Re: [PATCH v2.6.33 2/2] jme: Adding lock to protect vlgrp structure.

[Eric Dumazet <eric.dumazet@gmail.com>]

Le mardi 16 mars 2010 à 15:15 +0800, Guo-Fu Tseng a écrit :
> On Mon, 15 Mar 2010 15:53:50 -0700 (PDT), David Miller wrote
> > From: "Guo-Fu Tseng" <cooldavid@cooldavid.org>
> > Date: Tue, 16 Mar 2010 03:13:33 +0800
> > 
> > > The vlan_rx_register is called through ioctl.
> > > And the packet feeding is called in the tasklet.
> > > I see no lock in register_vlan_dev(), register_vlan_device(), and vlan_ioctl_handler()
> > > which is related to the vlan_hwaccel_receive_skb(), vlan_hwaccel_rx().
> > > 
> > > It prevents the vlgrp pointer be modified while trying to feed the packet.
> > 
> > This is not how you fix this.  Adding a new lock to your hot code
> > path of packet receiving is the last thing you should be doing.
> > 
> > Instead, do what other drivers do, take down the device and bring it
> > back up again when changing the ->vlgrp pointer.
> > 
> > See drivers/net/tg3.c:tg3_vlan_rx_register() for an example.
> I see. I'll work on it.

Then if driver is NAPI enabled, another possibility would be to rely on
RCU, since vgrp are already freed after a RCU grace period.

Completely untested patch, for ease of discussion

diff --git a/drivers/net/jme.c b/drivers/net/jme.c
index 0f31497..e19de8d 100644
--- a/drivers/net/jme.c
+++ b/drivers/net/jme.c
@@ -942,8 +942,10 @@ jme_alloc_and_feed_skb(struct jme_adapter *jme, int idx)
 			skb->ip_summed = CHECKSUM_NONE;
 
 		if (rxdesc->descwb.flags & cpu_to_le16(RXWBFLAG_TAGON)) {
-			if (jme->vlgrp) {
-				jme->jme_vlan_rx(skb, jme->vlgrp,
+			struct vlan_group *grp = rcu_dereference(jme->vlgrp);
+
+			if (grp) {
+				jme->jme_vlan_rx(skb, grp,
 					le16_to_cpu(rxdesc->descwb.vlan));
 				NET_STAT(jme).rx_bytes += 4;
 			}