Spinlocking patch for in xprt.c

Spinlocking patch for in xprt.c

[Trond Myklebust]

The following patch (taken from the zero copy networking) upgrades the
spinlocking of the xprt_(up|down)_transmit() 'semaphores' in order to
work safely with the networking bottom halves. Several of the latter
(cf. xprt.c:tcp_write_space() & friends) do want to test the value of
'xprt->snd_task'.

Cheers,
  Trond

diff -u --recursive --new-file linux-2.4.1-fh_align/net/sunrpc/xprt.c linux-2.4.1-xprt/net/sunrpc/xprt.c
--- linux-2.4.1-fh_align/net/sunrpc/xprt.c	Wed Nov 29 07:34:01 2000
+++ linux-2.4.1-xprt/net/sunrpc/xprt.c	Fri Jan 12 11:58:42 2001
@@ -1116,7 +1116,7 @@
 	struct rpc_xprt *xprt = task->tk_rqstp->rq_xprt;
 	struct rpc_rqst	*req = task->tk_rqstp;
 
-	spin_lock(&xprt_lock);
+	spin_lock_bh(&xprt_sock_lock);
 	if (xprt->snd_task && xprt->snd_task != task) {
 		dprintk("RPC: %4d TCP write queue full (task %d)\n",
 			task->tk_pid, xprt->snd_task->tk_pid);
@@ -1130,7 +1130,7 @@
 #endif
 		req->rq_bytes_sent = 0;
 	}
-	spin_unlock(&xprt_lock);
+	spin_unlock_bh(&xprt_sock_lock);
 	return xprt->snd_task == task;
 }
 
@@ -1143,10 +1143,10 @@
 	struct rpc_xprt *xprt = task->tk_rqstp->rq_xprt;
 
 	if (xprt->snd_task && xprt->snd_task == task) {
-		spin_lock(&xprt_lock);
+		spin_lock_bh(&xprt_sock_lock);
 		xprt->snd_task = NULL;
 		rpc_wake_up_next(&xprt->sending);
-		spin_unlock(&xprt_lock);
+		spin_unlock_bh(&xprt_sock_lock);
 	}
 }
 
-
To unsubscribe from this list: send the line "unsubscribe linux-kernel" in
the body of a message to majordomo@vger.kernel.org
Please read the FAQ at http://www.tux.org/lkml/

Re: Spinlocking patch for in xprt.c

[David S. Miller]

Trond, did you actually look at how this code works before
you made modifications to my fixes?

xprt_lock serializes sleep/wakeup sequences in the xprt code, so you
cannot remove xprt_lock from the sections where I added holding of
xprt_sock_lock to protect the state of xprt->snd_task.  So for
example, this part of your patch is completely bogus and will create
new corruptions and crashes:

@@ -1143,10 +1143,10 @@
 	struct rpc_xprt *xprt = task->tk_rqstp->rq_xprt;
 
 	if (xprt->snd_task && xprt->snd_task == task) {
-		spin_lock(&xprt_lock);
+		spin_lock_bh(&xprt_sock_lock);
 		xprt->snd_task = NULL;
 		rpc_wake_up_next(&xprt->sending);
-		spin_unlock(&xprt_lock);
+		spin_unlock_bh(&xprt_sock_lock);
 	}
 }

You _must_ hold both xprt_lock and xprt_sock_lock in this
section of code, not just one or just the other.

Linus, please do not apply this patch until these issues
are addressed.

Later,
David S. Miller
davem@redhat.com
-
To unsubscribe from this list: send the line "unsubscribe linux-kernel" in
the body of a message to majordomo@vger.kernel.org
Please read the FAQ at http://www.tux.org/lkml/

Re: Spinlocking patch for in xprt.c

[Trond Myklebust]

>>>>> " " == David S Miller <davem@redhat.com> writes:

     > Trond, did you actually look at how this code works before you
     > made modifications to my fixes?

     > xprt_lock serializes sleep/wakeup sequences in the xprt code,
     > so you cannot remove xprt_lock from the sections where I added
     > holding of xprt_sock_lock to protect the state of
     > xprt->snd_task.  So for example, this part of your patch is
     > completely bogus and will create new corruptions and crashes:

IIRC xprt_lock is there for 2 purposes:

  - serialize access to the TCP connect code
  - gate access to the *socket* via the xprt_(up|down)_transmit() (and
    hence setting xprt->snd_task which is a pointer to the task that
    currently is allowed to access the socket.)

Those 2 tasks are completely orthogonal to one another, so we should
be quite free to drop xprt_lock in the second case.

I can see no other places where we're using xprt_lock to protect a
sleep/wakeup of xprt->snd_task unless you're introducing it? If so for
what purpose?

Cheers,
  Trond
-
To unsubscribe from this list: send the line "unsubscribe linux-kernel" in
the body of a message to majordomo@vger.kernel.org
Please read the FAQ at http://www.tux.org/lkml/