Re: [PATCH net] inet: bring NLM_DONE out to a separate recv() in inet_dump_ifaddr()

[Jakub Kicinski <kuba@kernel.org>]

On Sun, 2 Jun 2024 15:21:02 -0700 Jakub Kicinski wrote:
> Netlink is full of legacy behavior, the only way to make it usable
> in modern environments is to let new families not repeat the mistakes.
> That's why I'd really rather not add a workaround at the af_netlink
> level. Why would ethtool (which correctly coalesced NLM_DONE from day 1)
> suddenly start needed another recv(). A lot of the time the entire dump
> fits in one skb.
> 
> If you prefer to sacrifice all of rtnetlink (some of which, to be clear,
> has also been correctly coded from day 1) - we can add a trampoline for
> rtnetlink dump handlers?

Hi Eric, how do you feel about this approach? It would also let us
extract the "RTNL unlocked dump" handling from af_netlink.c, which
would be nice.

BTW it will probably need to be paired with fixing the
for_each_netdev_dump() foot gun, maybe (untested):

--- a/include/linux/netdevice.h
+++ b/include/linux/netdevice.h
@@ -3025,7 +3025,8 @@ int call_netdevice_notifiers_info(unsigned long val,
 #define net_device_entry(lh)	list_entry(lh, struct net_device, dev_list)
 
 #define for_each_netdev_dump(net, d, ifindex)				\
-	xa_for_each_start(&(net)->dev_by_index, (ifindex), (d), (ifindex))
+	for (; (d = xa_find(&(net)->dev_by_index, &ifindex,		\
+			    ULONG_MAX, XA_PRESENT)); ifindex++)
 
 static inline struct net_device *next_net_device(struct net_device *dev)
 {