[net-next,v2,00/16] mlxsw: Improve IPv6 route insertion rate

From: Ido Schimmel <idosch@mellanox.com>

Unlike IPv4, an IPv6 multipath route in the kernel is composed from
multiple sibling routes, each representing a single nexthop.

Therefore, an addition of a multipath route with N nexthops translates
to N in-kernel notifications. This is inefficient for device drivers
that need to program the route to the underlying device. Each time a new
nexthop is appended, a new nexthop group needs to be constructed and the
old one deleted.

This patchset improves the situation by sending a single notification
for a multipath route addition / deletion instead of one per-nexthop.
When adding thousands of multipath routes with 16 nexthops, I measured
an improvement of about x10 in the insertion rate.

Patches #1-#3 add a flag that indicates that in-kernel notifications
need to be suppressed and extend the IPv6 FIB notification info with
information about the number of sibling routes that are being notified.

Patches #4-#5 adjust the two current listeners to these notifications to
ignore notifications about IPv6 multipath routes.

Patches #6-#7 adds add / delete notifications for IPv6 multipath routes.

Patches #8-#14 do the same for mlxsw.

Patch #15 finally removes the limitations added in patches #4-#5 and
stops the kernel from sending a notification for each added / deleted
nexthop.

Patch #16 adds test cases.

v2 (David Ahern):
* Remove patch adjusting netdevsim to consume resources for each
  fib6_info. Instead, consume one resource for the entire multipath
  route
* Remove 'multipath_rt' usage in patch #10
* Remove 'multipath_rt' from 'struct fib6_entry_notifier_info' in patch
  #15. The member is only removed in this patch to prevent drivers from
  processing multipath routes twice during the series

Ido Schimmel (16):
  netlink: Document all fields of 'struct nl_info'
  netlink: Add field to skip in-kernel notifications
  ipv6: Extend notifier info for multipath routes
  mlxsw: spectrum_router: Ignore IPv6 multipath notifications
  netdevsim: Ignore IPv6 multipath notifications
  ipv6: Add IPv6 multipath notifications for add / replace
  ipv6: Add IPv6 multipath notification for route delete
  mlxsw: spectrum_router: Remove processing of IPv6 append notifications
  mlxsw: spectrum_router: Prepare function to return errors
  mlxsw: spectrum_router: Pass multiple routes to work item
  mlxsw: spectrum_router: Adjust IPv6 replace logic to new notifications
  mlxsw: spectrum_router: Pass array of routes to route handling
    functions
  mlxsw: spectrum_router: Add / delete multiple IPv6 nexthops
  mlxsw: spectrum_router: Create IPv6 multipath routes in one go
  ipv6: Stop sending in-kernel notifications for each nexthop
  selftests: mlxsw: Add a test for FIB offload indication

 .../ethernet/mellanox/mlxsw/spectrum_router.c | 209 ++++++++---
 include/net/ip6_fib.h                         |   6 +
 include/net/netlink.h                         |   6 +-
 net/ipv6/ip6_fib.c                            |  44 ++-
 net/ipv6/route.c                              |  21 ++
 .../drivers/net/mlxsw/fib_offload.sh          | 349 ++++++++++++++++++
 6 files changed, 567 insertions(+), 68 deletions(-)
 create mode 100755 tools/testing/selftests/drivers/net/mlxsw/fib_offload.sh

From: Ido Schimmel <idosch@idosch.org>
Date: Tue, 18 Jun 2019 18:12:42 +0300

> Unlike IPv4, an IPv6 multipath route in the kernel is composed from
> multiple sibling routes, each representing a single nexthop.
> 
> Therefore, an addition of a multipath route with N nexthops translates
> to N in-kernel notifications. This is inefficient for device drivers
> that need to program the route to the underlying device. Each time a new
> nexthop is appended, a new nexthop group needs to be constructed and the
> old one deleted.
> 
> This patchset improves the situation by sending a single notification
> for a multipath route addition / deletion instead of one per-nexthop.
> When adding thousands of multipath routes with 16 nexthops, I measured
> an improvement of about x10 in the insertion rate.
 ...

Series applied, thanks everyone.