Re: copy on write for splice() from file to pipe?

On Fri, Feb 10, 2023 at 9:23 AM Linus Torvalds
[off-list ref] wrote:

And when it comes to networking, in general things like TCP checksums
etc should be ok even with data that isn't stable.  When doing things
by hand, networking should always use the "copy-and-checksum"
functions that do the checksum while copying (so even if the source
data changes, the checksum is going to be the checksum for the data
that was copied).

And in many (most?) smarter network cards, the card itself does the
checksum, again on the data as it is transferred from memory.

So it's not like "networking needs a stable source" is some really
_fundamental_ requirement for things like that to work.

But it may well be that we have situations where some network driver
does the checksumming separately from then copying the data.

Ok, so I decided to try to take a look.

Somebody who actually does networking (and drivers in particular)
should probably check this, but it *looks* like the IPv4 TCP case
(just to pick the ony I looked at) gores through
tcp_sendpage_locked(), which does

        if (!(sk->sk_route_caps & NETIF_F_SG))
                return sock_no_sendpage_locked(sk, page, offset, size, flags);

which basically says "if you can't handle fragmented socket buffers,
do that 'no_sendpage' case".

So that will basically end up just falling back to a kernel
'sendmsg()', which does a copy and then it's stable.

But for the networks that *can* handle fragmented socket buffers, it
then calls do_tcp_sendpages() instead, which just creates a skb
fragment of the page (with tcp_build_frag()).

I wonder if that case should just require NETIF_F_HW_CSUM?

              Linus