skb length without fragments

skb length without fragments

[pradeep kumar nalla]

Hi

While testing my network driver with pktgen I could see an skb greater
than 16K without fragments in xmit function. This lead to a fix in my
driver that assumes when an SKB whose length is greater than 16K will
come with fragments. Apart from pktgen what are the chances or
possibilities of getting an SKB greater than 16K without fragments? .
When I tried with tools like iperf/iper3/netperf, didn’t see a single
incidence where the SKB length is greater than 16K and without frags.
Even socket layer I see function alloc_skb_with_frags, does this mean
all the larger packets come with frags.

Please pardon if it is not a correct mailing list to ask above question.

Thanks
Pradeep.

Re: skb length without fragments

[Eric Dumazet]

On 10/08/2018 11:02 AM, pradeep kumar nalla wrote:
> Hi
> 
> While testing my network driver with pktgen I could see an skb greater
> than 16K without fragments in xmit function. This lead to a fix in my
> driver that assumes when an SKB whose length is greater than 16K will
> come with fragments. Apart from pktgen what are the chances or
> possibilities of getting an SKB greater than 16K without fragments? .
> When I tried with tools like iperf/iper3/netperf, didn’t see a single
> incidence where the SKB length is greater than 16K and without frags.
> Even socket layer I see function alloc_skb_with_frags, does this mean
> all the larger packets come with frags.
>

There are cases where skb_linearize() calls happen, and then certainly
can feed a driver with a big linear skb.

Even if current tree would not hit this, you have to play safe and
do the check in the driver.