On 20.10.2018 00:25, Eric Dumazet wrote:

On 10/19/2018 02:58 PM, Eric Dumazet wrote:

quoted

On 10/16/2018 06:00 AM, Eric Dumazet wrote:

quoted

On Mon, Oct 15, 2018 at 11:30 PM Andre Tomt [off-list ref] wrote:

quoted

I've seen similar on several systems with mlx4 cards when using 4.18.x -
that is hw csum failure followed by some backtrace.

Only seems to happen on systems dealing with quite a bit of UDP.

Strange, because mlx4 on IPv6+UDP should not use CHECKSUM_COMPLETE,
but CHECKSUM_UNNECESSARY

I would be nice to track this a bit further, maybe by providing the
full packet content.

<snip>

quoted

As a matter of fact Dimitris found the issue in the patch and is working on a fix involving csum_block_sub()

Problems comes from trimming an odd number of bytes.

More exactly, trimming bytes starting at an odd offset.

No hw csum failures here since I deployed Dimitris fix on top of 4.18.16 
32 hours ago.

Thanks