[rds-commits] zab commits r127 - trunk/linux/net/rds

[svn-commits@oss.oracle.com]

Author: zab
Date: 2006-07-05 19:08:52 -0500 (Wed, 05 Jul 2006)
New Revision: 127

Modified:
   trunk/linux/net/rds/flow.c
Log:
Don't use read_lock_bh() to wake the RDS socket as messaages are dropped from
the flow.  Instead we use more conservative but suboptimal irqsave locking.


Modified: trunk/linux/net/rds/flow.c
===================================================================
--- trunk/linux/net/rds/flow.c	2006-07-05 22:40:23 UTC (rev 126)
+++ trunk/linux/net/rds/flow.c	2006-07-06 00:08:52 UTC (rev 127)
@@ -225,12 +225,18 @@
 	spin_unlock_irqrestore(&flow->f_conn->c_lock, flags);
 
 	if (wake) {
-		/* careful not to race with rds_release -> sock_orphan */
+		/* 
+		 * Careful not to race with rds_release -> sock_orphan which
+		 * clears sk_sleep.  _bh() isn't OK here, we're called 
+		 * from interrupt handlers.  It's probably OK to wake the
+		 * waitqueue after sk_sleep is clear as we hold a flow ref
+		 * which holds a sock ref, but this seems more conservative.
+		 */
 		struct sock *sk = rds_rs_to_sk(flow->f_rs);
-		read_lock_bh(&sk->sk_callback_lock);
+		read_lock_irqsave(&sk->sk_callback_lock, flags);
 		if (!sock_flag(sk, SOCK_DEAD))
 			wake_up(sk->sk_sleep);
-		read_unlock_bh(&sk->sk_callback_lock);
+		read_unlock_irqrestore(&sk->sk_callback_lock, flags);
 	}
 }