[IA64-SGI] get XPC to cleanly disengage from remote memory references

When XPC is being shutdown (i.e., rmmod, reboot) it doesn't ensure that
other partitions with whom it was connected have completely disengaged
from any attempt at cross-partition memory references. This can lead to
MCAs in any of these other partitions when the partition is reset.

Signed-off-by: Dean Nelson <dcn@sgi.com>
Signed-off-by: Tony Luck <tony.luck@intel.com>

1 files changed, 115 insertions, 101 deletions

diff --git a/arch/ia64/sn/kernel/xpc_channel.c b/arch/ia64/sn/kernel/xpc_channel.c
index 94698bea7be0..195ac1b8e262 100644
--- a/arch/ia64/sn/kernel/xpc_channel.c
+++ b/arch/ia64/sn/kernel/xpc_channel.c
@@ -57,6 +57,7 @@ xpc_initialize_channels(struct xpc_partition *part, partid_t partid)
 
 		spin_lock_init(&ch->lock);
 		sema_init(&ch->msg_to_pull_sema, 1);	/* mutex */
+		sema_init(&ch->wdisconnect_sema, 0);	/* event wait */
 
 		atomic_set(&ch->n_on_msg_allocate_wq, 0);
 		init_waitqueue_head(&ch->msg_allocate_wq);
@@ -166,6 +167,7 @@ xpc_setup_infrastructure(struct xpc_partition *part)
 	xpc_initialize_channels(part, partid);
 
 	atomic_set(&part->nchannels_active, 0);
+	atomic_set(&part->nchannels_engaged, 0);
 
 
 	/* local_IPI_amo were set to 0 by an earlier memset() */
@@ -555,8 +557,6 @@ xpc_allocate_msgqueues(struct xpc_channel *ch)
 		sema_init(&ch->notify_queue[i].sema, 0);
 	}
 
-	sema_init(&ch->teardown_sema, 0);	/* event wait */
-
 	spin_lock_irqsave(&ch->lock, irq_flags);
 	ch->flags |= XPC_C_SETUP;
 	spin_unlock_irqrestore(&ch->lock, irq_flags);
@@ -626,6 +626,55 @@ xpc_process_connect(struct xpc_channel *ch, unsigned long *irq_flags)
 
 
 /*
+ * Notify those who wanted to be notified upon delivery of their message.
+ */
+static void
+xpc_notify_senders(struct xpc_channel *ch, enum xpc_retval reason, s64 put)
+{
+	struct xpc_notify *notify;
+	u8 notify_type;
+	s64 get = ch->w_remote_GP.get - 1;
+
+
+	while (++get < put && atomic_read(&ch->n_to_notify) > 0) {
+
+		notify = &ch->notify_queue[get % ch->local_nentries];
+
+		/*
+		 * See if the notify entry indicates it was associated with
+		 * a message who's sender wants to be notified. It is possible
+		 * that it is, but someone else is doing or has done the
+		 * notification.
+		 */
+		notify_type = notify->type;
+		if (notify_type == 0 ||
+				cmpxchg(&notify->type, notify_type, 0) !=
+								notify_type) {
+			continue;
+		}
+
+		DBUG_ON(notify_type != XPC_N_CALL);
+
+		atomic_dec(&ch->n_to_notify);
+
+		if (notify->func != NULL) {
+			dev_dbg(xpc_chan, "notify->func() called, notify=0x%p, "
+				"msg_number=%ld, partid=%d, channel=%d\n",
+				(void *) notify, get, ch->partid, ch->number);
+
+			notify->func(reason, ch->partid, ch->number,
+								notify->key);
+
+			dev_dbg(xpc_chan, "notify->func() returned, "
+				"notify=0x%p, msg_number=%ld, partid=%d, "
+				"channel=%d\n", (void *) notify, get,
+				ch->partid, ch->number);
+		}
+	}
+}
+
+
+/*
  * Free up message queues and other stuff that were allocated for the specified
  * channel.
  *
@@ -669,9 +718,6 @@ xpc_free_msgqueues(struct xpc_channel *ch)
 		ch->remote_msgqueue = NULL;
 		kfree(ch->notify_queue);
 		ch->notify_queue = NULL;
-
-		/* in case someone is waiting for the teardown to complete */
-		up(&ch->teardown_sema);
 	}
 }
 
@@ -683,7 +729,7 @@ static void
 xpc_process_disconnect(struct xpc_channel *ch, unsigned long *irq_flags)
 {
 	struct xpc_partition *part = &xpc_partitions[ch->partid];
-	u32 ch_flags = ch->flags;
+	u32 channel_was_connected = (ch->flags & XPC_C_WASCONNECTED);
 
