Re: [RFC][PATCHSET] more iov_iter conversion in net/*

[David Miller <davem@davemloft.net>]

From: Al Viro <viro@ZenIV.linux.org.uk>
Date: Mon, 2 Feb 2015 06:53:39 +0000

> Point...  Are you OK with tcp_send_syn_data() change in 8/18?  Basically,
> tcp_sendmsg() treats short copy from userland as "send as much as we can,
> stop at the first point where copy_from_user() fails", same as write(),
> etc.; that much is unchanged, but the current mainline has a strange behaviour
> in case when short copy happens within what would be packed into SYN packet -
> if that happens, it still sends as much as possible, but it falls back to
> separate SYN.  With this patch it simply sends shorter SYN+data packet instead.
> The reason I went that way is that I wanted to avoid copying the same data
> from userland twice and it was easy to do; I can preserve the current mainline
> behaviour, but it'll cost making a backup copy of ->msg_iter in
> tcp_send_syn_data(), only to never use it in normal case *and* do piles
> of extra work (extra packet to send, repeated copying from userland) in case
> of short copy.  Is there any reason not to combine SYN+data in case of short
> copy?  Again, we do send exact same data, return the same value, etc. - you
> need tcpdump to see the difference.

I think dropping down to plain SYN in the short copy case was just based
upon a blanket decision to not do SYN+DATA if anything out of the
ordinary happens, rathe than an explicit policy decision.  So that change
should be OK.