diff mbox

[Xen-devel] xen-netback notify DomU to send ARP.

Message ID 50ED6255.7020006@oracle.com
State RFC, archived
Delegated to: David Miller
Headers show

Commit Message

jianhai luan Jan. 9, 2013, 12:28 p.m. UTC 
On 2013-1-9 18:06, Jan Beulich wrote:
>>>> On 09.01.13 at 08:39, jianhai luan <jianhai.luan@oracle.com> wrote:
>> @@ -34,11 +35,42 @@ static void connect(struct backend_info *);
>> static void backend_create_xenvif(struct backend_info *be);
>> static void unregister_hotplug_status_watch(struct backend_info *be);
>>
>> +#define nb_to_backend(nb) container_of(nb, struct backend_info, vif_notifier)
>> +/**
>> + * When network condition of vif change, notify the frontend.
>> + */
>> +static int netback_netdev_event(struct notifier_block *this,
>> +				unsigned long event, void *ptr)
>> +{
>> +	struct net_device *event_dev = (struct net_device *)ptr;
> Pointless cast.
>
>> +	struct backend_info *be = nb_to_backend(this);
>> +
>> +	pr_debug("event_dev: %s, event: %lx\n",
>> +		 event_dev ? event_dev->name : "None", event);
>> +
>> +	if (!be->vif)
>> +		goto out;
>> +
>> +	switch (event) {
>> +	case NETDEV_NOTIFY_PEERS:
>> +		/* Notify frontend to Send gratuitous ARP */
>> +		xenbus_switch_state(be->dev, XenbusStateInitialised);
>> +		xenbus_switch_state(be->dev, );
> This is the sort of change that clearly isn't acceptable, as I don't
> think you have ways to check _all_ existing frontends for their
> compatibility with this. A connected -> connected transition
> might be acceptable (that was done in the block frontend too, for
> implementing dynamic resize), but will likely need to be
> accompanied by a frontend side patch to handle that (which so
> far should be a no-op).
The latest xen net-frontent driver have handled the condition. State 
XenbusStateInitialised will do nothing,
but change to XenbusStateConnected will trigger 
netdev_notify_peers(netdev) to send ARP.
>
>> +		break;
>> +	default:
>> +		break;
> Pointless default case.
>
>> +	}
>> +
>> +out:
> I don't think you really need the label (and the goto above) - just
> put a return there.
>
>> +	return NOTIFY_DONE;
>> +}
>> +
>> static int netback_remove(struct xenbus_device *dev)
>> {
>> 	struct backend_info *be = dev_get_drvdata(&dev->dev);
>>
>> 	unregister_hotplug_status_watch(be);
>> +	unregister_netdevice_notifier(&be->vif_notifier);
>> 	if (be->vif) {
>> 		kobject_uevent(&dev->dev.kobj, KOBJ_OFFLINE);
>> 		xenbus_rm(XBT_NIL, dev->nodename, "hotplug-status");
>> @@ -129,6 +161,10 @@ static int netback_probe(struct xenbus_device *dev,
>> 	/* This kicks hotplug scripts, so do it immediately. */
>> 	backend_create_xenvif(be);
>>
>> +	/*   Event Notify */
>> +	(be->vif_notifier).notifier_call = netback_netdev_event;
> Pointless parentheses.
>
> Jan
>
>> +	register_netdevice_notifier(&be->vif_notifier);
>> +
>> 	return 0;
>>
>> abort_transaction:
>
From a64d80cc0c780bee7a8d6e842126cb5f7d17f0d2 Mon Sep 17 00:00:00 2001
From: Jason Luan <jianhai.luan@oracle.com>
Date: Fri, 28 Dec 2012 15:43:06 +0800
Subject: [PATCH] xen-netback notify DomU to send ARP.

When Xen Dom0's network circumstance changed, DomU
should be notified in some special condition. For
example the below circumstance:
  ping from Guest A to DomU:
  Guest A --> eth0 - bond0 - xenbr0 --VIF(DOMU)
              eth1 /
  when eth0 inactive, and eth1 active.
  Guest A --> eth0   bond0 - xenbr0 --VIF(DOMU)
              eth1 /
  Guest A will don't reach to DomU. After Guest A
  send ARP request and DomU respond, Guest A will
  reach DomU. But some more second will be elapsed.
              eth0   bond0 - xenbr0 --VIF(DOMU)
  Guest A --> eth1/

If Xen netback watch the network change, will notify
DomU by change it own status. So netfront will watch
netback's change, and DomU send ARP initiative.