 
 	DBUG_ON(!spin_is_locked(&ch->lock));
@@ -701,12 +747,13 @@ xpc_process_disconnect(struct xpc_channel *ch, unsigned long *irq_flags)
 	}
 	DBUG_ON(atomic_read(&ch->kthreads_assigned) != 0);
 
-	/* it's now safe to free the channel's message queues */
-
-	xpc_free_msgqueues(ch);
-	DBUG_ON(ch->flags & XPC_C_SETUP);
+	if (part->act_state == XPC_P_DEACTIVATING) {
+		/* can't proceed until the other side disengages from us */
+		if (xpc_partition_engaged(1UL << ch->partid)) {
+			return;
+		}
 
-	if (part->act_state != XPC_P_DEACTIVATING) {
+	} else {
 
 		/* as long as the other side is up do the full protocol */
 
@@ -724,16 +771,33 @@ xpc_process_disconnect(struct xpc_channel *ch, unsigned long *irq_flags)
 		}
 	}
 
+	/* wake those waiting for notify completion */
+	if (atomic_read(&ch->n_to_notify) > 0) {
+		/* >>> we do callout while holding ch->lock */
+		xpc_notify_senders(ch, ch->reason, ch->w_local_GP.put);
+	}
+
 	/* both sides are disconnected now */
 
-	ch->flags = XPC_C_DISCONNECTED;	/* clear all flags, but this one */
+	/* it's now safe to free the channel's message queues */
+	xpc_free_msgqueues(ch);
+
+	/* mark disconnected, clear all other flags except XPC_C_WDISCONNECT */
+	ch->flags = (XPC_C_DISCONNECTED | (ch->flags & XPC_C_WDISCONNECT));
 
 	atomic_dec(&part->nchannels_active);
 
-	if (ch_flags & XPC_C_WASCONNECTED) {
+	if (channel_was_connected) {
 		dev_info(xpc_chan, "channel %d to partition %d disconnected, "
 			"reason=%d\n", ch->number, ch->partid, ch->reason);
 	}
+
+	/* wake the thread that is waiting for this channel to disconnect */
+	if (ch->flags & XPC_C_WDISCONNECT) {
+		spin_unlock_irqrestore(&ch->lock, *irq_flags);
+		up(&ch->wdisconnect_sema);
+		spin_lock_irqsave(&ch->lock, *irq_flags);
+	}
 }
 
 
@@ -764,7 +828,7 @@ xpc_process_openclose_IPI(struct xpc_partition *part, int ch_number,
 		/*
 		 * If RCLOSEREQUEST is set, we're probably waiting for
 		 * RCLOSEREPLY. We should find it and a ROPENREQUEST packed
-		 * with this RCLOSEQREUQEST in the IPI_flags.
+		 * with this RCLOSEREQUEST in the IPI_flags.
 		 */
 
 		if (ch->flags & XPC_C_RCLOSEREQUEST) {
@@ -852,7 +916,7 @@ xpc_process_openclose_IPI(struct xpc_partition *part, int ch_number,
 			"channel=%d\n", args->msg_size, args->local_nentries,
 			ch->partid, ch->number);
 
-		if ((ch->flags & XPC_C_DISCONNECTING) ||
+		if ((ch->flags & (XPC_C_DISCONNECTING | XPC_C_WDISCONNECT)) ||
 					part->act_state == XPC_P_DEACTIVATING) {
 			spin_unlock_irqrestore(&ch->lock, irq_flags);
 			return;
@@ -1040,55 +1104,6 @@ xpc_connect_channel(struct xpc_channel *ch)
 
 
 /*
- * Notify those who wanted to be notified upon delivery of their message.
- */
-static void
-xpc_notify_senders(struct xpc_channel *ch, enum xpc_retval reason, s64 put)
-{
-	struct xpc_notify *notify;
-	u8 notify_type;
-	s64 get = ch->w_remote_GP.get - 1;
-
-
-	while (++get < put && atomic_read(&ch->n_to_notify) > 0) {
-
-		notify = &ch->notify_queue[get % ch->local_nentries];
-
-		/*
-		 * See if the notify entry indicates it was associated with
-		 * a message who's sender wants to be notified. It is possible
-		 * that it is, but someone else is doing or has done the
-		 * notification.
-		 */
-		notify_type = notify->type;
-		if (notify_type == 0 ||
-				cmpxchg(&notify->type, notify_type, 0) !=
-								notify_type) {
-			continue;
-		}
-
-		DBUG_ON(notify_type != XPC_N_CALL);
-
-		atomic_dec(&ch->n_to_notify);
-
-		if (notify->func != NULL) {
-			dev_dbg(xpc_chan, "notify->func() called, notify=0x%p, "
-				"msg_number=%ld, partid=%d, channel=%d\n",
-				(void *) notify, get, ch->partid, ch->number);
-
-			notify->func(reason, ch->partid, ch->number,
-								notify->key);
-
-			dev_dbg(xpc_chan, "notify->func() returned, "
-				"notify=0x%p, msg_number=%ld, partid=%d, "
-				"channel=%d\n", (void *) notify, get,
-				ch->partid, ch->number);
-		}
-	}
-}
-
-
-/*
  * Clear some of the msg flags in the local message queue.
  */
 static inline void
@@ -1240,6 +1255,7 @@ xpc_process_channel_activity(struct xpc_partition *part)
 	u64 IPI_amo, IPI_flags;
 	struct xpc_channel *ch;
 	int ch_number;
+	u32 ch_flags;
 
