David Miller <da...@davemloft.net> wrote: > Why don't you just do an skb_unshare() at the beginning when you know that > you'll need to do that?
I was trying to defer any copying to process context rather than doing it in softirq context to spend less time in softirq context - plus that way I can use GFP_NOIO (kafs) or GFP_KERNEL (direct AF_RXRPC socket) rather than GFP_ATOMIC if the api supports it. I don't remember now why I used skb_cow_data() rather than skb_unshare() - but it was probably because the former leaves the sk_buff object itself intact, whereas the latter replaces it. I can switch to using skb_unshare() instead. Question for you: how likely is a newly received buffer, through a UDP socket, to be 'cloned'? David
