2013-08-15 08:37:11

by Tao Ma

Subject: [PATCH] bond: Don't set skb->queue_mapping in netpoll.

From: Tao Ma <[email protected]>

When we are using netpoll, we don't go through the normal
transmit process. In this case, bond_select_queue is not called
and qdisc_skb_cb(skb)->slave_dev_queue_mapping isn't set.

So when netpoll_send_skb_on_dev calls ndo_start_xmit and we
enter bond_dev_queue_xmit, we will set skb->queue_mapping to
an invalid value and in some cases cause the driver panic the
kernel(We meet with bnx2 panic because of a very large queue_mapping).

This patch skip skb->queue_mapping if we find we are in netpoll.

CC: "David S. Miller" <[email protected]>
CC: Cong Wang <[email protected]>
CC: Eric Dumazet <[email protected]>
Signed-off-by: Tao Ma <[email protected]>
---
drivers/net/bonding/bond_main.c | 5 +++--
1 files changed, 3 insertions(+), 2 deletions(-)

diff --git a/drivers/net/bonding/bond_main.c b/drivers/net/bonding/bond_main.c
index 07f257d..97b2f52 100644
--- a/drivers/net/bonding/bond_main.c
+++ b/drivers/net/bonding/bond_main.c
@@ -405,12 +405,13 @@ int bond_dev_queue_xmit(struct bonding *bond, struct sk_buff *skb,

BUILD_BUG_ON(sizeof(skb->queue_mapping) !=
sizeof(qdisc_skb_cb(skb)->slave_dev_queue_mapping));
- skb->queue_mapping = qdisc_skb_cb(skb)->slave_dev_queue_mapping;

if (unlikely(netpoll_tx_running(bond->dev)))
bond_netpoll_send_skb(bond_get_slave_by_dev(bond, slave_dev), skb);
- else
+ else {
+ skb->queue_mapping = qdisc_skb_cb(skb)->slave_dev_queue_mapping;
dev_queue_xmit(skb);
+ }

return 0;
}
--
1.7.0.4

2013-08-15 10:35:46

by Cong Wang

Subject: Re: [PATCH] bond: Don't set skb->queue_mapping in netpoll.

On Thu, 2013-08-15 at 16:36 +0800, Tao Ma wrote:
> From: Tao Ma <[email protected]>
>
> When we are using netpoll, we don't go through the normal
> transmit process. In this case, bond_select_queue is not called
> and qdisc_skb_cb(skb)->slave_dev_queue_mapping isn't set.
>

Yes?

In netpoll_send_skb_on_dev() we have:

if (skb_queue_len(&npinfo->txq) == 0 && !
netpoll_owner_active(dev)) {
struct netdev_queue *txq;

txq = netdev_pick_tx(dev, skb);
...

where netdev_pick_tx() will call ->ndo_select_queue().

2013-08-15 21:07:51

by Sergei Shtylyov

Subject: Re: [PATCH] bond: Don't set skb->queue_mapping in netpoll.

On 08/15/2013 12:36 PM, Tao Ma wrote:

> From: Tao Ma <[email protected]>

> When we are using netpoll, we don't go through the normal
> transmit process. In this case, bond_select_queue is not called
> and qdisc_skb_cb(skb)->slave_dev_queue_mapping isn't set.

> So when netpoll_send_skb_on_dev calls ndo_start_xmit and we
> enter bond_dev_queue_xmit, we will set skb->queue_mapping to
> an invalid value and in some cases cause the driver panic the
> kernel(We meet with bnx2 panic because of a very large queue_mapping).

> This patch skip skb->queue_mapping if we find we are in netpoll.

> CC: "David S. Miller" <[email protected]>
> CC: Cong Wang <[email protected]>
> CC: Eric Dumazet <[email protected]>
> Signed-off-by: Tao Ma <[email protected]>
> ---
> drivers/net/bonding/bond_main.c | 5 +++--
> 1 files changed, 3 insertions(+), 2 deletions(-)

> diff --git a/drivers/net/bonding/bond_main.c b/drivers/net/bonding/bond_main.c
> index 07f257d..97b2f52 100644
> --- a/drivers/net/bonding/bond_main.c
> +++ b/drivers/net/bonding/bond_main.c
> @@ -405,12 +405,13 @@ int bond_dev_queue_xmit(struct bonding *bond, struct sk_buff *skb,
>
> BUILD_BUG_ON(sizeof(skb->queue_mapping) !=
> sizeof(qdisc_skb_cb(skb)->slave_dev_queue_mapping));
> - skb->queue_mapping = qdisc_skb_cb(skb)->slave_dev_queue_mapping;
>
> if (unlikely(netpoll_tx_running(bond->dev)))
> bond_netpoll_send_skb(bond_get_slave_by_dev(bond, slave_dev), skb);
> - else
> + else {
> + skb->queue_mapping = qdisc_skb_cb(skb)->slave_dev_queue_mapping;
> dev_queue_xmit(skb);
> + }

According to Documentation/CodingStyle, both arms of the *if* statement
should have {} if one arm has them.

WBR, Sergei