[net-next,v8,0/6] Improve neigh_flush_dev performance

Message

Gilad Naaman Nov. 4, 2024, 8:04 a.m. UTC

This patchsets improves the performance of neigh_flush_dev.

Currently, the only way to implement it requires traversing
all neighbours known to the kernel, across all network-namespaces.

This means that some flows are slowed down as a function of neigh-scale,
even if the specific link they're handling has little to no neighbours.

In order to solve this, this patchset adds a netdev->neighbours list,
as well as making the original linked-list doubly-, so that it is
possible to unlink neighbours without traversing the hash-bucket to
obtain the previous neighbour.

The original use-case we encountered was mass-deletion of links (12K
VLANs) while there are 50K ARPs and 50K NDPs in the system; though the
slowdowns would also appear when the links are set down.

Changes in v8:

 - Rebase and absorb alloc_pages->kvzalloc changes

Gilad Naaman (6):
  neighbour: Add hlist_node to struct neighbour
  neighbour: Define neigh_for_each_in_bucket
  neighbour: Convert seq_file functions to use hlist
  neighbour: Convert iteration to use hlist+macro
  neighbour: Remove bare neighbour::next pointer
  neighbour: Create netdev->neighbour association

 .../networking/net_cachelines/net_device.rst  |   1 +
 include/linux/netdevice.h                     |   7 +
 include/net/neighbour.h                       |  24 +-
 include/net/neighbour_tables.h                |  12 +
 net/core/neighbour.c                          | 325 ++++++++----------
 net/ipv4/arp.c                                |   2 +-
 6 files changed, 168 insertions(+), 203 deletions(-)
 create mode 100644 include/net/neighbour_tables.h