On Thu, 5 Sep 2019 14:49:44 -0700
Alexander Duyck [off-list ref] wrote:
I would change the order of the tests you use here so that we can
eliminate the possibility of needing to perform many tests for the
more common cases. You could probably swap "list_skb" and "mss !=
GSO_BY_FRAGS" since list_skb is more likely to be false for many of
the common cases such as a standard TSO send from a socket. You might
even consider moving the GSO_BY_FRAGS check toward the end of your
checks since SCTP is the only protocol that I believe uses it and the
likelihood of encountering it is much lower compared to other
protocols.
You could probably test for !list_skb->head_frag before seeing if
there is a headlen since many NICs would be generating frames using
head_frag, so in the GRO case you mentioned above it could probably
save you some effort on a number of NICs.
You might also consider moving this code up before we push the mac
header back on and instead of setting sg to false you could just clear
the NETIF_F_SG flag from features. It would save you from having to
then remove doffset in your last check.
Thanks Alexander for the input. Will encorporate as much as possible
into a v2 patch.
BTW, I've strugged with myself regarding order of tests, and came
up with this suggestion, as my motivation was to have the tests order
tell a coherent logical story when read top-to-bottom left-to-right, FWIW.
For example, although 'mss != skb_headlen(head_skb)' could be tested
earlier, the condition by itself isn't meaningful unless we have an
existing frag_list and with a !head_frag.
Best
Shmulik