signo];
+    }
+
+  assert (__sigismember (&ss->pending, signo));
+  __sigdelset (&ss->pending, signo);
+  return ss->pending_data[signo];
+}
+libc_hidden_def (_hurd_sigstate_lock)
+libc_hidden_def (_hurd_sigstate_unlock)
+
+/* Retreive a thread's action vector.  SS must be locked.  */
+struct sigaction *
+_hurd_sigstate_actions (struct hurd_sigstate *ss)
+{
+  if (sigstate_is_global_rcv (ss))
+    return _hurd_global_sigstate->actions;
+  else
+    return ss->actions;
+}
+libc_hidden_def (_hurd_sigstate_pending)
+
 
 /* Signal delivery itself is on this page.  */
 
@@ -219,6 +322,8 @@ static void
 abort_thread (struct hurd_sigstate *ss, struct machine_thread_all_state *state,
 	      void (*reply) (void))
 {
+  assert (ss->thread != MACH_PORT_NULL);
+
   if (!(state->set & THREAD_ABORTED))
     {
       error_t err = __thread_abort (ss->thread);
@@ -367,7 +472,7 @@ _hurdsig_abort_rpcs (struct hurd_sigstate *ss, int signo, int sigthread,
 	   call above will retry their RPCs unless we clear SS->intr_port.
 	   So we clear it for the thread taking a signal when SA_RESTART is
 	   clear, so that its call returns EINTR.  */
-	if (! signo || !(ss->actions[signo].sa_flags & SA_RESTART))
+	if (! signo || !(_hurd_sigstate_actions (ss) [signo].sa_flags & SA_RESTART))
 	  ss->intr_port = MACH_PORT_NULL;
       }
 
@@ -490,9 +595,11 @@ weak_alias (_hurdsig_preemptors, _hurdsig_preempters)
 		  | sigmask (SIGSTOP) | sigmask (SIGTSTP))
 
 /* Actual delivery of a single signal.  Called with SS unlocked.  When
-   the signal is delivered, return 1 with SS locked.  If the signal is
-   being traced, return 0 with SS unlocked.   */
-static int
+   the signal is delivered, return SS, locked (or, if SS was originally
+   _hurd_global_sigstate, the sigstate of the actual thread the signal
+   was delivered to).  If the signal is being traced, return NULL with
+   SS unlocked.   */
+static struct hurd_sigstate *
 post_signal (struct hurd_sigstate *ss,
 	     int signo, struct hurd_signal_detail *detail,
 	     int untraced, void (*reply) (void))
@@ -545,8 +652,12 @@ post_signal (struct hurd_sigstate *ss,
       assert_perror (err);
       for (i = 0; i < nthreads; ++i)
 	{
-	  if (threads[i] != _hurd_msgport_thread
-	      && (act != handle || threads[i] != ss->thread))
+	  if (act == handle && threads[i] == ss->thread)
+	    {
+	      /* The thread that will run the handler is kept suspended.  */
+	      ss_suspended = 1;
+	    }
+	  else if (threads[i] != _hurd_msgport_thread)
 	    {
 	      err = __thread_resume (threads[i]);
 	      assert_perror (err);
@@ -559,9 +670,6 @@ post_signal (struct hurd_sigstate *ss,
 		       (vm_address_t) threads,
 		       nthreads * sizeof *threads);
       _hurd_stopped = 0;
-      if (act == handle)
-	/* The thread that will run the handler is already suspended.  */
-	ss_suspended = 1;
     }
 
   error_t err;
@@ -577,13 +685,43 @@ post_signal (struct hurd_sigstate *ss,
 	}
 
       /* This call is just to check for pending signals.  */
-      __spin_lock (&ss->lock);
-      return 1;
+      _hurd_sigstate_lock (ss);
+      return ss;
     }
 
   thread_state.set = 0;		/* We know nothing.  */
 
-  __spin_lock (&ss->lock);
+  _hurd_sigstate_lock (ss);
+
+  /* If this is a global signal, try to find a thread ready to accept
+     it right away.  This is especially important for untraced signals,
+     since going through the global pending mask would de-untrace them.  */
+  if (ss->thread == MACH_PORT_NULL)
+  {
+    struct hurd_sigstate *rss;
+
+    __mutex_lock (&_hurd_siglock);
+    for (rss = _hurd_sigstates; rss != NULL; rss = rss->next)
+      {
+	if (! sigstate_is_global_rcv (rss))
+	  continue;
+
+	/* The global sigstate is already locked.  */
+	__spin_lock (&rss->lock);
+	if (! __sigismember (&rss->blocked, signo))
+	  {
+	    ss = rss;
+	    break;
+	  }
+	__spin_unlock (&rss->lock);
+      }
+    __mutex_unlock (&_hurd_siglock);
+  }
+
+  /* We want the preemptors to be able to update the blocking mask
+     without affecting the delivery of this signal, so we save the
+     current value to test against later.  */
+  sigset_t blocked = ss->blocked;
 
