Currently, when inserting a new filter that needs to sit at the head
of chain 0, it will first update the heads pointer on all devices using
the (shared) block, and only then complete the initialization of the new
element so that it has a "next" element.

This can lead to a situation that the chain 0 head is propagated to
another CPU before the "next" initialization is done. When this race
condition is triggered, packets being matched on that CPU will simply
miss all other filters, and will flow through the stack as if there were
no other filters installed. If the system is using OVS + TC, such
packets will get handled by vswitchd via upcall, which results in much
higher latency and reordering. For other applications it may result in
packet drops.

This is reproducible with a tc only setup, but it varies from system to
system. It could be reproduced with a shared block amongst 10 veth
tunnels, and an ingress filter mirroring packets to another veth.
That's because using the last added veth tunnel to the shared block to
do the actual traffic, it makes the race window bigger and easier to
trigger.

The fix is rather simple, to just initialize the next pointer of the new
filter instance (tp) before propagating the head change.

The fixes tag is pointing to the original code though this issue should
only be observed when using it unlocked.

Fixes: 2190d1d0 ("net: sched: introduce helpers to work with filter chains")
Signed-off-by: Marcelo Ricardo Leitner <marcelo.leitner@gmail.com>
Signed-off-by: Vlad Buslov <vladbu@nvidia.com>
Reviewed-by: Davide Caratti <dcaratti@redhat.com>
Link: https://lore.kernel.org/r/b97d5f4eaffeeb9d058155bcab63347527261abf.1649341369.git.marcelo.leitner@gmail.comSigned-off-by: Jakub Kicinski <kuba@kernel.org>