 
 	IPI_amo = xpc_get_IPI_flags(part);
@@ -1266,8 +1282,9 @@ xpc_process_channel_activity(struct xpc_partition *part)
 			xpc_process_openclose_IPI(part, ch_number, IPI_flags);
 		}
 
+		ch_flags = ch->flags;	/* need an atomic snapshot of flags */
 
-		if (ch->flags & XPC_C_DISCONNECTING) {
+		if (ch_flags & XPC_C_DISCONNECTING) {
 			spin_lock_irqsave(&ch->lock, irq_flags);
 			xpc_process_disconnect(ch, &irq_flags);
 			spin_unlock_irqrestore(&ch->lock, irq_flags);
@@ -1278,9 +1295,9 @@ xpc_process_channel_activity(struct xpc_partition *part)
 			continue;
 		}
 
-		if (!(ch->flags & XPC_C_CONNECTED)) {
-			if (!(ch->flags & XPC_C_OPENREQUEST)) {
-				DBUG_ON(ch->flags & XPC_C_SETUP);
+		if (!(ch_flags & XPC_C_CONNECTED)) {
+			if (!(ch_flags & XPC_C_OPENREQUEST)) {
+				DBUG_ON(ch_flags & XPC_C_SETUP);
 				(void) xpc_connect_channel(ch);
 			} else {
 				spin_lock_irqsave(&ch->lock, irq_flags);
@@ -1305,8 +1322,8 @@ xpc_process_channel_activity(struct xpc_partition *part)
 
 
 /*
- * XPC's heartbeat code calls this function to inform XPC that a partition has
- * gone down.  XPC responds by tearing down the XPartition Communication
+ * XPC's heartbeat code calls this function to inform XPC that a partition is
+ * going down.  XPC responds by tearing down the XPartition Communication
  * infrastructure used for the just downed partition.
  *
  * XPC's heartbeat code will never call this function and xpc_partition_up()
@@ -1314,7 +1331,7 @@ xpc_process_channel_activity(struct xpc_partition *part)
  * at the same time.
  */
 void
-xpc_partition_down(struct xpc_partition *part, enum xpc_retval reason)
+xpc_partition_going_down(struct xpc_partition *part, enum xpc_retval reason)
 {
 	unsigned long irq_flags;
 	int ch_number;
@@ -1330,12 +1347,11 @@ xpc_partition_down(struct xpc_partition *part, enum xpc_retval reason)
 	}
 
 
-	/* disconnect all channels associated with the downed partition */
+	/* disconnect channels associated with the partition going down */
 
 	for (ch_number = 0; ch_number < part->nchannels; ch_number++) {
 		ch = &part->channels[ch_number];
 
-
 		xpc_msgqueue_ref(ch);
 		spin_lock_irqsave(&ch->lock, irq_flags);
 
@@ -1370,6 +1386,7 @@ xpc_teardown_infrastructure(struct xpc_partition *part)
 	 * this partition.
 	 */
 
+	DBUG_ON(atomic_read(&part->nchannels_engaged) != 0);
 	DBUG_ON(atomic_read(&part->nchannels_active) != 0);
 	DBUG_ON(part->setup_state != XPC_P_SETUP);
 	part->setup_state = XPC_P_WTEARDOWN;
@@ -1506,8 +1523,12 @@ xpc_initiate_disconnect(int ch_number)
 
 			spin_lock_irqsave(&ch->lock, irq_flags);
 
-			XPC_DISCONNECT_CHANNEL(ch, xpcUnregistering,
+			if (!(ch->flags & XPC_C_DISCONNECTED)) {
+				ch->flags |= XPC_C_WDISCONNECT;
+
+				XPC_DISCONNECT_CHANNEL(ch, xpcUnregistering,
 								&irq_flags);
+			}
 
 			spin_unlock_irqrestore(&ch->lock, irq_flags);
 
