Merge branch 'udp_tunnel-add-NIC-RX-port-offload-infrastructure'

Jakub Kicinski says:

====================
udp_tunnel: add NIC RX port offload infrastructure

Kernel has a facility to notify drivers about the UDP tunnel ports
so that devices can recognize tunneled packets. This is important
mostly for RX - devices which don't support CHECKSUM_COMPLETE can
report checksums of inner packets, and compute RSS over inner headers.
Some drivers also match the UDP tunnel ports also for TX, although
doing so may lead to false positives and negatives.

Unfortunately the user experience when trying to take adavantage
of these facilities is suboptimal. First of all there is no way
for users to check which ports are offloaded. Many drivers resort
to printing messages to aid debugging, other use debugfs. Even worse
the availability of the RX features (NETIF_F_RX_UDP_TUNNEL_PORT)
is established purely on the basis of the driver having the ndos
installed. For most drivers, however, the ability to perform offloads
is contingent on device capabilities (driver support multiple device
and firmware versions). Unless driver resorts to hackish clearing
of features set incorrectly by the core - users are left guessing
whether their device really supports UDP tunnel port offload or not.

There is currently no way to indicate or configure whether RX
features include just the checksum offload or checksum and using
inner headers for RSS. Many drivers default to not using inner
headers for RSS because most implementations populate the source
port with entropy from the inner headers. This, however, is not
always the case, for example certain switches are only able to
use a fixed source port during encapsulation.

We have also seen many driver authors get the intricacies of UDP
tunnel port offloads wrong. Most commonly the drivers forget to
perform reference counting, or take sleeping locks in the callbacks.

This work tries to improve the situation by pulling the UDP tunnel
port table maintenance out of the drivers. It turns out that almost
all drivers maintain a fixed size table of ports (in most cases one
per tunnel type), so we can take care of all the refcounting in the
core, and let the driver specify if they need to sleep in the
callbacks or not. The new common implementation will also support
replacing ports - when a port is removed from a full table it will
try to find a previously missing port to take its place.

This patch only implements the core functionality along with a few
drivers I was hoping to test manually [1] along with a test based
on a netdevsim implementation. Following patches will convert all
the drivers. Once that's complete we can remove the ndos, and rely
directly on the new infrastrucutre.

Then after RSS (RXFH) is converted to netlink we can add the ability
to configure the use of inner RSS headers for UDP tunnels.

[1] Unfortunately I wasn't able to, turns out 2 of the devices
I had access to were older generation or had old FW, and they
did not actually support UDP tunnel port notifications (see
the second paragraph). The thrid device appears to program
the UDP ports correctly but it generates bad UDP checksums with
or without these patches. Long story short - I'd appreciate
reviews and testing here..

v4:
 - better build fix (hopefully this one does it..)
v3:
 - fix build issue;
 - improve bnxt changes.
====================

Signed-off-by: David S. Miller <davem@davemloft.net>

0 files changed, 0 insertions, 0 deletions