On Tue, Aug 22, 2017 at 10:50:41AM +0800, Jason Wang wrote: > > Perhaps the descriptor pool should also be > > revised to allow out of order completions. Then there is no need to > > copy zerocopy packets whenever they may experience delay. > > Yes, but as replied in the referenced thread, windows driver may treat out > of order completion as a bug.
That would be a windows driver bug then, but I don't think it makes this assumption. What the referenced thread (https://patchwork.kernel.org/patch/3787671/) is saying is that host must use any buffers made available on a tx vq within a reasonable timeframe otherwise windows guests panic. Ideally we would detect that a packet is actually experiencing delay and trigger the copy at that point e.g. by calling skb_linearize. But it isn't easy to track these packets though and even harder to do a data copy without races. Which reminds me that skb_linearize in net core seems to be fundamentally racy - I suspect that if skb is cloned, and someone is trying to use the shared frags while another thread calls skb_linearize, we get some use after free bugs which likely mostly go undetected because the corrupted packets mostly go on wire and get dropped by checksum code. -- MST
