Thread (12 messages) 12 messages, 3 authors, 5d ago

Re: [PATCH net-next v3 0/9] netconsole: stop charging netpoll users for netconsole-only data

From: Simon Horman <horms@kernel.org>
Date: 2026-07-16 13:21:53
Also in: lkml

On Fri, Jul 10, 2026 at 04:38:52AM -0700, Breno Leitao wrote:
This work continue to untangle netconsole and netpoll, improving
memory usage for netpoll users that are not netconsole.

struct netpoll is shared by every netpoll consumer in the tree:
bonding, bridge, team, vlan, macvlan, dsa and netconsole. The non-
netconsole users only need a per-port handle that lets them transmit
SKBs.

Yet struct netpoll has accumulated a pile of fields that exist only
to serve netconsole's printk path:

  - skb_pool / refill_wq: the fallback skb pool find_skb() falls back
    on when alloc_skb() returns NULL, plus the workqueue that refills
    it.
  - local_port / remote_port / remote_mac: the source and destination
    UDP ports and the destination ethernet address used to build
    outgoing log packets.

Every netpoll user that allocates a struct netpoll pays for all of
these fields, even though it never queues a single skb on the pool,
never sends a UDP frame and never looks at any of the addressing.

Move this netconsole-only state out of the shared struct into struct
netconsole_target, where its only consumer already lives.

After the series every passive netpoll consumer drops
sizeof(sk_buff_head) + sizeof(work_struct) + 2 * sizeof(u16) +
ETH_ALEN bytes per port (plus padding), and struct netpoll keeps only
what belongs to the generic poll/transmit abstraction or what netpoll
core still reads itself.

Follow-ups
----------
local_ip, remote_ip and ipv6 are still in struct netpoll. Moving the
addressing additionally requires relocating netpoll_setup() -- whose
only caller is netconsole -- and its IPv4/IPv6 helpers, so it is left
for a follow-up to keep this series focused on a single responsibility
transfer.

This is pure code motion with no functional change. The pre-existing
target/device teardown races reported against v1 have since been fixed
independently (as suggested by Jakub) and are already in the tree, so,
re-posting this one.

--

Changes in v3:
- Added a fix for a pre-existing cleanup-skip leak (released/joined targets).
- Added an skb-pool init-once patch closing an init/teardown race; rebased.
- Link to v2: https://lore.kernel.org/r/20260702-netconsole_move_more-v2-0-1ebedd921dcb@debian.org (local)

Changes in v2:
- Rebase on net-next, now that the prerequisite netconsole
  teardown-race fixes are merged.
- Squash the local_port and remote_port moves into a single patch.
- v1: https://lore.kernel.org/r/20260524-netconsole_move_more-v1-0-909d1ab398b4@debian.org (local)
For the series:

Reviewed-by: Simon Horman <horms@kernel.org>
Keyboard shortcuts
hback out one level
jnext message in thread
kprevious message in thread
ldrill in
Escclose help / fold thread tree
?toggle this help