[PATCH 1/4] bonding: don't call slave_xxx_netpoll under spinlocks

[PATCH 1/4] bonding: don't call slave_xxx_netpoll under spinlocks

[Ding Tianhong]

the slave_xxx_netpoll may sleep, so it should't be called under spinlocks.

the slave point of the bonding will not be changed outside rtnl lock,
so rtnl lock is enough here.

Signed-off-by: Ding Tianhong <dingtianhong@huawei.com>
---
 drivers/net/bonding/bond_main.c | 15 +++------------
 1 file changed, 3 insertions(+), 12 deletions(-)

diff --git a/drivers/net/bonding/bond_main.c b/drivers/net/bonding/bond_main.c
index 07f257d4..5eb75ef 100644
--- a/drivers/net/bonding/bond_main.c
+++ b/drivers/net/bonding/bond_main.c
@@ -1249,8 +1249,9 @@ static void bond_poll_controller(struct net_device *bond_dev)
 {
 }
 
-static void __bond_netpoll_cleanup(struct bonding *bond)
+static void bond_netpoll_cleanup(struct net_device *bond_dev)
 {
+	struct bonding *bond = netdev_priv(bond_dev);
 	struct slave *slave;
 	int i;
 
@@ -1258,14 +1259,6 @@ static void __bond_netpoll_cleanup(struct bonding *bond)
 		if (IS_UP(slave->dev))
 			slave_disable_netpoll(slave);
 }
-static void bond_netpoll_cleanup(struct net_device *bond_dev)
-{
-	struct bonding *bond = netdev_priv(bond_dev);
-
-	read_lock(&bond->lock);
-	__bond_netpoll_cleanup(bond);
-	read_unlock(&bond->lock);
-}
 
 static int bond_netpoll_setup(struct net_device *dev, struct netpoll_info *ni, gfp_t gfp)
 {
@@ -1273,15 +1266,13 @@ static int bond_netpoll_setup(struct net_device *dev, struct netpoll_info *ni, g
 	struct slave *slave;
 	int i, err = 0;
 
-	read_lock(&bond->lock);
 	bond_for_each_slave(bond, slave, i) {
 		err = slave_enable_netpoll(slave);
 		if (err) {
-			__bond_netpoll_cleanup(bond);
+			bond_netpoll_cleanup(dev);
 			break;
 		}
 	}
-	read_unlock(&bond->lock);
 	return err;
 }
 
-- 
1.8.0

Re: [1/4] bonding: don't call slave_xxx_netpoll under spinlocks

[Veaceslav Falico]

On Sat, Jul 20, 2013 at 03:23:47PM +0800, dingtianhong wrote:
>the slave_xxx_netpoll may sleep, so it should't be called under spinlocks.

I don't really see how it may sleep, it was specifically changed to not
sleep actually. However, see below...

>
>the slave point of the bonding will not be changed outside rtnl lock,
>so rtnl lock is enough here.

Yep, as far as I see there's really no need to take the lock, both the
slave list and the netpoll part are always protected by rtnl lock, unless
I'm missing something, and indeed .ndo_netpoll_setup() is always called
under rtnl.

BTW, bond_netpoll_cleanup() has the same problem - maybe you could check if
we can remove the bond->lock from there also and update the patch?

>
>Signed-off-by: Ding Tianhong <dingtianhong@huawei.com>
>
>---
>drivers/net/bonding/bond_main.c | 15 +++------------
> 1 file changed, 3 insertions(+), 12 deletions(-)
>
>diff --git a/drivers/net/bonding/bond_main.c b/drivers/net/bonding/bond_main.c
>index 07f257d4..5eb75ef 100644
>--- a/drivers/net/bonding/bond_main.c
>+++ b/drivers/net/bonding/bond_main.c
>@@ -1249,8 +1249,9 @@ static void bond_poll_controller(struct net_device *bond_dev)
> {
> }
>
>-static void __bond_netpoll_cleanup(struct bonding *bond)
>+static void bond_netpoll_cleanup(struct net_device *bond_dev)
> {
>+	struct bonding *bond = netdev_priv(bond_dev);
> 	struct slave *slave;
> 	int i;
>
>@@ -1258,14 +1259,6 @@ static void __bond_netpoll_cleanup(struct bonding *bond)
> 		if (IS_UP(slave->dev))
> 			slave_disable_netpoll(slave);
> }
>-static void bond_netpoll_cleanup(struct net_device *bond_dev)
>-{
>-	struct bonding *bond = netdev_priv(bond_dev);
>-
>-	read_lock(&bond->lock);
>-	__bond_netpoll_cleanup(bond);
>-	read_unlock(&bond->lock);
>-}
>
> static int bond_netpoll_setup(struct net_device *dev, struct netpoll_info *ni, gfp_t gfp)
> {
>@@ -1273,15 +1266,13 @@ static int bond_netpoll_setup(struct net_device *dev, struct netpoll_info *ni, g
> 	struct slave *slave;
> 	int i, err = 0;
>
>-	read_lock(&bond->lock);
> 	bond_for_each_slave(bond, slave, i) {
> 		err = slave_enable_netpoll(slave);
> 		if (err) {
>-			__bond_netpoll_cleanup(bond);
>+			bond_netpoll_cleanup(dev);
> 			break;
> 		}
> 	}
>-	read_unlock(&bond->lock);
> 	return err;
> }
>

