On 03/29/2017 05:29 PM, Eric Dumazet wrote:
On Wed, 2017-03-29 at 16:55 -0700, Tushar Dave wrote:
There are events seen where skb->queue_mapping value is greater than
adapter->num_tx_queue. In such cases, adapter->tx_ring becomes invalid
(null) and xmit results in kernel panic.
One such event is running netconsole and enabling VF on the same device.
Or running netconsole and changing number of tx queues via ethtool on
same device.
This patch adds check for skb->queue_mapping and uses simple arithmetic
to select available tx queue from driver.
We are not going to fix all multi queue drivers, for something that
seems to be a generic issue higher in the stack.
I have fixed it in ixgbe,i40e drivers because I see Intel igb driver
does similar thing. But I agree that the core issue lies up in the stack.
For the information, I found that this issue only affecting skbs that
are coming from direct xmit path (netdev_start_xmit) that netconsole,
pktgen , af_packet etc,. uses. skbs coming from qdisc already have this
fixed in stack. (http://lists.openwall.net/netdev/2010/07/01/58)
Thanks.
-Tushar
