On Mon, Feb 1, 2021 at 8:42 PM Jakub Kicinski <kuba@xxxxxxxxxx> wrote: > > On Mon, 1 Feb 2021 08:14:31 -0800 Xie He wrote: > > On Mon, Feb 1, 2021 at 6:10 AM Julian Wiedmann <jwi@xxxxxxxxxxxxx> wrote: > > > This sounds a bit like you want skb_cow_head() ... ? > > > > Calling "skb_cow_head" before we call "skb_clone" would indeed solve > > the problem of writes to our clones affecting clones in other parts of > > the system. But since we are still writing to the skb after > > "skb_clone", it'd still be better to replace "skb_clone" with > > "skb_copy" to avoid interference between our own clones. > > Why call skb_cow_head() before skb_clone()? skb_cow_head should be > called before the data in skb head is modified. I'm assuming you're only > modifying "front" of the frame, right? skb_cow_head() should do nicely > in that case. The modification happens after skb_clone. If we call skb_cow_head after skb_clone (before the modification), then skb_cow_head would always see that the skb is a clone and would always copy it. Therefore skb_clone + skb_cow_head is equivalent to skb_copy.
