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

[David Ahern <dsahern@kernel.org>]

On 4/11/24 12:02 PM, Jakub Kicinski wrote:
> Commit under Fixes optimized the number of recv() calls
> needed during RTM_GETROUTE dumps, but we got multiple
> reports of applications hanging on recv() calls.
> Applications expect that a route dump will be terminated
> with a recv() reading an individual NLM_DONE message.
> 
> Coalescing NLM_DONE is perfectly legal in netlink,
> but even tho reporters fixed the code in respective
> projects, chances are it will take time for those
> applications to get updated. So revert to old behavior
> (for now)?
> 
> Old kernel (5.19):
> 
>  $ ./cli.py --dbg-small-recv 4096 --spec netlink/specs/rt_route.yaml \
>             --dump getroute --json '{"rtm-family": 2}'
>  Recv: read 692 bytes, 11 messages
>    nl_len = 68 (52) nl_flags = 0x22 nl_type = 24
>  ...
>    nl_len = 60 (44) nl_flags = 0x22 nl_type = 24
>  Recv: read 20 bytes, 1 messages
>    nl_len = 20 (4) nl_flags = 0x2 nl_type = 3
> 
> Before (6.9-rc2):
> 
>  $ ./cli.py --dbg-small-recv 4096 --spec netlink/specs/rt_route.yaml \
>             --dump getroute --json '{"rtm-family": 2}'
>  Recv: read 712 bytes, 12 messages
>    nl_len = 68 (52) nl_flags = 0x22 nl_type = 24
>  ...
>    nl_len = 60 (44) nl_flags = 0x22 nl_type = 24
>    nl_len = 20 (4) nl_flags = 0x2 nl_type = 3
> 
> After:
> 
>  $ ./cli.py --dbg-small-recv 4096 --spec netlink/specs/rt_route.yaml \
>             --dump getroute --json '{"rtm-family": 2}'
>  Recv: read 692 bytes, 11 messages
>    nl_len = 68 (52) nl_flags = 0x22 nl_type = 24
>  ...
>    nl_len = 60 (44) nl_flags = 0x22 nl_type = 24
>  Recv: read 20 bytes, 1 messages
>    nl_len = 20 (4) nl_flags = 0x2 nl_type = 3
> 
> Reported-by: Stefano Brivio <sbrivio@redhat.com>
> Link: https://lore.kernel.org/all/20240315124808.033ff58d@elisabeth
> Reported-by: Ilya Maximets <i.maximets@ovn.org>
> Link: https://lore.kernel.org/all/02b50aae-f0e9-47a4-8365-a977a85975d3@ovn.org
> Fixes: 4ce5dc9316de ("inet: switch inet_dump_fib() to RCU protection")
> Signed-off-by: Jakub Kicinski <kuba@kernel.org>
> ---
> CC: dsahern@kernel.org
> CC: donald.hunter@gmail.com
> ---
>  net/ipv4/fib_frontend.c | 5 +++++
>  1 file changed, 5 insertions(+)
> 
> diff --git a/net/ipv4/fib_frontend.c b/net/ipv4/fib_frontend.c
> index 48741352a88a..c484b1c0fc00 100644
> --- a/net/ipv4/fib_frontend.c
> +++ b/net/ipv4/fib_frontend.c
> @@ -1050,6 +1050,11 @@ static int inet_dump_fib(struct sk_buff *skb, struct netlink_callback *cb)
>  			e++;
>  		}
>  	}
> +
> +	/* Don't let NLM_DONE coalesce into a message, even if it could.
> +	 * Some user space expects NLM_DONE in a separate recv().

that's unfortunate

> +	 */
> +	err = skb->len;
>  out:
>  
>  	cb->args[1] = e;


Reviewed-by: David Ahern <dsahern@kernel.org>