@@ -1523,8 +1544,9 @@ xpc_initiate_disconnect(int ch_number)
 /*
  * To disconnect a channel, and reflect it back to all who may be waiting.
  *
- * >>> An OPEN is not allowed until XPC_C_DISCONNECTING is cleared by
- * >>> xpc_free_msgqueues().
+ * An OPEN is not allowed until XPC_C_DISCONNECTING is cleared by
+ * xpc_process_disconnect(), and if set, XPC_C_WDISCONNECT is cleared by
+ * xpc_disconnect_wait().
  *
  * THE CHANNEL IS TO BE LOCKED BY THE CALLER AND WILL REMAIN LOCKED UPON RETURN.
  */
@@ -1532,7 +1554,7 @@ void
 xpc_disconnect_channel(const int line, struct xpc_channel *ch,
 			enum xpc_retval reason, unsigned long *irq_flags)
 {
-	u32 flags;
+	u32 channel_was_connected = (ch->flags & XPC_C_CONNECTED);
 
 
 	DBUG_ON(!spin_is_locked(&ch->lock));
@@ -1547,61 +1569,53 @@ xpc_disconnect_channel(const int line, struct xpc_channel *ch,
 
 	XPC_SET_REASON(ch, reason, line);
 
-	flags = ch->flags;
+	ch->flags |= (XPC_C_CLOSEREQUEST | XPC_C_DISCONNECTING);
 	/* some of these may not have been set */
 	ch->flags &= ~(XPC_C_OPENREQUEST | XPC_C_OPENREPLY |
 			XPC_C_ROPENREQUEST | XPC_C_ROPENREPLY |
 			XPC_C_CONNECTING | XPC_C_CONNECTED);
 
-	ch->flags |= (XPC_C_CLOSEREQUEST | XPC_C_DISCONNECTING);
 	xpc_IPI_send_closerequest(ch, irq_flags);
 
-	if (flags & XPC_C_CONNECTED) {
+	if (channel_was_connected) {
 		ch->flags |= XPC_C_WASCONNECTED;
 	}
 
+	spin_unlock_irqrestore(&ch->lock, *irq_flags);
+
+	/* wake all idle kthreads so they can exit */
 	if (atomic_read(&ch->kthreads_idle) > 0) {
-		/* wake all idle kthreads so they can exit */
 		wake_up_all(&ch->idle_wq);
 	}
 
-	spin_unlock_irqrestore(&ch->lock, *irq_flags);
-
-
 	/* wake those waiting to allocate an entry from the local msg queue */
-
 	if (atomic_read(&ch->n_on_msg_allocate_wq) > 0) {
 		wake_up(&ch->msg_allocate_wq);
 	}
 
-	/* wake those waiting for notify completion */
-
-	if (atomic_read(&ch->n_to_notify) > 0) {
-		xpc_notify_senders(ch, reason, ch->w_local_GP.put);
-	}
-
 	spin_lock_irqsave(&ch->lock, *irq_flags);
 }
 
 
 void
-xpc_disconnected_callout(struct xpc_channel *ch)
+xpc_disconnecting_callout(struct xpc_channel *ch)
 {
 	/*
-	 * Let the channel's registerer know that the channel is now
+	 * Let the channel's registerer know that the channel is being
 	 * disconnected. We don't want to do this if the registerer was never
-	 * informed of a connection being made, unless the disconnect was for
-	 * abnormal reasons.
+	 * informed of a connection being made.
 	 */
 
 	if (ch->func != NULL) {
-		dev_dbg(xpc_chan, "ch->func() called, reason=%d, partid=%d, "
-			"channel=%d\n", ch->reason, ch->partid, ch->number);
+		dev_dbg(xpc_chan, "ch->func() called, reason=xpcDisconnecting,"
+			" partid=%d, channel=%d\n", ch->partid, ch->number);
 
-		ch->func(ch->reason, ch->partid, ch->number, NULL, ch->key);
+		ch->func(xpcDisconnecting, ch->partid, ch->number, NULL,
+								ch->key);
 
-		dev_dbg(xpc_chan, "ch->func() returned, reason=%d, partid=%d, "
-			"channel=%d\n", ch->reason, ch->partid, ch->number);
+		dev_dbg(xpc_chan, "ch->func() returned, reason="
+			"xpcDisconnecting, partid=%d, channel=%d\n",
+			ch->partid, ch->number);
 	}
 }
 
@@ -1848,7 +1862,7 @@ xpc_send_msg(struct xpc_channel *ch, struct xpc_msg *msg, u8 notify_type,
 			xpc_notify_func func, void *key)
 {
 	enum xpc_retval ret = xpcSuccess;
-	struct xpc_notify *notify = NULL;   // >>> to keep the compiler happy!!
+	struct xpc_notify *notify = notify;
 	s64 put, msg_number = msg->number;