skb_reset: do not free socket memory with preemption disabled

It turns out that calling skb_orphan (via skb_scrub_packet(xnet=true))
might result in a call to schedule(), since it will wake up waiting
socket writers in userspace. This means that everything explodes, since
we're calling this with preemption disabled.

Work around this by removing the spinlocks in the consumers -- they're
the sole consumers and so don't need the spinlocks anyway -- and then
move skb_orphan to the consumption thread, which doesn't have preemption
disabled. (When we move to a proper lockless mpmc ring buffer, we won't
even have spinlocks anyway.)

Since we don't actually need to call skb_orphan except on tx, so that
net_cls does the right thing (see net/netfilter/xt_cgroup.c and its
check of skb->sk), just explictly orphan the packet in the one place
where that's actually necessary, and otherwise don't orphan it. This
leaves us without a call to skb_orphan on the rx path, where we don't
need it and can't use it anyway because preemption is disabled for
netif_receive_skb.

This effectively reworks 740319127f14793a13ad385e8150cd98c715c20c.

0 files changed, 0 insertions, 0 deletions