Re: [1/4] bonding: don't call slave_xxx_netpoll under spinlocks

[Ding Tianhong]

On 2013/7/20 18:38, Veaceslav Falico wrote:
> On Sat, Jul 20, 2013 at 03:23:47PM +0800, dingtianhong wrote:
>> the slave_xxx_netpoll may sleep, so it should't be called under spinlocks.
> 
> I don't really see how it may sleep, it was specifically changed to not
> sleep actually. However, see below...
>

I think the synchronize_rcu_bh() in slave disable_netpoll will sched and speed,so spinlock
should not used here.

>>
>> the slave point of the bonding will not be changed outside rtnl lock,
>> so rtnl lock is enough here.
> 
> Yep, as far as I see there's really no need to take the lock, both the
> slave list and the netpoll part are always protected by rtnl lock, unless
> I'm missing something, and indeed .ndo_netpoll_setup() is always called
> under rtnl.
> 
> BTW, bond_netpoll_cleanup() has the same problem - maybe you could check if
> we can remove the bond->lock from there also and update the patch?
> 

yes, this patch has remove bond_netpoll_cleanup(), and change _bond_netpoll_cleanup() to bond_netpoll_cleanup(), rtnl lock is enough here.

>>
>> Signed-off-by: Ding Tianhong <dingtianhong@huawei.com>
>>
>> ---
>> drivers/net/bonding/bond_main.c | 15 +++------------
>> 1 file changed, 3 insertions(+), 12 deletions(-)
>>
>> diff --git a/drivers/net/bonding/bond_main.c b/drivers/net/bonding/bond_main.c
>> index 07f257d4..5eb75ef 100644
>> --- a/drivers/net/bonding/bond_main.c
>> +++ b/drivers/net/bonding/bond_main.c
>> @@ -1249,8 +1249,9 @@ static void bond_poll_controller(struct net_device *bond_dev)
>> {
>> }
>>
>> -static void __bond_netpoll_cleanup(struct bonding *bond)
>> +static void bond_netpoll_cleanup(struct net_device *bond_dev)
>> {
>> +    struct bonding *bond = netdev_priv(bond_dev);
>>     struct slave *slave;
>>     int i;
>>
>> @@ -1258,14 +1259,6 @@ static void __bond_netpoll_cleanup(struct bonding *bond)
>>         if (IS_UP(slave->dev))
>>             slave_disable_netpoll(slave);
>> }
>> -static void bond_netpoll_cleanup(struct net_device *bond_dev)
>> -{
>> -    struct bonding *bond = netdev_priv(bond_dev);
>> -
>> -    read_lock(&bond->lock);
>> -    __bond_netpoll_cleanup(bond);
>> -    read_unlock(&bond->lock);
>> -}
>>
>> static int bond_netpoll_setup(struct net_device *dev, struct netpoll_info *ni, gfp_t gfp)
>> {
>> @@ -1273,15 +1266,13 @@ static int bond_netpoll_setup(struct net_device *dev, struct netpoll_info *ni, g
>>     struct slave *slave;
>>     int i, err = 0;
>>
>> -    read_lock(&bond->lock);
>>     bond_for_each_slave(bond, slave, i) {
>>         err = slave_enable_netpoll(slave);
>>         if (err) {
>> -            __bond_netpoll_cleanup(bond);
>> +            bond_netpoll_cleanup(dev);
>>             break;
>>         }
>>     }
>> -    read_unlock(&bond->lock);
>>     return err;
>> }
>>
> 
> .
> 

Re: [PATCH 1/4] bonding: don't call slave_xxx_netpoll under spinlocks

[David Miller]

I'm waiting for a reposting of this entire series with the requested changes
made to patch #1.

Thanks.