Signed-off-by: Jason Luan <jianhai.luan@oracle.com>
---
 drivers/net/xen-netback/xenbus.c |   33 +++++++++++++++++++++++++++++++++
 1 files changed, 33 insertions(+), 0 deletions(-)

Comments

Jan Beulich Jan. 9, 2013, 1:44 p.m. UTC | #1 
>>> On 09.01.13 at 13:28, jianhai luan <jianhai.luan@oracle.com> wrote:
>>> +	switch (event) {
>>> +	case NETDEV_NOTIFY_PEERS:
>>> +		/* Notify frontend to Send gratuitous ARP */
>>> +		xenbus_switch_state(be->dev, XenbusStateInitialised);
>>> +		xenbus_switch_state(be->dev, );
>> This is the sort of change that clearly isn't acceptable, as I don't
>> think you have ways to check _all_ existing frontends for their
>> compatibility with this. A connected -> connected transition
>> might be acceptable (that was done in the block frontend too, for
>> implementing dynamic resize), but will likely need to be
>> accompanied by a frontend side patch to handle that (which so
>> far should be a no-op).
> The latest xen net-frontent driver have handled the condition. State 
> XenbusStateInitialised will do nothing,
> but change to XenbusStateConnected will trigger 
> netdev_notify_peers(netdev) to send ARP.

Did you read my earlier reply carefully? You still only talk about
(upstream) Linux netfront, but this is not the only (possible)
frontend. You should not invoke state transitions that can -
even if only theoretically - blow up frontends. And afaict the
only thing you can safely assume frontends ought to tolerate
are transitions from Connected to Connected (or more
generally from one state to the same one, but the other
states aren't useful here, except maybe the Reconfigur* ones).

Jan

--
To unsubscribe from this list: send the line "unsubscribe netdev" in
the body of a message to majordomo@vger.kernel.org
More majordomo info at  http://vger.kernel.org/majordomo-info.html
Jason Luan Jan. 9, 2013, 3:37 p.m. UTC | #2 
于 2013年01月09日 21:44, Jan Beulich 写道:
>>>> On 09.01.13 at 13:28, jianhai luan <jianhai.luan@oracle.com> wrote:
>>>> +	switch (event) {
>>>> +	case NETDEV_NOTIFY_PEERS:
>>>> +		/* Notify frontend to Send gratuitous ARP */
>>>> +		xenbus_switch_state(be->dev, XenbusStateInitialised);
>>>> +		xenbus_switch_state(be->dev, );
>>> This is the sort of change that clearly isn't acceptable, as I don't
>>> think you have ways to check _all_ existing frontends for their
>>> compatibility with this. A connected -> connected transition
>>> might be acceptable (that was done in the block frontend too, for
>>> implementing dynamic resize), but will likely need to be
>>> accompanied by a frontend side patch to handle that (which so
>>> far should be a no-op).
>> The latest xen net-frontent driver have handled the condition. State
>> XenbusStateInitialised will do nothing,
>> but change to XenbusStateConnected will trigger
>> netdev_notify_peers(netdev) to send ARP.
> Did you read my earlier reply carefully? You still only talk about
> (upstream) Linux netfront, but this is not the only (possible)
> frontend. You should not invoke state transitions that can -
> even if only theoretically - blow up frontends. And afaict the
I only want to notify xen-netfront. I don't know what is better way?
To attainthegoal, i try to modify virtual interrupt, but the way is
morecomplicated,modified and working. So, i give up the way.
Would you like to give some suggestion about how to notify xen-netfront?
> only thing you can safely assume frontends ought to tolerate
> are transitions from Connected to Connected (or more
> generally from one state to the same one, but the other
> states aren't useful here, except maybe the Reconfigur* ones).
Sorry for that. At the beginning the patch be applied in kernel 2.6.18 to
fixed one issue. Only XenbusStateInitialised and XenbusStateClosed ( Not
Reconfigure* one) don't any thing, so i choose the XenbusStateInitialised.
Do you suggestion that i choose Reconfigure*?
>
> Jan
>
>
> _______________________________________________
> Xen-devel mailing list
> Xen-devel@lists.xen.org
> http://lists.xen.org/xen-devel


