[net-next] bonding: fix system hang due to fast igmp timer rescheduling

From: Nikolay Aleksandrov <Nikolay Aleksandrov nikolay@redhat.com>

After commit 4aa5dee4d9 ("net: convert resend IGMP to notifier event")
we try to acquire rtnl in bond_resend_igmp_join_requests but it can be
scheduled with rtnl already held (e.g. when bond_change_active_slave is
called with rtnl) causing a loop of immediate reschedules + calls because
rtnl_trylock fails each time since it's being already held.
For me this issue leads to system hangs very easy:
modprobe bonding; ifconfig bond0 up; ifenslave bond0 eth0; rmmod
bonding;

The fix is to introduce a small (1 jiffy) delay which is enough for the
sections holding rtnl to finish without putting any strain on the system.

Signed-off-by: Nikolay Aleksandrov <nikolay@redhat.com>
---
 drivers/net/bonding/bond_main.c | 2 +-
 1 file changed, 1 insertion(+), 1 deletion(-)

Nikolay Aleksandrov <nikolay@redhat.com> wrote:

>From: Nikolay Aleksandrov <Nikolay Aleksandrov nikolay@redhat.com>
>
>After commit 4aa5dee4d9 ("net: convert resend IGMP to notifier event")
>we try to acquire rtnl in bond_resend_igmp_join_requests but it can be
>scheduled with rtnl already held (e.g. when bond_change_active_slave is
>called with rtnl) causing a loop of immediate reschedules + calls because
>rtnl_trylock fails each time since it's being already held.
>For me this issue leads to system hangs very easy:
>modprobe bonding; ifconfig bond0 up; ifenslave bond0 eth0; rmmod
>bonding;

	I believe that bond_change_active_slave is always called with
rtnl held, and it is the only caller of bond_resend_igmp_join_requests
(well, "caller" in the sense that it queues the delayed work for
mcast_work that runs the function, currently with delay of 0).

>The fix is to introduce a small (1 jiffy) delay which is enough for the
>sections holding rtnl to finish without putting any strain on the system.

	Should the delay also be in the bond_change_active_slave queue
work call as well, to eliminate one loop of the "rtnl_trylock failing ->
queue_delayed_work" sequence in bond_resend_igmp_join_requests?

	-J

>Signed-off-by: Nikolay Aleksandrov <nikolay@redhat.com>
>---
> drivers/net/bonding/bond_main.c | 2 +-
> 1 file changed, 1 insertion(+), 1 deletion(-)
>
>diff --git a/drivers/net/bonding/bond_main.c b/drivers/net/bonding/bond_main.c
>index da3af63..9d94313 100644
>--- a/drivers/net/bonding/bond_main.c
>+++ b/drivers/net/bonding/bond_main.c
>@@ -723,7 +723,7 @@ static int bond_set_allmulti(struct bonding *bond, int inc)
> static void bond_resend_igmp_join_requests(struct bonding *bond)
> {
> 	if (!rtnl_trylock()) {
>-		queue_delayed_work(bond->wq, &bond->mcast_work, 0);
>+		queue_delayed_work(bond->wq, &bond->mcast_work, 1);
> 		return;
> 	}
> 	call_netdevice_notifiers(NETDEV_RESEND_IGMP, bond->dev);
>-- 
>1.8.1.4

---
	-Jay Vosburgh, IBM Linux Technology Center, fubar@us.ibm.com

--
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

On 07/30/2013 08:40 PM, Jay Vosburgh wrote:
> Nikolay Aleksandrov <nikolay@redhat.com> wrote:
> 
>> From: Nikolay Aleksandrov <Nikolay Aleksandrov nikolay@redhat.com>
>>
>> After commit 4aa5dee4d9 ("net: convert resend IGMP to notifier event")
>> we try to acquire rtnl in bond_resend_igmp_join_requests but it can be
>> scheduled with rtnl already held (e.g. when bond_change_active_slave is
>> called with rtnl) causing a loop of immediate reschedules + calls because
>> rtnl_trylock fails each time since it's being already held.
>> For me this issue leads to system hangs very easy:
>> modprobe bonding; ifconfig bond0 up; ifenslave bond0 eth0; rmmod
>> bonding;
> 
> 	I believe that bond_change_active_slave is always called with
> rtnl held, and it is the only caller of bond_resend_igmp_join_requests
> (well, "caller" in the sense that it queues the delayed work for
> mcast_work that runs the function, currently with delay of 0).
> 
>> The fix is to introduce a small (1 jiffy) delay which is enough for the
>> sections holding rtnl to finish without putting any strain on the system.
> 
> 	Should the delay also be in the bond_change_active_slave queue
> work call as well, to eliminate one loop of the "rtnl_trylock failing ->
> queue_delayed_work" sequence in bond_resend_igmp_join_requests?
> 
> 	-J
> 
Hi Jay,
I actually think there's one way to call bond_change_active_slave without rtnl:
bond_select_active_slave (and thus bond_change_active_slave) called from
bond_loadbalance_arp_mon with read_lock(&bond->lock) &
write_lock_bh(&bond->curr_slave_lock) only.

But you're right that most of the time (i.e. all other callers) it's called
with rtnl held, I will adjust its timer as well and send a v2.

Since I've prepared the initial RCU conversion of the bonding which I think
to submit tomorrow I don't want to make this patch a part of that series,
so I will submit it alone (and thus apart).
If this is not the protocol in such cases, or if there's anything else, I
can wait for it to go in and submit the series then, just let me know.

Thanks,
 Nik

--
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