   /* Check for a preempted signal.  Preempted signals can arrive during
      critical sections.  */
@@ -641,12 +779,12 @@ post_signal (struct hurd_sigstate *ss,
 	    mark_pending ();
 	  else
 	    suspend ();
-	  __spin_unlock (&ss->lock);
+	  _hurd_sigstate_unlock (ss);
 	  reply ();
-	  return 0;
+	  return NULL;
 	}
 
-      handler = ss->actions[signo].sa_handler;
+      handler = _hurd_sigstate_actions (ss) [signo].sa_handler;
 
       if (handler == SIG_DFL)
 	/* Figure out the default action for this signal.  */
@@ -739,9 +877,7 @@ post_signal (struct hurd_sigstate *ss,
     }
 
   /* Handle receipt of a blocked signal, or any signal while stopped.  */
-  if (act != ignore		/* Signals ignored now are forgotten now.  */
-      && __sigismember (&ss->blocked, signo)
-      || (signo != SIGKILL && _hurd_stopped))
+  if (__sigismember (&blocked, signo) || (signo != SIGKILL && _hurd_stopped))
     {
       mark_pending ();
       act = ignore;
@@ -776,6 +912,7 @@ post_signal (struct hurd_sigstate *ss,
 	 now's the time to set it going. */
       if (ss_suspended)
 	{
+	  assert (ss->thread != MACH_PORT_NULL);
 	  err = __thread_resume (ss->thread);
 	  assert_perror (err);
 	  ss_suspended = 0;
@@ -820,6 +957,8 @@ post_signal (struct hurd_sigstate *ss,
 	struct sigcontext *scp, ocontext;
 	int wait_for_reply, state_changed;
 
+	assert (ss->thread != MACH_PORT_NULL);
+
 	/* Stop the thread and abort its pending RPC operations.  */
 	if (! ss_suspended)
 	  {
@@ -956,23 +1095,25 @@ post_signal (struct hurd_sigstate *ss,
 	    }
 	}
 
+	struct sigaction *action = & _hurd_sigstate_actions (ss) [signo];
+
 	/* Backdoor extra argument to signal handler.  */
 	scp->sc_error = detail->error;
 
 	/* Block requested signals while running the handler.  */
 	scp->sc_mask = ss->blocked;
-	__sigorset (&ss->blocked, &ss->blocked, &ss->actions[signo].sa_mask);
+	__sigorset (&ss->blocked, &ss->blocked, &action->sa_mask);
 
 	/* Also block SIGNO unless we're asked not to.  */
-	if (! (ss->actions[signo].sa_flags & (SA_RESETHAND | SA_NODEFER)))
+	if (! (action->sa_flags & (SA_RESETHAND | SA_NODEFER)))
 	  __sigaddset (&ss->blocked, signo);
 
 	/* Reset to SIG_DFL if requested.  SIGILL and SIGTRAP cannot
            be automatically reset when delivered; the system silently
            enforces this restriction.  */
-	if (ss->actions[signo].sa_flags & SA_RESETHAND
+	if (action->sa_flags & SA_RESETHAND
 	    && signo != SIGILL && signo != SIGTRAP)
-	  ss->actions[signo].sa_handler = SIG_DFL;
+	  action->sa_handler = SIG_DFL;
 
 	/* Any sigsuspend call must return after the handler does.  */
 	wake_sigsuspend (ss);
@@ -990,7 +1131,7 @@ post_signal (struct hurd_sigstate *ss,
       }
     }
 
-  return 1;
+  return ss;
 }
 
 /* Return the set of pending signals in SS which should be delivered. */
@@ -1005,7 +1146,7 @@ pending_signals (struct hurd_sigstate *ss)
   if (_hurd_stopped || __spin_lock_locked (&ss->critical_section_lock))
     return 0;
 
-  return ss->pending & ~ss->blocked;
+  return _hurd_sigstate_pending (ss) & ~ss->blocked;
 }
 
 /* Post the specified pending signals in SS and return 1.  If one of
@@ -1017,12 +1158,15 @@ post_pending (struct hurd_sigstate *ss, sigset_t pending, void (*reply) (void))
   int signo;
   struct hurd_signal_detail detail;
 
+  /* Make sure SS corresponds to an actual thread, since we assume it won't
+     change in post_signal. */
+  assert (ss->thread != MACH_PORT_NULL);
+
   for (signo = 1; signo < NSIG; ++signo)
     if (__sigismember (&pending, signo))
       {
-	__sigdelset (&ss->pending, signo);
-	detail = ss->pending_data[signo];
-	__spin_unlock (&ss->lock);
+	detail = sigstate_clear_pending (ss, signo);
+	_hurd_sigstate_unlock (ss);
 
 	/* Will reacquire the lock, except if the signal is traced.  */
 	if (! post_signal (ss, signo, &detail, 0, reply))
@@ -1030,7 +1174,7 @@ post_pending (struct hurd_sigstate *ss, sigset_t pending, void (*reply) (void))
       }
 
   /* No more signals pending; SS->lock is still locked.  */
-  __spin_unlock (&ss->lock);
+  _hurd_sigstate_unlock (ss);
 
   return 1;
 }
@@ -1048,14 +1192,14 @@ post_all_pending_signals (void (*reply) (void))
       __mutex_lock (&_hurd_siglock);
       for (ss = _hurd_sigstates; ss != NULL; ss = ss->next)
         {
-	  __spin_lock (&ss->lock);
+	  _hurd_sigstate_lock (ss);
 
 	  pending = pending_signals (ss);
 	  if (pending)
 	    /* post_pending() below will unlock SS. */
 	    break;
 
-	  __spin_unlock (&ss->lock);
+	  _hurd_sigstate_unlock (ss);
 	}
       __mutex_unlock (&_hurd_siglock);
 
@@ -1088,11 +1232,12 @@ _hurd_internal_post_signal (struct hurd_sigstate *ss,
 	assert_perror (err);
     }
 
-  if (! post_signal (ss, signo, detail, untraced, reply))
+  ss = post_signal (ss, signo, detail, untraced, reply);
+  if (! ss)
     return;
 
   /* The signal was neither fatal nor traced.  We still hold SS->lock.  */
-  if (signo != 0)
+  if (signo != 0 && ss->thread != MACH_PORT_NULL)
     {
       /* The signal has either been ignored or is now being handled.  We can
 	 consider it delivered and reply to the killer.  */
@@ -1104,8 +1249,9 @@ _hurd_internal_post_signal (struct hurd_sigstate *ss,
     }
   else
     {
-      /* We need to check for pending signals for all threads.  */
-      __spin_unlock (&ss->lock);
+      /* If this was a process-wide signal or a poll request, we need
+	 to check for pending signals for all threads.  */
+      _hurd_sigstate_unlock (ss);
       if (! post_all_pending_signals (reply))
 	return;
 
@@ -1231,9 +1377,10 @@ _S_msg_sig_post (mach_port_t me,
   d.code = sigcode;
   d.exc = 0;
 
-  /* Post the signal to the designated signal-receiving thread.  This will
-     reply when the signal can be considered delivered.  */
-  _hurd_internal_post_signal (_hurd_thread_sigstate (_hurd_sigthread),
+  /* Post the signal to a global receiver thread (or mark it pending in
+     the global sigstate).  This will reply when the signal can be
+     considered delivered.  */
+  _hurd_internal_post_signal (_hurd_global_sigstate,
 			      signo, &d, reply_port, reply_port_type,
 			      0); /* Stop if traced.  */
 
@@ -1261,7 +1408,7 @@ _S_msg_sig_post_untraced (mach_port_t me,
 
   /* Post the signal to the designated signal-receiving thread.  This will
      reply when the signal can be considered delivered.  */
-  _hurd_internal_post_signal (_hurd_thread_sigstate (_hurd_sigthread),
+  _hurd_internal_post_signal (_hurd_global_sigstate,
 			      signo, &d, reply_port, reply_port_type,
 			      1); /* Untraced flag. */
 
@@ -1272,8 +1419,8 @@ extern void __mig_init (void *);
 
 #include <mach/task_special_ports.h>
 
-/* Initialize the message port and _hurd_sigthread and start the signal
-   thread.  */
+/* Initialize the message port, _hurd_global_sigstate, and start the
+   signal thread.  */
 
 void
 _hurdsig_init (const int *intarray, size_t intarraysize)
@@ -1296,27 +1443,34 @@ _hurdsig_init (const int *intarray, size_t intarraysize)
 				  MACH_MSG_TYPE_MAKE_SEND);
   assert_perror (err);
 
+  /* Initialize the global signal state.  */
+  _hurd_global_sigstate = _hurd_thread_sigstate (MACH_PORT_NULL);
+
+  /* We block all signals, and let actual threads pull them from the
+     pending mask.  */
+  __sigfillset(& _hurd_global_sigstate->blocked);
+
   /* Initialize the main thread's signal state.  */
   ss = _hurd_self_sigstate ();
 
-  /* Copy inherited values from our parent (or pre-exec process state)
-     into the signal settings of the main thread.  */
+  /* Mark it as a process-wide signal receiver.  Threads in this set use
+     the common action vector in _hurd_global_sigstate.  */
+  _hurd_sigstate_set_global_rcv (ss);
+
+  /* Copy inherited signal settings from our parent (or pre-exec process
+     state) */
   if (intarraysize > INIT_SIGMASK)
     ss->blocked = intarray[INIT_SIGMASK];
   if (intarraysize > INIT_SIGPENDING)
-    ss->pending = intarray[INIT_SIGPENDING];
+    _hurd_global_sigstate->pending = intarray[INIT_SIGPENDING];
   if (intarraysize > INIT_SIGIGN && intarray[INIT_SIGIGN] != 0)
     {
       int signo;
       for (signo = 1; signo < NSIG; ++signo)
 	if (intarray[INIT_SIGIGN] & __sigmask(signo))
-	  ss->actions[signo].sa_handler = SIG_IGN;
+	  _hurd_global_sigstate->actions[signo].sa_handler = SIG_IGN;
     }
 
-  /* Set the default thread to receive task-global signals
-     to this one, the main (first) user thread.  */
-  _hurd_sigthread = ss->thread;
-
   /* Start the signal thread listening on the message port.  */
 
 #pragma weak __cthread_fork