--
To unsubscribe from this list: send the line "unsubscribe netdev" in
the body of a message to majordomo@vger.kernel.org
More majordomo info at  http://vger.kernel.org/majordomo-info.html
Jan Beulich Jan. 9, 2013, 3:44 p.m. UTC | #3 
>>> On 09.01.13 at 16:37, Jason Luan <luanjianhai@163.com> wrote:
> 于 2013年01月09日 21:44, Jan Beulich 写道:
>>>>> On 09.01.13 at 13:28, jianhai luan <jianhai.luan@oracle.com> wrote:
>>>>> +	switch (event) {
>>>>> +	case NETDEV_NOTIFY_PEERS:
>>>>> +		/* Notify frontend to Send gratuitous ARP */
>>>>> +		xenbus_switch_state(be->dev, XenbusStateInitialised);
>>>>> +		xenbus_switch_state(be->dev, );
>>>> This is the sort of change that clearly isn't acceptable, as I don't
>>>> think you have ways to check _all_ existing frontends for their
>>>> compatibility with this. A connected -> connected transition
>>>> might be acceptable (that was done in the block frontend too, for
>>>> implementing dynamic resize), but will likely need to be
>>>> accompanied by a frontend side patch to handle that (which so
>>>> far should be a no-op).
>>> The latest xen net-frontent driver have handled the condition. State
>>> XenbusStateInitialised will do nothing,
>>> but change to XenbusStateConnected will trigger
>>> netdev_notify_peers(netdev) to send ARP.
>> Did you read my earlier reply carefully? You still only talk about
>> (upstream) Linux netfront, but this is not the only (possible)
>> frontend. You should not invoke state transitions that can -
>> even if only theoretically - blow up frontends. And afaict the
> I only want to notify xen-netfront. I don't know what is better way?
> To attainthegoal, i try to modify virtual interrupt, but the way is
> morecomplicated,modified and working. So, i give up the way.
> Would you like to give some suggestion about how to notify xen-netfront?
>> only thing you can safely assume frontends ought to tolerate
>> are transitions from Connected to Connected (or more
>> generally from one state to the same one, but the other
>> states aren't useful here, except maybe the Reconfigur* ones).
> Sorry for that. At the beginning the patch be applied in kernel 2.6.18 to
> fixed one issue. Only XenbusStateInitialised and XenbusStateClosed ( Not
> Reconfigure* one) don't any thing, so i choose the XenbusStateInitialised.
> Do you suggestion that i choose Reconfigure*?

I already said that I think that only a Connected->Connected
transition is to be considered here. Bringing up Reconfigur* was
just to point out that there are other states available that could
be used, but you'd first need to make sure that (a) they aren't
used for anything else and (b) frontends can reasonably be
expected to deal with (ignore) them when not originally aware
of them (as opposed to a Connected->Connected transition,
which all frontends ought to be able to deal with afaict).

Also, when you reply to earlier mails, could you - in order to make
the result readable - insert blank lines between the quoted text
and your responses, please?

Jan
--
To unsubscribe from this list: send the line "unsubscribe netdev" in
the body of a message to majordomo@vger.kernel.org
More majordomo info at  http://vger.kernel.org/majordomo-info.html
diff mbox

Patch

diff --git a/drivers/net/xen-netback/xenbus.c b/drivers/net/xen-netback/xenbus.c
index 410018c..17a3990 100644
--- a/drivers/net/xen-netback/xenbus.c
+++ b/drivers/net/xen-netback/xenbus.c
@@ -26,6 +26,7 @@  struct backend_info {
 	struct xenvif *vif;
 	enum xenbus_state frontend_state;
 	struct xenbus_watch hotplug_status_watch;
+	struct notifier_block vif_notifier;
 	u8 have_hotplug_status_watch:1;
 };
 
@@ -34,11 +35,39 @@  static void connect(struct backend_info *);
 static void backend_create_xenvif(struct backend_info *be);
 static void unregister_hotplug_status_watch(struct backend_info *be);
 
+#define nb_to_backend(nb) container_of(nb, struct backend_info, vif_notifier)
+/**
+ * When network condition of vif change, notify the frontend.
+ */
+static int netback_netdev_event(struct notifier_block *this,
+				unsigned long event, void *ptr)
+{
+	struct net_device *event_dev = ptr;
+	struct backend_info *be = nb_to_backend(this);
+
+	pr_debug("event_dev: %s, event: %lx\n",
+		 event_dev ? event_dev->name : "None", event);
+
+	if (!be->vif)
+		return NOTIFY_DONE;
+
+	switch (event) {
+	case NETDEV_NOTIFY_PEERS:
+		/* Notify frontend to Send gratuitous ARP */
+		xenbus_switch_state(be->dev, XenbusStateInitialised);
+		xenbus_switch_state(be->dev, XenbusStateConnected);
+		break;
+	}
+
+	return NOTIFY_DONE;
+}
+
 static int netback_remove(struct xenbus_device *dev)
 {
 	struct backend_info *be = dev_get_drvdata(&dev->dev);
 
 	unregister_hotplug_status_watch(be);
+	unregister_netdevice_notifier(&be->vif_notifier);
 	if (be->vif) {
 		kobject_uevent(&dev->dev.kobj, KOBJ_OFFLINE);
 		xenbus_rm(XBT_NIL, dev->nodename, "hotplug-status");
@@ -129,6 +158,10 @@  static int netback_probe(struct xenbus_device *dev,
 	/* This kicks hotplug scripts, so do it immediately. */
 	backend_create_xenvif(be);
 
+	/* Register Frontend Event Notify */
+	be->vif_notifier.notifier_call = netback_netdev_event;
+	register_netdevice_notifier(&be->vif_notifier);
+
 	return 0;
 
 abort_transaction: